1 /* regcomp.c
2 */
3
4 /*
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"]
8 */
9
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
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.
18 */
19
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
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
32 */
33
34 #ifdef PERL_EXT_RE_BUILD
35 #include "re_top.h"
36 #endif
37
38 /*
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
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 ****
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
64 ****
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
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
74 /* Note on debug output:
75 *
76 * This is set up so that -Dr turns on debugging like all other flags that are
77 * enabled by -DDEBUGGING. -Drv gives more verbose output. This applies to
78 * all regular expressions encountered in a program, and gives a huge amount of
79 * output for all but the shortest programs.
80 *
81 * The ability to output pattern debugging information lexically, and with much
82 * finer grained control was added, with 'use re qw(Debug ....);' available even
83 * in non-DEBUGGING builds. This is accomplished by copying the contents of
84 * regcomp.c to ext/re/re_comp.c, and regexec.c is copied to ext/re/re_exec.c.
85 * Those files are compiled and linked into the perl executable, and they are
86 * compiled essentially as if DEBUGGING were enabled, and controlled by calls
87 * to re.pm.
88 *
89 * That would normally mean linking errors when two functions of the same name
90 * are attempted to be placed into the same executable. That is solved in one
91 * of four ways:
92 * 1) Static functions aren't known outside the file they are in, so for the
93 * many functions of that type in this file, it just isn't a problem.
94 * 2) Most externally known functions are enclosed in
95 * #ifndef PERL_IN_XSUB_RE
96 * ...
97 * #endif
98 * blocks, so there is only one defintion for them in the whole
99 * executable, the one in regcomp.c (or regexec.c). The implication of
100 * that is any debugging info that comes from them is controlled only by
101 * -Dr. Further, any static function they call will also be the version
102 * in regcomp.c (or regexec.c), so its debugging will also be by -Dr.
103 * 3) About a dozen external functions are re-#defined in ext/re/re_top.h, to
104 * have different names, so that what gets loaded in the executable is
105 * 'Perl_foo' from regcomp.c (and regexec.c), and the identical function
106 * from re_comp.c (and re_exec.c), but with the name 'my_foo' Debugging
107 * in the 'Perl_foo' versions is controlled by -Dr, but the 'my_foo'
108 * versions and their callees are under control of re.pm. The catch is
109 * that references to all these go through the regexp_engine structure,
110 * which is initialized in regcomp.h to the Perl_foo versions, and
111 * substituted out in lexical scopes where 'use re' is in effect to the
112 * 'my_foo' ones. That structure is public API, so it would be a hard
113 * sell to add any additional members.
114 * 4) For functions in regcomp.c and re_comp.c that are called only from,
115 * respectively, regexec.c and re_exec.c, they can have two different
116 * names, depending on #ifdef'ing PERL_IN_XSUB_RE, in both regexec.c and
117 * embed.fnc.
118 *
119 * The bottom line is that if you add code to one of the public functions
120 * listed in ext/re/re_top.h, debugging automagically works. But if you write
121 * a new function that needs to do debugging or there is a chain of calls from
122 * it that need to do debugging, all functions in the chain should use options
123 * 2) or 4) above.
124 *
125 * A function may have to be split so that debugging stuff is static, but it
126 * calls out to some other function that only gets compiled in regcomp.c to
127 * access data that we don't want to duplicate.
128 */
129
130 #include "EXTERN.h"
131 #define PERL_IN_REGCOMP_C
132 #include "perl.h"
133
134 #define REG_COMP_C
135 #ifdef PERL_IN_XSUB_RE
136 # include "re_comp.h"
137 EXTERN_C const struct regexp_engine my_reg_engine;
138 EXTERN_C const struct regexp_engine wild_reg_engine;
139 #else
140 # include "regcomp.h"
141 #endif
142
143 #include "invlist_inline.h"
144 #include "unicode_constants.h"
145
146 #ifndef STATIC
147 #define STATIC static
148 #endif
149
150 /* this is a chain of data about sub patterns we are processing that
151 need to be handled separately/specially in study_chunk. Its so
152 we can simulate recursion without losing state. */
153 struct scan_frame;
154 typedef struct scan_frame {
155 regnode *last_regnode; /* last node to process in this frame */
156 regnode *next_regnode; /* next node to process when last is reached */
157 U32 prev_recursed_depth;
158 I32 stopparen; /* what stopparen do we use */
159 bool in_gosub; /* this or an outer frame is for GOSUB */
160
161 struct scan_frame *this_prev_frame; /* this previous frame */
162 struct scan_frame *prev_frame; /* previous frame */
163 struct scan_frame *next_frame; /* next frame */
164 } scan_frame;
165
166 /* Certain characters are output as a sequence with the first being a
167 * backslash. */
168 #define isBACKSLASHED_PUNCT(c) memCHRs("-[]\\^", c)
169
170
171 struct RExC_state_t {
172 U32 flags; /* RXf_* are we folding, multilining? */
173 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
174 char *precomp; /* uncompiled string. */
175 char *precomp_end; /* pointer to end of uncompiled string. */
176 REGEXP *rx_sv; /* The SV that is the regexp. */
177 regexp *rx; /* perl core regexp structure */
178 regexp_internal *rxi; /* internal data for regexp object
179 pprivate field */
180 char *start; /* Start of input for compile */
181 char *end; /* End of input for compile */
182 char *parse; /* Input-scan pointer. */
183 char *copy_start; /* start of copy of input within
184 constructed parse string */
185 char *save_copy_start; /* Provides one level of saving
186 and restoring 'copy_start' */
187 char *copy_start_in_input; /* Position in input string
188 corresponding to copy_start */
189 SSize_t whilem_seen; /* number of WHILEM in this expr */
190 regnode *emit_start; /* Start of emitted-code area */
191 regnode_offset emit; /* Code-emit pointer */
192 I32 naughty; /* How bad is this pattern? */
193 I32 sawback; /* Did we see \1, ...? */
194 SSize_t size; /* Number of regnode equivalents in
195 pattern */
196 Size_t sets_depth; /* Counts recursion depth of already-
197 compiled regex set patterns */
198 U32 seen;
199
200 I32 parens_buf_size; /* #slots malloced open/close_parens */
201 regnode_offset *open_parens; /* offsets to open parens */
202 regnode_offset *close_parens; /* offsets to close parens */
203 HV *paren_names; /* Paren names */
204
205 /* position beyond 'precomp' of the warning message furthest away from
206 * 'precomp'. During the parse, no warnings are raised for any problems
207 * earlier in the parse than this position. This works if warnings are
208 * raised the first time a given spot is parsed, and if only one
209 * independent warning is raised for any given spot */
210 Size_t latest_warn_offset;
211
212 I32 npar; /* Capture buffer count so far in the
213 parse, (OPEN) plus one. ("par" 0 is
214 the whole pattern)*/
215 I32 total_par; /* During initial parse, is either 0,
216 or -1; the latter indicating a
217 reparse is needed. After that pass,
218 it is what 'npar' became after the
219 pass. Hence, it being > 0 indicates
220 we are in a reparse situation */
221 I32 nestroot; /* root parens we are in - used by
222 accept */
223 I32 seen_zerolen;
224 regnode *end_op; /* END node in program */
225 I32 utf8; /* whether the pattern is utf8 or not */
226 I32 orig_utf8; /* whether the pattern was originally in utf8 */
227 /* XXX use this for future optimisation of case
228 * where pattern must be upgraded to utf8. */
229 I32 uni_semantics; /* If a d charset modifier should use unicode
230 rules, even if the pattern is not in
231 utf8 */
232
233 I32 recurse_count; /* Number of recurse regops we have generated */
234 regnode **recurse; /* Recurse regops */
235 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
236 through */
237 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
238 I32 in_lookaround;
239 I32 contains_locale;
240 I32 override_recoding;
241 I32 recode_x_to_native;
242 I32 in_multi_char_class;
243 int code_index; /* next code_blocks[] slot */
244 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
245 within pattern */
246 SSize_t maxlen; /* mininum possible number of chars in string to match */
247 scan_frame *frame_head;
248 scan_frame *frame_last;
249 U32 frame_count;
250 AV *warn_text;
251 HV *unlexed_names;
252 SV *runtime_code_qr; /* qr with the runtime code blocks */
253 #ifdef DEBUGGING
254 const char *lastparse;
255 I32 lastnum;
256 U32 study_chunk_recursed_count;
257 AV *paren_name_list; /* idx -> name */
258 SV *mysv1;
259 SV *mysv2;
260
261 #define RExC_lastparse (pRExC_state->lastparse)
262 #define RExC_lastnum (pRExC_state->lastnum)
263 #define RExC_paren_name_list (pRExC_state->paren_name_list)
264 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
265 #define RExC_mysv (pRExC_state->mysv1)
266 #define RExC_mysv1 (pRExC_state->mysv1)
267 #define RExC_mysv2 (pRExC_state->mysv2)
268
269 #endif
270 bool seen_d_op;
271 bool strict;
272 bool study_started;
273 bool in_script_run;
274 bool use_BRANCHJ;
275 bool sWARN_EXPERIMENTAL__VLB;
276 bool sWARN_EXPERIMENTAL__REGEX_SETS;
277 };
278
279 #define RExC_flags (pRExC_state->flags)
280 #define RExC_pm_flags (pRExC_state->pm_flags)
281 #define RExC_precomp (pRExC_state->precomp)
282 #define RExC_copy_start_in_input (pRExC_state->copy_start_in_input)
283 #define RExC_copy_start_in_constructed (pRExC_state->copy_start)
284 #define RExC_save_copy_start_in_constructed (pRExC_state->save_copy_start)
285 #define RExC_precomp_end (pRExC_state->precomp_end)
286 #define RExC_rx_sv (pRExC_state->rx_sv)
287 #define RExC_rx (pRExC_state->rx)
288 #define RExC_rxi (pRExC_state->rxi)
289 #define RExC_start (pRExC_state->start)
290 #define RExC_end (pRExC_state->end)
291 #define RExC_parse (pRExC_state->parse)
292 #define RExC_latest_warn_offset (pRExC_state->latest_warn_offset )
293 #define RExC_whilem_seen (pRExC_state->whilem_seen)
294 #define RExC_seen_d_op (pRExC_state->seen_d_op) /* Seen something that differs
295 under /d from /u ? */
296
297 #ifdef RE_TRACK_PATTERN_OFFSETS
298 # define RExC_offsets (RExC_rxi->u.offsets) /* I am not like the
299 others */
300 #endif
301 #define RExC_emit (pRExC_state->emit)
302 #define RExC_emit_start (pRExC_state->emit_start)
303 #define RExC_sawback (pRExC_state->sawback)
304 #define RExC_seen (pRExC_state->seen)
305 #define RExC_size (pRExC_state->size)
306 #define RExC_maxlen (pRExC_state->maxlen)
307 #define RExC_npar (pRExC_state->npar)
308 #define RExC_total_parens (pRExC_state->total_par)
309 #define RExC_parens_buf_size (pRExC_state->parens_buf_size)
310 #define RExC_nestroot (pRExC_state->nestroot)
311 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
312 #define RExC_utf8 (pRExC_state->utf8)
313 #define RExC_uni_semantics (pRExC_state->uni_semantics)
314 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
315 #define RExC_open_parens (pRExC_state->open_parens)
316 #define RExC_close_parens (pRExC_state->close_parens)
317 #define RExC_end_op (pRExC_state->end_op)
318 #define RExC_paren_names (pRExC_state->paren_names)
319 #define RExC_recurse (pRExC_state->recurse)
320 #define RExC_recurse_count (pRExC_state->recurse_count)
321 #define RExC_sets_depth (pRExC_state->sets_depth)
322 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
323 #define RExC_study_chunk_recursed_bytes \
324 (pRExC_state->study_chunk_recursed_bytes)
325 #define RExC_in_lookaround (pRExC_state->in_lookaround)
326 #define RExC_contains_locale (pRExC_state->contains_locale)
327 #define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
328
329 #ifdef EBCDIC
330 # define SET_recode_x_to_native(x) \
331 STMT_START { RExC_recode_x_to_native = (x); } STMT_END
332 #else
333 # define SET_recode_x_to_native(x) NOOP
334 #endif
335
336 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
337 #define RExC_frame_head (pRExC_state->frame_head)
338 #define RExC_frame_last (pRExC_state->frame_last)
339 #define RExC_frame_count (pRExC_state->frame_count)
340 #define RExC_strict (pRExC_state->strict)
341 #define RExC_study_started (pRExC_state->study_started)
342 #define RExC_warn_text (pRExC_state->warn_text)
343 #define RExC_in_script_run (pRExC_state->in_script_run)
344 #define RExC_use_BRANCHJ (pRExC_state->use_BRANCHJ)
345 #define RExC_warned_WARN_EXPERIMENTAL__VLB (pRExC_state->sWARN_EXPERIMENTAL__VLB)
346 #define RExC_warned_WARN_EXPERIMENTAL__REGEX_SETS (pRExC_state->sWARN_EXPERIMENTAL__REGEX_SETS)
347 #define RExC_unlexed_names (pRExC_state->unlexed_names)
348
349 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
350 * a flag to disable back-off on the fixed/floating substrings - if it's
351 * a high complexity pattern we assume the benefit of avoiding a full match
352 * is worth the cost of checking for the substrings even if they rarely help.
353 */
354 #define RExC_naughty (pRExC_state->naughty)
355 #define TOO_NAUGHTY (10)
356 #define MARK_NAUGHTY(add) \
357 if (RExC_naughty < TOO_NAUGHTY) \
358 RExC_naughty += (add)
359 #define MARK_NAUGHTY_EXP(exp, add) \
360 if (RExC_naughty < TOO_NAUGHTY) \
361 RExC_naughty += RExC_naughty / (exp) + (add)
362
363 #define isNON_BRACE_QUANTIFIER(c) ((c) == '*' || (c) == '+' || (c) == '?')
364 #define isQUANTIFIER(s,e) ( isNON_BRACE_QUANTIFIER(*s) \
365 || ((*s) == '{' && regcurly(s, e, NULL)))
366
367 /*
368 * Flags to be passed up.
369 */
370 #define HASWIDTH 0x01 /* Known to not match null strings, could match
371 non-null ones. */
372 #define SIMPLE 0x02 /* Exactly one character wide */
373 /* (or LNBREAK as a special case) */
374 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
375 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
376 #define RESTART_PARSE 0x20 /* Need to redo the parse */
377 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PARSE, need to
378 calcuate sizes as UTF-8 */
379
380 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
381
382 /* whether trie related optimizations are enabled */
383 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
384 #define TRIE_STUDY_OPT
385 #define FULL_TRIE_STUDY
386 #define TRIE_STCLASS
387 #endif
388
389
390
391 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
392 #define PBITVAL(paren) (1 << ((paren) & 7))
393 #define PAREN_OFFSET(depth) \
394 (RExC_study_chunk_recursed + (depth) * RExC_study_chunk_recursed_bytes)
395 #define PAREN_TEST(depth, paren) \
396 (PBYTE(PAREN_OFFSET(depth), paren) & PBITVAL(paren))
397 #define PAREN_SET(depth, paren) \
398 (PBYTE(PAREN_OFFSET(depth), paren) |= PBITVAL(paren))
399 #define PAREN_UNSET(depth, paren) \
400 (PBYTE(PAREN_OFFSET(depth), paren) &= ~PBITVAL(paren))
401
402 #define REQUIRE_UTF8(flagp) STMT_START { \
403 if (!UTF) { \
404 *flagp = RESTART_PARSE|NEED_UTF8; \
405 return 0; \
406 } \
407 } STMT_END
408
409 /* /u is to be chosen if we are supposed to use Unicode rules, or if the
410 * pattern is in UTF-8. This latter condition is in case the outermost rules
411 * are locale. See GH #17278 */
412 #define toUSE_UNI_CHARSET_NOT_DEPENDS (RExC_uni_semantics || UTF)
413
414 /* Change from /d into /u rules, and restart the parse. RExC_uni_semantics is
415 * a flag that indicates we need to override /d with /u as a result of
416 * something in the pattern. It should only be used in regards to calling
417 * set_regex_charset() or get_regex_charset() */
418 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
419 STMT_START { \
420 if (DEPENDS_SEMANTICS) { \
421 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
422 RExC_uni_semantics = 1; \
423 if (RExC_seen_d_op && LIKELY(! IN_PARENS_PASS)) { \
424 /* No need to restart the parse if we haven't seen \
425 * anything that differs between /u and /d, and no need \
426 * to restart immediately if we're going to reparse \
427 * anyway to count parens */ \
428 *flagp |= RESTART_PARSE; \
429 return restart_retval; \
430 } \
431 } \
432 } STMT_END
433
434 #define REQUIRE_BRANCHJ(flagp, restart_retval) \
435 STMT_START { \
436 RExC_use_BRANCHJ = 1; \
437 *flagp |= RESTART_PARSE; \
438 return restart_retval; \
439 } STMT_END
440
441 /* Until we have completed the parse, we leave RExC_total_parens at 0 or
442 * less. After that, it must always be positive, because the whole re is
443 * considered to be surrounded by virtual parens. Setting it to negative
444 * indicates there is some construct that needs to know the actual number of
445 * parens to be properly handled. And that means an extra pass will be
446 * required after we've counted them all */
447 #define ALL_PARENS_COUNTED (RExC_total_parens > 0)
448 #define REQUIRE_PARENS_PASS \
449 STMT_START { /* No-op if have completed a pass */ \
450 if (! ALL_PARENS_COUNTED) RExC_total_parens = -1; \
451 } STMT_END
452 #define IN_PARENS_PASS (RExC_total_parens < 0)
453
454
455 /* This is used to return failure (zero) early from the calling function if
456 * various flags in 'flags' are set. Two flags always cause a return:
457 * 'RESTART_PARSE' and 'NEED_UTF8'. 'extra' can be used to specify any
458 * additional flags that should cause a return; 0 if none. If the return will
459 * be done, '*flagp' is first set to be all of the flags that caused the
460 * return. */
461 #define RETURN_FAIL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
462 STMT_START { \
463 if ((flags) & (RESTART_PARSE|NEED_UTF8|(extra))) { \
464 *(flagp) = (flags) & (RESTART_PARSE|NEED_UTF8|(extra)); \
465 return 0; \
466 } \
467 } STMT_END
468
469 #define MUST_RESTART(flags) ((flags) & (RESTART_PARSE))
470
471 #define RETURN_FAIL_ON_RESTART(flags,flagp) \
472 RETURN_FAIL_ON_RESTART_OR_FLAGS( flags, flagp, 0)
473 #define RETURN_FAIL_ON_RESTART_FLAGP(flagp) \
474 if (MUST_RESTART(*(flagp))) return 0
475
476 /* This converts the named class defined in regcomp.h to its equivalent class
477 * number defined in handy.h. */
478 #define namedclass_to_classnum(class) ((int) ((class) / 2))
479 #define classnum_to_namedclass(classnum) ((classnum) * 2)
480
481 #define _invlist_union_complement_2nd(a, b, output) \
482 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
483 #define _invlist_intersection_complement_2nd(a, b, output) \
484 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
485
486 /* We add a marker if we are deferring expansion of a property that is both
487 * 1) potentiallly user-defined; and
488 * 2) could also be an official Unicode property.
489 *
490 * Without this marker, any deferred expansion can only be for a user-defined
491 * one. This marker shouldn't conflict with any that could be in a legal name,
492 * and is appended to its name to indicate this. There is a string and
493 * character form */
494 #define DEFERRED_COULD_BE_OFFICIAL_MARKERs "~"
495 #define DEFERRED_COULD_BE_OFFICIAL_MARKERc '~'
496
497 /* What is infinity for optimization purposes */
498 #define OPTIMIZE_INFTY SSize_t_MAX
499
500 /* About scan_data_t.
501
502 During optimisation we recurse through the regexp program performing
503 various inplace (keyhole style) optimisations. In addition study_chunk
504 and scan_commit populate this data structure with information about
505 what strings MUST appear in the pattern. We look for the longest
506 string that must appear at a fixed location, and we look for the
507 longest string that may appear at a floating location. So for instance
508 in the pattern:
509
510 /FOO[xX]A.*B[xX]BAR/
511
512 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
513 strings (because they follow a .* construct). study_chunk will identify
514 both FOO and BAR as being the longest fixed and floating strings respectively.
515
516 The strings can be composites, for instance
517
518 /(f)(o)(o)/
519
520 will result in a composite fixed substring 'foo'.
521
522 For each string some basic information is maintained:
523
524 - min_offset
525 This is the position the string must appear at, or not before.
526 It also implicitly (when combined with minlenp) tells us how many
527 characters must match before the string we are searching for.
528 Likewise when combined with minlenp and the length of the string it
529 tells us how many characters must appear after the string we have
530 found.
531
532 - max_offset
533 Only used for floating strings. This is the rightmost point that
534 the string can appear at. If set to OPTIMIZE_INFTY it indicates that the
535 string can occur infinitely far to the right.
536 For fixed strings, it is equal to min_offset.
537
538 - minlenp
539 A pointer to the minimum number of characters of the pattern that the
540 string was found inside. This is important as in the case of positive
541 lookahead or positive lookbehind we can have multiple patterns
542 involved. Consider
543
544 /(?=FOO).*F/
545
546 The minimum length of the pattern overall is 3, the minimum length
547 of the lookahead part is 3, but the minimum length of the part that
548 will actually match is 1. So 'FOO's minimum length is 3, but the
549 minimum length for the F is 1. This is important as the minimum length
550 is used to determine offsets in front of and behind the string being
551 looked for. Since strings can be composites this is the length of the
552 pattern at the time it was committed with a scan_commit. Note that
553 the length is calculated by study_chunk, so that the minimum lengths
554 are not known until the full pattern has been compiled, thus the
555 pointer to the value.
556
557 - lookbehind
558
559 In the case of lookbehind the string being searched for can be
560 offset past the start point of the final matching string.
561 If this value was just blithely removed from the min_offset it would
562 invalidate some of the calculations for how many chars must match
563 before or after (as they are derived from min_offset and minlen and
564 the length of the string being searched for).
565 When the final pattern is compiled and the data is moved from the
566 scan_data_t structure into the regexp structure the information
567 about lookbehind is factored in, with the information that would
568 have been lost precalculated in the end_shift field for the
569 associated string.
570
571 The fields pos_min and pos_delta are used to store the minimum offset
572 and the delta to the maximum offset at the current point in the pattern.
573
574 */
575
576 struct scan_data_substrs {
577 SV *str; /* longest substring found in pattern */
578 SSize_t min_offset; /* earliest point in string it can appear */
579 SSize_t max_offset; /* latest point in string it can appear */
580 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
581 SSize_t lookbehind; /* is the pos of the string modified by LB */
582 I32 flags; /* per substring SF_* and SCF_* flags */
583 };
584
585 typedef struct scan_data_t {
586 /*I32 len_min; unused */
587 /*I32 len_delta; unused */
588 SSize_t pos_min;
589 SSize_t pos_delta;
590 SV *last_found;
591 SSize_t last_end; /* min value, <0 unless valid. */
592 SSize_t last_start_min;
593 SSize_t last_start_max;
594 U8 cur_is_floating; /* whether the last_* values should be set as
595 * the next fixed (0) or floating (1)
596 * substring */
597
598 /* [0] is longest fixed substring so far, [1] is longest float so far */
599 struct scan_data_substrs substrs[2];
600
601 I32 flags; /* common SF_* and SCF_* flags */
602 I32 whilem_c;
603 SSize_t *last_closep;
604 regnode_ssc *start_class;
605 } scan_data_t;
606
607 /*
608 * Forward declarations for pregcomp()'s friends.
609 */
610
611 static const scan_data_t zero_scan_data = {
612 0, 0, NULL, 0, 0, 0, 0,
613 {
614 { NULL, 0, 0, 0, 0, 0 },
615 { NULL, 0, 0, 0, 0, 0 },
616 },
617 0, 0, NULL, NULL
618 };
619
620 /* study flags */
621
622 #define SF_BEFORE_SEOL 0x0001
623 #define SF_BEFORE_MEOL 0x0002
624 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
625
626 #define SF_IS_INF 0x0040
627 #define SF_HAS_PAR 0x0080
628 #define SF_IN_PAR 0x0100
629 #define SF_HAS_EVAL 0x0200
630
631
632 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
633 * longest substring in the pattern. When it is not set the optimiser keeps
634 * track of position, but does not keep track of the actual strings seen,
635 *
636 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
637 * /foo/i will not.
638 *
639 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
640 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
641 * turned off because of the alternation (BRANCH). */
642 #define SCF_DO_SUBSTR 0x0400
643
644 #define SCF_DO_STCLASS_AND 0x0800
645 #define SCF_DO_STCLASS_OR 0x1000
646 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
647 #define SCF_WHILEM_VISITED_POS 0x2000
648
649 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
650 #define SCF_SEEN_ACCEPT 0x8000
651 #define SCF_TRIE_DOING_RESTUDY 0x10000
652 #define SCF_IN_DEFINE 0x20000
653
654
655
656
657 #define UTF cBOOL(RExC_utf8)
658
659 /* The enums for all these are ordered so things work out correctly */
660 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
661 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
662 == REGEX_DEPENDS_CHARSET)
663 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
664 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
665 >= REGEX_UNICODE_CHARSET)
666 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
667 == REGEX_ASCII_RESTRICTED_CHARSET)
668 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
669 >= REGEX_ASCII_RESTRICTED_CHARSET)
670 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
671 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
672
673 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
674
675 /* For programs that want to be strictly Unicode compatible by dying if any
676 * attempt is made to match a non-Unicode code point against a Unicode
677 * property. */
678 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
679
680 #define OOB_NAMEDCLASS -1
681
682 /* There is no code point that is out-of-bounds, so this is problematic. But
683 * its only current use is to initialize a variable that is always set before
684 * looked at. */
685 #define OOB_UNICODE 0xDEADBEEF
686
687 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
688
689
690 /* length of regex to show in messages that don't mark a position within */
691 #define RegexLengthToShowInErrorMessages 127
692
693 /*
694 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
695 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
696 * op/pragma/warn/regcomp.
697 */
698 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
699 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
700
701 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
702 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
703
704 /* The code in this file in places uses one level of recursion with parsing
705 * rebased to an alternate string constructed by us in memory. This can take
706 * the form of something that is completely different from the input, or
707 * something that uses the input as part of the alternate. In the first case,
708 * there should be no possibility of an error, as we are in complete control of
709 * the alternate string. But in the second case we don't completely control
710 * the input portion, so there may be errors in that. Here's an example:
711 * /[abc\x{DF}def]/ui
712 * is handled specially because \x{df} folds to a sequence of more than one
713 * character: 'ss'. What is done is to create and parse an alternate string,
714 * which looks like this:
715 * /(?:\x{DF}|[abc\x{DF}def])/ui
716 * where it uses the input unchanged in the middle of something it constructs,
717 * which is a branch for the DF outside the character class, and clustering
718 * parens around the whole thing. (It knows enough to skip the DF inside the
719 * class while in this substitute parse.) 'abc' and 'def' may have errors that
720 * need to be reported. The general situation looks like this:
721 *
722 * |<------- identical ------>|
723 * sI tI xI eI
724 * Input: ---------------------------------------------------------------
725 * Constructed: ---------------------------------------------------
726 * sC tC xC eC EC
727 * |<------- identical ------>|
728 *
729 * sI..eI is the portion of the input pattern we are concerned with here.
730 * sC..EC is the constructed substitute parse string.
731 * sC..tC is constructed by us
732 * tC..eC is an exact duplicate of the portion of the input pattern tI..eI.
733 * In the diagram, these are vertically aligned.
734 * eC..EC is also constructed by us.
735 * xC is the position in the substitute parse string where we found a
736 * problem.
737 * xI is the position in the original pattern corresponding to xC.
738 *
739 * We want to display a message showing the real input string. Thus we need to
740 * translate from xC to xI. We know that xC >= tC, since the portion of the
741 * string sC..tC has been constructed by us, and so shouldn't have errors. We
742 * get:
743 * xI = tI + (xC - tC)
744 *
745 * When the substitute parse is constructed, the code needs to set:
746 * RExC_start (sC)
747 * RExC_end (eC)
748 * RExC_copy_start_in_input (tI)
749 * RExC_copy_start_in_constructed (tC)
750 * and restore them when done.
751 *
752 * During normal processing of the input pattern, both
753 * 'RExC_copy_start_in_input' and 'RExC_copy_start_in_constructed' are set to
754 * sI, so that xC equals xI.
755 */
756
757 #define sI RExC_precomp
758 #define eI RExC_precomp_end
759 #define sC RExC_start
760 #define eC RExC_end
761 #define tI RExC_copy_start_in_input
762 #define tC RExC_copy_start_in_constructed
763 #define xI(xC) (tI + (xC - tC))
764 #define xI_offset(xC) (xI(xC) - sI)
765
766 #define REPORT_LOCATION_ARGS(xC) \
767 UTF8fARG(UTF, \
768 (xI(xC) > eI) /* Don't run off end */ \
769 ? eI - sI /* Length before the <--HERE */ \
770 : ((xI_offset(xC) >= 0) \
771 ? xI_offset(xC) \
772 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
773 IVdf " trying to output message for " \
774 " pattern %.*s", \
775 __FILE__, __LINE__, (IV) xI_offset(xC), \
776 ((int) (eC - sC)), sC), 0)), \
777 sI), /* The input pattern printed up to the <--HERE */ \
778 UTF8fARG(UTF, \
779 (xI(xC) > eI) ? 0 : eI - xI(xC), /* Length after <--HERE */ \
780 (xI(xC) > eI) ? eI : xI(xC)) /* pattern after <--HERE */
781
782 /* Used to point after bad bytes for an error message, but avoid skipping
783 * past a nul byte. */
784 #define SKIP_IF_CHAR(s, e) (!*(s) ? 0 : UTF ? UTF8_SAFE_SKIP(s, e) : 1)
785
786 /* Set up to clean up after our imminent demise */
787 #define PREPARE_TO_DIE \
788 STMT_START { \
789 if (RExC_rx_sv) \
790 SAVEFREESV(RExC_rx_sv); \
791 if (RExC_open_parens) \
792 SAVEFREEPV(RExC_open_parens); \
793 if (RExC_close_parens) \
794 SAVEFREEPV(RExC_close_parens); \
795 } STMT_END
796
797 /*
798 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
799 * arg. Show regex, up to a maximum length. If it's too long, chop and add
800 * "...".
801 */
802 #define _FAIL(code) STMT_START { \
803 const char *ellipses = ""; \
804 IV len = RExC_precomp_end - RExC_precomp; \
805 \
806 PREPARE_TO_DIE; \
807 if (len > RegexLengthToShowInErrorMessages) { \
808 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
809 len = RegexLengthToShowInErrorMessages - 10; \
810 ellipses = "..."; \
811 } \
812 code; \
813 } STMT_END
814
815 #define FAIL(msg) _FAIL( \
816 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
817 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
818
819 #define FAIL2(msg,arg) _FAIL( \
820 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
821 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
822
823 #define FAIL3(msg,arg1,arg2) _FAIL( \
824 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
825 arg1, arg2, UTF8fARG(UTF, len, RExC_precomp), ellipses))
826
827 /*
828 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
829 */
830 #define Simple_vFAIL(m) STMT_START { \
831 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
832 m, REPORT_LOCATION_ARGS(RExC_parse)); \
833 } STMT_END
834
835 /*
836 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
837 */
838 #define vFAIL(m) STMT_START { \
839 PREPARE_TO_DIE; \
840 Simple_vFAIL(m); \
841 } STMT_END
842
843 /*
844 * Like Simple_vFAIL(), but accepts two arguments.
845 */
846 #define Simple_vFAIL2(m,a1) STMT_START { \
847 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, \
848 REPORT_LOCATION_ARGS(RExC_parse)); \
849 } STMT_END
850
851 /*
852 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
853 */
854 #define vFAIL2(m,a1) STMT_START { \
855 PREPARE_TO_DIE; \
856 Simple_vFAIL2(m, a1); \
857 } STMT_END
858
859
860 /*
861 * Like Simple_vFAIL(), but accepts three arguments.
862 */
863 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
864 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, \
865 REPORT_LOCATION_ARGS(RExC_parse)); \
866 } STMT_END
867
868 /*
869 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
870 */
871 #define vFAIL3(m,a1,a2) STMT_START { \
872 PREPARE_TO_DIE; \
873 Simple_vFAIL3(m, a1, a2); \
874 } STMT_END
875
876 /*
877 * Like Simple_vFAIL(), but accepts four arguments.
878 */
879 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
880 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, a3, \
881 REPORT_LOCATION_ARGS(RExC_parse)); \
882 } STMT_END
883
884 #define vFAIL4(m,a1,a2,a3) STMT_START { \
885 PREPARE_TO_DIE; \
886 Simple_vFAIL4(m, a1, a2, a3); \
887 } STMT_END
888
889 /* A specialized version of vFAIL2 that works with UTF8f */
890 #define vFAIL2utf8f(m, a1) STMT_START { \
891 PREPARE_TO_DIE; \
892 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, \
893 REPORT_LOCATION_ARGS(RExC_parse)); \
894 } STMT_END
895
896 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
897 PREPARE_TO_DIE; \
898 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, \
899 REPORT_LOCATION_ARGS(RExC_parse)); \
900 } STMT_END
901
902 /* Setting this to NULL is a signal to not output warnings */
903 #define TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE \
904 STMT_START { \
905 RExC_save_copy_start_in_constructed = RExC_copy_start_in_constructed;\
906 RExC_copy_start_in_constructed = NULL; \
907 } STMT_END
908 #define RESTORE_WARNINGS \
909 RExC_copy_start_in_constructed = RExC_save_copy_start_in_constructed
910
911 /* Since a warning can be generated multiple times as the input is reparsed, we
912 * output it the first time we come to that point in the parse, but suppress it
913 * otherwise. 'RExC_copy_start_in_constructed' being NULL is a flag to not
914 * generate any warnings */
915 #define TO_OUTPUT_WARNINGS(loc) \
916 ( RExC_copy_start_in_constructed \
917 && ((xI(loc)) - RExC_precomp) > (Ptrdiff_t) RExC_latest_warn_offset)
918
919 /* After we've emitted a warning, we save the position in the input so we don't
920 * output it again */
921 #define UPDATE_WARNINGS_LOC(loc) \
922 STMT_START { \
923 if (TO_OUTPUT_WARNINGS(loc)) { \
924 RExC_latest_warn_offset = MAX(sI, MIN(eI, xI(loc))) \
925 - RExC_precomp; \
926 } \
927 } STMT_END
928
929 /* 'warns' is the output of the packWARNx macro used in 'code' */
930 #define _WARN_HELPER(loc, warns, code) \
931 STMT_START { \
932 if (! RExC_copy_start_in_constructed) { \
933 Perl_croak( aTHX_ "panic! %s: %d: Tried to warn when none" \
934 " expected at '%s'", \
935 __FILE__, __LINE__, loc); \
936 } \
937 if (TO_OUTPUT_WARNINGS(loc)) { \
938 if (ckDEAD(warns)) \
939 PREPARE_TO_DIE; \
940 code; \
941 UPDATE_WARNINGS_LOC(loc); \
942 } \
943 } STMT_END
944
945 /* m is not necessarily a "literal string", in this macro */
946 #define warn_non_literal_string(loc, packed_warn, m) \
947 _WARN_HELPER(loc, packed_warn, \
948 Perl_warner(aTHX_ packed_warn, \
949 "%s" REPORT_LOCATION, \
950 m, REPORT_LOCATION_ARGS(loc)))
951 #define reg_warn_non_literal_string(loc, m) \
952 warn_non_literal_string(loc, packWARN(WARN_REGEXP), m)
953
954 #define ckWARN2_non_literal_string(loc, packwarn, m, a1) \
955 STMT_START { \
956 char * format; \
957 Size_t format_size = strlen(m) + strlen(REPORT_LOCATION)+ 1;\
958 Newx(format, format_size, char); \
959 my_strlcpy(format, m, format_size); \
960 my_strlcat(format, REPORT_LOCATION, format_size); \
961 SAVEFREEPV(format); \
962 _WARN_HELPER(loc, packwarn, \
963 Perl_ck_warner(aTHX_ packwarn, \
964 format, \
965 a1, REPORT_LOCATION_ARGS(loc))); \
966 } STMT_END
967
968 #define ckWARNreg(loc,m) \
969 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
970 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
971 m REPORT_LOCATION, \
972 REPORT_LOCATION_ARGS(loc)))
973
974 #define vWARN(loc, m) \
975 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
976 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
977 m REPORT_LOCATION, \
978 REPORT_LOCATION_ARGS(loc))) \
979
980 #define vWARN_dep(loc, m) \
981 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
982 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
983 m REPORT_LOCATION, \
984 REPORT_LOCATION_ARGS(loc)))
985
986 #define ckWARNdep(loc,m) \
987 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
988 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
989 m REPORT_LOCATION, \
990 REPORT_LOCATION_ARGS(loc)))
991
992 #define ckWARNregdep(loc,m) \
993 _WARN_HELPER(loc, packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
994 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
995 WARN_REGEXP), \
996 m REPORT_LOCATION, \
997 REPORT_LOCATION_ARGS(loc)))
998
999 #define ckWARN2reg_d(loc,m, a1) \
1000 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1001 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
1002 m REPORT_LOCATION, \
1003 a1, REPORT_LOCATION_ARGS(loc)))
1004
1005 #define ckWARN2reg(loc, m, a1) \
1006 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1007 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1008 m REPORT_LOCATION, \
1009 a1, REPORT_LOCATION_ARGS(loc)))
1010
1011 #define vWARN3(loc, m, a1, a2) \
1012 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1013 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1014 m REPORT_LOCATION, \
1015 a1, a2, REPORT_LOCATION_ARGS(loc)))
1016
1017 #define ckWARN3reg(loc, m, a1, a2) \
1018 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1019 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1020 m REPORT_LOCATION, \
1021 a1, a2, \
1022 REPORT_LOCATION_ARGS(loc)))
1023
1024 #define vWARN4(loc, m, a1, a2, a3) \
1025 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1026 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1027 m REPORT_LOCATION, \
1028 a1, a2, a3, \
1029 REPORT_LOCATION_ARGS(loc)))
1030
1031 #define ckWARN4reg(loc, m, a1, a2, a3) \
1032 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1033 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1034 m REPORT_LOCATION, \
1035 a1, a2, a3, \
1036 REPORT_LOCATION_ARGS(loc)))
1037
1038 #define vWARN5(loc, m, a1, a2, a3, a4) \
1039 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1040 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1041 m REPORT_LOCATION, \
1042 a1, a2, a3, a4, \
1043 REPORT_LOCATION_ARGS(loc)))
1044
1045 #define ckWARNexperimental(loc, class, m) \
1046 STMT_START { \
1047 if (! RExC_warned_ ## class) { /* warn once per compilation */ \
1048 RExC_warned_ ## class = 1; \
1049 _WARN_HELPER(loc, packWARN(class), \
1050 Perl_ck_warner_d(aTHX_ packWARN(class), \
1051 m REPORT_LOCATION, \
1052 REPORT_LOCATION_ARGS(loc)));\
1053 } \
1054 } STMT_END
1055
1056 /* Convert between a pointer to a node and its offset from the beginning of the
1057 * program */
1058 #define REGNODE_p(offset) (RExC_emit_start + (offset))
1059 #define REGNODE_OFFSET(node) ((node) - RExC_emit_start)
1060
1061 /* Macros for recording node offsets. 20001227 mjd@plover.com
1062 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
1063 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
1064 * Element 0 holds the number n.
1065 * Position is 1 indexed.
1066 */
1067 #ifndef RE_TRACK_PATTERN_OFFSETS
1068 #define Set_Node_Offset_To_R(offset,byte)
1069 #define Set_Node_Offset(node,byte)
1070 #define Set_Cur_Node_Offset
1071 #define Set_Node_Length_To_R(node,len)
1072 #define Set_Node_Length(node,len)
1073 #define Set_Node_Cur_Length(node,start)
1074 #define Node_Offset(n)
1075 #define Node_Length(n)
1076 #define Set_Node_Offset_Length(node,offset,len)
1077 #define ProgLen(ri) ri->u.proglen
1078 #define SetProgLen(ri,x) ri->u.proglen = x
1079 #define Track_Code(code)
1080 #else
1081 #define ProgLen(ri) ri->u.offsets[0]
1082 #define SetProgLen(ri,x) ri->u.offsets[0] = x
1083 #define Set_Node_Offset_To_R(offset,byte) STMT_START { \
1084 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
1085 __LINE__, (int)(offset), (int)(byte))); \
1086 if((offset) < 0) { \
1087 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
1088 (int)(offset)); \
1089 } else { \
1090 RExC_offsets[2*(offset)-1] = (byte); \
1091 } \
1092 } STMT_END
1093
1094 #define Set_Node_Offset(node,byte) \
1095 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (byte)-RExC_start)
1096 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
1097
1098 #define Set_Node_Length_To_R(node,len) STMT_START { \
1099 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
1100 __LINE__, (int)(node), (int)(len))); \
1101 if((node) < 0) { \
1102 Perl_croak(aTHX_ "value of node is %d in Length macro", \
1103 (int)(node)); \
1104 } else { \
1105 RExC_offsets[2*(node)] = (len); \
1106 } \
1107 } STMT_END
1108
1109 #define Set_Node_Length(node,len) \
1110 Set_Node_Length_To_R(REGNODE_OFFSET(node), len)
1111 #define Set_Node_Cur_Length(node, start) \
1112 Set_Node_Length(node, RExC_parse - start)
1113
1114 /* Get offsets and lengths */
1115 #define Node_Offset(n) (RExC_offsets[2*(REGNODE_OFFSET(n))-1])
1116 #define Node_Length(n) (RExC_offsets[2*(REGNODE_OFFSET(n))])
1117
1118 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
1119 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (offset)); \
1120 Set_Node_Length_To_R(REGNODE_OFFSET(node), (len)); \
1121 } STMT_END
1122
1123 #define Track_Code(code) STMT_START { code } STMT_END
1124 #endif
1125
1126 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
1127 #define EXPERIMENTAL_INPLACESCAN
1128 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
1129
1130 #ifdef DEBUGGING
1131 int
Perl_re_printf(pTHX_ const char * fmt,...)1132 Perl_re_printf(pTHX_ const char *fmt, ...)
1133 {
1134 va_list ap;
1135 int result;
1136 PerlIO *f= Perl_debug_log;
1137 PERL_ARGS_ASSERT_RE_PRINTF;
1138 va_start(ap, fmt);
1139 result = PerlIO_vprintf(f, fmt, ap);
1140 va_end(ap);
1141 return result;
1142 }
1143
1144 int
Perl_re_indentf(pTHX_ const char * fmt,U32 depth,...)1145 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
1146 {
1147 va_list ap;
1148 int result;
1149 PerlIO *f= Perl_debug_log;
1150 PERL_ARGS_ASSERT_RE_INDENTF;
1151 va_start(ap, depth);
1152 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
1153 result = PerlIO_vprintf(f, fmt, ap);
1154 va_end(ap);
1155 return result;
1156 }
1157 #endif /* DEBUGGING */
1158
1159 #define DEBUG_RExC_seen() \
1160 DEBUG_OPTIMISE_MORE_r({ \
1161 Perl_re_printf( aTHX_ "RExC_seen: "); \
1162 \
1163 if (RExC_seen & REG_ZERO_LEN_SEEN) \
1164 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
1165 \
1166 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
1167 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
1168 \
1169 if (RExC_seen & REG_GPOS_SEEN) \
1170 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
1171 \
1172 if (RExC_seen & REG_RECURSE_SEEN) \
1173 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
1174 \
1175 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
1176 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
1177 \
1178 if (RExC_seen & REG_VERBARG_SEEN) \
1179 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
1180 \
1181 if (RExC_seen & REG_CUTGROUP_SEEN) \
1182 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
1183 \
1184 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
1185 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
1186 \
1187 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
1188 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
1189 \
1190 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
1191 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
1192 \
1193 Perl_re_printf( aTHX_ "\n"); \
1194 });
1195
1196 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
1197 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
1198
1199
1200 #ifdef DEBUGGING
1201 static void
S_debug_show_study_flags(pTHX_ U32 flags,const char * open_str,const char * close_str)1202 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1203 const char *close_str)
1204 {
1205 if (!flags)
1206 return;
1207
1208 Perl_re_printf( aTHX_ "%s", open_str);
1209 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1210 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
1211 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1212 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1213 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1214 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1215 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1216 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1217 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1218 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1219 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1220 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1221 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1222 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1223 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1224 Perl_re_printf( aTHX_ "%s", close_str);
1225 }
1226
1227
1228 static void
S_debug_studydata(pTHX_ const char * where,scan_data_t * data,U32 depth,int is_inf)1229 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1230 U32 depth, int is_inf)
1231 {
1232 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1233
1234 DEBUG_OPTIMISE_MORE_r({
1235 if (!data)
1236 return;
1237 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1238 depth,
1239 where,
1240 (IV)data->pos_min,
1241 (IV)data->pos_delta,
1242 (UV)data->flags
1243 );
1244
1245 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1246
1247 Perl_re_printf( aTHX_
1248 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1249 (IV)data->whilem_c,
1250 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1251 is_inf ? "INF " : ""
1252 );
1253
1254 if (data->last_found) {
1255 int i;
1256 Perl_re_printf(aTHX_
1257 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1258 SvPVX_const(data->last_found),
1259 (IV)data->last_end,
1260 (IV)data->last_start_min,
1261 (IV)data->last_start_max
1262 );
1263
1264 for (i = 0; i < 2; i++) {
1265 Perl_re_printf(aTHX_
1266 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1267 data->cur_is_floating == i ? "*" : "",
1268 i ? "Float" : "Fixed",
1269 SvPVX_const(data->substrs[i].str),
1270 (IV)data->substrs[i].min_offset,
1271 (IV)data->substrs[i].max_offset
1272 );
1273 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1274 }
1275 }
1276
1277 Perl_re_printf( aTHX_ "\n");
1278 });
1279 }
1280
1281
1282 static void
S_debug_peep(pTHX_ const char * str,const RExC_state_t * pRExC_state,regnode * scan,U32 depth,U32 flags)1283 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1284 regnode *scan, U32 depth, U32 flags)
1285 {
1286 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1287
1288 DEBUG_OPTIMISE_r({
1289 regnode *Next;
1290
1291 if (!scan)
1292 return;
1293 Next = regnext(scan);
1294 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1295 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1296 depth,
1297 str,
1298 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1299 Next ? (REG_NODE_NUM(Next)) : 0 );
1300 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1301 Perl_re_printf( aTHX_ "\n");
1302 });
1303 }
1304
1305
1306 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1307 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1308
1309 # define DEBUG_PEEP(str, scan, depth, flags) \
1310 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1311
1312 #else
1313 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1314 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1315 #endif
1316
1317
1318 /* =========================================================
1319 * BEGIN edit_distance stuff.
1320 *
1321 * This calculates how many single character changes of any type are needed to
1322 * transform a string into another one. It is taken from version 3.1 of
1323 *
1324 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1325 */
1326
1327 /* Our unsorted dictionary linked list. */
1328 /* Note we use UVs, not chars. */
1329
1330 struct dictionary{
1331 UV key;
1332 UV value;
1333 struct dictionary* next;
1334 };
1335 typedef struct dictionary item;
1336
1337
1338 PERL_STATIC_INLINE item*
push(UV key,item * curr)1339 push(UV key, item* curr)
1340 {
1341 item* head;
1342 Newx(head, 1, item);
1343 head->key = key;
1344 head->value = 0;
1345 head->next = curr;
1346 return head;
1347 }
1348
1349
1350 PERL_STATIC_INLINE item*
find(item * head,UV key)1351 find(item* head, UV key)
1352 {
1353 item* iterator = head;
1354 while (iterator){
1355 if (iterator->key == key){
1356 return iterator;
1357 }
1358 iterator = iterator->next;
1359 }
1360
1361 return NULL;
1362 }
1363
1364 PERL_STATIC_INLINE item*
uniquePush(item * head,UV key)1365 uniquePush(item* head, UV key)
1366 {
1367 item* iterator = head;
1368
1369 while (iterator){
1370 if (iterator->key == key) {
1371 return head;
1372 }
1373 iterator = iterator->next;
1374 }
1375
1376 return push(key, head);
1377 }
1378
1379 PERL_STATIC_INLINE void
dict_free(item * head)1380 dict_free(item* head)
1381 {
1382 item* iterator = head;
1383
1384 while (iterator) {
1385 item* temp = iterator;
1386 iterator = iterator->next;
1387 Safefree(temp);
1388 }
1389
1390 head = NULL;
1391 }
1392
1393 /* End of Dictionary Stuff */
1394
1395 /* All calculations/work are done here */
1396 STATIC int
S_edit_distance(const UV * src,const UV * tgt,const STRLEN x,const STRLEN y,const SSize_t maxDistance)1397 S_edit_distance(const UV* src,
1398 const UV* tgt,
1399 const STRLEN x, /* length of src[] */
1400 const STRLEN y, /* length of tgt[] */
1401 const SSize_t maxDistance
1402 )
1403 {
1404 item *head = NULL;
1405 UV swapCount, swapScore, targetCharCount, i, j;
1406 UV *scores;
1407 UV score_ceil = x + y;
1408
1409 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1410
1411 /* intialize matrix start values */
1412 Newx(scores, ( (x + 2) * (y + 2)), UV);
1413 scores[0] = score_ceil;
1414 scores[1 * (y + 2) + 0] = score_ceil;
1415 scores[0 * (y + 2) + 1] = score_ceil;
1416 scores[1 * (y + 2) + 1] = 0;
1417 head = uniquePush(uniquePush(head, src[0]), tgt[0]);
1418
1419 /* work loops */
1420 /* i = src index */
1421 /* j = tgt index */
1422 for (i=1;i<=x;i++) {
1423 if (i < x)
1424 head = uniquePush(head, src[i]);
1425 scores[(i+1) * (y + 2) + 1] = i;
1426 scores[(i+1) * (y + 2) + 0] = score_ceil;
1427 swapCount = 0;
1428
1429 for (j=1;j<=y;j++) {
1430 if (i == 1) {
1431 if(j < y)
1432 head = uniquePush(head, tgt[j]);
1433 scores[1 * (y + 2) + (j + 1)] = j;
1434 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1435 }
1436
1437 targetCharCount = find(head, tgt[j-1])->value;
1438 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1439
1440 if (src[i-1] != tgt[j-1]){
1441 scores[(i+1) * (y + 2) + (j + 1)] = MIN(swapScore,(MIN(scores[i * (y + 2) + j], MIN(scores[(i+1) * (y + 2) + j], scores[i * (y + 2) + (j + 1)])) + 1));
1442 }
1443 else {
1444 swapCount = j;
1445 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1446 }
1447 }
1448
1449 find(head, src[i-1])->value = i;
1450 }
1451
1452 {
1453 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1454 dict_free(head);
1455 Safefree(scores);
1456 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1457 }
1458 }
1459
1460 /* END of edit_distance() stuff
1461 * ========================================================= */
1462
1463 /* Mark that we cannot extend a found fixed substring at this point.
1464 Update the longest found anchored substring or the longest found
1465 floating substrings if needed. */
1466
1467 STATIC void
S_scan_commit(pTHX_ const RExC_state_t * pRExC_state,scan_data_t * data,SSize_t * minlenp,int is_inf)1468 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1469 SSize_t *minlenp, int is_inf)
1470 {
1471 const STRLEN l = CHR_SVLEN(data->last_found);
1472 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1473 const STRLEN old_l = CHR_SVLEN(longest_sv);
1474 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1475
1476 PERL_ARGS_ASSERT_SCAN_COMMIT;
1477
1478 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1479 const U8 i = data->cur_is_floating;
1480 SvSetMagicSV(longest_sv, data->last_found);
1481 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1482
1483 if (!i) /* fixed */
1484 data->substrs[0].max_offset = data->substrs[0].min_offset;
1485 else { /* float */
1486 data->substrs[1].max_offset =
1487 (is_inf)
1488 ? OPTIMIZE_INFTY
1489 : (l
1490 ? data->last_start_max
1491 /* temporary underflow guard for 5.32 */
1492 : data->pos_delta < 0 ? OPTIMIZE_INFTY
1493 : (data->pos_delta > OPTIMIZE_INFTY - data->pos_min
1494 ? OPTIMIZE_INFTY
1495 : data->pos_min + data->pos_delta));
1496 }
1497
1498 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1499 data->substrs[i].flags |= data->flags & SF_BEFORE_EOL;
1500 data->substrs[i].minlenp = minlenp;
1501 data->substrs[i].lookbehind = 0;
1502 }
1503
1504 SvCUR_set(data->last_found, 0);
1505 {
1506 SV * const sv = data->last_found;
1507 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1508 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1509 if (mg)
1510 mg->mg_len = 0;
1511 }
1512 }
1513 data->last_end = -1;
1514 data->flags &= ~SF_BEFORE_EOL;
1515 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1516 }
1517
1518 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1519 * list that describes which code points it matches */
1520
1521 STATIC void
S_ssc_anything(pTHX_ regnode_ssc * ssc)1522 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1523 {
1524 /* Set the SSC 'ssc' to match an empty string or any code point */
1525
1526 PERL_ARGS_ASSERT_SSC_ANYTHING;
1527
1528 assert(is_ANYOF_SYNTHETIC(ssc));
1529
1530 /* mortalize so won't leak */
1531 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1532 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1533 }
1534
1535 STATIC int
S_ssc_is_anything(const regnode_ssc * ssc)1536 S_ssc_is_anything(const regnode_ssc *ssc)
1537 {
1538 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1539 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1540 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1541 * in any way, so there's no point in using it */
1542
1543 UV start, end;
1544 bool ret;
1545
1546 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1547
1548 assert(is_ANYOF_SYNTHETIC(ssc));
1549
1550 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1551 return FALSE;
1552 }
1553
1554 /* See if the list consists solely of the range 0 - Infinity */
1555 invlist_iterinit(ssc->invlist);
1556 ret = invlist_iternext(ssc->invlist, &start, &end)
1557 && start == 0
1558 && end == UV_MAX;
1559
1560 invlist_iterfinish(ssc->invlist);
1561
1562 if (ret) {
1563 return TRUE;
1564 }
1565
1566 /* If e.g., both \w and \W are set, matches everything */
1567 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1568 int i;
1569 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1570 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1571 return TRUE;
1572 }
1573 }
1574 }
1575
1576 return FALSE;
1577 }
1578
1579 STATIC void
S_ssc_init(pTHX_ const RExC_state_t * pRExC_state,regnode_ssc * ssc)1580 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1581 {
1582 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1583 * string, any code point, or any posix class under locale */
1584
1585 PERL_ARGS_ASSERT_SSC_INIT;
1586
1587 Zero(ssc, 1, regnode_ssc);
1588 set_ANYOF_SYNTHETIC(ssc);
1589 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1590 ssc_anything(ssc);
1591
1592 /* If any portion of the regex is to operate under locale rules that aren't
1593 * fully known at compile time, initialization includes it. The reason
1594 * this isn't done for all regexes is that the optimizer was written under
1595 * the assumption that locale was all-or-nothing. Given the complexity and
1596 * lack of documentation in the optimizer, and that there are inadequate
1597 * test cases for locale, many parts of it may not work properly, it is
1598 * safest to avoid locale unless necessary. */
1599 if (RExC_contains_locale) {
1600 ANYOF_POSIXL_SETALL(ssc);
1601 }
1602 else {
1603 ANYOF_POSIXL_ZERO(ssc);
1604 }
1605 }
1606
1607 STATIC int
S_ssc_is_cp_posixl_init(const RExC_state_t * pRExC_state,const regnode_ssc * ssc)1608 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1609 const regnode_ssc *ssc)
1610 {
1611 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1612 * to the list of code points matched, and locale posix classes; hence does
1613 * not check its flags) */
1614
1615 UV start, end;
1616 bool ret;
1617
1618 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1619
1620 assert(is_ANYOF_SYNTHETIC(ssc));
1621
1622 invlist_iterinit(ssc->invlist);
1623 ret = invlist_iternext(ssc->invlist, &start, &end)
1624 && start == 0
1625 && end == UV_MAX;
1626
1627 invlist_iterfinish(ssc->invlist);
1628
1629 if (! ret) {
1630 return FALSE;
1631 }
1632
1633 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1634 return FALSE;
1635 }
1636
1637 return TRUE;
1638 }
1639
1640 #define INVLIST_INDEX 0
1641 #define ONLY_LOCALE_MATCHES_INDEX 1
1642 #define DEFERRED_USER_DEFINED_INDEX 2
1643
1644 STATIC SV*
S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t * pRExC_state,const regnode_charclass * const node)1645 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1646 const regnode_charclass* const node)
1647 {
1648 /* Returns a mortal inversion list defining which code points are matched
1649 * by 'node', which is of type ANYOF. Handles complementing the result if
1650 * appropriate. If some code points aren't knowable at this time, the
1651 * returned list must, and will, contain every code point that is a
1652 * possibility. */
1653
1654 SV* invlist = NULL;
1655 SV* only_utf8_locale_invlist = NULL;
1656 unsigned int i;
1657 const U32 n = ARG(node);
1658 bool new_node_has_latin1 = FALSE;
1659 const U8 flags = (inRANGE(OP(node), ANYOFH, ANYOFRb))
1660 ? 0
1661 : ANYOF_FLAGS(node);
1662
1663 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1664
1665 /* Look at the data structure created by S_set_ANYOF_arg() */
1666 if (n != ANYOF_ONLY_HAS_BITMAP) {
1667 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1668 AV * const av = MUTABLE_AV(SvRV(rv));
1669 SV **const ary = AvARRAY(av);
1670 assert(RExC_rxi->data->what[n] == 's');
1671
1672 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
1673
1674 /* Here there are things that won't be known until runtime -- we
1675 * have to assume it could be anything */
1676 invlist = sv_2mortal(_new_invlist(1));
1677 return _add_range_to_invlist(invlist, 0, UV_MAX);
1678 }
1679 else if (ary[INVLIST_INDEX]) {
1680
1681 /* Use the node's inversion list */
1682 invlist = sv_2mortal(invlist_clone(ary[INVLIST_INDEX], NULL));
1683 }
1684
1685 /* Get the code points valid only under UTF-8 locales */
1686 if ( (flags & ANYOFL_FOLD)
1687 && av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX)
1688 {
1689 only_utf8_locale_invlist = ary[ONLY_LOCALE_MATCHES_INDEX];
1690 }
1691 }
1692
1693 if (! invlist) {
1694 invlist = sv_2mortal(_new_invlist(0));
1695 }
1696
1697 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1698 * code points, and an inversion list for the others, but if there are code
1699 * points that should match only conditionally on the target string being
1700 * UTF-8, those are placed in the inversion list, and not the bitmap.
1701 * Since there are circumstances under which they could match, they are
1702 * included in the SSC. But if the ANYOF node is to be inverted, we have
1703 * to exclude them here, so that when we invert below, the end result
1704 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1705 * have to do this here before we add the unconditionally matched code
1706 * points */
1707 if (flags & ANYOF_INVERT) {
1708 _invlist_intersection_complement_2nd(invlist,
1709 PL_UpperLatin1,
1710 &invlist);
1711 }
1712
1713 /* Add in the points from the bit map */
1714 if (! inRANGE(OP(node), ANYOFH, ANYOFRb)) {
1715 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1716 if (ANYOF_BITMAP_TEST(node, i)) {
1717 unsigned int start = i++;
1718
1719 for (; i < NUM_ANYOF_CODE_POINTS
1720 && ANYOF_BITMAP_TEST(node, i); ++i)
1721 {
1722 /* empty */
1723 }
1724 invlist = _add_range_to_invlist(invlist, start, i-1);
1725 new_node_has_latin1 = TRUE;
1726 }
1727 }
1728 }
1729
1730 /* If this can match all upper Latin1 code points, have to add them
1731 * as well. But don't add them if inverting, as when that gets done below,
1732 * it would exclude all these characters, including the ones it shouldn't
1733 * that were added just above */
1734 if (! (flags & ANYOF_INVERT) && OP(node) == ANYOFD
1735 && (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1736 {
1737 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1738 }
1739
1740 /* Similarly for these */
1741 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1742 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1743 }
1744
1745 if (flags & ANYOF_INVERT) {
1746 _invlist_invert(invlist);
1747 }
1748 else if (flags & ANYOFL_FOLD) {
1749 if (new_node_has_latin1) {
1750
1751 /* Under /li, any 0-255 could fold to any other 0-255, depending on
1752 * the locale. We can skip this if there are no 0-255 at all. */
1753 _invlist_union(invlist, PL_Latin1, &invlist);
1754
1755 invlist = add_cp_to_invlist(invlist, LATIN_SMALL_LETTER_DOTLESS_I);
1756 invlist = add_cp_to_invlist(invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
1757 }
1758 else {
1759 if (_invlist_contains_cp(invlist, LATIN_SMALL_LETTER_DOTLESS_I)) {
1760 invlist = add_cp_to_invlist(invlist, 'I');
1761 }
1762 if (_invlist_contains_cp(invlist,
1763 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE))
1764 {
1765 invlist = add_cp_to_invlist(invlist, 'i');
1766 }
1767 }
1768 }
1769
1770 /* Similarly add the UTF-8 locale possible matches. These have to be
1771 * deferred until after the non-UTF-8 locale ones are taken care of just
1772 * above, or it leads to wrong results under ANYOF_INVERT */
1773 if (only_utf8_locale_invlist) {
1774 _invlist_union_maybe_complement_2nd(invlist,
1775 only_utf8_locale_invlist,
1776 flags & ANYOF_INVERT,
1777 &invlist);
1778 }
1779
1780 return invlist;
1781 }
1782
1783 /* These two functions currently do the exact same thing */
1784 #define ssc_init_zero ssc_init
1785
1786 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1787 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1788
1789 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1790 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1791 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1792
1793 STATIC void
S_ssc_and(pTHX_ const RExC_state_t * pRExC_state,regnode_ssc * ssc,const regnode_charclass * and_with)1794 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1795 const regnode_charclass *and_with)
1796 {
1797 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1798 * another SSC or a regular ANYOF class. Can create false positives. */
1799
1800 SV* anded_cp_list;
1801 U8 and_with_flags = inRANGE(OP(and_with), ANYOFH, ANYOFRb)
1802 ? 0
1803 : ANYOF_FLAGS(and_with);
1804 U8 anded_flags;
1805
1806 PERL_ARGS_ASSERT_SSC_AND;
1807
1808 assert(is_ANYOF_SYNTHETIC(ssc));
1809
1810 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1811 * the code point inversion list and just the relevant flags */
1812 if (is_ANYOF_SYNTHETIC(and_with)) {
1813 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1814 anded_flags = and_with_flags;
1815
1816 /* XXX This is a kludge around what appears to be deficiencies in the
1817 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1818 * there are paths through the optimizer where it doesn't get weeded
1819 * out when it should. And if we don't make some extra provision for
1820 * it like the code just below, it doesn't get added when it should.
1821 * This solution is to add it only when AND'ing, which is here, and
1822 * only when what is being AND'ed is the pristine, original node
1823 * matching anything. Thus it is like adding it to ssc_anything() but
1824 * only when the result is to be AND'ed. Probably the same solution
1825 * could be adopted for the same problem we have with /l matching,
1826 * which is solved differently in S_ssc_init(), and that would lead to
1827 * fewer false positives than that solution has. But if this solution
1828 * creates bugs, the consequences are only that a warning isn't raised
1829 * that should be; while the consequences for having /l bugs is
1830 * incorrect matches */
1831 if (ssc_is_anything((regnode_ssc *)and_with)) {
1832 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1833 }
1834 }
1835 else {
1836 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1837 if (OP(and_with) == ANYOFD) {
1838 anded_flags = and_with_flags & ANYOF_COMMON_FLAGS;
1839 }
1840 else {
1841 anded_flags = and_with_flags
1842 &( ANYOF_COMMON_FLAGS
1843 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1844 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1845 if (ANYOFL_UTF8_LOCALE_REQD(and_with_flags)) {
1846 anded_flags &=
1847 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1848 }
1849 }
1850 }
1851
1852 ANYOF_FLAGS(ssc) &= anded_flags;
1853
1854 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1855 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1856 * 'and_with' may be inverted. When not inverted, we have the situation of
1857 * computing:
1858 * (C1 | P1) & (C2 | P2)
1859 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1860 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1861 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1862 * <= ((C1 & C2) | P1 | P2)
1863 * Alternatively, the last few steps could be:
1864 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1865 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1866 * <= (C1 | C2 | (P1 & P2))
1867 * We favor the second approach if either P1 or P2 is non-empty. This is
1868 * because these components are a barrier to doing optimizations, as what
1869 * they match cannot be known until the moment of matching as they are
1870 * dependent on the current locale, 'AND"ing them likely will reduce or
1871 * eliminate them.
1872 * But we can do better if we know that C1,P1 are in their initial state (a
1873 * frequent occurrence), each matching everything:
1874 * (<everything>) & (C2 | P2) = C2 | P2
1875 * Similarly, if C2,P2 are in their initial state (again a frequent
1876 * occurrence), the result is a no-op
1877 * (C1 | P1) & (<everything>) = C1 | P1
1878 *
1879 * Inverted, we have
1880 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1881 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1882 * <= (C1 & ~C2) | (P1 & ~P2)
1883 * */
1884
1885 if ((and_with_flags & ANYOF_INVERT)
1886 && ! is_ANYOF_SYNTHETIC(and_with))
1887 {
1888 unsigned int i;
1889
1890 ssc_intersection(ssc,
1891 anded_cp_list,
1892 FALSE /* Has already been inverted */
1893 );
1894
1895 /* If either P1 or P2 is empty, the intersection will be also; can skip
1896 * the loop */
1897 if (! (and_with_flags & ANYOF_MATCHES_POSIXL)) {
1898 ANYOF_POSIXL_ZERO(ssc);
1899 }
1900 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1901
1902 /* Note that the Posix class component P from 'and_with' actually
1903 * looks like:
1904 * P = Pa | Pb | ... | Pn
1905 * where each component is one posix class, such as in [\w\s].
1906 * Thus
1907 * ~P = ~(Pa | Pb | ... | Pn)
1908 * = ~Pa & ~Pb & ... & ~Pn
1909 * <= ~Pa | ~Pb | ... | ~Pn
1910 * The last is something we can easily calculate, but unfortunately
1911 * is likely to have many false positives. We could do better
1912 * in some (but certainly not all) instances if two classes in
1913 * P have known relationships. For example
1914 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1915 * So
1916 * :lower: & :print: = :lower:
1917 * And similarly for classes that must be disjoint. For example,
1918 * since \s and \w can have no elements in common based on rules in
1919 * the POSIX standard,
1920 * \w & ^\S = nothing
1921 * Unfortunately, some vendor locales do not meet the Posix
1922 * standard, in particular almost everything by Microsoft.
1923 * The loop below just changes e.g., \w into \W and vice versa */
1924
1925 regnode_charclass_posixl temp;
1926 int add = 1; /* To calculate the index of the complement */
1927
1928 Zero(&temp, 1, regnode_charclass_posixl);
1929 ANYOF_POSIXL_ZERO(&temp);
1930 for (i = 0; i < ANYOF_MAX; i++) {
1931 assert(i % 2 != 0
1932 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1933 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1934
1935 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1936 ANYOF_POSIXL_SET(&temp, i + add);
1937 }
1938 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1939 }
1940 ANYOF_POSIXL_AND(&temp, ssc);
1941
1942 } /* else ssc already has no posixes */
1943 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1944 in its initial state */
1945 else if (! is_ANYOF_SYNTHETIC(and_with)
1946 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1947 {
1948 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1949 * copy it over 'ssc' */
1950 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1951 if (is_ANYOF_SYNTHETIC(and_with)) {
1952 StructCopy(and_with, ssc, regnode_ssc);
1953 }
1954 else {
1955 ssc->invlist = anded_cp_list;
1956 ANYOF_POSIXL_ZERO(ssc);
1957 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1958 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1959 }
1960 }
1961 }
1962 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1963 || (and_with_flags & ANYOF_MATCHES_POSIXL))
1964 {
1965 /* One or the other of P1, P2 is non-empty. */
1966 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1967 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1968 }
1969 ssc_union(ssc, anded_cp_list, FALSE);
1970 }
1971 else { /* P1 = P2 = empty */
1972 ssc_intersection(ssc, anded_cp_list, FALSE);
1973 }
1974 }
1975 }
1976
1977 STATIC void
S_ssc_or(pTHX_ const RExC_state_t * pRExC_state,regnode_ssc * ssc,const regnode_charclass * or_with)1978 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1979 const regnode_charclass *or_with)
1980 {
1981 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1982 * another SSC or a regular ANYOF class. Can create false positives if
1983 * 'or_with' is to be inverted. */
1984
1985 SV* ored_cp_list;
1986 U8 ored_flags;
1987 U8 or_with_flags = inRANGE(OP(or_with), ANYOFH, ANYOFRb)
1988 ? 0
1989 : ANYOF_FLAGS(or_with);
1990
1991 PERL_ARGS_ASSERT_SSC_OR;
1992
1993 assert(is_ANYOF_SYNTHETIC(ssc));
1994
1995 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1996 * the code point inversion list and just the relevant flags */
1997 if (is_ANYOF_SYNTHETIC(or_with)) {
1998 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1999 ored_flags = or_with_flags;
2000 }
2001 else {
2002 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
2003 ored_flags = or_with_flags & ANYOF_COMMON_FLAGS;
2004 if (OP(or_with) != ANYOFD) {
2005 ored_flags
2006 |= or_with_flags
2007 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2008 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
2009 if (ANYOFL_UTF8_LOCALE_REQD(or_with_flags)) {
2010 ored_flags |=
2011 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
2012 }
2013 }
2014 }
2015
2016 ANYOF_FLAGS(ssc) |= ored_flags;
2017
2018 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
2019 * C2 is the list of code points in 'or-with'; P2, its posix classes.
2020 * 'or_with' may be inverted. When not inverted, we have the simple
2021 * situation of computing:
2022 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
2023 * If P1|P2 yields a situation with both a class and its complement are
2024 * set, like having both \w and \W, this matches all code points, and we
2025 * can delete these from the P component of the ssc going forward. XXX We
2026 * might be able to delete all the P components, but I (khw) am not certain
2027 * about this, and it is better to be safe.
2028 *
2029 * Inverted, we have
2030 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
2031 * <= (C1 | P1) | ~C2
2032 * <= (C1 | ~C2) | P1
2033 * (which results in actually simpler code than the non-inverted case)
2034 * */
2035
2036 if ((or_with_flags & ANYOF_INVERT)
2037 && ! is_ANYOF_SYNTHETIC(or_with))
2038 {
2039 /* We ignore P2, leaving P1 going forward */
2040 } /* else Not inverted */
2041 else if (or_with_flags & ANYOF_MATCHES_POSIXL) {
2042 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
2043 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2044 unsigned int i;
2045 for (i = 0; i < ANYOF_MAX; i += 2) {
2046 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
2047 {
2048 ssc_match_all_cp(ssc);
2049 ANYOF_POSIXL_CLEAR(ssc, i);
2050 ANYOF_POSIXL_CLEAR(ssc, i+1);
2051 }
2052 }
2053 }
2054 }
2055
2056 ssc_union(ssc,
2057 ored_cp_list,
2058 FALSE /* Already has been inverted */
2059 );
2060 }
2061
2062 STATIC void
S_ssc_union(pTHX_ regnode_ssc * ssc,SV * const invlist,const bool invert2nd)2063 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
2064 {
2065 PERL_ARGS_ASSERT_SSC_UNION;
2066
2067 assert(is_ANYOF_SYNTHETIC(ssc));
2068
2069 _invlist_union_maybe_complement_2nd(ssc->invlist,
2070 invlist,
2071 invert2nd,
2072 &ssc->invlist);
2073 }
2074
2075 STATIC void
S_ssc_intersection(pTHX_ regnode_ssc * ssc,SV * const invlist,const bool invert2nd)2076 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
2077 SV* const invlist,
2078 const bool invert2nd)
2079 {
2080 PERL_ARGS_ASSERT_SSC_INTERSECTION;
2081
2082 assert(is_ANYOF_SYNTHETIC(ssc));
2083
2084 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
2085 invlist,
2086 invert2nd,
2087 &ssc->invlist);
2088 }
2089
2090 STATIC void
S_ssc_add_range(pTHX_ regnode_ssc * ssc,const UV start,const UV end)2091 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
2092 {
2093 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
2094
2095 assert(is_ANYOF_SYNTHETIC(ssc));
2096
2097 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
2098 }
2099
2100 STATIC void
S_ssc_cp_and(pTHX_ regnode_ssc * ssc,const UV cp)2101 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
2102 {
2103 /* AND just the single code point 'cp' into the SSC 'ssc' */
2104
2105 SV* cp_list = _new_invlist(2);
2106
2107 PERL_ARGS_ASSERT_SSC_CP_AND;
2108
2109 assert(is_ANYOF_SYNTHETIC(ssc));
2110
2111 cp_list = add_cp_to_invlist(cp_list, cp);
2112 ssc_intersection(ssc, cp_list,
2113 FALSE /* Not inverted */
2114 );
2115 SvREFCNT_dec_NN(cp_list);
2116 }
2117
2118 STATIC void
S_ssc_clear_locale(regnode_ssc * ssc)2119 S_ssc_clear_locale(regnode_ssc *ssc)
2120 {
2121 /* Set the SSC 'ssc' to not match any locale things */
2122 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
2123
2124 assert(is_ANYOF_SYNTHETIC(ssc));
2125
2126 ANYOF_POSIXL_ZERO(ssc);
2127 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
2128 }
2129
2130 STATIC bool
S_is_ssc_worth_it(const RExC_state_t * pRExC_state,const regnode_ssc * ssc)2131 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
2132 {
2133 /* The synthetic start class is used to hopefully quickly winnow down
2134 * places where a pattern could start a match in the target string. If it
2135 * doesn't really narrow things down that much, there isn't much point to
2136 * having the overhead of using it. This function uses some very crude
2137 * heuristics to decide if to use the ssc or not.
2138 *
2139 * It returns TRUE if 'ssc' rules out more than half what it considers to
2140 * be the "likely" possible matches, but of course it doesn't know what the
2141 * actual things being matched are going to be; these are only guesses
2142 *
2143 * For /l matches, it assumes that the only likely matches are going to be
2144 * in the 0-255 range, uniformly distributed, so half of that is 127
2145 * For /a and /d matches, it assumes that the likely matches will be just
2146 * the ASCII range, so half of that is 63
2147 * For /u and there isn't anything matching above the Latin1 range, it
2148 * assumes that that is the only range likely to be matched, and uses
2149 * half that as the cut-off: 127. If anything matches above Latin1,
2150 * it assumes that all of Unicode could match (uniformly), except for
2151 * non-Unicode code points and things in the General Category "Other"
2152 * (unassigned, private use, surrogates, controls and formats). This
2153 * is a much large number. */
2154
2155 U32 count = 0; /* Running total of number of code points matched by
2156 'ssc' */
2157 UV start, end; /* Start and end points of current range in inversion
2158 XXX outdated. UTF-8 locales are common, what about invert? list */
2159 const U32 max_code_points = (LOC)
2160 ? 256
2161 : (( ! UNI_SEMANTICS
2162 || invlist_highest(ssc->invlist) < 256)
2163 ? 128
2164 : NON_OTHER_COUNT);
2165 const U32 max_match = max_code_points / 2;
2166
2167 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
2168
2169 invlist_iterinit(ssc->invlist);
2170 while (invlist_iternext(ssc->invlist, &start, &end)) {
2171 if (start >= max_code_points) {
2172 break;
2173 }
2174 end = MIN(end, max_code_points - 1);
2175 count += end - start + 1;
2176 if (count >= max_match) {
2177 invlist_iterfinish(ssc->invlist);
2178 return FALSE;
2179 }
2180 }
2181
2182 return TRUE;
2183 }
2184
2185
2186 STATIC void
S_ssc_finalize(pTHX_ RExC_state_t * pRExC_state,regnode_ssc * ssc)2187 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2188 {
2189 /* The inversion list in the SSC is marked mortal; now we need a more
2190 * permanent copy, which is stored the same way that is done in a regular
2191 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2192 * map */
2193
2194 SV* invlist = invlist_clone(ssc->invlist, NULL);
2195
2196 PERL_ARGS_ASSERT_SSC_FINALIZE;
2197
2198 assert(is_ANYOF_SYNTHETIC(ssc));
2199
2200 /* The code in this file assumes that all but these flags aren't relevant
2201 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2202 * by the time we reach here */
2203 assert(! (ANYOF_FLAGS(ssc)
2204 & ~( ANYOF_COMMON_FLAGS
2205 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2206 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
2207
2208 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2209
2210 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist, NULL, NULL);
2211 SvREFCNT_dec(invlist);
2212
2213 /* Make sure is clone-safe */
2214 ssc->invlist = NULL;
2215
2216 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2217 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2218 OP(ssc) = ANYOFPOSIXL;
2219 }
2220 else if (RExC_contains_locale) {
2221 OP(ssc) = ANYOFL;
2222 }
2223
2224 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2225 }
2226
2227 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2228 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2229 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2230 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2231 ? (TRIE_LIST_CUR( idx ) - 1) \
2232 : 0 )
2233
2234
2235 #ifdef DEBUGGING
2236 /*
2237 dump_trie(trie,widecharmap,revcharmap)
2238 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2239 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2240
2241 These routines dump out a trie in a somewhat readable format.
2242 The _interim_ variants are used for debugging the interim
2243 tables that are used to generate the final compressed
2244 representation which is what dump_trie expects.
2245
2246 Part of the reason for their existence is to provide a form
2247 of documentation as to how the different representations function.
2248
2249 */
2250
2251 /*
2252 Dumps the final compressed table form of the trie to Perl_debug_log.
2253 Used for debugging make_trie().
2254 */
2255
2256 STATIC void
S_dump_trie(pTHX_ const struct _reg_trie_data * trie,HV * widecharmap,AV * revcharmap,U32 depth)2257 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2258 AV *revcharmap, U32 depth)
2259 {
2260 U32 state;
2261 SV *sv=sv_newmortal();
2262 int colwidth= widecharmap ? 6 : 4;
2263 U16 word;
2264 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2265
2266 PERL_ARGS_ASSERT_DUMP_TRIE;
2267
2268 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2269 depth+1, "Match","Base","Ofs" );
2270
2271 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2272 SV ** const tmp = av_fetch( revcharmap, state, 0);
2273 if ( tmp ) {
2274 Perl_re_printf( aTHX_ "%*s",
2275 colwidth,
2276 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2277 PL_colors[0], PL_colors[1],
2278 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2279 PERL_PV_ESCAPE_FIRSTCHAR
2280 )
2281 );
2282 }
2283 }
2284 Perl_re_printf( aTHX_ "\n");
2285 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2286
2287 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2288 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2289 Perl_re_printf( aTHX_ "\n");
2290
2291 for( state = 1 ; state < trie->statecount ; state++ ) {
2292 const U32 base = trie->states[ state ].trans.base;
2293
2294 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2295
2296 if ( trie->states[ state ].wordnum ) {
2297 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2298 } else {
2299 Perl_re_printf( aTHX_ "%6s", "" );
2300 }
2301
2302 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2303
2304 if ( base ) {
2305 U32 ofs = 0;
2306
2307 while( ( base + ofs < trie->uniquecharcount ) ||
2308 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2309 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2310 != state))
2311 ofs++;
2312
2313 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2314
2315 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2316 if ( ( base + ofs >= trie->uniquecharcount )
2317 && ( base + ofs - trie->uniquecharcount
2318 < trie->lasttrans )
2319 && trie->trans[ base + ofs
2320 - trie->uniquecharcount ].check == state )
2321 {
2322 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2323 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2324 );
2325 } else {
2326 Perl_re_printf( aTHX_ "%*s", colwidth," ." );
2327 }
2328 }
2329
2330 Perl_re_printf( aTHX_ "]");
2331
2332 }
2333 Perl_re_printf( aTHX_ "\n" );
2334 }
2335 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2336 depth);
2337 for (word=1; word <= trie->wordcount; word++) {
2338 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2339 (int)word, (int)(trie->wordinfo[word].prev),
2340 (int)(trie->wordinfo[word].len));
2341 }
2342 Perl_re_printf( aTHX_ "\n" );
2343 }
2344 /*
2345 Dumps a fully constructed but uncompressed trie in list form.
2346 List tries normally only are used for construction when the number of
2347 possible chars (trie->uniquecharcount) is very high.
2348 Used for debugging make_trie().
2349 */
2350 STATIC void
S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data * trie,HV * widecharmap,AV * revcharmap,U32 next_alloc,U32 depth)2351 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2352 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2353 U32 depth)
2354 {
2355 U32 state;
2356 SV *sv=sv_newmortal();
2357 int colwidth= widecharmap ? 6 : 4;
2358 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2359
2360 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2361
2362 /* print out the table precompression. */
2363 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2364 depth+1 );
2365 Perl_re_indentf( aTHX_ "%s",
2366 depth+1, "------:-----+-----------------\n" );
2367
2368 for( state=1 ; state < next_alloc ; state ++ ) {
2369 U16 charid;
2370
2371 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2372 depth+1, (UV)state );
2373 if ( ! trie->states[ state ].wordnum ) {
2374 Perl_re_printf( aTHX_ "%5s| ","");
2375 } else {
2376 Perl_re_printf( aTHX_ "W%4x| ",
2377 trie->states[ state ].wordnum
2378 );
2379 }
2380 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2381 SV ** const tmp = av_fetch( revcharmap,
2382 TRIE_LIST_ITEM(state, charid).forid, 0);
2383 if ( tmp ) {
2384 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2385 colwidth,
2386 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2387 colwidth,
2388 PL_colors[0], PL_colors[1],
2389 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2390 | PERL_PV_ESCAPE_FIRSTCHAR
2391 ) ,
2392 TRIE_LIST_ITEM(state, charid).forid,
2393 (UV)TRIE_LIST_ITEM(state, charid).newstate
2394 );
2395 if (!(charid % 10))
2396 Perl_re_printf( aTHX_ "\n%*s| ",
2397 (int)((depth * 2) + 14), "");
2398 }
2399 }
2400 Perl_re_printf( aTHX_ "\n");
2401 }
2402 }
2403
2404 /*
2405 Dumps a fully constructed but uncompressed trie in table form.
2406 This is the normal DFA style state transition table, with a few
2407 twists to facilitate compression later.
2408 Used for debugging make_trie().
2409 */
2410 STATIC void
S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data * trie,HV * widecharmap,AV * revcharmap,U32 next_alloc,U32 depth)2411 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2412 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2413 U32 depth)
2414 {
2415 U32 state;
2416 U16 charid;
2417 SV *sv=sv_newmortal();
2418 int colwidth= widecharmap ? 6 : 4;
2419 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2420
2421 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2422
2423 /*
2424 print out the table precompression so that we can do a visual check
2425 that they are identical.
2426 */
2427
2428 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2429
2430 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2431 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2432 if ( tmp ) {
2433 Perl_re_printf( aTHX_ "%*s",
2434 colwidth,
2435 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2436 PL_colors[0], PL_colors[1],
2437 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2438 PERL_PV_ESCAPE_FIRSTCHAR
2439 )
2440 );
2441 }
2442 }
2443
2444 Perl_re_printf( aTHX_ "\n");
2445 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2446
2447 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2448 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2449 }
2450
2451 Perl_re_printf( aTHX_ "\n" );
2452
2453 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2454
2455 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2456 depth+1,
2457 (UV)TRIE_NODENUM( state ) );
2458
2459 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2460 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2461 if (v)
2462 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2463 else
2464 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2465 }
2466 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2467 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2468 (UV)trie->trans[ state ].check );
2469 } else {
2470 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2471 (UV)trie->trans[ state ].check,
2472 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2473 }
2474 }
2475 }
2476
2477 #endif
2478
2479
2480 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2481 startbranch: the first branch in the whole branch sequence
2482 first : start branch of sequence of branch-exact nodes.
2483 May be the same as startbranch
2484 last : Thing following the last branch.
2485 May be the same as tail.
2486 tail : item following the branch sequence
2487 count : words in the sequence
2488 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2489 depth : indent depth
2490
2491 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2492
2493 A trie is an N'ary tree where the branches are determined by digital
2494 decomposition of the key. IE, at the root node you look up the 1st character and
2495 follow that branch repeat until you find the end of the branches. Nodes can be
2496 marked as "accepting" meaning they represent a complete word. Eg:
2497
2498 /he|she|his|hers/
2499
2500 would convert into the following structure. Numbers represent states, letters
2501 following numbers represent valid transitions on the letter from that state, if
2502 the number is in square brackets it represents an accepting state, otherwise it
2503 will be in parenthesis.
2504
2505 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2506 | |
2507 | (2)
2508 | |
2509 (1) +-i->(6)-+-s->[7]
2510 |
2511 +-s->(3)-+-h->(4)-+-e->[5]
2512
2513 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2514
2515 This shows that when matching against the string 'hers' we will begin at state 1
2516 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2517 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2518 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2519 single traverse. We store a mapping from accepting to state to which word was
2520 matched, and then when we have multiple possibilities we try to complete the
2521 rest of the regex in the order in which they occurred in the alternation.
2522
2523 The only prior NFA like behaviour that would be changed by the TRIE support is
2524 the silent ignoring of duplicate alternations which are of the form:
2525
2526 / (DUPE|DUPE) X? (?{ ... }) Y /x
2527
2528 Thus EVAL blocks following a trie may be called a different number of times with
2529 and without the optimisation. With the optimisations dupes will be silently
2530 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2531 the following demonstrates:
2532
2533 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2534
2535 which prints out 'word' three times, but
2536
2537 'words'=~/(word|word|word)(?{ print $1 })S/
2538
2539 which doesnt print it out at all. This is due to other optimisations kicking in.
2540
2541 Example of what happens on a structural level:
2542
2543 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2544
2545 1: CURLYM[1] {1,32767}(18)
2546 5: BRANCH(8)
2547 6: EXACT <ac>(16)
2548 8: BRANCH(11)
2549 9: EXACT <ad>(16)
2550 11: BRANCH(14)
2551 12: EXACT <ab>(16)
2552 16: SUCCEED(0)
2553 17: NOTHING(18)
2554 18: END(0)
2555
2556 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2557 and should turn into:
2558
2559 1: CURLYM[1] {1,32767}(18)
2560 5: TRIE(16)
2561 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2562 <ac>
2563 <ad>
2564 <ab>
2565 16: SUCCEED(0)
2566 17: NOTHING(18)
2567 18: END(0)
2568
2569 Cases where tail != last would be like /(?foo|bar)baz/:
2570
2571 1: BRANCH(4)
2572 2: EXACT <foo>(8)
2573 4: BRANCH(7)
2574 5: EXACT <bar>(8)
2575 7: TAIL(8)
2576 8: EXACT <baz>(10)
2577 10: END(0)
2578
2579 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2580 and would end up looking like:
2581
2582 1: TRIE(8)
2583 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2584 <foo>
2585 <bar>
2586 7: TAIL(8)
2587 8: EXACT <baz>(10)
2588 10: END(0)
2589
2590 d = uvchr_to_utf8_flags(d, uv, 0);
2591
2592 is the recommended Unicode-aware way of saying
2593
2594 *(d++) = uv;
2595 */
2596
2597 #define TRIE_STORE_REVCHAR(val) \
2598 STMT_START { \
2599 if (UTF) { \
2600 SV *zlopp = newSV(UTF8_MAXBYTES); \
2601 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2602 unsigned char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2603 *kapow = '\0'; \
2604 SvCUR_set(zlopp, kapow - flrbbbbb); \
2605 SvPOK_on(zlopp); \
2606 SvUTF8_on(zlopp); \
2607 av_push(revcharmap, zlopp); \
2608 } else { \
2609 char ooooff = (char)val; \
2610 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2611 } \
2612 } STMT_END
2613
2614 /* This gets the next character from the input, folding it if not already
2615 * folded. */
2616 #define TRIE_READ_CHAR STMT_START { \
2617 wordlen++; \
2618 if ( UTF ) { \
2619 /* if it is UTF then it is either already folded, or does not need \
2620 * folding */ \
2621 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2622 } \
2623 else if (folder == PL_fold_latin1) { \
2624 /* This folder implies Unicode rules, which in the range expressible \
2625 * by not UTF is the lower case, with the two exceptions, one of \
2626 * which should have been taken care of before calling this */ \
2627 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2628 uvc = toLOWER_L1(*uc); \
2629 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2630 len = 1; \
2631 } else { \
2632 /* raw data, will be folded later if needed */ \
2633 uvc = (U32)*uc; \
2634 len = 1; \
2635 } \
2636 } STMT_END
2637
2638
2639
2640 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2641 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2642 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2643 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2644 TRIE_LIST_LEN( state ) = ging; \
2645 } \
2646 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2647 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2648 TRIE_LIST_CUR( state )++; \
2649 } STMT_END
2650
2651 #define TRIE_LIST_NEW(state) STMT_START { \
2652 Newx( trie->states[ state ].trans.list, \
2653 4, reg_trie_trans_le ); \
2654 TRIE_LIST_CUR( state ) = 1; \
2655 TRIE_LIST_LEN( state ) = 4; \
2656 } STMT_END
2657
2658 #define TRIE_HANDLE_WORD(state) STMT_START { \
2659 U16 dupe= trie->states[ state ].wordnum; \
2660 regnode * const noper_next = regnext( noper ); \
2661 \
2662 DEBUG_r({ \
2663 /* store the word for dumping */ \
2664 SV* tmp; \
2665 if (OP(noper) != NOTHING) \
2666 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2667 else \
2668 tmp = newSVpvn_utf8( "", 0, UTF ); \
2669 av_push( trie_words, tmp ); \
2670 }); \
2671 \
2672 curword++; \
2673 trie->wordinfo[curword].prev = 0; \
2674 trie->wordinfo[curword].len = wordlen; \
2675 trie->wordinfo[curword].accept = state; \
2676 \
2677 if ( noper_next < tail ) { \
2678 if (!trie->jump) \
2679 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2680 sizeof(U16) ); \
2681 trie->jump[curword] = (U16)(noper_next - convert); \
2682 if (!jumper) \
2683 jumper = noper_next; \
2684 if (!nextbranch) \
2685 nextbranch= regnext(cur); \
2686 } \
2687 \
2688 if ( dupe ) { \
2689 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2690 /* chain, so that when the bits of chain are later */\
2691 /* linked together, the dups appear in the chain */\
2692 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2693 trie->wordinfo[dupe].prev = curword; \
2694 } else { \
2695 /* we haven't inserted this word yet. */ \
2696 trie->states[ state ].wordnum = curword; \
2697 } \
2698 } STMT_END
2699
2700
2701 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2702 ( ( base + charid >= ucharcount \
2703 && base + charid < ubound \
2704 && state == trie->trans[ base - ucharcount + charid ].check \
2705 && trie->trans[ base - ucharcount + charid ].next ) \
2706 ? trie->trans[ base - ucharcount + charid ].next \
2707 : ( state==1 ? special : 0 ) \
2708 )
2709
2710 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2711 STMT_START { \
2712 TRIE_BITMAP_SET(trie, uvc); \
2713 /* store the folded codepoint */ \
2714 if ( folder ) \
2715 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2716 \
2717 if ( !UTF ) { \
2718 /* store first byte of utf8 representation of */ \
2719 /* variant codepoints */ \
2720 if (! UVCHR_IS_INVARIANT(uvc)) { \
2721 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2722 } \
2723 } \
2724 } STMT_END
2725 #define MADE_TRIE 1
2726 #define MADE_JUMP_TRIE 2
2727 #define MADE_EXACT_TRIE 4
2728
2729 STATIC I32
S_make_trie(pTHX_ RExC_state_t * pRExC_state,regnode * startbranch,regnode * first,regnode * last,regnode * tail,U32 word_count,U32 flags,U32 depth)2730 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2731 regnode *first, regnode *last, regnode *tail,
2732 U32 word_count, U32 flags, U32 depth)
2733 {
2734 /* first pass, loop through and scan words */
2735 reg_trie_data *trie;
2736 HV *widecharmap = NULL;
2737 AV *revcharmap = newAV();
2738 regnode *cur;
2739 STRLEN len = 0;
2740 UV uvc = 0;
2741 U16 curword = 0;
2742 U32 next_alloc = 0;
2743 regnode *jumper = NULL;
2744 regnode *nextbranch = NULL;
2745 regnode *convert = NULL;
2746 U32 *prev_states; /* temp array mapping each state to previous one */
2747 /* we just use folder as a flag in utf8 */
2748 const U8 * folder = NULL;
2749
2750 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2751 * which stands for one trie structure, one hash, optionally followed
2752 * by two arrays */
2753 #ifdef DEBUGGING
2754 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2755 AV *trie_words = NULL;
2756 /* along with revcharmap, this only used during construction but both are
2757 * useful during debugging so we store them in the struct when debugging.
2758 */
2759 #else
2760 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2761 STRLEN trie_charcount=0;
2762 #endif
2763 SV *re_trie_maxbuff;
2764 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2765
2766 PERL_ARGS_ASSERT_MAKE_TRIE;
2767 #ifndef DEBUGGING
2768 PERL_UNUSED_ARG(depth);
2769 #endif
2770
2771 switch (flags) {
2772 case EXACT: case EXACT_REQ8: case EXACTL: break;
2773 case EXACTFAA:
2774 case EXACTFUP:
2775 case EXACTFU:
2776 case EXACTFLU8: folder = PL_fold_latin1; break;
2777 case EXACTF: folder = PL_fold; break;
2778 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2779 }
2780
2781 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2782 trie->refcount = 1;
2783 trie->startstate = 1;
2784 trie->wordcount = word_count;
2785 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2786 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2787 if (flags == EXACT || flags == EXACT_REQ8 || flags == EXACTL)
2788 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2789 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2790 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2791
2792 DEBUG_r({
2793 trie_words = newAV();
2794 });
2795
2796 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, GV_ADD);
2797 assert(re_trie_maxbuff);
2798 if (!SvIOK(re_trie_maxbuff)) {
2799 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2800 }
2801 DEBUG_TRIE_COMPILE_r({
2802 Perl_re_indentf( aTHX_
2803 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2804 depth+1,
2805 REG_NODE_NUM(startbranch), REG_NODE_NUM(first),
2806 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2807 });
2808
2809 /* Find the node we are going to overwrite */
2810 if ( first == startbranch && OP( last ) != BRANCH ) {
2811 /* whole branch chain */
2812 convert = first;
2813 } else {
2814 /* branch sub-chain */
2815 convert = NEXTOPER( first );
2816 }
2817
2818 /* -- First loop and Setup --
2819
2820 We first traverse the branches and scan each word to determine if it
2821 contains widechars, and how many unique chars there are, this is
2822 important as we have to build a table with at least as many columns as we
2823 have unique chars.
2824
2825 We use an array of integers to represent the character codes 0..255
2826 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2827 the native representation of the character value as the key and IV's for
2828 the coded index.
2829
2830 *TODO* If we keep track of how many times each character is used we can
2831 remap the columns so that the table compression later on is more
2832 efficient in terms of memory by ensuring the most common value is in the
2833 middle and the least common are on the outside. IMO this would be better
2834 than a most to least common mapping as theres a decent chance the most
2835 common letter will share a node with the least common, meaning the node
2836 will not be compressible. With a middle is most common approach the worst
2837 case is when we have the least common nodes twice.
2838
2839 */
2840
2841 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2842 regnode *noper = NEXTOPER( cur );
2843 const U8 *uc;
2844 const U8 *e;
2845 int foldlen = 0;
2846 U32 wordlen = 0; /* required init */
2847 STRLEN minchars = 0;
2848 STRLEN maxchars = 0;
2849 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2850 bitmap?*/
2851
2852 if (OP(noper) == NOTHING) {
2853 /* skip past a NOTHING at the start of an alternation
2854 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2855 *
2856 * If the next node is not something we are supposed to process
2857 * we will just ignore it due to the condition guarding the
2858 * next block.
2859 */
2860
2861 regnode *noper_next= regnext(noper);
2862 if (noper_next < tail)
2863 noper= noper_next;
2864 }
2865
2866 if ( noper < tail
2867 && ( OP(noper) == flags
2868 || (flags == EXACT && OP(noper) == EXACT_REQ8)
2869 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
2870 || OP(noper) == EXACTFUP))))
2871 {
2872 uc= (U8*)STRING(noper);
2873 e= uc + STR_LEN(noper);
2874 } else {
2875 trie->minlen= 0;
2876 continue;
2877 }
2878
2879
2880 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2881 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2882 regardless of encoding */
2883 if (OP( noper ) == EXACTFUP) {
2884 /* false positives are ok, so just set this */
2885 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2886 }
2887 }
2888
2889 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2890 branch */
2891 TRIE_CHARCOUNT(trie)++;
2892 TRIE_READ_CHAR;
2893
2894 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2895 * is in effect. Under /i, this character can match itself, or
2896 * anything that folds to it. If not under /i, it can match just
2897 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2898 * all fold to k, and all are single characters. But some folds
2899 * expand to more than one character, so for example LATIN SMALL
2900 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2901 * the string beginning at 'uc' is 'ffi', it could be matched by
2902 * three characters, or just by the one ligature character. (It
2903 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2904 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2905 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2906 * match.) The trie needs to know the minimum and maximum number
2907 * of characters that could match so that it can use size alone to
2908 * quickly reject many match attempts. The max is simple: it is
2909 * the number of folded characters in this branch (since a fold is
2910 * never shorter than what folds to it. */
2911
2912 maxchars++;
2913
2914 /* And the min is equal to the max if not under /i (indicated by
2915 * 'folder' being NULL), or there are no multi-character folds. If
2916 * there is a multi-character fold, the min is incremented just
2917 * once, for the character that folds to the sequence. Each
2918 * character in the sequence needs to be added to the list below of
2919 * characters in the trie, but we count only the first towards the
2920 * min number of characters needed. This is done through the
2921 * variable 'foldlen', which is returned by the macros that look
2922 * for these sequences as the number of bytes the sequence
2923 * occupies. Each time through the loop, we decrement 'foldlen' by
2924 * how many bytes the current char occupies. Only when it reaches
2925 * 0 do we increment 'minchars' or look for another multi-character
2926 * sequence. */
2927 if (folder == NULL) {
2928 minchars++;
2929 }
2930 else if (foldlen > 0) {
2931 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2932 }
2933 else {
2934 minchars++;
2935
2936 /* See if *uc is the beginning of a multi-character fold. If
2937 * so, we decrement the length remaining to look at, to account
2938 * for the current character this iteration. (We can use 'uc'
2939 * instead of the fold returned by TRIE_READ_CHAR because the
2940 * macro is smart enough to account for any unfolded
2941 * characters. */
2942 if (UTF) {
2943 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2944 foldlen -= UTF8SKIP(uc);
2945 }
2946 }
2947 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2948 foldlen--;
2949 }
2950 }
2951
2952 /* The current character (and any potential folds) should be added
2953 * to the possible matching characters for this position in this
2954 * branch */
2955 if ( uvc < 256 ) {
2956 if ( folder ) {
2957 U8 folded= folder[ (U8) uvc ];
2958 if ( !trie->charmap[ folded ] ) {
2959 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2960 TRIE_STORE_REVCHAR( folded );
2961 }
2962 }
2963 if ( !trie->charmap[ uvc ] ) {
2964 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2965 TRIE_STORE_REVCHAR( uvc );
2966 }
2967 if ( set_bit ) {
2968 /* store the codepoint in the bitmap, and its folded
2969 * equivalent. */
2970 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2971 set_bit = 0; /* We've done our bit :-) */
2972 }
2973 } else {
2974
2975 /* XXX We could come up with the list of code points that fold
2976 * to this using PL_utf8_foldclosures, except not for
2977 * multi-char folds, as there may be multiple combinations
2978 * there that could work, which needs to wait until runtime to
2979 * resolve (The comment about LIGATURE FFI above is such an
2980 * example */
2981
2982 SV** svpp;
2983 if ( !widecharmap )
2984 widecharmap = newHV();
2985
2986 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2987
2988 if ( !svpp )
2989 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2990
2991 if ( !SvTRUE( *svpp ) ) {
2992 sv_setiv( *svpp, ++trie->uniquecharcount );
2993 TRIE_STORE_REVCHAR(uvc);
2994 }
2995 }
2996 } /* end loop through characters in this branch of the trie */
2997
2998 /* We take the min and max for this branch and combine to find the min
2999 * and max for all branches processed so far */
3000 if( cur == first ) {
3001 trie->minlen = minchars;
3002 trie->maxlen = maxchars;
3003 } else if (minchars < trie->minlen) {
3004 trie->minlen = minchars;
3005 } else if (maxchars > trie->maxlen) {
3006 trie->maxlen = maxchars;
3007 }
3008 } /* end first pass */
3009 DEBUG_TRIE_COMPILE_r(
3010 Perl_re_indentf( aTHX_
3011 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
3012 depth+1,
3013 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
3014 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
3015 (int)trie->minlen, (int)trie->maxlen )
3016 );
3017
3018 /*
3019 We now know what we are dealing with in terms of unique chars and
3020 string sizes so we can calculate how much memory a naive
3021 representation using a flat table will take. If it's over a reasonable
3022 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
3023 conservative but potentially much slower representation using an array
3024 of lists.
3025
3026 At the end we convert both representations into the same compressed
3027 form that will be used in regexec.c for matching with. The latter
3028 is a form that cannot be used to construct with but has memory
3029 properties similar to the list form and access properties similar
3030 to the table form making it both suitable for fast searches and
3031 small enough that its feasable to store for the duration of a program.
3032
3033 See the comment in the code where the compressed table is produced
3034 inplace from the flat tabe representation for an explanation of how
3035 the compression works.
3036
3037 */
3038
3039
3040 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
3041 prev_states[1] = 0;
3042
3043 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
3044 > SvIV(re_trie_maxbuff) )
3045 {
3046 /*
3047 Second Pass -- Array Of Lists Representation
3048
3049 Each state will be represented by a list of charid:state records
3050 (reg_trie_trans_le) the first such element holds the CUR and LEN
3051 points of the allocated array. (See defines above).
3052
3053 We build the initial structure using the lists, and then convert
3054 it into the compressed table form which allows faster lookups
3055 (but cant be modified once converted).
3056 */
3057
3058 STRLEN transcount = 1;
3059
3060 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
3061 depth+1));
3062
3063 trie->states = (reg_trie_state *)
3064 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3065 sizeof(reg_trie_state) );
3066 TRIE_LIST_NEW(1);
3067 next_alloc = 2;
3068
3069 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3070
3071 regnode *noper = NEXTOPER( cur );
3072 U32 state = 1; /* required init */
3073 U16 charid = 0; /* sanity init */
3074 U32 wordlen = 0; /* required init */
3075
3076 if (OP(noper) == NOTHING) {
3077 regnode *noper_next= regnext(noper);
3078 if (noper_next < tail)
3079 noper= noper_next;
3080 /* we will undo this assignment if noper does not
3081 * point at a trieable type in the else clause of
3082 * the following statement. */
3083 }
3084
3085 if ( noper < tail
3086 && ( OP(noper) == flags
3087 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3088 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
3089 || OP(noper) == EXACTFUP))))
3090 {
3091 const U8 *uc= (U8*)STRING(noper);
3092 const U8 *e= uc + STR_LEN(noper);
3093
3094 for ( ; uc < e ; uc += len ) {
3095
3096 TRIE_READ_CHAR;
3097
3098 if ( uvc < 256 ) {
3099 charid = trie->charmap[ uvc ];
3100 } else {
3101 SV** const svpp = hv_fetch( widecharmap,
3102 (char*)&uvc,
3103 sizeof( UV ),
3104 0);
3105 if ( !svpp ) {
3106 charid = 0;
3107 } else {
3108 charid=(U16)SvIV( *svpp );
3109 }
3110 }
3111 /* charid is now 0 if we dont know the char read, or
3112 * nonzero if we do */
3113 if ( charid ) {
3114
3115 U16 check;
3116 U32 newstate = 0;
3117
3118 charid--;
3119 if ( !trie->states[ state ].trans.list ) {
3120 TRIE_LIST_NEW( state );
3121 }
3122 for ( check = 1;
3123 check <= TRIE_LIST_USED( state );
3124 check++ )
3125 {
3126 if ( TRIE_LIST_ITEM( state, check ).forid
3127 == charid )
3128 {
3129 newstate = TRIE_LIST_ITEM( state, check ).newstate;
3130 break;
3131 }
3132 }
3133 if ( ! newstate ) {
3134 newstate = next_alloc++;
3135 prev_states[newstate] = state;
3136 TRIE_LIST_PUSH( state, charid, newstate );
3137 transcount++;
3138 }
3139 state = newstate;
3140 } else {
3141 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3142 }
3143 }
3144 } else {
3145 /* If we end up here it is because we skipped past a NOTHING, but did not end up
3146 * on a trieable type. So we need to reset noper back to point at the first regop
3147 * in the branch before we call TRIE_HANDLE_WORD()
3148 */
3149 noper= NEXTOPER(cur);
3150 }
3151 TRIE_HANDLE_WORD(state);
3152
3153 } /* end second pass */
3154
3155 /* next alloc is the NEXT state to be allocated */
3156 trie->statecount = next_alloc;
3157 trie->states = (reg_trie_state *)
3158 PerlMemShared_realloc( trie->states,
3159 next_alloc
3160 * sizeof(reg_trie_state) );
3161
3162 /* and now dump it out before we compress it */
3163 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
3164 revcharmap, next_alloc,
3165 depth+1)
3166 );
3167
3168 trie->trans = (reg_trie_trans *)
3169 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
3170 {
3171 U32 state;
3172 U32 tp = 0;
3173 U32 zp = 0;
3174
3175
3176 for( state=1 ; state < next_alloc ; state ++ ) {
3177 U32 base=0;
3178
3179 /*
3180 DEBUG_TRIE_COMPILE_MORE_r(
3181 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
3182 );
3183 */
3184
3185 if (trie->states[state].trans.list) {
3186 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
3187 U16 maxid=minid;
3188 U16 idx;
3189
3190 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3191 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3192 if ( forid < minid ) {
3193 minid=forid;
3194 } else if ( forid > maxid ) {
3195 maxid=forid;
3196 }
3197 }
3198 if ( transcount < tp + maxid - minid + 1) {
3199 transcount *= 2;
3200 trie->trans = (reg_trie_trans *)
3201 PerlMemShared_realloc( trie->trans,
3202 transcount
3203 * sizeof(reg_trie_trans) );
3204 Zero( trie->trans + (transcount / 2),
3205 transcount / 2,
3206 reg_trie_trans );
3207 }
3208 base = trie->uniquecharcount + tp - minid;
3209 if ( maxid == minid ) {
3210 U32 set = 0;
3211 for ( ; zp < tp ; zp++ ) {
3212 if ( ! trie->trans[ zp ].next ) {
3213 base = trie->uniquecharcount + zp - minid;
3214 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3215 1).newstate;
3216 trie->trans[ zp ].check = state;
3217 set = 1;
3218 break;
3219 }
3220 }
3221 if ( !set ) {
3222 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3223 1).newstate;
3224 trie->trans[ tp ].check = state;
3225 tp++;
3226 zp = tp;
3227 }
3228 } else {
3229 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3230 const U32 tid = base
3231 - trie->uniquecharcount
3232 + TRIE_LIST_ITEM( state, idx ).forid;
3233 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3234 idx ).newstate;
3235 trie->trans[ tid ].check = state;
3236 }
3237 tp += ( maxid - minid + 1 );
3238 }
3239 Safefree(trie->states[ state ].trans.list);
3240 }
3241 /*
3242 DEBUG_TRIE_COMPILE_MORE_r(
3243 Perl_re_printf( aTHX_ " base: %d\n",base);
3244 );
3245 */
3246 trie->states[ state ].trans.base=base;
3247 }
3248 trie->lasttrans = tp + 1;
3249 }
3250 } else {
3251 /*
3252 Second Pass -- Flat Table Representation.
3253
3254 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3255 each. We know that we will need Charcount+1 trans at most to store
3256 the data (one row per char at worst case) So we preallocate both
3257 structures assuming worst case.
3258
3259 We then construct the trie using only the .next slots of the entry
3260 structs.
3261
3262 We use the .check field of the first entry of the node temporarily
3263 to make compression both faster and easier by keeping track of how
3264 many non zero fields are in the node.
3265
3266 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3267 transition.
3268
3269 There are two terms at use here: state as a TRIE_NODEIDX() which is
3270 a number representing the first entry of the node, and state as a
3271 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3272 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3273 if there are 2 entrys per node. eg:
3274
3275 A B A B
3276 1. 2 4 1. 3 7
3277 2. 0 3 3. 0 5
3278 3. 0 0 5. 0 0
3279 4. 0 0 7. 0 0
3280
3281 The table is internally in the right hand, idx form. However as we
3282 also have to deal with the states array which is indexed by nodenum
3283 we have to use TRIE_NODENUM() to convert.
3284
3285 */
3286 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3287 depth+1));
3288
3289 trie->trans = (reg_trie_trans *)
3290 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3291 * trie->uniquecharcount + 1,
3292 sizeof(reg_trie_trans) );
3293 trie->states = (reg_trie_state *)
3294 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3295 sizeof(reg_trie_state) );
3296 next_alloc = trie->uniquecharcount + 1;
3297
3298
3299 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3300
3301 regnode *noper = NEXTOPER( cur );
3302
3303 U32 state = 1; /* required init */
3304
3305 U16 charid = 0; /* sanity init */
3306 U32 accept_state = 0; /* sanity init */
3307
3308 U32 wordlen = 0; /* required init */
3309
3310 if (OP(noper) == NOTHING) {
3311 regnode *noper_next= regnext(noper);
3312 if (noper_next < tail)
3313 noper= noper_next;
3314 /* we will undo this assignment if noper does not
3315 * point at a trieable type in the else clause of
3316 * the following statement. */
3317 }
3318
3319 if ( noper < tail
3320 && ( OP(noper) == flags
3321 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3322 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
3323 || OP(noper) == EXACTFUP))))
3324 {
3325 const U8 *uc= (U8*)STRING(noper);
3326 const U8 *e= uc + STR_LEN(noper);
3327
3328 for ( ; uc < e ; uc += len ) {
3329
3330 TRIE_READ_CHAR;
3331
3332 if ( uvc < 256 ) {
3333 charid = trie->charmap[ uvc ];
3334 } else {
3335 SV* const * const svpp = hv_fetch( widecharmap,
3336 (char*)&uvc,
3337 sizeof( UV ),
3338 0);
3339 charid = svpp ? (U16)SvIV(*svpp) : 0;
3340 }
3341 if ( charid ) {
3342 charid--;
3343 if ( !trie->trans[ state + charid ].next ) {
3344 trie->trans[ state + charid ].next = next_alloc;
3345 trie->trans[ state ].check++;
3346 prev_states[TRIE_NODENUM(next_alloc)]
3347 = TRIE_NODENUM(state);
3348 next_alloc += trie->uniquecharcount;
3349 }
3350 state = trie->trans[ state + charid ].next;
3351 } else {
3352 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3353 }
3354 /* charid is now 0 if we dont know the char read, or
3355 * nonzero if we do */
3356 }
3357 } else {
3358 /* If we end up here it is because we skipped past a NOTHING, but did not end up
3359 * on a trieable type. So we need to reset noper back to point at the first regop
3360 * in the branch before we call TRIE_HANDLE_WORD().
3361 */
3362 noper= NEXTOPER(cur);
3363 }
3364 accept_state = TRIE_NODENUM( state );
3365 TRIE_HANDLE_WORD(accept_state);
3366
3367 } /* end second pass */
3368
3369 /* and now dump it out before we compress it */
3370 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3371 revcharmap,
3372 next_alloc, depth+1));
3373
3374 {
3375 /*
3376 * Inplace compress the table.*
3377
3378 For sparse data sets the table constructed by the trie algorithm will
3379 be mostly 0/FAIL transitions or to put it another way mostly empty.
3380 (Note that leaf nodes will not contain any transitions.)
3381
3382 This algorithm compresses the tables by eliminating most such
3383 transitions, at the cost of a modest bit of extra work during lookup:
3384
3385 - Each states[] entry contains a .base field which indicates the
3386 index in the state[] array wheres its transition data is stored.
3387
3388 - If .base is 0 there are no valid transitions from that node.
3389
3390 - If .base is nonzero then charid is added to it to find an entry in
3391 the trans array.
3392
3393 -If trans[states[state].base+charid].check!=state then the
3394 transition is taken to be a 0/Fail transition. Thus if there are fail
3395 transitions at the front of the node then the .base offset will point
3396 somewhere inside the previous nodes data (or maybe even into a node
3397 even earlier), but the .check field determines if the transition is
3398 valid.
3399
3400 XXX - wrong maybe?
3401 The following process inplace converts the table to the compressed
3402 table: We first do not compress the root node 1,and mark all its
3403 .check pointers as 1 and set its .base pointer as 1 as well. This
3404 allows us to do a DFA construction from the compressed table later,
3405 and ensures that any .base pointers we calculate later are greater
3406 than 0.
3407
3408 - We set 'pos' to indicate the first entry of the second node.
3409
3410 - We then iterate over the columns of the node, finding the first and
3411 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3412 and set the .check pointers accordingly, and advance pos
3413 appropriately and repreat for the next node. Note that when we copy
3414 the next pointers we have to convert them from the original
3415 NODEIDX form to NODENUM form as the former is not valid post
3416 compression.
3417
3418 - If a node has no transitions used we mark its base as 0 and do not
3419 advance the pos pointer.
3420
3421 - If a node only has one transition we use a second pointer into the
3422 structure to fill in allocated fail transitions from other states.
3423 This pointer is independent of the main pointer and scans forward
3424 looking for null transitions that are allocated to a state. When it
3425 finds one it writes the single transition into the "hole". If the
3426 pointer doesnt find one the single transition is appended as normal.
3427
3428 - Once compressed we can Renew/realloc the structures to release the
3429 excess space.
3430
3431 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3432 specifically Fig 3.47 and the associated pseudocode.
3433
3434 demq
3435 */
3436 const U32 laststate = TRIE_NODENUM( next_alloc );
3437 U32 state, charid;
3438 U32 pos = 0, zp=0;
3439 trie->statecount = laststate;
3440
3441 for ( state = 1 ; state < laststate ; state++ ) {
3442 U8 flag = 0;
3443 const U32 stateidx = TRIE_NODEIDX( state );
3444 const U32 o_used = trie->trans[ stateidx ].check;
3445 U32 used = trie->trans[ stateidx ].check;
3446 trie->trans[ stateidx ].check = 0;
3447
3448 for ( charid = 0;
3449 used && charid < trie->uniquecharcount;
3450 charid++ )
3451 {
3452 if ( flag || trie->trans[ stateidx + charid ].next ) {
3453 if ( trie->trans[ stateidx + charid ].next ) {
3454 if (o_used == 1) {
3455 for ( ; zp < pos ; zp++ ) {
3456 if ( ! trie->trans[ zp ].next ) {
3457 break;
3458 }
3459 }
3460 trie->states[ state ].trans.base
3461 = zp
3462 + trie->uniquecharcount
3463 - charid ;
3464 trie->trans[ zp ].next
3465 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3466 + charid ].next );
3467 trie->trans[ zp ].check = state;
3468 if ( ++zp > pos ) pos = zp;
3469 break;
3470 }
3471 used--;
3472 }
3473 if ( !flag ) {
3474 flag = 1;
3475 trie->states[ state ].trans.base
3476 = pos + trie->uniquecharcount - charid ;
3477 }
3478 trie->trans[ pos ].next
3479 = SAFE_TRIE_NODENUM(
3480 trie->trans[ stateidx + charid ].next );
3481 trie->trans[ pos ].check = state;
3482 pos++;
3483 }
3484 }
3485 }
3486 trie->lasttrans = pos + 1;
3487 trie->states = (reg_trie_state *)
3488 PerlMemShared_realloc( trie->states, laststate
3489 * sizeof(reg_trie_state) );
3490 DEBUG_TRIE_COMPILE_MORE_r(
3491 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3492 depth+1,
3493 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3494 + 1 ),
3495 (IV)next_alloc,
3496 (IV)pos,
3497 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3498 );
3499
3500 } /* end table compress */
3501 }
3502 DEBUG_TRIE_COMPILE_MORE_r(
3503 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3504 depth+1,
3505 (UV)trie->statecount,
3506 (UV)trie->lasttrans)
3507 );
3508 /* resize the trans array to remove unused space */
3509 trie->trans = (reg_trie_trans *)
3510 PerlMemShared_realloc( trie->trans, trie->lasttrans
3511 * sizeof(reg_trie_trans) );
3512
3513 { /* Modify the program and insert the new TRIE node */
3514 U8 nodetype =(U8) flags;
3515 char *str=NULL;
3516
3517 #ifdef DEBUGGING
3518 regnode *optimize = NULL;
3519 #ifdef RE_TRACK_PATTERN_OFFSETS
3520
3521 U32 mjd_offset = 0;
3522 U32 mjd_nodelen = 0;
3523 #endif /* RE_TRACK_PATTERN_OFFSETS */
3524 #endif /* DEBUGGING */
3525 /*
3526 This means we convert either the first branch or the first Exact,
3527 depending on whether the thing following (in 'last') is a branch
3528 or not and whther first is the startbranch (ie is it a sub part of
3529 the alternation or is it the whole thing.)
3530 Assuming its a sub part we convert the EXACT otherwise we convert
3531 the whole branch sequence, including the first.
3532 */
3533 /* Find the node we are going to overwrite */
3534 if ( first != startbranch || OP( last ) == BRANCH ) {
3535 /* branch sub-chain */
3536 NEXT_OFF( first ) = (U16)(last - first);
3537 #ifdef RE_TRACK_PATTERN_OFFSETS
3538 DEBUG_r({
3539 mjd_offset= Node_Offset((convert));
3540 mjd_nodelen= Node_Length((convert));
3541 });
3542 #endif
3543 /* whole branch chain */
3544 }
3545 #ifdef RE_TRACK_PATTERN_OFFSETS
3546 else {
3547 DEBUG_r({
3548 const regnode *nop = NEXTOPER( convert );
3549 mjd_offset= Node_Offset((nop));
3550 mjd_nodelen= Node_Length((nop));
3551 });
3552 }
3553 DEBUG_OPTIMISE_r(
3554 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3555 depth+1,
3556 (UV)mjd_offset, (UV)mjd_nodelen)
3557 );
3558 #endif
3559 /* But first we check to see if there is a common prefix we can
3560 split out as an EXACT and put in front of the TRIE node. */
3561 trie->startstate= 1;
3562 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3563 /* we want to find the first state that has more than
3564 * one transition, if that state is not the first state
3565 * then we have a common prefix which we can remove.
3566 */
3567 U32 state;
3568 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3569 U32 ofs = 0;
3570 I32 first_ofs = -1; /* keeps track of the ofs of the first
3571 transition, -1 means none */
3572 U32 count = 0;
3573 const U32 base = trie->states[ state ].trans.base;
3574
3575 /* does this state terminate an alternation? */
3576 if ( trie->states[state].wordnum )
3577 count = 1;
3578
3579 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3580 if ( ( base + ofs >= trie->uniquecharcount ) &&
3581 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3582 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3583 {
3584 if ( ++count > 1 ) {
3585 /* we have more than one transition */
3586 SV **tmp;
3587 U8 *ch;
3588 /* if this is the first state there is no common prefix
3589 * to extract, so we can exit */
3590 if ( state == 1 ) break;
3591 tmp = av_fetch( revcharmap, ofs, 0);
3592 ch = (U8*)SvPV_nolen_const( *tmp );
3593
3594 /* if we are on count 2 then we need to initialize the
3595 * bitmap, and store the previous char if there was one
3596 * in it*/
3597 if ( count == 2 ) {
3598 /* clear the bitmap */
3599 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3600 DEBUG_OPTIMISE_r(
3601 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3602 depth+1,
3603 (UV)state));
3604 if (first_ofs >= 0) {
3605 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3606 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3607
3608 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3609 DEBUG_OPTIMISE_r(
3610 Perl_re_printf( aTHX_ "%s", (char*)ch)
3611 );
3612 }
3613 }
3614 /* store the current firstchar in the bitmap */
3615 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3616 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3617 }
3618 first_ofs = ofs;
3619 }
3620 }
3621 if ( count == 1 ) {
3622 /* This state has only one transition, its transition is part
3623 * of a common prefix - we need to concatenate the char it
3624 * represents to what we have so far. */
3625 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3626 STRLEN len;
3627 char *ch = SvPV( *tmp, len );
3628 DEBUG_OPTIMISE_r({
3629 SV *sv=sv_newmortal();
3630 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3631 depth+1,
3632 (UV)state, (UV)first_ofs,
3633 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3634 PL_colors[0], PL_colors[1],
3635 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3636 PERL_PV_ESCAPE_FIRSTCHAR
3637 )
3638 );
3639 });
3640 if ( state==1 ) {
3641 OP( convert ) = nodetype;
3642 str=STRING(convert);
3643 setSTR_LEN(convert, 0);
3644 }
3645 assert( ( STR_LEN(convert) + len ) < 256 );
3646 setSTR_LEN(convert, (U8)(STR_LEN(convert) + len));
3647 while (len--)
3648 *str++ = *ch++;
3649 } else {
3650 #ifdef DEBUGGING
3651 if (state>1)
3652 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3653 #endif
3654 break;
3655 }
3656 }
3657 trie->prefixlen = (state-1);
3658 if (str) {
3659 regnode *n = convert+NODE_SZ_STR(convert);
3660 assert( NODE_SZ_STR(convert) <= U16_MAX );
3661 NEXT_OFF(convert) = (U16)(NODE_SZ_STR(convert));
3662 trie->startstate = state;
3663 trie->minlen -= (state - 1);
3664 trie->maxlen -= (state - 1);
3665 #ifdef DEBUGGING
3666 /* At least the UNICOS C compiler choked on this
3667 * being argument to DEBUG_r(), so let's just have
3668 * it right here. */
3669 if (
3670 #ifdef PERL_EXT_RE_BUILD
3671 1
3672 #else
3673 DEBUG_r_TEST
3674 #endif
3675 ) {
3676 regnode *fix = convert;
3677 U32 word = trie->wordcount;
3678 #ifdef RE_TRACK_PATTERN_OFFSETS
3679 mjd_nodelen++;
3680 #endif
3681 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3682 while( ++fix < n ) {
3683 Set_Node_Offset_Length(fix, 0, 0);
3684 }
3685 while (word--) {
3686 SV ** const tmp = av_fetch( trie_words, word, 0 );
3687 if (tmp) {
3688 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3689 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3690 else
3691 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3692 }
3693 }
3694 }
3695 #endif
3696 if (trie->maxlen) {
3697 convert = n;
3698 } else {
3699 NEXT_OFF(convert) = (U16)(tail - convert);
3700 DEBUG_r(optimize= n);
3701 }
3702 }
3703 }
3704 if (!jumper)
3705 jumper = last;
3706 if ( trie->maxlen ) {
3707 NEXT_OFF( convert ) = (U16)(tail - convert);
3708 ARG_SET( convert, data_slot );
3709 /* Store the offset to the first unabsorbed branch in
3710 jump[0], which is otherwise unused by the jump logic.
3711 We use this when dumping a trie and during optimisation. */
3712 if (trie->jump)
3713 trie->jump[0] = (U16)(nextbranch - convert);
3714
3715 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3716 * and there is a bitmap
3717 * and the first "jump target" node we found leaves enough room
3718 * then convert the TRIE node into a TRIEC node, with the bitmap
3719 * embedded inline in the opcode - this is hypothetically faster.
3720 */
3721 if ( !trie->states[trie->startstate].wordnum
3722 && trie->bitmap
3723 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3724 {
3725 OP( convert ) = TRIEC;
3726 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3727 PerlMemShared_free(trie->bitmap);
3728 trie->bitmap= NULL;
3729 } else
3730 OP( convert ) = TRIE;
3731
3732 /* store the type in the flags */
3733 convert->flags = nodetype;
3734 DEBUG_r({
3735 optimize = convert
3736 + NODE_STEP_REGNODE
3737 + regarglen[ OP( convert ) ];
3738 });
3739 /* XXX We really should free up the resource in trie now,
3740 as we won't use them - (which resources?) dmq */
3741 }
3742 /* needed for dumping*/
3743 DEBUG_r(if (optimize) {
3744 regnode *opt = convert;
3745
3746 while ( ++opt < optimize) {
3747 Set_Node_Offset_Length(opt, 0, 0);
3748 }
3749 /*
3750 Try to clean up some of the debris left after the
3751 optimisation.
3752 */
3753 while( optimize < jumper ) {
3754 Track_Code( mjd_nodelen += Node_Length((optimize)); );
3755 OP( optimize ) = OPTIMIZED;
3756 Set_Node_Offset_Length(optimize, 0, 0);
3757 optimize++;
3758 }
3759 Set_Node_Offset_Length(convert, mjd_offset, mjd_nodelen);
3760 });
3761 } /* end node insert */
3762
3763 /* Finish populating the prev field of the wordinfo array. Walk back
3764 * from each accept state until we find another accept state, and if
3765 * so, point the first word's .prev field at the second word. If the
3766 * second already has a .prev field set, stop now. This will be the
3767 * case either if we've already processed that word's accept state,
3768 * or that state had multiple words, and the overspill words were
3769 * already linked up earlier.
3770 */
3771 {
3772 U16 word;
3773 U32 state;
3774 U16 prev;
3775
3776 for (word=1; word <= trie->wordcount; word++) {
3777 prev = 0;
3778 if (trie->wordinfo[word].prev)
3779 continue;
3780 state = trie->wordinfo[word].accept;
3781 while (state) {
3782 state = prev_states[state];
3783 if (!state)
3784 break;
3785 prev = trie->states[state].wordnum;
3786 if (prev)
3787 break;
3788 }
3789 trie->wordinfo[word].prev = prev;
3790 }
3791 Safefree(prev_states);
3792 }
3793
3794
3795 /* and now dump out the compressed format */
3796 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3797
3798 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3799 #ifdef DEBUGGING
3800 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3801 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3802 #else
3803 SvREFCNT_dec_NN(revcharmap);
3804 #endif
3805 return trie->jump
3806 ? MADE_JUMP_TRIE
3807 : trie->startstate>1
3808 ? MADE_EXACT_TRIE
3809 : MADE_TRIE;
3810 }
3811
3812 STATIC regnode *
S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t * pRExC_state,regnode * source,U32 depth)3813 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3814 {
3815 /* The Trie is constructed and compressed now so we can build a fail array if
3816 * it's needed
3817
3818 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3819 3.32 in the
3820 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3821 Ullman 1985/88
3822 ISBN 0-201-10088-6
3823
3824 We find the fail state for each state in the trie, this state is the longest
3825 proper suffix of the current state's 'word' that is also a proper prefix of
3826 another word in our trie. State 1 represents the word '' and is thus the
3827 default fail state. This allows the DFA not to have to restart after its
3828 tried and failed a word at a given point, it simply continues as though it
3829 had been matching the other word in the first place.
3830 Consider
3831 'abcdgu'=~/abcdefg|cdgu/
3832 When we get to 'd' we are still matching the first word, we would encounter
3833 'g' which would fail, which would bring us to the state representing 'd' in
3834 the second word where we would try 'g' and succeed, proceeding to match
3835 'cdgu'.
3836 */
3837 /* add a fail transition */
3838 const U32 trie_offset = ARG(source);
3839 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3840 U32 *q;
3841 const U32 ucharcount = trie->uniquecharcount;
3842 const U32 numstates = trie->statecount;
3843 const U32 ubound = trie->lasttrans + ucharcount;
3844 U32 q_read = 0;
3845 U32 q_write = 0;
3846 U32 charid;
3847 U32 base = trie->states[ 1 ].trans.base;
3848 U32 *fail;
3849 reg_ac_data *aho;
3850 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3851 regnode *stclass;
3852 DECLARE_AND_GET_RE_DEBUG_FLAGS;
3853
3854 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3855 PERL_UNUSED_CONTEXT;
3856 #ifndef DEBUGGING
3857 PERL_UNUSED_ARG(depth);
3858 #endif
3859
3860 if ( OP(source) == TRIE ) {
3861 struct regnode_1 *op = (struct regnode_1 *)
3862 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3863 StructCopy(source, op, struct regnode_1);
3864 stclass = (regnode *)op;
3865 } else {
3866 struct regnode_charclass *op = (struct regnode_charclass *)
3867 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3868 StructCopy(source, op, struct regnode_charclass);
3869 stclass = (regnode *)op;
3870 }
3871 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3872
3873 ARG_SET( stclass, data_slot );
3874 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3875 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3876 aho->trie=trie_offset;
3877 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3878 Copy( trie->states, aho->states, numstates, reg_trie_state );
3879 Newx( q, numstates, U32);
3880 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3881 aho->refcount = 1;
3882 fail = aho->fail;
3883 /* initialize fail[0..1] to be 1 so that we always have
3884 a valid final fail state */
3885 fail[ 0 ] = fail[ 1 ] = 1;
3886
3887 for ( charid = 0; charid < ucharcount ; charid++ ) {
3888 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3889 if ( newstate ) {
3890 q[ q_write ] = newstate;
3891 /* set to point at the root */
3892 fail[ q[ q_write++ ] ]=1;
3893 }
3894 }
3895 while ( q_read < q_write) {
3896 const U32 cur = q[ q_read++ % numstates ];
3897 base = trie->states[ cur ].trans.base;
3898
3899 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3900 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3901 if (ch_state) {
3902 U32 fail_state = cur;
3903 U32 fail_base;
3904 do {
3905 fail_state = fail[ fail_state ];
3906 fail_base = aho->states[ fail_state ].trans.base;
3907 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3908
3909 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3910 fail[ ch_state ] = fail_state;
3911 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3912 {
3913 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3914 }
3915 q[ q_write++ % numstates] = ch_state;
3916 }
3917 }
3918 }
3919 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3920 when we fail in state 1, this allows us to use the
3921 charclass scan to find a valid start char. This is based on the principle
3922 that theres a good chance the string being searched contains lots of stuff
3923 that cant be a start char.
3924 */
3925 fail[ 0 ] = fail[ 1 ] = 0;
3926 DEBUG_TRIE_COMPILE_r({
3927 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3928 depth, (UV)numstates
3929 );
3930 for( q_read=1; q_read<numstates; q_read++ ) {
3931 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3932 }
3933 Perl_re_printf( aTHX_ "\n");
3934 });
3935 Safefree(q);
3936 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3937 return stclass;
3938 }
3939
3940
3941 /* The below joins as many adjacent EXACTish nodes as possible into a single
3942 * one. The regop may be changed if the node(s) contain certain sequences that
3943 * require special handling. The joining is only done if:
3944 * 1) there is room in the current conglomerated node to entirely contain the
3945 * next one.
3946 * 2) they are compatible node types
3947 *
3948 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3949 * these get optimized out
3950 *
3951 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3952 * as possible, even if that means splitting an existing node so that its first
3953 * part is moved to the preceding node. This would maximise the efficiency of
3954 * memEQ during matching.
3955 *
3956 * If a node is to match under /i (folded), the number of characters it matches
3957 * can be different than its character length if it contains a multi-character
3958 * fold. *min_subtract is set to the total delta number of characters of the
3959 * input nodes.
3960 *
3961 * And *unfolded_multi_char is set to indicate whether or not the node contains
3962 * an unfolded multi-char fold. This happens when it won't be known until
3963 * runtime whether the fold is valid or not; namely
3964 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3965 * target string being matched against turns out to be UTF-8 is that fold
3966 * valid; or
3967 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3968 * runtime.
3969 * (Multi-char folds whose components are all above the Latin1 range are not
3970 * run-time locale dependent, and have already been folded by the time this
3971 * function is called.)
3972 *
3973 * This is as good a place as any to discuss the design of handling these
3974 * multi-character fold sequences. It's been wrong in Perl for a very long
3975 * time. There are three code points in Unicode whose multi-character folds
3976 * were long ago discovered to mess things up. The previous designs for
3977 * dealing with these involved assigning a special node for them. This
3978 * approach doesn't always work, as evidenced by this example:
3979 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3980 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3981 * would match just the \xDF, it won't be able to handle the case where a
3982 * successful match would have to cross the node's boundary. The new approach
3983 * that hopefully generally solves the problem generates an EXACTFUP node
3984 * that is "sss" in this case.
3985 *
3986 * It turns out that there are problems with all multi-character folds, and not
3987 * just these three. Now the code is general, for all such cases. The
3988 * approach taken is:
3989 * 1) This routine examines each EXACTFish node that could contain multi-
3990 * character folded sequences. Since a single character can fold into
3991 * such a sequence, the minimum match length for this node is less than
3992 * the number of characters in the node. This routine returns in
3993 * *min_subtract how many characters to subtract from the actual
3994 * length of the string to get a real minimum match length; it is 0 if
3995 * there are no multi-char foldeds. This delta is used by the caller to
3996 * adjust the min length of the match, and the delta between min and max,
3997 * so that the optimizer doesn't reject these possibilities based on size
3998 * constraints.
3999 *
4000 * 2) For the sequence involving the LATIN SMALL LETTER SHARP S (U+00DF)
4001 * under /u, we fold it to 'ss' in regatom(), and in this routine, after
4002 * joining, we scan for occurrences of the sequence 'ss' in non-UTF-8
4003 * EXACTFU nodes. The node type of such nodes is then changed to
4004 * EXACTFUP, indicating it is problematic, and needs careful handling.
4005 * (The procedures in step 1) above are sufficient to handle this case in
4006 * UTF-8 encoded nodes.) The reason this is problematic is that this is
4007 * the only case where there is a possible fold length change in non-UTF-8
4008 * patterns. By reserving a special node type for problematic cases, the
4009 * far more common regular EXACTFU nodes can be processed faster.
4010 * regexec.c takes advantage of this.
4011 *
4012 * EXACTFUP has been created as a grab-bag for (hopefully uncommon)
4013 * problematic cases. These all only occur when the pattern is not
4014 * UTF-8. In addition to the 'ss' sequence where there is a possible fold
4015 * length change, it handles the situation where the string cannot be
4016 * entirely folded. The strings in an EXACTFish node are folded as much
4017 * as possible during compilation in regcomp.c. This saves effort in
4018 * regex matching. By using an EXACTFUP node when it is not possible to
4019 * fully fold at compile time, regexec.c can know that everything in an
4020 * EXACTFU node is folded, so folding can be skipped at runtime. The only
4021 * case where folding in EXACTFU nodes can't be done at compile time is
4022 * the presumably uncommon MICRO SIGN, when the pattern isn't UTF-8. This
4023 * is because its fold requires UTF-8 to represent. Thus EXACTFUP nodes
4024 * handle two very different cases. Alternatively, there could have been
4025 * a node type where there are length changes, one for unfolded, and one
4026 * for both. If yet another special case needed to be created, the number
4027 * of required node types would have to go to 7. khw figures that even
4028 * though there are plenty of node types to spare, that the maintenance
4029 * cost wasn't worth the small speedup of doing it that way, especially
4030 * since he thinks the MICRO SIGN is rarely encountered in practice.
4031 *
4032 * There are other cases where folding isn't done at compile time, but
4033 * none of them are under /u, and hence not for EXACTFU nodes. The folds
4034 * in EXACTFL nodes aren't known until runtime, and vary as the locale
4035 * changes. Some folds in EXACTF depend on if the runtime target string
4036 * is UTF-8 or not. (regatom() will create an EXACTFU node even under /di
4037 * when no fold in it depends on the UTF-8ness of the target string.)
4038 *
4039 * 3) A problem remains for unfolded multi-char folds. (These occur when the
4040 * validity of the fold won't be known until runtime, and so must remain
4041 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
4042 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
4043 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
4044 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
4045 * The reason this is a problem is that the optimizer part of regexec.c
4046 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
4047 * that a character in the pattern corresponds to at most a single
4048 * character in the target string. (And I do mean character, and not byte
4049 * here, unlike other parts of the documentation that have never been
4050 * updated to account for multibyte Unicode.) Sharp s in EXACTF and
4051 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
4052 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
4053 * EXACTFL nodes, violate the assumption, and they are the only instances
4054 * where it is violated. I'm reluctant to try to change the assumption,
4055 * as the code involved is impenetrable to me (khw), so instead the code
4056 * here punts. This routine examines EXACTFL nodes, and (when the pattern
4057 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
4058 * boolean indicating whether or not the node contains such a fold. When
4059 * it is true, the caller sets a flag that later causes the optimizer in
4060 * this file to not set values for the floating and fixed string lengths,
4061 * and thus avoids the optimizer code in regexec.c that makes the invalid
4062 * assumption. Thus, there is no optimization based on string lengths for
4063 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
4064 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
4065 * assumption is wrong only in these cases is that all other non-UTF-8
4066 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
4067 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
4068 * EXACTF nodes because we don't know at compile time if it actually
4069 * matches 'ss' or not. For EXACTF nodes it will match iff the target
4070 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
4071 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
4072 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
4073 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
4074 * string would require the pattern to be forced into UTF-8, the overhead
4075 * of which we want to avoid. Similarly the unfolded multi-char folds in
4076 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
4077 * locale.)
4078 *
4079 * Similarly, the code that generates tries doesn't currently handle
4080 * not-already-folded multi-char folds, and it looks like a pain to change
4081 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
4082 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
4083 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
4084 * using /iaa matching will be doing so almost entirely with ASCII
4085 * strings, so this should rarely be encountered in practice */
4086
4087 STATIC U32
S_join_exact(pTHX_ RExC_state_t * pRExC_state,regnode * scan,UV * min_subtract,bool * unfolded_multi_char,U32 flags,regnode * val,U32 depth)4088 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
4089 UV *min_subtract, bool *unfolded_multi_char,
4090 U32 flags, regnode *val, U32 depth)
4091 {
4092 /* Merge several consecutive EXACTish nodes into one. */
4093
4094 regnode *n = regnext(scan);
4095 U32 stringok = 1;
4096 regnode *next = scan + NODE_SZ_STR(scan);
4097 U32 merged = 0;
4098 U32 stopnow = 0;
4099 #ifdef DEBUGGING
4100 regnode *stop = scan;
4101 DECLARE_AND_GET_RE_DEBUG_FLAGS;
4102 #else
4103 PERL_UNUSED_ARG(depth);
4104 #endif
4105
4106 PERL_ARGS_ASSERT_JOIN_EXACT;
4107 #ifndef EXPERIMENTAL_INPLACESCAN
4108 PERL_UNUSED_ARG(flags);
4109 PERL_UNUSED_ARG(val);
4110 #endif
4111 DEBUG_PEEP("join", scan, depth, 0);
4112
4113 assert(PL_regkind[OP(scan)] == EXACT);
4114
4115 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
4116 * EXACT ones that are mergeable to the current one. */
4117 while ( n
4118 && ( PL_regkind[OP(n)] == NOTHING
4119 || (stringok && PL_regkind[OP(n)] == EXACT))
4120 && NEXT_OFF(n)
4121 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
4122 {
4123
4124 if (OP(n) == TAIL || n > next)
4125 stringok = 0;
4126 if (PL_regkind[OP(n)] == NOTHING) {
4127 DEBUG_PEEP("skip:", n, depth, 0);
4128 NEXT_OFF(scan) += NEXT_OFF(n);
4129 next = n + NODE_STEP_REGNODE;
4130 #ifdef DEBUGGING
4131 if (stringok)
4132 stop = n;
4133 #endif
4134 n = regnext(n);
4135 }
4136 else if (stringok) {
4137 const unsigned int oldl = STR_LEN(scan);
4138 regnode * const nnext = regnext(n);
4139
4140 /* XXX I (khw) kind of doubt that this works on platforms (should
4141 * Perl ever run on one) where U8_MAX is above 255 because of lots
4142 * of other assumptions */
4143 /* Don't join if the sum can't fit into a single node */
4144 if (oldl + STR_LEN(n) > U8_MAX)
4145 break;
4146
4147 /* Joining something that requires UTF-8 with something that
4148 * doesn't, means the result requires UTF-8. */
4149 if (OP(scan) == EXACT && (OP(n) == EXACT_REQ8)) {
4150 OP(scan) = EXACT_REQ8;
4151 }
4152 else if (OP(scan) == EXACT_REQ8 && (OP(n) == EXACT)) {
4153 ; /* join is compatible, no need to change OP */
4154 }
4155 else if ((OP(scan) == EXACTFU) && (OP(n) == EXACTFU_REQ8)) {
4156 OP(scan) = EXACTFU_REQ8;
4157 }
4158 else if ((OP(scan) == EXACTFU_REQ8) && (OP(n) == EXACTFU)) {
4159 ; /* join is compatible, no need to change OP */
4160 }
4161 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU) {
4162 ; /* join is compatible, no need to change OP */
4163 }
4164 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU_S_EDGE) {
4165
4166 /* Under /di, temporary EXACTFU_S_EDGE nodes are generated,
4167 * which can join with EXACTFU ones. We check for this case
4168 * here. These need to be resolved to either EXACTFU or
4169 * EXACTF at joining time. They have nothing in them that
4170 * would forbid them from being the more desirable EXACTFU
4171 * nodes except that they begin and/or end with a single [Ss].
4172 * The reason this is problematic is because they could be
4173 * joined in this loop with an adjacent node that ends and/or
4174 * begins with [Ss] which would then form the sequence 'ss',
4175 * which matches differently under /di than /ui, in which case
4176 * EXACTFU can't be used. If the 'ss' sequence doesn't get
4177 * formed, the nodes get absorbed into any adjacent EXACTFU
4178 * node. And if the only adjacent node is EXACTF, they get
4179 * absorbed into that, under the theory that a longer node is
4180 * better than two shorter ones, even if one is EXACTFU. Note
4181 * that EXACTFU_REQ8 is generated only for UTF-8 patterns,
4182 * and the EXACTFU_S_EDGE ones only for non-UTF-8. */
4183
4184 if (STRING(n)[STR_LEN(n)-1] == 's') {
4185
4186 /* Here the joined node would end with 's'. If the node
4187 * following the combination is an EXACTF one, it's better to
4188 * join this trailing edge 's' node with that one, leaving the
4189 * current one in 'scan' be the more desirable EXACTFU */
4190 if (OP(nnext) == EXACTF) {
4191 break;
4192 }
4193
4194 OP(scan) = EXACTFU_S_EDGE;
4195
4196 } /* Otherwise, the beginning 's' of the 2nd node just
4197 becomes an interior 's' in 'scan' */
4198 }
4199 else if (OP(scan) == EXACTF && OP(n) == EXACTF) {
4200 ; /* join is compatible, no need to change OP */
4201 }
4202 else if (OP(scan) == EXACTF && OP(n) == EXACTFU_S_EDGE) {
4203
4204 /* EXACTF nodes are compatible for joining with EXACTFU_S_EDGE
4205 * nodes. But the latter nodes can be also joined with EXACTFU
4206 * ones, and that is a better outcome, so if the node following
4207 * 'n' is EXACTFU, quit now so that those two can be joined
4208 * later */
4209 if (OP(nnext) == EXACTFU) {
4210 break;
4211 }
4212
4213 /* The join is compatible, and the combined node will be
4214 * EXACTF. (These don't care if they begin or end with 's' */
4215 }
4216 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU_S_EDGE) {
4217 if ( STRING(scan)[STR_LEN(scan)-1] == 's'
4218 && STRING(n)[0] == 's')
4219 {
4220 /* When combined, we have the sequence 'ss', which means we
4221 * have to remain /di */
4222 OP(scan) = EXACTF;
4223 }
4224 }
4225 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU) {
4226 if (STRING(n)[0] == 's') {
4227 ; /* Here the join is compatible and the combined node
4228 starts with 's', no need to change OP */
4229 }
4230 else { /* Now the trailing 's' is in the interior */
4231 OP(scan) = EXACTFU;
4232 }
4233 }
4234 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTF) {
4235
4236 /* The join is compatible, and the combined node will be
4237 * EXACTF. (These don't care if they begin or end with 's' */
4238 OP(scan) = EXACTF;
4239 }
4240 else if (OP(scan) != OP(n)) {
4241
4242 /* The only other compatible joinings are the same node type */
4243 break;
4244 }
4245
4246 DEBUG_PEEP("merg", n, depth, 0);
4247 merged++;
4248
4249 NEXT_OFF(scan) += NEXT_OFF(n);
4250 assert( ( STR_LEN(scan) + STR_LEN(n) ) < 256 );
4251 setSTR_LEN(scan, (U8)(STR_LEN(scan) + STR_LEN(n)));
4252 next = n + NODE_SZ_STR(n);
4253 /* Now we can overwrite *n : */
4254 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
4255 #ifdef DEBUGGING
4256 stop = next - 1;
4257 #endif
4258 n = nnext;
4259 if (stopnow) break;
4260 }
4261
4262 #ifdef EXPERIMENTAL_INPLACESCAN
4263 if (flags && !NEXT_OFF(n)) {
4264 DEBUG_PEEP("atch", val, depth, 0);
4265 if (reg_off_by_arg[OP(n)]) {
4266 ARG_SET(n, val - n);
4267 }
4268 else {
4269 NEXT_OFF(n) = val - n;
4270 }
4271 stopnow = 1;
4272 }
4273 #endif
4274 }
4275
4276 /* This temporary node can now be turned into EXACTFU, and must, as
4277 * regexec.c doesn't handle it */
4278 if (OP(scan) == EXACTFU_S_EDGE) {
4279 OP(scan) = EXACTFU;
4280 }
4281
4282 *min_subtract = 0;
4283 *unfolded_multi_char = FALSE;
4284
4285 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
4286 * can now analyze for sequences of problematic code points. (Prior to
4287 * this final joining, sequences could have been split over boundaries, and
4288 * hence missed). The sequences only happen in folding, hence for any
4289 * non-EXACT EXACTish node */
4290 if (OP(scan) != EXACT && OP(scan) != EXACT_REQ8 && OP(scan) != EXACTL) {
4291 U8* s0 = (U8*) STRING(scan);
4292 U8* s = s0;
4293 U8* s_end = s0 + STR_LEN(scan);
4294
4295 int total_count_delta = 0; /* Total delta number of characters that
4296 multi-char folds expand to */
4297
4298 /* One pass is made over the node's string looking for all the
4299 * possibilities. To avoid some tests in the loop, there are two main
4300 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
4301 * non-UTF-8 */
4302 if (UTF) {
4303 U8* folded = NULL;
4304
4305 if (OP(scan) == EXACTFL) {
4306 U8 *d;
4307
4308 /* An EXACTFL node would already have been changed to another
4309 * node type unless there is at least one character in it that
4310 * is problematic; likely a character whose fold definition
4311 * won't be known until runtime, and so has yet to be folded.
4312 * For all but the UTF-8 locale, folds are 1-1 in length, but
4313 * to handle the UTF-8 case, we need to create a temporary
4314 * folded copy using UTF-8 locale rules in order to analyze it.
4315 * This is because our macros that look to see if a sequence is
4316 * a multi-char fold assume everything is folded (otherwise the
4317 * tests in those macros would be too complicated and slow).
4318 * Note that here, the non-problematic folds will have already
4319 * been done, so we can just copy such characters. We actually
4320 * don't completely fold the EXACTFL string. We skip the
4321 * unfolded multi-char folds, as that would just create work
4322 * below to figure out the size they already are */
4323
4324 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
4325 d = folded;
4326 while (s < s_end) {
4327 STRLEN s_len = UTF8SKIP(s);
4328 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
4329 Copy(s, d, s_len, U8);
4330 d += s_len;
4331 }
4332 else if (is_FOLDS_TO_MULTI_utf8(s)) {
4333 *unfolded_multi_char = TRUE;
4334 Copy(s, d, s_len, U8);
4335 d += s_len;
4336 }
4337 else if (isASCII(*s)) {
4338 *(d++) = toFOLD(*s);
4339 }
4340 else {
4341 STRLEN len;
4342 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
4343 d += len;
4344 }
4345 s += s_len;
4346 }
4347
4348 /* Point the remainder of the routine to look at our temporary
4349 * folded copy */
4350 s = folded;
4351 s_end = d;
4352 } /* End of creating folded copy of EXACTFL string */
4353
4354 /* Examine the string for a multi-character fold sequence. UTF-8
4355 * patterns have all characters pre-folded by the time this code is
4356 * executed */
4357 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
4358 length sequence we are looking for is 2 */
4359 {
4360 int count = 0; /* How many characters in a multi-char fold */
4361 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
4362 if (! len) { /* Not a multi-char fold: get next char */
4363 s += UTF8SKIP(s);
4364 continue;
4365 }
4366
4367 { /* Here is a generic multi-char fold. */
4368 U8* multi_end = s + len;
4369
4370 /* Count how many characters are in it. In the case of
4371 * /aa, no folds which contain ASCII code points are
4372 * allowed, so check for those, and skip if found. */
4373 if (OP(scan) != EXACTFAA && OP(scan) != EXACTFAA_NO_TRIE) {
4374 count = utf8_length(s, multi_end);
4375 s = multi_end;
4376 }
4377 else {
4378 while (s < multi_end) {
4379 if (isASCII(*s)) {
4380 s++;
4381 goto next_iteration;
4382 }
4383 else {
4384 s += UTF8SKIP(s);
4385 }
4386 count++;
4387 }
4388 }
4389 }
4390
4391 /* The delta is how long the sequence is minus 1 (1 is how long
4392 * the character that folds to the sequence is) */
4393 total_count_delta += count - 1;
4394 next_iteration: ;
4395 }
4396
4397 /* We created a temporary folded copy of the string in EXACTFL
4398 * nodes. Therefore we need to be sure it doesn't go below zero,
4399 * as the real string could be shorter */
4400 if (OP(scan) == EXACTFL) {
4401 int total_chars = utf8_length((U8*) STRING(scan),
4402 (U8*) STRING(scan) + STR_LEN(scan));
4403 if (total_count_delta > total_chars) {
4404 total_count_delta = total_chars;
4405 }
4406 }
4407
4408 *min_subtract += total_count_delta;
4409 Safefree(folded);
4410 }
4411 else if (OP(scan) == EXACTFAA) {
4412
4413 /* Non-UTF-8 pattern, EXACTFAA node. There can't be a multi-char
4414 * fold to the ASCII range (and there are no existing ones in the
4415 * upper latin1 range). But, as outlined in the comments preceding
4416 * this function, we need to flag any occurrences of the sharp s.
4417 * This character forbids trie formation (because of added
4418 * complexity) */
4419 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
4420 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
4421 || UNICODE_DOT_DOT_VERSION > 0)
4422 while (s < s_end) {
4423 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
4424 OP(scan) = EXACTFAA_NO_TRIE;
4425 *unfolded_multi_char = TRUE;
4426 break;
4427 }
4428 s++;
4429 }
4430 }
4431 else if (OP(scan) != EXACTFAA_NO_TRIE) {
4432
4433 /* Non-UTF-8 pattern, not EXACTFAA node. Look for the multi-char
4434 * folds that are all Latin1. As explained in the comments
4435 * preceding this function, we look also for the sharp s in EXACTF
4436 * and EXACTFL nodes; it can be in the final position. Otherwise
4437 * we can stop looking 1 byte earlier because have to find at least
4438 * two characters for a multi-fold */
4439 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
4440 ? s_end
4441 : s_end -1;
4442
4443 while (s < upper) {
4444 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
4445 if (! len) { /* Not a multi-char fold. */
4446 if (*s == LATIN_SMALL_LETTER_SHARP_S
4447 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4448 {
4449 *unfolded_multi_char = TRUE;
4450 }
4451 s++;
4452 continue;
4453 }
4454
4455 if (len == 2
4456 && isALPHA_FOLD_EQ(*s, 's')
4457 && isALPHA_FOLD_EQ(*(s+1), 's'))
4458 {
4459
4460 /* EXACTF nodes need to know that the minimum length
4461 * changed so that a sharp s in the string can match this
4462 * ss in the pattern, but they remain EXACTF nodes, as they
4463 * won't match this unless the target string is in UTF-8,
4464 * which we don't know until runtime. EXACTFL nodes can't
4465 * transform into EXACTFU nodes */
4466 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4467 OP(scan) = EXACTFUP;
4468 }
4469 }
4470
4471 *min_subtract += len - 1;
4472 s += len;
4473 }
4474 #endif
4475 }
4476 }
4477
4478 #ifdef DEBUGGING
4479 /* Allow dumping but overwriting the collection of skipped
4480 * ops and/or strings with fake optimized ops */
4481 n = scan + NODE_SZ_STR(scan);
4482 while (n <= stop) {
4483 OP(n) = OPTIMIZED;
4484 FLAGS(n) = 0;
4485 NEXT_OFF(n) = 0;
4486 n++;
4487 }
4488 #endif
4489 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl", scan, depth, 0);});
4490 return stopnow;
4491 }
4492
4493 /* REx optimizer. Converts nodes into quicker variants "in place".
4494 Finds fixed substrings. */
4495
4496 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4497 to the position after last scanned or to NULL. */
4498
4499 #define INIT_AND_WITHP \
4500 assert(!and_withp); \
4501 Newx(and_withp, 1, regnode_ssc); \
4502 SAVEFREEPV(and_withp)
4503
4504
4505 static void
S_unwind_scan_frames(pTHX_ const void * p)4506 S_unwind_scan_frames(pTHX_ const void *p)
4507 {
4508 scan_frame *f= (scan_frame *)p;
4509 do {
4510 scan_frame *n= f->next_frame;
4511 Safefree(f);
4512 f= n;
4513 } while (f);
4514 }
4515
4516 /* Follow the next-chain of the current node and optimize away
4517 all the NOTHINGs from it.
4518 */
4519 STATIC void
S_rck_elide_nothing(pTHX_ regnode * node)4520 S_rck_elide_nothing(pTHX_ regnode *node)
4521 {
4522 PERL_ARGS_ASSERT_RCK_ELIDE_NOTHING;
4523
4524 if (OP(node) != CURLYX) {
4525 const int max = (reg_off_by_arg[OP(node)]
4526 ? I32_MAX
4527 /* I32 may be smaller than U16 on CRAYs! */
4528 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4529 int off = (reg_off_by_arg[OP(node)] ? ARG(node) : NEXT_OFF(node));
4530 int noff;
4531 regnode *n = node;
4532
4533 /* Skip NOTHING and LONGJMP. */
4534 while (
4535 (n = regnext(n))
4536 && (
4537 (PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4538 || ((OP(n) == LONGJMP) && (noff = ARG(n)))
4539 )
4540 && off + noff < max
4541 ) {
4542 off += noff;
4543 }
4544 if (reg_off_by_arg[OP(node)])
4545 ARG(node) = off;
4546 else
4547 NEXT_OFF(node) = off;
4548 }
4549 return;
4550 }
4551
4552 /* the return from this sub is the minimum length that could possibly match */
4553 STATIC SSize_t
S_study_chunk(pTHX_ RExC_state_t * pRExC_state,regnode ** scanp,SSize_t * minlenp,SSize_t * deltap,regnode * last,scan_data_t * data,I32 stopparen,U32 recursed_depth,regnode_ssc * and_withp,U32 flags,U32 depth,bool was_mutate_ok)4554 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4555 SSize_t *minlenp, SSize_t *deltap,
4556 regnode *last,
4557 scan_data_t *data,
4558 I32 stopparen,
4559 U32 recursed_depth,
4560 regnode_ssc *and_withp,
4561 U32 flags, U32 depth, bool was_mutate_ok)
4562 /* scanp: Start here (read-write). */
4563 /* deltap: Write maxlen-minlen here. */
4564 /* last: Stop before this one. */
4565 /* data: string data about the pattern */
4566 /* stopparen: treat close N as END */
4567 /* recursed: which subroutines have we recursed into */
4568 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4569 {
4570 SSize_t final_minlen;
4571 /* There must be at least this number of characters to match */
4572 SSize_t min = 0;
4573 I32 pars = 0, code;
4574 regnode *scan = *scanp, *next;
4575 SSize_t delta = 0;
4576 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4577 int is_inf_internal = 0; /* The studied chunk is infinite */
4578 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4579 scan_data_t data_fake;
4580 SV *re_trie_maxbuff = NULL;
4581 regnode *first_non_open = scan;
4582 SSize_t stopmin = OPTIMIZE_INFTY;
4583 scan_frame *frame = NULL;
4584 DECLARE_AND_GET_RE_DEBUG_FLAGS;
4585
4586 PERL_ARGS_ASSERT_STUDY_CHUNK;
4587 RExC_study_started= 1;
4588
4589 Zero(&data_fake, 1, scan_data_t);
4590
4591 if ( depth == 0 ) {
4592 while (first_non_open && OP(first_non_open) == OPEN)
4593 first_non_open=regnext(first_non_open);
4594 }
4595
4596
4597 fake_study_recurse:
4598 DEBUG_r(
4599 RExC_study_chunk_recursed_count++;
4600 );
4601 DEBUG_OPTIMISE_MORE_r(
4602 {
4603 Perl_re_indentf( aTHX_ "study_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
4604 depth, (long)stopparen,
4605 (unsigned long)RExC_study_chunk_recursed_count,
4606 (unsigned long)depth, (unsigned long)recursed_depth,
4607 scan,
4608 last);
4609 if (recursed_depth) {
4610 U32 i;
4611 U32 j;
4612 for ( j = 0 ; j < recursed_depth ; j++ ) {
4613 for ( i = 0 ; i < (U32)RExC_total_parens ; i++ ) {
4614 if (PAREN_TEST(j, i) && (!j || !PAREN_TEST(j - 1, i))) {
4615 Perl_re_printf( aTHX_ " %d",(int)i);
4616 break;
4617 }
4618 }
4619 if ( j + 1 < recursed_depth ) {
4620 Perl_re_printf( aTHX_ ",");
4621 }
4622 }
4623 }
4624 Perl_re_printf( aTHX_ "\n");
4625 }
4626 );
4627 while ( scan && OP(scan) != END && scan < last ){
4628 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
4629 node length to get a real minimum (because
4630 the folded version may be shorter) */
4631 bool unfolded_multi_char = FALSE;
4632 /* avoid mutating ops if we are anywhere within the recursed or
4633 * enframed handling for a GOSUB: the outermost level will handle it.
4634 */
4635 bool mutate_ok = was_mutate_ok && !(frame && frame->in_gosub);
4636 /* Peephole optimizer: */
4637 DEBUG_STUDYDATA("Peep", data, depth, is_inf);
4638 DEBUG_PEEP("Peep", scan, depth, flags);
4639
4640
4641 /* The reason we do this here is that we need to deal with things like
4642 * /(?:f)(?:o)(?:o)/ which cant be dealt with by the normal EXACT
4643 * parsing code, as each (?:..) is handled by a different invocation of
4644 * reg() -- Yves
4645 */
4646 if (PL_regkind[OP(scan)] == EXACT
4647 && OP(scan) != LEXACT
4648 && OP(scan) != LEXACT_REQ8
4649 && mutate_ok
4650 ) {
4651 join_exact(pRExC_state, scan, &min_subtract, &unfolded_multi_char,
4652 0, NULL, depth + 1);
4653 }
4654
4655 /* Follow the next-chain of the current node and optimize
4656 away all the NOTHINGs from it.
4657 */
4658 rck_elide_nothing(scan);
4659
4660 /* The principal pseudo-switch. Cannot be a switch, since we look into
4661 * several different things. */
4662 if ( OP(scan) == DEFINEP ) {
4663 SSize_t minlen = 0;
4664 SSize_t deltanext = 0;
4665 SSize_t fake_last_close = 0;
4666 I32 f = SCF_IN_DEFINE;
4667
4668 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4669 scan = regnext(scan);
4670 assert( OP(scan) == IFTHEN );
4671 DEBUG_PEEP("expect IFTHEN", scan, depth, flags);
4672
4673 data_fake.last_closep= &fake_last_close;
4674 minlen = *minlenp;
4675 next = regnext(scan);
4676 scan = NEXTOPER(NEXTOPER(scan));
4677 DEBUG_PEEP("scan", scan, depth, flags);
4678 DEBUG_PEEP("next", next, depth, flags);
4679
4680 /* we suppose the run is continuous, last=next...
4681 * NOTE we dont use the return here! */
4682 /* DEFINEP study_chunk() recursion */
4683 (void)study_chunk(pRExC_state, &scan, &minlen,
4684 &deltanext, next, &data_fake, stopparen,
4685 recursed_depth, NULL, f, depth+1, mutate_ok);
4686
4687 scan = next;
4688 } else
4689 if (
4690 OP(scan) == BRANCH ||
4691 OP(scan) == BRANCHJ ||
4692 OP(scan) == IFTHEN
4693 ) {
4694 next = regnext(scan);
4695 code = OP(scan);
4696
4697 /* The op(next)==code check below is to see if we
4698 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
4699 * IFTHEN is special as it might not appear in pairs.
4700 * Not sure whether BRANCH-BRANCHJ is possible, regardless
4701 * we dont handle it cleanly. */
4702 if (OP(next) == code || code == IFTHEN) {
4703 /* NOTE - There is similar code to this block below for
4704 * handling TRIE nodes on a re-study. If you change stuff here
4705 * check there too. */
4706 SSize_t max1 = 0, min1 = OPTIMIZE_INFTY, num = 0;
4707 regnode_ssc accum;
4708 regnode * const startbranch=scan;
4709
4710 if (flags & SCF_DO_SUBSTR) {
4711 /* Cannot merge strings after this. */
4712 scan_commit(pRExC_state, data, minlenp, is_inf);
4713 }
4714
4715 if (flags & SCF_DO_STCLASS)
4716 ssc_init_zero(pRExC_state, &accum);
4717
4718 while (OP(scan) == code) {
4719 SSize_t deltanext, minnext, fake;
4720 I32 f = 0;
4721 regnode_ssc this_class;
4722
4723 DEBUG_PEEP("Branch", scan, depth, flags);
4724
4725 num++;
4726 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4727 if (data) {
4728 data_fake.whilem_c = data->whilem_c;
4729 data_fake.last_closep = data->last_closep;
4730 }
4731 else
4732 data_fake.last_closep = &fake;
4733
4734 data_fake.pos_delta = delta;
4735 next = regnext(scan);
4736
4737 scan = NEXTOPER(scan); /* everything */
4738 if (code != BRANCH) /* everything but BRANCH */
4739 scan = NEXTOPER(scan);
4740
4741 if (flags & SCF_DO_STCLASS) {
4742 ssc_init(pRExC_state, &this_class);
4743 data_fake.start_class = &this_class;
4744 f = SCF_DO_STCLASS_AND;
4745 }
4746 if (flags & SCF_WHILEM_VISITED_POS)
4747 f |= SCF_WHILEM_VISITED_POS;
4748
4749 /* we suppose the run is continuous, last=next...*/
4750 /* recurse study_chunk() for each BRANCH in an alternation */
4751 minnext = study_chunk(pRExC_state, &scan, minlenp,
4752 &deltanext, next, &data_fake, stopparen,
4753 recursed_depth, NULL, f, depth+1,
4754 mutate_ok);
4755
4756 if (min1 > minnext)
4757 min1 = minnext;
4758 if (deltanext == OPTIMIZE_INFTY) {
4759 is_inf = is_inf_internal = 1;
4760 max1 = OPTIMIZE_INFTY;
4761 } else if (max1 < minnext + deltanext)
4762 max1 = minnext + deltanext;
4763 scan = next;
4764 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4765 pars++;
4766 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4767 if ( stopmin > minnext)
4768 stopmin = min + min1;
4769 flags &= ~SCF_DO_SUBSTR;
4770 if (data)
4771 data->flags |= SCF_SEEN_ACCEPT;
4772 }
4773 if (data) {
4774 if (data_fake.flags & SF_HAS_EVAL)
4775 data->flags |= SF_HAS_EVAL;
4776 data->whilem_c = data_fake.whilem_c;
4777 }
4778 if (flags & SCF_DO_STCLASS)
4779 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4780 }
4781 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4782 min1 = 0;
4783 if (flags & SCF_DO_SUBSTR) {
4784 data->pos_min += min1;
4785 if (data->pos_delta >= OPTIMIZE_INFTY - (max1 - min1))
4786 data->pos_delta = OPTIMIZE_INFTY;
4787 else
4788 data->pos_delta += max1 - min1;
4789 if (max1 != min1 || is_inf)
4790 data->cur_is_floating = 1;
4791 }
4792 min += min1;
4793 if (delta == OPTIMIZE_INFTY
4794 || OPTIMIZE_INFTY - delta - (max1 - min1) < 0)
4795 delta = OPTIMIZE_INFTY;
4796 else
4797 delta += max1 - min1;
4798 if (flags & SCF_DO_STCLASS_OR) {
4799 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4800 if (min1) {
4801 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4802 flags &= ~SCF_DO_STCLASS;
4803 }
4804 }
4805 else if (flags & SCF_DO_STCLASS_AND) {
4806 if (min1) {
4807 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4808 flags &= ~SCF_DO_STCLASS;
4809 }
4810 else {
4811 /* Switch to OR mode: cache the old value of
4812 * data->start_class */
4813 INIT_AND_WITHP;
4814 StructCopy(data->start_class, and_withp, regnode_ssc);
4815 flags &= ~SCF_DO_STCLASS_AND;
4816 StructCopy(&accum, data->start_class, regnode_ssc);
4817 flags |= SCF_DO_STCLASS_OR;
4818 }
4819 }
4820
4821 if (PERL_ENABLE_TRIE_OPTIMISATION
4822 && OP(startbranch) == BRANCH
4823 && mutate_ok
4824 ) {
4825 /* demq.
4826
4827 Assuming this was/is a branch we are dealing with: 'scan'
4828 now points at the item that follows the branch sequence,
4829 whatever it is. We now start at the beginning of the
4830 sequence and look for subsequences of
4831
4832 BRANCH->EXACT=>x1
4833 BRANCH->EXACT=>x2
4834 tail
4835
4836 which would be constructed from a pattern like
4837 /A|LIST|OF|WORDS/
4838
4839 If we can find such a subsequence we need to turn the first
4840 element into a trie and then add the subsequent branch exact
4841 strings to the trie.
4842
4843 We have two cases
4844
4845 1. patterns where the whole set of branches can be
4846 converted.
4847
4848 2. patterns where only a subset can be converted.
4849
4850 In case 1 we can replace the whole set with a single regop
4851 for the trie. In case 2 we need to keep the start and end
4852 branches so
4853
4854 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4855 becomes BRANCH TRIE; BRANCH X;
4856
4857 There is an additional case, that being where there is a
4858 common prefix, which gets split out into an EXACT like node
4859 preceding the TRIE node.
4860
4861 If x(1..n)==tail then we can do a simple trie, if not we make
4862 a "jump" trie, such that when we match the appropriate word
4863 we "jump" to the appropriate tail node. Essentially we turn
4864 a nested if into a case structure of sorts.
4865
4866 */
4867
4868 int made=0;
4869 if (!re_trie_maxbuff) {
4870 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4871 if (!SvIOK(re_trie_maxbuff))
4872 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4873 }
4874 if ( SvIV(re_trie_maxbuff)>=0 ) {
4875 regnode *cur;
4876 regnode *first = (regnode *)NULL;
4877 regnode *prev = (regnode *)NULL;
4878 regnode *tail = scan;
4879 U8 trietype = 0;
4880 U32 count=0;
4881
4882 /* var tail is used because there may be a TAIL
4883 regop in the way. Ie, the exacts will point to the
4884 thing following the TAIL, but the last branch will
4885 point at the TAIL. So we advance tail. If we
4886 have nested (?:) we may have to move through several
4887 tails.
4888 */
4889
4890 while ( OP( tail ) == TAIL ) {
4891 /* this is the TAIL generated by (?:) */
4892 tail = regnext( tail );
4893 }
4894
4895
4896 DEBUG_TRIE_COMPILE_r({
4897 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4898 Perl_re_indentf( aTHX_ "%s %" UVuf ":%s\n",
4899 depth+1,
4900 "Looking for TRIE'able sequences. Tail node is ",
4901 (UV) REGNODE_OFFSET(tail),
4902 SvPV_nolen_const( RExC_mysv )
4903 );
4904 });
4905
4906 /*
4907
4908 Step through the branches
4909 cur represents each branch,
4910 noper is the first thing to be matched as part
4911 of that branch
4912 noper_next is the regnext() of that node.
4913
4914 We normally handle a case like this
4915 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4916 support building with NOJUMPTRIE, which restricts
4917 the trie logic to structures like /FOO|BAR/.
4918
4919 If noper is a trieable nodetype then the branch is
4920 a possible optimization target. If we are building
4921 under NOJUMPTRIE then we require that noper_next is
4922 the same as scan (our current position in the regex
4923 program).
4924
4925 Once we have two or more consecutive such branches
4926 we can create a trie of the EXACT's contents and
4927 stitch it in place into the program.
4928
4929 If the sequence represents all of the branches in
4930 the alternation we replace the entire thing with a
4931 single TRIE node.
4932
4933 Otherwise when it is a subsequence we need to
4934 stitch it in place and replace only the relevant
4935 branches. This means the first branch has to remain
4936 as it is used by the alternation logic, and its
4937 next pointer, and needs to be repointed at the item
4938 on the branch chain following the last branch we
4939 have optimized away.
4940
4941 This could be either a BRANCH, in which case the
4942 subsequence is internal, or it could be the item
4943 following the branch sequence in which case the
4944 subsequence is at the end (which does not
4945 necessarily mean the first node is the start of the
4946 alternation).
4947
4948 TRIE_TYPE(X) is a define which maps the optype to a
4949 trietype.
4950
4951 optype | trietype
4952 ----------------+-----------
4953 NOTHING | NOTHING
4954 EXACT | EXACT
4955 EXACT_REQ8 | EXACT
4956 EXACTFU | EXACTFU
4957 EXACTFU_REQ8 | EXACTFU
4958 EXACTFUP | EXACTFU
4959 EXACTFAA | EXACTFAA
4960 EXACTL | EXACTL
4961 EXACTFLU8 | EXACTFLU8
4962
4963
4964 */
4965 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4966 ? NOTHING \
4967 : ( EXACT == (X) || EXACT_REQ8 == (X) ) \
4968 ? EXACT \
4969 : ( EXACTFU == (X) \
4970 || EXACTFU_REQ8 == (X) \
4971 || EXACTFUP == (X) ) \
4972 ? EXACTFU \
4973 : ( EXACTFAA == (X) ) \
4974 ? EXACTFAA \
4975 : ( EXACTL == (X) ) \
4976 ? EXACTL \
4977 : ( EXACTFLU8 == (X) ) \
4978 ? EXACTFLU8 \
4979 : 0 )
4980
4981 /* dont use tail as the end marker for this traverse */
4982 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4983 regnode * const noper = NEXTOPER( cur );
4984 U8 noper_type = OP( noper );
4985 U8 noper_trietype = TRIE_TYPE( noper_type );
4986 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4987 regnode * const noper_next = regnext( noper );
4988 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4989 U8 noper_next_trietype = (noper_next && noper_next < tail) ? TRIE_TYPE( noper_next_type ) :0;
4990 #endif
4991
4992 DEBUG_TRIE_COMPILE_r({
4993 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4994 Perl_re_indentf( aTHX_ "- %d:%s (%d)",
4995 depth+1,
4996 REG_NODE_NUM(cur), SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4997
4998 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4999 Perl_re_printf( aTHX_ " -> %d:%s",
5000 REG_NODE_NUM(noper), SvPV_nolen_const(RExC_mysv));
5001
5002 if ( noper_next ) {
5003 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
5004 Perl_re_printf( aTHX_ "\t=> %d:%s\t",
5005 REG_NODE_NUM(noper_next), SvPV_nolen_const(RExC_mysv));
5006 }
5007 Perl_re_printf( aTHX_ "(First==%d,Last==%d,Cur==%d,tt==%s,ntt==%s,nntt==%s)\n",
5008 REG_NODE_NUM(first), REG_NODE_NUM(prev), REG_NODE_NUM(cur),
5009 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
5010 );
5011 });
5012
5013 /* Is noper a trieable nodetype that can be merged
5014 * with the current trie (if there is one)? */
5015 if ( noper_trietype
5016 &&
5017 (
5018 ( noper_trietype == NOTHING )
5019 || ( trietype == NOTHING )
5020 || ( trietype == noper_trietype )
5021 )
5022 #ifdef NOJUMPTRIE
5023 && noper_next >= tail
5024 #endif
5025 && count < U16_MAX)
5026 {
5027 /* Handle mergable triable node Either we are
5028 * the first node in a new trieable sequence,
5029 * in which case we do some bookkeeping,
5030 * otherwise we update the end pointer. */
5031 if ( !first ) {
5032 first = cur;
5033 if ( noper_trietype == NOTHING ) {
5034 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
5035 regnode * const noper_next = regnext( noper );
5036 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
5037 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
5038 #endif
5039
5040 if ( noper_next_trietype ) {
5041 trietype = noper_next_trietype;
5042 } else if (noper_next_type) {
5043 /* a NOTHING regop is 1 regop wide.
5044 * We need at least two for a trie
5045 * so we can't merge this in */
5046 first = NULL;
5047 }
5048 } else {
5049 trietype = noper_trietype;
5050 }
5051 } else {
5052 if ( trietype == NOTHING )
5053 trietype = noper_trietype;
5054 prev = cur;
5055 }
5056 if (first)
5057 count++;
5058 } /* end handle mergable triable node */
5059 else {
5060 /* handle unmergable node -
5061 * noper may either be a triable node which can
5062 * not be tried together with the current trie,
5063 * or a non triable node */
5064 if ( prev ) {
5065 /* If last is set and trietype is not
5066 * NOTHING then we have found at least two
5067 * triable branch sequences in a row of a
5068 * similar trietype so we can turn them
5069 * into a trie. If/when we allow NOTHING to
5070 * start a trie sequence this condition
5071 * will be required, and it isn't expensive
5072 * so we leave it in for now. */
5073 if ( trietype && trietype != NOTHING )
5074 make_trie( pRExC_state,
5075 startbranch, first, cur, tail,
5076 count, trietype, depth+1 );
5077 prev = NULL; /* note: we clear/update
5078 first, trietype etc below,
5079 so we dont do it here */
5080 }
5081 if ( noper_trietype
5082 #ifdef NOJUMPTRIE
5083 && noper_next >= tail
5084 #endif
5085 ){
5086 /* noper is triable, so we can start a new
5087 * trie sequence */
5088 count = 1;
5089 first = cur;
5090 trietype = noper_trietype;
5091 } else if (first) {
5092 /* if we already saw a first but the
5093 * current node is not triable then we have
5094 * to reset the first information. */
5095 count = 0;
5096 first = NULL;
5097 trietype = 0;
5098 }
5099 } /* end handle unmergable node */
5100 } /* loop over branches */
5101 DEBUG_TRIE_COMPILE_r({
5102 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
5103 Perl_re_indentf( aTHX_ "- %s (%d) <SCAN FINISHED> ",
5104 depth+1, SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur));
5105 Perl_re_printf( aTHX_ "(First==%d, Last==%d, Cur==%d, tt==%s)\n",
5106 REG_NODE_NUM(first), REG_NODE_NUM(prev), REG_NODE_NUM(cur),
5107 PL_reg_name[trietype]
5108 );
5109
5110 });
5111 if ( prev && trietype ) {
5112 if ( trietype != NOTHING ) {
5113 /* the last branch of the sequence was part of
5114 * a trie, so we have to construct it here
5115 * outside of the loop */
5116 made= make_trie( pRExC_state, startbranch,
5117 first, scan, tail, count,
5118 trietype, depth+1 );
5119 #ifdef TRIE_STUDY_OPT
5120 if ( ((made == MADE_EXACT_TRIE &&
5121 startbranch == first)
5122 || ( first_non_open == first )) &&
5123 depth==0 ) {
5124 flags |= SCF_TRIE_RESTUDY;
5125 if ( startbranch == first
5126 && scan >= tail )
5127 {
5128 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
5129 }
5130 }
5131 #endif
5132 } else {
5133 /* at this point we know whatever we have is a
5134 * NOTHING sequence/branch AND if 'startbranch'
5135 * is 'first' then we can turn the whole thing
5136 * into a NOTHING
5137 */
5138 if ( startbranch == first ) {
5139 regnode *opt;
5140 /* the entire thing is a NOTHING sequence,
5141 * something like this: (?:|) So we can
5142 * turn it into a plain NOTHING op. */
5143 DEBUG_TRIE_COMPILE_r({
5144 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
5145 Perl_re_indentf( aTHX_ "- %s (%d) <NOTHING BRANCH SEQUENCE>\n",
5146 depth+1,
5147 SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur));
5148
5149 });
5150 OP(startbranch)= NOTHING;
5151 NEXT_OFF(startbranch)= tail - startbranch;
5152 for ( opt= startbranch + 1; opt < tail ; opt++ )
5153 OP(opt)= OPTIMIZED;
5154 }
5155 }
5156 } /* end if ( prev) */
5157 } /* TRIE_MAXBUF is non zero */
5158 } /* do trie */
5159
5160 }
5161 else if ( code == BRANCHJ ) { /* single branch is optimized. */
5162 scan = NEXTOPER(NEXTOPER(scan));
5163 } else /* single branch is optimized. */
5164 scan = NEXTOPER(scan);
5165 continue;
5166 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB) {
5167 I32 paren = 0;
5168 regnode *start = NULL;
5169 regnode *end = NULL;
5170 U32 my_recursed_depth= recursed_depth;
5171
5172 if (OP(scan) != SUSPEND) { /* GOSUB */
5173 /* Do setup, note this code has side effects beyond
5174 * the rest of this block. Specifically setting
5175 * RExC_recurse[] must happen at least once during
5176 * study_chunk(). */
5177 paren = ARG(scan);
5178 RExC_recurse[ARG2L(scan)] = scan;
5179 start = REGNODE_p(RExC_open_parens[paren]);
5180 end = REGNODE_p(RExC_close_parens[paren]);
5181
5182 /* NOTE we MUST always execute the above code, even
5183 * if we do nothing with a GOSUB */
5184 if (
5185 ( flags & SCF_IN_DEFINE )
5186 ||
5187 (
5188 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
5189 &&
5190 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
5191 )
5192 ) {
5193 /* no need to do anything here if we are in a define. */
5194 /* or we are after some kind of infinite construct
5195 * so we can skip recursing into this item.
5196 * Since it is infinite we will not change the maxlen
5197 * or delta, and if we miss something that might raise
5198 * the minlen it will merely pessimise a little.
5199 *
5200 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
5201 * might result in a minlen of 1 and not of 4,
5202 * but this doesn't make us mismatch, just try a bit
5203 * harder than we should.
5204 *
5205 * However we must assume this GOSUB is infinite, to
5206 * avoid wrongly applying other optimizations in the
5207 * enclosing scope - see GH 18096, for example.
5208 */
5209 is_inf = is_inf_internal = 1;
5210 scan= regnext(scan);
5211 continue;
5212 }
5213
5214 if (
5215 !recursed_depth
5216 || !PAREN_TEST(recursed_depth - 1, paren)
5217 ) {
5218 /* it is quite possible that there are more efficient ways
5219 * to do this. We maintain a bitmap per level of recursion
5220 * of which patterns we have entered so we can detect if a
5221 * pattern creates a possible infinite loop. When we
5222 * recurse down a level we copy the previous levels bitmap
5223 * down. When we are at recursion level 0 we zero the top
5224 * level bitmap. It would be nice to implement a different
5225 * more efficient way of doing this. In particular the top
5226 * level bitmap may be unnecessary.
5227 */
5228 if (!recursed_depth) {
5229 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
5230 } else {
5231 Copy(PAREN_OFFSET(recursed_depth - 1),
5232 PAREN_OFFSET(recursed_depth),
5233 RExC_study_chunk_recursed_bytes, U8);
5234 }
5235 /* we havent recursed into this paren yet, so recurse into it */
5236 DEBUG_STUDYDATA("gosub-set", data, depth, is_inf);
5237 PAREN_SET(recursed_depth, paren);
5238 my_recursed_depth= recursed_depth + 1;
5239 } else {
5240 DEBUG_STUDYDATA("gosub-inf", data, depth, is_inf);
5241 /* some form of infinite recursion, assume infinite length
5242 * */
5243 if (flags & SCF_DO_SUBSTR) {
5244 scan_commit(pRExC_state, data, minlenp, is_inf);
5245 data->cur_is_floating = 1;
5246 }
5247 is_inf = is_inf_internal = 1;
5248 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5249 ssc_anything(data->start_class);
5250 flags &= ~SCF_DO_STCLASS;
5251
5252 start= NULL; /* reset start so we dont recurse later on. */
5253 }
5254 } else {
5255 paren = stopparen;
5256 start = scan + 2;
5257 end = regnext(scan);
5258 }
5259 if (start) {
5260 scan_frame *newframe;
5261 assert(end);
5262 if (!RExC_frame_last) {
5263 Newxz(newframe, 1, scan_frame);
5264 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
5265 RExC_frame_head= newframe;
5266 RExC_frame_count++;
5267 } else if (!RExC_frame_last->next_frame) {
5268 Newxz(newframe, 1, scan_frame);
5269 RExC_frame_last->next_frame= newframe;
5270 newframe->prev_frame= RExC_frame_last;
5271 RExC_frame_count++;
5272 } else {
5273 newframe= RExC_frame_last->next_frame;
5274 }
5275 RExC_frame_last= newframe;
5276
5277 newframe->next_regnode = regnext(scan);
5278 newframe->last_regnode = last;
5279 newframe->stopparen = stopparen;
5280 newframe->prev_recursed_depth = recursed_depth;
5281 newframe->this_prev_frame= frame;
5282 newframe->in_gosub = (
5283 (frame && frame->in_gosub) || OP(scan) == GOSUB
5284 );
5285
5286 DEBUG_STUDYDATA("frame-new", data, depth, is_inf);
5287 DEBUG_PEEP("fnew", scan, depth, flags);
5288
5289 frame = newframe;
5290 scan = start;
5291 stopparen = paren;
5292 last = end;
5293 depth = depth + 1;
5294 recursed_depth= my_recursed_depth;
5295
5296 continue;
5297 }
5298 }
5299 else if (PL_regkind[OP(scan)] == EXACT && ! isEXACTFish(OP(scan))) {
5300 SSize_t bytelen = STR_LEN(scan), charlen;
5301 UV uc;
5302 assert(bytelen);
5303 if (UTF) {
5304 const U8 * const s = (U8*)STRING(scan);
5305 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
5306 charlen = utf8_length(s, s + bytelen);
5307 } else {
5308 uc = *((U8*)STRING(scan));
5309 charlen = bytelen;
5310 }
5311 min += charlen;
5312 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
5313 /* The code below prefers earlier match for fixed
5314 offset, later match for variable offset. */
5315 if (data->last_end == -1) { /* Update the start info. */
5316 data->last_start_min = data->pos_min;
5317 data->last_start_max =
5318 is_inf ? OPTIMIZE_INFTY
5319 : (data->pos_delta > OPTIMIZE_INFTY - data->pos_min)
5320 ? OPTIMIZE_INFTY : data->pos_min + data->pos_delta;
5321 }
5322 sv_catpvn(data->last_found, STRING(scan), bytelen);
5323 if (UTF)
5324 SvUTF8_on(data->last_found);
5325 {
5326 SV * const sv = data->last_found;
5327 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5328 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5329 if (mg && mg->mg_len >= 0)
5330 mg->mg_len += charlen;
5331 }
5332 data->last_end = data->pos_min + charlen;
5333 data->pos_min += charlen; /* As in the first entry. */
5334 data->flags &= ~SF_BEFORE_EOL;
5335 }
5336
5337 /* ANDing the code point leaves at most it, and not in locale, and
5338 * can't match null string */
5339 if (flags & SCF_DO_STCLASS_AND) {
5340 ssc_cp_and(data->start_class, uc);
5341 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5342 ssc_clear_locale(data->start_class);
5343 }
5344 else if (flags & SCF_DO_STCLASS_OR) {
5345 ssc_add_cp(data->start_class, uc);
5346 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5347
5348 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5349 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5350 }
5351 flags &= ~SCF_DO_STCLASS;
5352 }
5353 else if (PL_regkind[OP(scan)] == EXACT) {
5354 /* But OP != EXACT!, so is EXACTFish */
5355 SSize_t bytelen = STR_LEN(scan), charlen;
5356 const U8 * s = (U8*)STRING(scan);
5357
5358 /* Replace a length 1 ASCII fold pair node with an ANYOFM node,
5359 * with the mask set to the complement of the bit that differs
5360 * between upper and lower case, and the lowest code point of the
5361 * pair (which the '&' forces) */
5362 if ( bytelen == 1
5363 && isALPHA_A(*s)
5364 && ( OP(scan) == EXACTFAA
5365 || ( OP(scan) == EXACTFU
5366 && ! HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(*s)))
5367 && mutate_ok
5368 ) {
5369 U8 mask = ~ ('A' ^ 'a'); /* These differ in just one bit */
5370
5371 OP(scan) = ANYOFM;
5372 ARG_SET(scan, *s & mask);
5373 FLAGS(scan) = mask;
5374 /* we're not EXACTFish any more, so restudy */
5375 continue;
5376 }
5377
5378 /* Search for fixed substrings supports EXACT only. */
5379 if (flags & SCF_DO_SUBSTR) {
5380 assert(data);
5381 scan_commit(pRExC_state, data, minlenp, is_inf);
5382 }
5383 charlen = UTF ? (SSize_t) utf8_length(s, s + bytelen) : bytelen;
5384 if (unfolded_multi_char) {
5385 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
5386 }
5387 min += charlen - min_subtract;
5388 assert (min >= 0);
5389 if ((SSize_t)min_subtract < OPTIMIZE_INFTY
5390 && delta < OPTIMIZE_INFTY - (SSize_t)min_subtract
5391 ) {
5392 delta += min_subtract;
5393 } else {
5394 delta = OPTIMIZE_INFTY;
5395 }
5396 if (flags & SCF_DO_SUBSTR) {
5397 data->pos_min += charlen - min_subtract;
5398 if (data->pos_min < 0) {
5399 data->pos_min = 0;
5400 }
5401 if ((SSize_t)min_subtract < OPTIMIZE_INFTY
5402 && data->pos_delta < OPTIMIZE_INFTY - (SSize_t)min_subtract
5403 ) {
5404 data->pos_delta += min_subtract;
5405 } else {
5406 data->pos_delta = OPTIMIZE_INFTY;
5407 }
5408 if (min_subtract) {
5409 data->cur_is_floating = 1; /* float */
5410 }
5411 }
5412
5413 if (flags & SCF_DO_STCLASS) {
5414 SV* EXACTF_invlist = make_exactf_invlist(pRExC_state, scan);
5415
5416 assert(EXACTF_invlist);
5417 if (flags & SCF_DO_STCLASS_AND) {
5418 if (OP(scan) != EXACTFL)
5419 ssc_clear_locale(data->start_class);
5420 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5421 ANYOF_POSIXL_ZERO(data->start_class);
5422 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
5423 }
5424 else { /* SCF_DO_STCLASS_OR */
5425 ssc_union(data->start_class, EXACTF_invlist, FALSE);
5426 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5427
5428 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5429 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5430 }
5431 flags &= ~SCF_DO_STCLASS;
5432 SvREFCNT_dec(EXACTF_invlist);
5433 }
5434 }
5435 else if (REGNODE_VARIES(OP(scan))) {
5436 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
5437 I32 fl = 0, f = flags;
5438 regnode * const oscan = scan;
5439 regnode_ssc this_class;
5440 regnode_ssc *oclass = NULL;
5441 I32 next_is_eval = 0;
5442
5443 switch (PL_regkind[OP(scan)]) {
5444 case WHILEM: /* End of (?:...)* . */
5445 scan = NEXTOPER(scan);
5446 goto finish;
5447 case PLUS:
5448 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
5449 next = NEXTOPER(scan);
5450 if ( ( PL_regkind[OP(next)] == EXACT
5451 && ! isEXACTFish(OP(next)))
5452 || (flags & SCF_DO_STCLASS))
5453 {
5454 mincount = 1;
5455 maxcount = REG_INFTY;
5456 next = regnext(scan);
5457 scan = NEXTOPER(scan);
5458 goto do_curly;
5459 }
5460 }
5461 if (flags & SCF_DO_SUBSTR)
5462 data->pos_min++;
5463 /* This will bypass the formal 'min += minnext * mincount'
5464 * calculation in the do_curly path, so assumes min width
5465 * of the PLUS payload is exactly one. */
5466 min++;
5467 /* FALLTHROUGH */
5468 case STAR:
5469 next = NEXTOPER(scan);
5470
5471 /* This temporary node can now be turned into EXACTFU, and
5472 * must, as regexec.c doesn't handle it */
5473 if (OP(next) == EXACTFU_S_EDGE && mutate_ok) {
5474 OP(next) = EXACTFU;
5475 }
5476
5477 if ( STR_LEN(next) == 1
5478 && isALPHA_A(* STRING(next))
5479 && ( OP(next) == EXACTFAA
5480 || ( OP(next) == EXACTFU
5481 && ! HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(* STRING(next))))
5482 && mutate_ok
5483 ) {
5484 /* These differ in just one bit */
5485 U8 mask = ~ ('A' ^ 'a');
5486
5487 assert(isALPHA_A(* STRING(next)));
5488
5489 /* Then replace it by an ANYOFM node, with
5490 * the mask set to the complement of the
5491 * bit that differs between upper and lower
5492 * case, and the lowest code point of the
5493 * pair (which the '&' forces) */
5494 OP(next) = ANYOFM;
5495 ARG_SET(next, *STRING(next) & mask);
5496 FLAGS(next) = mask;
5497 }
5498
5499 if (flags & SCF_DO_STCLASS) {
5500 mincount = 0;
5501 maxcount = REG_INFTY;
5502 next = regnext(scan);
5503 scan = NEXTOPER(scan);
5504 goto do_curly;
5505 }
5506 if (flags & SCF_DO_SUBSTR) {
5507 scan_commit(pRExC_state, data, minlenp, is_inf);
5508 /* Cannot extend fixed substrings */
5509 data->cur_is_floating = 1; /* float */
5510 }
5511 is_inf = is_inf_internal = 1;
5512 scan = regnext(scan);
5513 goto optimize_curly_tail;
5514 case CURLY:
5515 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
5516 && (scan->flags == stopparen))
5517 {
5518 mincount = 1;
5519 maxcount = 1;
5520 } else {
5521 mincount = ARG1(scan);
5522 maxcount = ARG2(scan);
5523 }
5524 next = regnext(scan);
5525 if (OP(scan) == CURLYX) {
5526 I32 lp = (data ? *(data->last_closep) : 0);
5527 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
5528 }
5529 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
5530 next_is_eval = (OP(scan) == EVAL);
5531 do_curly:
5532 if (flags & SCF_DO_SUBSTR) {
5533 if (mincount == 0)
5534 scan_commit(pRExC_state, data, minlenp, is_inf);
5535 /* Cannot extend fixed substrings */
5536 pos_before = data->pos_min;
5537 }
5538 if (data) {
5539 fl = data->flags;
5540 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
5541 if (is_inf)
5542 data->flags |= SF_IS_INF;
5543 }
5544 if (flags & SCF_DO_STCLASS) {
5545 ssc_init(pRExC_state, &this_class);
5546 oclass = data->start_class;
5547 data->start_class = &this_class;
5548 f |= SCF_DO_STCLASS_AND;
5549 f &= ~SCF_DO_STCLASS_OR;
5550 }
5551 /* Exclude from super-linear cache processing any {n,m}
5552 regops for which the combination of input pos and regex
5553 pos is not enough information to determine if a match
5554 will be possible.
5555
5556 For example, in the regex /foo(bar\s*){4,8}baz/ with the
5557 regex pos at the \s*, the prospects for a match depend not
5558 only on the input position but also on how many (bar\s*)
5559 repeats into the {4,8} we are. */
5560 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
5561 f &= ~SCF_WHILEM_VISITED_POS;
5562
5563 /* This will finish on WHILEM, setting scan, or on NULL: */
5564 /* recurse study_chunk() on loop bodies */
5565 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
5566 last, data, stopparen, recursed_depth, NULL,
5567 (mincount == 0
5568 ? (f & ~SCF_DO_SUBSTR)
5569 : f)
5570 , depth+1, mutate_ok);
5571
5572 if (flags & SCF_DO_STCLASS)
5573 data->start_class = oclass;
5574 if (mincount == 0 || minnext == 0) {
5575 if (flags & SCF_DO_STCLASS_OR) {
5576 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5577 }
5578 else if (flags & SCF_DO_STCLASS_AND) {
5579 /* Switch to OR mode: cache the old value of
5580 * data->start_class */
5581 INIT_AND_WITHP;
5582 StructCopy(data->start_class, and_withp, regnode_ssc);
5583 flags &= ~SCF_DO_STCLASS_AND;
5584 StructCopy(&this_class, data->start_class, regnode_ssc);
5585 flags |= SCF_DO_STCLASS_OR;
5586 ANYOF_FLAGS(data->start_class)
5587 |= SSC_MATCHES_EMPTY_STRING;
5588 }
5589 } else { /* Non-zero len */
5590 if (flags & SCF_DO_STCLASS_OR) {
5591 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5592 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5593 }
5594 else if (flags & SCF_DO_STCLASS_AND)
5595 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5596 flags &= ~SCF_DO_STCLASS;
5597 }
5598 if (!scan) /* It was not CURLYX, but CURLY. */
5599 scan = next;
5600 if (((flags & (SCF_TRIE_DOING_RESTUDY|SCF_DO_SUBSTR))==SCF_DO_SUBSTR)
5601 /* ? quantifier ok, except for (?{ ... }) */
5602 && (next_is_eval || !(mincount == 0 && maxcount == 1))
5603 && (minnext == 0) && (deltanext == 0)
5604 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
5605 && maxcount <= REG_INFTY/3) /* Complement check for big
5606 count */
5607 {
5608 _WARN_HELPER(RExC_precomp_end, packWARN(WARN_REGEXP),
5609 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
5610 "Quantifier unexpected on zero-length expression "
5611 "in regex m/%" UTF8f "/",
5612 UTF8fARG(UTF, RExC_precomp_end - RExC_precomp,
5613 RExC_precomp)));
5614 }
5615
5616 if ( ( minnext > 0 && mincount >= SSize_t_MAX / minnext )
5617 || min >= SSize_t_MAX - minnext * mincount )
5618 {
5619 FAIL("Regexp out of space");
5620 }
5621
5622 min += minnext * mincount;
5623 is_inf_internal |= deltanext == OPTIMIZE_INFTY
5624 || (maxcount == REG_INFTY && minnext + deltanext > 0);
5625 is_inf |= is_inf_internal;
5626 if (is_inf) {
5627 delta = OPTIMIZE_INFTY;
5628 } else {
5629 delta += (minnext + deltanext) * maxcount
5630 - minnext * mincount;
5631 }
5632 /* Try powerful optimization CURLYX => CURLYN. */
5633 if ( OP(oscan) == CURLYX && data
5634 && data->flags & SF_IN_PAR
5635 && !(data->flags & SF_HAS_EVAL)
5636 && !deltanext && minnext == 1
5637 && mutate_ok
5638 ) {
5639 /* Try to optimize to CURLYN. */
5640 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
5641 regnode * const nxt1 = nxt;
5642 #ifdef DEBUGGING
5643 regnode *nxt2;
5644 #endif
5645
5646 /* Skip open. */
5647 nxt = regnext(nxt);
5648 if (!REGNODE_SIMPLE(OP(nxt))
5649 && !(PL_regkind[OP(nxt)] == EXACT
5650 && STR_LEN(nxt) == 1))
5651 goto nogo;
5652 #ifdef DEBUGGING
5653 nxt2 = nxt;
5654 #endif
5655 nxt = regnext(nxt);
5656 if (OP(nxt) != CLOSE)
5657 goto nogo;
5658 if (RExC_open_parens) {
5659
5660 /*open->CURLYM*/
5661 RExC_open_parens[ARG(nxt1)] = REGNODE_OFFSET(oscan);
5662
5663 /*close->while*/
5664 RExC_close_parens[ARG(nxt1)] = REGNODE_OFFSET(nxt) + 2;
5665 }
5666 /* Now we know that nxt2 is the only contents: */
5667 oscan->flags = (U8)ARG(nxt);
5668 OP(oscan) = CURLYN;
5669 OP(nxt1) = NOTHING; /* was OPEN. */
5670
5671 #ifdef DEBUGGING
5672 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5673 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
5674 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
5675 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5676 OP(nxt + 1) = OPTIMIZED; /* was count. */
5677 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
5678 #endif
5679 }
5680 nogo:
5681
5682 /* Try optimization CURLYX => CURLYM. */
5683 if ( OP(oscan) == CURLYX && data
5684 && !(data->flags & SF_HAS_PAR)
5685 && !(data->flags & SF_HAS_EVAL)
5686 && !deltanext /* atom is fixed width */
5687 && minnext != 0 /* CURLYM can't handle zero width */
5688 /* Nor characters whose fold at run-time may be
5689 * multi-character */
5690 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
5691 && mutate_ok
5692 ) {
5693 /* XXXX How to optimize if data == 0? */
5694 /* Optimize to a simpler form. */
5695 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
5696 regnode *nxt2;
5697
5698 OP(oscan) = CURLYM;
5699 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
5700 && (OP(nxt2) != WHILEM))
5701 nxt = nxt2;
5702 OP(nxt2) = SUCCEED; /* Whas WHILEM */
5703 /* Need to optimize away parenths. */
5704 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
5705 /* Set the parenth number. */
5706 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
5707
5708 oscan->flags = (U8)ARG(nxt);
5709 if (RExC_open_parens) {
5710 /*open->CURLYM*/
5711 RExC_open_parens[ARG(nxt1)] = REGNODE_OFFSET(oscan);
5712
5713 /*close->NOTHING*/
5714 RExC_close_parens[ARG(nxt1)] = REGNODE_OFFSET(nxt2)
5715 + 1;
5716 }
5717 OP(nxt1) = OPTIMIZED; /* was OPEN. */
5718 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5719
5720 #ifdef DEBUGGING
5721 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5722 OP(nxt + 1) = OPTIMIZED; /* was count. */
5723 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
5724 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
5725 #endif
5726 #if 0
5727 while ( nxt1 && (OP(nxt1) != WHILEM)) {
5728 regnode *nnxt = regnext(nxt1);
5729 if (nnxt == nxt) {
5730 if (reg_off_by_arg[OP(nxt1)])
5731 ARG_SET(nxt1, nxt2 - nxt1);
5732 else if (nxt2 - nxt1 < U16_MAX)
5733 NEXT_OFF(nxt1) = nxt2 - nxt1;
5734 else
5735 OP(nxt) = NOTHING; /* Cannot beautify */
5736 }
5737 nxt1 = nnxt;
5738 }
5739 #endif
5740 /* Optimize again: */
5741 /* recurse study_chunk() on optimised CURLYX => CURLYM */
5742 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
5743 NULL, stopparen, recursed_depth, NULL, 0,
5744 depth+1, mutate_ok);
5745 }
5746 else
5747 oscan->flags = 0;
5748 }
5749 else if ((OP(oscan) == CURLYX)
5750 && (flags & SCF_WHILEM_VISITED_POS)
5751 /* See the comment on a similar expression above.
5752 However, this time it's not a subexpression
5753 we care about, but the expression itself. */
5754 && (maxcount == REG_INFTY)
5755 && data) {
5756 /* This stays as CURLYX, we can put the count/of pair. */
5757 /* Find WHILEM (as in regexec.c) */
5758 regnode *nxt = oscan + NEXT_OFF(oscan);
5759
5760 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
5761 nxt += ARG(nxt);
5762 nxt = PREVOPER(nxt);
5763 if (nxt->flags & 0xf) {
5764 /* we've already set whilem count on this node */
5765 } else if (++data->whilem_c < 16) {
5766 assert(data->whilem_c <= RExC_whilem_seen);
5767 nxt->flags = (U8)(data->whilem_c
5768 | (RExC_whilem_seen << 4)); /* On WHILEM */
5769 }
5770 }
5771 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
5772 pars++;
5773 if (flags & SCF_DO_SUBSTR) {
5774 SV *last_str = NULL;
5775 STRLEN last_chrs = 0;
5776 int counted = mincount != 0;
5777
5778 if (data->last_end > 0 && mincount != 0) { /* Ends with a
5779 string. */
5780 SSize_t b = pos_before >= data->last_start_min
5781 ? pos_before : data->last_start_min;
5782 STRLEN l;
5783 const char * const s = SvPV_const(data->last_found, l);
5784 SSize_t old = b - data->last_start_min;
5785 assert(old >= 0);
5786
5787 if (UTF)
5788 old = utf8_hop_forward((U8*)s, old,
5789 (U8 *) SvEND(data->last_found))
5790 - (U8*)s;
5791 l -= old;
5792 /* Get the added string: */
5793 last_str = newSVpvn_utf8(s + old, l, UTF);
5794 last_chrs = UTF ? utf8_length((U8*)(s + old),
5795 (U8*)(s + old + l)) : l;
5796 if (deltanext == 0 && pos_before == b) {
5797 /* What was added is a constant string */
5798 if (mincount > 1) {
5799
5800 SvGROW(last_str, (mincount * l) + 1);
5801 repeatcpy(SvPVX(last_str) + l,
5802 SvPVX_const(last_str), l,
5803 mincount - 1);
5804 SvCUR_set(last_str, SvCUR(last_str) * mincount);
5805 /* Add additional parts. */
5806 SvCUR_set(data->last_found,
5807 SvCUR(data->last_found) - l);
5808 sv_catsv(data->last_found, last_str);
5809 {
5810 SV * sv = data->last_found;
5811 MAGIC *mg =
5812 SvUTF8(sv) && SvMAGICAL(sv) ?
5813 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5814 if (mg && mg->mg_len >= 0)
5815 mg->mg_len += last_chrs * (mincount-1);
5816 }
5817 last_chrs *= mincount;
5818 data->last_end += l * (mincount - 1);
5819 }
5820 } else {
5821 /* start offset must point into the last copy */
5822 data->last_start_min += minnext * (mincount - 1);
5823 data->last_start_max =
5824 is_inf
5825 ? OPTIMIZE_INFTY
5826 : data->last_start_max +
5827 (maxcount - 1) * (minnext + data->pos_delta);
5828 }
5829 }
5830 /* It is counted once already... */
5831 data->pos_min += minnext * (mincount - counted);
5832 #if 0
5833 Perl_re_printf( aTHX_ "counted=%" UVuf " deltanext=%" UVuf
5834 " OPTIMIZE_INFTY=%" UVuf " minnext=%" UVuf
5835 " maxcount=%" UVuf " mincount=%" UVuf
5836 " data->pos_delta=%" UVuf "\n",
5837 (UV)counted, (UV)deltanext, (UV)OPTIMIZE_INFTY, (UV)minnext,
5838 (UV)maxcount, (UV)mincount, (UV)data->pos_delta);
5839 if (deltanext != OPTIMIZE_INFTY)
5840 Perl_re_printf( aTHX_ "LHS=%" UVuf " RHS=%" UVuf "\n",
5841 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5842 - minnext * mincount), (UV)(OPTIMIZE_INFTY - data->pos_delta));
5843 #endif
5844 if (deltanext == OPTIMIZE_INFTY
5845 || data->pos_delta == OPTIMIZE_INFTY
5846 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= OPTIMIZE_INFTY - data->pos_delta)
5847 data->pos_delta = OPTIMIZE_INFTY;
5848 else
5849 data->pos_delta += - counted * deltanext +
5850 (minnext + deltanext) * maxcount - minnext * mincount;
5851 if (mincount != maxcount) {
5852 /* Cannot extend fixed substrings found inside
5853 the group. */
5854 scan_commit(pRExC_state, data, minlenp, is_inf);
5855 if (mincount && last_str) {
5856 SV * const sv = data->last_found;
5857 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5858 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5859
5860 if (mg)
5861 mg->mg_len = -1;
5862 sv_setsv(sv, last_str);
5863 data->last_end = data->pos_min;
5864 data->last_start_min = data->pos_min - last_chrs;
5865 data->last_start_max = is_inf
5866 ? OPTIMIZE_INFTY
5867 : data->pos_min + data->pos_delta - last_chrs;
5868 }
5869 data->cur_is_floating = 1; /* float */
5870 }
5871 SvREFCNT_dec(last_str);
5872 }
5873 if (data && (fl & SF_HAS_EVAL))
5874 data->flags |= SF_HAS_EVAL;
5875 optimize_curly_tail:
5876 rck_elide_nothing(oscan);
5877 continue;
5878
5879 default:
5880 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5881 OP(scan));
5882 case REF:
5883 case CLUMP:
5884 if (flags & SCF_DO_SUBSTR) {
5885 /* Cannot expect anything... */
5886 scan_commit(pRExC_state, data, minlenp, is_inf);
5887 data->cur_is_floating = 1; /* float */
5888 }
5889 is_inf = is_inf_internal = 1;
5890 if (flags & SCF_DO_STCLASS_OR) {
5891 if (OP(scan) == CLUMP) {
5892 /* Actually is any start char, but very few code points
5893 * aren't start characters */
5894 ssc_match_all_cp(data->start_class);
5895 }
5896 else {
5897 ssc_anything(data->start_class);
5898 }
5899 }
5900 flags &= ~SCF_DO_STCLASS;
5901 break;
5902 }
5903 }
5904 else if (OP(scan) == LNBREAK) {
5905 if (flags & SCF_DO_STCLASS) {
5906 if (flags & SCF_DO_STCLASS_AND) {
5907 ssc_intersection(data->start_class,
5908 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5909 ssc_clear_locale(data->start_class);
5910 ANYOF_FLAGS(data->start_class)
5911 &= ~SSC_MATCHES_EMPTY_STRING;
5912 }
5913 else if (flags & SCF_DO_STCLASS_OR) {
5914 ssc_union(data->start_class,
5915 PL_XPosix_ptrs[_CC_VERTSPACE],
5916 FALSE);
5917 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5918
5919 /* See commit msg for
5920 * 749e076fceedeb708a624933726e7989f2302f6a */
5921 ANYOF_FLAGS(data->start_class)
5922 &= ~SSC_MATCHES_EMPTY_STRING;
5923 }
5924 flags &= ~SCF_DO_STCLASS;
5925 }
5926 min++;
5927 if (delta != OPTIMIZE_INFTY)
5928 delta++; /* Because of the 2 char string cr-lf */
5929 if (flags & SCF_DO_SUBSTR) {
5930 /* Cannot expect anything... */
5931 scan_commit(pRExC_state, data, minlenp, is_inf);
5932 data->pos_min += 1;
5933 if (data->pos_delta != OPTIMIZE_INFTY) {
5934 data->pos_delta += 1;
5935 }
5936 data->cur_is_floating = 1; /* float */
5937 }
5938 }
5939 else if (REGNODE_SIMPLE(OP(scan))) {
5940
5941 if (flags & SCF_DO_SUBSTR) {
5942 scan_commit(pRExC_state, data, minlenp, is_inf);
5943 data->pos_min++;
5944 }
5945 min++;
5946 if (flags & SCF_DO_STCLASS) {
5947 bool invert = 0;
5948 SV* my_invlist = NULL;
5949 U8 namedclass;
5950
5951 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5952 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5953
5954 /* Some of the logic below assumes that switching
5955 locale on will only add false positives. */
5956 switch (OP(scan)) {
5957
5958 default:
5959 #ifdef DEBUGGING
5960 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5961 OP(scan));
5962 #endif
5963 case SANY:
5964 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5965 ssc_match_all_cp(data->start_class);
5966 break;
5967
5968 case REG_ANY:
5969 {
5970 SV* REG_ANY_invlist = _new_invlist(2);
5971 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5972 '\n');
5973 if (flags & SCF_DO_STCLASS_OR) {
5974 ssc_union(data->start_class,
5975 REG_ANY_invlist,
5976 TRUE /* TRUE => invert, hence all but \n
5977 */
5978 );
5979 }
5980 else if (flags & SCF_DO_STCLASS_AND) {
5981 ssc_intersection(data->start_class,
5982 REG_ANY_invlist,
5983 TRUE /* TRUE => invert */
5984 );
5985 ssc_clear_locale(data->start_class);
5986 }
5987 SvREFCNT_dec_NN(REG_ANY_invlist);
5988 }
5989 break;
5990
5991 case ANYOFD:
5992 case ANYOFL:
5993 case ANYOFPOSIXL:
5994 case ANYOFH:
5995 case ANYOFHb:
5996 case ANYOFHr:
5997 case ANYOFHs:
5998 case ANYOF:
5999 if (flags & SCF_DO_STCLASS_AND)
6000 ssc_and(pRExC_state, data->start_class,
6001 (regnode_charclass *) scan);
6002 else
6003 ssc_or(pRExC_state, data->start_class,
6004 (regnode_charclass *) scan);
6005 break;
6006
6007 case NANYOFM: /* NANYOFM already contains the inversion of the
6008 input ANYOF data, so, unlike things like
6009 NPOSIXA, don't change 'invert' to TRUE */
6010 /* FALLTHROUGH */
6011 case ANYOFM:
6012 {
6013 SV* cp_list = get_ANYOFM_contents(scan);
6014
6015 if (flags & SCF_DO_STCLASS_OR) {
6016 ssc_union(data->start_class, cp_list, invert);
6017 }
6018 else if (flags & SCF_DO_STCLASS_AND) {
6019 ssc_intersection(data->start_class, cp_list, invert);
6020 }
6021
6022 SvREFCNT_dec_NN(cp_list);
6023 break;
6024 }
6025
6026 case ANYOFR:
6027 case ANYOFRb:
6028 {
6029 SV* cp_list = NULL;
6030
6031 cp_list = _add_range_to_invlist(cp_list,
6032 ANYOFRbase(scan),
6033 ANYOFRbase(scan) + ANYOFRdelta(scan));
6034
6035 if (flags & SCF_DO_STCLASS_OR) {
6036 ssc_union(data->start_class, cp_list, invert);
6037 }
6038 else if (flags & SCF_DO_STCLASS_AND) {
6039 ssc_intersection(data->start_class, cp_list, invert);
6040 }
6041
6042 SvREFCNT_dec_NN(cp_list);
6043 break;
6044 }
6045
6046 case NPOSIXL:
6047 invert = 1;
6048 /* FALLTHROUGH */
6049
6050 case POSIXL:
6051 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
6052 if (flags & SCF_DO_STCLASS_AND) {
6053 bool was_there = cBOOL(
6054 ANYOF_POSIXL_TEST(data->start_class,
6055 namedclass));
6056 ANYOF_POSIXL_ZERO(data->start_class);
6057 if (was_there) { /* Do an AND */
6058 ANYOF_POSIXL_SET(data->start_class, namedclass);
6059 }
6060 /* No individual code points can now match */
6061 data->start_class->invlist
6062 = sv_2mortal(_new_invlist(0));
6063 }
6064 else {
6065 int complement = namedclass + ((invert) ? -1 : 1);
6066
6067 assert(flags & SCF_DO_STCLASS_OR);
6068
6069 /* If the complement of this class was already there,
6070 * the result is that they match all code points,
6071 * (\d + \D == everything). Remove the classes from
6072 * future consideration. Locale is not relevant in
6073 * this case */
6074 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
6075 ssc_match_all_cp(data->start_class);
6076 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
6077 ANYOF_POSIXL_CLEAR(data->start_class, complement);
6078 }
6079 else { /* The usual case; just add this class to the
6080 existing set */
6081 ANYOF_POSIXL_SET(data->start_class, namedclass);
6082 }
6083 }
6084 break;
6085
6086 case NPOSIXA: /* For these, we always know the exact set of
6087 what's matched */
6088 invert = 1;
6089 /* FALLTHROUGH */
6090 case POSIXA:
6091 my_invlist = invlist_clone(PL_Posix_ptrs[FLAGS(scan)], NULL);
6092 goto join_posix_and_ascii;
6093
6094 case NPOSIXD:
6095 case NPOSIXU:
6096 invert = 1;
6097 /* FALLTHROUGH */
6098 case POSIXD:
6099 case POSIXU:
6100 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)], NULL);
6101
6102 /* NPOSIXD matches all upper Latin1 code points unless the
6103 * target string being matched is UTF-8, which is
6104 * unknowable until match time. Since we are going to
6105 * invert, we want to get rid of all of them so that the
6106 * inversion will match all */
6107 if (OP(scan) == NPOSIXD) {
6108 _invlist_subtract(my_invlist, PL_UpperLatin1,
6109 &my_invlist);
6110 }
6111
6112 join_posix_and_ascii:
6113
6114 if (flags & SCF_DO_STCLASS_AND) {
6115 ssc_intersection(data->start_class, my_invlist, invert);
6116 ssc_clear_locale(data->start_class);
6117 }
6118 else {
6119 assert(flags & SCF_DO_STCLASS_OR);
6120 ssc_union(data->start_class, my_invlist, invert);
6121 }
6122 SvREFCNT_dec(my_invlist);
6123 }
6124 if (flags & SCF_DO_STCLASS_OR)
6125 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6126 flags &= ~SCF_DO_STCLASS;
6127 }
6128 }
6129 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
6130 data->flags |= (OP(scan) == MEOL
6131 ? SF_BEFORE_MEOL
6132 : SF_BEFORE_SEOL);
6133 scan_commit(pRExC_state, data, minlenp, is_inf);
6134
6135 }
6136 else if ( PL_regkind[OP(scan)] == BRANCHJ
6137 /* Lookbehind, or need to calculate parens/evals/stclass: */
6138 && (scan->flags || data || (flags & SCF_DO_STCLASS))
6139 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
6140 {
6141 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
6142 || OP(scan) == UNLESSM )
6143 {
6144 /* Negative Lookahead/lookbehind
6145 In this case we can't do fixed string optimisation.
6146 */
6147
6148 SSize_t deltanext, minnext, fake = 0;
6149 regnode *nscan;
6150 regnode_ssc intrnl;
6151 int f = 0;
6152
6153 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
6154 if (data) {
6155 data_fake.whilem_c = data->whilem_c;
6156 data_fake.last_closep = data->last_closep;
6157 }
6158 else
6159 data_fake.last_closep = &fake;
6160 data_fake.pos_delta = delta;
6161 if ( flags & SCF_DO_STCLASS && !scan->flags
6162 && OP(scan) == IFMATCH ) { /* Lookahead */
6163 ssc_init(pRExC_state, &intrnl);
6164 data_fake.start_class = &intrnl;
6165 f |= SCF_DO_STCLASS_AND;
6166 }
6167 if (flags & SCF_WHILEM_VISITED_POS)
6168 f |= SCF_WHILEM_VISITED_POS;
6169 next = regnext(scan);
6170 nscan = NEXTOPER(NEXTOPER(scan));
6171
6172 /* recurse study_chunk() for lookahead body */
6173 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
6174 last, &data_fake, stopparen,
6175 recursed_depth, NULL, f, depth+1,
6176 mutate_ok);
6177 if (scan->flags) {
6178 if ( deltanext < 0
6179 || deltanext > (I32) U8_MAX
6180 || minnext > (I32)U8_MAX
6181 || minnext + deltanext > (I32)U8_MAX)
6182 {
6183 FAIL2("Lookbehind longer than %" UVuf " not implemented",
6184 (UV)U8_MAX);
6185 }
6186
6187 /* The 'next_off' field has been repurposed to count the
6188 * additional starting positions to try beyond the initial
6189 * one. (This leaves it at 0 for non-variable length
6190 * matches to avoid breakage for those not using this
6191 * extension) */
6192 if (deltanext) {
6193 scan->next_off = deltanext;
6194 ckWARNexperimental(RExC_parse,
6195 WARN_EXPERIMENTAL__VLB,
6196 "Variable length lookbehind is experimental");
6197 }
6198 scan->flags = (U8)minnext + deltanext;
6199 }
6200 if (data) {
6201 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6202 pars++;
6203 if (data_fake.flags & SF_HAS_EVAL)
6204 data->flags |= SF_HAS_EVAL;
6205 data->whilem_c = data_fake.whilem_c;
6206 }
6207 if (f & SCF_DO_STCLASS_AND) {
6208 if (flags & SCF_DO_STCLASS_OR) {
6209 /* OR before, AND after: ideally we would recurse with
6210 * data_fake to get the AND applied by study of the
6211 * remainder of the pattern, and then derecurse;
6212 * *** HACK *** for now just treat as "no information".
6213 * See [perl #56690].
6214 */
6215 ssc_init(pRExC_state, data->start_class);
6216 } else {
6217 /* AND before and after: combine and continue. These
6218 * assertions are zero-length, so can match an EMPTY
6219 * string */
6220 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
6221 ANYOF_FLAGS(data->start_class)
6222 |= SSC_MATCHES_EMPTY_STRING;
6223 }
6224 }
6225 }
6226 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
6227 else {
6228 /* Positive Lookahead/lookbehind
6229 In this case we can do fixed string optimisation,
6230 but we must be careful about it. Note in the case of
6231 lookbehind the positions will be offset by the minimum
6232 length of the pattern, something we won't know about
6233 until after the recurse.
6234 */
6235 SSize_t deltanext, fake = 0;
6236 regnode *nscan;
6237 regnode_ssc intrnl;
6238 int f = 0;
6239 /* We use SAVEFREEPV so that when the full compile
6240 is finished perl will clean up the allocated
6241 minlens when it's all done. This way we don't
6242 have to worry about freeing them when we know
6243 they wont be used, which would be a pain.
6244 */
6245 SSize_t *minnextp;
6246 Newx( minnextp, 1, SSize_t );
6247 SAVEFREEPV(minnextp);
6248
6249 if (data) {
6250 StructCopy(data, &data_fake, scan_data_t);
6251 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
6252 f |= SCF_DO_SUBSTR;
6253 if (scan->flags)
6254 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
6255 data_fake.last_found=newSVsv(data->last_found);
6256 }
6257 }
6258 else
6259 data_fake.last_closep = &fake;
6260 data_fake.flags = 0;
6261 data_fake.substrs[0].flags = 0;
6262 data_fake.substrs[1].flags = 0;
6263 data_fake.pos_delta = delta;
6264 if (is_inf)
6265 data_fake.flags |= SF_IS_INF;
6266 if ( flags & SCF_DO_STCLASS && !scan->flags
6267 && OP(scan) == IFMATCH ) { /* Lookahead */
6268 ssc_init(pRExC_state, &intrnl);
6269 data_fake.start_class = &intrnl;
6270 f |= SCF_DO_STCLASS_AND;
6271 }
6272 if (flags & SCF_WHILEM_VISITED_POS)
6273 f |= SCF_WHILEM_VISITED_POS;
6274 next = regnext(scan);
6275 nscan = NEXTOPER(NEXTOPER(scan));
6276
6277 /* positive lookahead study_chunk() recursion */
6278 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
6279 &deltanext, last, &data_fake,
6280 stopparen, recursed_depth, NULL,
6281 f, depth+1, mutate_ok);
6282 if (scan->flags) {
6283 assert(0); /* This code has never been tested since this
6284 is normally not compiled */
6285 if ( deltanext < 0
6286 || deltanext > (I32) U8_MAX
6287 || *minnextp > (I32)U8_MAX
6288 || *minnextp + deltanext > (I32)U8_MAX)
6289 {
6290 FAIL2("Lookbehind longer than %" UVuf " not implemented",
6291 (UV)U8_MAX);
6292 }
6293
6294 if (deltanext) {
6295 scan->next_off = deltanext;
6296 }
6297 scan->flags = (U8)*minnextp + deltanext;
6298 }
6299
6300 *minnextp += min;
6301
6302 if (f & SCF_DO_STCLASS_AND) {
6303 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
6304 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
6305 }
6306 if (data) {
6307 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6308 pars++;
6309 if (data_fake.flags & SF_HAS_EVAL)
6310 data->flags |= SF_HAS_EVAL;
6311 data->whilem_c = data_fake.whilem_c;
6312 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
6313 int i;
6314 if (RExC_rx->minlen<*minnextp)
6315 RExC_rx->minlen=*minnextp;
6316 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
6317 SvREFCNT_dec_NN(data_fake.last_found);
6318
6319 for (i = 0; i < 2; i++) {
6320 if (data_fake.substrs[i].minlenp != minlenp) {
6321 data->substrs[i].min_offset =
6322 data_fake.substrs[i].min_offset;
6323 data->substrs[i].max_offset =
6324 data_fake.substrs[i].max_offset;
6325 data->substrs[i].minlenp =
6326 data_fake.substrs[i].minlenp;
6327 data->substrs[i].lookbehind += scan->flags;
6328 }
6329 }
6330 }
6331 }
6332 }
6333 #endif
6334 }
6335 else if (OP(scan) == OPEN) {
6336 if (stopparen != (I32)ARG(scan))
6337 pars++;
6338 }
6339 else if (OP(scan) == CLOSE) {
6340 if (stopparen == (I32)ARG(scan)) {
6341 break;
6342 }
6343 if ((I32)ARG(scan) == is_par) {
6344 next = regnext(scan);
6345
6346 if ( next && (OP(next) != WHILEM) && next < last)
6347 is_par = 0; /* Disable optimization */
6348 }
6349 if (data)
6350 *(data->last_closep) = ARG(scan);
6351 }
6352 else if (OP(scan) == EVAL) {
6353 if (data)
6354 data->flags |= SF_HAS_EVAL;
6355 }
6356 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
6357 if (flags & SCF_DO_SUBSTR) {
6358 scan_commit(pRExC_state, data, minlenp, is_inf);
6359 flags &= ~SCF_DO_SUBSTR;
6360 }
6361 if (OP(scan)==ACCEPT) {
6362 /* m{(*ACCEPT)x} does not have to start with 'x' */
6363 flags &= ~SCF_DO_STCLASS;
6364 if (data) {
6365 data->flags |= SCF_SEEN_ACCEPT;
6366 if (stopmin > min)
6367 stopmin = min;
6368 }
6369 }
6370 }
6371 else if (OP(scan) == COMMIT) {
6372 /* gh18770: m{abc(*COMMIT)xyz} must fail on "abc abcxyz", so we
6373 * must not end up with "abcxyz" as a fixed substring else we'll
6374 * skip straight to attempting to match at offset 4.
6375 */
6376 if (flags & SCF_DO_SUBSTR) {
6377 scan_commit(pRExC_state, data, minlenp, is_inf);
6378 flags &= ~SCF_DO_SUBSTR;
6379 }
6380 }
6381 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
6382 {
6383 if (flags & SCF_DO_SUBSTR) {
6384 scan_commit(pRExC_state, data, minlenp, is_inf);
6385 data->cur_is_floating = 1; /* float */
6386 }
6387 is_inf = is_inf_internal = 1;
6388 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
6389 ssc_anything(data->start_class);
6390 flags &= ~SCF_DO_STCLASS;
6391 }
6392 else if (OP(scan) == GPOS) {
6393 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
6394 !(delta || is_inf || (data && data->pos_delta)))
6395 {
6396 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
6397 RExC_rx->intflags |= PREGf_ANCH_GPOS;
6398 if (RExC_rx->gofs < (STRLEN)min)
6399 RExC_rx->gofs = min;
6400 } else {
6401 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
6402 RExC_rx->gofs = 0;
6403 }
6404 }
6405 #ifdef TRIE_STUDY_OPT
6406 #ifdef FULL_TRIE_STUDY
6407 else if (PL_regkind[OP(scan)] == TRIE) {
6408 /* NOTE - There is similar code to this block above for handling
6409 BRANCH nodes on the initial study. If you change stuff here
6410 check there too. */
6411 regnode *trie_node= scan;
6412 regnode *tail= regnext(scan);
6413 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6414 SSize_t max1 = 0, min1 = OPTIMIZE_INFTY;
6415 regnode_ssc accum;
6416
6417 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
6418 /* Cannot merge strings after this. */
6419 scan_commit(pRExC_state, data, minlenp, is_inf);
6420 }
6421 if (flags & SCF_DO_STCLASS)
6422 ssc_init_zero(pRExC_state, &accum);
6423
6424 if (!trie->jump) {
6425 min1= trie->minlen;
6426 max1= trie->maxlen;
6427 } else {
6428 const regnode *nextbranch= NULL;
6429 U32 word;
6430
6431 for ( word=1 ; word <= trie->wordcount ; word++)
6432 {
6433 SSize_t deltanext=0, minnext=0, f = 0, fake;
6434 regnode_ssc this_class;
6435
6436 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
6437 if (data) {
6438 data_fake.whilem_c = data->whilem_c;
6439 data_fake.last_closep = data->last_closep;
6440 }
6441 else
6442 data_fake.last_closep = &fake;
6443 data_fake.pos_delta = delta;
6444 if (flags & SCF_DO_STCLASS) {
6445 ssc_init(pRExC_state, &this_class);
6446 data_fake.start_class = &this_class;
6447 f = SCF_DO_STCLASS_AND;
6448 }
6449 if (flags & SCF_WHILEM_VISITED_POS)
6450 f |= SCF_WHILEM_VISITED_POS;
6451
6452 if (trie->jump[word]) {
6453 if (!nextbranch)
6454 nextbranch = trie_node + trie->jump[0];
6455 scan= trie_node + trie->jump[word];
6456 /* We go from the jump point to the branch that follows
6457 it. Note this means we need the vestigal unused
6458 branches even though they arent otherwise used. */
6459 /* optimise study_chunk() for TRIE */
6460 minnext = study_chunk(pRExC_state, &scan, minlenp,
6461 &deltanext, (regnode *)nextbranch, &data_fake,
6462 stopparen, recursed_depth, NULL, f, depth+1,
6463 mutate_ok);
6464 }
6465 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
6466 nextbranch= regnext((regnode*)nextbranch);
6467
6468 if (min1 > (SSize_t)(minnext + trie->minlen))
6469 min1 = minnext + trie->minlen;
6470 if (deltanext == OPTIMIZE_INFTY) {
6471 is_inf = is_inf_internal = 1;
6472 max1 = OPTIMIZE_INFTY;
6473 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
6474 max1 = minnext + deltanext + trie->maxlen;
6475
6476 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6477 pars++;
6478 if (data_fake.flags & SCF_SEEN_ACCEPT) {
6479 if ( stopmin > min + min1)
6480 stopmin = min + min1;
6481 flags &= ~SCF_DO_SUBSTR;
6482 if (data)
6483 data->flags |= SCF_SEEN_ACCEPT;
6484 }
6485 if (data) {
6486 if (data_fake.flags & SF_HAS_EVAL)
6487 data->flags |= SF_HAS_EVAL;
6488 data->whilem_c = data_fake.whilem_c;
6489 }
6490 if (flags & SCF_DO_STCLASS)
6491 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
6492 }
6493 }
6494 if (flags & SCF_DO_SUBSTR) {
6495 data->pos_min += min1;
6496 data->pos_delta += max1 - min1;
6497 if (max1 != min1 || is_inf)
6498 data->cur_is_floating = 1; /* float */
6499 }
6500 min += min1;
6501 if (delta != OPTIMIZE_INFTY) {
6502 if (OPTIMIZE_INFTY - (max1 - min1) >= delta)
6503 delta += max1 - min1;
6504 else
6505 delta = OPTIMIZE_INFTY;
6506 }
6507 if (flags & SCF_DO_STCLASS_OR) {
6508 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6509 if (min1) {
6510 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6511 flags &= ~SCF_DO_STCLASS;
6512 }
6513 }
6514 else if (flags & SCF_DO_STCLASS_AND) {
6515 if (min1) {
6516 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6517 flags &= ~SCF_DO_STCLASS;
6518 }
6519 else {
6520 /* Switch to OR mode: cache the old value of
6521 * data->start_class */
6522 INIT_AND_WITHP;
6523 StructCopy(data->start_class, and_withp, regnode_ssc);
6524 flags &= ~SCF_DO_STCLASS_AND;
6525 StructCopy(&accum, data->start_class, regnode_ssc);
6526 flags |= SCF_DO_STCLASS_OR;
6527 }
6528 }
6529 scan= tail;
6530 continue;
6531 }
6532 #else
6533 else if (PL_regkind[OP(scan)] == TRIE) {
6534 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6535 U8*bang=NULL;
6536
6537 min += trie->minlen;
6538 delta += (trie->maxlen - trie->minlen);
6539 flags &= ~SCF_DO_STCLASS; /* xxx */
6540 if (flags & SCF_DO_SUBSTR) {
6541 /* Cannot expect anything... */
6542 scan_commit(pRExC_state, data, minlenp, is_inf);
6543 data->pos_min += trie->minlen;
6544 data->pos_delta += (trie->maxlen - trie->minlen);
6545 if (trie->maxlen != trie->minlen)
6546 data->cur_is_floating = 1; /* float */
6547 }
6548 if (trie->jump) /* no more substrings -- for now /grr*/
6549 flags &= ~SCF_DO_SUBSTR;
6550 }
6551
6552 #endif /* old or new */
6553 #endif /* TRIE_STUDY_OPT */
6554
6555 else if (OP(scan) == REGEX_SET) {
6556 Perl_croak(aTHX_ "panic: %s regnode should be resolved"
6557 " before optimization", PL_reg_name[REGEX_SET]);
6558 }
6559
6560 /* Else: zero-length, ignore. */
6561 scan = regnext(scan);
6562 }
6563
6564 finish:
6565 if (frame) {
6566 /* we need to unwind recursion. */
6567 depth = depth - 1;
6568
6569 DEBUG_STUDYDATA("frame-end", data, depth, is_inf);
6570 DEBUG_PEEP("fend", scan, depth, flags);
6571
6572 /* restore previous context */
6573 last = frame->last_regnode;
6574 scan = frame->next_regnode;
6575 stopparen = frame->stopparen;
6576 recursed_depth = frame->prev_recursed_depth;
6577
6578 RExC_frame_last = frame->prev_frame;
6579 frame = frame->this_prev_frame;
6580 goto fake_study_recurse;
6581 }
6582
6583 assert(!frame);
6584 DEBUG_STUDYDATA("pre-fin", data, depth, is_inf);
6585
6586 *scanp = scan;
6587 *deltap = is_inf_internal ? OPTIMIZE_INFTY : delta;
6588
6589 if (flags & SCF_DO_SUBSTR && is_inf)
6590 data->pos_delta = OPTIMIZE_INFTY - data->pos_min;
6591 if (is_par > (I32)U8_MAX)
6592 is_par = 0;
6593 if (is_par && pars==1 && data) {
6594 data->flags |= SF_IN_PAR;
6595 data->flags &= ~SF_HAS_PAR;
6596 }
6597 else if (pars && data) {
6598 data->flags |= SF_HAS_PAR;
6599 data->flags &= ~SF_IN_PAR;
6600 }
6601 if (flags & SCF_DO_STCLASS_OR)
6602 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6603 if (flags & SCF_TRIE_RESTUDY)
6604 data->flags |= SCF_TRIE_RESTUDY;
6605
6606 DEBUG_STUDYDATA("post-fin", data, depth, is_inf);
6607
6608 final_minlen = min < stopmin
6609 ? min : stopmin;
6610
6611 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
6612 if (final_minlen > OPTIMIZE_INFTY - delta)
6613 RExC_maxlen = OPTIMIZE_INFTY;
6614 else if (RExC_maxlen < final_minlen + delta)
6615 RExC_maxlen = final_minlen + delta;
6616 }
6617 return final_minlen;
6618 }
6619
6620 STATIC U32
S_add_data(RExC_state_t * const pRExC_state,const char * const s,const U32 n)6621 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
6622 {
6623 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
6624
6625 PERL_ARGS_ASSERT_ADD_DATA;
6626
6627 Renewc(RExC_rxi->data,
6628 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
6629 char, struct reg_data);
6630 if(count)
6631 Renew(RExC_rxi->data->what, count + n, U8);
6632 else
6633 Newx(RExC_rxi->data->what, n, U8);
6634 RExC_rxi->data->count = count + n;
6635 Copy(s, RExC_rxi->data->what + count, n, U8);
6636 return count;
6637 }
6638
6639 /*XXX: todo make this not included in a non debugging perl, but appears to be
6640 * used anyway there, in 'use re' */
6641 #ifndef PERL_IN_XSUB_RE
6642 void
Perl_reginitcolors(pTHX)6643 Perl_reginitcolors(pTHX)
6644 {
6645 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6646 if (s) {
6647 char *t = savepv(s);
6648 int i = 0;
6649 PL_colors[0] = t;
6650 while (++i < 6) {
6651 t = strchr(t, '\t');
6652 if (t) {
6653 *t = '\0';
6654 PL_colors[i] = ++t;
6655 }
6656 else
6657 PL_colors[i] = t = (char *)"";
6658 }
6659 } else {
6660 int i = 0;
6661 while (i < 6)
6662 PL_colors[i++] = (char *)"";
6663 }
6664 PL_colorset = 1;
6665 }
6666 #endif
6667
6668
6669 #ifdef TRIE_STUDY_OPT
6670 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6671 STMT_START { \
6672 if ( \
6673 (data.flags & SCF_TRIE_RESTUDY) \
6674 && ! restudied++ \
6675 ) { \
6676 dOsomething; \
6677 goto reStudy; \
6678 } \
6679 } STMT_END
6680 #else
6681 #define CHECK_RESTUDY_GOTO_butfirst
6682 #endif
6683
6684 /*
6685 * pregcomp - compile a regular expression into internal code
6686 *
6687 * Decides which engine's compiler to call based on the hint currently in
6688 * scope
6689 */
6690
6691 #ifndef PERL_IN_XSUB_RE
6692
6693 /* return the currently in-scope regex engine (or the default if none) */
6694
6695 regexp_engine const *
Perl_current_re_engine(pTHX)6696 Perl_current_re_engine(pTHX)
6697 {
6698 if (IN_PERL_COMPILETIME) {
6699 HV * const table = GvHV(PL_hintgv);
6700 SV **ptr;
6701
6702 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6703 return &PL_core_reg_engine;
6704 ptr = hv_fetchs(table, "regcomp", FALSE);
6705 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6706 return &PL_core_reg_engine;
6707 return INT2PTR(regexp_engine*, SvIV(*ptr));
6708 }
6709 else {
6710 SV *ptr;
6711 if (!PL_curcop->cop_hints_hash)
6712 return &PL_core_reg_engine;
6713 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6714 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6715 return &PL_core_reg_engine;
6716 return INT2PTR(regexp_engine*, SvIV(ptr));
6717 }
6718 }
6719
6720
6721 REGEXP *
Perl_pregcomp(pTHX_ SV * const pattern,const U32 flags)6722 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6723 {
6724 regexp_engine const *eng = current_re_engine();
6725 DECLARE_AND_GET_RE_DEBUG_FLAGS;
6726
6727 PERL_ARGS_ASSERT_PREGCOMP;
6728
6729 /* Dispatch a request to compile a regexp to correct regexp engine. */
6730 DEBUG_COMPILE_r({
6731 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6732 PTR2UV(eng));
6733 });
6734 return CALLREGCOMP_ENG(eng, pattern, flags);
6735 }
6736 #endif
6737
6738 /* public(ish) entry point for the perl core's own regex compiling code.
6739 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6740 * pattern rather than a list of OPs, and uses the internal engine rather
6741 * than the current one */
6742
6743 REGEXP *
Perl_re_compile(pTHX_ SV * const pattern,U32 rx_flags)6744 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6745 {
6746 SV *pat = pattern; /* defeat constness! */
6747
6748 PERL_ARGS_ASSERT_RE_COMPILE;
6749
6750 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6751 #ifdef PERL_IN_XSUB_RE
6752 &my_reg_engine,
6753 #else
6754 &PL_core_reg_engine,
6755 #endif
6756 NULL, NULL, rx_flags, 0);
6757 }
6758
6759 static void
S_free_codeblocks(pTHX_ struct reg_code_blocks * cbs)6760 S_free_codeblocks(pTHX_ struct reg_code_blocks *cbs)
6761 {
6762 int n;
6763
6764 if (--cbs->refcnt > 0)
6765 return;
6766 for (n = 0; n < cbs->count; n++) {
6767 REGEXP *rx = cbs->cb[n].src_regex;
6768 if (rx) {
6769 cbs->cb[n].src_regex = NULL;
6770 SvREFCNT_dec_NN(rx);
6771 }
6772 }
6773 Safefree(cbs->cb);
6774 Safefree(cbs);
6775 }
6776
6777
6778 static struct reg_code_blocks *
S_alloc_code_blocks(pTHX_ int ncode)6779 S_alloc_code_blocks(pTHX_ int ncode)
6780 {
6781 struct reg_code_blocks *cbs;
6782 Newx(cbs, 1, struct reg_code_blocks);
6783 cbs->count = ncode;
6784 cbs->refcnt = 1;
6785 SAVEDESTRUCTOR_X(S_free_codeblocks, cbs);
6786 if (ncode)
6787 Newx(cbs->cb, ncode, struct reg_code_block);
6788 else
6789 cbs->cb = NULL;
6790 return cbs;
6791 }
6792
6793
6794 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6795 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6796 * point to the realloced string and length.
6797 *
6798 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6799 * stuff added */
6800
6801 static void
S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,char ** pat_p,STRLEN * plen_p,int num_code_blocks)6802 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6803 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6804 {
6805 U8 *const src = (U8*)*pat_p;
6806 U8 *dst, *d;
6807 int n=0;
6808 STRLEN s = 0;
6809 bool do_end = 0;
6810 DECLARE_AND_GET_RE_DEBUG_FLAGS;
6811
6812 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6813 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6814
6815 /* 1 for each byte + 1 for each byte that expands to two, + trailing NUL */
6816 Newx(dst, *plen_p + variant_under_utf8_count(src, src + *plen_p) + 1, U8);
6817 d = dst;
6818
6819 while (s < *plen_p) {
6820 append_utf8_from_native_byte(src[s], &d);
6821
6822 if (n < num_code_blocks) {
6823 assert(pRExC_state->code_blocks);
6824 if (!do_end && pRExC_state->code_blocks->cb[n].start == s) {
6825 pRExC_state->code_blocks->cb[n].start = d - dst - 1;
6826 assert(*(d - 1) == '(');
6827 do_end = 1;
6828 }
6829 else if (do_end && pRExC_state->code_blocks->cb[n].end == s) {
6830 pRExC_state->code_blocks->cb[n].end = d - dst - 1;
6831 assert(*(d - 1) == ')');
6832 do_end = 0;
6833 n++;
6834 }
6835 }
6836 s++;
6837 }
6838 *d = '\0';
6839 *plen_p = d - dst;
6840 *pat_p = (char*) dst;
6841 SAVEFREEPV(*pat_p);
6842 RExC_orig_utf8 = RExC_utf8 = 1;
6843 }
6844
6845
6846
6847 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6848 * while recording any code block indices, and handling overloading,
6849 * nested qr// objects etc. If pat is null, it will allocate a new
6850 * string, or just return the first arg, if there's only one.
6851 *
6852 * Returns the malloced/updated pat.
6853 * patternp and pat_count is the array of SVs to be concatted;
6854 * oplist is the optional list of ops that generated the SVs;
6855 * recompile_p is a pointer to a boolean that will be set if
6856 * the regex will need to be recompiled.
6857 * delim, if non-null is an SV that will be inserted between each element
6858 */
6859
6860 static SV*
S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,SV * pat,SV ** const patternp,int pat_count,OP * oplist,bool * recompile_p,SV * delim)6861 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6862 SV *pat, SV ** const patternp, int pat_count,
6863 OP *oplist, bool *recompile_p, SV *delim)
6864 {
6865 SV **svp;
6866 int n = 0;
6867 bool use_delim = FALSE;
6868 bool alloced = FALSE;
6869
6870 /* if we know we have at least two args, create an empty string,
6871 * then concatenate args to that. For no args, return an empty string */
6872 if (!pat && pat_count != 1) {
6873 pat = newSVpvs("");
6874 SAVEFREESV(pat);
6875 alloced = TRUE;
6876 }
6877
6878 for (svp = patternp; svp < patternp + pat_count; svp++) {
6879 SV *sv;
6880 SV *rx = NULL;
6881 STRLEN orig_patlen = 0;
6882 bool code = 0;
6883 SV *msv = use_delim ? delim : *svp;
6884 if (!msv) msv = &PL_sv_undef;
6885
6886 /* if we've got a delimiter, we go round the loop twice for each
6887 * svp slot (except the last), using the delimiter the second
6888 * time round */
6889 if (use_delim) {
6890 svp--;
6891 use_delim = FALSE;
6892 }
6893 else if (delim)
6894 use_delim = TRUE;
6895
6896 if (SvTYPE(msv) == SVt_PVAV) {
6897 /* we've encountered an interpolated array within
6898 * the pattern, e.g. /...@a..../. Expand the list of elements,
6899 * then recursively append elements.
6900 * The code in this block is based on S_pushav() */
6901
6902 AV *const av = (AV*)msv;
6903 const SSize_t maxarg = AvFILL(av) + 1;
6904 SV **array;
6905
6906 if (oplist) {
6907 assert(oplist->op_type == OP_PADAV
6908 || oplist->op_type == OP_RV2AV);
6909 oplist = OpSIBLING(oplist);
6910 }
6911
6912 if (SvRMAGICAL(av)) {
6913 SSize_t i;
6914
6915 Newx(array, maxarg, SV*);
6916 SAVEFREEPV(array);
6917 for (i=0; i < maxarg; i++) {
6918 SV ** const svp = av_fetch(av, i, FALSE);
6919 array[i] = svp ? *svp : &PL_sv_undef;
6920 }
6921 }
6922 else
6923 array = AvARRAY(av);
6924
6925 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6926 array, maxarg, NULL, recompile_p,
6927 /* $" */
6928 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6929
6930 continue;
6931 }
6932
6933
6934 /* we make the assumption here that each op in the list of
6935 * op_siblings maps to one SV pushed onto the stack,
6936 * except for code blocks, with have both an OP_NULL and
6937 * an OP_CONST.
6938 * This allows us to match up the list of SVs against the
6939 * list of OPs to find the next code block.
6940 *
6941 * Note that PUSHMARK PADSV PADSV ..
6942 * is optimised to
6943 * PADRANGE PADSV PADSV ..
6944 * so the alignment still works. */
6945
6946 if (oplist) {
6947 if (oplist->op_type == OP_NULL
6948 && (oplist->op_flags & OPf_SPECIAL))
6949 {
6950 assert(n < pRExC_state->code_blocks->count);
6951 pRExC_state->code_blocks->cb[n].start = pat ? SvCUR(pat) : 0;
6952 pRExC_state->code_blocks->cb[n].block = oplist;
6953 pRExC_state->code_blocks->cb[n].src_regex = NULL;
6954 n++;
6955 code = 1;
6956 oplist = OpSIBLING(oplist); /* skip CONST */
6957 assert(oplist);
6958 }
6959 oplist = OpSIBLING(oplist);;
6960 }
6961
6962 /* apply magic and QR overloading to arg */
6963
6964 SvGETMAGIC(msv);
6965 if (SvROK(msv) && SvAMAGIC(msv)) {
6966 SV *sv = AMG_CALLunary(msv, regexp_amg);
6967 if (sv) {
6968 if (SvROK(sv))
6969 sv = SvRV(sv);
6970 if (SvTYPE(sv) != SVt_REGEXP)
6971 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6972 msv = sv;
6973 }
6974 }
6975
6976 /* try concatenation overload ... */
6977 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6978 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6979 {
6980 sv_setsv(pat, sv);
6981 /* overloading involved: all bets are off over literal
6982 * code. Pretend we haven't seen it */
6983 if (n)
6984 pRExC_state->code_blocks->count -= n;
6985 n = 0;
6986 }
6987 else {
6988 /* ... or failing that, try "" overload */
6989 while (SvAMAGIC(msv)
6990 && (sv = AMG_CALLunary(msv, string_amg))
6991 && sv != msv
6992 && !( SvROK(msv)
6993 && SvROK(sv)
6994 && SvRV(msv) == SvRV(sv))
6995 ) {
6996 msv = sv;
6997 SvGETMAGIC(msv);
6998 }
6999 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
7000 msv = SvRV(msv);
7001
7002 if (pat) {
7003 /* this is a partially unrolled
7004 * sv_catsv_nomg(pat, msv);
7005 * that allows us to adjust code block indices if
7006 * needed */
7007 STRLEN dlen;
7008 char *dst = SvPV_force_nomg(pat, dlen);
7009 orig_patlen = dlen;
7010 if (SvUTF8(msv) && !SvUTF8(pat)) {
7011 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
7012 sv_setpvn(pat, dst, dlen);
7013 SvUTF8_on(pat);
7014 }
7015 sv_catsv_nomg(pat, msv);
7016 rx = msv;
7017 }
7018 else {
7019 /* We have only one SV to process, but we need to verify
7020 * it is properly null terminated or we will fail asserts
7021 * later. In theory we probably shouldn't get such SV's,
7022 * but if we do we should handle it gracefully. */
7023 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) || SvIsCOW_shared_hash(msv) ) {
7024 /* not a string, or a string with a trailing null */
7025 pat = msv;
7026 } else {
7027 /* a string with no trailing null, we need to copy it
7028 * so it has a trailing null */
7029 pat = sv_2mortal(newSVsv(msv));
7030 }
7031 }
7032
7033 if (code)
7034 pRExC_state->code_blocks->cb[n-1].end = SvCUR(pat)-1;
7035 }
7036
7037 /* extract any code blocks within any embedded qr//'s */
7038 if (rx && SvTYPE(rx) == SVt_REGEXP
7039 && RX_ENGINE((REGEXP*)rx)->op_comp)
7040 {
7041
7042 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
7043 if (ri->code_blocks && ri->code_blocks->count) {
7044 int i;
7045 /* the presence of an embedded qr// with code means
7046 * we should always recompile: the text of the
7047 * qr// may not have changed, but it may be a
7048 * different closure than last time */
7049 *recompile_p = 1;
7050 if (pRExC_state->code_blocks) {
7051 int new_count = pRExC_state->code_blocks->count
7052 + ri->code_blocks->count;
7053 Renew(pRExC_state->code_blocks->cb,
7054 new_count, struct reg_code_block);
7055 pRExC_state->code_blocks->count = new_count;
7056 }
7057 else
7058 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_
7059 ri->code_blocks->count);
7060
7061 for (i=0; i < ri->code_blocks->count; i++) {
7062 struct reg_code_block *src, *dst;
7063 STRLEN offset = orig_patlen
7064 + ReANY((REGEXP *)rx)->pre_prefix;
7065 assert(n < pRExC_state->code_blocks->count);
7066 src = &ri->code_blocks->cb[i];
7067 dst = &pRExC_state->code_blocks->cb[n];
7068 dst->start = src->start + offset;
7069 dst->end = src->end + offset;
7070 dst->block = src->block;
7071 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
7072 src->src_regex
7073 ? src->src_regex
7074 : (REGEXP*)rx);
7075 n++;
7076 }
7077 }
7078 }
7079 }
7080 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
7081 if (alloced)
7082 SvSETMAGIC(pat);
7083
7084 return pat;
7085 }
7086
7087
7088
7089 /* see if there are any run-time code blocks in the pattern.
7090 * False positives are allowed */
7091
7092 static bool
S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,char * pat,STRLEN plen)7093 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
7094 char *pat, STRLEN plen)
7095 {
7096 int n = 0;
7097 STRLEN s;
7098
7099 PERL_UNUSED_CONTEXT;
7100
7101 for (s = 0; s < plen; s++) {
7102 if ( pRExC_state->code_blocks
7103 && n < pRExC_state->code_blocks->count
7104 && s == pRExC_state->code_blocks->cb[n].start)
7105 {
7106 s = pRExC_state->code_blocks->cb[n].end;
7107 n++;
7108 continue;
7109 }
7110 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
7111 * positives here */
7112 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
7113 (pat[s+2] == '{'
7114 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
7115 )
7116 return 1;
7117 }
7118 return 0;
7119 }
7120
7121 /* Handle run-time code blocks. We will already have compiled any direct
7122 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
7123 * copy of it, but with any literal code blocks blanked out and
7124 * appropriate chars escaped; then feed it into
7125 *
7126 * eval "qr'modified_pattern'"
7127 *
7128 * For example,
7129 *
7130 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
7131 *
7132 * becomes
7133 *
7134 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
7135 *
7136 * After eval_sv()-ing that, grab any new code blocks from the returned qr
7137 * and merge them with any code blocks of the original regexp.
7138 *
7139 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
7140 * instead, just save the qr and return FALSE; this tells our caller that
7141 * the original pattern needs upgrading to utf8.
7142 */
7143
7144 static bool
S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,char * pat,STRLEN plen)7145 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
7146 char *pat, STRLEN plen)
7147 {
7148 SV *qr;
7149
7150 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7151
7152 if (pRExC_state->runtime_code_qr) {
7153 /* this is the second time we've been called; this should
7154 * only happen if the main pattern got upgraded to utf8
7155 * during compilation; re-use the qr we compiled first time
7156 * round (which should be utf8 too)
7157 */
7158 qr = pRExC_state->runtime_code_qr;
7159 pRExC_state->runtime_code_qr = NULL;
7160 assert(RExC_utf8 && SvUTF8(qr));
7161 }
7162 else {
7163 int n = 0;
7164 STRLEN s;
7165 char *p, *newpat;
7166 int newlen = plen + 7; /* allow for "qr''xx\0" extra chars */
7167 SV *sv, *qr_ref;
7168 dSP;
7169
7170 /* determine how many extra chars we need for ' and \ escaping */
7171 for (s = 0; s < plen; s++) {
7172 if (pat[s] == '\'' || pat[s] == '\\')
7173 newlen++;
7174 }
7175
7176 Newx(newpat, newlen, char);
7177 p = newpat;
7178 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
7179
7180 for (s = 0; s < plen; s++) {
7181 if ( pRExC_state->code_blocks
7182 && n < pRExC_state->code_blocks->count
7183 && s == pRExC_state->code_blocks->cb[n].start)
7184 {
7185 /* blank out literal code block so that they aren't
7186 * recompiled: eg change from/to:
7187 * /(?{xyz})/
7188 * /(?=====)/
7189 * and
7190 * /(??{xyz})/
7191 * /(?======)/
7192 * and
7193 * /(?(?{xyz}))/
7194 * /(?(?=====))/
7195 */
7196 assert(pat[s] == '(');
7197 assert(pat[s+1] == '?');
7198 *p++ = '(';
7199 *p++ = '?';
7200 s += 2;
7201 while (s < pRExC_state->code_blocks->cb[n].end) {
7202 *p++ = '=';
7203 s++;
7204 }
7205 *p++ = ')';
7206 n++;
7207 continue;
7208 }
7209 if (pat[s] == '\'' || pat[s] == '\\')
7210 *p++ = '\\';
7211 *p++ = pat[s];
7212 }
7213 *p++ = '\'';
7214 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
7215 *p++ = 'x';
7216 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
7217 *p++ = 'x';
7218 }
7219 }
7220 *p++ = '\0';
7221 DEBUG_COMPILE_r({
7222 Perl_re_printf( aTHX_
7223 "%sre-parsing pattern for runtime code:%s %s\n",
7224 PL_colors[4], PL_colors[5], newpat);
7225 });
7226
7227 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
7228 Safefree(newpat);
7229
7230 ENTER;
7231 SAVETMPS;
7232 save_re_context();
7233 PUSHSTACKi(PERLSI_REQUIRE);
7234 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
7235 * parsing qr''; normally only q'' does this. It also alters
7236 * hints handling */
7237 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
7238 SvREFCNT_dec_NN(sv);
7239 SPAGAIN;
7240 qr_ref = POPs;
7241 PUTBACK;
7242 {
7243 SV * const errsv = ERRSV;
7244 if (SvTRUE_NN(errsv))
7245 /* use croak_sv ? */
7246 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
7247 }
7248 assert(SvROK(qr_ref));
7249 qr = SvRV(qr_ref);
7250 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
7251 /* the leaving below frees the tmp qr_ref.
7252 * Give qr a life of its own */
7253 SvREFCNT_inc(qr);
7254 POPSTACK;
7255 FREETMPS;
7256 LEAVE;
7257
7258 }
7259
7260 if (!RExC_utf8 && SvUTF8(qr)) {
7261 /* first time through; the pattern got upgraded; save the
7262 * qr for the next time through */
7263 assert(!pRExC_state->runtime_code_qr);
7264 pRExC_state->runtime_code_qr = qr;
7265 return 0;
7266 }
7267
7268
7269 /* extract any code blocks within the returned qr// */
7270
7271
7272 /* merge the main (r1) and run-time (r2) code blocks into one */
7273 {
7274 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
7275 struct reg_code_block *new_block, *dst;
7276 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
7277 int i1 = 0, i2 = 0;
7278 int r1c, r2c;
7279
7280 if (!r2->code_blocks || !r2->code_blocks->count) /* we guessed wrong */
7281 {
7282 SvREFCNT_dec_NN(qr);
7283 return 1;
7284 }
7285
7286 if (!r1->code_blocks)
7287 r1->code_blocks = S_alloc_code_blocks(aTHX_ 0);
7288
7289 r1c = r1->code_blocks->count;
7290 r2c = r2->code_blocks->count;
7291
7292 Newx(new_block, r1c + r2c, struct reg_code_block);
7293
7294 dst = new_block;
7295
7296 while (i1 < r1c || i2 < r2c) {
7297 struct reg_code_block *src;
7298 bool is_qr = 0;
7299
7300 if (i1 == r1c) {
7301 src = &r2->code_blocks->cb[i2++];
7302 is_qr = 1;
7303 }
7304 else if (i2 == r2c)
7305 src = &r1->code_blocks->cb[i1++];
7306 else if ( r1->code_blocks->cb[i1].start
7307 < r2->code_blocks->cb[i2].start)
7308 {
7309 src = &r1->code_blocks->cb[i1++];
7310 assert(src->end < r2->code_blocks->cb[i2].start);
7311 }
7312 else {
7313 assert( r1->code_blocks->cb[i1].start
7314 > r2->code_blocks->cb[i2].start);
7315 src = &r2->code_blocks->cb[i2++];
7316 is_qr = 1;
7317 assert(src->end < r1->code_blocks->cb[i1].start);
7318 }
7319
7320 assert(pat[src->start] == '(');
7321 assert(pat[src->end] == ')');
7322 dst->start = src->start;
7323 dst->end = src->end;
7324 dst->block = src->block;
7325 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
7326 : src->src_regex;
7327 dst++;
7328 }
7329 r1->code_blocks->count += r2c;
7330 Safefree(r1->code_blocks->cb);
7331 r1->code_blocks->cb = new_block;
7332 }
7333
7334 SvREFCNT_dec_NN(qr);
7335 return 1;
7336 }
7337
7338
7339 STATIC bool
S_setup_longest(pTHX_ RExC_state_t * pRExC_state,struct reg_substr_datum * rsd,struct scan_data_substrs * sub,STRLEN longest_length)7340 S_setup_longest(pTHX_ RExC_state_t *pRExC_state,
7341 struct reg_substr_datum *rsd,
7342 struct scan_data_substrs *sub,
7343 STRLEN longest_length)
7344 {
7345 /* This is the common code for setting up the floating and fixed length
7346 * string data extracted from Perl_re_op_compile() below. Returns a boolean
7347 * as to whether succeeded or not */
7348
7349 I32 t;
7350 SSize_t ml;
7351 bool eol = cBOOL(sub->flags & SF_BEFORE_EOL);
7352 bool meol = cBOOL(sub->flags & SF_BEFORE_MEOL);
7353
7354 if (! (longest_length
7355 || (eol /* Can't have SEOL and MULTI */
7356 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
7357 )
7358 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
7359 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
7360 {
7361 return FALSE;
7362 }
7363
7364 /* copy the information about the longest from the reg_scan_data
7365 over to the program. */
7366 if (SvUTF8(sub->str)) {
7367 rsd->substr = NULL;
7368 rsd->utf8_substr = sub->str;
7369 } else {
7370 rsd->substr = sub->str;
7371 rsd->utf8_substr = NULL;
7372 }
7373 /* end_shift is how many chars that must be matched that
7374 follow this item. We calculate it ahead of time as once the
7375 lookbehind offset is added in we lose the ability to correctly
7376 calculate it.*/
7377 ml = sub->minlenp ? *(sub->minlenp) : (SSize_t)longest_length;
7378 rsd->end_shift = ml - sub->min_offset
7379 - longest_length
7380 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
7381 * intead? - DAPM
7382 + (SvTAIL(sub->str) != 0)
7383 */
7384 + sub->lookbehind;
7385
7386 t = (eol/* Can't have SEOL and MULTI */
7387 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
7388 fbm_compile(sub->str, t ? FBMcf_TAIL : 0);
7389
7390 return TRUE;
7391 }
7392
7393 STATIC void
S_set_regex_pv(pTHX_ RExC_state_t * pRExC_state,REGEXP * Rx)7394 S_set_regex_pv(pTHX_ RExC_state_t *pRExC_state, REGEXP *Rx)
7395 {
7396 /* Calculates and sets in the compiled pattern 'Rx' the string to compile,
7397 * properly wrapped with the right modifiers */
7398
7399 bool has_p = ((RExC_rx->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7400 bool has_charset = RExC_utf8 || (get_regex_charset(RExC_rx->extflags)
7401 != REGEX_DEPENDS_CHARSET);
7402
7403 /* The caret is output if there are any defaults: if not all the STD
7404 * flags are set, or if no character set specifier is needed */
7405 bool has_default =
7406 (((RExC_rx->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7407 || ! has_charset);
7408 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7409 == REG_RUN_ON_COMMENT_SEEN);
7410 U8 reganch = (U8)((RExC_rx->extflags & RXf_PMf_STD_PMMOD)
7411 >> RXf_PMf_STD_PMMOD_SHIFT);
7412 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7413 char *p;
7414 STRLEN pat_len = RExC_precomp_end - RExC_precomp;
7415
7416 /* We output all the necessary flags; we never output a minus, as all
7417 * those are defaults, so are
7418 * covered by the caret */
7419 const STRLEN wraplen = pat_len + has_p + has_runon
7420 + has_default /* If needs a caret */
7421 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7422
7423 /* If needs a character set specifier */
7424 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7425 + (sizeof("(?:)") - 1);
7426
7427 PERL_ARGS_ASSERT_SET_REGEX_PV;
7428
7429 /* make sure PL_bitcount bounds not exceeded */
7430 STATIC_ASSERT_STMT(sizeof(STD_PAT_MODS) <= 8);
7431
7432 p = sv_grow(MUTABLE_SV(Rx), wraplen + 1); /* +1 for the ending NUL */
7433 SvPOK_on(Rx);
7434 if (RExC_utf8)
7435 SvFLAGS(Rx) |= SVf_UTF8;
7436 *p++='('; *p++='?';
7437
7438 /* If a default, cover it using the caret */
7439 if (has_default) {
7440 *p++= DEFAULT_PAT_MOD;
7441 }
7442 if (has_charset) {
7443 STRLEN len;
7444 const char* name;
7445
7446 name = get_regex_charset_name(RExC_rx->extflags, &len);
7447 if (strEQ(name, DEPENDS_PAT_MODS)) { /* /d under UTF-8 => /u */
7448 assert(RExC_utf8);
7449 name = UNICODE_PAT_MODS;
7450 len = sizeof(UNICODE_PAT_MODS) - 1;
7451 }
7452 Copy(name, p, len, char);
7453 p += len;
7454 }
7455 if (has_p)
7456 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7457 {
7458 char ch;
7459 while((ch = *fptr++)) {
7460 if(reganch & 1)
7461 *p++ = ch;
7462 reganch >>= 1;
7463 }
7464 }
7465
7466 *p++ = ':';
7467 Copy(RExC_precomp, p, pat_len, char);
7468 assert ((RX_WRAPPED(Rx) - p) < 16);
7469 RExC_rx->pre_prefix = p - RX_WRAPPED(Rx);
7470 p += pat_len;
7471
7472 /* Adding a trailing \n causes this to compile properly:
7473 my $R = qr / A B C # D E/x; /($R)/
7474 Otherwise the parens are considered part of the comment */
7475 if (has_runon)
7476 *p++ = '\n';
7477 *p++ = ')';
7478 *p = 0;
7479 SvCUR_set(Rx, p - RX_WRAPPED(Rx));
7480 }
7481
7482 /*
7483 * Perl_re_op_compile - the perl internal RE engine's function to compile a
7484 * regular expression into internal code.
7485 * The pattern may be passed either as:
7486 * a list of SVs (patternp plus pat_count)
7487 * a list of OPs (expr)
7488 * If both are passed, the SV list is used, but the OP list indicates
7489 * which SVs are actually pre-compiled code blocks
7490 *
7491 * The SVs in the list have magic and qr overloading applied to them (and
7492 * the list may be modified in-place with replacement SVs in the latter
7493 * case).
7494 *
7495 * If the pattern hasn't changed from old_re, then old_re will be
7496 * returned.
7497 *
7498 * eng is the current engine. If that engine has an op_comp method, then
7499 * handle directly (i.e. we assume that op_comp was us); otherwise, just
7500 * do the initial concatenation of arguments and pass on to the external
7501 * engine.
7502 *
7503 * If is_bare_re is not null, set it to a boolean indicating whether the
7504 * arg list reduced (after overloading) to a single bare regex which has
7505 * been returned (i.e. /$qr/).
7506 *
7507 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
7508 *
7509 * pm_flags contains the PMf_* flags, typically based on those from the
7510 * pm_flags field of the related PMOP. Currently we're only interested in
7511 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL, PMf_WILDCARD.
7512 *
7513 * For many years this code had an initial sizing pass that calculated
7514 * (sometimes incorrectly, leading to security holes) the size needed for the
7515 * compiled pattern. That was changed by commit
7516 * 7c932d07cab18751bfc7515b4320436273a459e2 in 5.29, which reallocs the size, a
7517 * node at a time, as parsing goes along. Patches welcome to fix any obsolete
7518 * references to this sizing pass.
7519 *
7520 * Now, an initial crude guess as to the size needed is made, based on the
7521 * length of the pattern. Patches welcome to improve that guess. That amount
7522 * of space is malloc'd and then immediately freed, and then clawed back node
7523 * by node. This design is to minimze, to the extent possible, memory churn
7524 * when doing the reallocs.
7525 *
7526 * A separate parentheses counting pass may be needed in some cases.
7527 * (Previously the sizing pass did this.) Patches welcome to reduce the number
7528 * of these cases.
7529 *
7530 * The existence of a sizing pass necessitated design decisions that are no
7531 * longer needed. There are potential areas of simplification.
7532 *
7533 * Beware that the optimization-preparation code in here knows about some
7534 * of the structure of the compiled regexp. [I'll say.]
7535 */
7536
7537 REGEXP *
Perl_re_op_compile(pTHX_ SV ** const patternp,int pat_count,OP * expr,const regexp_engine * eng,REGEXP * old_re,bool * is_bare_re,const U32 orig_rx_flags,const U32 pm_flags)7538 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
7539 OP *expr, const regexp_engine* eng, REGEXP *old_re,
7540 bool *is_bare_re, const U32 orig_rx_flags, const U32 pm_flags)
7541 {
7542 REGEXP *Rx; /* Capital 'R' means points to a REGEXP */
7543 STRLEN plen;
7544 char *exp;
7545 regnode *scan;
7546 I32 flags;
7547 SSize_t minlen = 0;
7548 U32 rx_flags;
7549 SV *pat;
7550 SV** new_patternp = patternp;
7551
7552 /* these are all flags - maybe they should be turned
7553 * into a single int with different bit masks */
7554 I32 sawlookahead = 0;
7555 I32 sawplus = 0;
7556 I32 sawopen = 0;
7557 I32 sawminmod = 0;
7558
7559 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
7560 bool recompile = 0;
7561 bool runtime_code = 0;
7562 scan_data_t data;
7563 RExC_state_t RExC_state;
7564 RExC_state_t * const pRExC_state = &RExC_state;
7565 #ifdef TRIE_STUDY_OPT
7566 int restudied = 0;
7567 RExC_state_t copyRExC_state;
7568 #endif
7569 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7570
7571 PERL_ARGS_ASSERT_RE_OP_COMPILE;
7572
7573 DEBUG_r(if (!PL_colorset) reginitcolors());
7574
7575
7576 pRExC_state->warn_text = NULL;
7577 pRExC_state->unlexed_names = NULL;
7578 pRExC_state->code_blocks = NULL;
7579
7580 if (is_bare_re)
7581 *is_bare_re = FALSE;
7582
7583 if (expr && (expr->op_type == OP_LIST ||
7584 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
7585 /* allocate code_blocks if needed */
7586 OP *o;
7587 int ncode = 0;
7588
7589 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
7590 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
7591 ncode++; /* count of DO blocks */
7592
7593 if (ncode)
7594 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_ ncode);
7595 }
7596
7597 if (!pat_count) {
7598 /* compile-time pattern with just OP_CONSTs and DO blocks */
7599
7600 int n;
7601 OP *o;
7602
7603 /* find how many CONSTs there are */
7604 assert(expr);
7605 n = 0;
7606 if (expr->op_type == OP_CONST)
7607 n = 1;
7608 else
7609 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7610 if (o->op_type == OP_CONST)
7611 n++;
7612 }
7613
7614 /* fake up an SV array */
7615
7616 assert(!new_patternp);
7617 Newx(new_patternp, n, SV*);
7618 SAVEFREEPV(new_patternp);
7619 pat_count = n;
7620
7621 n = 0;
7622 if (expr->op_type == OP_CONST)
7623 new_patternp[n] = cSVOPx_sv(expr);
7624 else
7625 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7626 if (o->op_type == OP_CONST)
7627 new_patternp[n++] = cSVOPo_sv;
7628 }
7629
7630 }
7631
7632 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7633 "Assembling pattern from %d elements%s\n", pat_count,
7634 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7635
7636 /* set expr to the first arg op */
7637
7638 if (pRExC_state->code_blocks && pRExC_state->code_blocks->count
7639 && expr->op_type != OP_CONST)
7640 {
7641 expr = cLISTOPx(expr)->op_first;
7642 assert( expr->op_type == OP_PUSHMARK
7643 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
7644 || expr->op_type == OP_PADRANGE);
7645 expr = OpSIBLING(expr);
7646 }
7647
7648 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
7649 expr, &recompile, NULL);
7650
7651 /* handle bare (possibly after overloading) regex: foo =~ $re */
7652 {
7653 SV *re = pat;
7654 if (SvROK(re))
7655 re = SvRV(re);
7656 if (SvTYPE(re) == SVt_REGEXP) {
7657 if (is_bare_re)
7658 *is_bare_re = TRUE;
7659 SvREFCNT_inc(re);
7660 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7661 "Precompiled pattern%s\n",
7662 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7663
7664 return (REGEXP*)re;
7665 }
7666 }
7667
7668 exp = SvPV_nomg(pat, plen);
7669
7670 if (!eng->op_comp) {
7671 if ((SvUTF8(pat) && IN_BYTES)
7672 || SvGMAGICAL(pat) || SvAMAGIC(pat))
7673 {
7674 /* make a temporary copy; either to convert to bytes,
7675 * or to avoid repeating get-magic / overloaded stringify */
7676 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
7677 (IN_BYTES ? 0 : SvUTF8(pat)));
7678 }
7679 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
7680 }
7681
7682 /* ignore the utf8ness if the pattern is 0 length */
7683 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
7684 RExC_uni_semantics = 0;
7685 RExC_contains_locale = 0;
7686 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
7687 RExC_in_script_run = 0;
7688 RExC_study_started = 0;
7689 pRExC_state->runtime_code_qr = NULL;
7690 RExC_frame_head= NULL;
7691 RExC_frame_last= NULL;
7692 RExC_frame_count= 0;
7693 RExC_latest_warn_offset = 0;
7694 RExC_use_BRANCHJ = 0;
7695 RExC_warned_WARN_EXPERIMENTAL__VLB = 0;
7696 RExC_warned_WARN_EXPERIMENTAL__REGEX_SETS = 0;
7697 RExC_total_parens = 0;
7698 RExC_open_parens = NULL;
7699 RExC_close_parens = NULL;
7700 RExC_paren_names = NULL;
7701 RExC_size = 0;
7702 RExC_seen_d_op = FALSE;
7703 #ifdef DEBUGGING
7704 RExC_paren_name_list = NULL;
7705 #endif
7706
7707 DEBUG_r({
7708 RExC_mysv1= sv_newmortal();
7709 RExC_mysv2= sv_newmortal();
7710 });
7711
7712 DEBUG_COMPILE_r({
7713 SV *dsv= sv_newmortal();
7714 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7715 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
7716 PL_colors[4], PL_colors[5], s);
7717 });
7718
7719 /* we jump here if we have to recompile, e.g., from upgrading the pattern
7720 * to utf8 */
7721
7722 if ((pm_flags & PMf_USE_RE_EVAL)
7723 /* this second condition covers the non-regex literal case,
7724 * i.e. $foo =~ '(?{})'. */
7725 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
7726 )
7727 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
7728
7729 redo_parse:
7730 /* return old regex if pattern hasn't changed */
7731 /* XXX: note in the below we have to check the flags as well as the
7732 * pattern.
7733 *
7734 * Things get a touch tricky as we have to compare the utf8 flag
7735 * independently from the compile flags. */
7736
7737 if ( old_re
7738 && !recompile
7739 && !!RX_UTF8(old_re) == !!RExC_utf8
7740 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
7741 && RX_PRECOMP(old_re)
7742 && RX_PRELEN(old_re) == plen
7743 && memEQ(RX_PRECOMP(old_re), exp, plen)
7744 && !runtime_code /* with runtime code, always recompile */ )
7745 {
7746 DEBUG_COMPILE_r({
7747 SV *dsv= sv_newmortal();
7748 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7749 Perl_re_printf( aTHX_ "%sSkipping recompilation of unchanged REx%s %s\n",
7750 PL_colors[4], PL_colors[5], s);
7751 });
7752 return old_re;
7753 }
7754
7755 /* Allocate the pattern's SV */
7756 RExC_rx_sv = Rx = (REGEXP*) newSV_type(SVt_REGEXP);
7757 RExC_rx = ReANY(Rx);
7758 if ( RExC_rx == NULL )
7759 FAIL("Regexp out of space");
7760
7761 rx_flags = orig_rx_flags;
7762
7763 if ( toUSE_UNI_CHARSET_NOT_DEPENDS
7764 && initial_charset == REGEX_DEPENDS_CHARSET)
7765 {
7766
7767 /* Set to use unicode semantics if the pattern is in utf8 and has the
7768 * 'depends' charset specified, as it means unicode when utf8 */
7769 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7770 RExC_uni_semantics = 1;
7771 }
7772
7773 RExC_pm_flags = pm_flags;
7774
7775 if (runtime_code) {
7776 assert(TAINTING_get || !TAINT_get);
7777 if (TAINT_get)
7778 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
7779
7780 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
7781 /* whoops, we have a non-utf8 pattern, whilst run-time code
7782 * got compiled as utf8. Try again with a utf8 pattern */
7783 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7784 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7785 goto redo_parse;
7786 }
7787 }
7788 assert(!pRExC_state->runtime_code_qr);
7789
7790 RExC_sawback = 0;
7791
7792 RExC_seen = 0;
7793 RExC_maxlen = 0;
7794 RExC_in_lookaround = 0;
7795 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7796 RExC_recode_x_to_native = 0;
7797 RExC_in_multi_char_class = 0;
7798
7799 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = RExC_precomp = exp;
7800 RExC_precomp_end = RExC_end = exp + plen;
7801 RExC_nestroot = 0;
7802 RExC_whilem_seen = 0;
7803 RExC_end_op = NULL;
7804 RExC_recurse = NULL;
7805 RExC_study_chunk_recursed = NULL;
7806 RExC_study_chunk_recursed_bytes= 0;
7807 RExC_recurse_count = 0;
7808 RExC_sets_depth = 0;
7809 pRExC_state->code_index = 0;
7810
7811 /* Initialize the string in the compiled pattern. This is so that there is
7812 * something to output if necessary */
7813 set_regex_pv(pRExC_state, Rx);
7814
7815 DEBUG_PARSE_r({
7816 Perl_re_printf( aTHX_
7817 "Starting parse and generation\n");
7818 RExC_lastnum=0;
7819 RExC_lastparse=NULL;
7820 });
7821
7822 /* Allocate space and zero-initialize. Note, the two step process
7823 of zeroing when in debug mode, thus anything assigned has to
7824 happen after that */
7825 if (! RExC_size) {
7826
7827 /* On the first pass of the parse, we guess how big this will be. Then
7828 * we grow in one operation to that amount and then give it back. As
7829 * we go along, we re-allocate what we need.
7830 *
7831 * XXX Currently the guess is essentially that the pattern will be an
7832 * EXACT node with one byte input, one byte output. This is crude, and
7833 * better heuristics are welcome.
7834 *
7835 * On any subsequent passes, we guess what we actually computed in the
7836 * latest earlier pass. Such a pass probably didn't complete so is
7837 * missing stuff. We could improve those guesses by knowing where the
7838 * parse stopped, and use the length so far plus apply the above
7839 * assumption to what's left. */
7840 RExC_size = STR_SZ(RExC_end - RExC_start);
7841 }
7842
7843 Newxc(RExC_rxi, sizeof(regexp_internal) + RExC_size, char, regexp_internal);
7844 if ( RExC_rxi == NULL )
7845 FAIL("Regexp out of space");
7846
7847 Zero(RExC_rxi, sizeof(regexp_internal) + RExC_size, char);
7848 RXi_SET( RExC_rx, RExC_rxi );
7849
7850 /* We start from 0 (over from 0 in the case this is a reparse. The first
7851 * node parsed will give back any excess memory we have allocated so far).
7852 * */
7853 RExC_size = 0;
7854
7855 /* non-zero initialization begins here */
7856 RExC_rx->engine= eng;
7857 RExC_rx->extflags = rx_flags;
7858 RXp_COMPFLAGS(RExC_rx) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7859
7860 if (pm_flags & PMf_IS_QR) {
7861 RExC_rxi->code_blocks = pRExC_state->code_blocks;
7862 if (RExC_rxi->code_blocks) {
7863 RExC_rxi->code_blocks->refcnt++;
7864 }
7865 }
7866
7867 RExC_rx->intflags = 0;
7868
7869 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7870 RExC_parse = exp;
7871
7872 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7873 * code makes sure the final byte is an uncounted NUL. But should this
7874 * ever not be the case, lots of things could read beyond the end of the
7875 * buffer: loops like
7876 * while(isFOO(*RExC_parse)) RExC_parse++;
7877 * strchr(RExC_parse, "foo");
7878 * etc. So it is worth noting. */
7879 assert(*RExC_end == '\0');
7880
7881 RExC_naughty = 0;
7882 RExC_npar = 1;
7883 RExC_parens_buf_size = 0;
7884 RExC_emit_start = RExC_rxi->program;
7885 pRExC_state->code_index = 0;
7886
7887 *((char*) RExC_emit_start) = (char) REG_MAGIC;
7888 RExC_emit = 1;
7889
7890 /* Do the parse */
7891 if (reg(pRExC_state, 0, &flags, 1)) {
7892
7893 /* Success!, But we may need to redo the parse knowing how many parens
7894 * there actually are */
7895 if (IN_PARENS_PASS) {
7896 flags |= RESTART_PARSE;
7897 }
7898
7899 /* We have that number in RExC_npar */
7900 RExC_total_parens = RExC_npar;
7901 }
7902 else if (! MUST_RESTART(flags)) {
7903 ReREFCNT_dec(Rx);
7904 Perl_croak(aTHX_ "panic: reg returned failure to re_op_compile, flags=%#" UVxf, (UV) flags);
7905 }
7906
7907 /* Here, we either have success, or we have to redo the parse for some reason */
7908 if (MUST_RESTART(flags)) {
7909
7910 /* It's possible to write a regexp in ascii that represents Unicode
7911 codepoints outside of the byte range, such as via \x{100}. If we
7912 detect such a sequence we have to convert the entire pattern to utf8
7913 and then recompile, as our sizing calculation will have been based
7914 on 1 byte == 1 character, but we will need to use utf8 to encode
7915 at least some part of the pattern, and therefore must convert the whole
7916 thing.
7917 -- dmq */
7918 if (flags & NEED_UTF8) {
7919
7920 /* We have stored the offset of the final warning output so far.
7921 * That must be adjusted. Any variant characters between the start
7922 * of the pattern and this warning count for 2 bytes in the final,
7923 * so just add them again */
7924 if (UNLIKELY(RExC_latest_warn_offset > 0)) {
7925 RExC_latest_warn_offset +=
7926 variant_under_utf8_count((U8 *) exp, (U8 *) exp
7927 + RExC_latest_warn_offset);
7928 }
7929 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7930 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7931 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse after upgrade\n"));
7932 }
7933 else {
7934 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse\n"));
7935 }
7936
7937 if (ALL_PARENS_COUNTED) {
7938 /* Make enough room for all the known parens, and zero it */
7939 Renew(RExC_open_parens, RExC_total_parens, regnode_offset);
7940 Zero(RExC_open_parens, RExC_total_parens, regnode_offset);
7941 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
7942
7943 Renew(RExC_close_parens, RExC_total_parens, regnode_offset);
7944 Zero(RExC_close_parens, RExC_total_parens, regnode_offset);
7945 }
7946 else { /* Parse did not complete. Reinitialize the parentheses
7947 structures */
7948 RExC_total_parens = 0;
7949 if (RExC_open_parens) {
7950 Safefree(RExC_open_parens);
7951 RExC_open_parens = NULL;
7952 }
7953 if (RExC_close_parens) {
7954 Safefree(RExC_close_parens);
7955 RExC_close_parens = NULL;
7956 }
7957 }
7958
7959 /* Clean up what we did in this parse */
7960 SvREFCNT_dec_NN(RExC_rx_sv);
7961
7962 goto redo_parse;
7963 }
7964
7965 /* Here, we have successfully parsed and generated the pattern's program
7966 * for the regex engine. We are ready to finish things up and look for
7967 * optimizations. */
7968
7969 /* Update the string to compile, with correct modifiers, etc */
7970 set_regex_pv(pRExC_state, Rx);
7971
7972 RExC_rx->nparens = RExC_total_parens - 1;
7973
7974 /* Uses the upper 4 bits of the FLAGS field, so keep within that size */
7975 if (RExC_whilem_seen > 15)
7976 RExC_whilem_seen = 15;
7977
7978 DEBUG_PARSE_r({
7979 Perl_re_printf( aTHX_
7980 "Required size %" IVdf " nodes\n", (IV)RExC_size);
7981 RExC_lastnum=0;
7982 RExC_lastparse=NULL;
7983 });
7984
7985 #ifdef RE_TRACK_PATTERN_OFFSETS
7986 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7987 "%s %" UVuf " bytes for offset annotations.\n",
7988 RExC_offsets ? "Got" : "Couldn't get",
7989 (UV)((RExC_offsets[0] * 2 + 1))));
7990 DEBUG_OFFSETS_r(if (RExC_offsets) {
7991 const STRLEN len = RExC_offsets[0];
7992 STRLEN i;
7993 DECLARE_AND_GET_RE_DEBUG_FLAGS;
7994 Perl_re_printf( aTHX_
7995 "Offsets: [%" UVuf "]\n\t", (UV)RExC_offsets[0]);
7996 for (i = 1; i <= len; i++) {
7997 if (RExC_offsets[i*2-1] || RExC_offsets[i*2])
7998 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7999 (UV)i, (UV)RExC_offsets[i*2-1], (UV)RExC_offsets[i*2]);
8000 }
8001 Perl_re_printf( aTHX_ "\n");
8002 });
8003
8004 #else
8005 SetProgLen(RExC_rxi,RExC_size);
8006 #endif
8007
8008 DEBUG_DUMP_PRE_OPTIMIZE_r({
8009 SV * const sv = sv_newmortal();
8010 RXi_GET_DECL(RExC_rx, ri);
8011 DEBUG_RExC_seen();
8012 Perl_re_printf( aTHX_ "Program before optimization:\n");
8013
8014 (void)dumpuntil(RExC_rx, ri->program, ri->program + 1, NULL, NULL,
8015 sv, 0, 0);
8016 });
8017
8018 DEBUG_OPTIMISE_r(
8019 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
8020 );
8021
8022 /* XXXX To minimize changes to RE engine we always allocate
8023 3-units-long substrs field. */
8024 Newx(RExC_rx->substrs, 1, struct reg_substr_data);
8025 if (RExC_recurse_count) {
8026 Newx(RExC_recurse, RExC_recurse_count, regnode *);
8027 SAVEFREEPV(RExC_recurse);
8028 }
8029
8030 if (RExC_seen & REG_RECURSE_SEEN) {
8031 /* Note, RExC_total_parens is 1 + the number of parens in a pattern.
8032 * So its 1 if there are no parens. */
8033 RExC_study_chunk_recursed_bytes= (RExC_total_parens >> 3) +
8034 ((RExC_total_parens & 0x07) != 0);
8035 Newx(RExC_study_chunk_recursed,
8036 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
8037 SAVEFREEPV(RExC_study_chunk_recursed);
8038 }
8039
8040 reStudy:
8041 RExC_rx->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
8042 DEBUG_r(
8043 RExC_study_chunk_recursed_count= 0;
8044 );
8045 Zero(RExC_rx->substrs, 1, struct reg_substr_data);
8046 if (RExC_study_chunk_recursed) {
8047 Zero(RExC_study_chunk_recursed,
8048 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
8049 }
8050
8051
8052 #ifdef TRIE_STUDY_OPT
8053 if (!restudied) {
8054 StructCopy(&zero_scan_data, &data, scan_data_t);
8055 copyRExC_state = RExC_state;
8056 } else {
8057 U32 seen=RExC_seen;
8058 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
8059
8060 RExC_state = copyRExC_state;
8061 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
8062 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
8063 else
8064 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
8065 StructCopy(&zero_scan_data, &data, scan_data_t);
8066 }
8067 #else
8068 StructCopy(&zero_scan_data, &data, scan_data_t);
8069 #endif
8070
8071 /* Dig out information for optimizations. */
8072 RExC_rx->extflags = RExC_flags; /* was pm_op */
8073 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
8074
8075 if (UTF)
8076 SvUTF8_on(Rx); /* Unicode in it? */
8077 RExC_rxi->regstclass = NULL;
8078 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
8079 RExC_rx->intflags |= PREGf_NAUGHTY;
8080 scan = RExC_rxi->program + 1; /* First BRANCH. */
8081
8082 /* testing for BRANCH here tells us whether there is "must appear"
8083 data in the pattern. If there is then we can use it for optimisations */
8084 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
8085 */
8086 SSize_t fake;
8087 STRLEN longest_length[2];
8088 regnode_ssc ch_class; /* pointed to by data */
8089 int stclass_flag;
8090 SSize_t last_close = 0; /* pointed to by data */
8091 regnode *first= scan;
8092 regnode *first_next= regnext(first);
8093 int i;
8094
8095 /*
8096 * Skip introductions and multiplicators >= 1
8097 * so that we can extract the 'meat' of the pattern that must
8098 * match in the large if() sequence following.
8099 * NOTE that EXACT is NOT covered here, as it is normally
8100 * picked up by the optimiser separately.
8101 *
8102 * This is unfortunate as the optimiser isnt handling lookahead
8103 * properly currently.
8104 *
8105 */
8106 while ((OP(first) == OPEN && (sawopen = 1)) ||
8107 /* An OR of *one* alternative - should not happen now. */
8108 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
8109 /* for now we can't handle lookbehind IFMATCH*/
8110 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
8111 (OP(first) == PLUS) ||
8112 (OP(first) == MINMOD) ||
8113 /* An {n,m} with n>0 */
8114 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
8115 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
8116 {
8117 /*
8118 * the only op that could be a regnode is PLUS, all the rest
8119 * will be regnode_1 or regnode_2.
8120 *
8121 * (yves doesn't think this is true)
8122 */
8123 if (OP(first) == PLUS)
8124 sawplus = 1;
8125 else {
8126 if (OP(first) == MINMOD)
8127 sawminmod = 1;
8128 first += regarglen[OP(first)];
8129 }
8130 first = NEXTOPER(first);
8131 first_next= regnext(first);
8132 }
8133
8134 /* Starting-point info. */
8135 again:
8136 DEBUG_PEEP("first:", first, 0, 0);
8137 /* Ignore EXACT as we deal with it later. */
8138 if (PL_regkind[OP(first)] == EXACT) {
8139 if (! isEXACTFish(OP(first))) {
8140 NOOP; /* Empty, get anchored substr later. */
8141 }
8142 else
8143 RExC_rxi->regstclass = first;
8144 }
8145 #ifdef TRIE_STCLASS
8146 else if (PL_regkind[OP(first)] == TRIE &&
8147 ((reg_trie_data *)RExC_rxi->data->data[ ARG(first) ])->minlen>0)
8148 {
8149 /* this can happen only on restudy */
8150 RExC_rxi->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
8151 }
8152 #endif
8153 else if (REGNODE_SIMPLE(OP(first)))
8154 RExC_rxi->regstclass = first;
8155 else if (PL_regkind[OP(first)] == BOUND ||
8156 PL_regkind[OP(first)] == NBOUND)
8157 RExC_rxi->regstclass = first;
8158 else if (PL_regkind[OP(first)] == BOL) {
8159 RExC_rx->intflags |= (OP(first) == MBOL
8160 ? PREGf_ANCH_MBOL
8161 : PREGf_ANCH_SBOL);
8162 first = NEXTOPER(first);
8163 goto again;
8164 }
8165 else if (OP(first) == GPOS) {
8166 RExC_rx->intflags |= PREGf_ANCH_GPOS;
8167 first = NEXTOPER(first);
8168 goto again;
8169 }
8170 else if ((!sawopen || !RExC_sawback) &&
8171 !sawlookahead &&
8172 (OP(first) == STAR &&
8173 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
8174 !(RExC_rx->intflags & PREGf_ANCH) && !pRExC_state->code_blocks)
8175 {
8176 /* turn .* into ^.* with an implied $*=1 */
8177 const int type =
8178 (OP(NEXTOPER(first)) == REG_ANY)
8179 ? PREGf_ANCH_MBOL
8180 : PREGf_ANCH_SBOL;
8181 RExC_rx->intflags |= (type | PREGf_IMPLICIT);
8182 first = NEXTOPER(first);
8183 goto again;
8184 }
8185 if (sawplus && !sawminmod && !sawlookahead
8186 && (!sawopen || !RExC_sawback)
8187 && !pRExC_state->code_blocks) /* May examine pos and $& */
8188 /* x+ must match at the 1st pos of run of x's */
8189 RExC_rx->intflags |= PREGf_SKIP;
8190
8191 /* Scan is after the zeroth branch, first is atomic matcher. */
8192 #ifdef TRIE_STUDY_OPT
8193 DEBUG_PARSE_r(
8194 if (!restudied)
8195 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8196 (IV)(first - scan + 1))
8197 );
8198 #else
8199 DEBUG_PARSE_r(
8200 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8201 (IV)(first - scan + 1))
8202 );
8203 #endif
8204
8205
8206 /*
8207 * If there's something expensive in the r.e., find the
8208 * longest literal string that must appear and make it the
8209 * regmust. Resolve ties in favor of later strings, since
8210 * the regstart check works with the beginning of the r.e.
8211 * and avoiding duplication strengthens checking. Not a
8212 * strong reason, but sufficient in the absence of others.
8213 * [Now we resolve ties in favor of the earlier string if
8214 * it happens that c_offset_min has been invalidated, since the
8215 * earlier string may buy us something the later one won't.]
8216 */
8217
8218 data.substrs[0].str = newSVpvs("");
8219 data.substrs[1].str = newSVpvs("");
8220 data.last_found = newSVpvs("");
8221 data.cur_is_floating = 0; /* initially any found substring is fixed */
8222 ENTER_with_name("study_chunk");
8223 SAVEFREESV(data.substrs[0].str);
8224 SAVEFREESV(data.substrs[1].str);
8225 SAVEFREESV(data.last_found);
8226 first = scan;
8227 if (!RExC_rxi->regstclass) {
8228 ssc_init(pRExC_state, &ch_class);
8229 data.start_class = &ch_class;
8230 stclass_flag = SCF_DO_STCLASS_AND;
8231 } else /* XXXX Check for BOUND? */
8232 stclass_flag = 0;
8233 data.last_closep = &last_close;
8234
8235 DEBUG_RExC_seen();
8236 /*
8237 * MAIN ENTRY FOR study_chunk() FOR m/PATTERN/
8238 * (NO top level branches)
8239 */
8240 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
8241 scan + RExC_size, /* Up to end */
8242 &data, -1, 0, NULL,
8243 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
8244 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
8245 0, TRUE);
8246
8247
8248 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
8249
8250
8251 if ( RExC_total_parens == 1 && !data.cur_is_floating
8252 && data.last_start_min == 0 && data.last_end > 0
8253 && !RExC_seen_zerolen
8254 && !(RExC_seen & REG_VERBARG_SEEN)
8255 && !(RExC_seen & REG_GPOS_SEEN)
8256 ){
8257 RExC_rx->extflags |= RXf_CHECK_ALL;
8258 }
8259 scan_commit(pRExC_state, &data,&minlen, 0);
8260
8261
8262 /* XXX this is done in reverse order because that's the way the
8263 * code was before it was parameterised. Don't know whether it
8264 * actually needs doing in reverse order. DAPM */
8265 for (i = 1; i >= 0; i--) {
8266 longest_length[i] = CHR_SVLEN(data.substrs[i].str);
8267
8268 if ( !( i
8269 && SvCUR(data.substrs[0].str) /* ok to leave SvCUR */
8270 && data.substrs[0].min_offset
8271 == data.substrs[1].min_offset
8272 && SvCUR(data.substrs[0].str)
8273 == SvCUR(data.substrs[1].str)
8274 )
8275 && S_setup_longest (aTHX_ pRExC_state,
8276 &(RExC_rx->substrs->data[i]),
8277 &(data.substrs[i]),
8278 longest_length[i]))
8279 {
8280 RExC_rx->substrs->data[i].min_offset =
8281 data.substrs[i].min_offset - data.substrs[i].lookbehind;
8282
8283 RExC_rx->substrs->data[i].max_offset = data.substrs[i].max_offset;
8284 /* Don't offset infinity */
8285 if (data.substrs[i].max_offset < OPTIMIZE_INFTY)
8286 RExC_rx->substrs->data[i].max_offset -= data.substrs[i].lookbehind;
8287 SvREFCNT_inc_simple_void_NN(data.substrs[i].str);
8288 }
8289 else {
8290 RExC_rx->substrs->data[i].substr = NULL;
8291 RExC_rx->substrs->data[i].utf8_substr = NULL;
8292 longest_length[i] = 0;
8293 }
8294 }
8295
8296 LEAVE_with_name("study_chunk");
8297
8298 if (RExC_rxi->regstclass
8299 && (OP(RExC_rxi->regstclass) == REG_ANY || OP(RExC_rxi->regstclass) == SANY))
8300 RExC_rxi->regstclass = NULL;
8301
8302 if ((!(RExC_rx->substrs->data[0].substr || RExC_rx->substrs->data[0].utf8_substr)
8303 || RExC_rx->substrs->data[0].min_offset)
8304 && stclass_flag
8305 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8306 && is_ssc_worth_it(pRExC_state, data.start_class))
8307 {
8308 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8309
8310 ssc_finalize(pRExC_state, data.start_class);
8311
8312 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8313 StructCopy(data.start_class,
8314 (regnode_ssc*)RExC_rxi->data->data[n],
8315 regnode_ssc);
8316 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8317 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8318 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
8319 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8320 Perl_re_printf( aTHX_
8321 "synthetic stclass \"%s\".\n",
8322 SvPVX_const(sv));});
8323 data.start_class = NULL;
8324 }
8325
8326 /* A temporary algorithm prefers floated substr to fixed one of
8327 * same length to dig more info. */
8328 i = (longest_length[0] <= longest_length[1]);
8329 RExC_rx->substrs->check_ix = i;
8330 RExC_rx->check_end_shift = RExC_rx->substrs->data[i].end_shift;
8331 RExC_rx->check_substr = RExC_rx->substrs->data[i].substr;
8332 RExC_rx->check_utf8 = RExC_rx->substrs->data[i].utf8_substr;
8333 RExC_rx->check_offset_min = RExC_rx->substrs->data[i].min_offset;
8334 RExC_rx->check_offset_max = RExC_rx->substrs->data[i].max_offset;
8335 if (!i && (RExC_rx->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS)))
8336 RExC_rx->intflags |= PREGf_NOSCAN;
8337
8338 if ((RExC_rx->check_substr || RExC_rx->check_utf8) ) {
8339 RExC_rx->extflags |= RXf_USE_INTUIT;
8340 if (SvTAIL(RExC_rx->check_substr ? RExC_rx->check_substr : RExC_rx->check_utf8))
8341 RExC_rx->extflags |= RXf_INTUIT_TAIL;
8342 }
8343
8344 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
8345 if ( (STRLEN)minlen < longest_length[1] )
8346 minlen= longest_length[1];
8347 if ( (STRLEN)minlen < longest_length[0] )
8348 minlen= longest_length[0];
8349 */
8350 }
8351 else {
8352 /* Several toplevels. Best we can is to set minlen. */
8353 SSize_t fake;
8354 regnode_ssc ch_class;
8355 SSize_t last_close = 0;
8356
8357 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
8358
8359 scan = RExC_rxi->program + 1;
8360 ssc_init(pRExC_state, &ch_class);
8361 data.start_class = &ch_class;
8362 data.last_closep = &last_close;
8363
8364 DEBUG_RExC_seen();
8365 /*
8366 * MAIN ENTRY FOR study_chunk() FOR m/P1|P2|.../
8367 * (patterns WITH top level branches)
8368 */
8369 minlen = study_chunk(pRExC_state,
8370 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
8371 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
8372 ? SCF_TRIE_DOING_RESTUDY
8373 : 0),
8374 0, TRUE);
8375
8376 CHECK_RESTUDY_GOTO_butfirst(NOOP);
8377
8378 RExC_rx->check_substr = NULL;
8379 RExC_rx->check_utf8 = NULL;
8380 RExC_rx->substrs->data[0].substr = NULL;
8381 RExC_rx->substrs->data[0].utf8_substr = NULL;
8382 RExC_rx->substrs->data[1].substr = NULL;
8383 RExC_rx->substrs->data[1].utf8_substr = NULL;
8384
8385 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8386 && is_ssc_worth_it(pRExC_state, data.start_class))
8387 {
8388 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8389
8390 ssc_finalize(pRExC_state, data.start_class);
8391
8392 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8393 StructCopy(data.start_class,
8394 (regnode_ssc*)RExC_rxi->data->data[n],
8395 regnode_ssc);
8396 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8397 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8398 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
8399 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8400 Perl_re_printf( aTHX_
8401 "synthetic stclass \"%s\".\n",
8402 SvPVX_const(sv));});
8403 data.start_class = NULL;
8404 }
8405 }
8406
8407 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
8408 RExC_rx->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
8409 RExC_rx->maxlen = REG_INFTY;
8410 }
8411 else {
8412 RExC_rx->maxlen = RExC_maxlen;
8413 }
8414
8415 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
8416 the "real" pattern. */
8417 DEBUG_OPTIMISE_r({
8418 Perl_re_printf( aTHX_ "minlen: %" IVdf " RExC_rx->minlen:%" IVdf " maxlen:%" IVdf "\n",
8419 (IV)minlen, (IV)RExC_rx->minlen, (IV)RExC_maxlen);
8420 });
8421 RExC_rx->minlenret = minlen;
8422 if (RExC_rx->minlen < minlen)
8423 RExC_rx->minlen = minlen;
8424
8425 if (RExC_seen & REG_RECURSE_SEEN ) {
8426 RExC_rx->intflags |= PREGf_RECURSE_SEEN;
8427 Newx(RExC_rx->recurse_locinput, RExC_rx->nparens + 1, char *);
8428 }
8429 if (RExC_seen & REG_GPOS_SEEN)
8430 RExC_rx->intflags |= PREGf_GPOS_SEEN;
8431 if (RExC_seen & REG_LOOKBEHIND_SEEN)
8432 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
8433 lookbehind */
8434 if (pRExC_state->code_blocks)
8435 RExC_rx->extflags |= RXf_EVAL_SEEN;
8436 if (RExC_seen & REG_VERBARG_SEEN)
8437 {
8438 RExC_rx->intflags |= PREGf_VERBARG_SEEN;
8439 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
8440 }
8441 if (RExC_seen & REG_CUTGROUP_SEEN)
8442 RExC_rx->intflags |= PREGf_CUTGROUP_SEEN;
8443 if (pm_flags & PMf_USE_RE_EVAL)
8444 RExC_rx->intflags |= PREGf_USE_RE_EVAL;
8445 if (RExC_paren_names)
8446 RXp_PAREN_NAMES(RExC_rx) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
8447 else
8448 RXp_PAREN_NAMES(RExC_rx) = NULL;
8449
8450 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
8451 * so it can be used in pp.c */
8452 if (RExC_rx->intflags & PREGf_ANCH)
8453 RExC_rx->extflags |= RXf_IS_ANCHORED;
8454
8455
8456 {
8457 /* this is used to identify "special" patterns that might result
8458 * in Perl NOT calling the regex engine and instead doing the match "itself",
8459 * particularly special cases in split//. By having the regex compiler
8460 * do this pattern matching at a regop level (instead of by inspecting the pattern)
8461 * we avoid weird issues with equivalent patterns resulting in different behavior,
8462 * AND we allow non Perl engines to get the same optimizations by the setting the
8463 * flags appropriately - Yves */
8464 regnode *first = RExC_rxi->program + 1;
8465 U8 fop = OP(first);
8466 regnode *next = NEXTOPER(first);
8467 /* It's safe to read through *next only if OP(first) is a regop of
8468 * the right type (not EXACT, for example).
8469 */
8470 U8 nop = (fop == NOTHING || fop == MBOL || fop == SBOL || fop == PLUS)
8471 ? OP(next) : 0;
8472
8473 if (PL_regkind[fop] == NOTHING && nop == END)
8474 RExC_rx->extflags |= RXf_NULL;
8475 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
8476 /* when fop is SBOL first->flags will be true only when it was
8477 * produced by parsing /\A/, and not when parsing /^/. This is
8478 * very important for the split code as there we want to
8479 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
8480 * See rt #122761 for more details. -- Yves */
8481 RExC_rx->extflags |= RXf_START_ONLY;
8482 else if (fop == PLUS
8483 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
8484 && OP(regnext(first)) == END)
8485 RExC_rx->extflags |= RXf_WHITE;
8486 else if ( RExC_rx->extflags & RXf_SPLIT
8487 && (PL_regkind[fop] == EXACT && ! isEXACTFish(fop))
8488 && STR_LEN(first) == 1
8489 && *(STRING(first)) == ' '
8490 && OP(regnext(first)) == END )
8491 RExC_rx->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
8492
8493 }
8494
8495 if (RExC_contains_locale) {
8496 RXp_EXTFLAGS(RExC_rx) |= RXf_TAINTED;
8497 }
8498
8499 #ifdef DEBUGGING
8500 if (RExC_paren_names) {
8501 RExC_rxi->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
8502 RExC_rxi->data->data[RExC_rxi->name_list_idx]
8503 = (void*)SvREFCNT_inc(RExC_paren_name_list);
8504 } else
8505 #endif
8506 RExC_rxi->name_list_idx = 0;
8507
8508 while ( RExC_recurse_count > 0 ) {
8509 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
8510 /*
8511 * This data structure is set up in study_chunk() and is used
8512 * to calculate the distance between a GOSUB regopcode and
8513 * the OPEN/CURLYM (CURLYM's are special and can act like OPEN's)
8514 * it refers to.
8515 *
8516 * If for some reason someone writes code that optimises
8517 * away a GOSUB opcode then the assert should be changed to
8518 * an if(scan) to guard the ARG2L_SET() - Yves
8519 *
8520 */
8521 assert(scan && OP(scan) == GOSUB);
8522 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - REGNODE_OFFSET(scan));
8523 }
8524
8525 Newxz(RExC_rx->offs, RExC_total_parens, regexp_paren_pair);
8526 /* assume we don't need to swap parens around before we match */
8527 DEBUG_TEST_r({
8528 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
8529 (unsigned long)RExC_study_chunk_recursed_count);
8530 });
8531 DEBUG_DUMP_r({
8532 DEBUG_RExC_seen();
8533 Perl_re_printf( aTHX_ "Final program:\n");
8534 regdump(RExC_rx);
8535 });
8536
8537 if (RExC_open_parens) {
8538 Safefree(RExC_open_parens);
8539 RExC_open_parens = NULL;
8540 }
8541 if (RExC_close_parens) {
8542 Safefree(RExC_close_parens);
8543 RExC_close_parens = NULL;
8544 }
8545
8546 #ifdef USE_ITHREADS
8547 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
8548 * by setting the regexp SV to readonly-only instead. If the
8549 * pattern's been recompiled, the USEDness should remain. */
8550 if (old_re && SvREADONLY(old_re))
8551 SvREADONLY_on(Rx);
8552 #endif
8553 return Rx;
8554 }
8555
8556
8557 SV*
Perl_reg_named_buff(pTHX_ REGEXP * const rx,SV * const key,SV * const value,const U32 flags)8558 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
8559 const U32 flags)
8560 {
8561 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
8562
8563 PERL_UNUSED_ARG(value);
8564
8565 if (flags & RXapif_FETCH) {
8566 return reg_named_buff_fetch(rx, key, flags);
8567 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
8568 Perl_croak_no_modify();
8569 return NULL;
8570 } else if (flags & RXapif_EXISTS) {
8571 return reg_named_buff_exists(rx, key, flags)
8572 ? &PL_sv_yes
8573 : &PL_sv_no;
8574 } else if (flags & RXapif_REGNAMES) {
8575 return reg_named_buff_all(rx, flags);
8576 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
8577 return reg_named_buff_scalar(rx, flags);
8578 } else {
8579 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
8580 return NULL;
8581 }
8582 }
8583
8584 SV*
Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx,const SV * const lastkey,const U32 flags)8585 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
8586 const U32 flags)
8587 {
8588 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
8589 PERL_UNUSED_ARG(lastkey);
8590
8591 if (flags & RXapif_FIRSTKEY)
8592 return reg_named_buff_firstkey(rx, flags);
8593 else if (flags & RXapif_NEXTKEY)
8594 return reg_named_buff_nextkey(rx, flags);
8595 else {
8596 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
8597 (int)flags);
8598 return NULL;
8599 }
8600 }
8601
8602 SV*
Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r,SV * const namesv,const U32 flags)8603 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
8604 const U32 flags)
8605 {
8606 SV *ret;
8607 struct regexp *const rx = ReANY(r);
8608
8609 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
8610
8611 if (rx && RXp_PAREN_NAMES(rx)) {
8612 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
8613 if (he_str) {
8614 IV i;
8615 SV* sv_dat=HeVAL(he_str);
8616 I32 *nums=(I32*)SvPVX(sv_dat);
8617 AV * const retarray = (flags & RXapif_ALL) ? newAV() : NULL;
8618 for ( i=0; i<SvIVX(sv_dat); i++ ) {
8619 if ((I32)(rx->nparens) >= nums[i]
8620 && rx->offs[nums[i]].start != -1
8621 && rx->offs[nums[i]].end != -1)
8622 {
8623 ret = newSVpvs("");
8624 CALLREG_NUMBUF_FETCH(r, nums[i], ret);
8625 if (!retarray)
8626 return ret;
8627 } else {
8628 if (retarray)
8629 ret = newSVsv(&PL_sv_undef);
8630 }
8631 if (retarray)
8632 av_push(retarray, ret);
8633 }
8634 if (retarray)
8635 return newRV_noinc(MUTABLE_SV(retarray));
8636 }
8637 }
8638 return NULL;
8639 }
8640
8641 bool
Perl_reg_named_buff_exists(pTHX_ REGEXP * const r,SV * const key,const U32 flags)8642 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
8643 const U32 flags)
8644 {
8645 struct regexp *const rx = ReANY(r);
8646
8647 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
8648
8649 if (rx && RXp_PAREN_NAMES(rx)) {
8650 if (flags & RXapif_ALL) {
8651 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
8652 } else {
8653 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
8654 if (sv) {
8655 SvREFCNT_dec_NN(sv);
8656 return TRUE;
8657 } else {
8658 return FALSE;
8659 }
8660 }
8661 } else {
8662 return FALSE;
8663 }
8664 }
8665
8666 SV*
Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r,const U32 flags)8667 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
8668 {
8669 struct regexp *const rx = ReANY(r);
8670
8671 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
8672
8673 if ( rx && RXp_PAREN_NAMES(rx) ) {
8674 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
8675
8676 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
8677 } else {
8678 return FALSE;
8679 }
8680 }
8681
8682 SV*
Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r,const U32 flags)8683 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
8684 {
8685 struct regexp *const rx = ReANY(r);
8686 DECLARE_AND_GET_RE_DEBUG_FLAGS;
8687
8688 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
8689
8690 if (rx && RXp_PAREN_NAMES(rx)) {
8691 HV *hv = RXp_PAREN_NAMES(rx);
8692 HE *temphe;
8693 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8694 IV i;
8695 IV parno = 0;
8696 SV* sv_dat = HeVAL(temphe);
8697 I32 *nums = (I32*)SvPVX(sv_dat);
8698 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8699 if ((I32)(rx->lastparen) >= nums[i] &&
8700 rx->offs[nums[i]].start != -1 &&
8701 rx->offs[nums[i]].end != -1)
8702 {
8703 parno = nums[i];
8704 break;
8705 }
8706 }
8707 if (parno || flags & RXapif_ALL) {
8708 return newSVhek(HeKEY_hek(temphe));
8709 }
8710 }
8711 }
8712 return NULL;
8713 }
8714
8715 SV*
Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r,const U32 flags)8716 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
8717 {
8718 SV *ret;
8719 AV *av;
8720 SSize_t length;
8721 struct regexp *const rx = ReANY(r);
8722
8723 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
8724
8725 if (rx && RXp_PAREN_NAMES(rx)) {
8726 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
8727 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
8728 } else if (flags & RXapif_ONE) {
8729 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
8730 av = MUTABLE_AV(SvRV(ret));
8731 length = av_count(av);
8732 SvREFCNT_dec_NN(ret);
8733 return newSViv(length);
8734 } else {
8735 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8736 (int)flags);
8737 return NULL;
8738 }
8739 }
8740 return &PL_sv_undef;
8741 }
8742
8743 SV*
Perl_reg_named_buff_all(pTHX_ REGEXP * const r,const U32 flags)8744 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8745 {
8746 struct regexp *const rx = ReANY(r);
8747 AV *av = newAV();
8748
8749 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8750
8751 if (rx && RXp_PAREN_NAMES(rx)) {
8752 HV *hv= RXp_PAREN_NAMES(rx);
8753 HE *temphe;
8754 (void)hv_iterinit(hv);
8755 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8756 IV i;
8757 IV parno = 0;
8758 SV* sv_dat = HeVAL(temphe);
8759 I32 *nums = (I32*)SvPVX(sv_dat);
8760 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8761 if ((I32)(rx->lastparen) >= nums[i] &&
8762 rx->offs[nums[i]].start != -1 &&
8763 rx->offs[nums[i]].end != -1)
8764 {
8765 parno = nums[i];
8766 break;
8767 }
8768 }
8769 if (parno || flags & RXapif_ALL) {
8770 av_push(av, newSVhek(HeKEY_hek(temphe)));
8771 }
8772 }
8773 }
8774
8775 return newRV_noinc(MUTABLE_SV(av));
8776 }
8777
8778 void
Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r,const I32 paren,SV * const sv)8779 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8780 SV * const sv)
8781 {
8782 struct regexp *const rx = ReANY(r);
8783 char *s = NULL;
8784 SSize_t i = 0;
8785 SSize_t s1, t1;
8786 I32 n = paren;
8787
8788 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8789
8790 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8791 || n == RX_BUFF_IDX_CARET_FULLMATCH
8792 || n == RX_BUFF_IDX_CARET_POSTMATCH
8793 )
8794 {
8795 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8796 if (!keepcopy) {
8797 /* on something like
8798 * $r = qr/.../;
8799 * /$qr/p;
8800 * the KEEPCOPY is set on the PMOP rather than the regex */
8801 if (PL_curpm && r == PM_GETRE(PL_curpm))
8802 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8803 }
8804 if (!keepcopy)
8805 goto ret_undef;
8806 }
8807
8808 if (!rx->subbeg)
8809 goto ret_undef;
8810
8811 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8812 /* no need to distinguish between them any more */
8813 n = RX_BUFF_IDX_FULLMATCH;
8814
8815 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8816 && rx->offs[0].start != -1)
8817 {
8818 /* $`, ${^PREMATCH} */
8819 i = rx->offs[0].start;
8820 s = rx->subbeg;
8821 }
8822 else
8823 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8824 && rx->offs[0].end != -1)
8825 {
8826 /* $', ${^POSTMATCH} */
8827 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8828 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8829 }
8830 else
8831 if (inRANGE(n, 0, (I32)rx->nparens) &&
8832 (s1 = rx->offs[n].start) != -1 &&
8833 (t1 = rx->offs[n].end) != -1)
8834 {
8835 /* $&, ${^MATCH}, $1 ... */
8836 i = t1 - s1;
8837 s = rx->subbeg + s1 - rx->suboffset;
8838 } else {
8839 goto ret_undef;
8840 }
8841
8842 assert(s >= rx->subbeg);
8843 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8844 if (i >= 0) {
8845 #ifdef NO_TAINT_SUPPORT
8846 sv_setpvn(sv, s, i);
8847 #else
8848 const int oldtainted = TAINT_get;
8849 TAINT_NOT;
8850 sv_setpvn(sv, s, i);
8851 TAINT_set(oldtainted);
8852 #endif
8853 if (RXp_MATCH_UTF8(rx))
8854 SvUTF8_on(sv);
8855 else
8856 SvUTF8_off(sv);
8857 if (TAINTING_get) {
8858 if (RXp_MATCH_TAINTED(rx)) {
8859 if (SvTYPE(sv) >= SVt_PVMG) {
8860 MAGIC* const mg = SvMAGIC(sv);
8861 MAGIC* mgt;
8862 TAINT;
8863 SvMAGIC_set(sv, mg->mg_moremagic);
8864 SvTAINT(sv);
8865 if ((mgt = SvMAGIC(sv))) {
8866 mg->mg_moremagic = mgt;
8867 SvMAGIC_set(sv, mg);
8868 }
8869 } else {
8870 TAINT;
8871 SvTAINT(sv);
8872 }
8873 } else
8874 SvTAINTED_off(sv);
8875 }
8876 } else {
8877 ret_undef:
8878 sv_set_undef(sv);
8879 return;
8880 }
8881 }
8882
8883 void
Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx,const I32 paren,SV const * const value)8884 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8885 SV const * const value)
8886 {
8887 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8888
8889 PERL_UNUSED_ARG(rx);
8890 PERL_UNUSED_ARG(paren);
8891 PERL_UNUSED_ARG(value);
8892
8893 if (!PL_localizing)
8894 Perl_croak_no_modify();
8895 }
8896
8897 I32
Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r,const SV * const sv,const I32 paren)8898 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8899 const I32 paren)
8900 {
8901 struct regexp *const rx = ReANY(r);
8902 I32 i;
8903 I32 s1, t1;
8904
8905 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8906
8907 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8908 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8909 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8910 )
8911 {
8912 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8913 if (!keepcopy) {
8914 /* on something like
8915 * $r = qr/.../;
8916 * /$qr/p;
8917 * the KEEPCOPY is set on the PMOP rather than the regex */
8918 if (PL_curpm && r == PM_GETRE(PL_curpm))
8919 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8920 }
8921 if (!keepcopy)
8922 goto warn_undef;
8923 }
8924
8925 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8926 switch (paren) {
8927 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8928 case RX_BUFF_IDX_PREMATCH: /* $` */
8929 if (rx->offs[0].start != -1) {
8930 i = rx->offs[0].start;
8931 if (i > 0) {
8932 s1 = 0;
8933 t1 = i;
8934 goto getlen;
8935 }
8936 }
8937 return 0;
8938
8939 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8940 case RX_BUFF_IDX_POSTMATCH: /* $' */
8941 if (rx->offs[0].end != -1) {
8942 i = rx->sublen - rx->offs[0].end;
8943 if (i > 0) {
8944 s1 = rx->offs[0].end;
8945 t1 = rx->sublen;
8946 goto getlen;
8947 }
8948 }
8949 return 0;
8950
8951 default: /* $& / ${^MATCH}, $1, $2, ... */
8952 if (paren <= (I32)rx->nparens &&
8953 (s1 = rx->offs[paren].start) != -1 &&
8954 (t1 = rx->offs[paren].end) != -1)
8955 {
8956 i = t1 - s1;
8957 goto getlen;
8958 } else {
8959 warn_undef:
8960 if (ckWARN(WARN_UNINITIALIZED))
8961 report_uninit((const SV *)sv);
8962 return 0;
8963 }
8964 }
8965 getlen:
8966 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8967 const char * const s = rx->subbeg - rx->suboffset + s1;
8968 const U8 *ep;
8969 STRLEN el;
8970
8971 i = t1 - s1;
8972 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8973 i = el;
8974 }
8975 return i;
8976 }
8977
8978 SV*
Perl_reg_qr_package(pTHX_ REGEXP * const rx)8979 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8980 {
8981 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8982 PERL_UNUSED_ARG(rx);
8983 if (0)
8984 return NULL;
8985 else
8986 return newSVpvs("Regexp");
8987 }
8988
8989 /* Scans the name of a named buffer from the pattern.
8990 * If flags is REG_RSN_RETURN_NULL returns null.
8991 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8992 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8993 * to the parsed name as looked up in the RExC_paren_names hash.
8994 * If there is an error throws a vFAIL().. type exception.
8995 */
8996
8997 #define REG_RSN_RETURN_NULL 0
8998 #define REG_RSN_RETURN_NAME 1
8999 #define REG_RSN_RETURN_DATA 2
9000
9001 STATIC SV*
S_reg_scan_name(pTHX_ RExC_state_t * pRExC_state,U32 flags)9002 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
9003 {
9004 char *name_start = RExC_parse;
9005 SV* sv_name;
9006
9007 PERL_ARGS_ASSERT_REG_SCAN_NAME;
9008
9009 assert (RExC_parse <= RExC_end);
9010 if (RExC_parse == RExC_end) NOOP;
9011 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
9012 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
9013 * using do...while */
9014 if (UTF)
9015 do {
9016 RExC_parse += UTF8SKIP(RExC_parse);
9017 } while ( RExC_parse < RExC_end
9018 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
9019 else
9020 do {
9021 RExC_parse++;
9022 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
9023 } else {
9024 RExC_parse++; /* so the <- from the vFAIL is after the offending
9025 character */
9026 vFAIL("Group name must start with a non-digit word character");
9027 }
9028 sv_name = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
9029 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
9030 if ( flags == REG_RSN_RETURN_NAME)
9031 return sv_name;
9032 else if (flags==REG_RSN_RETURN_DATA) {
9033 HE *he_str = NULL;
9034 SV *sv_dat = NULL;
9035 if ( ! sv_name ) /* should not happen*/
9036 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
9037 if (RExC_paren_names)
9038 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
9039 if ( he_str )
9040 sv_dat = HeVAL(he_str);
9041 if ( ! sv_dat ) { /* Didn't find group */
9042
9043 /* It might be a forward reference; we can't fail until we
9044 * know, by completing the parse to get all the groups, and
9045 * then reparsing */
9046 if (ALL_PARENS_COUNTED) {
9047 vFAIL("Reference to nonexistent named group");
9048 }
9049 else {
9050 REQUIRE_PARENS_PASS;
9051 }
9052 }
9053 return sv_dat;
9054 }
9055
9056 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
9057 (unsigned long) flags);
9058 }
9059
9060 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
9061 if (RExC_lastparse!=RExC_parse) { \
9062 Perl_re_printf( aTHX_ "%s", \
9063 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
9064 RExC_end - RExC_parse, 16, \
9065 "", "", \
9066 PERL_PV_ESCAPE_UNI_DETECT | \
9067 PERL_PV_PRETTY_ELLIPSES | \
9068 PERL_PV_PRETTY_LTGT | \
9069 PERL_PV_ESCAPE_RE | \
9070 PERL_PV_PRETTY_EXACTSIZE \
9071 ) \
9072 ); \
9073 } else \
9074 Perl_re_printf( aTHX_ "%16s",""); \
9075 \
9076 if (RExC_lastnum!=RExC_emit) \
9077 Perl_re_printf( aTHX_ "|%4zu", RExC_emit); \
9078 else \
9079 Perl_re_printf( aTHX_ "|%4s",""); \
9080 Perl_re_printf( aTHX_ "|%*s%-4s", \
9081 (int)((depth*2)), "", \
9082 (funcname) \
9083 ); \
9084 RExC_lastnum=RExC_emit; \
9085 RExC_lastparse=RExC_parse; \
9086 })
9087
9088
9089
9090 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
9091 DEBUG_PARSE_MSG((funcname)); \
9092 Perl_re_printf( aTHX_ "%4s","\n"); \
9093 })
9094 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
9095 DEBUG_PARSE_MSG((funcname)); \
9096 Perl_re_printf( aTHX_ fmt "\n",args); \
9097 })
9098
9099 /* This section of code defines the inversion list object and its methods. The
9100 * interfaces are highly subject to change, so as much as possible is static to
9101 * this file. An inversion list is here implemented as a malloc'd C UV array
9102 * as an SVt_INVLIST scalar.
9103 *
9104 * An inversion list for Unicode is an array of code points, sorted by ordinal
9105 * number. Each element gives the code point that begins a range that extends
9106 * up-to but not including the code point given by the next element. The final
9107 * element gives the first code point of a range that extends to the platform's
9108 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
9109 * ...) give ranges whose code points are all in the inversion list. We say
9110 * that those ranges are in the set. The odd-numbered elements give ranges
9111 * whose code points are not in the inversion list, and hence not in the set.
9112 * Thus, element [0] is the first code point in the list. Element [1]
9113 * is the first code point beyond that not in the list; and element [2] is the
9114 * first code point beyond that that is in the list. In other words, the first
9115 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
9116 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
9117 * all code points in that range are not in the inversion list. The third
9118 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
9119 * list, and so forth. Thus every element whose index is divisible by two
9120 * gives the beginning of a range that is in the list, and every element whose
9121 * index is not divisible by two gives the beginning of a range not in the
9122 * list. If the final element's index is divisible by two, the inversion list
9123 * extends to the platform's infinity; otherwise the highest code point in the
9124 * inversion list is the contents of that element minus 1.
9125 *
9126 * A range that contains just a single code point N will look like
9127 * invlist[i] == N
9128 * invlist[i+1] == N+1
9129 *
9130 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
9131 * impossible to represent, so element [i+1] is omitted. The single element
9132 * inversion list
9133 * invlist[0] == UV_MAX
9134 * contains just UV_MAX, but is interpreted as matching to infinity.
9135 *
9136 * Taking the complement (inverting) an inversion list is quite simple, if the
9137 * first element is 0, remove it; otherwise add a 0 element at the beginning.
9138 * This implementation reserves an element at the beginning of each inversion
9139 * list to always contain 0; there is an additional flag in the header which
9140 * indicates if the list begins at the 0, or is offset to begin at the next
9141 * element. This means that the inversion list can be inverted without any
9142 * copying; just flip the flag.
9143 *
9144 * More about inversion lists can be found in "Unicode Demystified"
9145 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
9146 *
9147 * The inversion list data structure is currently implemented as an SV pointing
9148 * to an array of UVs that the SV thinks are bytes. This allows us to have an
9149 * array of UV whose memory management is automatically handled by the existing
9150 * facilities for SV's.
9151 *
9152 * Some of the methods should always be private to the implementation, and some
9153 * should eventually be made public */
9154
9155 /* The header definitions are in F<invlist_inline.h> */
9156
9157 #ifndef PERL_IN_XSUB_RE
9158
9159 PERL_STATIC_INLINE UV*
S__invlist_array_init(SV * const invlist,const bool will_have_0)9160 S__invlist_array_init(SV* const invlist, const bool will_have_0)
9161 {
9162 /* Returns a pointer to the first element in the inversion list's array.
9163 * This is called upon initialization of an inversion list. Where the
9164 * array begins depends on whether the list has the code point U+0000 in it
9165 * or not. The other parameter tells it whether the code that follows this
9166 * call is about to put a 0 in the inversion list or not. The first
9167 * element is either the element reserved for 0, if TRUE, or the element
9168 * after it, if FALSE */
9169
9170 bool* offset = get_invlist_offset_addr(invlist);
9171 UV* zero_addr = (UV *) SvPVX(invlist);
9172
9173 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
9174
9175 /* Must be empty */
9176 assert(! _invlist_len(invlist));
9177
9178 *zero_addr = 0;
9179
9180 /* 1^1 = 0; 1^0 = 1 */
9181 *offset = 1 ^ will_have_0;
9182 return zero_addr + *offset;
9183 }
9184
9185 STATIC void
S_invlist_replace_list_destroys_src(pTHX_ SV * dest,SV * src)9186 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
9187 {
9188 /* Replaces the inversion list in 'dest' with the one from 'src'. It
9189 * steals the list from 'src', so 'src' is made to have a NULL list. This
9190 * is similar to what SvSetMagicSV() would do, if it were implemented on
9191 * inversion lists, though this routine avoids a copy */
9192
9193 const UV src_len = _invlist_len(src);
9194 const bool src_offset = *get_invlist_offset_addr(src);
9195 const STRLEN src_byte_len = SvLEN(src);
9196 char * array = SvPVX(src);
9197
9198 const int oldtainted = TAINT_get;
9199
9200 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
9201
9202 assert(is_invlist(src));
9203 assert(is_invlist(dest));
9204 assert(! invlist_is_iterating(src));
9205 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
9206
9207 /* Make sure it ends in the right place with a NUL, as our inversion list
9208 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
9209 * asserts it */
9210 array[src_byte_len - 1] = '\0';
9211
9212 TAINT_NOT; /* Otherwise it breaks */
9213 sv_usepvn_flags(dest,
9214 (char *) array,
9215 src_byte_len - 1,
9216
9217 /* This flag is documented to cause a copy to be avoided */
9218 SV_HAS_TRAILING_NUL);
9219 TAINT_set(oldtainted);
9220 SvPV_set(src, 0);
9221 SvLEN_set(src, 0);
9222 SvCUR_set(src, 0);
9223
9224 /* Finish up copying over the other fields in an inversion list */
9225 *get_invlist_offset_addr(dest) = src_offset;
9226 invlist_set_len(dest, src_len, src_offset);
9227 *get_invlist_previous_index_addr(dest) = 0;
9228 invlist_iterfinish(dest);
9229 }
9230
9231 PERL_STATIC_INLINE IV*
S_get_invlist_previous_index_addr(SV * invlist)9232 S_get_invlist_previous_index_addr(SV* invlist)
9233 {
9234 /* Return the address of the IV that is reserved to hold the cached index
9235 * */
9236 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
9237
9238 assert(is_invlist(invlist));
9239
9240 return &(((XINVLIST*) SvANY(invlist))->prev_index);
9241 }
9242
9243 PERL_STATIC_INLINE IV
S_invlist_previous_index(SV * const invlist)9244 S_invlist_previous_index(SV* const invlist)
9245 {
9246 /* Returns cached index of previous search */
9247
9248 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
9249
9250 return *get_invlist_previous_index_addr(invlist);
9251 }
9252
9253 PERL_STATIC_INLINE void
S_invlist_set_previous_index(SV * const invlist,const IV index)9254 S_invlist_set_previous_index(SV* const invlist, const IV index)
9255 {
9256 /* Caches <index> for later retrieval */
9257
9258 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
9259
9260 assert(index == 0 || index < (int) _invlist_len(invlist));
9261
9262 *get_invlist_previous_index_addr(invlist) = index;
9263 }
9264
9265 PERL_STATIC_INLINE void
S_invlist_trim(SV * invlist)9266 S_invlist_trim(SV* invlist)
9267 {
9268 /* Free the not currently-being-used space in an inversion list */
9269
9270 /* But don't free up the space needed for the 0 UV that is always at the
9271 * beginning of the list, nor the trailing NUL */
9272 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
9273
9274 PERL_ARGS_ASSERT_INVLIST_TRIM;
9275
9276 assert(is_invlist(invlist));
9277
9278 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
9279 }
9280
9281 PERL_STATIC_INLINE void
S_invlist_clear(pTHX_ SV * invlist)9282 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
9283 {
9284 PERL_ARGS_ASSERT_INVLIST_CLEAR;
9285
9286 assert(is_invlist(invlist));
9287
9288 invlist_set_len(invlist, 0, 0);
9289 invlist_trim(invlist);
9290 }
9291
9292 #endif /* ifndef PERL_IN_XSUB_RE */
9293
9294 PERL_STATIC_INLINE bool
S_invlist_is_iterating(SV * const invlist)9295 S_invlist_is_iterating(SV* const invlist)
9296 {
9297 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
9298
9299 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
9300 }
9301
9302 #ifndef PERL_IN_XSUB_RE
9303
9304 PERL_STATIC_INLINE UV
S_invlist_max(SV * const invlist)9305 S_invlist_max(SV* const invlist)
9306 {
9307 /* Returns the maximum number of elements storable in the inversion list's
9308 * array, without having to realloc() */
9309
9310 PERL_ARGS_ASSERT_INVLIST_MAX;
9311
9312 assert(is_invlist(invlist));
9313
9314 /* Assumes worst case, in which the 0 element is not counted in the
9315 * inversion list, so subtracts 1 for that */
9316 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
9317 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
9318 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
9319 }
9320
9321 STATIC void
S_initialize_invlist_guts(pTHX_ SV * invlist,const Size_t initial_size)9322 S_initialize_invlist_guts(pTHX_ SV* invlist, const Size_t initial_size)
9323 {
9324 PERL_ARGS_ASSERT_INITIALIZE_INVLIST_GUTS;
9325
9326 /* First 1 is in case the zero element isn't in the list; second 1 is for
9327 * trailing NUL */
9328 SvGROW(invlist, TO_INTERNAL_SIZE(initial_size + 1) + 1);
9329 invlist_set_len(invlist, 0, 0);
9330
9331 /* Force iterinit() to be used to get iteration to work */
9332 invlist_iterfinish(invlist);
9333
9334 *get_invlist_previous_index_addr(invlist) = 0;
9335 SvPOK_on(invlist); /* This allows B to extract the PV */
9336 }
9337
9338 SV*
Perl__new_invlist(pTHX_ IV initial_size)9339 Perl__new_invlist(pTHX_ IV initial_size)
9340 {
9341
9342 /* Return a pointer to a newly constructed inversion list, with enough
9343 * space to store 'initial_size' elements. If that number is negative, a
9344 * system default is used instead */
9345
9346 SV* new_list;
9347
9348 if (initial_size < 0) {
9349 initial_size = 10;
9350 }
9351
9352 new_list = newSV_type(SVt_INVLIST);
9353 initialize_invlist_guts(new_list, initial_size);
9354
9355 return new_list;
9356 }
9357
9358 SV*
Perl__new_invlist_C_array(pTHX_ const UV * const list)9359 Perl__new_invlist_C_array(pTHX_ const UV* const list)
9360 {
9361 /* Return a pointer to a newly constructed inversion list, initialized to
9362 * point to <list>, which has to be in the exact correct inversion list
9363 * form, including internal fields. Thus this is a dangerous routine that
9364 * should not be used in the wrong hands. The passed in 'list' contains
9365 * several header fields at the beginning that are not part of the
9366 * inversion list body proper */
9367
9368 const STRLEN length = (STRLEN) list[0];
9369 const UV version_id = list[1];
9370 const bool offset = cBOOL(list[2]);
9371 #define HEADER_LENGTH 3
9372 /* If any of the above changes in any way, you must change HEADER_LENGTH
9373 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
9374 * perl -E 'say int(rand 2**31-1)'
9375 */
9376 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
9377 data structure type, so that one being
9378 passed in can be validated to be an
9379 inversion list of the correct vintage.
9380 */
9381
9382 SV* invlist = newSV_type(SVt_INVLIST);
9383
9384 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
9385
9386 if (version_id != INVLIST_VERSION_ID) {
9387 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
9388 }
9389
9390 /* The generated array passed in includes header elements that aren't part
9391 * of the list proper, so start it just after them */
9392 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
9393
9394 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
9395 shouldn't touch it */
9396
9397 *(get_invlist_offset_addr(invlist)) = offset;
9398
9399 /* The 'length' passed to us is the physical number of elements in the
9400 * inversion list. But if there is an offset the logical number is one
9401 * less than that */
9402 invlist_set_len(invlist, length - offset, offset);
9403
9404 invlist_set_previous_index(invlist, 0);
9405
9406 /* Initialize the iteration pointer. */
9407 invlist_iterfinish(invlist);
9408
9409 SvREADONLY_on(invlist);
9410 SvPOK_on(invlist);
9411
9412 return invlist;
9413 }
9414
9415 STATIC void
S__append_range_to_invlist(pTHX_ SV * const invlist,const UV start,const UV end)9416 S__append_range_to_invlist(pTHX_ SV* const invlist,
9417 const UV start, const UV end)
9418 {
9419 /* Subject to change or removal. Append the range from 'start' to 'end' at
9420 * the end of the inversion list. The range must be above any existing
9421 * ones. */
9422
9423 UV* array;
9424 UV max = invlist_max(invlist);
9425 UV len = _invlist_len(invlist);
9426 bool offset;
9427
9428 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
9429
9430 if (len == 0) { /* Empty lists must be initialized */
9431 offset = start != 0;
9432 array = _invlist_array_init(invlist, ! offset);
9433 }
9434 else {
9435 /* Here, the existing list is non-empty. The current max entry in the
9436 * list is generally the first value not in the set, except when the
9437 * set extends to the end of permissible values, in which case it is
9438 * the first entry in that final set, and so this call is an attempt to
9439 * append out-of-order */
9440
9441 UV final_element = len - 1;
9442 array = invlist_array(invlist);
9443 if ( array[final_element] > start
9444 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
9445 {
9446 Perl_croak(aTHX_ "panic: attempting to append to an inversion list, but wasn't at the end of the list, final=%" UVuf ", start=%" UVuf ", match=%c",
9447 array[final_element], start,
9448 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
9449 }
9450
9451 /* Here, it is a legal append. If the new range begins 1 above the end
9452 * of the range below it, it is extending the range below it, so the
9453 * new first value not in the set is one greater than the newly
9454 * extended range. */
9455 offset = *get_invlist_offset_addr(invlist);
9456 if (array[final_element] == start) {
9457 if (end != UV_MAX) {
9458 array[final_element] = end + 1;
9459 }
9460 else {
9461 /* But if the end is the maximum representable on the machine,
9462 * assume that infinity was actually what was meant. Just let
9463 * the range that this would extend to have no end */
9464 invlist_set_len(invlist, len - 1, offset);
9465 }
9466 return;
9467 }
9468 }
9469
9470 /* Here the new range doesn't extend any existing set. Add it */
9471
9472 len += 2; /* Includes an element each for the start and end of range */
9473
9474 /* If wll overflow the existing space, extend, which may cause the array to
9475 * be moved */
9476 if (max < len) {
9477 invlist_extend(invlist, len);
9478
9479 /* Have to set len here to avoid assert failure in invlist_array() */
9480 invlist_set_len(invlist, len, offset);
9481
9482 array = invlist_array(invlist);
9483 }
9484 else {
9485 invlist_set_len(invlist, len, offset);
9486 }
9487
9488 /* The next item on the list starts the range, the one after that is
9489 * one past the new range. */
9490 array[len - 2] = start;
9491 if (end != UV_MAX) {
9492 array[len - 1] = end + 1;
9493 }
9494 else {
9495 /* But if the end is the maximum representable on the machine, just let
9496 * the range have no end */
9497 invlist_set_len(invlist, len - 1, offset);
9498 }
9499 }
9500
9501 SSize_t
Perl__invlist_search(SV * const invlist,const UV cp)9502 Perl__invlist_search(SV* const invlist, const UV cp)
9503 {
9504 /* Searches the inversion list for the entry that contains the input code
9505 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
9506 * return value is the index into the list's array of the range that
9507 * contains <cp>, that is, 'i' such that
9508 * array[i] <= cp < array[i+1]
9509 */
9510
9511 IV low = 0;
9512 IV mid;
9513 IV high = _invlist_len(invlist);
9514 const IV highest_element = high - 1;
9515 const UV* array;
9516
9517 PERL_ARGS_ASSERT__INVLIST_SEARCH;
9518
9519 /* If list is empty, return failure. */
9520 if (UNLIKELY(high == 0)) {
9521 return -1;
9522 }
9523
9524 /* (We can't get the array unless we know the list is non-empty) */
9525 array = invlist_array(invlist);
9526
9527 mid = invlist_previous_index(invlist);
9528 assert(mid >=0);
9529 if (UNLIKELY(mid > highest_element)) {
9530 mid = highest_element;
9531 }
9532
9533 /* <mid> contains the cache of the result of the previous call to this
9534 * function (0 the first time). See if this call is for the same result,
9535 * or if it is for mid-1. This is under the theory that calls to this
9536 * function will often be for related code points that are near each other.
9537 * And benchmarks show that caching gives better results. We also test
9538 * here if the code point is within the bounds of the list. These tests
9539 * replace others that would have had to be made anyway to make sure that
9540 * the array bounds were not exceeded, and these give us extra information
9541 * at the same time */
9542 if (cp >= array[mid]) {
9543 if (cp >= array[highest_element]) {
9544 return highest_element;
9545 }
9546
9547 /* Here, array[mid] <= cp < array[highest_element]. This means that
9548 * the final element is not the answer, so can exclude it; it also
9549 * means that <mid> is not the final element, so can refer to 'mid + 1'
9550 * safely */
9551 if (cp < array[mid + 1]) {
9552 return mid;
9553 }
9554 high--;
9555 low = mid + 1;
9556 }
9557 else { /* cp < aray[mid] */
9558 if (cp < array[0]) { /* Fail if outside the array */
9559 return -1;
9560 }
9561 high = mid;
9562 if (cp >= array[mid - 1]) {
9563 goto found_entry;
9564 }
9565 }
9566
9567 /* Binary search. What we are looking for is <i> such that
9568 * array[i] <= cp < array[i+1]
9569 * The loop below converges on the i+1. Note that there may not be an
9570 * (i+1)th element in the array, and things work nonetheless */
9571 while (low < high) {
9572 mid = (low + high) / 2;
9573 assert(mid <= highest_element);
9574 if (array[mid] <= cp) { /* cp >= array[mid] */
9575 low = mid + 1;
9576
9577 /* We could do this extra test to exit the loop early.
9578 if (cp < array[low]) {
9579 return mid;
9580 }
9581 */
9582 }
9583 else { /* cp < array[mid] */
9584 high = mid;
9585 }
9586 }
9587
9588 found_entry:
9589 high--;
9590 invlist_set_previous_index(invlist, high);
9591 return high;
9592 }
9593
9594 void
Perl__invlist_union_maybe_complement_2nd(pTHX_ SV * const a,SV * const b,const bool complement_b,SV ** output)9595 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9596 const bool complement_b, SV** output)
9597 {
9598 /* Take the union of two inversion lists and point '*output' to it. On
9599 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9600 * even 'a' or 'b'). If to an inversion list, the contents of the original
9601 * list will be replaced by the union. The first list, 'a', may be
9602 * NULL, in which case a copy of the second list is placed in '*output'.
9603 * If 'complement_b' is TRUE, the union is taken of the complement
9604 * (inversion) of 'b' instead of b itself.
9605 *
9606 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9607 * Richard Gillam, published by Addison-Wesley, and explained at some
9608 * length there. The preface says to incorporate its examples into your
9609 * code at your own risk.
9610 *
9611 * The algorithm is like a merge sort. */
9612
9613 const UV* array_a; /* a's array */
9614 const UV* array_b;
9615 UV len_a; /* length of a's array */
9616 UV len_b;
9617
9618 SV* u; /* the resulting union */
9619 UV* array_u;
9620 UV len_u = 0;
9621
9622 UV i_a = 0; /* current index into a's array */
9623 UV i_b = 0;
9624 UV i_u = 0;
9625
9626 /* running count, as explained in the algorithm source book; items are
9627 * stopped accumulating and are output when the count changes to/from 0.
9628 * The count is incremented when we start a range that's in an input's set,
9629 * and decremented when we start a range that's not in a set. So this
9630 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9631 * and hence nothing goes into the union; 1, just one of the inputs is in
9632 * its set (and its current range gets added to the union); and 2 when both
9633 * inputs are in their sets. */
9634 UV count = 0;
9635
9636 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9637 assert(a != b);
9638 assert(*output == NULL || is_invlist(*output));
9639
9640 len_b = _invlist_len(b);
9641 if (len_b == 0) {
9642
9643 /* Here, 'b' is empty, hence it's complement is all possible code
9644 * points. So if the union includes the complement of 'b', it includes
9645 * everything, and we need not even look at 'a'. It's easiest to
9646 * create a new inversion list that matches everything. */
9647 if (complement_b) {
9648 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9649
9650 if (*output == NULL) { /* If the output didn't exist, just point it
9651 at the new list */
9652 *output = everything;
9653 }
9654 else { /* Otherwise, replace its contents with the new list */
9655 invlist_replace_list_destroys_src(*output, everything);
9656 SvREFCNT_dec_NN(everything);
9657 }
9658
9659 return;
9660 }
9661
9662 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9663 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9664 * output will be empty */
9665
9666 if (a == NULL || _invlist_len(a) == 0) {
9667 if (*output == NULL) {
9668 *output = _new_invlist(0);
9669 }
9670 else {
9671 invlist_clear(*output);
9672 }
9673 return;
9674 }
9675
9676 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9677 * union. We can just return a copy of 'a' if '*output' doesn't point
9678 * to an existing list */
9679 if (*output == NULL) {
9680 *output = invlist_clone(a, NULL);
9681 return;
9682 }
9683
9684 /* If the output is to overwrite 'a', we have a no-op, as it's
9685 * already in 'a' */
9686 if (*output == a) {
9687 return;
9688 }
9689
9690 /* Here, '*output' is to be overwritten by 'a' */
9691 u = invlist_clone(a, NULL);
9692 invlist_replace_list_destroys_src(*output, u);
9693 SvREFCNT_dec_NN(u);
9694
9695 return;
9696 }
9697
9698 /* Here 'b' is not empty. See about 'a' */
9699
9700 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9701
9702 /* Here, 'a' is empty (and b is not). That means the union will come
9703 * entirely from 'b'. If '*output' is NULL, we can directly return a
9704 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9705 * the clone */
9706
9707 SV ** dest = (*output == NULL) ? output : &u;
9708 *dest = invlist_clone(b, NULL);
9709 if (complement_b) {
9710 _invlist_invert(*dest);
9711 }
9712
9713 if (dest == &u) {
9714 invlist_replace_list_destroys_src(*output, u);
9715 SvREFCNT_dec_NN(u);
9716 }
9717
9718 return;
9719 }
9720
9721 /* Here both lists exist and are non-empty */
9722 array_a = invlist_array(a);
9723 array_b = invlist_array(b);
9724
9725 /* If are to take the union of 'a' with the complement of b, set it
9726 * up so are looking at b's complement. */
9727 if (complement_b) {
9728
9729 /* To complement, we invert: if the first element is 0, remove it. To
9730 * do this, we just pretend the array starts one later */
9731 if (array_b[0] == 0) {
9732 array_b++;
9733 len_b--;
9734 }
9735 else {
9736
9737 /* But if the first element is not zero, we pretend the list starts
9738 * at the 0 that is always stored immediately before the array. */
9739 array_b--;
9740 len_b++;
9741 }
9742 }
9743
9744 /* Size the union for the worst case: that the sets are completely
9745 * disjoint */
9746 u = _new_invlist(len_a + len_b);
9747
9748 /* Will contain U+0000 if either component does */
9749 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9750 || (len_b > 0 && array_b[0] == 0));
9751
9752 /* Go through each input list item by item, stopping when have exhausted
9753 * one of them */
9754 while (i_a < len_a && i_b < len_b) {
9755 UV cp; /* The element to potentially add to the union's array */
9756 bool cp_in_set; /* is it in the input list's set or not */
9757
9758 /* We need to take one or the other of the two inputs for the union.
9759 * Since we are merging two sorted lists, we take the smaller of the
9760 * next items. In case of a tie, we take first the one that is in its
9761 * set. If we first took the one not in its set, it would decrement
9762 * the count, possibly to 0 which would cause it to be output as ending
9763 * the range, and the next time through we would take the same number,
9764 * and output it again as beginning the next range. By doing it the
9765 * opposite way, there is no possibility that the count will be
9766 * momentarily decremented to 0, and thus the two adjoining ranges will
9767 * be seamlessly merged. (In a tie and both are in the set or both not
9768 * in the set, it doesn't matter which we take first.) */
9769 if ( array_a[i_a] < array_b[i_b]
9770 || ( array_a[i_a] == array_b[i_b]
9771 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9772 {
9773 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9774 cp = array_a[i_a++];
9775 }
9776 else {
9777 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9778 cp = array_b[i_b++];
9779 }
9780
9781 /* Here, have chosen which of the two inputs to look at. Only output
9782 * if the running count changes to/from 0, which marks the
9783 * beginning/end of a range that's in the set */
9784 if (cp_in_set) {
9785 if (count == 0) {
9786 array_u[i_u++] = cp;
9787 }
9788 count++;
9789 }
9790 else {
9791 count--;
9792 if (count == 0) {
9793 array_u[i_u++] = cp;
9794 }
9795 }
9796 }
9797
9798
9799 /* The loop above increments the index into exactly one of the input lists
9800 * each iteration, and ends when either index gets to its list end. That
9801 * means the other index is lower than its end, and so something is
9802 * remaining in that one. We decrement 'count', as explained below, if
9803 * that list is in its set. (i_a and i_b each currently index the element
9804 * beyond the one we care about.) */
9805 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9806 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9807 {
9808 count--;
9809 }
9810
9811 /* Above we decremented 'count' if the list that had unexamined elements in
9812 * it was in its set. This has made it so that 'count' being non-zero
9813 * means there isn't anything left to output; and 'count' equal to 0 means
9814 * that what is left to output is precisely that which is left in the
9815 * non-exhausted input list.
9816 *
9817 * To see why, note first that the exhausted input obviously has nothing
9818 * left to add to the union. If it was in its set at its end, that means
9819 * the set extends from here to the platform's infinity, and hence so does
9820 * the union and the non-exhausted set is irrelevant. The exhausted set
9821 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9822 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9823 * 'count' remains at 1. This is consistent with the decremented 'count'
9824 * != 0 meaning there's nothing left to add to the union.
9825 *
9826 * But if the exhausted input wasn't in its set, it contributed 0 to
9827 * 'count', and the rest of the union will be whatever the other input is.
9828 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9829 * otherwise it gets decremented to 0. This is consistent with 'count'
9830 * == 0 meaning the remainder of the union is whatever is left in the
9831 * non-exhausted list. */
9832 if (count != 0) {
9833 len_u = i_u;
9834 }
9835 else {
9836 IV copy_count = len_a - i_a;
9837 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9838 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9839 }
9840 else { /* The non-exhausted input is b */
9841 copy_count = len_b - i_b;
9842 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9843 }
9844 len_u = i_u + copy_count;
9845 }
9846
9847 /* Set the result to the final length, which can change the pointer to
9848 * array_u, so re-find it. (Note that it is unlikely that this will
9849 * change, as we are shrinking the space, not enlarging it) */
9850 if (len_u != _invlist_len(u)) {
9851 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9852 invlist_trim(u);
9853 array_u = invlist_array(u);
9854 }
9855
9856 if (*output == NULL) { /* Simply return the new inversion list */
9857 *output = u;
9858 }
9859 else {
9860 /* Otherwise, overwrite the inversion list that was in '*output'. We
9861 * could instead free '*output', and then set it to 'u', but experience
9862 * has shown [perl #127392] that if the input is a mortal, we can get a
9863 * huge build-up of these during regex compilation before they get
9864 * freed. */
9865 invlist_replace_list_destroys_src(*output, u);
9866 SvREFCNT_dec_NN(u);
9867 }
9868
9869 return;
9870 }
9871
9872 void
Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV * const a,SV * const b,const bool complement_b,SV ** i)9873 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9874 const bool complement_b, SV** i)
9875 {
9876 /* Take the intersection of two inversion lists and point '*i' to it. On
9877 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9878 * even 'a' or 'b'). If to an inversion list, the contents of the original
9879 * list will be replaced by the intersection. The first list, 'a', may be
9880 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9881 * TRUE, the result will be the intersection of 'a' and the complement (or
9882 * inversion) of 'b' instead of 'b' directly.
9883 *
9884 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9885 * Richard Gillam, published by Addison-Wesley, and explained at some
9886 * length there. The preface says to incorporate its examples into your
9887 * code at your own risk. In fact, it had bugs
9888 *
9889 * The algorithm is like a merge sort, and is essentially the same as the
9890 * union above
9891 */
9892
9893 const UV* array_a; /* a's array */
9894 const UV* array_b;
9895 UV len_a; /* length of a's array */
9896 UV len_b;
9897
9898 SV* r; /* the resulting intersection */
9899 UV* array_r;
9900 UV len_r = 0;
9901
9902 UV i_a = 0; /* current index into a's array */
9903 UV i_b = 0;
9904 UV i_r = 0;
9905
9906 /* running count of how many of the two inputs are postitioned at ranges
9907 * that are in their sets. As explained in the algorithm source book,
9908 * items are stopped accumulating and are output when the count changes
9909 * to/from 2. The count is incremented when we start a range that's in an
9910 * input's set, and decremented when we start a range that's not in a set.
9911 * Only when it is 2 are we in the intersection. */
9912 UV count = 0;
9913
9914 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9915 assert(a != b);
9916 assert(*i == NULL || is_invlist(*i));
9917
9918 /* Special case if either one is empty */
9919 len_a = (a == NULL) ? 0 : _invlist_len(a);
9920 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9921 if (len_a != 0 && complement_b) {
9922
9923 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9924 * must be empty. Here, also we are using 'b's complement, which
9925 * hence must be every possible code point. Thus the intersection
9926 * is simply 'a'. */
9927
9928 if (*i == a) { /* No-op */
9929 return;
9930 }
9931
9932 if (*i == NULL) {
9933 *i = invlist_clone(a, NULL);
9934 return;
9935 }
9936
9937 r = invlist_clone(a, NULL);
9938 invlist_replace_list_destroys_src(*i, r);
9939 SvREFCNT_dec_NN(r);
9940 return;
9941 }
9942
9943 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9944 * intersection must be empty */
9945 if (*i == NULL) {
9946 *i = _new_invlist(0);
9947 return;
9948 }
9949
9950 invlist_clear(*i);
9951 return;
9952 }
9953
9954 /* Here both lists exist and are non-empty */
9955 array_a = invlist_array(a);
9956 array_b = invlist_array(b);
9957
9958 /* If are to take the intersection of 'a' with the complement of b, set it
9959 * up so are looking at b's complement. */
9960 if (complement_b) {
9961
9962 /* To complement, we invert: if the first element is 0, remove it. To
9963 * do this, we just pretend the array starts one later */
9964 if (array_b[0] == 0) {
9965 array_b++;
9966 len_b--;
9967 }
9968 else {
9969
9970 /* But if the first element is not zero, we pretend the list starts
9971 * at the 0 that is always stored immediately before the array. */
9972 array_b--;
9973 len_b++;
9974 }
9975 }
9976
9977 /* Size the intersection for the worst case: that the intersection ends up
9978 * fragmenting everything to be completely disjoint */
9979 r= _new_invlist(len_a + len_b);
9980
9981 /* Will contain U+0000 iff both components do */
9982 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9983 && len_b > 0 && array_b[0] == 0);
9984
9985 /* Go through each list item by item, stopping when have exhausted one of
9986 * them */
9987 while (i_a < len_a && i_b < len_b) {
9988 UV cp; /* The element to potentially add to the intersection's
9989 array */
9990 bool cp_in_set; /* Is it in the input list's set or not */
9991
9992 /* We need to take one or the other of the two inputs for the
9993 * intersection. Since we are merging two sorted lists, we take the
9994 * smaller of the next items. In case of a tie, we take first the one
9995 * that is not in its set (a difference from the union algorithm). If
9996 * we first took the one in its set, it would increment the count,
9997 * possibly to 2 which would cause it to be output as starting a range
9998 * in the intersection, and the next time through we would take that
9999 * same number, and output it again as ending the set. By doing the
10000 * opposite of this, there is no possibility that the count will be
10001 * momentarily incremented to 2. (In a tie and both are in the set or
10002 * both not in the set, it doesn't matter which we take first.) */
10003 if ( array_a[i_a] < array_b[i_b]
10004 || ( array_a[i_a] == array_b[i_b]
10005 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
10006 {
10007 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
10008 cp = array_a[i_a++];
10009 }
10010 else {
10011 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
10012 cp= array_b[i_b++];
10013 }
10014
10015 /* Here, have chosen which of the two inputs to look at. Only output
10016 * if the running count changes to/from 2, which marks the
10017 * beginning/end of a range that's in the intersection */
10018 if (cp_in_set) {
10019 count++;
10020 if (count == 2) {
10021 array_r[i_r++] = cp;
10022 }
10023 }
10024 else {
10025 if (count == 2) {
10026 array_r[i_r++] = cp;
10027 }
10028 count--;
10029 }
10030
10031 }
10032
10033 /* The loop above increments the index into exactly one of the input lists
10034 * each iteration, and ends when either index gets to its list end. That
10035 * means the other index is lower than its end, and so something is
10036 * remaining in that one. We increment 'count', as explained below, if the
10037 * exhausted list was in its set. (i_a and i_b each currently index the
10038 * element beyond the one we care about.) */
10039 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
10040 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
10041 {
10042 count++;
10043 }
10044
10045 /* Above we incremented 'count' if the exhausted list was in its set. This
10046 * has made it so that 'count' being below 2 means there is nothing left to
10047 * output; otheriwse what's left to add to the intersection is precisely
10048 * that which is left in the non-exhausted input list.
10049 *
10050 * To see why, note first that the exhausted input obviously has nothing
10051 * left to affect the intersection. If it was in its set at its end, that
10052 * means the set extends from here to the platform's infinity, and hence
10053 * anything in the non-exhausted's list will be in the intersection, and
10054 * anything not in it won't be. Hence, the rest of the intersection is
10055 * precisely what's in the non-exhausted list The exhausted set also
10056 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
10057 * it means 'count' is now at least 2. This is consistent with the
10058 * incremented 'count' being >= 2 means to add the non-exhausted list to
10059 * the intersection.
10060 *
10061 * But if the exhausted input wasn't in its set, it contributed 0 to
10062 * 'count', and the intersection can't include anything further; the
10063 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
10064 * incremented. This is consistent with 'count' being < 2 meaning nothing
10065 * further to add to the intersection. */
10066 if (count < 2) { /* Nothing left to put in the intersection. */
10067 len_r = i_r;
10068 }
10069 else { /* copy the non-exhausted list, unchanged. */
10070 IV copy_count = len_a - i_a;
10071 if (copy_count > 0) { /* a is the one with stuff left */
10072 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
10073 }
10074 else { /* b is the one with stuff left */
10075 copy_count = len_b - i_b;
10076 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
10077 }
10078 len_r = i_r + copy_count;
10079 }
10080
10081 /* Set the result to the final length, which can change the pointer to
10082 * array_r, so re-find it. (Note that it is unlikely that this will
10083 * change, as we are shrinking the space, not enlarging it) */
10084 if (len_r != _invlist_len(r)) {
10085 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
10086 invlist_trim(r);
10087 array_r = invlist_array(r);
10088 }
10089
10090 if (*i == NULL) { /* Simply return the calculated intersection */
10091 *i = r;
10092 }
10093 else { /* Otherwise, replace the existing inversion list in '*i'. We could
10094 instead free '*i', and then set it to 'r', but experience has
10095 shown [perl #127392] that if the input is a mortal, we can get a
10096 huge build-up of these during regex compilation before they get
10097 freed. */
10098 if (len_r) {
10099 invlist_replace_list_destroys_src(*i, r);
10100 }
10101 else {
10102 invlist_clear(*i);
10103 }
10104 SvREFCNT_dec_NN(r);
10105 }
10106
10107 return;
10108 }
10109
10110 SV*
Perl__add_range_to_invlist(pTHX_ SV * invlist,UV start,UV end)10111 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
10112 {
10113 /* Add the range from 'start' to 'end' inclusive to the inversion list's
10114 * set. A pointer to the inversion list is returned. This may actually be
10115 * a new list, in which case the passed in one has been destroyed. The
10116 * passed-in inversion list can be NULL, in which case a new one is created
10117 * with just the one range in it. The new list is not necessarily
10118 * NUL-terminated. Space is not freed if the inversion list shrinks as a
10119 * result of this function. The gain would not be large, and in many
10120 * cases, this is called multiple times on a single inversion list, so
10121 * anything freed may almost immediately be needed again.
10122 *
10123 * This used to mostly call the 'union' routine, but that is much more
10124 * heavyweight than really needed for a single range addition */
10125
10126 UV* array; /* The array implementing the inversion list */
10127 UV len; /* How many elements in 'array' */
10128 SSize_t i_s; /* index into the invlist array where 'start'
10129 should go */
10130 SSize_t i_e = 0; /* And the index where 'end' should go */
10131 UV cur_highest; /* The highest code point in the inversion list
10132 upon entry to this function */
10133
10134 /* This range becomes the whole inversion list if none already existed */
10135 if (invlist == NULL) {
10136 invlist = _new_invlist(2);
10137 _append_range_to_invlist(invlist, start, end);
10138 return invlist;
10139 }
10140
10141 /* Likewise, if the inversion list is currently empty */
10142 len = _invlist_len(invlist);
10143 if (len == 0) {
10144 _append_range_to_invlist(invlist, start, end);
10145 return invlist;
10146 }
10147
10148 /* Starting here, we have to know the internals of the list */
10149 array = invlist_array(invlist);
10150
10151 /* If the new range ends higher than the current highest ... */
10152 cur_highest = invlist_highest(invlist);
10153 if (end > cur_highest) {
10154
10155 /* If the whole range is higher, we can just append it */
10156 if (start > cur_highest) {
10157 _append_range_to_invlist(invlist, start, end);
10158 return invlist;
10159 }
10160
10161 /* Otherwise, add the portion that is higher ... */
10162 _append_range_to_invlist(invlist, cur_highest + 1, end);
10163
10164 /* ... and continue on below to handle the rest. As a result of the
10165 * above append, we know that the index of the end of the range is the
10166 * final even numbered one of the array. Recall that the final element
10167 * always starts a range that extends to infinity. If that range is in
10168 * the set (meaning the set goes from here to infinity), it will be an
10169 * even index, but if it isn't in the set, it's odd, and the final
10170 * range in the set is one less, which is even. */
10171 if (end == UV_MAX) {
10172 i_e = len;
10173 }
10174 else {
10175 i_e = len - 2;
10176 }
10177 }
10178
10179 /* We have dealt with appending, now see about prepending. If the new
10180 * range starts lower than the current lowest ... */
10181 if (start < array[0]) {
10182
10183 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
10184 * Let the union code handle it, rather than having to know the
10185 * trickiness in two code places. */
10186 if (UNLIKELY(start == 0)) {
10187 SV* range_invlist;
10188
10189 range_invlist = _new_invlist(2);
10190 _append_range_to_invlist(range_invlist, start, end);
10191
10192 _invlist_union(invlist, range_invlist, &invlist);
10193
10194 SvREFCNT_dec_NN(range_invlist);
10195
10196 return invlist;
10197 }
10198
10199 /* If the whole new range comes before the first entry, and doesn't
10200 * extend it, we have to insert it as an additional range */
10201 if (end < array[0] - 1) {
10202 i_s = i_e = -1;
10203 goto splice_in_new_range;
10204 }
10205
10206 /* Here the new range adjoins the existing first range, extending it
10207 * downwards. */
10208 array[0] = start;
10209
10210 /* And continue on below to handle the rest. We know that the index of
10211 * the beginning of the range is the first one of the array */
10212 i_s = 0;
10213 }
10214 else { /* Not prepending any part of the new range to the existing list.
10215 * Find where in the list it should go. This finds i_s, such that:
10216 * invlist[i_s] <= start < array[i_s+1]
10217 */
10218 i_s = _invlist_search(invlist, start);
10219 }
10220
10221 /* At this point, any extending before the beginning of the inversion list
10222 * and/or after the end has been done. This has made it so that, in the
10223 * code below, each endpoint of the new range is either in a range that is
10224 * in the set, or is in a gap between two ranges that are. This means we
10225 * don't have to worry about exceeding the array bounds.
10226 *
10227 * Find where in the list the new range ends (but we can skip this if we
10228 * have already determined what it is, or if it will be the same as i_s,
10229 * which we already have computed) */
10230 if (i_e == 0) {
10231 i_e = (start == end)
10232 ? i_s
10233 : _invlist_search(invlist, end);
10234 }
10235
10236 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
10237 * is a range that goes to infinity there is no element at invlist[i_e+1],
10238 * so only the first relation holds. */
10239
10240 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10241
10242 /* Here, the ranges on either side of the beginning of the new range
10243 * are in the set, and this range starts in the gap between them.
10244 *
10245 * The new range extends the range above it downwards if the new range
10246 * ends at or above that range's start */
10247 const bool extends_the_range_above = ( end == UV_MAX
10248 || end + 1 >= array[i_s+1]);
10249
10250 /* The new range extends the range below it upwards if it begins just
10251 * after where that range ends */
10252 if (start == array[i_s]) {
10253
10254 /* If the new range fills the entire gap between the other ranges,
10255 * they will get merged together. Other ranges may also get
10256 * merged, depending on how many of them the new range spans. In
10257 * the general case, we do the merge later, just once, after we
10258 * figure out how many to merge. But in the case where the new
10259 * range exactly spans just this one gap (possibly extending into
10260 * the one above), we do the merge here, and an early exit. This
10261 * is done here to avoid having to special case later. */
10262 if (i_e - i_s <= 1) {
10263
10264 /* If i_e - i_s == 1, it means that the new range terminates
10265 * within the range above, and hence 'extends_the_range_above'
10266 * must be true. (If the range above it extends to infinity,
10267 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
10268 * will be 0, so no harm done.) */
10269 if (extends_the_range_above) {
10270 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
10271 invlist_set_len(invlist,
10272 len - 2,
10273 *(get_invlist_offset_addr(invlist)));
10274 return invlist;
10275 }
10276
10277 /* Here, i_e must == i_s. We keep them in sync, as they apply
10278 * to the same range, and below we are about to decrement i_s
10279 * */
10280 i_e--;
10281 }
10282
10283 /* Here, the new range is adjacent to the one below. (It may also
10284 * span beyond the range above, but that will get resolved later.)
10285 * Extend the range below to include this one. */
10286 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
10287 i_s--;
10288 start = array[i_s];
10289 }
10290 else if (extends_the_range_above) {
10291
10292 /* Here the new range only extends the range above it, but not the
10293 * one below. It merges with the one above. Again, we keep i_e
10294 * and i_s in sync if they point to the same range */
10295 if (i_e == i_s) {
10296 i_e++;
10297 }
10298 i_s++;
10299 array[i_s] = start;
10300 }
10301 }
10302
10303 /* Here, we've dealt with the new range start extending any adjoining
10304 * existing ranges.
10305 *
10306 * If the new range extends to infinity, it is now the final one,
10307 * regardless of what was there before */
10308 if (UNLIKELY(end == UV_MAX)) {
10309 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
10310 return invlist;
10311 }
10312
10313 /* If i_e started as == i_s, it has also been dealt with,
10314 * and been updated to the new i_s, which will fail the following if */
10315 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
10316
10317 /* Here, the ranges on either side of the end of the new range are in
10318 * the set, and this range ends in the gap between them.
10319 *
10320 * If this range is adjacent to (hence extends) the range above it, it
10321 * becomes part of that range; likewise if it extends the range below,
10322 * it becomes part of that range */
10323 if (end + 1 == array[i_e+1]) {
10324 i_e++;
10325 array[i_e] = start;
10326 }
10327 else if (start <= array[i_e]) {
10328 array[i_e] = end + 1;
10329 i_e--;
10330 }
10331 }
10332
10333 if (i_s == i_e) {
10334
10335 /* If the range fits entirely in an existing range (as possibly already
10336 * extended above), it doesn't add anything new */
10337 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10338 return invlist;
10339 }
10340
10341 /* Here, no part of the range is in the list. Must add it. It will
10342 * occupy 2 more slots */
10343 splice_in_new_range:
10344
10345 invlist_extend(invlist, len + 2);
10346 array = invlist_array(invlist);
10347 /* Move the rest of the array down two slots. Don't include any
10348 * trailing NUL */
10349 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
10350
10351 /* Do the actual splice */
10352 array[i_e+1] = start;
10353 array[i_e+2] = end + 1;
10354 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
10355 return invlist;
10356 }
10357
10358 /* Here the new range crossed the boundaries of a pre-existing range. The
10359 * code above has adjusted things so that both ends are in ranges that are
10360 * in the set. This means everything in between must also be in the set.
10361 * Just squash things together */
10362 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
10363 invlist_set_len(invlist,
10364 len - i_e + i_s,
10365 *(get_invlist_offset_addr(invlist)));
10366
10367 return invlist;
10368 }
10369
10370 SV*
Perl__setup_canned_invlist(pTHX_ const STRLEN size,const UV element0,UV ** other_elements_ptr)10371 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
10372 UV** other_elements_ptr)
10373 {
10374 /* Create and return an inversion list whose contents are to be populated
10375 * by the caller. The caller gives the number of elements (in 'size') and
10376 * the very first element ('element0'). This function will set
10377 * '*other_elements_ptr' to an array of UVs, where the remaining elements
10378 * are to be placed.
10379 *
10380 * Obviously there is some trust involved that the caller will properly
10381 * fill in the other elements of the array.
10382 *
10383 * (The first element needs to be passed in, as the underlying code does
10384 * things differently depending on whether it is zero or non-zero) */
10385
10386 SV* invlist = _new_invlist(size);
10387 bool offset;
10388
10389 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
10390
10391 invlist = add_cp_to_invlist(invlist, element0);
10392 offset = *get_invlist_offset_addr(invlist);
10393
10394 invlist_set_len(invlist, size, offset);
10395 *other_elements_ptr = invlist_array(invlist) + 1;
10396 return invlist;
10397 }
10398
10399 #endif
10400
10401 #ifndef PERL_IN_XSUB_RE
10402 void
Perl__invlist_invert(pTHX_ SV * const invlist)10403 Perl__invlist_invert(pTHX_ SV* const invlist)
10404 {
10405 /* Complement the input inversion list. This adds a 0 if the list didn't
10406 * have a zero; removes it otherwise. As described above, the data
10407 * structure is set up so that this is very efficient */
10408
10409 PERL_ARGS_ASSERT__INVLIST_INVERT;
10410
10411 assert(! invlist_is_iterating(invlist));
10412
10413 /* The inverse of matching nothing is matching everything */
10414 if (_invlist_len(invlist) == 0) {
10415 _append_range_to_invlist(invlist, 0, UV_MAX);
10416 return;
10417 }
10418
10419 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
10420 }
10421
10422 SV*
Perl_invlist_clone(pTHX_ SV * const invlist,SV * new_invlist)10423 Perl_invlist_clone(pTHX_ SV* const invlist, SV* new_invlist)
10424 {
10425 /* Return a new inversion list that is a copy of the input one, which is
10426 * unchanged. The new list will not be mortal even if the old one was. */
10427
10428 const STRLEN nominal_length = _invlist_len(invlist);
10429 const STRLEN physical_length = SvCUR(invlist);
10430 const bool offset = *(get_invlist_offset_addr(invlist));
10431
10432 PERL_ARGS_ASSERT_INVLIST_CLONE;
10433
10434 if (new_invlist == NULL) {
10435 new_invlist = _new_invlist(nominal_length);
10436 }
10437 else {
10438 sv_upgrade(new_invlist, SVt_INVLIST);
10439 initialize_invlist_guts(new_invlist, nominal_length);
10440 }
10441
10442 *(get_invlist_offset_addr(new_invlist)) = offset;
10443 invlist_set_len(new_invlist, nominal_length, offset);
10444 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
10445
10446 return new_invlist;
10447 }
10448
10449 #endif
10450
10451 PERL_STATIC_INLINE UV
S_invlist_lowest(SV * const invlist)10452 S_invlist_lowest(SV* const invlist)
10453 {
10454 /* Returns the lowest code point that matches an inversion list. This API
10455 * has an ambiguity, as it returns 0 under either the lowest is actually
10456 * 0, or if the list is empty. If this distinction matters to you, check
10457 * for emptiness before calling this function */
10458
10459 UV len = _invlist_len(invlist);
10460 UV *array;
10461
10462 PERL_ARGS_ASSERT_INVLIST_LOWEST;
10463
10464 if (len == 0) {
10465 return 0;
10466 }
10467
10468 array = invlist_array(invlist);
10469
10470 return array[0];
10471 }
10472
10473 STATIC SV *
S_invlist_contents(pTHX_ SV * const invlist,const bool traditional_style)10474 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
10475 {
10476 /* Get the contents of an inversion list into a string SV so that they can
10477 * be printed out. If 'traditional_style' is TRUE, it uses the format
10478 * traditionally done for debug tracing; otherwise it uses a format
10479 * suitable for just copying to the output, with blanks between ranges and
10480 * a dash between range components */
10481
10482 UV start, end;
10483 SV* output;
10484 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
10485 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
10486
10487 if (traditional_style) {
10488 output = newSVpvs("\n");
10489 }
10490 else {
10491 output = newSVpvs("");
10492 }
10493
10494 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
10495
10496 assert(! invlist_is_iterating(invlist));
10497
10498 invlist_iterinit(invlist);
10499 while (invlist_iternext(invlist, &start, &end)) {
10500 if (end == UV_MAX) {
10501 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFTY%c",
10502 start, intra_range_delimiter,
10503 inter_range_delimiter);
10504 }
10505 else if (end != start) {
10506 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
10507 start,
10508 intra_range_delimiter,
10509 end, inter_range_delimiter);
10510 }
10511 else {
10512 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
10513 start, inter_range_delimiter);
10514 }
10515 }
10516
10517 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10518 SvCUR_set(output, SvCUR(output) - 1);
10519 }
10520
10521 return output;
10522 }
10523
10524 #ifndef PERL_IN_XSUB_RE
10525 void
Perl__invlist_dump(pTHX_ PerlIO * file,I32 level,const char * const indent,SV * const invlist)10526 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10527 const char * const indent, SV* const invlist)
10528 {
10529 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10530 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10531 * the string 'indent'. The output looks like this:
10532 [0] 0x000A .. 0x000D
10533 [2] 0x0085
10534 [4] 0x2028 .. 0x2029
10535 [6] 0x3104 .. INFTY
10536 * This means that the first range of code points matched by the list are
10537 * 0xA through 0xD; the second range contains only the single code point
10538 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10539 * are used to define each range (except if the final range extends to
10540 * infinity, only a single element is needed). The array index of the
10541 * first element for the corresponding range is given in brackets. */
10542
10543 UV start, end;
10544 STRLEN count = 0;
10545
10546 PERL_ARGS_ASSERT__INVLIST_DUMP;
10547
10548 if (invlist_is_iterating(invlist)) {
10549 Perl_dump_indent(aTHX_ level, file,
10550 "%sCan't dump inversion list because is in middle of iterating\n",
10551 indent);
10552 return;
10553 }
10554
10555 invlist_iterinit(invlist);
10556 while (invlist_iternext(invlist, &start, &end)) {
10557 if (end == UV_MAX) {
10558 Perl_dump_indent(aTHX_ level, file,
10559 "%s[%" UVuf "] 0x%04" UVXf " .. INFTY\n",
10560 indent, (UV)count, start);
10561 }
10562 else if (end != start) {
10563 Perl_dump_indent(aTHX_ level, file,
10564 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10565 indent, (UV)count, start, end);
10566 }
10567 else {
10568 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10569 indent, (UV)count, start);
10570 }
10571 count += 2;
10572 }
10573 }
10574
10575 #endif
10576
10577 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10578 bool
Perl__invlistEQ(pTHX_ SV * const a,SV * const b,const bool complement_b)10579 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10580 {
10581 /* Return a boolean as to if the two passed in inversion lists are
10582 * identical. The final argument, if TRUE, says to take the complement of
10583 * the second inversion list before doing the comparison */
10584
10585 const UV len_a = _invlist_len(a);
10586 UV len_b = _invlist_len(b);
10587
10588 const UV* array_a = NULL;
10589 const UV* array_b = NULL;
10590
10591 PERL_ARGS_ASSERT__INVLISTEQ;
10592
10593 /* This code avoids accessing the arrays unless it knows the length is
10594 * non-zero */
10595
10596 if (len_a == 0) {
10597 if (len_b == 0) {
10598 return ! complement_b;
10599 }
10600 }
10601 else {
10602 array_a = invlist_array(a);
10603 }
10604
10605 if (len_b != 0) {
10606 array_b = invlist_array(b);
10607 }
10608
10609 /* If are to compare 'a' with the complement of b, set it
10610 * up so are looking at b's complement. */
10611 if (complement_b) {
10612
10613 /* The complement of nothing is everything, so <a> would have to have
10614 * just one element, starting at zero (ending at infinity) */
10615 if (len_b == 0) {
10616 return (len_a == 1 && array_a[0] == 0);
10617 }
10618 if (array_b[0] == 0) {
10619
10620 /* Otherwise, to complement, we invert. Here, the first element is
10621 * 0, just remove it. To do this, we just pretend the array starts
10622 * one later */
10623
10624 array_b++;
10625 len_b--;
10626 }
10627 else {
10628
10629 /* But if the first element is not zero, we pretend the list starts
10630 * at the 0 that is always stored immediately before the array. */
10631 array_b--;
10632 len_b++;
10633 }
10634 }
10635
10636 return len_a == len_b
10637 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10638
10639 }
10640 #endif
10641
10642 /*
10643 * As best we can, determine the characters that can match the start of
10644 * the given EXACTF-ish node. This is for use in creating ssc nodes, so there
10645 * can be false positive matches
10646 *
10647 * Returns the invlist as a new SV*; it is the caller's responsibility to
10648 * call SvREFCNT_dec() when done with it.
10649 */
10650 STATIC SV*
S_make_exactf_invlist(pTHX_ RExC_state_t * pRExC_state,regnode * node)10651 S_make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10652 {
10653 const U8 * s = (U8*)STRING(node);
10654 SSize_t bytelen = STR_LEN(node);
10655 UV uc;
10656 /* Start out big enough for 2 separate code points */
10657 SV* invlist = _new_invlist(4);
10658
10659 PERL_ARGS_ASSERT_MAKE_EXACTF_INVLIST;
10660
10661 if (! UTF) {
10662 uc = *s;
10663
10664 /* We punt and assume can match anything if the node begins
10665 * with a multi-character fold. Things are complicated. For
10666 * example, /ffi/i could match any of:
10667 * "\N{LATIN SMALL LIGATURE FFI}"
10668 * "\N{LATIN SMALL LIGATURE FF}I"
10669 * "F\N{LATIN SMALL LIGATURE FI}"
10670 * plus several other things; and making sure we have all the
10671 * possibilities is hard. */
10672 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10673 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10674 }
10675 else {
10676 /* Any Latin1 range character can potentially match any
10677 * other depending on the locale, and in Turkic locales, U+130 and
10678 * U+131 */
10679 if (OP(node) == EXACTFL) {
10680 _invlist_union(invlist, PL_Latin1, &invlist);
10681 invlist = add_cp_to_invlist(invlist,
10682 LATIN_SMALL_LETTER_DOTLESS_I);
10683 invlist = add_cp_to_invlist(invlist,
10684 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10685 }
10686 else {
10687 /* But otherwise, it matches at least itself. We can
10688 * quickly tell if it has a distinct fold, and if so,
10689 * it matches that as well */
10690 invlist = add_cp_to_invlist(invlist, uc);
10691 if (IS_IN_SOME_FOLD_L1(uc))
10692 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10693 }
10694
10695 /* Some characters match above-Latin1 ones under /i. This
10696 * is true of EXACTFL ones when the locale is UTF-8 */
10697 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10698 && (! isASCII(uc) || ! inRANGE(OP(node), EXACTFAA,
10699 EXACTFAA_NO_TRIE)))
10700 {
10701 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10702 }
10703 }
10704 }
10705 else { /* Pattern is UTF-8 */
10706 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10707 const U8* e = s + bytelen;
10708 IV fc;
10709
10710 fc = uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10711
10712 /* The only code points that aren't folded in a UTF EXACTFish
10713 * node are the problematic ones in EXACTFL nodes */
10714 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10715 /* We need to check for the possibility that this EXACTFL
10716 * node begins with a multi-char fold. Therefore we fold
10717 * the first few characters of it so that we can make that
10718 * check */
10719 U8 *d = folded;
10720 int i;
10721
10722 fc = -1;
10723 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10724 if (isASCII(*s)) {
10725 *(d++) = (U8) toFOLD(*s);
10726 if (fc < 0) { /* Save the first fold */
10727 fc = *(d-1);
10728 }
10729 s++;
10730 }
10731 else {
10732 STRLEN len;
10733 UV fold = toFOLD_utf8_safe(s, e, d, &len);
10734 if (fc < 0) { /* Save the first fold */
10735 fc = fold;
10736 }
10737 d += len;
10738 s += UTF8SKIP(s);
10739 }
10740 }
10741
10742 /* And set up so the code below that looks in this folded
10743 * buffer instead of the node's string */
10744 e = d;
10745 s = folded;
10746 }
10747
10748 /* When we reach here 's' points to the fold of the first
10749 * character(s) of the node; and 'e' points to far enough along
10750 * the folded string to be just past any possible multi-char
10751 * fold.
10752 *
10753 * Like the non-UTF case above, we punt if the node begins with a
10754 * multi-char fold */
10755
10756 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10757 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10758 }
10759 else { /* Single char fold */
10760 unsigned int k;
10761 U32 first_fold;
10762 const U32 * remaining_folds;
10763 Size_t folds_count;
10764
10765 /* It matches itself */
10766 invlist = add_cp_to_invlist(invlist, fc);
10767
10768 /* ... plus all the things that fold to it, which are found in
10769 * PL_utf8_foldclosures */
10770 folds_count = _inverse_folds(fc, &first_fold,
10771 &remaining_folds);
10772 for (k = 0; k < folds_count; k++) {
10773 UV c = (k == 0) ? first_fold : remaining_folds[k-1];
10774
10775 /* /aa doesn't allow folds between ASCII and non- */
10776 if ( inRANGE(OP(node), EXACTFAA, EXACTFAA_NO_TRIE)
10777 && isASCII(c) != isASCII(fc))
10778 {
10779 continue;
10780 }
10781
10782 invlist = add_cp_to_invlist(invlist, c);
10783 }
10784
10785 if (OP(node) == EXACTFL) {
10786
10787 /* If either [iI] are present in an EXACTFL node the above code
10788 * should have added its normal case pair, but under a Turkish
10789 * locale they could match instead the case pairs from it. Add
10790 * those as potential matches as well */
10791 if (isALPHA_FOLD_EQ(fc, 'I')) {
10792 invlist = add_cp_to_invlist(invlist,
10793 LATIN_SMALL_LETTER_DOTLESS_I);
10794 invlist = add_cp_to_invlist(invlist,
10795 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10796 }
10797 else if (fc == LATIN_SMALL_LETTER_DOTLESS_I) {
10798 invlist = add_cp_to_invlist(invlist, 'I');
10799 }
10800 else if (fc == LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE) {
10801 invlist = add_cp_to_invlist(invlist, 'i');
10802 }
10803 }
10804 }
10805 }
10806
10807 return invlist;
10808 }
10809
10810 #undef HEADER_LENGTH
10811 #undef TO_INTERNAL_SIZE
10812 #undef FROM_INTERNAL_SIZE
10813 #undef INVLIST_VERSION_ID
10814
10815 /* End of inversion list object */
10816
10817 STATIC void
S_parse_lparen_question_flags(pTHX_ RExC_state_t * pRExC_state)10818 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10819 {
10820 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10821 * constructs, and updates RExC_flags with them. On input, RExC_parse
10822 * should point to the first flag; it is updated on output to point to the
10823 * final ')' or ':'. There needs to be at least one flag, or this will
10824 * abort */
10825
10826 /* for (?g), (?gc), and (?o) warnings; warning
10827 about (?c) will warn about (?g) -- japhy */
10828
10829 #define WASTED_O 0x01
10830 #define WASTED_G 0x02
10831 #define WASTED_C 0x04
10832 #define WASTED_GC (WASTED_G|WASTED_C)
10833 I32 wastedflags = 0x00;
10834 U32 posflags = 0, negflags = 0;
10835 U32 *flagsp = &posflags;
10836 char has_charset_modifier = '\0';
10837 regex_charset cs;
10838 bool has_use_defaults = FALSE;
10839 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10840 int x_mod_count = 0;
10841
10842 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10843
10844 /* '^' as an initial flag sets certain defaults */
10845 if (UCHARAT(RExC_parse) == '^') {
10846 RExC_parse++;
10847 has_use_defaults = TRUE;
10848 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10849 cs = (toUSE_UNI_CHARSET_NOT_DEPENDS)
10850 ? REGEX_UNICODE_CHARSET
10851 : REGEX_DEPENDS_CHARSET;
10852 set_regex_charset(&RExC_flags, cs);
10853 }
10854 else {
10855 cs = get_regex_charset(RExC_flags);
10856 if ( cs == REGEX_DEPENDS_CHARSET
10857 && (toUSE_UNI_CHARSET_NOT_DEPENDS))
10858 {
10859 cs = REGEX_UNICODE_CHARSET;
10860 }
10861 }
10862
10863 while (RExC_parse < RExC_end) {
10864 /* && memCHRs("iogcmsx", *RExC_parse) */
10865 /* (?g), (?gc) and (?o) are useless here
10866 and must be globally applied -- japhy */
10867 if ((RExC_pm_flags & PMf_WILDCARD)) {
10868 if (flagsp == & negflags) {
10869 if (*RExC_parse == 'm') {
10870 RExC_parse++;
10871 /* diag_listed_as: Use of %s is not allowed in Unicode
10872 property wildcard subpatterns in regex; marked by <--
10873 HERE in m/%s/ */
10874 vFAIL("Use of modifier '-m' is not allowed in Unicode"
10875 " property wildcard subpatterns");
10876 }
10877 }
10878 else {
10879 if (*RExC_parse == 's') {
10880 goto modifier_illegal_in_wildcard;
10881 }
10882 }
10883 }
10884
10885 switch (*RExC_parse) {
10886
10887 /* Code for the imsxn flags */
10888 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10889
10890 case LOCALE_PAT_MOD:
10891 if (has_charset_modifier) {
10892 goto excess_modifier;
10893 }
10894 else if (flagsp == &negflags) {
10895 goto neg_modifier;
10896 }
10897 cs = REGEX_LOCALE_CHARSET;
10898 has_charset_modifier = LOCALE_PAT_MOD;
10899 break;
10900 case UNICODE_PAT_MOD:
10901 if (has_charset_modifier) {
10902 goto excess_modifier;
10903 }
10904 else if (flagsp == &negflags) {
10905 goto neg_modifier;
10906 }
10907 cs = REGEX_UNICODE_CHARSET;
10908 has_charset_modifier = UNICODE_PAT_MOD;
10909 break;
10910 case ASCII_RESTRICT_PAT_MOD:
10911 if (flagsp == &negflags) {
10912 goto neg_modifier;
10913 }
10914 if (has_charset_modifier) {
10915 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10916 goto excess_modifier;
10917 }
10918 /* Doubled modifier implies more restricted */
10919 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10920 }
10921 else {
10922 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10923 }
10924 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10925 break;
10926 case DEPENDS_PAT_MOD:
10927 if (has_use_defaults) {
10928 goto fail_modifiers;
10929 }
10930 else if (flagsp == &negflags) {
10931 goto neg_modifier;
10932 }
10933 else if (has_charset_modifier) {
10934 goto excess_modifier;
10935 }
10936
10937 /* The dual charset means unicode semantics if the
10938 * pattern (or target, not known until runtime) are
10939 * utf8, or something in the pattern indicates unicode
10940 * semantics */
10941 cs = (toUSE_UNI_CHARSET_NOT_DEPENDS)
10942 ? REGEX_UNICODE_CHARSET
10943 : REGEX_DEPENDS_CHARSET;
10944 has_charset_modifier = DEPENDS_PAT_MOD;
10945 break;
10946 excess_modifier:
10947 RExC_parse++;
10948 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10949 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10950 }
10951 else if (has_charset_modifier == *(RExC_parse - 1)) {
10952 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10953 *(RExC_parse - 1));
10954 }
10955 else {
10956 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10957 }
10958 NOT_REACHED; /*NOTREACHED*/
10959 neg_modifier:
10960 RExC_parse++;
10961 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10962 *(RExC_parse - 1));
10963 NOT_REACHED; /*NOTREACHED*/
10964 case GLOBAL_PAT_MOD: /* 'g' */
10965 if (RExC_pm_flags & PMf_WILDCARD) {
10966 goto modifier_illegal_in_wildcard;
10967 }
10968 /*FALLTHROUGH*/
10969 case ONCE_PAT_MOD: /* 'o' */
10970 if (ckWARN(WARN_REGEXP)) {
10971 const I32 wflagbit = *RExC_parse == 'o'
10972 ? WASTED_O
10973 : WASTED_G;
10974 if (! (wastedflags & wflagbit) ) {
10975 wastedflags |= wflagbit;
10976 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10977 vWARN5(
10978 RExC_parse + 1,
10979 "Useless (%s%c) - %suse /%c modifier",
10980 flagsp == &negflags ? "?-" : "?",
10981 *RExC_parse,
10982 flagsp == &negflags ? "don't " : "",
10983 *RExC_parse
10984 );
10985 }
10986 }
10987 break;
10988
10989 case CONTINUE_PAT_MOD: /* 'c' */
10990 if (RExC_pm_flags & PMf_WILDCARD) {
10991 goto modifier_illegal_in_wildcard;
10992 }
10993 if (ckWARN(WARN_REGEXP)) {
10994 if (! (wastedflags & WASTED_C) ) {
10995 wastedflags |= WASTED_GC;
10996 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10997 vWARN3(
10998 RExC_parse + 1,
10999 "Useless (%sc) - %suse /gc modifier",
11000 flagsp == &negflags ? "?-" : "?",
11001 flagsp == &negflags ? "don't " : ""
11002 );
11003 }
11004 }
11005 break;
11006 case KEEPCOPY_PAT_MOD: /* 'p' */
11007 if (RExC_pm_flags & PMf_WILDCARD) {
11008 goto modifier_illegal_in_wildcard;
11009 }
11010 if (flagsp == &negflags) {
11011 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
11012 } else {
11013 *flagsp |= RXf_PMf_KEEPCOPY;
11014 }
11015 break;
11016 case '-':
11017 /* A flag is a default iff it is following a minus, so
11018 * if there is a minus, it means will be trying to
11019 * re-specify a default which is an error */
11020 if (has_use_defaults || flagsp == &negflags) {
11021 goto fail_modifiers;
11022 }
11023 flagsp = &negflags;
11024 wastedflags = 0; /* reset so (?g-c) warns twice */
11025 x_mod_count = 0;
11026 break;
11027 case ':':
11028 case ')':
11029
11030 if ( (RExC_pm_flags & PMf_WILDCARD)
11031 && cs != REGEX_ASCII_MORE_RESTRICTED_CHARSET)
11032 {
11033 RExC_parse++;
11034 /* diag_listed_as: Use of %s is not allowed in Unicode
11035 property wildcard subpatterns in regex; marked by <--
11036 HERE in m/%s/ */
11037 vFAIL2("Use of modifier '%c' is not allowed in Unicode"
11038 " property wildcard subpatterns",
11039 has_charset_modifier);
11040 }
11041
11042 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
11043 negflags |= RXf_PMf_EXTENDED_MORE;
11044 }
11045 RExC_flags |= posflags;
11046
11047 if (negflags & RXf_PMf_EXTENDED) {
11048 negflags |= RXf_PMf_EXTENDED_MORE;
11049 }
11050 RExC_flags &= ~negflags;
11051 set_regex_charset(&RExC_flags, cs);
11052
11053 return;
11054 default:
11055 fail_modifiers:
11056 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11057 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11058 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
11059 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11060 NOT_REACHED; /*NOTREACHED*/
11061 }
11062
11063 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11064 }
11065
11066 vFAIL("Sequence (?... not terminated");
11067
11068 modifier_illegal_in_wildcard:
11069 RExC_parse++;
11070 /* diag_listed_as: Use of %s is not allowed in Unicode property wildcard
11071 subpatterns in regex; marked by <-- HERE in m/%s/ */
11072 vFAIL2("Use of modifier '%c' is not allowed in Unicode property wildcard"
11073 " subpatterns", *(RExC_parse - 1));
11074 }
11075
11076 /*
11077 - reg - regular expression, i.e. main body or parenthesized thing
11078 *
11079 * Caller must absorb opening parenthesis.
11080 *
11081 * Combining parenthesis handling with the base level of regular expression
11082 * is a trifle forced, but the need to tie the tails of the branches to what
11083 * follows makes it hard to avoid.
11084 */
11085 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
11086 #ifdef DEBUGGING
11087 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
11088 #else
11089 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
11090 #endif
11091
11092 STATIC regnode_offset
S_handle_named_backref(pTHX_ RExC_state_t * pRExC_state,I32 * flagp,char * parse_start,char ch)11093 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
11094 I32 *flagp,
11095 char * parse_start,
11096 char ch
11097 )
11098 {
11099 regnode_offset ret;
11100 char* name_start = RExC_parse;
11101 U32 num = 0;
11102 SV *sv_dat = reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
11103 DECLARE_AND_GET_RE_DEBUG_FLAGS;
11104
11105 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
11106
11107 if (RExC_parse != name_start && ch == '}') {
11108 while (isBLANK(*RExC_parse)) {
11109 RExC_parse++;
11110 }
11111 }
11112 if (RExC_parse == name_start || *RExC_parse != ch) {
11113 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11114 vFAIL2("Sequence %.3s... not terminated", parse_start);
11115 }
11116
11117 if (sv_dat) {
11118 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11119 RExC_rxi->data->data[num]=(void*)sv_dat;
11120 SvREFCNT_inc_simple_void_NN(sv_dat);
11121 }
11122 RExC_sawback = 1;
11123 ret = reganode(pRExC_state,
11124 ((! FOLD)
11125 ? REFN
11126 : (ASCII_FOLD_RESTRICTED)
11127 ? REFFAN
11128 : (AT_LEAST_UNI_SEMANTICS)
11129 ? REFFUN
11130 : (LOC)
11131 ? REFFLN
11132 : REFFN),
11133 num);
11134 *flagp |= HASWIDTH;
11135
11136 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
11137 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
11138
11139 nextchar(pRExC_state);
11140 return ret;
11141 }
11142
11143 /* Below are the main parsing routines.
11144 *
11145 * S_reg() parses a whole pattern or subpattern. It itself handles things
11146 * like the 'xyz' in '(?xyz:...)', and calls S_regbranch for each
11147 * alternation '|' in the '...' pattern.
11148 * S_regbranch() effectively implements the concatenation operator, handling
11149 * one alternative of '|', repeatedly calling S_regpiece on each
11150 * segment of the input.
11151 * S_regpiece() calls S_regatom to handle the next atomic chunk of the input,
11152 * and then adds any quantifier for that chunk.
11153 * S_regatom() parses the next chunk of the input, returning when it
11154 * determines it has found a complete atomic chunk. The chunk may
11155 * be a nested subpattern, in which case S_reg is called
11156 * recursively
11157 *
11158 * The functions generate regnodes as they go along, appending each to the
11159 * pattern data structure so far. They return the offset of the current final
11160 * node into that structure, or 0 on failure.
11161 *
11162 * There are three parameters common to all of them:
11163 * pRExC_state is a structure with much information about the current
11164 * state of the parse. It's easy to add new elements to
11165 * convey new information, but beware that an error return may
11166 * require clearing the element.
11167 * flagp is a pointer to bit flags set in a lower level to pass up
11168 * to higher levels information, such as the cause of a
11169 * failure, or some characteristic about the generated node
11170 * depth is roughly the recursion depth, mostly unused except for
11171 * pretty printing debugging info.
11172 *
11173 * There are ancillary functions that these may farm work out to, using the
11174 * same parameters.
11175 *
11176 * The protocol for handling flags is that each function will, before
11177 * returning, add into *flagp the flags it needs to pass up. Each function has
11178 * a second flags variable, typically named 'flags', which it sets and clears
11179 * at will. Flag bits in it are used in that function, and it calls the next
11180 * layer down with its 'flagp' parameter set to '&flags'. Thus, upon return,
11181 * 'flags' will contain whatever it had before the call, plus whatever that
11182 * function passed up. If it wants to pass any of these up to its caller, it
11183 * has to add them to its *flagp. This means that it takes extra steps to keep
11184 * passing a flag upwards, and otherwise the flag bit is cleared for higher
11185 * functions.
11186 */
11187
11188 /* On success, returns the offset at which any next node should be placed into
11189 * the regex engine program being compiled.
11190 *
11191 * Returns 0 otherwise, with *flagp set to indicate why:
11192 * TRYAGAIN at the end of (?) that only sets flags.
11193 * RESTART_PARSE if the parse needs to be restarted, or'd with
11194 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
11195 * Otherwise would only return 0 if regbranch() returns 0, which cannot
11196 * happen. */
11197 STATIC regnode_offset
S_reg(pTHX_ RExC_state_t * pRExC_state,I32 paren,I32 * flagp,U32 depth)11198 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp, U32 depth)
11199 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
11200 * 2 is like 1, but indicates that nextchar() has been called to advance
11201 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
11202 * this flag alerts us to the need to check for that */
11203 {
11204 regnode_offset ret = 0; /* Will be the head of the group. */
11205 regnode_offset br;
11206 regnode_offset lastbr;
11207 regnode_offset ender = 0;
11208 I32 parno = 0;
11209 I32 flags;
11210 U32 oregflags = RExC_flags;
11211 bool have_branch = 0;
11212 bool is_open = 0;
11213 I32 freeze_paren = 0;
11214 I32 after_freeze = 0;
11215 I32 num; /* numeric backreferences */
11216 SV * max_open; /* Max number of unclosed parens */
11217 I32 was_in_lookaround = RExC_in_lookaround;
11218
11219 char * parse_start = RExC_parse; /* MJD */
11220 char * const oregcomp_parse = RExC_parse;
11221
11222 DECLARE_AND_GET_RE_DEBUG_FLAGS;
11223
11224 PERL_ARGS_ASSERT_REG;
11225 DEBUG_PARSE("reg ");
11226
11227 max_open = get_sv(RE_COMPILE_RECURSION_LIMIT, GV_ADD);
11228 assert(max_open);
11229 if (!SvIOK(max_open)) {
11230 sv_setiv(max_open, RE_COMPILE_RECURSION_INIT);
11231 }
11232 if (depth > 4 * (UV) SvIV(max_open)) { /* We increase depth by 4 for each
11233 open paren */
11234 vFAIL("Too many nested open parens");
11235 }
11236
11237 *flagp = 0; /* Initialize. */
11238
11239 /* Having this true makes it feasible to have a lot fewer tests for the
11240 * parse pointer being in scope. For example, we can write
11241 * while(isFOO(*RExC_parse)) RExC_parse++;
11242 * instead of
11243 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
11244 */
11245 assert(*RExC_end == '\0');
11246
11247 /* Make an OPEN node, if parenthesized. */
11248 if (paren) {
11249
11250 /* Under /x, space and comments can be gobbled up between the '(' and
11251 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
11252 * intervening space, as the sequence is a token, and a token should be
11253 * indivisible */
11254 bool has_intervening_patws = (paren == 2)
11255 && *(RExC_parse - 1) != '(';
11256
11257 if (RExC_parse >= RExC_end) {
11258 vFAIL("Unmatched (");
11259 }
11260
11261 if (paren == 'r') { /* Atomic script run */
11262 paren = '>';
11263 goto parse_rest;
11264 }
11265 else if ( *RExC_parse == '*') { /* (*VERB:ARG), (*construct:...) */
11266 char *start_verb = RExC_parse + 1;
11267 STRLEN verb_len;
11268 char *start_arg = NULL;
11269 unsigned char op = 0;
11270 int arg_required = 0;
11271 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
11272 bool has_upper = FALSE;
11273
11274 if (has_intervening_patws) {
11275 RExC_parse++; /* past the '*' */
11276
11277 /* For strict backwards compatibility, don't change the message
11278 * now that we also have lowercase operands */
11279 if (isUPPER(*RExC_parse)) {
11280 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
11281 }
11282 else {
11283 vFAIL("In '(*...)', the '(' and '*' must be adjacent");
11284 }
11285 }
11286 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
11287 if ( *RExC_parse == ':' ) {
11288 start_arg = RExC_parse + 1;
11289 break;
11290 }
11291 else if (! UTF) {
11292 if (isUPPER(*RExC_parse)) {
11293 has_upper = TRUE;
11294 }
11295 RExC_parse++;
11296 }
11297 else {
11298 RExC_parse += UTF8SKIP(RExC_parse);
11299 }
11300 }
11301 verb_len = RExC_parse - start_verb;
11302 if ( start_arg ) {
11303 if (RExC_parse >= RExC_end) {
11304 goto unterminated_verb_pattern;
11305 }
11306
11307 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11308 while ( RExC_parse < RExC_end && *RExC_parse != ')' ) {
11309 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11310 }
11311 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11312 unterminated_verb_pattern:
11313 if (has_upper) {
11314 vFAIL("Unterminated verb pattern argument");
11315 }
11316 else {
11317 vFAIL("Unterminated '(*...' argument");
11318 }
11319 }
11320 } else {
11321 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11322 if (has_upper) {
11323 vFAIL("Unterminated verb pattern");
11324 }
11325 else {
11326 vFAIL("Unterminated '(*...' construct");
11327 }
11328 }
11329 }
11330
11331 /* Here, we know that RExC_parse < RExC_end */
11332
11333 switch ( *start_verb ) {
11334 case 'A': /* (*ACCEPT) */
11335 if ( memEQs(start_verb, verb_len,"ACCEPT") ) {
11336 op = ACCEPT;
11337 internal_argval = RExC_nestroot;
11338 }
11339 break;
11340 case 'C': /* (*COMMIT) */
11341 if ( memEQs(start_verb, verb_len,"COMMIT") )
11342 op = COMMIT;
11343 break;
11344 case 'F': /* (*FAIL) */
11345 if ( verb_len==1 || memEQs(start_verb, verb_len,"FAIL") ) {
11346 op = OPFAIL;
11347 }
11348 break;
11349 case ':': /* (*:NAME) */
11350 case 'M': /* (*MARK:NAME) */
11351 if ( verb_len==0 || memEQs(start_verb, verb_len,"MARK") ) {
11352 op = MARKPOINT;
11353 arg_required = 1;
11354 }
11355 break;
11356 case 'P': /* (*PRUNE) */
11357 if ( memEQs(start_verb, verb_len,"PRUNE") )
11358 op = PRUNE;
11359 break;
11360 case 'S': /* (*SKIP) */
11361 if ( memEQs(start_verb, verb_len,"SKIP") )
11362 op = SKIP;
11363 break;
11364 case 'T': /* (*THEN) */
11365 /* [19:06] <TimToady> :: is then */
11366 if ( memEQs(start_verb, verb_len,"THEN") ) {
11367 op = CUTGROUP;
11368 RExC_seen |= REG_CUTGROUP_SEEN;
11369 }
11370 break;
11371 case 'a':
11372 if ( memEQs(start_verb, verb_len, "asr")
11373 || memEQs(start_verb, verb_len, "atomic_script_run"))
11374 {
11375 paren = 'r'; /* Mnemonic: recursed run */
11376 goto script_run;
11377 }
11378 else if (memEQs(start_verb, verb_len, "atomic")) {
11379 paren = 't'; /* AtOMIC */
11380 goto alpha_assertions;
11381 }
11382 break;
11383 case 'p':
11384 if ( memEQs(start_verb, verb_len, "plb")
11385 || memEQs(start_verb, verb_len, "positive_lookbehind"))
11386 {
11387 paren = 'b';
11388 goto lookbehind_alpha_assertions;
11389 }
11390 else if ( memEQs(start_verb, verb_len, "pla")
11391 || memEQs(start_verb, verb_len, "positive_lookahead"))
11392 {
11393 paren = 'a';
11394 goto alpha_assertions;
11395 }
11396 break;
11397 case 'n':
11398 if ( memEQs(start_verb, verb_len, "nlb")
11399 || memEQs(start_verb, verb_len, "negative_lookbehind"))
11400 {
11401 paren = 'B';
11402 goto lookbehind_alpha_assertions;
11403 }
11404 else if ( memEQs(start_verb, verb_len, "nla")
11405 || memEQs(start_verb, verb_len, "negative_lookahead"))
11406 {
11407 paren = 'A';
11408 goto alpha_assertions;
11409 }
11410 break;
11411 case 's':
11412 if ( memEQs(start_verb, verb_len, "sr")
11413 || memEQs(start_verb, verb_len, "script_run"))
11414 {
11415 regnode_offset atomic;
11416
11417 paren = 's';
11418
11419 script_run:
11420
11421 /* This indicates Unicode rules. */
11422 REQUIRE_UNI_RULES(flagp, 0);
11423
11424 if (! start_arg) {
11425 goto no_colon;
11426 }
11427
11428 RExC_parse = start_arg;
11429
11430 if (RExC_in_script_run) {
11431
11432 /* Nested script runs are treated as no-ops, because
11433 * if the nested one fails, the outer one must as
11434 * well. It could fail sooner, and avoid (??{} with
11435 * side effects, but that is explicitly documented as
11436 * undefined behavior. */
11437
11438 ret = 0;
11439
11440 if (paren == 's') {
11441 paren = ':';
11442 goto parse_rest;
11443 }
11444
11445 /* But, the atomic part of a nested atomic script run
11446 * isn't a no-op, but can be treated just like a '(?>'
11447 * */
11448 paren = '>';
11449 goto parse_rest;
11450 }
11451
11452 if (paren == 's') {
11453 /* Here, we're starting a new regular script run */
11454 ret = reg_node(pRExC_state, SROPEN);
11455 RExC_in_script_run = 1;
11456 is_open = 1;
11457 goto parse_rest;
11458 }
11459
11460 /* Here, we are starting an atomic script run. This is
11461 * handled by recursing to deal with the atomic portion
11462 * separately, enclosed in SROPEN ... SRCLOSE nodes */
11463
11464 ret = reg_node(pRExC_state, SROPEN);
11465
11466 RExC_in_script_run = 1;
11467
11468 atomic = reg(pRExC_state, 'r', &flags, depth);
11469 if (flags & (RESTART_PARSE|NEED_UTF8)) {
11470 *flagp = flags & (RESTART_PARSE|NEED_UTF8);
11471 return 0;
11472 }
11473
11474 if (! REGTAIL(pRExC_state, ret, atomic)) {
11475 REQUIRE_BRANCHJ(flagp, 0);
11476 }
11477
11478 if (! REGTAIL(pRExC_state, atomic, reg_node(pRExC_state,
11479 SRCLOSE)))
11480 {
11481 REQUIRE_BRANCHJ(flagp, 0);
11482 }
11483
11484 RExC_in_script_run = 0;
11485 return ret;
11486 }
11487
11488 break;
11489
11490 lookbehind_alpha_assertions:
11491 RExC_seen |= REG_LOOKBEHIND_SEEN;
11492 /*FALLTHROUGH*/
11493
11494 alpha_assertions:
11495
11496 RExC_in_lookaround++;
11497 RExC_seen_zerolen++;
11498
11499 if (! start_arg) {
11500 goto no_colon;
11501 }
11502
11503 /* An empty negative lookahead assertion simply is failure */
11504 if (paren == 'A' && RExC_parse == start_arg) {
11505 ret=reganode(pRExC_state, OPFAIL, 0);
11506 nextchar(pRExC_state);
11507 return ret;
11508 }
11509
11510 RExC_parse = start_arg;
11511 goto parse_rest;
11512
11513 no_colon:
11514 vFAIL2utf8f(
11515 "'(*%" UTF8f "' requires a terminating ':'",
11516 UTF8fARG(UTF, verb_len, start_verb));
11517 NOT_REACHED; /*NOTREACHED*/
11518
11519 } /* End of switch */
11520 if ( ! op ) {
11521 RExC_parse += UTF
11522 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
11523 : 1;
11524 if (has_upper || verb_len == 0) {
11525 vFAIL2utf8f(
11526 "Unknown verb pattern '%" UTF8f "'",
11527 UTF8fARG(UTF, verb_len, start_verb));
11528 }
11529 else {
11530 vFAIL2utf8f(
11531 "Unknown '(*...)' construct '%" UTF8f "'",
11532 UTF8fARG(UTF, verb_len, start_verb));
11533 }
11534 }
11535 if ( RExC_parse == start_arg ) {
11536 start_arg = NULL;
11537 }
11538 if ( arg_required && !start_arg ) {
11539 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
11540 (int) verb_len, start_verb);
11541 }
11542 if (internal_argval == -1) {
11543 ret = reganode(pRExC_state, op, 0);
11544 } else {
11545 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
11546 }
11547 RExC_seen |= REG_VERBARG_SEEN;
11548 if (start_arg) {
11549 SV *sv = newSVpvn( start_arg,
11550 RExC_parse - start_arg);
11551 ARG(REGNODE_p(ret)) = add_data( pRExC_state,
11552 STR_WITH_LEN("S"));
11553 RExC_rxi->data->data[ARG(REGNODE_p(ret))]=(void*)sv;
11554 FLAGS(REGNODE_p(ret)) = 1;
11555 } else {
11556 FLAGS(REGNODE_p(ret)) = 0;
11557 }
11558 if ( internal_argval != -1 )
11559 ARG2L_SET(REGNODE_p(ret), internal_argval);
11560 nextchar(pRExC_state);
11561 return ret;
11562 }
11563 else if (*RExC_parse == '?') { /* (?...) */
11564 bool is_logical = 0;
11565 const char * const seqstart = RExC_parse;
11566 const char * endptr;
11567 const char non_existent_group_msg[]
11568 = "Reference to nonexistent group";
11569 const char impossible_group[] = "Invalid reference to group";
11570
11571 if (has_intervening_patws) {
11572 RExC_parse++;
11573 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
11574 }
11575
11576 RExC_parse++; /* past the '?' */
11577 paren = *RExC_parse; /* might be a trailing NUL, if not
11578 well-formed */
11579 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11580 if (RExC_parse > RExC_end) {
11581 paren = '\0';
11582 }
11583 ret = 0; /* For look-ahead/behind. */
11584 switch (paren) {
11585
11586 case 'P': /* (?P...) variants for those used to PCRE/Python */
11587 paren = *RExC_parse;
11588 if ( paren == '<') { /* (?P<...>) named capture */
11589 RExC_parse++;
11590 if (RExC_parse >= RExC_end) {
11591 vFAIL("Sequence (?P<... not terminated");
11592 }
11593 goto named_capture;
11594 }
11595 else if (paren == '>') { /* (?P>name) named recursion */
11596 RExC_parse++;
11597 if (RExC_parse >= RExC_end) {
11598 vFAIL("Sequence (?P>... not terminated");
11599 }
11600 goto named_recursion;
11601 }
11602 else if (paren == '=') { /* (?P=...) named backref */
11603 RExC_parse++;
11604 return handle_named_backref(pRExC_state, flagp,
11605 parse_start, ')');
11606 }
11607 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11608 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11609 vFAIL3("Sequence (%.*s...) not recognized",
11610 (int) (RExC_parse - seqstart), seqstart);
11611 NOT_REACHED; /*NOTREACHED*/
11612 case '<': /* (?<...) */
11613 /* If you want to support (?<*...), first reconcile with GH #17363 */
11614 if (*RExC_parse == '!')
11615 paren = ',';
11616 else if (*RExC_parse != '=')
11617 named_capture:
11618 { /* (?<...>) */
11619 char *name_start;
11620 SV *svname;
11621 paren= '>';
11622 /* FALLTHROUGH */
11623 case '\'': /* (?'...') */
11624 name_start = RExC_parse;
11625 svname = reg_scan_name(pRExC_state, REG_RSN_RETURN_NAME);
11626 if ( RExC_parse == name_start
11627 || RExC_parse >= RExC_end
11628 || *RExC_parse != paren)
11629 {
11630 vFAIL2("Sequence (?%c... not terminated",
11631 paren=='>' ? '<' : (char) paren);
11632 }
11633 {
11634 HE *he_str;
11635 SV *sv_dat = NULL;
11636 if (!svname) /* shouldn't happen */
11637 Perl_croak(aTHX_
11638 "panic: reg_scan_name returned NULL");
11639 if (!RExC_paren_names) {
11640 RExC_paren_names= newHV();
11641 sv_2mortal(MUTABLE_SV(RExC_paren_names));
11642 #ifdef DEBUGGING
11643 RExC_paren_name_list= newAV();
11644 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
11645 #endif
11646 }
11647 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
11648 if ( he_str )
11649 sv_dat = HeVAL(he_str);
11650 if ( ! sv_dat ) {
11651 /* croak baby croak */
11652 Perl_croak(aTHX_
11653 "panic: paren_name hash element allocation failed");
11654 } else if ( SvPOK(sv_dat) ) {
11655 /* (?|...) can mean we have dupes so scan to check
11656 its already been stored. Maybe a flag indicating
11657 we are inside such a construct would be useful,
11658 but the arrays are likely to be quite small, so
11659 for now we punt -- dmq */
11660 IV count = SvIV(sv_dat);
11661 I32 *pv = (I32*)SvPVX(sv_dat);
11662 IV i;
11663 for ( i = 0 ; i < count ; i++ ) {
11664 if ( pv[i] == RExC_npar ) {
11665 count = 0;
11666 break;
11667 }
11668 }
11669 if ( count ) {
11670 pv = (I32*)SvGROW(sv_dat,
11671 SvCUR(sv_dat) + sizeof(I32)+1);
11672 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
11673 pv[count] = RExC_npar;
11674 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
11675 }
11676 } else {
11677 (void)SvUPGRADE(sv_dat, SVt_PVNV);
11678 sv_setpvn(sv_dat, (char *)&(RExC_npar),
11679 sizeof(I32));
11680 SvIOK_on(sv_dat);
11681 SvIV_set(sv_dat, 1);
11682 }
11683 #ifdef DEBUGGING
11684 /* Yes this does cause a memory leak in debugging Perls
11685 * */
11686 if (!av_store(RExC_paren_name_list,
11687 RExC_npar, SvREFCNT_inc_NN(svname)))
11688 SvREFCNT_dec_NN(svname);
11689 #endif
11690
11691 /*sv_dump(sv_dat);*/
11692 }
11693 nextchar(pRExC_state);
11694 paren = 1;
11695 goto capturing_parens;
11696 }
11697
11698 RExC_seen |= REG_LOOKBEHIND_SEEN;
11699 RExC_in_lookaround++;
11700 RExC_parse++;
11701 if (RExC_parse >= RExC_end) {
11702 vFAIL("Sequence (?... not terminated");
11703 }
11704 RExC_seen_zerolen++;
11705 break;
11706 case '=': /* (?=...) */
11707 RExC_seen_zerolen++;
11708 RExC_in_lookaround++;
11709 break;
11710 case '!': /* (?!...) */
11711 RExC_seen_zerolen++;
11712 /* check if we're really just a "FAIL" assertion */
11713 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11714 FALSE /* Don't force to /x */ );
11715 if (*RExC_parse == ')') {
11716 ret=reganode(pRExC_state, OPFAIL, 0);
11717 nextchar(pRExC_state);
11718 return ret;
11719 }
11720 RExC_in_lookaround++;
11721 break;
11722 case '|': /* (?|...) */
11723 /* branch reset, behave like a (?:...) except that
11724 buffers in alternations share the same numbers */
11725 paren = ':';
11726 after_freeze = freeze_paren = RExC_npar;
11727
11728 /* XXX This construct currently requires an extra pass.
11729 * Investigation would be required to see if that could be
11730 * changed */
11731 REQUIRE_PARENS_PASS;
11732 break;
11733 case ':': /* (?:...) */
11734 case '>': /* (?>...) */
11735 break;
11736 case '$': /* (?$...) */
11737 case '@': /* (?@...) */
11738 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
11739 break;
11740 case '0' : /* (?0) */
11741 case 'R' : /* (?R) */
11742 if (RExC_parse == RExC_end || *RExC_parse != ')')
11743 FAIL("Sequence (?R) not terminated");
11744 num = 0;
11745 RExC_seen |= REG_RECURSE_SEEN;
11746
11747 /* XXX These constructs currently require an extra pass.
11748 * It probably could be changed */
11749 REQUIRE_PARENS_PASS;
11750
11751 *flagp |= POSTPONED;
11752 goto gen_recurse_regop;
11753 /*notreached*/
11754 /* named and numeric backreferences */
11755 case '&': /* (?&NAME) */
11756 parse_start = RExC_parse - 1;
11757 named_recursion:
11758 {
11759 SV *sv_dat = reg_scan_name(pRExC_state,
11760 REG_RSN_RETURN_DATA);
11761 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
11762 }
11763 if (RExC_parse >= RExC_end || *RExC_parse != ')')
11764 vFAIL("Sequence (?&... not terminated");
11765 goto gen_recurse_regop;
11766 /* NOTREACHED */
11767 case '+':
11768 if (! inRANGE(RExC_parse[0], '1', '9')) {
11769 RExC_parse++;
11770 vFAIL("Illegal pattern");
11771 }
11772 goto parse_recursion;
11773 /* NOTREACHED*/
11774 case '-': /* (?-1) */
11775 if (! inRANGE(RExC_parse[0], '1', '9')) {
11776 RExC_parse--; /* rewind to let it be handled later */
11777 goto parse_flags;
11778 }
11779 /* FALLTHROUGH */
11780 case '1': case '2': case '3': case '4': /* (?1) */
11781 case '5': case '6': case '7': case '8': case '9':
11782 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
11783 parse_recursion:
11784 {
11785 bool is_neg = FALSE;
11786 UV unum;
11787 parse_start = RExC_parse - 1; /* MJD */
11788 if (*RExC_parse == '-') {
11789 RExC_parse++;
11790 is_neg = TRUE;
11791 }
11792 endptr = RExC_end;
11793 if (grok_atoUV(RExC_parse, &unum, &endptr)
11794 && unum <= I32_MAX
11795 ) {
11796 num = (I32)unum;
11797 RExC_parse = (char*)endptr;
11798 }
11799 else { /* Overflow, or something like that. Position
11800 beyond all digits for the message */
11801 while (RExC_parse < RExC_end && isDIGIT(*RExC_parse)) {
11802 RExC_parse++;
11803 }
11804 vFAIL(impossible_group);
11805 }
11806 if (is_neg) {
11807 /* -num is always representable on 1 and 2's complement
11808 * machines */
11809 num = -num;
11810 }
11811 }
11812 if (*RExC_parse!=')')
11813 vFAIL("Expecting close bracket");
11814
11815 gen_recurse_regop:
11816 if (paren == '-' || paren == '+') {
11817
11818 /* Don't overflow */
11819 if (UNLIKELY(I32_MAX - RExC_npar < num)) {
11820 RExC_parse++;
11821 vFAIL(impossible_group);
11822 }
11823
11824 /*
11825 Diagram of capture buffer numbering.
11826 Top line is the normal capture buffer numbers
11827 Bottom line is the negative indexing as from
11828 the X (the (?-2))
11829
11830 1 2 3 4 5 X Y 6 7
11831 /(a(x)y)(a(b(c(?+2)d)e)f)(g(h))/
11832 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
11833 - 5 4 3 2 1 X Y x x
11834
11835 Resolve to absolute group. Recall that RExC_npar is +1 of
11836 the actual parenthesis group number. For lookahead, we
11837 have to compensate for that. Using the above example, when
11838 we get to Y in the parse, num is 2 and RExC_npar is 6. We
11839 want 7 for +2, and 4 for -2.
11840 */
11841 if ( paren == '+' ) {
11842 num--;
11843 }
11844
11845 num += RExC_npar;
11846
11847 if (paren == '-' && num < 1) {
11848 RExC_parse++;
11849 vFAIL(non_existent_group_msg);
11850 }
11851 }
11852
11853 if (num >= RExC_npar) {
11854
11855 /* It might be a forward reference; we can't fail until we
11856 * know, by completing the parse to get all the groups, and
11857 * then reparsing */
11858 if (ALL_PARENS_COUNTED) {
11859 if (num >= RExC_total_parens) {
11860 RExC_parse++;
11861 vFAIL(non_existent_group_msg);
11862 }
11863 }
11864 else {
11865 REQUIRE_PARENS_PASS;
11866 }
11867 }
11868
11869 /* We keep track how many GOSUB items we have produced.
11870 To start off the ARG2L() of the GOSUB holds its "id",
11871 which is used later in conjunction with RExC_recurse
11872 to calculate the offset we need to jump for the GOSUB,
11873 which it will store in the final representation.
11874 We have to defer the actual calculation until much later
11875 as the regop may move.
11876 */
11877 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
11878 RExC_recurse_count++;
11879 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11880 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
11881 22, "| |", (int)(depth * 2 + 1), "",
11882 (UV)ARG(REGNODE_p(ret)),
11883 (IV)ARG2L(REGNODE_p(ret))));
11884 RExC_seen |= REG_RECURSE_SEEN;
11885
11886 Set_Node_Length(REGNODE_p(ret),
11887 1 + regarglen[OP(REGNODE_p(ret))]); /* MJD */
11888 Set_Node_Offset(REGNODE_p(ret), parse_start); /* MJD */
11889
11890 *flagp |= POSTPONED;
11891 assert(*RExC_parse == ')');
11892 nextchar(pRExC_state);
11893 return ret;
11894
11895 /* NOTREACHED */
11896
11897 case '?': /* (??...) */
11898 is_logical = 1;
11899 if (*RExC_parse != '{') {
11900 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11901 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11902 vFAIL2utf8f(
11903 "Sequence (%" UTF8f "...) not recognized",
11904 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11905 NOT_REACHED; /*NOTREACHED*/
11906 }
11907 *flagp |= POSTPONED;
11908 paren = '{';
11909 RExC_parse++;
11910 /* FALLTHROUGH */
11911 case '{': /* (?{...}) */
11912 {
11913 U32 n = 0;
11914 struct reg_code_block *cb;
11915 OP * o;
11916
11917 RExC_seen_zerolen++;
11918
11919 if ( !pRExC_state->code_blocks
11920 || pRExC_state->code_index
11921 >= pRExC_state->code_blocks->count
11922 || pRExC_state->code_blocks->cb[pRExC_state->code_index].start
11923 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11924 - RExC_start)
11925 ) {
11926 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11927 FAIL("panic: Sequence (?{...}): no code block found\n");
11928 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11929 }
11930 /* this is a pre-compiled code block (?{...}) */
11931 cb = &pRExC_state->code_blocks->cb[pRExC_state->code_index];
11932 RExC_parse = RExC_start + cb->end;
11933 o = cb->block;
11934 if (cb->src_regex) {
11935 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11936 RExC_rxi->data->data[n] =
11937 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11938 RExC_rxi->data->data[n+1] = (void*)o;
11939 }
11940 else {
11941 n = add_data(pRExC_state,
11942 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11943 RExC_rxi->data->data[n] = (void*)o;
11944 }
11945 pRExC_state->code_index++;
11946 nextchar(pRExC_state);
11947
11948 if (is_logical) {
11949 regnode_offset eval;
11950 ret = reg_node(pRExC_state, LOGICAL);
11951
11952 eval = reg2Lanode(pRExC_state, EVAL,
11953 n,
11954
11955 /* for later propagation into (??{})
11956 * return value */
11957 RExC_flags & RXf_PMf_COMPILETIME
11958 );
11959 FLAGS(REGNODE_p(ret)) = 2;
11960 if (! REGTAIL(pRExC_state, ret, eval)) {
11961 REQUIRE_BRANCHJ(flagp, 0);
11962 }
11963 /* deal with the length of this later - MJD */
11964 return ret;
11965 }
11966 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11967 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1);
11968 Set_Node_Offset(REGNODE_p(ret), parse_start);
11969 return ret;
11970 }
11971 case '(': /* (?(?{...})...) and (?(?=...)...) */
11972 {
11973 int is_define= 0;
11974 const int DEFINE_len = sizeof("DEFINE") - 1;
11975 if ( RExC_parse < RExC_end - 1
11976 && ( ( RExC_parse[0] == '?' /* (?(?...)) */
11977 && ( RExC_parse[1] == '='
11978 || RExC_parse[1] == '!'
11979 || RExC_parse[1] == '<'
11980 || RExC_parse[1] == '{'))
11981 || ( RExC_parse[0] == '*' /* (?(*...)) */
11982 && ( memBEGINs(RExC_parse + 1,
11983 (Size_t) (RExC_end - (RExC_parse + 1)),
11984 "pla:")
11985 || memBEGINs(RExC_parse + 1,
11986 (Size_t) (RExC_end - (RExC_parse + 1)),
11987 "plb:")
11988 || memBEGINs(RExC_parse + 1,
11989 (Size_t) (RExC_end - (RExC_parse + 1)),
11990 "nla:")
11991 || memBEGINs(RExC_parse + 1,
11992 (Size_t) (RExC_end - (RExC_parse + 1)),
11993 "nlb:")
11994 || memBEGINs(RExC_parse + 1,
11995 (Size_t) (RExC_end - (RExC_parse + 1)),
11996 "positive_lookahead:")
11997 || memBEGINs(RExC_parse + 1,
11998 (Size_t) (RExC_end - (RExC_parse + 1)),
11999 "positive_lookbehind:")
12000 || memBEGINs(RExC_parse + 1,
12001 (Size_t) (RExC_end - (RExC_parse + 1)),
12002 "negative_lookahead:")
12003 || memBEGINs(RExC_parse + 1,
12004 (Size_t) (RExC_end - (RExC_parse + 1)),
12005 "negative_lookbehind:"))))
12006 ) { /* Lookahead or eval. */
12007 I32 flag;
12008 regnode_offset tail;
12009
12010 ret = reg_node(pRExC_state, LOGICAL);
12011 FLAGS(REGNODE_p(ret)) = 1;
12012
12013 tail = reg(pRExC_state, 1, &flag, depth+1);
12014 RETURN_FAIL_ON_RESTART(flag, flagp);
12015 if (! REGTAIL(pRExC_state, ret, tail)) {
12016 REQUIRE_BRANCHJ(flagp, 0);
12017 }
12018 goto insert_if;
12019 }
12020 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
12021 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
12022 {
12023 char ch = RExC_parse[0] == '<' ? '>' : '\'';
12024 char *name_start= RExC_parse++;
12025 U32 num = 0;
12026 SV *sv_dat=reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
12027 if ( RExC_parse == name_start
12028 || RExC_parse >= RExC_end
12029 || *RExC_parse != ch)
12030 {
12031 vFAIL2("Sequence (?(%c... not terminated",
12032 (ch == '>' ? '<' : ch));
12033 }
12034 RExC_parse++;
12035 if (sv_dat) {
12036 num = add_data( pRExC_state, STR_WITH_LEN("S"));
12037 RExC_rxi->data->data[num]=(void*)sv_dat;
12038 SvREFCNT_inc_simple_void_NN(sv_dat);
12039 }
12040 ret = reganode(pRExC_state, GROUPPN, num);
12041 goto insert_if_check_paren;
12042 }
12043 else if (memBEGINs(RExC_parse,
12044 (STRLEN) (RExC_end - RExC_parse),
12045 "DEFINE"))
12046 {
12047 ret = reganode(pRExC_state, DEFINEP, 0);
12048 RExC_parse += DEFINE_len;
12049 is_define = 1;
12050 goto insert_if_check_paren;
12051 }
12052 else if (RExC_parse[0] == 'R') {
12053 RExC_parse++;
12054 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
12055 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
12056 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
12057 */
12058 parno = 0;
12059 if (RExC_parse[0] == '0') {
12060 parno = 1;
12061 RExC_parse++;
12062 }
12063 else if (inRANGE(RExC_parse[0], '1', '9')) {
12064 UV uv;
12065 endptr = RExC_end;
12066 if (grok_atoUV(RExC_parse, &uv, &endptr)
12067 && uv <= I32_MAX
12068 ) {
12069 parno = (I32)uv + 1;
12070 RExC_parse = (char*)endptr;
12071 }
12072 /* else "Switch condition not recognized" below */
12073 } else if (RExC_parse[0] == '&') {
12074 SV *sv_dat;
12075 RExC_parse++;
12076 sv_dat = reg_scan_name(pRExC_state,
12077 REG_RSN_RETURN_DATA);
12078 if (sv_dat)
12079 parno = 1 + *((I32 *)SvPVX(sv_dat));
12080 }
12081 ret = reganode(pRExC_state, INSUBP, parno);
12082 goto insert_if_check_paren;
12083 }
12084 else if (inRANGE(RExC_parse[0], '1', '9')) {
12085 /* (?(1)...) */
12086 char c;
12087 UV uv;
12088 endptr = RExC_end;
12089 if (grok_atoUV(RExC_parse, &uv, &endptr)
12090 && uv <= I32_MAX
12091 ) {
12092 parno = (I32)uv;
12093 RExC_parse = (char*)endptr;
12094 }
12095 else {
12096 vFAIL("panic: grok_atoUV returned FALSE");
12097 }
12098 ret = reganode(pRExC_state, GROUPP, parno);
12099
12100 insert_if_check_paren:
12101 if (UCHARAT(RExC_parse) != ')') {
12102 RExC_parse += UTF
12103 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
12104 : 1;
12105 vFAIL("Switch condition not recognized");
12106 }
12107 nextchar(pRExC_state);
12108 insert_if:
12109 if (! REGTAIL(pRExC_state, ret, reganode(pRExC_state,
12110 IFTHEN, 0)))
12111 {
12112 REQUIRE_BRANCHJ(flagp, 0);
12113 }
12114 br = regbranch(pRExC_state, &flags, 1, depth+1);
12115 if (br == 0) {
12116 RETURN_FAIL_ON_RESTART(flags,flagp);
12117 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
12118 (UV) flags);
12119 } else
12120 if (! REGTAIL(pRExC_state, br, reganode(pRExC_state,
12121 LONGJMP, 0)))
12122 {
12123 REQUIRE_BRANCHJ(flagp, 0);
12124 }
12125 c = UCHARAT(RExC_parse);
12126 nextchar(pRExC_state);
12127 if (flags&HASWIDTH)
12128 *flagp |= HASWIDTH;
12129 if (c == '|') {
12130 if (is_define)
12131 vFAIL("(?(DEFINE)....) does not allow branches");
12132
12133 /* Fake one for optimizer. */
12134 lastbr = reganode(pRExC_state, IFTHEN, 0);
12135
12136 if (!regbranch(pRExC_state, &flags, 1, depth+1)) {
12137 RETURN_FAIL_ON_RESTART(flags, flagp);
12138 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
12139 (UV) flags);
12140 }
12141 if (! REGTAIL(pRExC_state, ret, lastbr)) {
12142 REQUIRE_BRANCHJ(flagp, 0);
12143 }
12144 if (flags&HASWIDTH)
12145 *flagp |= HASWIDTH;
12146 c = UCHARAT(RExC_parse);
12147 nextchar(pRExC_state);
12148 }
12149 else
12150 lastbr = 0;
12151 if (c != ')') {
12152 if (RExC_parse >= RExC_end)
12153 vFAIL("Switch (?(condition)... not terminated");
12154 else
12155 vFAIL("Switch (?(condition)... contains too many branches");
12156 }
12157 ender = reg_node(pRExC_state, TAIL);
12158 if (! REGTAIL(pRExC_state, br, ender)) {
12159 REQUIRE_BRANCHJ(flagp, 0);
12160 }
12161 if (lastbr) {
12162 if (! REGTAIL(pRExC_state, lastbr, ender)) {
12163 REQUIRE_BRANCHJ(flagp, 0);
12164 }
12165 if (! REGTAIL(pRExC_state,
12166 REGNODE_OFFSET(
12167 NEXTOPER(
12168 NEXTOPER(REGNODE_p(lastbr)))),
12169 ender))
12170 {
12171 REQUIRE_BRANCHJ(flagp, 0);
12172 }
12173 }
12174 else
12175 if (! REGTAIL(pRExC_state, ret, ender)) {
12176 REQUIRE_BRANCHJ(flagp, 0);
12177 }
12178 #if 0 /* Removing this doesn't cause failures in the test suite -- khw */
12179 RExC_size++; /* XXX WHY do we need this?!!
12180 For large programs it seems to be required
12181 but I can't figure out why. -- dmq*/
12182 #endif
12183 return ret;
12184 }
12185 RExC_parse += UTF
12186 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
12187 : 1;
12188 vFAIL("Unknown switch condition (?(...))");
12189 }
12190 case '[': /* (?[ ... ]) */
12191 return handle_regex_sets(pRExC_state, NULL, flagp, depth+1,
12192 oregcomp_parse);
12193 case 0: /* A NUL */
12194 RExC_parse--; /* for vFAIL to print correctly */
12195 vFAIL("Sequence (? incomplete");
12196 break;
12197
12198 case ')':
12199 if (RExC_strict) { /* [perl #132851] */
12200 ckWARNreg(RExC_parse, "Empty (?) without any modifiers");
12201 }
12202 /* FALLTHROUGH */
12203 case '*': /* If you want to support (?*...), first reconcile with GH #17363 */
12204 /* FALLTHROUGH */
12205 default: /* e.g., (?i) */
12206 RExC_parse = (char *) seqstart + 1;
12207 parse_flags:
12208 parse_lparen_question_flags(pRExC_state);
12209 if (UCHARAT(RExC_parse) != ':') {
12210 if (RExC_parse < RExC_end)
12211 nextchar(pRExC_state);
12212 *flagp = TRYAGAIN;
12213 return 0;
12214 }
12215 paren = ':';
12216 nextchar(pRExC_state);
12217 ret = 0;
12218 goto parse_rest;
12219 } /* end switch */
12220 }
12221 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
12222 capturing_parens:
12223 parno = RExC_npar;
12224 RExC_npar++;
12225 if (! ALL_PARENS_COUNTED) {
12226 /* If we are in our first pass through (and maybe only pass),
12227 * we need to allocate memory for the capturing parentheses
12228 * data structures.
12229 */
12230
12231 if (!RExC_parens_buf_size) {
12232 /* first guess at number of parens we might encounter */
12233 RExC_parens_buf_size = 10;
12234
12235 /* setup RExC_open_parens, which holds the address of each
12236 * OPEN tag, and to make things simpler for the 0 index the
12237 * start of the program - this is used later for offsets */
12238 Newxz(RExC_open_parens, RExC_parens_buf_size,
12239 regnode_offset);
12240 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
12241
12242 /* setup RExC_close_parens, which holds the address of each
12243 * CLOSE tag, and to make things simpler for the 0 index
12244 * the end of the program - this is used later for offsets
12245 * */
12246 Newxz(RExC_close_parens, RExC_parens_buf_size,
12247 regnode_offset);
12248 /* we dont know where end op starts yet, so we dont need to
12249 * set RExC_close_parens[0] like we do RExC_open_parens[0]
12250 * above */
12251 }
12252 else if (RExC_npar > RExC_parens_buf_size) {
12253 I32 old_size = RExC_parens_buf_size;
12254
12255 RExC_parens_buf_size *= 2;
12256
12257 Renew(RExC_open_parens, RExC_parens_buf_size,
12258 regnode_offset);
12259 Zero(RExC_open_parens + old_size,
12260 RExC_parens_buf_size - old_size, regnode_offset);
12261
12262 Renew(RExC_close_parens, RExC_parens_buf_size,
12263 regnode_offset);
12264 Zero(RExC_close_parens + old_size,
12265 RExC_parens_buf_size - old_size, regnode_offset);
12266 }
12267 }
12268
12269 ret = reganode(pRExC_state, OPEN, parno);
12270 if (!RExC_nestroot)
12271 RExC_nestroot = parno;
12272 if (RExC_open_parens && !RExC_open_parens[parno])
12273 {
12274 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12275 "%*s%*s Setting open paren #%" IVdf " to %zu\n",
12276 22, "| |", (int)(depth * 2 + 1), "",
12277 (IV)parno, ret));
12278 RExC_open_parens[parno]= ret;
12279 }
12280
12281 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
12282 Set_Node_Offset(REGNODE_p(ret), RExC_parse); /* MJD */
12283 is_open = 1;
12284 } else {
12285 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
12286 paren = ':';
12287 ret = 0;
12288 }
12289 }
12290 else /* ! paren */
12291 ret = 0;
12292
12293 parse_rest:
12294 /* Pick up the branches, linking them together. */
12295 parse_start = RExC_parse; /* MJD */
12296 br = regbranch(pRExC_state, &flags, 1, depth+1);
12297
12298 /* branch_len = (paren != 0); */
12299
12300 if (br == 0) {
12301 RETURN_FAIL_ON_RESTART(flags, flagp);
12302 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12303 }
12304 if (*RExC_parse == '|') {
12305 if (RExC_use_BRANCHJ) {
12306 reginsert(pRExC_state, BRANCHJ, br, depth+1);
12307 }
12308 else { /* MJD */
12309 reginsert(pRExC_state, BRANCH, br, depth+1);
12310 Set_Node_Length(REGNODE_p(br), paren != 0);
12311 Set_Node_Offset_To_R(br, parse_start-RExC_start);
12312 }
12313 have_branch = 1;
12314 }
12315 else if (paren == ':') {
12316 *flagp |= flags&SIMPLE;
12317 }
12318 if (is_open) { /* Starts with OPEN. */
12319 if (! REGTAIL(pRExC_state, ret, br)) { /* OPEN -> first. */
12320 REQUIRE_BRANCHJ(flagp, 0);
12321 }
12322 }
12323 else if (paren != '?') /* Not Conditional */
12324 ret = br;
12325 *flagp |= flags & (HASWIDTH | POSTPONED);
12326 lastbr = br;
12327 while (*RExC_parse == '|') {
12328 if (RExC_use_BRANCHJ) {
12329 bool shut_gcc_up;
12330
12331 ender = reganode(pRExC_state, LONGJMP, 0);
12332
12333 /* Append to the previous. */
12334 shut_gcc_up = REGTAIL(pRExC_state,
12335 REGNODE_OFFSET(NEXTOPER(NEXTOPER(REGNODE_p(lastbr)))),
12336 ender);
12337 PERL_UNUSED_VAR(shut_gcc_up);
12338 }
12339 nextchar(pRExC_state);
12340 if (freeze_paren) {
12341 if (RExC_npar > after_freeze)
12342 after_freeze = RExC_npar;
12343 RExC_npar = freeze_paren;
12344 }
12345 br = regbranch(pRExC_state, &flags, 0, depth+1);
12346
12347 if (br == 0) {
12348 RETURN_FAIL_ON_RESTART(flags, flagp);
12349 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12350 }
12351 if (! REGTAIL(pRExC_state, lastbr, br)) { /* BRANCH -> BRANCH. */
12352 REQUIRE_BRANCHJ(flagp, 0);
12353 }
12354 lastbr = br;
12355 *flagp |= flags & (HASWIDTH | POSTPONED);
12356 }
12357
12358 if (have_branch || paren != ':') {
12359 regnode * br;
12360
12361 /* Make a closing node, and hook it on the end. */
12362 switch (paren) {
12363 case ':':
12364 ender = reg_node(pRExC_state, TAIL);
12365 break;
12366 case 1: case 2:
12367 ender = reganode(pRExC_state, CLOSE, parno);
12368 if ( RExC_close_parens ) {
12369 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12370 "%*s%*s Setting close paren #%" IVdf " to %zu\n",
12371 22, "| |", (int)(depth * 2 + 1), "",
12372 (IV)parno, ender));
12373 RExC_close_parens[parno]= ender;
12374 if (RExC_nestroot == parno)
12375 RExC_nestroot = 0;
12376 }
12377 Set_Node_Offset(REGNODE_p(ender), RExC_parse+1); /* MJD */
12378 Set_Node_Length(REGNODE_p(ender), 1); /* MJD */
12379 break;
12380 case 's':
12381 ender = reg_node(pRExC_state, SRCLOSE);
12382 RExC_in_script_run = 0;
12383 break;
12384 case '<':
12385 case 'a':
12386 case 'A':
12387 case 'b':
12388 case 'B':
12389 case ',':
12390 case '=':
12391 case '!':
12392 *flagp &= ~HASWIDTH;
12393 /* FALLTHROUGH */
12394 case 't': /* aTomic */
12395 case '>':
12396 ender = reg_node(pRExC_state, SUCCEED);
12397 break;
12398 case 0:
12399 ender = reg_node(pRExC_state, END);
12400 assert(!RExC_end_op); /* there can only be one! */
12401 RExC_end_op = REGNODE_p(ender);
12402 if (RExC_close_parens) {
12403 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12404 "%*s%*s Setting close paren #0 (END) to %zu\n",
12405 22, "| |", (int)(depth * 2 + 1), "",
12406 ender));
12407
12408 RExC_close_parens[0]= ender;
12409 }
12410 break;
12411 }
12412 DEBUG_PARSE_r({
12413 DEBUG_PARSE_MSG("lsbr");
12414 regprop(RExC_rx, RExC_mysv1, REGNODE_p(lastbr), NULL, pRExC_state);
12415 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender), NULL, pRExC_state);
12416 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12417 SvPV_nolen_const(RExC_mysv1),
12418 (IV)lastbr,
12419 SvPV_nolen_const(RExC_mysv2),
12420 (IV)ender,
12421 (IV)(ender - lastbr)
12422 );
12423 });
12424 if (! REGTAIL(pRExC_state, lastbr, ender)) {
12425 REQUIRE_BRANCHJ(flagp, 0);
12426 }
12427
12428 if (have_branch) {
12429 char is_nothing= 1;
12430 if (depth==1)
12431 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
12432
12433 /* Hook the tails of the branches to the closing node. */
12434 for (br = REGNODE_p(ret); br; br = regnext(br)) {
12435 const U8 op = PL_regkind[OP(br)];
12436 if (op == BRANCH) {
12437 if (! REGTAIL_STUDY(pRExC_state,
12438 REGNODE_OFFSET(NEXTOPER(br)),
12439 ender))
12440 {
12441 REQUIRE_BRANCHJ(flagp, 0);
12442 }
12443 if ( OP(NEXTOPER(br)) != NOTHING
12444 || regnext(NEXTOPER(br)) != REGNODE_p(ender))
12445 is_nothing= 0;
12446 }
12447 else if (op == BRANCHJ) {
12448 bool shut_gcc_up = REGTAIL_STUDY(pRExC_state,
12449 REGNODE_OFFSET(NEXTOPER(NEXTOPER(br))),
12450 ender);
12451 PERL_UNUSED_VAR(shut_gcc_up);
12452 /* for now we always disable this optimisation * /
12453 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
12454 || regnext(NEXTOPER(NEXTOPER(br))) != REGNODE_p(ender))
12455 */
12456 is_nothing= 0;
12457 }
12458 }
12459 if (is_nothing) {
12460 regnode * ret_as_regnode = REGNODE_p(ret);
12461 br= PL_regkind[OP(ret_as_regnode)] != BRANCH
12462 ? regnext(ret_as_regnode)
12463 : ret_as_regnode;
12464 DEBUG_PARSE_r({
12465 DEBUG_PARSE_MSG("NADA");
12466 regprop(RExC_rx, RExC_mysv1, ret_as_regnode,
12467 NULL, pRExC_state);
12468 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender),
12469 NULL, pRExC_state);
12470 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12471 SvPV_nolen_const(RExC_mysv1),
12472 (IV)REG_NODE_NUM(ret_as_regnode),
12473 SvPV_nolen_const(RExC_mysv2),
12474 (IV)ender,
12475 (IV)(ender - ret)
12476 );
12477 });
12478 OP(br)= NOTHING;
12479 if (OP(REGNODE_p(ender)) == TAIL) {
12480 NEXT_OFF(br)= 0;
12481 RExC_emit= REGNODE_OFFSET(br) + 1;
12482 } else {
12483 regnode *opt;
12484 for ( opt= br + 1; opt < REGNODE_p(ender) ; opt++ )
12485 OP(opt)= OPTIMIZED;
12486 NEXT_OFF(br)= REGNODE_p(ender) - br;
12487 }
12488 }
12489 }
12490 }
12491
12492 {
12493 const char *p;
12494 /* Even/odd or x=don't care: 010101x10x */
12495 static const char parens[] = "=!aA<,>Bbt";
12496 /* flag below is set to 0 up through 'A'; 1 for larger */
12497
12498 if (paren && (p = strchr(parens, paren))) {
12499 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
12500 int flag = (p - parens) > 3;
12501
12502 if (paren == '>' || paren == 't') {
12503 node = SUSPEND, flag = 0;
12504 }
12505
12506 reginsert(pRExC_state, node, ret, depth+1);
12507 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12508 Set_Node_Offset(REGNODE_p(ret), parse_start + 1);
12509 FLAGS(REGNODE_p(ret)) = flag;
12510 if (! REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL)))
12511 {
12512 REQUIRE_BRANCHJ(flagp, 0);
12513 }
12514 }
12515 }
12516
12517 /* Check for proper termination. */
12518 if (paren) {
12519 /* restore original flags, but keep (?p) and, if we've encountered
12520 * something in the parse that changes /d rules into /u, keep the /u */
12521 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
12522 if (DEPENDS_SEMANTICS && toUSE_UNI_CHARSET_NOT_DEPENDS) {
12523 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
12524 }
12525 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
12526 RExC_parse = oregcomp_parse;
12527 vFAIL("Unmatched (");
12528 }
12529 nextchar(pRExC_state);
12530 }
12531 else if (!paren && RExC_parse < RExC_end) {
12532 if (*RExC_parse == ')') {
12533 RExC_parse++;
12534 vFAIL("Unmatched )");
12535 }
12536 else
12537 FAIL("Junk on end of regexp"); /* "Can't happen". */
12538 NOT_REACHED; /* NOTREACHED */
12539 }
12540
12541 if (after_freeze > RExC_npar)
12542 RExC_npar = after_freeze;
12543
12544 RExC_in_lookaround = was_in_lookaround;
12545
12546 return(ret);
12547 }
12548
12549 /*
12550 - regbranch - one alternative of an | operator
12551 *
12552 * Implements the concatenation operator.
12553 *
12554 * On success, returns the offset at which any next node should be placed into
12555 * the regex engine program being compiled.
12556 *
12557 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
12558 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
12559 * UTF-8
12560 */
12561 STATIC regnode_offset
S_regbranch(pTHX_ RExC_state_t * pRExC_state,I32 * flagp,I32 first,U32 depth)12562 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
12563 {
12564 regnode_offset ret;
12565 regnode_offset chain = 0;
12566 regnode_offset latest;
12567 I32 flags = 0, c = 0;
12568 DECLARE_AND_GET_RE_DEBUG_FLAGS;
12569
12570 PERL_ARGS_ASSERT_REGBRANCH;
12571
12572 DEBUG_PARSE("brnc");
12573
12574 if (first)
12575 ret = 0;
12576 else {
12577 if (RExC_use_BRANCHJ)
12578 ret = reganode(pRExC_state, BRANCHJ, 0);
12579 else {
12580 ret = reg_node(pRExC_state, BRANCH);
12581 Set_Node_Length(REGNODE_p(ret), 1);
12582 }
12583 }
12584
12585 *flagp = 0; /* Initialize. */
12586
12587 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
12588 FALSE /* Don't force to /x */ );
12589 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
12590 flags &= ~TRYAGAIN;
12591 latest = regpiece(pRExC_state, &flags, depth+1);
12592 if (latest == 0) {
12593 if (flags & TRYAGAIN)
12594 continue;
12595 RETURN_FAIL_ON_RESTART(flags, flagp);
12596 FAIL2("panic: regpiece returned failure, flags=%#" UVxf, (UV) flags);
12597 }
12598 else if (ret == 0)
12599 ret = latest;
12600 *flagp |= flags&(HASWIDTH|POSTPONED);
12601 if (chain != 0) {
12602 /* FIXME adding one for every branch after the first is probably
12603 * excessive now we have TRIE support. (hv) */
12604 MARK_NAUGHTY(1);
12605 if (! REGTAIL(pRExC_state, chain, latest)) {
12606 /* XXX We could just redo this branch, but figuring out what
12607 * bookkeeping needs to be reset is a pain, and it's likely
12608 * that other branches that goto END will also be too large */
12609 REQUIRE_BRANCHJ(flagp, 0);
12610 }
12611 }
12612 chain = latest;
12613 c++;
12614 }
12615 if (chain == 0) { /* Loop ran zero times. */
12616 chain = reg_node(pRExC_state, NOTHING);
12617 if (ret == 0)
12618 ret = chain;
12619 }
12620 if (c == 1) {
12621 *flagp |= flags&SIMPLE;
12622 }
12623
12624 return ret;
12625 }
12626
12627 #define RBRACE 0
12628 #define MIN_S 1
12629 #define MIN_E 2
12630 #define MAX_S 3
12631 #define MAX_E 4
12632
12633 #ifndef PERL_IN_XSUB_RE
12634 bool
Perl_regcurly(const char * s,const char * e,const char * result[5])12635 Perl_regcurly(const char *s, const char *e, const char * result[5])
12636 {
12637 /* This function matches a {m,n} quantifier. When called with a NULL final
12638 * argument, it simply parses the input from 's' up through 'e-1', and
12639 * returns a boolean as to whether or not this input is syntactically a
12640 * {m,n} quantifier.
12641 *
12642 * When called with a non-NULL final parameter, and when the function
12643 * returns TRUE, it additionally stores information into the array
12644 * specified by that parameter about what it found in the parse. The
12645 * parameter must be a pointer into a 5 element array of 'const char *'
12646 * elements. The returned information is as follows:
12647 * result[RBRACE] points to the closing brace
12648 * result[MIN_S] points to the first byte of the lower bound
12649 * result[MIN_E] points to one beyond the final byte of the lower bound
12650 * result[MAX_S] points to the first byte of the upper bound
12651 * result[MAX_E] points to one beyond the final byte of the upper bound
12652 *
12653 * If the quantifier is of the form {m,} (meaning an infinite upper
12654 * bound), result[MAX_E] is set to result[MAX_S]; what they actually point
12655 * to is irrelevant, just that it's the same place
12656 *
12657 * If instead the quantifier is of the form {m} there is actually only
12658 * one bound, and both the upper and lower result[] elements are set to
12659 * point to it.
12660 *
12661 * This function checks only for syntactic validity; it leaves checking for
12662 * semantic validity and raising any diagnostics to the caller. This
12663 * function is called in multiple places to check for syntax, but only from
12664 * one for semantics. It makes it as simple as possible for the
12665 * syntax-only callers, while furnishing just enough information for the
12666 * semantic caller.
12667 */
12668
12669 const char * min_start = NULL;
12670 const char * max_start = NULL;
12671 const char * min_end = NULL;
12672 const char * max_end = NULL;
12673
12674 bool has_comma = FALSE;
12675
12676 PERL_ARGS_ASSERT_REGCURLY;
12677
12678 if (s >= e || *s++ != '{')
12679 return FALSE;
12680
12681 while (s < e && isBLANK(*s)) {
12682 s++;
12683 }
12684
12685 if isDIGIT(*s) {
12686 min_start = s;
12687 do {
12688 s++;
12689 } while (s < e && isDIGIT(*s));
12690 min_end = s;
12691 }
12692
12693 while (s < e && isBLANK(*s)) {
12694 s++;
12695 }
12696
12697 if (*s == ',') {
12698 has_comma = TRUE;
12699 s++;
12700
12701 while (s < e && isBLANK(*s)) {
12702 s++;
12703 }
12704
12705 if isDIGIT(*s) {
12706 max_start = s;
12707 do {
12708 s++;
12709 } while (s < e && isDIGIT(*s));
12710 max_end = s;
12711 }
12712 }
12713
12714 while (s < e && isBLANK(*s)) {
12715 s++;
12716 }
12717 /* Need at least one number */
12718 if (s >= e || *s != '}' || (! min_start && ! max_end)) {
12719 return FALSE;
12720 }
12721
12722 if (result) {
12723
12724 result[RBRACE] = s;
12725
12726 result[MIN_S] = min_start;
12727 result[MIN_E] = min_end;
12728 if (has_comma) {
12729 if (max_start) {
12730 result[MAX_S] = max_start;
12731 result[MAX_E] = max_end;
12732 }
12733 else {
12734 /* Having no value after the comma is signalled by setting
12735 * start and end to the same value. What that value is isn't
12736 * relevant; NULL is chosen simply because it will fail if the
12737 * caller mistakenly uses it */
12738 result[MAX_S] = result[MAX_E] = NULL;
12739 }
12740 }
12741 else { /* No comma means lower and upper bounds are the same */
12742 result[MAX_S] = min_start;
12743 result[MAX_E] = min_end;
12744 }
12745 }
12746
12747 return TRUE;
12748 }
12749 #endif
12750
12751 U32
S_get_quantifier_value(pTHX_ RExC_state_t * pRExC_state,const char * start,const char * end)12752 S_get_quantifier_value(pTHX_ RExC_state_t *pRExC_state,
12753 const char * start, const char * end)
12754 {
12755 /* This is a helper function for regpiece() to compute, given the
12756 * quantifier {m,n}, the value of either m or n, based on the starting
12757 * position 'start' in the string, through the byte 'end-1', returning it
12758 * if valid, and failing appropriately if not. It knows the restrictions
12759 * imposed on quantifier values */
12760
12761 UV uv;
12762 STATIC_ASSERT_DECL(REG_INFTY <= U32_MAX);
12763
12764 PERL_ARGS_ASSERT_GET_QUANTIFIER_VALUE;
12765
12766 if (grok_atoUV(start, &uv, &end)) {
12767 if (uv < REG_INFTY) { /* A valid, small-enough number */
12768 return (U32) uv;
12769 }
12770 }
12771 else if (*start == '0') { /* grok_atoUV() fails for only two reasons:
12772 leading zeros or overflow */
12773 RExC_parse = (char * ) end;
12774
12775 /* Perhaps too generic a msg for what is only failure from having
12776 * leading zeros, but this is how it's always behaved. */
12777 vFAIL("Invalid quantifier in {,}");
12778 NOT_REACHED; /*NOTREACHED*/
12779 }
12780
12781 /* Here, found a quantifier, but was too large; either it overflowed or was
12782 * too big a legal number */
12783 RExC_parse = (char * ) end;
12784 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12785
12786 NOT_REACHED; /*NOTREACHED*/
12787 return U32_MAX; /* Perhaps some compilers will be expecting a return */
12788 }
12789
12790 /*
12791 - regpiece - something followed by possible quantifier * + ? {n,m}
12792 *
12793 * Note that the branching code sequences used for ? and the general cases
12794 * of * and + are somewhat optimized: they use the same NOTHING node as
12795 * both the endmarker for their branch list and the body of the last branch.
12796 * It might seem that this node could be dispensed with entirely, but the
12797 * endmarker role is not redundant.
12798 *
12799 * On success, returns the offset at which any next node should be placed into
12800 * the regex engine program being compiled.
12801 *
12802 * Returns 0 otherwise, with *flagp set to indicate why:
12803 * TRYAGAIN if regatom() returns 0 with TRYAGAIN.
12804 * RESTART_PARSE if the parse needs to be restarted, or'd with
12805 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
12806 */
12807 STATIC regnode_offset
S_regpiece(pTHX_ RExC_state_t * pRExC_state,I32 * flagp,U32 depth)12808 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12809 {
12810 regnode_offset ret;
12811 char op;
12812 I32 flags;
12813 const char * const origparse = RExC_parse;
12814 I32 min;
12815 I32 max = REG_INFTY;
12816 #ifdef RE_TRACK_PATTERN_OFFSETS
12817 char *parse_start;
12818 #endif
12819
12820 /* Save the original in case we change the emitted regop to a FAIL. */
12821 const regnode_offset orig_emit = RExC_emit;
12822
12823 DECLARE_AND_GET_RE_DEBUG_FLAGS;
12824
12825 PERL_ARGS_ASSERT_REGPIECE;
12826
12827 DEBUG_PARSE("piec");
12828
12829 ret = regatom(pRExC_state, &flags, depth+1);
12830 if (ret == 0) {
12831 RETURN_FAIL_ON_RESTART_OR_FLAGS(flags, flagp, TRYAGAIN);
12832 FAIL2("panic: regatom returned failure, flags=%#" UVxf, (UV) flags);
12833 }
12834
12835 #ifdef RE_TRACK_PATTERN_OFFSETS
12836 parse_start = RExC_parse;
12837 #endif
12838
12839 op = *RExC_parse;
12840 switch (op) {
12841 const char * regcurly_return[5];
12842
12843 case '*':
12844 nextchar(pRExC_state);
12845 min = 0;
12846 break;
12847
12848 case '+':
12849 nextchar(pRExC_state);
12850 min = 1;
12851 break;
12852
12853 case '?':
12854 nextchar(pRExC_state);
12855 min = 0; max = 1;
12856 break;
12857
12858 case '{': /* A '{' may or may not indicate a quantifier; call regcurly()
12859 to determine which */
12860 if (regcurly(RExC_parse, RExC_end, regcurly_return)) {
12861 const char * min_start = regcurly_return[MIN_S];
12862 const char * min_end = regcurly_return[MIN_E];
12863 const char * max_start = regcurly_return[MAX_S];
12864 const char * max_end = regcurly_return[MAX_E];
12865
12866 if (min_start) {
12867 min = get_quantifier_value(pRExC_state, min_start, min_end);
12868 }
12869 else {
12870 min = 0;
12871 }
12872
12873 if (max_start == max_end) { /* Was of the form {m,} */
12874 max = REG_INFTY;
12875 }
12876 else if (max_start == min_start) { /* Was of the form {m} */
12877 max = min;
12878 }
12879 else { /* Was of the form {m,n} */
12880 assert(max_end >= max_start);
12881
12882 max = get_quantifier_value(pRExC_state, max_start, max_end);
12883 }
12884
12885 RExC_parse = (char *) regcurly_return[RBRACE];
12886 nextchar(pRExC_state);
12887
12888 if (max < min) { /* If can't match, warn and optimize to fail
12889 unconditionally */
12890 reginsert(pRExC_state, OPFAIL, orig_emit, depth+1);
12891 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
12892 NEXT_OFF(REGNODE_p(orig_emit)) =
12893 regarglen[OPFAIL] + NODE_STEP_REGNODE;
12894 return ret;
12895 }
12896 else if (min == max && *RExC_parse == '?') {
12897 ckWARN2reg(RExC_parse + 1,
12898 "Useless use of greediness modifier '%c'",
12899 *RExC_parse);
12900 }
12901
12902 break;
12903 } /* End of is {m,n} */
12904
12905 /* Here was a '{', but what followed it didn't form a quantifier. */
12906 /* FALLTHROUGH */
12907
12908 default:
12909 *flagp = flags;
12910 return(ret);
12911 NOT_REACHED; /*NOTREACHED*/
12912 }
12913
12914 /* Here we have a quantifier, and have calculated 'min' and 'max'.
12915 *
12916 * Check and possibly adjust a zero width operand */
12917 if (! (flags & (HASWIDTH|POSTPONED))) {
12918 if (max > REG_INFTY/3) {
12919 if (origparse[0] == '\\' && origparse[1] == 'K') {
12920 vFAIL2utf8f(
12921 "%" UTF8f " is forbidden - matches null string"
12922 " many times",
12923 UTF8fARG(UTF, (RExC_parse >= origparse
12924 ? RExC_parse - origparse
12925 : 0),
12926 origparse));
12927 } else {
12928 ckWARN2reg(RExC_parse,
12929 "%" UTF8f " matches null string many times",
12930 UTF8fARG(UTF, (RExC_parse >= origparse
12931 ? RExC_parse - origparse
12932 : 0),
12933 origparse));
12934 }
12935 }
12936
12937 /* There's no point in trying to match something 0 length more than
12938 * once except for extra side effects, which we don't have here since
12939 * not POSTPONED */
12940 if (max > 1) {
12941 max = 1;
12942 if (min > max) {
12943 min = max;
12944 }
12945 }
12946 }
12947
12948 /* If this is a code block pass it up */
12949 *flagp |= (flags & POSTPONED);
12950
12951 if (max > 0) {
12952 *flagp |= (flags & HASWIDTH);
12953 if (max == REG_INFTY)
12954 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12955 }
12956
12957 /* 'SIMPLE' operands don't require full generality */
12958 if ((flags&SIMPLE)) {
12959 if (max == REG_INFTY) {
12960 if (min == 0) {
12961 if (UNLIKELY(RExC_pm_flags & PMf_WILDCARD)) {
12962 goto min0_maxINF_wildcard_forbidden;
12963 }
12964
12965 reginsert(pRExC_state, STAR, ret, depth+1);
12966 MARK_NAUGHTY(4);
12967 goto done_main_op;
12968 }
12969 else if (min == 1) {
12970 reginsert(pRExC_state, PLUS, ret, depth+1);
12971 MARK_NAUGHTY(3);
12972 goto done_main_op;
12973 }
12974 }
12975
12976 /* Here, SIMPLE, but not the '*' and '+' special cases */
12977
12978 MARK_NAUGHTY_EXP(2, 2);
12979 reginsert(pRExC_state, CURLY, ret, depth+1);
12980 Set_Node_Offset(REGNODE_p(ret), parse_start+1); /* MJD */
12981 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12982 }
12983 else { /* not SIMPLE */
12984 const regnode_offset w = reg_node(pRExC_state, WHILEM);
12985
12986 FLAGS(REGNODE_p(w)) = 0;
12987 if (! REGTAIL(pRExC_state, ret, w)) {
12988 REQUIRE_BRANCHJ(flagp, 0);
12989 }
12990 if (RExC_use_BRANCHJ) {
12991 reginsert(pRExC_state, LONGJMP, ret, depth+1);
12992 reginsert(pRExC_state, NOTHING, ret, depth+1);
12993 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over LONGJMP. */
12994 }
12995 reginsert(pRExC_state, CURLYX, ret, depth+1);
12996 /* MJD hk */
12997 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
12998 Set_Node_Length(REGNODE_p(ret),
12999 op == '{' ? (RExC_parse - parse_start) : 1);
13000
13001 if (RExC_use_BRANCHJ)
13002 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over NOTHING to
13003 LONGJMP. */
13004 if (! REGTAIL(pRExC_state, ret, reg_node(pRExC_state,
13005 NOTHING)))
13006 {
13007 REQUIRE_BRANCHJ(flagp, 0);
13008 }
13009 RExC_whilem_seen++;
13010 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
13011 }
13012
13013 /* Finish up the CURLY/CURLYX case */
13014 FLAGS(REGNODE_p(ret)) = 0;
13015
13016 ARG1_SET(REGNODE_p(ret), (U16)min);
13017 ARG2_SET(REGNODE_p(ret), (U16)max);
13018
13019 done_main_op:
13020
13021 /* Process any greediness modifiers */
13022 if (*RExC_parse == '?') {
13023 nextchar(pRExC_state);
13024 reginsert(pRExC_state, MINMOD, ret, depth+1);
13025 if (! REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE)) {
13026 REQUIRE_BRANCHJ(flagp, 0);
13027 }
13028 }
13029 else if (*RExC_parse == '+') {
13030 regnode_offset ender;
13031 nextchar(pRExC_state);
13032 ender = reg_node(pRExC_state, SUCCEED);
13033 if (! REGTAIL(pRExC_state, ret, ender)) {
13034 REQUIRE_BRANCHJ(flagp, 0);
13035 }
13036 reginsert(pRExC_state, SUSPEND, ret, depth+1);
13037 ender = reg_node(pRExC_state, TAIL);
13038 if (! REGTAIL(pRExC_state, ret, ender)) {
13039 REQUIRE_BRANCHJ(flagp, 0);
13040 }
13041 }
13042
13043 /* Forbid extra quantifiers */
13044 if (isQUANTIFIER(RExC_parse, RExC_end)) {
13045 RExC_parse++;
13046 vFAIL("Nested quantifiers");
13047 }
13048
13049 return(ret);
13050
13051 min0_maxINF_wildcard_forbidden:
13052
13053 /* Here we are in a wildcard match, and the minimum match length is 0, and
13054 * the max could be infinity. This is currently forbidden. The only
13055 * reason is to make it harder to write patterns that take a long long time
13056 * to halt, and because the use of this construct isn't necessary in
13057 * matching Unicode property values */
13058 RExC_parse++;
13059 /* diag_listed_as: Use of %s is not allowed in Unicode property wildcard
13060 subpatterns in regex; marked by <-- HERE in m/%s/
13061 */
13062 vFAIL("Use of quantifier '*' is not allowed in Unicode property wildcard"
13063 " subpatterns");
13064
13065 /* Note, don't need to worry about the input being '{0,}', as a '}' isn't
13066 * legal at all in wildcards, so can't get this far */
13067
13068 NOT_REACHED; /*NOTREACHED*/
13069 }
13070
13071 STATIC bool
S_grok_bslash_N(pTHX_ RExC_state_t * pRExC_state,regnode_offset * node_p,UV * code_point_p,int * cp_count,I32 * flagp,const bool strict,const U32 depth)13072 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
13073 regnode_offset * node_p,
13074 UV * code_point_p,
13075 int * cp_count,
13076 I32 * flagp,
13077 const bool strict,
13078 const U32 depth
13079 )
13080 {
13081 /* This routine teases apart the various meanings of \N and returns
13082 * accordingly. The input parameters constrain which meaning(s) is/are valid
13083 * in the current context.
13084 *
13085 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
13086 *
13087 * If <code_point_p> is not NULL, the context is expecting the result to be a
13088 * single code point. If this \N instance turns out to a single code point,
13089 * the function returns TRUE and sets *code_point_p to that code point.
13090 *
13091 * If <node_p> is not NULL, the context is expecting the result to be one of
13092 * the things representable by a regnode. If this \N instance turns out to be
13093 * one such, the function generates the regnode, returns TRUE and sets *node_p
13094 * to point to the offset of that regnode into the regex engine program being
13095 * compiled.
13096 *
13097 * If this instance of \N isn't legal in any context, this function will
13098 * generate a fatal error and not return.
13099 *
13100 * On input, RExC_parse should point to the first char following the \N at the
13101 * time of the call. On successful return, RExC_parse will have been updated
13102 * to point to just after the sequence identified by this routine. Also
13103 * *flagp has been updated as needed.
13104 *
13105 * When there is some problem with the current context and this \N instance,
13106 * the function returns FALSE, without advancing RExC_parse, nor setting
13107 * *node_p, nor *code_point_p, nor *flagp.
13108 *
13109 * If <cp_count> is not NULL, the caller wants to know the length (in code
13110 * points) that this \N sequence matches. This is set, and the input is
13111 * parsed for errors, even if the function returns FALSE, as detailed below.
13112 *
13113 * There are 6 possibilities here, as detailed in the next 6 paragraphs.
13114 *
13115 * Probably the most common case is for the \N to specify a single code point.
13116 * *cp_count will be set to 1, and *code_point_p will be set to that code
13117 * point.
13118 *
13119 * Another possibility is for the input to be an empty \N{}. This is no
13120 * longer accepted, and will generate a fatal error.
13121 *
13122 * Another possibility is for a custom charnames handler to be in effect which
13123 * translates the input name to an empty string. *cp_count will be set to 0.
13124 * *node_p will be set to a generated NOTHING node.
13125 *
13126 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
13127 * set to 0. *node_p will be set to a generated REG_ANY node.
13128 *
13129 * The fifth possibility is that \N resolves to a sequence of more than one
13130 * code points. *cp_count will be set to the number of code points in the
13131 * sequence. *node_p will be set to a generated node returned by this
13132 * function calling S_reg().
13133 *
13134 * The sixth and final possibility is that it is premature to be calling this
13135 * function; the parse needs to be restarted. This can happen when this
13136 * changes from /d to /u rules, or when the pattern needs to be upgraded to
13137 * UTF-8. The latter occurs only when the fifth possibility would otherwise
13138 * be in effect, and is because one of those code points requires the pattern
13139 * to be recompiled as UTF-8. The function returns FALSE, and sets the
13140 * RESTART_PARSE and NEED_UTF8 flags in *flagp, as appropriate. When this
13141 * happens, the caller needs to desist from continuing parsing, and return
13142 * this information to its caller. This is not set for when there is only one
13143 * code point, as this can be called as part of an ANYOF node, and they can
13144 * store above-Latin1 code points without the pattern having to be in UTF-8.
13145 *
13146 * For non-single-quoted regexes, the tokenizer has resolved character and
13147 * sequence names inside \N{...} into their Unicode values, normalizing the
13148 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
13149 * hex-represented code points in the sequence. This is done there because
13150 * the names can vary based on what charnames pragma is in scope at the time,
13151 * so we need a way to take a snapshot of what they resolve to at the time of
13152 * the original parse. [perl #56444].
13153 *
13154 * That parsing is skipped for single-quoted regexes, so here we may get
13155 * '\N{NAME}', which is parsed now. If the single-quoted regex is something
13156 * like '\N{U+41}', that code point is Unicode, and has to be translated into
13157 * the native character set for non-ASCII platforms. The other possibilities
13158 * are already native, so no translation is done. */
13159
13160 char * endbrace; /* points to '}' following the name */
13161 char * e; /* points to final non-blank before endbrace */
13162 char* p = RExC_parse; /* Temporary */
13163
13164 SV * substitute_parse = NULL;
13165 char *orig_end;
13166 char *save_start;
13167 I32 flags;
13168
13169 DECLARE_AND_GET_RE_DEBUG_FLAGS;
13170
13171 PERL_ARGS_ASSERT_GROK_BSLASH_N;
13172
13173 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
13174 assert(! (node_p && cp_count)); /* At most 1 should be set */
13175
13176 if (cp_count) { /* Initialize return for the most common case */
13177 *cp_count = 1;
13178 }
13179
13180 /* The [^\n] meaning of \N ignores spaces and comments under the /x
13181 * modifier. The other meanings do not (except blanks adjacent to and
13182 * within the braces), so use a temporary until we find out which we are
13183 * being called with */
13184 skip_to_be_ignored_text(pRExC_state, &p,
13185 FALSE /* Don't force to /x */ );
13186
13187 /* Disambiguate between \N meaning a named character versus \N meaning
13188 * [^\n]. The latter is assumed when the {...} following the \N is a legal
13189 * quantifier, or if there is no '{' at all */
13190 if (*p != '{' || regcurly(p, RExC_end, NULL)) {
13191 RExC_parse = p;
13192 if (cp_count) {
13193 *cp_count = -1;
13194 }
13195
13196 if (! node_p) {
13197 return FALSE;
13198 }
13199
13200 *node_p = reg_node(pRExC_state, REG_ANY);
13201 *flagp |= HASWIDTH|SIMPLE;
13202 MARK_NAUGHTY(1);
13203 Set_Node_Length(REGNODE_p(*(node_p)), 1); /* MJD */
13204 return TRUE;
13205 }
13206
13207 /* The test above made sure that the next real character is a '{', but
13208 * under the /x modifier, it could be separated by space (or a comment and
13209 * \n) and this is not allowed (for consistency with \x{...} and the
13210 * tokenizer handling of \N{NAME}). */
13211 if (*RExC_parse != '{') {
13212 vFAIL("Missing braces on \\N{}");
13213 }
13214
13215 RExC_parse++; /* Skip past the '{' */
13216
13217 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
13218 if (! endbrace) { /* no trailing brace */
13219 vFAIL2("Missing right brace on \\%c{}", 'N');
13220 }
13221
13222 /* Here, we have decided it should be a named character or sequence. These
13223 * imply Unicode semantics */
13224 REQUIRE_UNI_RULES(flagp, FALSE);
13225
13226 /* \N{_} is what toke.c returns to us to indicate a name that evaluates to
13227 * nothing at all (not allowed under strict) */
13228 if (endbrace - RExC_parse == 1 && *RExC_parse == '_') {
13229 RExC_parse = endbrace;
13230 if (strict) {
13231 RExC_parse++; /* Position after the "}" */
13232 vFAIL("Zero length \\N{}");
13233 }
13234
13235 if (cp_count) {
13236 *cp_count = 0;
13237 }
13238 nextchar(pRExC_state);
13239 if (! node_p) {
13240 return FALSE;
13241 }
13242
13243 *node_p = reg_node(pRExC_state, NOTHING);
13244 return TRUE;
13245 }
13246
13247 while (isBLANK(*RExC_parse)) {
13248 RExC_parse++;
13249 }
13250
13251 e = endbrace;
13252 while (RExC_parse < e && isBLANK(*(e-1))) {
13253 e--;
13254 }
13255
13256 if (e - RExC_parse < 2 || ! strBEGINs(RExC_parse, "U+")) {
13257
13258 /* Here, the name isn't of the form U+.... This can happen if the
13259 * pattern is single-quoted, so didn't get evaluated in toke.c. Now
13260 * is the time to find out what the name means */
13261
13262 const STRLEN name_len = e - RExC_parse;
13263 SV * value_sv; /* What does this name evaluate to */
13264 SV ** value_svp;
13265 const U8 * value; /* string of name's value */
13266 STRLEN value_len; /* and its length */
13267
13268 /* RExC_unlexed_names is a hash of names that weren't evaluated by
13269 * toke.c, and their values. Make sure is initialized */
13270 if (! RExC_unlexed_names) {
13271 RExC_unlexed_names = newHV();
13272 }
13273
13274 /* If we have already seen this name in this pattern, use that. This
13275 * allows us to only call the charnames handler once per name per
13276 * pattern. A broken or malicious handler could return something
13277 * different each time, which could cause the results to vary depending
13278 * on if something gets added or subtracted from the pattern that
13279 * causes the number of passes to change, for example */
13280 if ((value_svp = hv_fetch(RExC_unlexed_names, RExC_parse,
13281 name_len, 0)))
13282 {
13283 value_sv = *value_svp;
13284 }
13285 else { /* Otherwise we have to go out and get the name */
13286 const char * error_msg = NULL;
13287 value_sv = get_and_check_backslash_N_name(RExC_parse, e,
13288 UTF,
13289 &error_msg);
13290 if (error_msg) {
13291 RExC_parse = endbrace;
13292 vFAIL(error_msg);
13293 }
13294
13295 /* If no error message, should have gotten a valid return */
13296 assert (value_sv);
13297
13298 /* Save the name's meaning for later use */
13299 if (! hv_store(RExC_unlexed_names, RExC_parse, name_len,
13300 value_sv, 0))
13301 {
13302 Perl_croak(aTHX_ "panic: hv_store() unexpectedly failed");
13303 }
13304 }
13305
13306 /* Here, we have the value the name evaluates to in 'value_sv' */
13307 value = (U8 *) SvPV(value_sv, value_len);
13308
13309 /* See if the result is one code point vs 0 or multiple */
13310 if (inRANGE(value_len, 1, ((UV) SvUTF8(value_sv)
13311 ? UTF8SKIP(value)
13312 : 1)))
13313 {
13314 /* Here, exactly one code point. If that isn't what is wanted,
13315 * fail */
13316 if (! code_point_p) {
13317 RExC_parse = p;
13318 return FALSE;
13319 }
13320
13321 /* Convert from string to numeric code point */
13322 *code_point_p = (SvUTF8(value_sv))
13323 ? valid_utf8_to_uvchr(value, NULL)
13324 : *value;
13325
13326 /* Have parsed this entire single code point \N{...}. *cp_count
13327 * has already been set to 1, so don't do it again. */
13328 RExC_parse = endbrace;
13329 nextchar(pRExC_state);
13330 return TRUE;
13331 } /* End of is a single code point */
13332
13333 /* Count the code points, if caller desires. The API says to do this
13334 * even if we will later return FALSE */
13335 if (cp_count) {
13336 *cp_count = 0;
13337
13338 *cp_count = (SvUTF8(value_sv))
13339 ? utf8_length(value, value + value_len)
13340 : value_len;
13341 }
13342
13343 /* Fail if caller doesn't want to handle a multi-code-point sequence.
13344 * But don't back the pointer up if the caller wants to know how many
13345 * code points there are (they need to handle it themselves in this
13346 * case). */
13347 if (! node_p) {
13348 if (! cp_count) {
13349 RExC_parse = p;
13350 }
13351 return FALSE;
13352 }
13353
13354 /* Convert this to a sub-pattern of the form "(?: ... )", and then call
13355 * reg recursively to parse it. That way, it retains its atomicness,
13356 * while not having to worry about any special handling that some code
13357 * points may have. */
13358
13359 substitute_parse = newSVpvs("?:");
13360 sv_catsv(substitute_parse, value_sv);
13361 sv_catpv(substitute_parse, ")");
13362
13363 /* The value should already be native, so no need to convert on EBCDIC
13364 * platforms.*/
13365 assert(! RExC_recode_x_to_native);
13366
13367 }
13368 else { /* \N{U+...} */
13369 Size_t count = 0; /* code point count kept internally */
13370
13371 /* We can get to here when the input is \N{U+...} or when toke.c has
13372 * converted a name to the \N{U+...} form. This include changing a
13373 * name that evaluates to multiple code points to \N{U+c1.c2.c3 ...} */
13374
13375 RExC_parse += 2; /* Skip past the 'U+' */
13376
13377 /* Code points are separated by dots. The '}' terminates the whole
13378 * thing. */
13379
13380 do { /* Loop until the ending brace */
13381 I32 flags = PERL_SCAN_SILENT_OVERFLOW
13382 | PERL_SCAN_SILENT_ILLDIGIT
13383 | PERL_SCAN_NOTIFY_ILLDIGIT
13384 | PERL_SCAN_ALLOW_MEDIAL_UNDERSCORES
13385 | PERL_SCAN_DISALLOW_PREFIX;
13386 STRLEN len = e - RExC_parse;
13387 NV overflow_value;
13388 char * start_digit = RExC_parse;
13389 UV cp = grok_hex(RExC_parse, &len, &flags, &overflow_value);
13390
13391 if (len == 0) {
13392 RExC_parse++;
13393 bad_NU:
13394 vFAIL("Invalid hexadecimal number in \\N{U+...}");
13395 }
13396
13397 RExC_parse += len;
13398
13399 if (cp > MAX_LEGAL_CP) {
13400 vFAIL(form_cp_too_large_msg(16, start_digit, len, 0));
13401 }
13402
13403 if (RExC_parse >= e) { /* Got to the closing '}' */
13404 if (count) {
13405 goto do_concat;
13406 }
13407
13408 /* Here, is a single code point; fail if doesn't want that */
13409 if (! code_point_p) {
13410 RExC_parse = p;
13411 return FALSE;
13412 }
13413
13414 /* A single code point is easy to handle; just return it */
13415 *code_point_p = UNI_TO_NATIVE(cp);
13416 RExC_parse = endbrace;
13417 nextchar(pRExC_state);
13418 return TRUE;
13419 }
13420
13421 /* Here, the parse stopped bfore the ending brace. This is legal
13422 * only if that character is a dot separating code points, like a
13423 * multiple character sequence (of the form "\N{U+c1.c2. ... }".
13424 * So the next character must be a dot (and the one after that
13425 * can't be the ending brace, or we'd have something like
13426 * \N{U+100.} )
13427 * */
13428 if (*RExC_parse != '.' || RExC_parse + 1 >= e) {
13429 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
13430 ? UTF8SKIP(RExC_parse)
13431 : 1;
13432 RExC_parse = MIN(e, RExC_parse);/* Guard against malformed utf8
13433 */
13434 goto bad_NU;
13435 }
13436
13437 /* Here, looks like its really a multiple character sequence. Fail
13438 * if that's not what the caller wants. But continue with counting
13439 * and error checking if they still want a count */
13440 if (! node_p && ! cp_count) {
13441 return FALSE;
13442 }
13443
13444 /* What is done here is to convert this to a sub-pattern of the
13445 * form \x{char1}\x{char2}... and then call reg recursively to
13446 * parse it (enclosing in "(?: ... )" ). That way, it retains its
13447 * atomicness, while not having to worry about special handling
13448 * that some code points may have. We don't create a subpattern,
13449 * but go through the motions of code point counting and error
13450 * checking, if the caller doesn't want a node returned. */
13451
13452 if (node_p && ! substitute_parse) {
13453 substitute_parse = newSVpvs("?:");
13454 }
13455
13456 do_concat:
13457
13458 if (node_p) {
13459 /* Convert to notation the rest of the code understands */
13460 sv_catpvs(substitute_parse, "\\x{");
13461 sv_catpvn(substitute_parse, start_digit,
13462 RExC_parse - start_digit);
13463 sv_catpvs(substitute_parse, "}");
13464 }
13465
13466 /* Move to after the dot (or ending brace the final time through.)
13467 * */
13468 RExC_parse++;
13469 count++;
13470
13471 } while (RExC_parse < e);
13472
13473 if (! node_p) { /* Doesn't want the node */
13474 assert (cp_count);
13475
13476 *cp_count = count;
13477 return FALSE;
13478 }
13479
13480 sv_catpvs(substitute_parse, ")");
13481
13482 /* The values are Unicode, and therefore have to be converted to native
13483 * on a non-Unicode (meaning non-ASCII) platform. */
13484 SET_recode_x_to_native(1);
13485 }
13486
13487 /* Here, we have the string the name evaluates to, ready to be parsed,
13488 * stored in 'substitute_parse' as a series of valid "\x{...}\x{...}"
13489 * constructs. This can be called from within a substitute parse already.
13490 * The error reporting mechanism doesn't work for 2 levels of this, but the
13491 * code above has validated this new construct, so there should be no
13492 * errors generated by the below. And this isn't an exact copy, so the
13493 * mechanism to seamlessly deal with this won't work, so turn off warnings
13494 * during it */
13495 save_start = RExC_start;
13496 orig_end = RExC_end;
13497
13498 RExC_parse = RExC_start = SvPVX(substitute_parse);
13499 RExC_end = RExC_parse + SvCUR(substitute_parse);
13500 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
13501
13502 *node_p = reg(pRExC_state, 1, &flags, depth+1);
13503
13504 /* Restore the saved values */
13505 RESTORE_WARNINGS;
13506 RExC_start = save_start;
13507 RExC_parse = endbrace;
13508 RExC_end = orig_end;
13509 SET_recode_x_to_native(0);
13510
13511 SvREFCNT_dec_NN(substitute_parse);
13512
13513 if (! *node_p) {
13514 RETURN_FAIL_ON_RESTART(flags, flagp);
13515 FAIL2("panic: reg returned failure to grok_bslash_N, flags=%#" UVxf,
13516 (UV) flags);
13517 }
13518 *flagp |= flags&(HASWIDTH|SIMPLE|POSTPONED);
13519
13520 nextchar(pRExC_state);
13521
13522 return TRUE;
13523 }
13524
13525
13526 STATIC U8
S_compute_EXACTish(RExC_state_t * pRExC_state)13527 S_compute_EXACTish(RExC_state_t *pRExC_state)
13528 {
13529 U8 op;
13530
13531 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
13532
13533 if (! FOLD) {
13534 return (LOC)
13535 ? EXACTL
13536 : EXACT;
13537 }
13538
13539 op = get_regex_charset(RExC_flags);
13540 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
13541 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
13542 been, so there is no hole */
13543 }
13544
13545 return op + EXACTF;
13546 }
13547
13548 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
13549 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
13550
13551 static I32
S_backref_value(char * p,char * e)13552 S_backref_value(char *p, char *e)
13553 {
13554 const char* endptr = e;
13555 UV val;
13556 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
13557 return (I32)val;
13558 return I32_MAX;
13559 }
13560
13561
13562 /*
13563 - regatom - the lowest level
13564
13565 Try to identify anything special at the start of the current parse position.
13566 If there is, then handle it as required. This may involve generating a
13567 single regop, such as for an assertion; or it may involve recursing, such as
13568 to handle a () structure.
13569
13570 If the string doesn't start with something special then we gobble up
13571 as much literal text as we can. If we encounter a quantifier, we have to
13572 back off the final literal character, as that quantifier applies to just it
13573 and not to the whole string of literals.
13574
13575 Once we have been able to handle whatever type of thing started the
13576 sequence, we return the offset into the regex engine program being compiled
13577 at which any next regnode should be placed.
13578
13579 Returns 0, setting *flagp to TRYAGAIN if reg() returns 0 with TRYAGAIN.
13580 Returns 0, setting *flagp to RESTART_PARSE if the parse needs to be
13581 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
13582 Otherwise does not return 0.
13583
13584 Note: we have to be careful with escapes, as they can be both literal
13585 and special, and in the case of \10 and friends, context determines which.
13586
13587 A summary of the code structure is:
13588
13589 switch (first_byte) {
13590 cases for each special:
13591 handle this special;
13592 break;
13593 case '\\':
13594 switch (2nd byte) {
13595 cases for each unambiguous special:
13596 handle this special;
13597 break;
13598 cases for each ambigous special/literal:
13599 disambiguate;
13600 if (special) handle here
13601 else goto defchar;
13602 default: // unambiguously literal:
13603 goto defchar;
13604 }
13605 default: // is a literal char
13606 // FALL THROUGH
13607 defchar:
13608 create EXACTish node for literal;
13609 while (more input and node isn't full) {
13610 switch (input_byte) {
13611 cases for each special;
13612 make sure parse pointer is set so that the next call to
13613 regatom will see this special first
13614 goto loopdone; // EXACTish node terminated by prev. char
13615 default:
13616 append char to EXACTISH node;
13617 }
13618 get next input byte;
13619 }
13620 loopdone:
13621 }
13622 return the generated node;
13623
13624 Specifically there are two separate switches for handling
13625 escape sequences, with the one for handling literal escapes requiring
13626 a dummy entry for all of the special escapes that are actually handled
13627 by the other.
13628
13629 */
13630
13631 STATIC regnode_offset
S_regatom(pTHX_ RExC_state_t * pRExC_state,I32 * flagp,U32 depth)13632 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
13633 {
13634 regnode_offset ret = 0;
13635 I32 flags = 0;
13636 char *parse_start;
13637 U8 op;
13638 int invert = 0;
13639
13640 DECLARE_AND_GET_RE_DEBUG_FLAGS;
13641
13642 *flagp = 0; /* Initialize. */
13643
13644 DEBUG_PARSE("atom");
13645
13646 PERL_ARGS_ASSERT_REGATOM;
13647
13648 tryagain:
13649 parse_start = RExC_parse;
13650 assert(RExC_parse < RExC_end);
13651 switch ((U8)*RExC_parse) {
13652 case '^':
13653 RExC_seen_zerolen++;
13654 nextchar(pRExC_state);
13655 if (RExC_flags & RXf_PMf_MULTILINE)
13656 ret = reg_node(pRExC_state, MBOL);
13657 else
13658 ret = reg_node(pRExC_state, SBOL);
13659 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13660 break;
13661 case '$':
13662 nextchar(pRExC_state);
13663 if (*RExC_parse)
13664 RExC_seen_zerolen++;
13665 if (RExC_flags & RXf_PMf_MULTILINE)
13666 ret = reg_node(pRExC_state, MEOL);
13667 else
13668 ret = reg_node(pRExC_state, SEOL);
13669 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13670 break;
13671 case '.':
13672 nextchar(pRExC_state);
13673 if (RExC_flags & RXf_PMf_SINGLELINE)
13674 ret = reg_node(pRExC_state, SANY);
13675 else
13676 ret = reg_node(pRExC_state, REG_ANY);
13677 *flagp |= HASWIDTH|SIMPLE;
13678 MARK_NAUGHTY(1);
13679 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13680 break;
13681 case '[':
13682 {
13683 char * const oregcomp_parse = ++RExC_parse;
13684 ret = regclass(pRExC_state, flagp, depth+1,
13685 FALSE, /* means parse the whole char class */
13686 TRUE, /* allow multi-char folds */
13687 FALSE, /* don't silence non-portable warnings. */
13688 (bool) RExC_strict,
13689 TRUE, /* Allow an optimized regnode result */
13690 NULL);
13691 if (ret == 0) {
13692 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13693 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13694 (UV) *flagp);
13695 }
13696 if (*RExC_parse != ']') {
13697 RExC_parse = oregcomp_parse;
13698 vFAIL("Unmatched [");
13699 }
13700 nextchar(pRExC_state);
13701 Set_Node_Length(REGNODE_p(ret), RExC_parse - oregcomp_parse + 1); /* MJD */
13702 break;
13703 }
13704 case '(':
13705 nextchar(pRExC_state);
13706 ret = reg(pRExC_state, 2, &flags, depth+1);
13707 if (ret == 0) {
13708 if (flags & TRYAGAIN) {
13709 if (RExC_parse >= RExC_end) {
13710 /* Make parent create an empty node if needed. */
13711 *flagp |= TRYAGAIN;
13712 return(0);
13713 }
13714 goto tryagain;
13715 }
13716 RETURN_FAIL_ON_RESTART(flags, flagp);
13717 FAIL2("panic: reg returned failure to regatom, flags=%#" UVxf,
13718 (UV) flags);
13719 }
13720 *flagp |= flags&(HASWIDTH|SIMPLE|POSTPONED);
13721 break;
13722 case '|':
13723 case ')':
13724 if (flags & TRYAGAIN) {
13725 *flagp |= TRYAGAIN;
13726 return 0;
13727 }
13728 vFAIL("Internal urp");
13729 /* Supposed to be caught earlier. */
13730 break;
13731 case '?':
13732 case '+':
13733 case '*':
13734 RExC_parse++;
13735 vFAIL("Quantifier follows nothing");
13736 break;
13737 case '\\':
13738 /* Special Escapes
13739
13740 This switch handles escape sequences that resolve to some kind
13741 of special regop and not to literal text. Escape sequences that
13742 resolve to literal text are handled below in the switch marked
13743 "Literal Escapes".
13744
13745 Every entry in this switch *must* have a corresponding entry
13746 in the literal escape switch. However, the opposite is not
13747 required, as the default for this switch is to jump to the
13748 literal text handling code.
13749 */
13750 RExC_parse++;
13751 switch ((U8)*RExC_parse) {
13752 /* Special Escapes */
13753 case 'A':
13754 RExC_seen_zerolen++;
13755 /* Under wildcards, this is changed to match \n; should be
13756 * invisible to the user, as they have to compile under /m */
13757 if (RExC_pm_flags & PMf_WILDCARD) {
13758 ret = reg_node(pRExC_state, MBOL);
13759 }
13760 else {
13761 ret = reg_node(pRExC_state, SBOL);
13762 /* SBOL is shared with /^/ so we set the flags so we can tell
13763 * /\A/ from /^/ in split. */
13764 FLAGS(REGNODE_p(ret)) = 1;
13765 }
13766 goto finish_meta_pat;
13767 case 'G':
13768 if (RExC_pm_flags & PMf_WILDCARD) {
13769 RExC_parse++;
13770 /* diag_listed_as: Use of %s is not allowed in Unicode property
13771 wildcard subpatterns in regex; marked by <-- HERE in m/%s/
13772 */
13773 vFAIL("Use of '\\G' is not allowed in Unicode property"
13774 " wildcard subpatterns");
13775 }
13776 ret = reg_node(pRExC_state, GPOS);
13777 RExC_seen |= REG_GPOS_SEEN;
13778 goto finish_meta_pat;
13779 case 'K':
13780 if (!RExC_in_lookaround) {
13781 RExC_seen_zerolen++;
13782 ret = reg_node(pRExC_state, KEEPS);
13783 /* XXX:dmq : disabling in-place substitution seems to
13784 * be necessary here to avoid cases of memory corruption, as
13785 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
13786 */
13787 RExC_seen |= REG_LOOKBEHIND_SEEN;
13788 goto finish_meta_pat;
13789 }
13790 else {
13791 ++RExC_parse; /* advance past the 'K' */
13792 vFAIL("\\K not permitted in lookahead/lookbehind");
13793 }
13794 case 'Z':
13795 if (RExC_pm_flags & PMf_WILDCARD) {
13796 /* See comment under \A above */
13797 ret = reg_node(pRExC_state, MEOL);
13798 }
13799 else {
13800 ret = reg_node(pRExC_state, SEOL);
13801 }
13802 RExC_seen_zerolen++; /* Do not optimize RE away */
13803 goto finish_meta_pat;
13804 case 'z':
13805 if (RExC_pm_flags & PMf_WILDCARD) {
13806 /* See comment under \A above */
13807 ret = reg_node(pRExC_state, MEOL);
13808 }
13809 else {
13810 ret = reg_node(pRExC_state, EOS);
13811 }
13812 RExC_seen_zerolen++; /* Do not optimize RE away */
13813 goto finish_meta_pat;
13814 case 'C':
13815 vFAIL("\\C no longer supported");
13816 case 'X':
13817 ret = reg_node(pRExC_state, CLUMP);
13818 *flagp |= HASWIDTH;
13819 goto finish_meta_pat;
13820
13821 case 'B':
13822 invert = 1;
13823 /* FALLTHROUGH */
13824 case 'b':
13825 {
13826 U8 flags = 0;
13827 regex_charset charset = get_regex_charset(RExC_flags);
13828
13829 RExC_seen_zerolen++;
13830 RExC_seen |= REG_LOOKBEHIND_SEEN;
13831 op = BOUND + charset;
13832
13833 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
13834 flags = TRADITIONAL_BOUND;
13835 if (op > BOUNDA) { /* /aa is same as /a */
13836 op = BOUNDA;
13837 }
13838 }
13839 else {
13840 STRLEN length;
13841 char name = *RExC_parse;
13842 char * endbrace = (char *) memchr(RExC_parse, '}',
13843 RExC_end - RExC_parse);
13844 char * e = endbrace;
13845
13846 RExC_parse += 2;
13847
13848 if (! endbrace) {
13849 vFAIL2("Missing right brace on \\%c{}", name);
13850 }
13851
13852 while (isBLANK(*RExC_parse)) {
13853 RExC_parse++;
13854 }
13855
13856 while (RExC_parse < e && isBLANK(*(e - 1))) {
13857 e--;
13858 }
13859
13860 if (e == RExC_parse) {
13861 RExC_parse = endbrace + 1; /* After the '}' */
13862 vFAIL2("Empty \\%c{}", name);
13863 }
13864
13865 length = e - RExC_parse;
13866
13867 switch (*RExC_parse) {
13868 case 'g':
13869 if ( length != 1
13870 && (memNEs(RExC_parse + 1, length - 1, "cb")))
13871 {
13872 goto bad_bound_type;
13873 }
13874 flags = GCB_BOUND;
13875 break;
13876 case 'l':
13877 if (length != 2 || *(RExC_parse + 1) != 'b') {
13878 goto bad_bound_type;
13879 }
13880 flags = LB_BOUND;
13881 break;
13882 case 's':
13883 if (length != 2 || *(RExC_parse + 1) != 'b') {
13884 goto bad_bound_type;
13885 }
13886 flags = SB_BOUND;
13887 break;
13888 case 'w':
13889 if (length != 2 || *(RExC_parse + 1) != 'b') {
13890 goto bad_bound_type;
13891 }
13892 flags = WB_BOUND;
13893 break;
13894 default:
13895 bad_bound_type:
13896 RExC_parse = e;
13897 vFAIL2utf8f(
13898 "'%" UTF8f "' is an unknown bound type",
13899 UTF8fARG(UTF, length, e - length));
13900 NOT_REACHED; /*NOTREACHED*/
13901 }
13902 RExC_parse = endbrace;
13903 REQUIRE_UNI_RULES(flagp, 0);
13904
13905 if (op == BOUND) {
13906 op = BOUNDU;
13907 }
13908 else if (op >= BOUNDA) { /* /aa is same as /a */
13909 op = BOUNDU;
13910 length += 4;
13911
13912 /* Don't have to worry about UTF-8, in this message because
13913 * to get here the contents of the \b must be ASCII */
13914 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
13915 "Using /u for '%.*s' instead of /%s",
13916 (unsigned) length,
13917 endbrace - length + 1,
13918 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
13919 ? ASCII_RESTRICT_PAT_MODS
13920 : ASCII_MORE_RESTRICT_PAT_MODS);
13921 }
13922 }
13923
13924 if (op == BOUND) {
13925 RExC_seen_d_op = TRUE;
13926 }
13927 else if (op == BOUNDL) {
13928 RExC_contains_locale = 1;
13929 }
13930
13931 if (invert) {
13932 op += NBOUND - BOUND;
13933 }
13934
13935 ret = reg_node(pRExC_state, op);
13936 FLAGS(REGNODE_p(ret)) = flags;
13937
13938 goto finish_meta_pat;
13939 }
13940
13941 case 'R':
13942 ret = reg_node(pRExC_state, LNBREAK);
13943 *flagp |= HASWIDTH|SIMPLE;
13944 goto finish_meta_pat;
13945
13946 case 'd':
13947 case 'D':
13948 case 'h':
13949 case 'H':
13950 case 'p':
13951 case 'P':
13952 case 's':
13953 case 'S':
13954 case 'v':
13955 case 'V':
13956 case 'w':
13957 case 'W':
13958 /* These all have the same meaning inside [brackets], and it knows
13959 * how to do the best optimizations for them. So, pretend we found
13960 * these within brackets, and let it do the work */
13961 RExC_parse--;
13962
13963 ret = regclass(pRExC_state, flagp, depth+1,
13964 TRUE, /* means just parse this element */
13965 FALSE, /* don't allow multi-char folds */
13966 FALSE, /* don't silence non-portable warnings. It
13967 would be a bug if these returned
13968 non-portables */
13969 (bool) RExC_strict,
13970 TRUE, /* Allow an optimized regnode result */
13971 NULL);
13972 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13973 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
13974 * multi-char folds are allowed. */
13975 if (!ret)
13976 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13977 (UV) *flagp);
13978
13979 RExC_parse--; /* regclass() leaves this one too far ahead */
13980
13981 finish_meta_pat:
13982 /* The escapes above that don't take a parameter can't be
13983 * followed by a '{'. But 'pX', 'p{foo}' and
13984 * correspondingly 'P' can be */
13985 if ( RExC_parse - parse_start == 1
13986 && UCHARAT(RExC_parse + 1) == '{'
13987 && UNLIKELY(! regcurly(RExC_parse + 1, RExC_end, NULL)))
13988 {
13989 RExC_parse += 2;
13990 vFAIL("Unescaped left brace in regex is illegal here");
13991 }
13992 Set_Node_Offset(REGNODE_p(ret), parse_start);
13993 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1); /* MJD */
13994 nextchar(pRExC_state);
13995 break;
13996 case 'N':
13997 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
13998 * \N{...} evaluates to a sequence of more than one code points).
13999 * The function call below returns a regnode, which is our result.
14000 * The parameters cause it to fail if the \N{} evaluates to a
14001 * single code point; we handle those like any other literal. The
14002 * reason that the multicharacter case is handled here and not as
14003 * part of the EXACtish code is because of quantifiers. In
14004 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
14005 * this way makes that Just Happen. dmq.
14006 * join_exact() will join this up with adjacent EXACTish nodes
14007 * later on, if appropriate. */
14008 ++RExC_parse;
14009 if (grok_bslash_N(pRExC_state,
14010 &ret, /* Want a regnode returned */
14011 NULL, /* Fail if evaluates to a single code
14012 point */
14013 NULL, /* Don't need a count of how many code
14014 points */
14015 flagp,
14016 RExC_strict,
14017 depth)
14018 ) {
14019 break;
14020 }
14021
14022 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
14023
14024 /* Here, evaluates to a single code point. Go get that */
14025 RExC_parse = parse_start;
14026 goto defchar;
14027
14028 case 'k': /* Handle \k<NAME> and \k'NAME' and \k{NAME} */
14029 parse_named_seq: /* Also handle non-numeric \g{...} */
14030 {
14031 char ch;
14032 if ( RExC_parse >= RExC_end - 1
14033 || (( ch = RExC_parse[1]) != '<'
14034 && ch != '\''
14035 && ch != '{'))
14036 {
14037 RExC_parse++;
14038 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
14039 vFAIL2("Sequence %.2s... not terminated", parse_start);
14040 } else {
14041 RExC_parse += 2;
14042 if (ch == '{') {
14043 while (isBLANK(*RExC_parse)) {
14044 RExC_parse++;
14045 }
14046 }
14047 ret = handle_named_backref(pRExC_state,
14048 flagp,
14049 parse_start,
14050 (ch == '<')
14051 ? '>'
14052 : (ch == '{')
14053 ? '}'
14054 : '\'');
14055 }
14056 break;
14057 }
14058 case 'g':
14059 case '1': case '2': case '3': case '4':
14060 case '5': case '6': case '7': case '8': case '9':
14061 {
14062 I32 num;
14063 char * endbrace = NULL;
14064 char * s = RExC_parse;
14065 char * e = RExC_end;
14066
14067 if (*s == 'g') {
14068 bool isrel = 0;
14069
14070 s++;
14071 if (*s == '{') {
14072 endbrace = (char *) memchr(s, '}', RExC_end - s);
14073 if (! endbrace ) {
14074
14075 /* Missing '}'. Position after the number to give
14076 * a better indication to the user of where the
14077 * problem is. */
14078 s++;
14079 if (*s == '-') {
14080 s++;
14081 }
14082
14083 /* If it looks to be a name and not a number, go
14084 * handle it there */
14085 if (! isDIGIT(*s)) {
14086 goto parse_named_seq;
14087 }
14088
14089 do {
14090 s++;
14091 } while isDIGIT(*s);
14092
14093 RExC_parse = s;
14094 vFAIL("Unterminated \\g{...} pattern");
14095 }
14096
14097 s++; /* Past the '{' */
14098
14099 while (isBLANK(*s)) {
14100 s++;
14101 }
14102
14103 /* Ignore trailing blanks */
14104 e = endbrace;
14105 while (s < e && isBLANK(*(e - 1))) {
14106 e--;
14107 }
14108 }
14109
14110 /* Here, have isolated the meat of the construct from any
14111 * surrounding braces */
14112
14113 if (*s == '-') {
14114 isrel = 1;
14115 s++;
14116 }
14117
14118 if (endbrace && !isDIGIT(*s)) {
14119 goto parse_named_seq;
14120 }
14121
14122 RExC_parse = s;
14123 num = S_backref_value(RExC_parse, RExC_end);
14124 if (num == 0)
14125 vFAIL("Reference to invalid group 0");
14126 else if (num == I32_MAX) {
14127 if (isDIGIT(*RExC_parse))
14128 vFAIL("Reference to nonexistent group");
14129 else
14130 vFAIL("Unterminated \\g... pattern");
14131 }
14132
14133 if (isrel) {
14134 num = RExC_npar - num;
14135 if (num < 1)
14136 vFAIL("Reference to nonexistent or unclosed group");
14137 }
14138 }
14139 else {
14140 num = S_backref_value(RExC_parse, RExC_end);
14141 /* bare \NNN might be backref or octal - if it is larger
14142 * than or equal RExC_npar then it is assumed to be an
14143 * octal escape. Note RExC_npar is +1 from the actual
14144 * number of parens. */
14145 /* Note we do NOT check if num == I32_MAX here, as that is
14146 * handled by the RExC_npar check */
14147
14148 if ( /* any numeric escape < 10 is always a backref */
14149 num > 9
14150 /* any numeric escape < RExC_npar is a backref */
14151 && num >= RExC_npar
14152 /* cannot be an octal escape if it starts with [89]
14153 * */
14154 && ! inRANGE(*RExC_parse, '8', '9')
14155 ) {
14156 /* Probably not meant to be a backref, instead likely
14157 * to be an octal character escape, e.g. \35 or \777.
14158 * The above logic should make it obvious why using
14159 * octal escapes in patterns is problematic. - Yves */
14160 RExC_parse = parse_start;
14161 goto defchar;
14162 }
14163 }
14164
14165 /* At this point RExC_parse points at a numeric escape like
14166 * \12 or \88 or the digits in \g{34} or \g34 or something
14167 * similar, which we should NOT treat as an octal escape. It
14168 * may or may not be a valid backref escape. For instance
14169 * \88888888 is unlikely to be a valid backref.
14170 *
14171 * We've already figured out what value the digits represent.
14172 * Now, move the parse to beyond them. */
14173 if (endbrace) {
14174 RExC_parse = endbrace + 1;
14175 }
14176 else while (isDIGIT(*RExC_parse)) {
14177 RExC_parse++;
14178 }
14179
14180 if (num >= (I32)RExC_npar) {
14181
14182 /* It might be a forward reference; we can't fail until we
14183 * know, by completing the parse to get all the groups, and
14184 * then reparsing */
14185 if (ALL_PARENS_COUNTED) {
14186 if (num >= RExC_total_parens) {
14187 vFAIL("Reference to nonexistent group");
14188 }
14189 }
14190 else {
14191 REQUIRE_PARENS_PASS;
14192 }
14193 }
14194 RExC_sawback = 1;
14195 ret = reganode(pRExC_state,
14196 ((! FOLD)
14197 ? REF
14198 : (ASCII_FOLD_RESTRICTED)
14199 ? REFFA
14200 : (AT_LEAST_UNI_SEMANTICS)
14201 ? REFFU
14202 : (LOC)
14203 ? REFFL
14204 : REFF),
14205 num);
14206 if (OP(REGNODE_p(ret)) == REFF) {
14207 RExC_seen_d_op = TRUE;
14208 }
14209 *flagp |= HASWIDTH;
14210
14211 /* override incorrect value set in reganode MJD */
14212 Set_Node_Offset(REGNODE_p(ret), parse_start);
14213 Set_Node_Cur_Length(REGNODE_p(ret), parse_start-1);
14214 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14215 FALSE /* Don't force to /x */ );
14216 }
14217 break;
14218 case '\0':
14219 if (RExC_parse >= RExC_end)
14220 FAIL("Trailing \\");
14221 /* FALLTHROUGH */
14222 default:
14223 /* Do not generate "unrecognized" warnings here, we fall
14224 back into the quick-grab loop below */
14225 RExC_parse = parse_start;
14226 goto defchar;
14227 } /* end of switch on a \foo sequence */
14228 break;
14229
14230 case '#':
14231
14232 /* '#' comments should have been spaced over before this function was
14233 * called */
14234 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
14235 /*
14236 if (RExC_flags & RXf_PMf_EXTENDED) {
14237 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
14238 if (RExC_parse < RExC_end)
14239 goto tryagain;
14240 }
14241 */
14242
14243 /* FALLTHROUGH */
14244
14245 default:
14246 defchar: {
14247
14248 /* Here, we have determined that the next thing is probably a
14249 * literal character. RExC_parse points to the first byte of its
14250 * definition. (It still may be an escape sequence that evaluates
14251 * to a single character) */
14252
14253 STRLEN len = 0;
14254 UV ender = 0;
14255 char *p;
14256 char *s, *old_s = NULL, *old_old_s = NULL;
14257 char *s0;
14258 U32 max_string_len = 255;
14259
14260 /* We may have to reparse the node, artificially stopping filling
14261 * it early, based on info gleaned in the first parse. This
14262 * variable gives where we stop. Make it above the normal stopping
14263 * place first time through; otherwise it would stop too early */
14264 U32 upper_fill = max_string_len + 1;
14265
14266 /* We start out as an EXACT node, even if under /i, until we find a
14267 * character which is in a fold. The algorithm now segregates into
14268 * separate nodes, characters that fold from those that don't under
14269 * /i. (This hopefully will create nodes that are fixed strings
14270 * even under /i, giving the optimizer something to grab on to.)
14271 * So, if a node has something in it and the next character is in
14272 * the opposite category, that node is closed up, and the function
14273 * returns. Then regatom is called again, and a new node is
14274 * created for the new category. */
14275 U8 node_type = EXACT;
14276
14277 /* Assume the node will be fully used; the excess is given back at
14278 * the end. Under /i, we may need to temporarily add the fold of
14279 * an extra character or two at the end to check for splitting
14280 * multi-char folds, so allocate extra space for that. We can't
14281 * make any other length assumptions, as a byte input sequence
14282 * could shrink down. */
14283 Ptrdiff_t current_string_nodes = STR_SZ(max_string_len
14284 + ((! FOLD)
14285 ? 0
14286 : 2 * ((UTF)
14287 ? UTF8_MAXBYTES_CASE
14288 /* Max non-UTF-8 expansion is 2 */ : 2)));
14289
14290 bool next_is_quantifier;
14291 char * oldp = NULL;
14292
14293 /* We can convert EXACTF nodes to EXACTFU if they contain only
14294 * characters that match identically regardless of the target
14295 * string's UTF8ness. The reason to do this is that EXACTF is not
14296 * trie-able, EXACTFU is, and EXACTFU requires fewer operations at
14297 * runtime.
14298 *
14299 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
14300 * contain only above-Latin1 characters (hence must be in UTF8),
14301 * which don't participate in folds with Latin1-range characters,
14302 * as the latter's folds aren't known until runtime. */
14303 bool maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
14304
14305 /* Single-character EXACTish nodes are almost always SIMPLE. This
14306 * allows us to override this as encountered */
14307 U8 maybe_SIMPLE = SIMPLE;
14308
14309 /* Does this node contain something that can't match unless the
14310 * target string is (also) in UTF-8 */
14311 bool requires_utf8_target = FALSE;
14312
14313 /* The sequence 'ss' is problematic in non-UTF-8 patterns. */
14314 bool has_ss = FALSE;
14315
14316 /* So is the MICRO SIGN */
14317 bool has_micro_sign = FALSE;
14318
14319 /* Set when we fill up the current node and there is still more
14320 * text to process */
14321 bool overflowed;
14322
14323 /* Allocate an EXACT node. The node_type may change below to
14324 * another EXACTish node, but since the size of the node doesn't
14325 * change, it works */
14326 ret = regnode_guts(pRExC_state, node_type, current_string_nodes,
14327 "exact");
14328 FILL_NODE(ret, node_type);
14329 RExC_emit++;
14330
14331 s = STRING(REGNODE_p(ret));
14332
14333 s0 = s;
14334
14335 reparse:
14336
14337 p = RExC_parse;
14338 len = 0;
14339 s = s0;
14340 node_type = EXACT;
14341 oldp = NULL;
14342 maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
14343 maybe_SIMPLE = SIMPLE;
14344 requires_utf8_target = FALSE;
14345 has_ss = FALSE;
14346 has_micro_sign = FALSE;
14347
14348 continue_parse:
14349
14350 /* This breaks under rare circumstances. If folding, we do not
14351 * want to split a node at a character that is a non-final in a
14352 * multi-char fold, as an input string could just happen to want to
14353 * match across the node boundary. The code at the end of the loop
14354 * looks for this, and backs off until it finds not such a
14355 * character, but it is possible (though extremely, extremely
14356 * unlikely) for all characters in the node to be non-final fold
14357 * ones, in which case we just leave the node fully filled, and
14358 * hope that it doesn't match the string in just the wrong place */
14359
14360 assert( ! UTF /* Is at the beginning of a character */
14361 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
14362 || UTF8_IS_START(UCHARAT(RExC_parse)));
14363
14364 overflowed = FALSE;
14365
14366 /* Here, we have a literal character. Find the maximal string of
14367 * them in the input that we can fit into a single EXACTish node.
14368 * We quit at the first non-literal or when the node gets full, or
14369 * under /i the categorization of folding/non-folding character
14370 * changes */
14371 while (p < RExC_end && len < upper_fill) {
14372
14373 /* In most cases each iteration adds one byte to the output.
14374 * The exceptions override this */
14375 Size_t added_len = 1;
14376
14377 oldp = p;
14378 old_old_s = old_s;
14379 old_s = s;
14380
14381 /* White space has already been ignored */
14382 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
14383 || ! is_PATWS_safe((p), RExC_end, UTF));
14384
14385 switch ((U8)*p) {
14386 const char* message;
14387 U32 packed_warn;
14388 U8 grok_c_char;
14389
14390 case '^':
14391 case '$':
14392 case '.':
14393 case '[':
14394 case '(':
14395 case ')':
14396 case '|':
14397 goto loopdone;
14398 case '\\':
14399 /* Literal Escapes Switch
14400
14401 This switch is meant to handle escape sequences that
14402 resolve to a literal character.
14403
14404 Every escape sequence that represents something
14405 else, like an assertion or a char class, is handled
14406 in the switch marked 'Special Escapes' above in this
14407 routine, but also has an entry here as anything that
14408 isn't explicitly mentioned here will be treated as
14409 an unescaped equivalent literal.
14410 */
14411
14412 switch ((U8)*++p) {
14413
14414 /* These are all the special escapes. */
14415 case 'A': /* Start assertion */
14416 case 'b': case 'B': /* Word-boundary assertion*/
14417 case 'C': /* Single char !DANGEROUS! */
14418 case 'd': case 'D': /* digit class */
14419 case 'g': case 'G': /* generic-backref, pos assertion */
14420 case 'h': case 'H': /* HORIZWS */
14421 case 'k': case 'K': /* named backref, keep marker */
14422 case 'p': case 'P': /* Unicode property */
14423 case 'R': /* LNBREAK */
14424 case 's': case 'S': /* space class */
14425 case 'v': case 'V': /* VERTWS */
14426 case 'w': case 'W': /* word class */
14427 case 'X': /* eXtended Unicode "combining
14428 character sequence" */
14429 case 'z': case 'Z': /* End of line/string assertion */
14430 --p;
14431 goto loopdone;
14432
14433 /* Anything after here is an escape that resolves to a
14434 literal. (Except digits, which may or may not)
14435 */
14436 case 'n':
14437 ender = '\n';
14438 p++;
14439 break;
14440 case 'N': /* Handle a single-code point named character. */
14441 RExC_parse = p + 1;
14442 if (! grok_bslash_N(pRExC_state,
14443 NULL, /* Fail if evaluates to
14444 anything other than a
14445 single code point */
14446 &ender, /* The returned single code
14447 point */
14448 NULL, /* Don't need a count of
14449 how many code points */
14450 flagp,
14451 RExC_strict,
14452 depth)
14453 ) {
14454 if (*flagp & NEED_UTF8)
14455 FAIL("panic: grok_bslash_N set NEED_UTF8");
14456 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
14457
14458 /* Here, it wasn't a single code point. Go close
14459 * up this EXACTish node. The switch() prior to
14460 * this switch handles the other cases */
14461 RExC_parse = p = oldp;
14462 goto loopdone;
14463 }
14464 p = RExC_parse;
14465 RExC_parse = parse_start;
14466
14467 /* The \N{} means the pattern, if previously /d,
14468 * becomes /u. That means it can't be an EXACTF node,
14469 * but an EXACTFU */
14470 if (node_type == EXACTF) {
14471 node_type = EXACTFU;
14472
14473 /* If the node already contains something that
14474 * differs between EXACTF and EXACTFU, reparse it
14475 * as EXACTFU */
14476 if (! maybe_exactfu) {
14477 len = 0;
14478 s = s0;
14479 goto reparse;
14480 }
14481 }
14482
14483 break;
14484 case 'r':
14485 ender = '\r';
14486 p++;
14487 break;
14488 case 't':
14489 ender = '\t';
14490 p++;
14491 break;
14492 case 'f':
14493 ender = '\f';
14494 p++;
14495 break;
14496 case 'e':
14497 ender = ESC_NATIVE;
14498 p++;
14499 break;
14500 case 'a':
14501 ender = '\a';
14502 p++;
14503 break;
14504 case 'o':
14505 if (! grok_bslash_o(&p,
14506 RExC_end,
14507 &ender,
14508 &message,
14509 &packed_warn,
14510 (bool) RExC_strict,
14511 FALSE, /* No illegal cp's */
14512 UTF))
14513 {
14514 RExC_parse = p; /* going to die anyway; point to
14515 exact spot of failure */
14516 vFAIL(message);
14517 }
14518
14519 if (message && TO_OUTPUT_WARNINGS(p)) {
14520 warn_non_literal_string(p, packed_warn, message);
14521 }
14522 break;
14523 case 'x':
14524 if (! grok_bslash_x(&p,
14525 RExC_end,
14526 &ender,
14527 &message,
14528 &packed_warn,
14529 (bool) RExC_strict,
14530 FALSE, /* No illegal cp's */
14531 UTF))
14532 {
14533 RExC_parse = p; /* going to die anyway; point
14534 to exact spot of failure */
14535 vFAIL(message);
14536 }
14537
14538 if (message && TO_OUTPUT_WARNINGS(p)) {
14539 warn_non_literal_string(p, packed_warn, message);
14540 }
14541
14542 #ifdef EBCDIC
14543 if (ender < 0x100) {
14544 if (RExC_recode_x_to_native) {
14545 ender = LATIN1_TO_NATIVE(ender);
14546 }
14547 }
14548 #endif
14549 break;
14550 case 'c':
14551 p++;
14552 if (! grok_bslash_c(*p, &grok_c_char,
14553 &message, &packed_warn))
14554 {
14555 /* going to die anyway; point to exact spot of
14556 * failure */
14557 RExC_parse = p + ((UTF)
14558 ? UTF8_SAFE_SKIP(p, RExC_end)
14559 : 1);
14560 vFAIL(message);
14561 }
14562
14563 ender = grok_c_char;
14564 p++;
14565 if (message && TO_OUTPUT_WARNINGS(p)) {
14566 warn_non_literal_string(p, packed_warn, message);
14567 }
14568
14569 break;
14570 case '8': case '9': /* must be a backreference */
14571 --p;
14572 /* we have an escape like \8 which cannot be an octal escape
14573 * so we exit the loop, and let the outer loop handle this
14574 * escape which may or may not be a legitimate backref. */
14575 goto loopdone;
14576 case '1': case '2': case '3':case '4':
14577 case '5': case '6': case '7':
14578
14579 /* When we parse backslash escapes there is ambiguity
14580 * between backreferences and octal escapes. Any escape
14581 * from \1 - \9 is a backreference, any multi-digit
14582 * escape which does not start with 0 and which when
14583 * evaluated as decimal could refer to an already
14584 * parsed capture buffer is a back reference. Anything
14585 * else is octal.
14586 *
14587 * Note this implies that \118 could be interpreted as
14588 * 118 OR as "\11" . "8" depending on whether there
14589 * were 118 capture buffers defined already in the
14590 * pattern. */
14591
14592 /* NOTE, RExC_npar is 1 more than the actual number of
14593 * parens we have seen so far, hence the "<" as opposed
14594 * to "<=" */
14595 if ( !isDIGIT(p[1]) || S_backref_value(p, RExC_end) < RExC_npar)
14596 { /* Not to be treated as an octal constant, go
14597 find backref */
14598 p = oldp;
14599 goto loopdone;
14600 }
14601 /* FALLTHROUGH */
14602 case '0':
14603 {
14604 I32 flags = PERL_SCAN_SILENT_ILLDIGIT
14605 | PERL_SCAN_NOTIFY_ILLDIGIT;
14606 STRLEN numlen = 3;
14607 ender = grok_oct(p, &numlen, &flags, NULL);
14608 p += numlen;
14609 if ( (flags & PERL_SCAN_NOTIFY_ILLDIGIT)
14610 && isDIGIT(*p) /* like \08, \178 */
14611 && ckWARN(WARN_REGEXP))
14612 {
14613 reg_warn_non_literal_string(
14614 p + 1,
14615 form_alien_digit_msg(8, numlen, p,
14616 RExC_end, UTF, FALSE));
14617 }
14618 }
14619 break;
14620 case '\0':
14621 if (p >= RExC_end)
14622 FAIL("Trailing \\");
14623 /* FALLTHROUGH */
14624 default:
14625 if (isALPHANUMERIC(*p)) {
14626 /* An alpha followed by '{' is going to fail next
14627 * iteration, so don't output this warning in that
14628 * case */
14629 if (! isALPHA(*p) || *(p + 1) != '{') {
14630 ckWARN2reg(p + 1, "Unrecognized escape \\%.1s"
14631 " passed through", p);
14632 }
14633 }
14634 goto normal_default;
14635 } /* End of switch on '\' */
14636 break;
14637 case '{':
14638 /* Trying to gain new uses for '{' without breaking too
14639 * much existing code is hard. The solution currently
14640 * adopted is:
14641 * 1) If there is no ambiguity that a '{' should always
14642 * be taken literally, at the start of a construct, we
14643 * just do so.
14644 * 2) If the literal '{' conflicts with our desired use
14645 * of it as a metacharacter, we die. The deprecation
14646 * cycles for this have come and gone.
14647 * 3) If there is ambiguity, we raise a simple warning.
14648 * This could happen, for example, if the user
14649 * intended it to introduce a quantifier, but slightly
14650 * misspelled the quantifier. Without this warning,
14651 * the quantifier would silently be taken as a literal
14652 * string of characters instead of a meta construct */
14653 if (len || (p > RExC_start && isALPHA_A(*(p - 1)))) {
14654 if ( RExC_strict
14655 || ( p > parse_start + 1
14656 && isALPHA_A(*(p - 1))
14657 && *(p - 2) == '\\'))
14658 {
14659 RExC_parse = p + 1;
14660 vFAIL("Unescaped left brace in regex is "
14661 "illegal here");
14662 }
14663 ckWARNreg(p + 1, "Unescaped left brace in regex is"
14664 " passed through");
14665 }
14666 goto normal_default;
14667 case '}':
14668 case ']':
14669 if (p > RExC_parse && RExC_strict) {
14670 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
14671 }
14672 /*FALLTHROUGH*/
14673 default: /* A literal character */
14674 normal_default:
14675 if (! UTF8_IS_INVARIANT(*p) && UTF) {
14676 STRLEN numlen;
14677 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
14678 &numlen, UTF8_ALLOW_DEFAULT);
14679 p += numlen;
14680 }
14681 else
14682 ender = (U8) *p++;
14683 break;
14684 } /* End of switch on the literal */
14685
14686 /* Here, have looked at the literal character, and <ender>
14687 * contains its ordinal; <p> points to the character after it.
14688 * */
14689
14690 if (ender > 255) {
14691 REQUIRE_UTF8(flagp);
14692 if ( UNICODE_IS_PERL_EXTENDED(ender)
14693 && TO_OUTPUT_WARNINGS(p))
14694 {
14695 ckWARN2_non_literal_string(p,
14696 packWARN(WARN_PORTABLE),
14697 PL_extended_cp_format,
14698 ender);
14699 }
14700 }
14701
14702 /* We need to check if the next non-ignored thing is a
14703 * quantifier. Move <p> to after anything that should be
14704 * ignored, which, as a side effect, positions <p> for the next
14705 * loop iteration */
14706 skip_to_be_ignored_text(pRExC_state, &p,
14707 FALSE /* Don't force to /x */ );
14708
14709 /* If the next thing is a quantifier, it applies to this
14710 * character only, which means that this character has to be in
14711 * its own node and can't just be appended to the string in an
14712 * existing node, so if there are already other characters in
14713 * the node, close the node with just them, and set up to do
14714 * this character again next time through, when it will be the
14715 * only thing in its new node */
14716
14717 next_is_quantifier = LIKELY(p < RExC_end)
14718 && UNLIKELY(isQUANTIFIER(p, RExC_end));
14719
14720 if (next_is_quantifier && LIKELY(len)) {
14721 p = oldp;
14722 goto loopdone;
14723 }
14724
14725 /* Ready to add 'ender' to the node */
14726
14727 if (! FOLD) { /* The simple case, just append the literal */
14728 not_fold_common:
14729
14730 /* Don't output if it would overflow */
14731 if (UNLIKELY(len > max_string_len - ((UTF)
14732 ? UVCHR_SKIP(ender)
14733 : 1)))
14734 {
14735 overflowed = TRUE;
14736 break;
14737 }
14738
14739 if (UVCHR_IS_INVARIANT(ender) || ! UTF) {
14740 *(s++) = (char) ender;
14741 }
14742 else {
14743 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
14744 added_len = (char *) new_s - s;
14745 s = (char *) new_s;
14746
14747 if (ender > 255) {
14748 requires_utf8_target = TRUE;
14749 }
14750 }
14751 }
14752 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
14753
14754 /* Here are folding under /l, and the code point is
14755 * problematic. If this is the first character in the
14756 * node, change the node type to folding. Otherwise, if
14757 * this is the first problematic character, close up the
14758 * existing node, so can start a new node with this one */
14759 if (! len) {
14760 node_type = EXACTFL;
14761 RExC_contains_locale = 1;
14762 }
14763 else if (node_type == EXACT) {
14764 p = oldp;
14765 goto loopdone;
14766 }
14767
14768 /* This problematic code point means we can't simplify
14769 * things */
14770 maybe_exactfu = FALSE;
14771
14772 /* Although these two characters have folds that are
14773 * locale-problematic, they also have folds to above Latin1
14774 * that aren't a problem. Doing these now helps at
14775 * runtime. */
14776 if (UNLIKELY( ender == GREEK_CAPITAL_LETTER_MU
14777 || ender == LATIN_CAPITAL_LETTER_SHARP_S))
14778 {
14779 goto fold_anyway;
14780 }
14781
14782 /* Here, we are adding a problematic fold character.
14783 * "Problematic" in this context means that its fold isn't
14784 * known until runtime. (The non-problematic code points
14785 * are the above-Latin1 ones that fold to also all
14786 * above-Latin1. Their folds don't vary no matter what the
14787 * locale is.) But here we have characters whose fold
14788 * depends on the locale. We just add in the unfolded
14789 * character, and wait until runtime to fold it */
14790 goto not_fold_common;
14791 }
14792 else /* regular fold; see if actually is in a fold */
14793 if ( (ender < 256 && ! IS_IN_SOME_FOLD_L1(ender))
14794 || (ender > 255
14795 && ! _invlist_contains_cp(PL_in_some_fold, ender)))
14796 {
14797 /* Here, folding, but the character isn't in a fold.
14798 *
14799 * Start a new node if previous characters in the node were
14800 * folded */
14801 if (len && node_type != EXACT) {
14802 p = oldp;
14803 goto loopdone;
14804 }
14805
14806 /* Here, continuing a node with non-folded characters. Add
14807 * this one */
14808 goto not_fold_common;
14809 }
14810 else { /* Here, does participate in some fold */
14811
14812 /* If this is the first character in the node, change its
14813 * type to folding. Otherwise, if this is the first
14814 * folding character in the node, close up the existing
14815 * node, so can start a new node with this one. */
14816 if (! len) {
14817 node_type = compute_EXACTish(pRExC_state);
14818 }
14819 else if (node_type == EXACT) {
14820 p = oldp;
14821 goto loopdone;
14822 }
14823
14824 if (UTF) { /* Alway use the folded value for UTF-8
14825 patterns */
14826 if (UVCHR_IS_INVARIANT(ender)) {
14827 if (UNLIKELY(len + 1 > max_string_len)) {
14828 overflowed = TRUE;
14829 break;
14830 }
14831
14832 *(s)++ = (U8) toFOLD(ender);
14833 }
14834 else {
14835 UV folded;
14836
14837 fold_anyway:
14838 folded = _to_uni_fold_flags(
14839 ender,
14840 (U8 *) s, /* We have allocated extra space
14841 in 's' so can't run off the
14842 end */
14843 &added_len,
14844 FOLD_FLAGS_FULL
14845 | (( ASCII_FOLD_RESTRICTED
14846 || node_type == EXACTFL)
14847 ? FOLD_FLAGS_NOMIX_ASCII
14848 : 0));
14849 if (UNLIKELY(len + added_len > max_string_len)) {
14850 overflowed = TRUE;
14851 break;
14852 }
14853
14854 s += added_len;
14855
14856 if ( folded > 255
14857 && LIKELY(folded != GREEK_SMALL_LETTER_MU))
14858 {
14859 /* U+B5 folds to the MU, so its possible for a
14860 * non-UTF-8 target to match it */
14861 requires_utf8_target = TRUE;
14862 }
14863 }
14864 }
14865 else { /* Here is non-UTF8. */
14866
14867 /* The fold will be one or (rarely) two characters.
14868 * Check that there's room for at least a single one
14869 * before setting any flags, etc. Because otherwise an
14870 * overflowing character could cause a flag to be set
14871 * even though it doesn't end up in this node. (For
14872 * the two character fold, we check again, before
14873 * setting any flags) */
14874 if (UNLIKELY(len + 1 > max_string_len)) {
14875 overflowed = TRUE;
14876 break;
14877 }
14878
14879 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
14880 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
14881 || UNICODE_DOT_DOT_VERSION > 0)
14882
14883 /* On non-ancient Unicodes, check for the only possible
14884 * multi-char fold */
14885 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
14886
14887 /* This potential multi-char fold means the node
14888 * can't be simple (because it could match more
14889 * than a single char). And in some cases it will
14890 * match 'ss', so set that flag */
14891 maybe_SIMPLE = 0;
14892 has_ss = TRUE;
14893
14894 /* It can't change to be an EXACTFU (unless already
14895 * is one). We fold it iff under /u rules. */
14896 if (node_type != EXACTFU) {
14897 maybe_exactfu = FALSE;
14898 }
14899 else {
14900 if (UNLIKELY(len + 2 > max_string_len)) {
14901 overflowed = TRUE;
14902 break;
14903 }
14904
14905 *(s++) = 's';
14906 *(s++) = 's';
14907 added_len = 2;
14908
14909 goto done_with_this_char;
14910 }
14911 }
14912 else if ( UNLIKELY(isALPHA_FOLD_EQ(ender, 's'))
14913 && LIKELY(len > 0)
14914 && UNLIKELY(isALPHA_FOLD_EQ(*(s-1), 's')))
14915 {
14916 /* Also, the sequence 'ss' is special when not
14917 * under /u. If the target string is UTF-8, it
14918 * should match SHARP S; otherwise it won't. So,
14919 * here we have to exclude the possibility of this
14920 * node moving to /u.*/
14921 has_ss = TRUE;
14922 maybe_exactfu = FALSE;
14923 }
14924 #endif
14925 /* Here, the fold will be a single character */
14926
14927 if (UNLIKELY(ender == MICRO_SIGN)) {
14928 has_micro_sign = TRUE;
14929 }
14930 else if (PL_fold[ender] != PL_fold_latin1[ender]) {
14931
14932 /* If the character's fold differs between /d and
14933 * /u, this can't change to be an EXACTFU node */
14934 maybe_exactfu = FALSE;
14935 }
14936
14937 *(s++) = (DEPENDS_SEMANTICS)
14938 ? (char) toFOLD(ender)
14939
14940 /* Under /u, the fold of any character in
14941 * the 0-255 range happens to be its
14942 * lowercase equivalent, except for LATIN
14943 * SMALL LETTER SHARP S, which was handled
14944 * above, and the MICRO SIGN, whose fold
14945 * requires UTF-8 to represent. */
14946 : (char) toLOWER_L1(ender);
14947 }
14948 } /* End of adding current character to the node */
14949
14950 done_with_this_char:
14951
14952 len += added_len;
14953
14954 if (next_is_quantifier) {
14955
14956 /* Here, the next input is a quantifier, and to get here,
14957 * the current character is the only one in the node. */
14958 goto loopdone;
14959 }
14960
14961 } /* End of loop through literal characters */
14962
14963 /* Here we have either exhausted the input or run out of room in
14964 * the node. If the former, we are done. (If we encountered a
14965 * character that can't be in the node, transfer is made directly
14966 * to <loopdone>, and so we wouldn't have fallen off the end of the
14967 * loop.) */
14968 if (LIKELY(! overflowed)) {
14969 goto loopdone;
14970 }
14971
14972 /* Here we have run out of room. We can grow plain EXACT and
14973 * LEXACT nodes. If the pattern is gigantic enough, though,
14974 * eventually we'll have to artificially chunk the pattern into
14975 * multiple nodes. */
14976 if (! LOC && (node_type == EXACT || node_type == LEXACT)) {
14977 Size_t overhead = 1 + regarglen[OP(REGNODE_p(ret))];
14978 Size_t overhead_expansion = 0;
14979 char temp[256];
14980 Size_t max_nodes_for_string;
14981 Size_t achievable;
14982 SSize_t delta;
14983
14984 /* Here we couldn't fit the final character in the current
14985 * node, so it will have to be reparsed, no matter what else we
14986 * do */
14987 p = oldp;
14988
14989 /* If would have overflowed a regular EXACT node, switch
14990 * instead to an LEXACT. The code below is structured so that
14991 * the actual growing code is common to changing from an EXACT
14992 * or just increasing the LEXACT size. This means that we have
14993 * to save the string in the EXACT case before growing, and
14994 * then copy it afterwards to its new location */
14995 if (node_type == EXACT) {
14996 overhead_expansion = regarglen[LEXACT] - regarglen[EXACT];
14997 RExC_emit += overhead_expansion;
14998 Copy(s0, temp, len, char);
14999 }
15000
15001 /* Ready to grow. If it was a plain EXACT, the string was
15002 * saved, and the first few bytes of it overwritten by adding
15003 * an argument field. We assume, as we do elsewhere in this
15004 * file, that one byte of remaining input will translate into
15005 * one byte of output, and if that's too small, we grow again,
15006 * if too large the excess memory is freed at the end */
15007
15008 max_nodes_for_string = U16_MAX - overhead - overhead_expansion;
15009 achievable = MIN(max_nodes_for_string,
15010 current_string_nodes + STR_SZ(RExC_end - p));
15011 delta = achievable - current_string_nodes;
15012
15013 /* If there is just no more room, go finish up this chunk of
15014 * the pattern. */
15015 if (delta <= 0) {
15016 goto loopdone;
15017 }
15018
15019 change_engine_size(pRExC_state, delta + overhead_expansion);
15020 current_string_nodes += delta;
15021 max_string_len
15022 = sizeof(struct regnode) * current_string_nodes;
15023 upper_fill = max_string_len + 1;
15024
15025 /* If the length was small, we know this was originally an
15026 * EXACT node now converted to LEXACT, and the string has to be
15027 * restored. Otherwise the string was untouched. 260 is just
15028 * a number safely above 255 so don't have to worry about
15029 * getting it precise */
15030 if (len < 260) {
15031 node_type = LEXACT;
15032 FILL_NODE(ret, node_type);
15033 s0 = STRING(REGNODE_p(ret));
15034 Copy(temp, s0, len, char);
15035 s = s0 + len;
15036 }
15037
15038 goto continue_parse;
15039 }
15040 else if (FOLD) {
15041 bool splittable = FALSE;
15042 bool backed_up = FALSE;
15043 char * e; /* should this be U8? */
15044 char * s_start; /* should this be U8? */
15045
15046 /* Here is /i. Running out of room creates a problem if we are
15047 * folding, and the split happens in the middle of a
15048 * multi-character fold, as a match that should have occurred,
15049 * won't, due to the way nodes are matched, and our artificial
15050 * boundary. So back off until we aren't splitting such a
15051 * fold. If there is no such place to back off to, we end up
15052 * taking the entire node as-is. This can happen if the node
15053 * consists entirely of 'f' or entirely of 's' characters (or
15054 * things that fold to them) as 'ff' and 'ss' are
15055 * multi-character folds.
15056 *
15057 * The Unicode standard says that multi character folds consist
15058 * of either two or three characters. That means we would be
15059 * splitting one if the final character in the node is at the
15060 * beginning of either type, or is the second of a three
15061 * character fold.
15062 *
15063 * At this point:
15064 * ender is the code point of the character that won't fit
15065 * in the node
15066 * s points to just beyond the final byte in the node.
15067 * It's where we would place ender if there were
15068 * room, and where in fact we do place ender's fold
15069 * in the code below, as we've over-allocated space
15070 * for s0 (hence s) to allow for this
15071 * e starts at 's' and advances as we append things.
15072 * old_s is the same as 's'. (If ender had fit, 's' would
15073 * have been advanced to beyond it).
15074 * old_old_s points to the beginning byte of the final
15075 * character in the node
15076 * p points to the beginning byte in the input of the
15077 * character beyond 'ender'.
15078 * oldp points to the beginning byte in the input of
15079 * 'ender'.
15080 *
15081 * In the case of /il, we haven't folded anything that could be
15082 * affected by the locale. That means only above-Latin1
15083 * characters that fold to other above-latin1 characters get
15084 * folded at compile time. To check where a good place to
15085 * split nodes is, everything in it will have to be folded.
15086 * The boolean 'maybe_exactfu' keeps track in /il if there are
15087 * any unfolded characters in the node. */
15088 bool need_to_fold_loc = LOC && ! maybe_exactfu;
15089
15090 /* If we do need to fold the node, we need a place to store the
15091 * folded copy, and a way to map back to the unfolded original
15092 * */
15093 char * locfold_buf = NULL;
15094 Size_t * loc_correspondence = NULL;
15095
15096 if (! need_to_fold_loc) { /* The normal case. Just
15097 initialize to the actual node */
15098 e = s;
15099 s_start = s0;
15100 s = old_old_s; /* Point to the beginning of the final char
15101 that fits in the node */
15102 }
15103 else {
15104
15105 /* Here, we have filled a /il node, and there are unfolded
15106 * characters in it. If the runtime locale turns out to be
15107 * UTF-8, there are possible multi-character folds, just
15108 * like when not under /l. The node hence can't terminate
15109 * in the middle of such a fold. To determine this, we
15110 * have to create a folded copy of this node. That means
15111 * reparsing the node, folding everything assuming a UTF-8
15112 * locale. (If at runtime it isn't such a locale, the
15113 * actions here wouldn't have been necessary, but we have
15114 * to assume the worst case.) If we find we need to back
15115 * off the folded string, we do so, and then map that
15116 * position back to the original unfolded node, which then
15117 * gets output, truncated at that spot */
15118
15119 char * redo_p = RExC_parse;
15120 char * redo_e;
15121 char * old_redo_e;
15122
15123 /* Allow enough space assuming a single byte input folds to
15124 * a single byte output, plus assume that the two unparsed
15125 * characters (that we may need) fold to the largest number
15126 * of bytes possible, plus extra for one more worst case
15127 * scenario. In the loop below, if we start eating into
15128 * that final spare space, we enlarge this initial space */
15129 Size_t size = max_string_len + (3 * UTF8_MAXBYTES_CASE) + 1;
15130
15131 Newxz(locfold_buf, size, char);
15132 Newxz(loc_correspondence, size, Size_t);
15133
15134 /* Redo this node's parse, folding into 'locfold_buf' */
15135 redo_p = RExC_parse;
15136 old_redo_e = redo_e = locfold_buf;
15137 while (redo_p <= oldp) {
15138
15139 old_redo_e = redo_e;
15140 loc_correspondence[redo_e - locfold_buf]
15141 = redo_p - RExC_parse;
15142
15143 if (UTF) {
15144 Size_t added_len;
15145
15146 (void) _to_utf8_fold_flags((U8 *) redo_p,
15147 (U8 *) RExC_end,
15148 (U8 *) redo_e,
15149 &added_len,
15150 FOLD_FLAGS_FULL);
15151 redo_e += added_len;
15152 redo_p += UTF8SKIP(redo_p);
15153 }
15154 else {
15155
15156 /* Note that if this code is run on some ancient
15157 * Unicode versions, SHARP S doesn't fold to 'ss',
15158 * but rather than clutter the code with #ifdef's,
15159 * as is done above, we ignore that possibility.
15160 * This is ok because this code doesn't affect what
15161 * gets matched, but merely where the node gets
15162 * split */
15163 if (UCHARAT(redo_p) != LATIN_SMALL_LETTER_SHARP_S) {
15164 *redo_e++ = toLOWER_L1(UCHARAT(redo_p));
15165 }
15166 else {
15167 *redo_e++ = 's';
15168 *redo_e++ = 's';
15169 }
15170 redo_p++;
15171 }
15172
15173
15174 /* If we're getting so close to the end that a
15175 * worst-case fold in the next character would cause us
15176 * to overflow, increase, assuming one byte output byte
15177 * per one byte input one, plus room for another worst
15178 * case fold */
15179 if ( redo_p <= oldp
15180 && redo_e > locfold_buf + size
15181 - (UTF8_MAXBYTES_CASE + 1))
15182 {
15183 Size_t new_size = size
15184 + (oldp - redo_p)
15185 + UTF8_MAXBYTES_CASE + 1;
15186 Ptrdiff_t e_offset = redo_e - locfold_buf;
15187
15188 Renew(locfold_buf, new_size, char);
15189 Renew(loc_correspondence, new_size, Size_t);
15190 size = new_size;
15191
15192 redo_e = locfold_buf + e_offset;
15193 }
15194 }
15195
15196 /* Set so that things are in terms of the folded, temporary
15197 * string */
15198 s = old_redo_e;
15199 s_start = locfold_buf;
15200 e = redo_e;
15201
15202 }
15203
15204 /* Here, we have 's', 's_start' and 'e' set up to point to the
15205 * input that goes into the node, folded.
15206 *
15207 * If the final character of the node and the fold of ender
15208 * form the first two characters of a three character fold, we
15209 * need to peek ahead at the next (unparsed) character in the
15210 * input to determine if the three actually do form such a
15211 * fold. Just looking at that character is not generally
15212 * sufficient, as it could be, for example, an escape sequence
15213 * that evaluates to something else, and it needs to be folded.
15214 *
15215 * khw originally thought to just go through the parse loop one
15216 * extra time, but that doesn't work easily as that iteration
15217 * could cause things to think that the parse is over and to
15218 * goto loopdone. The character could be a '$' for example, or
15219 * the character beyond could be a quantifier, and other
15220 * glitches as well.
15221 *
15222 * The solution used here for peeking ahead is to look at that
15223 * next character. If it isn't ASCII punctuation, then it will
15224 * be something that would continue on in an EXACTish node if
15225 * there were space. We append the fold of it to s, having
15226 * reserved enough room in s0 for the purpose. If we can't
15227 * reasonably peek ahead, we instead assume the worst case:
15228 * that it is something that would form the completion of a
15229 * multi-char fold.
15230 *
15231 * If we can't split between s and ender, we work backwards
15232 * character-by-character down to s0. At each current point
15233 * see if we are at the beginning of a multi-char fold. If so,
15234 * that means we would be splitting the fold across nodes, and
15235 * so we back up one and try again.
15236 *
15237 * If we're not at the beginning, we still could be at the
15238 * final two characters of a (rare) three character fold. We
15239 * check if the sequence starting at the character before the
15240 * current position (and including the current and next
15241 * characters) is a three character fold. If not, the node can
15242 * be split here. If it is, we have to backup two characters
15243 * and try again.
15244 *
15245 * Otherwise, the node can be split at the current position.
15246 *
15247 * The same logic is used for UTF-8 patterns and not */
15248 if (UTF) {
15249 Size_t added_len;
15250
15251 /* Append the fold of ender */
15252 (void) _to_uni_fold_flags(
15253 ender,
15254 (U8 *) e,
15255 &added_len,
15256 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
15257 ? FOLD_FLAGS_NOMIX_ASCII
15258 : 0));
15259 e += added_len;
15260
15261 /* 's' and the character folded to by ender may be the
15262 * first two of a three-character fold, in which case the
15263 * node should not be split here. That may mean examining
15264 * the so-far unparsed character starting at 'p'. But if
15265 * ender folded to more than one character, we already have
15266 * three characters to look at. Also, we first check if
15267 * the sequence consisting of s and the next character form
15268 * the first two of some three character fold. If not,
15269 * there's no need to peek ahead. */
15270 if ( added_len <= UTF8SKIP(e - added_len)
15271 && UNLIKELY(is_THREE_CHAR_FOLD_HEAD_utf8_safe(s, e)))
15272 {
15273 /* Here, the two do form the beginning of a potential
15274 * three character fold. The unexamined character may
15275 * or may not complete it. Peek at it. It might be
15276 * something that ends the node or an escape sequence,
15277 * in which case we don't know without a lot of work
15278 * what it evaluates to, so we have to assume the worst
15279 * case: that it does complete the fold, and so we
15280 * can't split here. All such instances will have
15281 * that character be an ASCII punctuation character,
15282 * like a backslash. So, for that case, backup one and
15283 * drop down to try at that position */
15284 if (isPUNCT(*p)) {
15285 s = (char *) utf8_hop_back((U8 *) s, -1,
15286 (U8 *) s_start);
15287 backed_up = TRUE;
15288 }
15289 else {
15290 /* Here, since it's not punctuation, it must be a
15291 * real character, and we can append its fold to
15292 * 'e' (having deliberately reserved enough space
15293 * for this eventuality) and drop down to check if
15294 * the three actually do form a folded sequence */
15295 (void) _to_utf8_fold_flags(
15296 (U8 *) p, (U8 *) RExC_end,
15297 (U8 *) e,
15298 &added_len,
15299 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
15300 ? FOLD_FLAGS_NOMIX_ASCII
15301 : 0));
15302 e += added_len;
15303 }
15304 }
15305
15306 /* Here, we either have three characters available in
15307 * sequence starting at 's', or we have two characters and
15308 * know that the following one can't possibly be part of a
15309 * three character fold. We go through the node backwards
15310 * until we find a place where we can split it without
15311 * breaking apart a multi-character fold. At any given
15312 * point we have to worry about if such a fold begins at
15313 * the current 's', and also if a three-character fold
15314 * begins at s-1, (containing s and s+1). Splitting in
15315 * either case would break apart a fold */
15316 do {
15317 char *prev_s = (char *) utf8_hop_back((U8 *) s, -1,
15318 (U8 *) s_start);
15319
15320 /* If is a multi-char fold, can't split here. Backup
15321 * one char and try again */
15322 if (UNLIKELY(is_MULTI_CHAR_FOLD_utf8_safe(s, e))) {
15323 s = prev_s;
15324 backed_up = TRUE;
15325 continue;
15326 }
15327
15328 /* If the two characters beginning at 's' are part of a
15329 * three character fold starting at the character
15330 * before s, we can't split either before or after s.
15331 * Backup two chars and try again */
15332 if ( LIKELY(s > s_start)
15333 && UNLIKELY(is_THREE_CHAR_FOLD_utf8_safe(prev_s, e)))
15334 {
15335 s = prev_s;
15336 s = (char *) utf8_hop_back((U8 *) s, -1, (U8 *) s_start);
15337 backed_up = TRUE;
15338 continue;
15339 }
15340
15341 /* Here there's no multi-char fold between s and the
15342 * next character following it. We can split */
15343 splittable = TRUE;
15344 break;
15345
15346 } while (s > s_start); /* End of loops backing up through the node */
15347
15348 /* Here we either couldn't find a place to split the node,
15349 * or else we broke out of the loop setting 'splittable' to
15350 * true. In the latter case, the place to split is between
15351 * the first and second characters in the sequence starting
15352 * at 's' */
15353 if (splittable) {
15354 s += UTF8SKIP(s);
15355 }
15356 }
15357 else { /* Pattern not UTF-8 */
15358 if ( ender != LATIN_SMALL_LETTER_SHARP_S
15359 || ASCII_FOLD_RESTRICTED)
15360 {
15361 assert( toLOWER_L1(ender) < 256 );
15362 *e++ = (char)(toLOWER_L1(ender)); /* should e and the cast be U8? */
15363 }
15364 else {
15365 *e++ = 's';
15366 *e++ = 's';
15367 }
15368
15369 if ( e - s <= 1
15370 && UNLIKELY(is_THREE_CHAR_FOLD_HEAD_latin1_safe(s, e)))
15371 {
15372 if (isPUNCT(*p)) {
15373 s--;
15374 backed_up = TRUE;
15375 }
15376 else {
15377 if ( UCHARAT(p) != LATIN_SMALL_LETTER_SHARP_S
15378 || ASCII_FOLD_RESTRICTED)
15379 {
15380 assert( toLOWER_L1(ender) < 256 );
15381 *e++ = (char)(toLOWER_L1(ender)); /* should e and the cast be U8? */
15382 }
15383 else {
15384 *e++ = 's';
15385 *e++ = 's';
15386 }
15387 }
15388 }
15389
15390 do {
15391 if (UNLIKELY(is_MULTI_CHAR_FOLD_latin1_safe(s, e))) {
15392 s--;
15393 backed_up = TRUE;
15394 continue;
15395 }
15396
15397 if ( LIKELY(s > s_start)
15398 && UNLIKELY(is_THREE_CHAR_FOLD_latin1_safe(s - 1, e)))
15399 {
15400 s -= 2;
15401 backed_up = TRUE;
15402 continue;
15403 }
15404
15405 splittable = TRUE;
15406 break;
15407
15408 } while (s > s_start);
15409
15410 if (splittable) {
15411 s++;
15412 }
15413 }
15414
15415 /* Here, we are done backing up. If we didn't backup at all
15416 * (the likely case), just proceed */
15417 if (backed_up) {
15418
15419 /* If we did find a place to split, reparse the entire node
15420 * stopping where we have calculated. */
15421 if (splittable) {
15422
15423 /* If we created a temporary folded string under /l, we
15424 * have to map that back to the original */
15425 if (need_to_fold_loc) {
15426 upper_fill = loc_correspondence[s - s_start];
15427 if (upper_fill == 0) {
15428 FAIL2("panic: loc_correspondence[%d] is 0",
15429 (int) (s - s_start));
15430 }
15431 Safefree(locfold_buf);
15432 Safefree(loc_correspondence);
15433 }
15434 else {
15435 upper_fill = s - s0;
15436 }
15437 goto reparse;
15438 }
15439
15440 /* Here the node consists entirely of non-final multi-char
15441 * folds. (Likely it is all 'f's or all 's's.) There's no
15442 * decent place to split it, so give up and just take the
15443 * whole thing */
15444 len = old_s - s0;
15445 }
15446
15447 if (need_to_fold_loc) {
15448 Safefree(locfold_buf);
15449 Safefree(loc_correspondence);
15450 }
15451 } /* End of verifying node ends with an appropriate char */
15452
15453 /* We need to start the next node at the character that didn't fit
15454 * in this one */
15455 p = oldp;
15456
15457 loopdone: /* Jumped to when encounters something that shouldn't be
15458 in the node */
15459
15460 /* Free up any over-allocated space; cast is to silence bogus
15461 * warning in MS VC */
15462 change_engine_size(pRExC_state,
15463 - (Ptrdiff_t) (current_string_nodes - STR_SZ(len)));
15464
15465 /* I (khw) don't know if you can get here with zero length, but the
15466 * old code handled this situation by creating a zero-length EXACT
15467 * node. Might as well be NOTHING instead */
15468 if (len == 0) {
15469 OP(REGNODE_p(ret)) = NOTHING;
15470 }
15471 else {
15472
15473 /* If the node type is EXACT here, check to see if it
15474 * should be EXACTL, or EXACT_REQ8. */
15475 if (node_type == EXACT) {
15476 if (LOC) {
15477 node_type = EXACTL;
15478 }
15479 else if (requires_utf8_target) {
15480 node_type = EXACT_REQ8;
15481 }
15482 }
15483 else if (node_type == LEXACT) {
15484 if (requires_utf8_target) {
15485 node_type = LEXACT_REQ8;
15486 }
15487 }
15488 else if (FOLD) {
15489 if ( UNLIKELY(has_micro_sign || has_ss)
15490 && (node_type == EXACTFU || ( node_type == EXACTF
15491 && maybe_exactfu)))
15492 { /* These two conditions are problematic in non-UTF-8
15493 EXACTFU nodes. */
15494 assert(! UTF);
15495 node_type = EXACTFUP;
15496 }
15497 else if (node_type == EXACTFL) {
15498
15499 /* 'maybe_exactfu' is deliberately set above to
15500 * indicate this node type, where all code points in it
15501 * are above 255 */
15502 if (maybe_exactfu) {
15503 node_type = EXACTFLU8;
15504 }
15505 else if (UNLIKELY(
15506 _invlist_contains_cp(PL_HasMultiCharFold, ender)))
15507 {
15508 /* A character that folds to more than one will
15509 * match multiple characters, so can't be SIMPLE.
15510 * We don't have to worry about this with EXACTFLU8
15511 * nodes just above, as they have already been
15512 * folded (since the fold doesn't vary at run
15513 * time). Here, if the final character in the node
15514 * folds to multiple, it can't be simple. (This
15515 * only has an effect if the node has only a single
15516 * character, hence the final one, as elsewhere we
15517 * turn off simple for nodes whose length > 1 */
15518 maybe_SIMPLE = 0;
15519 }
15520 }
15521 else if (node_type == EXACTF) { /* Means is /di */
15522
15523 /* This intermediate variable is needed solely because
15524 * the asserts in the macro where used exceed Win32's
15525 * literal string capacity */
15526 char first_char = * STRING(REGNODE_p(ret));
15527
15528 /* If 'maybe_exactfu' is clear, then we need to stay
15529 * /di. If it is set, it means there are no code
15530 * points that match differently depending on UTF8ness
15531 * of the target string, so it can become an EXACTFU
15532 * node */
15533 if (! maybe_exactfu) {
15534 RExC_seen_d_op = TRUE;
15535 }
15536 else if ( isALPHA_FOLD_EQ(first_char, 's')
15537 || isALPHA_FOLD_EQ(ender, 's'))
15538 {
15539 /* But, if the node begins or ends in an 's' we
15540 * have to defer changing it into an EXACTFU, as
15541 * the node could later get joined with another one
15542 * that ends or begins with 's' creating an 'ss'
15543 * sequence which would then wrongly match the
15544 * sharp s without the target being UTF-8. We
15545 * create a special node that we resolve later when
15546 * we join nodes together */
15547
15548 node_type = EXACTFU_S_EDGE;
15549 }
15550 else {
15551 node_type = EXACTFU;
15552 }
15553 }
15554
15555 if (requires_utf8_target && node_type == EXACTFU) {
15556 node_type = EXACTFU_REQ8;
15557 }
15558 }
15559
15560 OP(REGNODE_p(ret)) = node_type;
15561 setSTR_LEN(REGNODE_p(ret), len);
15562 RExC_emit += STR_SZ(len);
15563
15564 /* If the node isn't a single character, it can't be SIMPLE */
15565 if (len > (Size_t) ((UTF) ? UTF8SKIP(STRING(REGNODE_p(ret))) : 1)) {
15566 maybe_SIMPLE = 0;
15567 }
15568
15569 *flagp |= HASWIDTH | maybe_SIMPLE;
15570 }
15571
15572 Set_Node_Length(REGNODE_p(ret), p - parse_start - 1);
15573 RExC_parse = p;
15574
15575 {
15576 /* len is STRLEN which is unsigned, need to copy to signed */
15577 IV iv = len;
15578 if (iv < 0)
15579 vFAIL("Internal disaster");
15580 }
15581
15582 } /* End of label 'defchar:' */
15583 break;
15584 } /* End of giant switch on input character */
15585
15586 /* Position parse to next real character */
15587 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
15588 FALSE /* Don't force to /x */ );
15589 if ( *RExC_parse == '{'
15590 && OP(REGNODE_p(ret)) != SBOL && ! regcurly(RExC_parse, RExC_end, NULL))
15591 {
15592 if (RExC_strict) {
15593 RExC_parse++;
15594 vFAIL("Unescaped left brace in regex is illegal here");
15595 }
15596 ckWARNreg(RExC_parse + 1, "Unescaped left brace in regex is"
15597 " passed through");
15598 }
15599
15600 return(ret);
15601 }
15602
15603
15604 STATIC void
S_populate_ANYOF_from_invlist(pTHX_ regnode * node,SV ** invlist_ptr)15605 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
15606 {
15607 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
15608 * sets up the bitmap and any flags, removing those code points from the
15609 * inversion list, setting it to NULL should it become completely empty */
15610
15611
15612 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
15613 assert(PL_regkind[OP(node)] == ANYOF);
15614
15615 /* There is no bitmap for this node type */
15616 if (inRANGE(OP(node), ANYOFH, ANYOFRb)) {
15617 return;
15618 }
15619
15620 ANYOF_BITMAP_ZERO(node);
15621 if (*invlist_ptr) {
15622
15623 /* This gets set if we actually need to modify things */
15624 bool change_invlist = FALSE;
15625
15626 UV start, end;
15627
15628 /* Start looking through *invlist_ptr */
15629 invlist_iterinit(*invlist_ptr);
15630 while (invlist_iternext(*invlist_ptr, &start, &end)) {
15631 UV high;
15632 int i;
15633
15634 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
15635 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
15636 }
15637
15638 /* Quit if are above what we should change */
15639 if (start >= NUM_ANYOF_CODE_POINTS) {
15640 break;
15641 }
15642
15643 change_invlist = TRUE;
15644
15645 /* Set all the bits in the range, up to the max that we are doing */
15646 high = (end < NUM_ANYOF_CODE_POINTS - 1)
15647 ? end
15648 : NUM_ANYOF_CODE_POINTS - 1;
15649 for (i = start; i <= (int) high; i++) {
15650 ANYOF_BITMAP_SET(node, i);
15651 }
15652 }
15653 invlist_iterfinish(*invlist_ptr);
15654
15655 /* Done with loop; remove any code points that are in the bitmap from
15656 * *invlist_ptr; similarly for code points above the bitmap if we have
15657 * a flag to match all of them anyways */
15658 if (change_invlist) {
15659 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
15660 }
15661 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
15662 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
15663 }
15664
15665 /* If have completely emptied it, remove it completely */
15666 if (_invlist_len(*invlist_ptr) == 0) {
15667 SvREFCNT_dec_NN(*invlist_ptr);
15668 *invlist_ptr = NULL;
15669 }
15670 }
15671 }
15672
15673 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
15674 Character classes ([:foo:]) can also be negated ([:^foo:]).
15675 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
15676 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
15677 but trigger failures because they are currently unimplemented. */
15678
15679 #define POSIXCC_DONE(c) ((c) == ':')
15680 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
15681 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
15682 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
15683
15684 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
15685 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
15686 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
15687
15688 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
15689
15690 /* 'posix_warnings' and 'warn_text' are names of variables in the following
15691 * routine. q.v. */
15692 #define ADD_POSIX_WARNING(p, text) STMT_START { \
15693 if (posix_warnings) { \
15694 if (! RExC_warn_text ) RExC_warn_text = \
15695 (AV *) sv_2mortal((SV *) newAV()); \
15696 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
15697 WARNING_PREFIX \
15698 text \
15699 REPORT_LOCATION, \
15700 REPORT_LOCATION_ARGS(p))); \
15701 } \
15702 } STMT_END
15703 #define CLEAR_POSIX_WARNINGS() \
15704 STMT_START { \
15705 if (posix_warnings && RExC_warn_text) \
15706 av_clear(RExC_warn_text); \
15707 } STMT_END
15708
15709 #define CLEAR_POSIX_WARNINGS_AND_RETURN(ret) \
15710 STMT_START { \
15711 CLEAR_POSIX_WARNINGS(); \
15712 return ret; \
15713 } STMT_END
15714
15715 STATIC int
S_handle_possible_posix(pTHX_ RExC_state_t * pRExC_state,const char * const s,char ** updated_parse_ptr,AV ** posix_warnings,const bool check_only)15716 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
15717
15718 const char * const s, /* Where the putative posix class begins.
15719 Normally, this is one past the '['. This
15720 parameter exists so it can be somewhere
15721 besides RExC_parse. */
15722 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
15723 NULL */
15724 AV ** posix_warnings, /* Where to place any generated warnings, or
15725 NULL */
15726 const bool check_only /* Don't die if error */
15727 )
15728 {
15729 /* This parses what the caller thinks may be one of the three POSIX
15730 * constructs:
15731 * 1) a character class, like [:blank:]
15732 * 2) a collating symbol, like [. .]
15733 * 3) an equivalence class, like [= =]
15734 * In the latter two cases, it croaks if it finds a syntactically legal
15735 * one, as these are not handled by Perl.
15736 *
15737 * The main purpose is to look for a POSIX character class. It returns:
15738 * a) the class number
15739 * if it is a completely syntactically and semantically legal class.
15740 * 'updated_parse_ptr', if not NULL, is set to point to just after the
15741 * closing ']' of the class
15742 * b) OOB_NAMEDCLASS
15743 * if it appears that one of the three POSIX constructs was meant, but
15744 * its specification was somehow defective. 'updated_parse_ptr', if
15745 * not NULL, is set to point to the character just after the end
15746 * character of the class. See below for handling of warnings.
15747 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
15748 * if it doesn't appear that a POSIX construct was intended.
15749 * 'updated_parse_ptr' is not changed. No warnings nor errors are
15750 * raised.
15751 *
15752 * In b) there may be errors or warnings generated. If 'check_only' is
15753 * TRUE, then any errors are discarded. Warnings are returned to the
15754 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
15755 * instead it is NULL, warnings are suppressed.
15756 *
15757 * The reason for this function, and its complexity is that a bracketed
15758 * character class can contain just about anything. But it's easy to
15759 * mistype the very specific posix class syntax but yielding a valid
15760 * regular bracketed class, so it silently gets compiled into something
15761 * quite unintended.
15762 *
15763 * The solution adopted here maintains backward compatibility except that
15764 * it adds a warning if it looks like a posix class was intended but
15765 * improperly specified. The warning is not raised unless what is input
15766 * very closely resembles one of the 14 legal posix classes. To do this,
15767 * it uses fuzzy parsing. It calculates how many single-character edits it
15768 * would take to transform what was input into a legal posix class. Only
15769 * if that number is quite small does it think that the intention was a
15770 * posix class. Obviously these are heuristics, and there will be cases
15771 * where it errs on one side or another, and they can be tweaked as
15772 * experience informs.
15773 *
15774 * The syntax for a legal posix class is:
15775 *
15776 * qr/(?xa: \[ : \^? [[:lower:]]{4,6} : \] )/
15777 *
15778 * What this routine considers syntactically to be an intended posix class
15779 * is this (the comments indicate some restrictions that the pattern
15780 * doesn't show):
15781 *
15782 * qr/(?x: \[? # The left bracket, possibly
15783 * # omitted
15784 * \h* # possibly followed by blanks
15785 * (?: \^ \h* )? # possibly a misplaced caret
15786 * [:;]? # The opening class character,
15787 * # possibly omitted. A typo
15788 * # semi-colon can also be used.
15789 * \h*
15790 * \^? # possibly a correctly placed
15791 * # caret, but not if there was also
15792 * # a misplaced one
15793 * \h*
15794 * .{3,15} # The class name. If there are
15795 * # deviations from the legal syntax,
15796 * # its edit distance must be close
15797 * # to a real class name in order
15798 * # for it to be considered to be
15799 * # an intended posix class.
15800 * \h*
15801 * [[:punct:]]? # The closing class character,
15802 * # possibly omitted. If not a colon
15803 * # nor semi colon, the class name
15804 * # must be even closer to a valid
15805 * # one
15806 * \h*
15807 * \]? # The right bracket, possibly
15808 * # omitted.
15809 * )/
15810 *
15811 * In the above, \h must be ASCII-only.
15812 *
15813 * These are heuristics, and can be tweaked as field experience dictates.
15814 * There will be cases when someone didn't intend to specify a posix class
15815 * that this warns as being so. The goal is to minimize these, while
15816 * maximizing the catching of things intended to be a posix class that
15817 * aren't parsed as such.
15818 */
15819
15820 const char* p = s;
15821 const char * const e = RExC_end;
15822 unsigned complement = 0; /* If to complement the class */
15823 bool found_problem = FALSE; /* Assume OK until proven otherwise */
15824 bool has_opening_bracket = FALSE;
15825 bool has_opening_colon = FALSE;
15826 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
15827 valid class */
15828 const char * possible_end = NULL; /* used for a 2nd parse pass */
15829 const char* name_start; /* ptr to class name first char */
15830
15831 /* If the number of single-character typos the input name is away from a
15832 * legal name is no more than this number, it is considered to have meant
15833 * the legal name */
15834 int max_distance = 2;
15835
15836 /* to store the name. The size determines the maximum length before we
15837 * decide that no posix class was intended. Should be at least
15838 * sizeof("alphanumeric") */
15839 UV input_text[15];
15840 STATIC_ASSERT_DECL(C_ARRAY_LENGTH(input_text) >= sizeof "alphanumeric");
15841
15842 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
15843
15844 CLEAR_POSIX_WARNINGS();
15845
15846 if (p >= e) {
15847 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
15848 }
15849
15850 if (*(p - 1) != '[') {
15851 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
15852 found_problem = TRUE;
15853 }
15854 else {
15855 has_opening_bracket = TRUE;
15856 }
15857
15858 /* They could be confused and think you can put spaces between the
15859 * components */
15860 if (isBLANK(*p)) {
15861 found_problem = TRUE;
15862
15863 do {
15864 p++;
15865 } while (p < e && isBLANK(*p));
15866
15867 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15868 }
15869
15870 /* For [. .] and [= =]. These are quite different internally from [: :],
15871 * so they are handled separately. */
15872 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
15873 and 1 for at least one char in it
15874 */
15875 {
15876 const char open_char = *p;
15877 const char * temp_ptr = p + 1;
15878
15879 /* These two constructs are not handled by perl, and if we find a
15880 * syntactically valid one, we croak. khw, who wrote this code, finds
15881 * this explanation of them very unclear:
15882 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
15883 * And searching the rest of the internet wasn't very helpful either.
15884 * It looks like just about any byte can be in these constructs,
15885 * depending on the locale. But unless the pattern is being compiled
15886 * under /l, which is very rare, Perl runs under the C or POSIX locale.
15887 * In that case, it looks like [= =] isn't allowed at all, and that
15888 * [. .] could be any single code point, but for longer strings the
15889 * constituent characters would have to be the ASCII alphabetics plus
15890 * the minus-hyphen. Any sensible locale definition would limit itself
15891 * to these. And any portable one definitely should. Trying to parse
15892 * the general case is a nightmare (see [perl #127604]). So, this code
15893 * looks only for interiors of these constructs that match:
15894 * qr/.|[-\w]{2,}/
15895 * Using \w relaxes the apparent rules a little, without adding much
15896 * danger of mistaking something else for one of these constructs.
15897 *
15898 * [. .] in some implementations described on the internet is usable to
15899 * escape a character that otherwise is special in bracketed character
15900 * classes. For example [.].] means a literal right bracket instead of
15901 * the ending of the class
15902 *
15903 * [= =] can legitimately contain a [. .] construct, but we don't
15904 * handle this case, as that [. .] construct will later get parsed
15905 * itself and croak then. And [= =] is checked for even when not under
15906 * /l, as Perl has long done so.
15907 *
15908 * The code below relies on there being a trailing NUL, so it doesn't
15909 * have to keep checking if the parse ptr < e.
15910 */
15911 if (temp_ptr[1] == open_char) {
15912 temp_ptr++;
15913 }
15914 else while ( temp_ptr < e
15915 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
15916 {
15917 temp_ptr++;
15918 }
15919
15920 if (*temp_ptr == open_char) {
15921 temp_ptr++;
15922 if (*temp_ptr == ']') {
15923 temp_ptr++;
15924 if (! found_problem && ! check_only) {
15925 RExC_parse = (char *) temp_ptr;
15926 vFAIL3("POSIX syntax [%c %c] is reserved for future "
15927 "extensions", open_char, open_char);
15928 }
15929
15930 /* Here, the syntax wasn't completely valid, or else the call
15931 * is to check-only */
15932 if (updated_parse_ptr) {
15933 *updated_parse_ptr = (char *) temp_ptr;
15934 }
15935
15936 CLEAR_POSIX_WARNINGS_AND_RETURN(OOB_NAMEDCLASS);
15937 }
15938 }
15939
15940 /* If we find something that started out to look like one of these
15941 * constructs, but isn't, we continue below so that it can be checked
15942 * for being a class name with a typo of '.' or '=' instead of a colon.
15943 * */
15944 }
15945
15946 /* Here, we think there is a possibility that a [: :] class was meant, and
15947 * we have the first real character. It could be they think the '^' comes
15948 * first */
15949 if (*p == '^') {
15950 found_problem = TRUE;
15951 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
15952 complement = 1;
15953 p++;
15954
15955 if (isBLANK(*p)) {
15956 found_problem = TRUE;
15957
15958 do {
15959 p++;
15960 } while (p < e && isBLANK(*p));
15961
15962 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15963 }
15964 }
15965
15966 /* But the first character should be a colon, which they could have easily
15967 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
15968 * distinguish from a colon, so treat that as a colon). */
15969 if (*p == ':') {
15970 p++;
15971 has_opening_colon = TRUE;
15972 }
15973 else if (*p == ';') {
15974 found_problem = TRUE;
15975 p++;
15976 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15977 has_opening_colon = TRUE;
15978 }
15979 else {
15980 found_problem = TRUE;
15981 ADD_POSIX_WARNING(p, "there must be a starting ':'");
15982
15983 /* Consider an initial punctuation (not one of the recognized ones) to
15984 * be a left terminator */
15985 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
15986 p++;
15987 }
15988 }
15989
15990 /* They may think that you can put spaces between the components */
15991 if (isBLANK(*p)) {
15992 found_problem = TRUE;
15993
15994 do {
15995 p++;
15996 } while (p < e && isBLANK(*p));
15997
15998 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15999 }
16000
16001 if (*p == '^') {
16002
16003 /* We consider something like [^:^alnum:]] to not have been intended to
16004 * be a posix class, but XXX maybe we should */
16005 if (complement) {
16006 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16007 }
16008
16009 complement = 1;
16010 p++;
16011 }
16012
16013 /* Again, they may think that you can put spaces between the components */
16014 if (isBLANK(*p)) {
16015 found_problem = TRUE;
16016
16017 do {
16018 p++;
16019 } while (p < e && isBLANK(*p));
16020
16021 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
16022 }
16023
16024 if (*p == ']') {
16025
16026 /* XXX This ']' may be a typo, and something else was meant. But
16027 * treating it as such creates enough complications, that that
16028 * possibility isn't currently considered here. So we assume that the
16029 * ']' is what is intended, and if we've already found an initial '[',
16030 * this leaves this construct looking like [:] or [:^], which almost
16031 * certainly weren't intended to be posix classes */
16032 if (has_opening_bracket) {
16033 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16034 }
16035
16036 /* But this function can be called when we parse the colon for
16037 * something like qr/[alpha:]]/, so we back up to look for the
16038 * beginning */
16039 p--;
16040
16041 if (*p == ';') {
16042 found_problem = TRUE;
16043 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
16044 }
16045 else if (*p != ':') {
16046
16047 /* XXX We are currently very restrictive here, so this code doesn't
16048 * consider the possibility that, say, /[alpha.]]/ was intended to
16049 * be a posix class. */
16050 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16051 }
16052
16053 /* Here we have something like 'foo:]'. There was no initial colon,
16054 * and we back up over 'foo. XXX Unlike the going forward case, we
16055 * don't handle typos of non-word chars in the middle */
16056 has_opening_colon = FALSE;
16057 p--;
16058
16059 while (p > RExC_start && isWORDCHAR(*p)) {
16060 p--;
16061 }
16062 p++;
16063
16064 /* Here, we have positioned ourselves to where we think the first
16065 * character in the potential class is */
16066 }
16067
16068 /* Now the interior really starts. There are certain key characters that
16069 * can end the interior, or these could just be typos. To catch both
16070 * cases, we may have to do two passes. In the first pass, we keep on
16071 * going unless we come to a sequence that matches
16072 * qr/ [[:punct:]] [[:blank:]]* \] /xa
16073 * This means it takes a sequence to end the pass, so two typos in a row if
16074 * that wasn't what was intended. If the class is perfectly formed, just
16075 * this one pass is needed. We also stop if there are too many characters
16076 * being accumulated, but this number is deliberately set higher than any
16077 * real class. It is set high enough so that someone who thinks that
16078 * 'alphanumeric' is a correct name would get warned that it wasn't.
16079 * While doing the pass, we keep track of where the key characters were in
16080 * it. If we don't find an end to the class, and one of the key characters
16081 * was found, we redo the pass, but stop when we get to that character.
16082 * Thus the key character was considered a typo in the first pass, but a
16083 * terminator in the second. If two key characters are found, we stop at
16084 * the second one in the first pass. Again this can miss two typos, but
16085 * catches a single one
16086 *
16087 * In the first pass, 'possible_end' starts as NULL, and then gets set to
16088 * point to the first key character. For the second pass, it starts as -1.
16089 * */
16090
16091 name_start = p;
16092 parse_name:
16093 {
16094 bool has_blank = FALSE;
16095 bool has_upper = FALSE;
16096 bool has_terminating_colon = FALSE;
16097 bool has_terminating_bracket = FALSE;
16098 bool has_semi_colon = FALSE;
16099 unsigned int name_len = 0;
16100 int punct_count = 0;
16101
16102 while (p < e) {
16103
16104 /* Squeeze out blanks when looking up the class name below */
16105 if (isBLANK(*p) ) {
16106 has_blank = TRUE;
16107 found_problem = TRUE;
16108 p++;
16109 continue;
16110 }
16111
16112 /* The name will end with a punctuation */
16113 if (isPUNCT(*p)) {
16114 const char * peek = p + 1;
16115
16116 /* Treat any non-']' punctuation followed by a ']' (possibly
16117 * with intervening blanks) as trying to terminate the class.
16118 * ']]' is very likely to mean a class was intended (but
16119 * missing the colon), but the warning message that gets
16120 * generated shows the error position better if we exit the
16121 * loop at the bottom (eventually), so skip it here. */
16122 if (*p != ']') {
16123 if (peek < e && isBLANK(*peek)) {
16124 has_blank = TRUE;
16125 found_problem = TRUE;
16126 do {
16127 peek++;
16128 } while (peek < e && isBLANK(*peek));
16129 }
16130
16131 if (peek < e && *peek == ']') {
16132 has_terminating_bracket = TRUE;
16133 if (*p == ':') {
16134 has_terminating_colon = TRUE;
16135 }
16136 else if (*p == ';') {
16137 has_semi_colon = TRUE;
16138 has_terminating_colon = TRUE;
16139 }
16140 else {
16141 found_problem = TRUE;
16142 }
16143 p = peek + 1;
16144 goto try_posix;
16145 }
16146 }
16147
16148 /* Here we have punctuation we thought didn't end the class.
16149 * Keep track of the position of the key characters that are
16150 * more likely to have been class-enders */
16151 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
16152
16153 /* Allow just one such possible class-ender not actually
16154 * ending the class. */
16155 if (possible_end) {
16156 break;
16157 }
16158 possible_end = p;
16159 }
16160
16161 /* If we have too many punctuation characters, no use in
16162 * keeping going */
16163 if (++punct_count > max_distance) {
16164 break;
16165 }
16166
16167 /* Treat the punctuation as a typo. */
16168 input_text[name_len++] = *p;
16169 p++;
16170 }
16171 else if (isUPPER(*p)) { /* Use lowercase for lookup */
16172 input_text[name_len++] = toLOWER(*p);
16173 has_upper = TRUE;
16174 found_problem = TRUE;
16175 p++;
16176 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
16177 input_text[name_len++] = *p;
16178 p++;
16179 }
16180 else {
16181 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
16182 p+= UTF8SKIP(p);
16183 }
16184
16185 /* The declaration of 'input_text' is how long we allow a potential
16186 * class name to be, before saying they didn't mean a class name at
16187 * all */
16188 if (name_len >= C_ARRAY_LENGTH(input_text)) {
16189 break;
16190 }
16191 }
16192
16193 /* We get to here when the possible class name hasn't been properly
16194 * terminated before:
16195 * 1) we ran off the end of the pattern; or
16196 * 2) found two characters, each of which might have been intended to
16197 * be the name's terminator
16198 * 3) found so many punctuation characters in the purported name,
16199 * that the edit distance to a valid one is exceeded
16200 * 4) we decided it was more characters than anyone could have
16201 * intended to be one. */
16202
16203 found_problem = TRUE;
16204
16205 /* In the final two cases, we know that looking up what we've
16206 * accumulated won't lead to a match, even a fuzzy one. */
16207 if ( name_len >= C_ARRAY_LENGTH(input_text)
16208 || punct_count > max_distance)
16209 {
16210 /* If there was an intermediate key character that could have been
16211 * an intended end, redo the parse, but stop there */
16212 if (possible_end && possible_end != (char *) -1) {
16213 possible_end = (char *) -1; /* Special signal value to say
16214 we've done a first pass */
16215 p = name_start;
16216 goto parse_name;
16217 }
16218
16219 /* Otherwise, it can't have meant to have been a class */
16220 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16221 }
16222
16223 /* If we ran off the end, and the final character was a punctuation
16224 * one, back up one, to look at that final one just below. Later, we
16225 * will restore the parse pointer if appropriate */
16226 if (name_len && p == e && isPUNCT(*(p-1))) {
16227 p--;
16228 name_len--;
16229 }
16230
16231 if (p < e && isPUNCT(*p)) {
16232 if (*p == ']') {
16233 has_terminating_bracket = TRUE;
16234
16235 /* If this is a 2nd ']', and the first one is just below this
16236 * one, consider that to be the real terminator. This gives a
16237 * uniform and better positioning for the warning message */
16238 if ( possible_end
16239 && possible_end != (char *) -1
16240 && *possible_end == ']'
16241 && name_len && input_text[name_len - 1] == ']')
16242 {
16243 name_len--;
16244 p = possible_end;
16245
16246 /* And this is actually equivalent to having done the 2nd
16247 * pass now, so set it to not try again */
16248 possible_end = (char *) -1;
16249 }
16250 }
16251 else {
16252 if (*p == ':') {
16253 has_terminating_colon = TRUE;
16254 }
16255 else if (*p == ';') {
16256 has_semi_colon = TRUE;
16257 has_terminating_colon = TRUE;
16258 }
16259 p++;
16260 }
16261 }
16262
16263 try_posix:
16264
16265 /* Here, we have a class name to look up. We can short circuit the
16266 * stuff below for short names that can't possibly be meant to be a
16267 * class name. (We can do this on the first pass, as any second pass
16268 * will yield an even shorter name) */
16269 if (name_len < 3) {
16270 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16271 }
16272
16273 /* Find which class it is. Initially switch on the length of the name.
16274 * */
16275 switch (name_len) {
16276 case 4:
16277 if (memEQs(name_start, 4, "word")) {
16278 /* this is not POSIX, this is the Perl \w */
16279 class_number = ANYOF_WORDCHAR;
16280 }
16281 break;
16282 case 5:
16283 /* Names all of length 5: alnum alpha ascii blank cntrl digit
16284 * graph lower print punct space upper
16285 * Offset 4 gives the best switch position. */
16286 switch (name_start[4]) {
16287 case 'a':
16288 if (memBEGINs(name_start, 5, "alph")) /* alpha */
16289 class_number = ANYOF_ALPHA;
16290 break;
16291 case 'e':
16292 if (memBEGINs(name_start, 5, "spac")) /* space */
16293 class_number = ANYOF_SPACE;
16294 break;
16295 case 'h':
16296 if (memBEGINs(name_start, 5, "grap")) /* graph */
16297 class_number = ANYOF_GRAPH;
16298 break;
16299 case 'i':
16300 if (memBEGINs(name_start, 5, "asci")) /* ascii */
16301 class_number = ANYOF_ASCII;
16302 break;
16303 case 'k':
16304 if (memBEGINs(name_start, 5, "blan")) /* blank */
16305 class_number = ANYOF_BLANK;
16306 break;
16307 case 'l':
16308 if (memBEGINs(name_start, 5, "cntr")) /* cntrl */
16309 class_number = ANYOF_CNTRL;
16310 break;
16311 case 'm':
16312 if (memBEGINs(name_start, 5, "alnu")) /* alnum */
16313 class_number = ANYOF_ALPHANUMERIC;
16314 break;
16315 case 'r':
16316 if (memBEGINs(name_start, 5, "lowe")) /* lower */
16317 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
16318 else if (memBEGINs(name_start, 5, "uppe")) /* upper */
16319 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
16320 break;
16321 case 't':
16322 if (memBEGINs(name_start, 5, "digi")) /* digit */
16323 class_number = ANYOF_DIGIT;
16324 else if (memBEGINs(name_start, 5, "prin")) /* print */
16325 class_number = ANYOF_PRINT;
16326 else if (memBEGINs(name_start, 5, "punc")) /* punct */
16327 class_number = ANYOF_PUNCT;
16328 break;
16329 }
16330 break;
16331 case 6:
16332 if (memEQs(name_start, 6, "xdigit"))
16333 class_number = ANYOF_XDIGIT;
16334 break;
16335 }
16336
16337 /* If the name exactly matches a posix class name the class number will
16338 * here be set to it, and the input almost certainly was meant to be a
16339 * posix class, so we can skip further checking. If instead the syntax
16340 * is exactly correct, but the name isn't one of the legal ones, we
16341 * will return that as an error below. But if neither of these apply,
16342 * it could be that no posix class was intended at all, or that one
16343 * was, but there was a typo. We tease these apart by doing fuzzy
16344 * matching on the name */
16345 if (class_number == OOB_NAMEDCLASS && found_problem) {
16346 const UV posix_names[][6] = {
16347 { 'a', 'l', 'n', 'u', 'm' },
16348 { 'a', 'l', 'p', 'h', 'a' },
16349 { 'a', 's', 'c', 'i', 'i' },
16350 { 'b', 'l', 'a', 'n', 'k' },
16351 { 'c', 'n', 't', 'r', 'l' },
16352 { 'd', 'i', 'g', 'i', 't' },
16353 { 'g', 'r', 'a', 'p', 'h' },
16354 { 'l', 'o', 'w', 'e', 'r' },
16355 { 'p', 'r', 'i', 'n', 't' },
16356 { 'p', 'u', 'n', 'c', 't' },
16357 { 's', 'p', 'a', 'c', 'e' },
16358 { 'u', 'p', 'p', 'e', 'r' },
16359 { 'w', 'o', 'r', 'd' },
16360 { 'x', 'd', 'i', 'g', 'i', 't' }
16361 };
16362 /* The names of the above all have added NULs to make them the same
16363 * size, so we need to also have the real lengths */
16364 const UV posix_name_lengths[] = {
16365 sizeof("alnum") - 1,
16366 sizeof("alpha") - 1,
16367 sizeof("ascii") - 1,
16368 sizeof("blank") - 1,
16369 sizeof("cntrl") - 1,
16370 sizeof("digit") - 1,
16371 sizeof("graph") - 1,
16372 sizeof("lower") - 1,
16373 sizeof("print") - 1,
16374 sizeof("punct") - 1,
16375 sizeof("space") - 1,
16376 sizeof("upper") - 1,
16377 sizeof("word") - 1,
16378 sizeof("xdigit")- 1
16379 };
16380 unsigned int i;
16381 int temp_max = max_distance; /* Use a temporary, so if we
16382 reparse, we haven't changed the
16383 outer one */
16384
16385 /* Use a smaller max edit distance if we are missing one of the
16386 * delimiters */
16387 if ( has_opening_bracket + has_opening_colon < 2
16388 || has_terminating_bracket + has_terminating_colon < 2)
16389 {
16390 temp_max--;
16391 }
16392
16393 /* See if the input name is close to a legal one */
16394 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
16395
16396 /* Short circuit call if the lengths are too far apart to be
16397 * able to match */
16398 if (abs( (int) (name_len - posix_name_lengths[i]))
16399 > temp_max)
16400 {
16401 continue;
16402 }
16403
16404 if (edit_distance(input_text,
16405 posix_names[i],
16406 name_len,
16407 posix_name_lengths[i],
16408 temp_max
16409 )
16410 > -1)
16411 { /* If it is close, it probably was intended to be a class */
16412 goto probably_meant_to_be;
16413 }
16414 }
16415
16416 /* Here the input name is not close enough to a valid class name
16417 * for us to consider it to be intended to be a posix class. If
16418 * we haven't already done so, and the parse found a character that
16419 * could have been terminators for the name, but which we absorbed
16420 * as typos during the first pass, repeat the parse, signalling it
16421 * to stop at that character */
16422 if (possible_end && possible_end != (char *) -1) {
16423 possible_end = (char *) -1;
16424 p = name_start;
16425 goto parse_name;
16426 }
16427
16428 /* Here neither pass found a close-enough class name */
16429 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
16430 }
16431
16432 probably_meant_to_be:
16433
16434 /* Here we think that a posix specification was intended. Update any
16435 * parse pointer */
16436 if (updated_parse_ptr) {
16437 *updated_parse_ptr = (char *) p;
16438 }
16439
16440 /* If a posix class name was intended but incorrectly specified, we
16441 * output or return the warnings */
16442 if (found_problem) {
16443
16444 /* We set flags for these issues in the parse loop above instead of
16445 * adding them to the list of warnings, because we can parse it
16446 * twice, and we only want one warning instance */
16447 if (has_upper) {
16448 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
16449 }
16450 if (has_blank) {
16451 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
16452 }
16453 if (has_semi_colon) {
16454 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
16455 }
16456 else if (! has_terminating_colon) {
16457 ADD_POSIX_WARNING(p, "there is no terminating ':'");
16458 }
16459 if (! has_terminating_bracket) {
16460 ADD_POSIX_WARNING(p, "there is no terminating ']'");
16461 }
16462
16463 if ( posix_warnings
16464 && RExC_warn_text
16465 && av_count(RExC_warn_text) > 0)
16466 {
16467 *posix_warnings = RExC_warn_text;
16468 }
16469 }
16470 else if (class_number != OOB_NAMEDCLASS) {
16471 /* If it is a known class, return the class. The class number
16472 * #defines are structured so each complement is +1 to the normal
16473 * one */
16474 CLEAR_POSIX_WARNINGS_AND_RETURN(class_number + complement);
16475 }
16476 else if (! check_only) {
16477
16478 /* Here, it is an unrecognized class. This is an error (unless the
16479 * call is to check only, which we've already handled above) */
16480 const char * const complement_string = (complement)
16481 ? "^"
16482 : "";
16483 RExC_parse = (char *) p;
16484 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
16485 complement_string,
16486 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
16487 }
16488 }
16489
16490 return OOB_NAMEDCLASS;
16491 }
16492 #undef ADD_POSIX_WARNING
16493
16494 STATIC unsigned int
S_regex_set_precedence(const U8 my_operator)16495 S_regex_set_precedence(const U8 my_operator) {
16496
16497 /* Returns the precedence in the (?[...]) construct of the input operator,
16498 * specified by its character representation. The precedence follows
16499 * general Perl rules, but it extends this so that ')' and ']' have (low)
16500 * precedence even though they aren't really operators */
16501
16502 switch (my_operator) {
16503 case '!':
16504 return 5;
16505 case '&':
16506 return 4;
16507 case '^':
16508 case '|':
16509 case '+':
16510 case '-':
16511 return 3;
16512 case ')':
16513 return 2;
16514 case ']':
16515 return 1;
16516 }
16517
16518 NOT_REACHED; /* NOTREACHED */
16519 return 0; /* Silence compiler warning */
16520 }
16521
16522 STATIC regnode_offset
S_handle_regex_sets(pTHX_ RExC_state_t * pRExC_state,SV ** return_invlist,I32 * flagp,U32 depth,char * const oregcomp_parse)16523 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
16524 I32 *flagp, U32 depth,
16525 char * const oregcomp_parse)
16526 {
16527 /* Handle the (?[...]) construct to do set operations */
16528
16529 U8 curchar; /* Current character being parsed */
16530 UV start, end; /* End points of code point ranges */
16531 SV* final = NULL; /* The end result inversion list */
16532 SV* result_string; /* 'final' stringified */
16533 AV* stack; /* stack of operators and operands not yet
16534 resolved */
16535 AV* fence_stack = NULL; /* A stack containing the positions in
16536 'stack' of where the undealt-with left
16537 parens would be if they were actually
16538 put there */
16539 /* The 'volatile' is a workaround for an optimiser bug
16540 * in Solaris Studio 12.3. See RT #127455 */
16541 volatile IV fence = 0; /* Position of where most recent undealt-
16542 with left paren in stack is; -1 if none.
16543 */
16544 STRLEN len; /* Temporary */
16545 regnode_offset node; /* Temporary, and final regnode returned by
16546 this function */
16547 const bool save_fold = FOLD; /* Temporary */
16548 char *save_end, *save_parse; /* Temporaries */
16549 const bool in_locale = LOC; /* we turn off /l during processing */
16550
16551 DECLARE_AND_GET_RE_DEBUG_FLAGS;
16552
16553 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
16554 PERL_UNUSED_ARG(oregcomp_parse); /* Only for Set_Node_Length */
16555
16556 DEBUG_PARSE("xcls");
16557
16558 if (in_locale) {
16559 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
16560 }
16561
16562 /* The use of this operator implies /u. This is required so that the
16563 * compile time values are valid in all runtime cases */
16564 REQUIRE_UNI_RULES(flagp, 0);
16565
16566 ckWARNexperimental(RExC_parse,
16567 WARN_EXPERIMENTAL__REGEX_SETS,
16568 "The regex_sets feature is experimental");
16569
16570 /* Everything in this construct is a metacharacter. Operands begin with
16571 * either a '\' (for an escape sequence), or a '[' for a bracketed
16572 * character class. Any other character should be an operator, or
16573 * parenthesis for grouping. Both types of operands are handled by calling
16574 * regclass() to parse them. It is called with a parameter to indicate to
16575 * return the computed inversion list. The parsing here is implemented via
16576 * a stack. Each entry on the stack is a single character representing one
16577 * of the operators; or else a pointer to an operand inversion list. */
16578
16579 #define IS_OPERATOR(a) SvIOK(a)
16580 #define IS_OPERAND(a) (! IS_OPERATOR(a))
16581
16582 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
16583 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
16584 * with pronouncing it called it Reverse Polish instead, but now that YOU
16585 * know how to pronounce it you can use the correct term, thus giving due
16586 * credit to the person who invented it, and impressing your geek friends.
16587 * Wikipedia says that the pronounciation of "Ł" has been changing so that
16588 * it is now more like an English initial W (as in wonk) than an L.)
16589 *
16590 * This means that, for example, 'a | b & c' is stored on the stack as
16591 *
16592 * c [4]
16593 * b [3]
16594 * & [2]
16595 * a [1]
16596 * | [0]
16597 *
16598 * where the numbers in brackets give the stack [array] element number.
16599 * In this implementation, parentheses are not stored on the stack.
16600 * Instead a '(' creates a "fence" so that the part of the stack below the
16601 * fence is invisible except to the corresponding ')' (this allows us to
16602 * replace testing for parens, by using instead subtraction of the fence
16603 * position). As new operands are processed they are pushed onto the stack
16604 * (except as noted in the next paragraph). New operators of higher
16605 * precedence than the current final one are inserted on the stack before
16606 * the lhs operand (so that when the rhs is pushed next, everything will be
16607 * in the correct positions shown above. When an operator of equal or
16608 * lower precedence is encountered in parsing, all the stacked operations
16609 * of equal or higher precedence are evaluated, leaving the result as the
16610 * top entry on the stack. This makes higher precedence operations
16611 * evaluate before lower precedence ones, and causes operations of equal
16612 * precedence to left associate.
16613 *
16614 * The only unary operator '!' is immediately pushed onto the stack when
16615 * encountered. When an operand is encountered, if the top of the stack is
16616 * a '!", the complement is immediately performed, and the '!' popped. The
16617 * resulting value is treated as a new operand, and the logic in the
16618 * previous paragraph is executed. Thus in the expression
16619 * [a] + ! [b]
16620 * the stack looks like
16621 *
16622 * !
16623 * a
16624 * +
16625 *
16626 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
16627 * becomes
16628 *
16629 * !b
16630 * a
16631 * +
16632 *
16633 * A ')' is treated as an operator with lower precedence than all the
16634 * aforementioned ones, which causes all operations on the stack above the
16635 * corresponding '(' to be evaluated down to a single resultant operand.
16636 * Then the fence for the '(' is removed, and the operand goes through the
16637 * algorithm above, without the fence.
16638 *
16639 * A separate stack is kept of the fence positions, so that the position of
16640 * the latest so-far unbalanced '(' is at the top of it.
16641 *
16642 * The ']' ending the construct is treated as the lowest operator of all,
16643 * so that everything gets evaluated down to a single operand, which is the
16644 * result */
16645
16646 sv_2mortal((SV *)(stack = newAV()));
16647 sv_2mortal((SV *)(fence_stack = newAV()));
16648
16649 while (RExC_parse < RExC_end) {
16650 I32 top_index; /* Index of top-most element in 'stack' */
16651 SV** top_ptr; /* Pointer to top 'stack' element */
16652 SV* current = NULL; /* To contain the current inversion list
16653 operand */
16654 SV* only_to_avoid_leaks;
16655
16656 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
16657 TRUE /* Force /x */ );
16658 if (RExC_parse >= RExC_end) { /* Fail */
16659 break;
16660 }
16661
16662 curchar = UCHARAT(RExC_parse);
16663
16664 redo_curchar:
16665
16666 #ifdef ENABLE_REGEX_SETS_DEBUGGING
16667 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
16668 DEBUG_U(dump_regex_sets_structures(pRExC_state,
16669 stack, fence, fence_stack));
16670 #endif
16671
16672 top_index = av_tindex_skip_len_mg(stack);
16673
16674 switch (curchar) {
16675 SV** stacked_ptr; /* Ptr to something already on 'stack' */
16676 char stacked_operator; /* The topmost operator on the 'stack'. */
16677 SV* lhs; /* Operand to the left of the operator */
16678 SV* rhs; /* Operand to the right of the operator */
16679 SV* fence_ptr; /* Pointer to top element of the fence
16680 stack */
16681 case '(':
16682
16683 if ( RExC_parse < RExC_end - 2
16684 && UCHARAT(RExC_parse + 1) == '?'
16685 && UCHARAT(RExC_parse + 2) == '^')
16686 {
16687 const regnode_offset orig_emit = RExC_emit;
16688 SV * resultant_invlist;
16689
16690 /* If is a '(?^', could be an embedded '(?^flags:(?[...])'.
16691 * This happens when we have some thing like
16692 *
16693 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
16694 * ...
16695 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
16696 *
16697 * Here we would be handling the interpolated
16698 * '$thai_or_lao'. We handle this by a recursive call to
16699 * reg which returns the inversion list the
16700 * interpolated expression evaluates to. Actually, the
16701 * return is a special regnode containing a pointer to that
16702 * inversion list. If the return isn't that regnode alone,
16703 * we know that this wasn't such an interpolation, which is
16704 * an error: we need to get a single inversion list back
16705 * from the recursion */
16706
16707 RExC_parse++;
16708 RExC_sets_depth++;
16709
16710 node = reg(pRExC_state, 2, flagp, depth+1);
16711 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16712
16713 if ( OP(REGNODE_p(node)) != REGEX_SET
16714 /* If more than a single node returned, the nested
16715 * parens evaluated to more than just a (?[...]),
16716 * which isn't legal */
16717 || RExC_emit != orig_emit
16718 + NODE_STEP_REGNODE
16719 + regarglen[REGEX_SET])
16720 {
16721 vFAIL("Expecting interpolated extended charclass");
16722 }
16723 resultant_invlist = (SV *) ARGp(REGNODE_p(node));
16724 current = invlist_clone(resultant_invlist, NULL);
16725 SvREFCNT_dec(resultant_invlist);
16726
16727 RExC_sets_depth--;
16728 RExC_emit = orig_emit;
16729 goto handle_operand;
16730 }
16731
16732 /* A regular '('. Look behind for illegal syntax */
16733 if (top_index - fence >= 0) {
16734 /* If the top entry on the stack is an operator, it had
16735 * better be a '!', otherwise the entry below the top
16736 * operand should be an operator */
16737 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
16738 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
16739 || ( IS_OPERAND(*top_ptr)
16740 && ( top_index - fence < 1
16741 || ! (stacked_ptr = av_fetch(stack,
16742 top_index - 1,
16743 FALSE))
16744 || ! IS_OPERATOR(*stacked_ptr))))
16745 {
16746 RExC_parse++;
16747 vFAIL("Unexpected '(' with no preceding operator");
16748 }
16749 }
16750
16751 /* Stack the position of this undealt-with left paren */
16752 av_push(fence_stack, newSViv(fence));
16753 fence = top_index + 1;
16754 break;
16755
16756 case '\\':
16757 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16758 * multi-char folds are allowed. */
16759 if (!regclass(pRExC_state, flagp, depth+1,
16760 TRUE, /* means parse just the next thing */
16761 FALSE, /* don't allow multi-char folds */
16762 FALSE, /* don't silence non-portable warnings. */
16763 TRUE, /* strict */
16764 FALSE, /* Require return to be an ANYOF */
16765 ¤t))
16766 {
16767 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16768 goto regclass_failed;
16769 }
16770
16771 assert(current);
16772
16773 /* regclass() will return with parsing just the \ sequence,
16774 * leaving the parse pointer at the next thing to parse */
16775 RExC_parse--;
16776 goto handle_operand;
16777
16778 case '[': /* Is a bracketed character class */
16779 {
16780 /* See if this is a [:posix:] class. */
16781 bool is_posix_class = (OOB_NAMEDCLASS
16782 < handle_possible_posix(pRExC_state,
16783 RExC_parse + 1,
16784 NULL,
16785 NULL,
16786 TRUE /* checking only */));
16787 /* If it is a posix class, leave the parse pointer at the '['
16788 * to fool regclass() into thinking it is part of a
16789 * '[[:posix:]]'. */
16790 if (! is_posix_class) {
16791 RExC_parse++;
16792 }
16793
16794 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16795 * multi-char folds are allowed. */
16796 if (!regclass(pRExC_state, flagp, depth+1,
16797 is_posix_class, /* parse the whole char
16798 class only if not a
16799 posix class */
16800 FALSE, /* don't allow multi-char folds */
16801 TRUE, /* silence non-portable warnings. */
16802 TRUE, /* strict */
16803 FALSE, /* Require return to be an ANYOF */
16804 ¤t))
16805 {
16806 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16807 goto regclass_failed;
16808 }
16809
16810 assert(current);
16811
16812 /* function call leaves parse pointing to the ']', except if we
16813 * faked it */
16814 if (is_posix_class) {
16815 RExC_parse--;
16816 }
16817
16818 goto handle_operand;
16819 }
16820
16821 case ']':
16822 if (top_index >= 1) {
16823 goto join_operators;
16824 }
16825
16826 /* Only a single operand on the stack: are done */
16827 goto done;
16828
16829 case ')':
16830 if (av_tindex_skip_len_mg(fence_stack) < 0) {
16831 if (UCHARAT(RExC_parse - 1) == ']') {
16832 break;
16833 }
16834 RExC_parse++;
16835 vFAIL("Unexpected ')'");
16836 }
16837
16838 /* If nothing after the fence, is missing an operand */
16839 if (top_index - fence < 0) {
16840 RExC_parse++;
16841 goto bad_syntax;
16842 }
16843 /* If at least two things on the stack, treat this as an
16844 * operator */
16845 if (top_index - fence >= 1) {
16846 goto join_operators;
16847 }
16848
16849 /* Here only a single thing on the fenced stack, and there is a
16850 * fence. Get rid of it */
16851 fence_ptr = av_pop(fence_stack);
16852 assert(fence_ptr);
16853 fence = SvIV(fence_ptr);
16854 SvREFCNT_dec_NN(fence_ptr);
16855 fence_ptr = NULL;
16856
16857 if (fence < 0) {
16858 fence = 0;
16859 }
16860
16861 /* Having gotten rid of the fence, we pop the operand at the
16862 * stack top and process it as a newly encountered operand */
16863 current = av_pop(stack);
16864 if (IS_OPERAND(current)) {
16865 goto handle_operand;
16866 }
16867
16868 RExC_parse++;
16869 goto bad_syntax;
16870
16871 case '&':
16872 case '|':
16873 case '+':
16874 case '-':
16875 case '^':
16876
16877 /* These binary operators should have a left operand already
16878 * parsed */
16879 if ( top_index - fence < 0
16880 || top_index - fence == 1
16881 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
16882 || ! IS_OPERAND(*top_ptr))
16883 {
16884 goto unexpected_binary;
16885 }
16886
16887 /* If only the one operand is on the part of the stack visible
16888 * to us, we just place this operator in the proper position */
16889 if (top_index - fence < 2) {
16890
16891 /* Place the operator before the operand */
16892
16893 SV* lhs = av_pop(stack);
16894 av_push(stack, newSVuv(curchar));
16895 av_push(stack, lhs);
16896 break;
16897 }
16898
16899 /* But if there is something else on the stack, we need to
16900 * process it before this new operator if and only if the
16901 * stacked operation has equal or higher precedence than the
16902 * new one */
16903
16904 join_operators:
16905
16906 /* The operator on the stack is supposed to be below both its
16907 * operands */
16908 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
16909 || IS_OPERAND(*stacked_ptr))
16910 {
16911 /* But if not, it's legal and indicates we are completely
16912 * done if and only if we're currently processing a ']',
16913 * which should be the final thing in the expression */
16914 if (curchar == ']') {
16915 goto done;
16916 }
16917
16918 unexpected_binary:
16919 RExC_parse++;
16920 vFAIL2("Unexpected binary operator '%c' with no "
16921 "preceding operand", curchar);
16922 }
16923 stacked_operator = (char) SvUV(*stacked_ptr);
16924
16925 if (regex_set_precedence(curchar)
16926 > regex_set_precedence(stacked_operator))
16927 {
16928 /* Here, the new operator has higher precedence than the
16929 * stacked one. This means we need to add the new one to
16930 * the stack to await its rhs operand (and maybe more
16931 * stuff). We put it before the lhs operand, leaving
16932 * untouched the stacked operator and everything below it
16933 * */
16934 lhs = av_pop(stack);
16935 assert(IS_OPERAND(lhs));
16936
16937 av_push(stack, newSVuv(curchar));
16938 av_push(stack, lhs);
16939 break;
16940 }
16941
16942 /* Here, the new operator has equal or lower precedence than
16943 * what's already there. This means the operation already
16944 * there should be performed now, before the new one. */
16945
16946 rhs = av_pop(stack);
16947 if (! IS_OPERAND(rhs)) {
16948
16949 /* This can happen when a ! is not followed by an operand,
16950 * like in /(?[\t &!])/ */
16951 goto bad_syntax;
16952 }
16953
16954 lhs = av_pop(stack);
16955
16956 if (! IS_OPERAND(lhs)) {
16957
16958 /* This can happen when there is an empty (), like in
16959 * /(?[[0]+()+])/ */
16960 goto bad_syntax;
16961 }
16962
16963 switch (stacked_operator) {
16964 case '&':
16965 _invlist_intersection(lhs, rhs, &rhs);
16966 break;
16967
16968 case '|':
16969 case '+':
16970 _invlist_union(lhs, rhs, &rhs);
16971 break;
16972
16973 case '-':
16974 _invlist_subtract(lhs, rhs, &rhs);
16975 break;
16976
16977 case '^': /* The union minus the intersection */
16978 {
16979 SV* i = NULL;
16980 SV* u = NULL;
16981
16982 _invlist_union(lhs, rhs, &u);
16983 _invlist_intersection(lhs, rhs, &i);
16984 _invlist_subtract(u, i, &rhs);
16985 SvREFCNT_dec_NN(i);
16986 SvREFCNT_dec_NN(u);
16987 break;
16988 }
16989 }
16990 SvREFCNT_dec(lhs);
16991
16992 /* Here, the higher precedence operation has been done, and the
16993 * result is in 'rhs'. We overwrite the stacked operator with
16994 * the result. Then we redo this code to either push the new
16995 * operator onto the stack or perform any higher precedence
16996 * stacked operation */
16997 only_to_avoid_leaks = av_pop(stack);
16998 SvREFCNT_dec(only_to_avoid_leaks);
16999 av_push(stack, rhs);
17000 goto redo_curchar;
17001
17002 case '!': /* Highest priority, right associative */
17003
17004 /* If what's already at the top of the stack is another '!",
17005 * they just cancel each other out */
17006 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
17007 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
17008 {
17009 only_to_avoid_leaks = av_pop(stack);
17010 SvREFCNT_dec(only_to_avoid_leaks);
17011 }
17012 else { /* Otherwise, since it's right associative, just push
17013 onto the stack */
17014 av_push(stack, newSVuv(curchar));
17015 }
17016 break;
17017
17018 default:
17019 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
17020 if (RExC_parse >= RExC_end) {
17021 break;
17022 }
17023 vFAIL("Unexpected character");
17024
17025 handle_operand:
17026
17027 /* Here 'current' is the operand. If something is already on the
17028 * stack, we have to check if it is a !. But first, the code above
17029 * may have altered the stack in the time since we earlier set
17030 * 'top_index'. */
17031
17032 top_index = av_tindex_skip_len_mg(stack);
17033 if (top_index - fence >= 0) {
17034 /* If the top entry on the stack is an operator, it had better
17035 * be a '!', otherwise the entry below the top operand should
17036 * be an operator */
17037 top_ptr = av_fetch(stack, top_index, FALSE);
17038 assert(top_ptr);
17039 if (IS_OPERATOR(*top_ptr)) {
17040
17041 /* The only permissible operator at the top of the stack is
17042 * '!', which is applied immediately to this operand. */
17043 curchar = (char) SvUV(*top_ptr);
17044 if (curchar != '!') {
17045 SvREFCNT_dec(current);
17046 vFAIL2("Unexpected binary operator '%c' with no "
17047 "preceding operand", curchar);
17048 }
17049
17050 _invlist_invert(current);
17051
17052 only_to_avoid_leaks = av_pop(stack);
17053 SvREFCNT_dec(only_to_avoid_leaks);
17054
17055 /* And we redo with the inverted operand. This allows
17056 * handling multiple ! in a row */
17057 goto handle_operand;
17058 }
17059 /* Single operand is ok only for the non-binary ')'
17060 * operator */
17061 else if ((top_index - fence == 0 && curchar != ')')
17062 || (top_index - fence > 0
17063 && (! (stacked_ptr = av_fetch(stack,
17064 top_index - 1,
17065 FALSE))
17066 || IS_OPERAND(*stacked_ptr))))
17067 {
17068 SvREFCNT_dec(current);
17069 vFAIL("Operand with no preceding operator");
17070 }
17071 }
17072
17073 /* Here there was nothing on the stack or the top element was
17074 * another operand. Just add this new one */
17075 av_push(stack, current);
17076
17077 } /* End of switch on next parse token */
17078
17079 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
17080 } /* End of loop parsing through the construct */
17081
17082 vFAIL("Syntax error in (?[...])");
17083
17084 done:
17085
17086 if (RExC_parse >= RExC_end || RExC_parse[1] != ')') {
17087 if (RExC_parse < RExC_end) {
17088 RExC_parse++;
17089 }
17090
17091 vFAIL("Unexpected ']' with no following ')' in (?[...");
17092 }
17093
17094 if (av_tindex_skip_len_mg(fence_stack) >= 0) {
17095 vFAIL("Unmatched (");
17096 }
17097
17098 if (av_tindex_skip_len_mg(stack) < 0 /* Was empty */
17099 || ((final = av_pop(stack)) == NULL)
17100 || ! IS_OPERAND(final)
17101 || ! is_invlist(final)
17102 || av_tindex_skip_len_mg(stack) >= 0) /* More left on stack */
17103 {
17104 bad_syntax:
17105 SvREFCNT_dec(final);
17106 vFAIL("Incomplete expression within '(?[ ])'");
17107 }
17108
17109 /* Here, 'final' is the resultant inversion list from evaluating the
17110 * expression. Return it if so requested */
17111 if (return_invlist) {
17112 *return_invlist = final;
17113 return END;
17114 }
17115
17116 if (RExC_sets_depth) { /* If within a recursive call, return in a special
17117 regnode */
17118 RExC_parse++;
17119 node = regpnode(pRExC_state, REGEX_SET, final);
17120 }
17121 else {
17122
17123 /* Otherwise generate a resultant node, based on 'final'. regclass()
17124 * is expecting a string of ranges and individual code points */
17125 invlist_iterinit(final);
17126 result_string = newSVpvs("");
17127 while (invlist_iternext(final, &start, &end)) {
17128 if (start == end) {
17129 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
17130 }
17131 else {
17132 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%"
17133 UVXf "}", start, end);
17134 }
17135 }
17136
17137 /* About to generate an ANYOF (or similar) node from the inversion list
17138 * we have calculated */
17139 save_parse = RExC_parse;
17140 RExC_parse = SvPV(result_string, len);
17141 save_end = RExC_end;
17142 RExC_end = RExC_parse + len;
17143 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
17144
17145 /* We turn off folding around the call, as the class we have
17146 * constructed already has all folding taken into consideration, and we
17147 * don't want regclass() to add to that */
17148 RExC_flags &= ~RXf_PMf_FOLD;
17149 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if multi-char
17150 * folds are allowed. */
17151 node = regclass(pRExC_state, flagp, depth+1,
17152 FALSE, /* means parse the whole char class */
17153 FALSE, /* don't allow multi-char folds */
17154 TRUE, /* silence non-portable warnings. The above may
17155 very well have generated non-portable code
17156 points, but they're valid on this machine */
17157 FALSE, /* similarly, no need for strict */
17158
17159 /* We can optimize into something besides an ANYOF,
17160 * except under /l, which needs to be ANYOF because of
17161 * runtime checks for locale sanity, etc */
17162 ! in_locale,
17163 NULL
17164 );
17165
17166 RESTORE_WARNINGS;
17167 RExC_parse = save_parse + 1;
17168 RExC_end = save_end;
17169 SvREFCNT_dec_NN(final);
17170 SvREFCNT_dec_NN(result_string);
17171
17172 if (save_fold) {
17173 RExC_flags |= RXf_PMf_FOLD;
17174 }
17175
17176 if (!node) {
17177 RETURN_FAIL_ON_RESTART(*flagp, flagp);
17178 goto regclass_failed;
17179 }
17180
17181 /* Fix up the node type if we are in locale. (We have pretended we are
17182 * under /u for the purposes of regclass(), as this construct will only
17183 * work under UTF-8 locales. But now we change the opcode to be ANYOFL
17184 * (so as to cause any warnings about bad locales to be output in
17185 * regexec.c), and add the flag that indicates to check if not in a
17186 * UTF-8 locale. The reason we above forbid optimization into
17187 * something other than an ANYOF node is simply to minimize the number
17188 * of code changes in regexec.c. Otherwise we would have to create new
17189 * EXACTish node types and deal with them. This decision could be
17190 * revisited should this construct become popular.
17191 *
17192 * (One might think we could look at the resulting ANYOF node and
17193 * suppress the flag if everything is above 255, as those would be
17194 * UTF-8 only, but this isn't true, as the components that led to that
17195 * result could have been locale-affected, and just happen to cancel
17196 * each other out under UTF-8 locales.) */
17197 if (in_locale) {
17198 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
17199
17200 assert(OP(REGNODE_p(node)) == ANYOF);
17201
17202 OP(REGNODE_p(node)) = ANYOFL;
17203 ANYOF_FLAGS(REGNODE_p(node))
17204 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
17205 }
17206 }
17207
17208 nextchar(pRExC_state);
17209 Set_Node_Length(REGNODE_p(node), RExC_parse - oregcomp_parse + 1); /* MJD */
17210 return node;
17211
17212 regclass_failed:
17213 FAIL2("panic: regclass returned failure to handle_sets, " "flags=%#" UVxf,
17214 (UV) *flagp);
17215 }
17216
17217 #ifdef ENABLE_REGEX_SETS_DEBUGGING
17218
17219 STATIC void
S_dump_regex_sets_structures(pTHX_ RExC_state_t * pRExC_state,AV * stack,const IV fence,AV * fence_stack)17220 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
17221 AV * stack, const IV fence, AV * fence_stack)
17222 { /* Dumps the stacks in handle_regex_sets() */
17223
17224 const SSize_t stack_top = av_tindex_skip_len_mg(stack);
17225 const SSize_t fence_stack_top = av_tindex_skip_len_mg(fence_stack);
17226 SSize_t i;
17227
17228 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
17229
17230 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
17231
17232 if (stack_top < 0) {
17233 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
17234 }
17235 else {
17236 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
17237 for (i = stack_top; i >= 0; i--) {
17238 SV ** element_ptr = av_fetch(stack, i, FALSE);
17239 if (! element_ptr) {
17240 }
17241
17242 if (IS_OPERATOR(*element_ptr)) {
17243 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
17244 (int) i, (int) SvIV(*element_ptr));
17245 }
17246 else {
17247 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
17248 sv_dump(*element_ptr);
17249 }
17250 }
17251 }
17252
17253 if (fence_stack_top < 0) {
17254 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
17255 }
17256 else {
17257 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
17258 for (i = fence_stack_top; i >= 0; i--) {
17259 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
17260 if (! element_ptr) {
17261 }
17262
17263 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
17264 (int) i, (int) SvIV(*element_ptr));
17265 }
17266 }
17267 }
17268
17269 #endif
17270
17271 #undef IS_OPERATOR
17272 #undef IS_OPERAND
17273
17274 STATIC void
S_add_above_Latin1_folds(pTHX_ RExC_state_t * pRExC_state,const U8 cp,SV ** invlist)17275 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
17276 {
17277 /* This adds the Latin1/above-Latin1 folding rules.
17278 *
17279 * This should be called only for a Latin1-range code points, cp, which is
17280 * known to be involved in a simple fold with other code points above
17281 * Latin1. It would give false results if /aa has been specified.
17282 * Multi-char folds are outside the scope of this, and must be handled
17283 * specially. */
17284
17285 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
17286
17287 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
17288
17289 /* The rules that are valid for all Unicode versions are hard-coded in */
17290 switch (cp) {
17291 case 'k':
17292 case 'K':
17293 *invlist =
17294 add_cp_to_invlist(*invlist, KELVIN_SIGN);
17295 break;
17296 case 's':
17297 case 'S':
17298 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
17299 break;
17300 case MICRO_SIGN:
17301 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
17302 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
17303 break;
17304 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
17305 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
17306 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
17307 break;
17308 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
17309 *invlist = add_cp_to_invlist(*invlist,
17310 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
17311 break;
17312
17313 default: /* Other code points are checked against the data for the
17314 current Unicode version */
17315 {
17316 Size_t folds_count;
17317 U32 first_fold;
17318 const U32 * remaining_folds;
17319 UV folded_cp;
17320
17321 if (isASCII(cp)) {
17322 folded_cp = toFOLD(cp);
17323 }
17324 else {
17325 U8 dummy_fold[UTF8_MAXBYTES_CASE+1];
17326 Size_t dummy_len;
17327 folded_cp = _to_fold_latin1(cp, dummy_fold, &dummy_len, 0);
17328 }
17329
17330 if (folded_cp > 255) {
17331 *invlist = add_cp_to_invlist(*invlist, folded_cp);
17332 }
17333
17334 folds_count = _inverse_folds(folded_cp, &first_fold,
17335 &remaining_folds);
17336 if (folds_count == 0) {
17337
17338 /* Use deprecated warning to increase the chances of this being
17339 * output */
17340 ckWARN2reg_d(RExC_parse,
17341 "Perl folding rules are not up-to-date for 0x%02X;"
17342 " please use the perlbug utility to report;", cp);
17343 }
17344 else {
17345 unsigned int i;
17346
17347 if (first_fold > 255) {
17348 *invlist = add_cp_to_invlist(*invlist, first_fold);
17349 }
17350 for (i = 0; i < folds_count - 1; i++) {
17351 if (remaining_folds[i] > 255) {
17352 *invlist = add_cp_to_invlist(*invlist,
17353 remaining_folds[i]);
17354 }
17355 }
17356 }
17357 break;
17358 }
17359 }
17360 }
17361
17362 STATIC void
S_output_posix_warnings(pTHX_ RExC_state_t * pRExC_state,AV * posix_warnings)17363 S_output_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings)
17364 {
17365 /* Output the elements of the array given by '*posix_warnings' as REGEXP
17366 * warnings. */
17367
17368 SV * msg;
17369 const bool first_is_fatal = ckDEAD(packWARN(WARN_REGEXP));
17370
17371 PERL_ARGS_ASSERT_OUTPUT_POSIX_WARNINGS;
17372
17373 if (! TO_OUTPUT_WARNINGS(RExC_parse)) {
17374 CLEAR_POSIX_WARNINGS();
17375 return;
17376 }
17377
17378 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
17379 if (first_is_fatal) { /* Avoid leaking this */
17380 av_undef(posix_warnings); /* This isn't necessary if the
17381 array is mortal, but is a
17382 fail-safe */
17383 (void) sv_2mortal(msg);
17384 PREPARE_TO_DIE;
17385 }
17386 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
17387 SvREFCNT_dec_NN(msg);
17388 }
17389
17390 UPDATE_WARNINGS_LOC(RExC_parse);
17391 }
17392
17393 PERL_STATIC_INLINE Size_t
S_find_first_differing_byte_pos(const U8 * s1,const U8 * s2,const Size_t max)17394 S_find_first_differing_byte_pos(const U8 * s1, const U8 * s2, const Size_t max)
17395 {
17396 const U8 * const start = s1;
17397 const U8 * const send = start + max;
17398
17399 PERL_ARGS_ASSERT_FIND_FIRST_DIFFERING_BYTE_POS;
17400
17401 while (s1 < send && *s1 == *s2) {
17402 s1++; s2++;
17403 }
17404
17405 return s1 - start;
17406 }
17407
17408
17409 STATIC AV *
S_add_multi_match(pTHX_ AV * multi_char_matches,SV * multi_string,const STRLEN cp_count)17410 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
17411 {
17412 /* This adds the string scalar <multi_string> to the array
17413 * <multi_char_matches>. <multi_string> is known to have exactly
17414 * <cp_count> code points in it. This is used when constructing a
17415 * bracketed character class and we find something that needs to match more
17416 * than a single character.
17417 *
17418 * <multi_char_matches> is actually an array of arrays. Each top-level
17419 * element is an array that contains all the strings known so far that are
17420 * the same length. And that length (in number of code points) is the same
17421 * as the index of the top-level array. Hence, the [2] element is an
17422 * array, each element thereof is a string containing TWO code points;
17423 * while element [3] is for strings of THREE characters, and so on. Since
17424 * this is for multi-char strings there can never be a [0] nor [1] element.
17425 *
17426 * When we rewrite the character class below, we will do so such that the
17427 * longest strings are written first, so that it prefers the longest
17428 * matching strings first. This is done even if it turns out that any
17429 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
17430 * Christiansen has agreed that this is ok. This makes the test for the
17431 * ligature 'ffi' come before the test for 'ff', for example */
17432
17433 AV* this_array;
17434 AV** this_array_ptr;
17435
17436 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
17437
17438 if (! multi_char_matches) {
17439 multi_char_matches = newAV();
17440 }
17441
17442 if (av_exists(multi_char_matches, cp_count)) {
17443 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
17444 this_array = *this_array_ptr;
17445 }
17446 else {
17447 this_array = newAV();
17448 av_store(multi_char_matches, cp_count,
17449 (SV*) this_array);
17450 }
17451 av_push(this_array, multi_string);
17452
17453 return multi_char_matches;
17454 }
17455
17456 /* The names of properties whose definitions are not known at compile time are
17457 * stored in this SV, after a constant heading. So if the length has been
17458 * changed since initialization, then there is a run-time definition. */
17459 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
17460 (SvCUR(listsv) != initial_listsv_len)
17461
17462 /* There is a restricted set of white space characters that are legal when
17463 * ignoring white space in a bracketed character class. This generates the
17464 * code to skip them.
17465 *
17466 * There is a line below that uses the same white space criteria but is outside
17467 * this macro. Both here and there must use the same definition */
17468 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p, stop_p) \
17469 STMT_START { \
17470 if (do_skip) { \
17471 while (p < stop_p && isBLANK_A(UCHARAT(p))) \
17472 { \
17473 p++; \
17474 } \
17475 } \
17476 } STMT_END
17477
17478 STATIC regnode_offset
S_regclass(pTHX_ RExC_state_t * pRExC_state,I32 * flagp,U32 depth,const bool stop_at_1,bool allow_mutiple_chars,const bool silence_non_portable,const bool strict,bool optimizable,SV ** ret_invlist)17479 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
17480 const bool stop_at_1, /* Just parse the next thing, don't
17481 look for a full character class */
17482 bool allow_mutiple_chars,
17483 const bool silence_non_portable, /* Don't output warnings
17484 about too large
17485 characters */
17486 const bool strict,
17487 bool optimizable, /* ? Allow a non-ANYOF return
17488 node */
17489 SV** ret_invlist /* Return an inversion list, not a node */
17490 )
17491 {
17492 /* parse a bracketed class specification. Most of these will produce an
17493 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
17494 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
17495 * under /i with multi-character folds: it will be rewritten following the
17496 * paradigm of this example, where the <multi-fold>s are characters which
17497 * fold to multiple character sequences:
17498 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
17499 * gets effectively rewritten as:
17500 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
17501 * reg() gets called (recursively) on the rewritten version, and this
17502 * function will return what it constructs. (Actually the <multi-fold>s
17503 * aren't physically removed from the [abcdefghi], it's just that they are
17504 * ignored in the recursion by means of a flag:
17505 * <RExC_in_multi_char_class>.)
17506 *
17507 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
17508 * characters, with the corresponding bit set if that character is in the
17509 * list. For characters above this, an inversion list is used. There
17510 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
17511 * determinable at compile time
17512 *
17513 * On success, returns the offset at which any next node should be placed
17514 * into the regex engine program being compiled.
17515 *
17516 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
17517 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
17518 * UTF-8
17519 */
17520
17521 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
17522 IV range = 0;
17523 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
17524 regnode_offset ret = -1; /* Initialized to an illegal value */
17525 STRLEN numlen;
17526 int namedclass = OOB_NAMEDCLASS;
17527 char *rangebegin = NULL;
17528 SV *listsv = NULL; /* List of \p{user-defined} whose definitions
17529 aren't available at the time this was called */
17530 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
17531 than just initialized. */
17532 SV* properties = NULL; /* Code points that match \p{} \P{} */
17533 SV* posixes = NULL; /* Code points that match classes like [:word:],
17534 extended beyond the Latin1 range. These have to
17535 be kept separate from other code points for much
17536 of this function because their handling is
17537 different under /i, and for most classes under
17538 /d as well */
17539 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
17540 separate for a while from the non-complemented
17541 versions because of complications with /d
17542 matching */
17543 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
17544 treated more simply than the general case,
17545 leading to less compilation and execution
17546 work */
17547 UV element_count = 0; /* Number of distinct elements in the class.
17548 Optimizations may be possible if this is tiny */
17549 AV * multi_char_matches = NULL; /* Code points that fold to more than one
17550 character; used under /i */
17551 UV n;
17552 char * stop_ptr = RExC_end; /* where to stop parsing */
17553
17554 /* ignore unescaped whitespace? */
17555 const bool skip_white = cBOOL( ret_invlist
17556 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
17557
17558 /* inversion list of code points this node matches only when the target
17559 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
17560 * /d) */
17561 SV* upper_latin1_only_utf8_matches = NULL;
17562
17563 /* Inversion list of code points this node matches regardless of things
17564 * like locale, folding, utf8ness of the target string */
17565 SV* cp_list = NULL;
17566
17567 /* Like cp_list, but code points on this list need to be checked for things
17568 * that fold to/from them under /i */
17569 SV* cp_foldable_list = NULL;
17570
17571 /* Like cp_list, but code points on this list are valid only when the
17572 * runtime locale is UTF-8 */
17573 SV* only_utf8_locale_list = NULL;
17574
17575 /* In a range, if one of the endpoints is non-character-set portable,
17576 * meaning that it hard-codes a code point that may mean a different
17577 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
17578 * mnemonic '\t' which each mean the same character no matter which
17579 * character set the platform is on. */
17580 unsigned int non_portable_endpoint = 0;
17581
17582 /* Is the range unicode? which means on a platform that isn't 1-1 native
17583 * to Unicode (i.e. non-ASCII), each code point in it should be considered
17584 * to be a Unicode value. */
17585 bool unicode_range = FALSE;
17586 bool invert = FALSE; /* Is this class to be complemented */
17587
17588 bool warn_super = ALWAYS_WARN_SUPER;
17589
17590 const char * orig_parse = RExC_parse;
17591
17592 /* This variable is used to mark where the end in the input is of something
17593 * that looks like a POSIX construct but isn't. During the parse, when
17594 * something looks like it could be such a construct is encountered, it is
17595 * checked for being one, but not if we've already checked this area of the
17596 * input. Only after this position is reached do we check again */
17597 char *not_posix_region_end = RExC_parse - 1;
17598
17599 AV* posix_warnings = NULL;
17600 const bool do_posix_warnings = ckWARN(WARN_REGEXP);
17601 U8 op = ANYOF; /* The returned node-type, initialized to the expected
17602 type. */
17603 U8 anyof_flags = 0; /* flag bits if the node is an ANYOF-type */
17604 U32 posixl = 0; /* bit field of posix classes matched under /l */
17605
17606
17607 /* Flags as to what things aren't knowable until runtime. (Note that these are
17608 * mutually exclusive.) */
17609 #define HAS_USER_DEFINED_PROPERTY 0x01 /* /u any user-defined properties that
17610 haven't been defined as of yet */
17611 #define HAS_D_RUNTIME_DEPENDENCY 0x02 /* /d if the target being matched is
17612 UTF-8 or not */
17613 #define HAS_L_RUNTIME_DEPENDENCY 0x04 /* /l what the posix classes match and
17614 what gets folded */
17615 U32 has_runtime_dependency = 0; /* OR of the above flags */
17616
17617 DECLARE_AND_GET_RE_DEBUG_FLAGS;
17618
17619 PERL_ARGS_ASSERT_REGCLASS;
17620 #ifndef DEBUGGING
17621 PERL_UNUSED_ARG(depth);
17622 #endif
17623
17624 assert(! (ret_invlist && allow_mutiple_chars));
17625
17626 /* If wants an inversion list returned, we can't optimize to something
17627 * else. */
17628 if (ret_invlist) {
17629 optimizable = FALSE;
17630 }
17631
17632 DEBUG_PARSE("clas");
17633
17634 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
17635 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
17636 && UNICODE_DOT_DOT_VERSION == 0)
17637 allow_mutiple_chars = FALSE;
17638 #endif
17639
17640 /* We include the /i status at the beginning of this so that we can
17641 * know it at runtime */
17642 listsv = sv_2mortal(Perl_newSVpvf(aTHX_ "#%d\n", cBOOL(FOLD)));
17643 initial_listsv_len = SvCUR(listsv);
17644 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
17645
17646 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
17647
17648 assert(RExC_parse <= RExC_end);
17649
17650 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
17651 RExC_parse++;
17652 invert = TRUE;
17653 allow_mutiple_chars = FALSE;
17654 MARK_NAUGHTY(1);
17655 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
17656 }
17657
17658 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
17659 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
17660 int maybe_class = handle_possible_posix(pRExC_state,
17661 RExC_parse,
17662 ¬_posix_region_end,
17663 NULL,
17664 TRUE /* checking only */);
17665 if (maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
17666 ckWARN4reg(not_posix_region_end,
17667 "POSIX syntax [%c %c] belongs inside character classes%s",
17668 *RExC_parse, *RExC_parse,
17669 (maybe_class == OOB_NAMEDCLASS)
17670 ? ((POSIXCC_NOTYET(*RExC_parse))
17671 ? " (but this one isn't implemented)"
17672 : " (but this one isn't fully valid)")
17673 : ""
17674 );
17675 }
17676 }
17677
17678 /* If the caller wants us to just parse a single element, accomplish this
17679 * by faking the loop ending condition */
17680 if (stop_at_1 && RExC_end > RExC_parse) {
17681 stop_ptr = RExC_parse + 1;
17682 }
17683
17684 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
17685 if (UCHARAT(RExC_parse) == ']')
17686 goto charclassloop;
17687
17688 while (1) {
17689
17690 if ( posix_warnings
17691 && av_tindex_skip_len_mg(posix_warnings) >= 0
17692 && RExC_parse > not_posix_region_end)
17693 {
17694 /* Warnings about posix class issues are considered tentative until
17695 * we are far enough along in the parse that we can no longer
17696 * change our mind, at which point we output them. This is done
17697 * each time through the loop so that a later class won't zap them
17698 * before they have been dealt with. */
17699 output_posix_warnings(pRExC_state, posix_warnings);
17700 }
17701
17702 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
17703
17704 if (RExC_parse >= stop_ptr) {
17705 break;
17706 }
17707
17708 if (UCHARAT(RExC_parse) == ']') {
17709 break;
17710 }
17711
17712 charclassloop:
17713
17714 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
17715 save_value = value;
17716 save_prevvalue = prevvalue;
17717
17718 if (!range) {
17719 rangebegin = RExC_parse;
17720 element_count++;
17721 non_portable_endpoint = 0;
17722 }
17723 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
17724 value = utf8n_to_uvchr((U8*)RExC_parse,
17725 RExC_end - RExC_parse,
17726 &numlen, UTF8_ALLOW_DEFAULT);
17727 RExC_parse += numlen;
17728 }
17729 else
17730 value = UCHARAT(RExC_parse++);
17731
17732 if (value == '[') {
17733 char * posix_class_end;
17734 namedclass = handle_possible_posix(pRExC_state,
17735 RExC_parse,
17736 &posix_class_end,
17737 do_posix_warnings ? &posix_warnings : NULL,
17738 FALSE /* die if error */);
17739 if (namedclass > OOB_NAMEDCLASS) {
17740
17741 /* If there was an earlier attempt to parse this particular
17742 * posix class, and it failed, it was a false alarm, as this
17743 * successful one proves */
17744 if ( posix_warnings
17745 && av_tindex_skip_len_mg(posix_warnings) >= 0
17746 && not_posix_region_end >= RExC_parse
17747 && not_posix_region_end <= posix_class_end)
17748 {
17749 av_undef(posix_warnings);
17750 }
17751
17752 RExC_parse = posix_class_end;
17753 }
17754 else if (namedclass == OOB_NAMEDCLASS) {
17755 not_posix_region_end = posix_class_end;
17756 }
17757 else {
17758 namedclass = OOB_NAMEDCLASS;
17759 }
17760 }
17761 else if ( RExC_parse - 1 > not_posix_region_end
17762 && MAYBE_POSIXCC(value))
17763 {
17764 (void) handle_possible_posix(
17765 pRExC_state,
17766 RExC_parse - 1, /* -1 because parse has already been
17767 advanced */
17768 ¬_posix_region_end,
17769 do_posix_warnings ? &posix_warnings : NULL,
17770 TRUE /* checking only */);
17771 }
17772 else if ( strict && ! skip_white
17773 && ( _generic_isCC(value, _CC_VERTSPACE)
17774 || is_VERTWS_cp_high(value)))
17775 {
17776 vFAIL("Literal vertical space in [] is illegal except under /x");
17777 }
17778 else if (value == '\\') {
17779 /* Is a backslash; get the code point of the char after it */
17780
17781 if (RExC_parse >= RExC_end) {
17782 vFAIL("Unmatched [");
17783 }
17784
17785 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
17786 value = utf8n_to_uvchr((U8*)RExC_parse,
17787 RExC_end - RExC_parse,
17788 &numlen, UTF8_ALLOW_DEFAULT);
17789 RExC_parse += numlen;
17790 }
17791 else
17792 value = UCHARAT(RExC_parse++);
17793
17794 /* Some compilers cannot handle switching on 64-bit integer
17795 * values, therefore value cannot be an UV. Yes, this will
17796 * be a problem later if we want switch on Unicode.
17797 * A similar issue a little bit later when switching on
17798 * namedclass. --jhi */
17799
17800 /* If the \ is escaping white space when white space is being
17801 * skipped, it means that that white space is wanted literally, and
17802 * is already in 'value'. Otherwise, need to translate the escape
17803 * into what it signifies. */
17804 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
17805 const char * message;
17806 U32 packed_warn;
17807 U8 grok_c_char;
17808
17809 case 'w': namedclass = ANYOF_WORDCHAR; break;
17810 case 'W': namedclass = ANYOF_NWORDCHAR; break;
17811 case 's': namedclass = ANYOF_SPACE; break;
17812 case 'S': namedclass = ANYOF_NSPACE; break;
17813 case 'd': namedclass = ANYOF_DIGIT; break;
17814 case 'D': namedclass = ANYOF_NDIGIT; break;
17815 case 'v': namedclass = ANYOF_VERTWS; break;
17816 case 'V': namedclass = ANYOF_NVERTWS; break;
17817 case 'h': namedclass = ANYOF_HORIZWS; break;
17818 case 'H': namedclass = ANYOF_NHORIZWS; break;
17819 case 'N': /* Handle \N{NAME} in class */
17820 {
17821 const char * const backslash_N_beg = RExC_parse - 2;
17822 int cp_count;
17823
17824 if (! grok_bslash_N(pRExC_state,
17825 NULL, /* No regnode */
17826 &value, /* Yes single value */
17827 &cp_count, /* Multiple code pt count */
17828 flagp,
17829 strict,
17830 depth)
17831 ) {
17832
17833 if (*flagp & NEED_UTF8)
17834 FAIL("panic: grok_bslash_N set NEED_UTF8");
17835
17836 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
17837
17838 if (cp_count < 0) {
17839 vFAIL("\\N in a character class must be a named character: \\N{...}");
17840 }
17841 else if (cp_count == 0) {
17842 ckWARNreg(RExC_parse,
17843 "Ignoring zero length \\N{} in character class");
17844 }
17845 else { /* cp_count > 1 */
17846 assert(cp_count > 1);
17847 if (! RExC_in_multi_char_class) {
17848 if ( ! allow_mutiple_chars
17849 || invert
17850 || range
17851 || *RExC_parse == '-')
17852 {
17853 if (strict) {
17854 RExC_parse--;
17855 vFAIL("\\N{} here is restricted to one character");
17856 }
17857 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
17858 break; /* <value> contains the first code
17859 point. Drop out of the switch to
17860 process it */
17861 }
17862 else {
17863 SV * multi_char_N = newSVpvn(backslash_N_beg,
17864 RExC_parse - backslash_N_beg);
17865 multi_char_matches
17866 = add_multi_match(multi_char_matches,
17867 multi_char_N,
17868 cp_count);
17869 }
17870 }
17871 } /* End of cp_count != 1 */
17872
17873 /* This element should not be processed further in this
17874 * class */
17875 element_count--;
17876 value = save_value;
17877 prevvalue = save_prevvalue;
17878 continue; /* Back to top of loop to get next char */
17879 }
17880
17881 /* Here, is a single code point, and <value> contains it */
17882 unicode_range = TRUE; /* \N{} are Unicode */
17883 }
17884 break;
17885 case 'p':
17886 case 'P':
17887 {
17888 char *e;
17889
17890 if (RExC_pm_flags & PMf_WILDCARD) {
17891 RExC_parse++;
17892 /* diag_listed_as: Use of %s is not allowed in Unicode
17893 property wildcard subpatterns in regex; marked by <--
17894 HERE in m/%s/ */
17895 vFAIL3("Use of '\\%c%c' is not allowed in Unicode property"
17896 " wildcard subpatterns", (char) value, *(RExC_parse - 1));
17897 }
17898
17899 /* \p means they want Unicode semantics */
17900 REQUIRE_UNI_RULES(flagp, 0);
17901
17902 if (RExC_parse >= RExC_end)
17903 vFAIL2("Empty \\%c", (U8)value);
17904 if (*RExC_parse == '{') {
17905 const U8 c = (U8)value;
17906 e = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
17907 if (!e) {
17908 RExC_parse++;
17909 vFAIL2("Missing right brace on \\%c{}", c);
17910 }
17911
17912 RExC_parse++;
17913
17914 /* White space is allowed adjacent to the braces and after
17915 * any '^', even when not under /x */
17916 while (isSPACE(*RExC_parse)) {
17917 RExC_parse++;
17918 }
17919
17920 if (UCHARAT(RExC_parse) == '^') {
17921
17922 /* toggle. (The rhs xor gets the single bit that
17923 * differs between P and p; the other xor inverts just
17924 * that bit) */
17925 value ^= 'P' ^ 'p';
17926
17927 RExC_parse++;
17928 while (isSPACE(*RExC_parse)) {
17929 RExC_parse++;
17930 }
17931 }
17932
17933 if (e == RExC_parse)
17934 vFAIL2("Empty \\%c{}", c);
17935
17936 n = e - RExC_parse;
17937 while (isSPACE(*(RExC_parse + n - 1)))
17938 n--;
17939
17940 } /* The \p isn't immediately followed by a '{' */
17941 else if (! isALPHA(*RExC_parse)) {
17942 RExC_parse += (UTF)
17943 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17944 : 1;
17945 vFAIL2("Character following \\%c must be '{' or a "
17946 "single-character Unicode property name",
17947 (U8) value);
17948 }
17949 else {
17950 e = RExC_parse;
17951 n = 1;
17952 }
17953 {
17954 char* name = RExC_parse;
17955
17956 /* Any message returned about expanding the definition */
17957 SV* msg = newSVpvs_flags("", SVs_TEMP);
17958
17959 /* If set TRUE, the property is user-defined as opposed to
17960 * official Unicode */
17961 bool user_defined = FALSE;
17962 AV * strings = NULL;
17963
17964 SV * prop_definition = parse_uniprop_string(
17965 name, n, UTF, FOLD,
17966 FALSE, /* This is compile-time */
17967
17968 /* We can't defer this defn when
17969 * the full result is required in
17970 * this call */
17971 ! cBOOL(ret_invlist),
17972
17973 &strings,
17974 &user_defined,
17975 msg,
17976 0 /* Base level */
17977 );
17978 if (SvCUR(msg)) { /* Assumes any error causes a msg */
17979 assert(prop_definition == NULL);
17980 RExC_parse = e + 1;
17981 if (SvUTF8(msg)) { /* msg being UTF-8 makes the whole
17982 thing so, or else the display is
17983 mojibake */
17984 RExC_utf8 = TRUE;
17985 }
17986 /* diag_listed_as: Can't find Unicode property definition "%s" in regex; marked by <-- HERE in m/%s/ */
17987 vFAIL2utf8f("%" UTF8f, UTF8fARG(SvUTF8(msg),
17988 SvCUR(msg), SvPVX(msg)));
17989 }
17990
17991 assert(prop_definition || strings);
17992
17993 if (strings) {
17994 if (ret_invlist) {
17995 if (! prop_definition) {
17996 RExC_parse = e + 1;
17997 vFAIL("Unicode string properties are not implemented in (?[...])");
17998 }
17999 else {
18000 ckWARNreg(e + 1,
18001 "Using just the single character results"
18002 " returned by \\p{} in (?[...])");
18003 }
18004 }
18005 else if (! RExC_in_multi_char_class) {
18006 if (invert ^ (value == 'P')) {
18007 RExC_parse = e + 1;
18008 vFAIL("Inverting a character class which contains"
18009 " a multi-character sequence is illegal");
18010 }
18011
18012 /* For each multi-character string ... */
18013 while (av_count(strings) > 0) {
18014 /* ... Each entry is itself an array of code
18015 * points. */
18016 AV * this_string = (AV *) av_shift( strings);
18017 STRLEN cp_count = av_count(this_string);
18018 SV * final = newSV(cp_count * 4);
18019 SvPVCLEAR(final);
18020
18021 /* Create another string of sequences of \x{...} */
18022 while (av_count(this_string) > 0) {
18023 SV * character = av_shift(this_string);
18024 UV cp = SvUV(character);
18025
18026 if (cp > 255) {
18027 REQUIRE_UTF8(flagp);
18028 }
18029 Perl_sv_catpvf(aTHX_ final, "\\x{%" UVXf "}",
18030 cp);
18031 SvREFCNT_dec_NN(character);
18032 }
18033 SvREFCNT_dec_NN(this_string);
18034
18035 /* And add that to the list of such things */
18036 multi_char_matches
18037 = add_multi_match(multi_char_matches,
18038 final,
18039 cp_count);
18040 }
18041 }
18042 SvREFCNT_dec_NN(strings);
18043 }
18044
18045 if (! prop_definition) { /* If we got only a string,
18046 this iteration didn't really
18047 find a character */
18048 element_count--;
18049 }
18050 else if (! is_invlist(prop_definition)) {
18051
18052 /* Here, the definition isn't known, so we have gotten
18053 * returned a string that will be evaluated if and when
18054 * encountered at runtime. We add it to the list of
18055 * such properties, along with whether it should be
18056 * complemented or not */
18057 if (value == 'P') {
18058 sv_catpvs(listsv, "!");
18059 }
18060 else {
18061 sv_catpvs(listsv, "+");
18062 }
18063 sv_catsv(listsv, prop_definition);
18064
18065 has_runtime_dependency |= HAS_USER_DEFINED_PROPERTY;
18066
18067 /* We don't know yet what this matches, so have to flag
18068 * it */
18069 anyof_flags |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
18070 }
18071 else {
18072 assert (prop_definition && is_invlist(prop_definition));
18073
18074 /* Here we do have the complete property definition
18075 *
18076 * Temporary workaround for [perl #133136]. For this
18077 * precise input that is in the .t that is failing,
18078 * load utf8.pm, which is what the test wants, so that
18079 * that .t passes */
18080 if ( memEQs(RExC_start, e + 1 - RExC_start,
18081 "foo\\p{Alnum}")
18082 && ! hv_common(GvHVn(PL_incgv),
18083 NULL,
18084 "utf8.pm", sizeof("utf8.pm") - 1,
18085 0, HV_FETCH_ISEXISTS, NULL, 0))
18086 {
18087 require_pv("utf8.pm");
18088 }
18089
18090 if (! user_defined &&
18091 /* We warn on matching an above-Unicode code point
18092 * if the match would return true, except don't
18093 * warn for \p{All}, which has exactly one element
18094 * = 0 */
18095 (_invlist_contains_cp(prop_definition, 0x110000)
18096 && (! (_invlist_len(prop_definition) == 1
18097 && *invlist_array(prop_definition) == 0))))
18098 {
18099 warn_super = TRUE;
18100 }
18101
18102 /* Invert if asking for the complement */
18103 if (value == 'P') {
18104 _invlist_union_complement_2nd(properties,
18105 prop_definition,
18106 &properties);
18107 }
18108 else {
18109 _invlist_union(properties, prop_definition, &properties);
18110 }
18111 }
18112 }
18113
18114 RExC_parse = e + 1;
18115 namedclass = ANYOF_UNIPROP; /* no official name, but it's
18116 named */
18117 }
18118 break;
18119 case 'n': value = '\n'; break;
18120 case 'r': value = '\r'; break;
18121 case 't': value = '\t'; break;
18122 case 'f': value = '\f'; break;
18123 case 'b': value = '\b'; break;
18124 case 'e': value = ESC_NATIVE; break;
18125 case 'a': value = '\a'; break;
18126 case 'o':
18127 RExC_parse--; /* function expects to be pointed at the 'o' */
18128 if (! grok_bslash_o(&RExC_parse,
18129 RExC_end,
18130 &value,
18131 &message,
18132 &packed_warn,
18133 strict,
18134 cBOOL(range), /* MAX_UV allowed for range
18135 upper limit */
18136 UTF))
18137 {
18138 vFAIL(message);
18139 }
18140 else if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
18141 warn_non_literal_string(RExC_parse, packed_warn, message);
18142 }
18143
18144 if (value < 256) {
18145 non_portable_endpoint++;
18146 }
18147 break;
18148 case 'x':
18149 RExC_parse--; /* function expects to be pointed at the 'x' */
18150 if (! grok_bslash_x(&RExC_parse,
18151 RExC_end,
18152 &value,
18153 &message,
18154 &packed_warn,
18155 strict,
18156 cBOOL(range), /* MAX_UV allowed for range
18157 upper limit */
18158 UTF))
18159 {
18160 vFAIL(message);
18161 }
18162 else if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
18163 warn_non_literal_string(RExC_parse, packed_warn, message);
18164 }
18165
18166 if (value < 256) {
18167 non_portable_endpoint++;
18168 }
18169 break;
18170 case 'c':
18171 if (! grok_bslash_c(*RExC_parse, &grok_c_char, &message,
18172 &packed_warn))
18173 {
18174 /* going to die anyway; point to exact spot of
18175 * failure */
18176 RExC_parse += (UTF)
18177 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
18178 : 1;
18179 vFAIL(message);
18180 }
18181
18182 value = grok_c_char;
18183 RExC_parse++;
18184 if (message && TO_OUTPUT_WARNINGS(RExC_parse)) {
18185 warn_non_literal_string(RExC_parse, packed_warn, message);
18186 }
18187
18188 non_portable_endpoint++;
18189 break;
18190 case '0': case '1': case '2': case '3': case '4':
18191 case '5': case '6': case '7':
18192 {
18193 /* Take 1-3 octal digits */
18194 I32 flags = PERL_SCAN_SILENT_ILLDIGIT
18195 | PERL_SCAN_NOTIFY_ILLDIGIT;
18196 numlen = (strict) ? 4 : 3;
18197 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
18198 RExC_parse += numlen;
18199 if (numlen != 3) {
18200 if (strict) {
18201 RExC_parse += (UTF)
18202 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
18203 : 1;
18204 vFAIL("Need exactly 3 octal digits");
18205 }
18206 else if ( (flags & PERL_SCAN_NOTIFY_ILLDIGIT)
18207 && RExC_parse < RExC_end
18208 && isDIGIT(*RExC_parse)
18209 && ckWARN(WARN_REGEXP))
18210 {
18211 reg_warn_non_literal_string(
18212 RExC_parse + 1,
18213 form_alien_digit_msg(8, numlen, RExC_parse,
18214 RExC_end, UTF, FALSE));
18215 }
18216 }
18217 if (value < 256) {
18218 non_portable_endpoint++;
18219 }
18220 break;
18221 }
18222 default:
18223 /* Allow \_ to not give an error */
18224 if (isWORDCHAR(value) && value != '_') {
18225 if (strict) {
18226 vFAIL2("Unrecognized escape \\%c in character class",
18227 (int)value);
18228 }
18229 else {
18230 ckWARN2reg(RExC_parse,
18231 "Unrecognized escape \\%c in character class passed through",
18232 (int)value);
18233 }
18234 }
18235 break;
18236 } /* End of switch on char following backslash */
18237 } /* end of handling backslash escape sequences */
18238
18239 /* Here, we have the current token in 'value' */
18240
18241 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
18242 U8 classnum;
18243
18244 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
18245 * literal, as is the character that began the false range, i.e.
18246 * the 'a' in the examples */
18247 if (range) {
18248 const int w = (RExC_parse >= rangebegin)
18249 ? RExC_parse - rangebegin
18250 : 0;
18251 if (strict) {
18252 vFAIL2utf8f(
18253 "False [] range \"%" UTF8f "\"",
18254 UTF8fARG(UTF, w, rangebegin));
18255 }
18256 else {
18257 ckWARN2reg(RExC_parse,
18258 "False [] range \"%" UTF8f "\"",
18259 UTF8fARG(UTF, w, rangebegin));
18260 cp_list = add_cp_to_invlist(cp_list, '-');
18261 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
18262 prevvalue);
18263 }
18264
18265 range = 0; /* this was not a true range */
18266 element_count += 2; /* So counts for three values */
18267 }
18268
18269 classnum = namedclass_to_classnum(namedclass);
18270
18271 if (LOC && namedclass < ANYOF_POSIXL_MAX
18272 #ifndef HAS_ISASCII
18273 && classnum != _CC_ASCII
18274 #endif
18275 ) {
18276 SV* scratch_list = NULL;
18277
18278 /* What the Posix classes (like \w, [:space:]) match isn't
18279 * generally knowable under locale until actual match time. A
18280 * special node is used for these which has extra space for a
18281 * bitmap, with a bit reserved for each named class that is to
18282 * be matched against. (This isn't needed for \p{} and
18283 * pseudo-classes, as they are not affected by locale, and
18284 * hence are dealt with separately.) However, if a named class
18285 * and its complement are both present, then it matches
18286 * everything, and there is no runtime dependency. Odd numbers
18287 * are the complements of the next lower number, so xor works.
18288 * (Note that something like [\w\D] should match everything,
18289 * because \d should be a proper subset of \w. But rather than
18290 * trust that the locale is well behaved, we leave this to
18291 * runtime to sort out) */
18292 if (POSIXL_TEST(posixl, namedclass ^ 1)) {
18293 cp_list = _add_range_to_invlist(cp_list, 0, UV_MAX);
18294 POSIXL_ZERO(posixl);
18295 has_runtime_dependency &= ~HAS_L_RUNTIME_DEPENDENCY;
18296 anyof_flags &= ~ANYOF_MATCHES_POSIXL;
18297 continue; /* We could ignore the rest of the class, but
18298 best to parse it for any errors */
18299 }
18300 else { /* Here, isn't the complement of any already parsed
18301 class */
18302 POSIXL_SET(posixl, namedclass);
18303 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
18304 anyof_flags |= ANYOF_MATCHES_POSIXL;
18305
18306 /* The above-Latin1 characters are not subject to locale
18307 * rules. Just add them to the unconditionally-matched
18308 * list */
18309
18310 /* Get the list of the above-Latin1 code points this
18311 * matches */
18312 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
18313 PL_XPosix_ptrs[classnum],
18314
18315 /* Odd numbers are complements,
18316 * like NDIGIT, NASCII, ... */
18317 namedclass % 2 != 0,
18318 &scratch_list);
18319 /* Checking if 'cp_list' is NULL first saves an extra
18320 * clone. Its reference count will be decremented at the
18321 * next union, etc, or if this is the only instance, at the
18322 * end of the routine */
18323 if (! cp_list) {
18324 cp_list = scratch_list;
18325 }
18326 else {
18327 _invlist_union(cp_list, scratch_list, &cp_list);
18328 SvREFCNT_dec_NN(scratch_list);
18329 }
18330 continue; /* Go get next character */
18331 }
18332 }
18333 else {
18334
18335 /* Here, is not /l, or is a POSIX class for which /l doesn't
18336 * matter (or is a Unicode property, which is skipped here). */
18337 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
18338 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
18339
18340 /* Here, should be \h, \H, \v, or \V. None of /d, /i
18341 * nor /l make a difference in what these match,
18342 * therefore we just add what they match to cp_list. */
18343 if (classnum != _CC_VERTSPACE) {
18344 assert( namedclass == ANYOF_HORIZWS
18345 || namedclass == ANYOF_NHORIZWS);
18346
18347 /* It turns out that \h is just a synonym for
18348 * XPosixBlank */
18349 classnum = _CC_BLANK;
18350 }
18351
18352 _invlist_union_maybe_complement_2nd(
18353 cp_list,
18354 PL_XPosix_ptrs[classnum],
18355 namedclass % 2 != 0, /* Complement if odd
18356 (NHORIZWS, NVERTWS)
18357 */
18358 &cp_list);
18359 }
18360 }
18361 else if ( AT_LEAST_UNI_SEMANTICS
18362 || classnum == _CC_ASCII
18363 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
18364 || classnum == _CC_XDIGIT)))
18365 {
18366 /* We usually have to worry about /d affecting what POSIX
18367 * classes match, with special code needed because we won't
18368 * know until runtime what all matches. But there is no
18369 * extra work needed under /u and /a; and [:ascii:] is
18370 * unaffected by /d; and :digit: and :xdigit: don't have
18371 * runtime differences under /d. So we can special case
18372 * these, and avoid some extra work below, and at runtime.
18373 * */
18374 _invlist_union_maybe_complement_2nd(
18375 simple_posixes,
18376 ((AT_LEAST_ASCII_RESTRICTED)
18377 ? PL_Posix_ptrs[classnum]
18378 : PL_XPosix_ptrs[classnum]),
18379 namedclass % 2 != 0,
18380 &simple_posixes);
18381 }
18382 else { /* Garden variety class. If is NUPPER, NALPHA, ...
18383 complement and use nposixes */
18384 SV** posixes_ptr = namedclass % 2 == 0
18385 ? &posixes
18386 : &nposixes;
18387 _invlist_union_maybe_complement_2nd(
18388 *posixes_ptr,
18389 PL_XPosix_ptrs[classnum],
18390 namedclass % 2 != 0,
18391 posixes_ptr);
18392 }
18393 }
18394 } /* end of namedclass \blah */
18395
18396 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse, RExC_end);
18397
18398 /* If 'range' is set, 'value' is the ending of a range--check its
18399 * validity. (If value isn't a single code point in the case of a
18400 * range, we should have figured that out above in the code that
18401 * catches false ranges). Later, we will handle each individual code
18402 * point in the range. If 'range' isn't set, this could be the
18403 * beginning of a range, so check for that by looking ahead to see if
18404 * the next real character to be processed is the range indicator--the
18405 * minus sign */
18406
18407 if (range) {
18408 #ifdef EBCDIC
18409 /* For unicode ranges, we have to test that the Unicode as opposed
18410 * to the native values are not decreasing. (Above 255, there is
18411 * no difference between native and Unicode) */
18412 if (unicode_range && prevvalue < 255 && value < 255) {
18413 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
18414 goto backwards_range;
18415 }
18416 }
18417 else
18418 #endif
18419 if (prevvalue > value) /* b-a */ {
18420 int w;
18421 #ifdef EBCDIC
18422 backwards_range:
18423 #endif
18424 w = RExC_parse - rangebegin;
18425 vFAIL2utf8f(
18426 "Invalid [] range \"%" UTF8f "\"",
18427 UTF8fARG(UTF, w, rangebegin));
18428 NOT_REACHED; /* NOTREACHED */
18429 }
18430 }
18431 else {
18432 prevvalue = value; /* save the beginning of the potential range */
18433 if (! stop_at_1 /* Can't be a range if parsing just one thing */
18434 && *RExC_parse == '-')
18435 {
18436 char* next_char_ptr = RExC_parse + 1;
18437
18438 /* Get the next real char after the '-' */
18439 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr, RExC_end);
18440
18441 /* If the '-' is at the end of the class (just before the ']',
18442 * it is a literal minus; otherwise it is a range */
18443 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
18444 RExC_parse = next_char_ptr;
18445
18446 /* a bad range like \w-, [:word:]- ? */
18447 if (namedclass > OOB_NAMEDCLASS) {
18448 if (strict || ckWARN(WARN_REGEXP)) {
18449 const int w = RExC_parse >= rangebegin
18450 ? RExC_parse - rangebegin
18451 : 0;
18452 if (strict) {
18453 vFAIL4("False [] range \"%*.*s\"",
18454 w, w, rangebegin);
18455 }
18456 else {
18457 vWARN4(RExC_parse,
18458 "False [] range \"%*.*s\"",
18459 w, w, rangebegin);
18460 }
18461 }
18462 cp_list = add_cp_to_invlist(cp_list, '-');
18463 element_count++;
18464 } else
18465 range = 1; /* yeah, it's a range! */
18466 continue; /* but do it the next time */
18467 }
18468 }
18469 }
18470
18471 if (namedclass > OOB_NAMEDCLASS) {
18472 continue;
18473 }
18474
18475 /* Here, we have a single value this time through the loop, and
18476 * <prevvalue> is the beginning of the range, if any; or <value> if
18477 * not. */
18478
18479 /* non-Latin1 code point implies unicode semantics. */
18480 if (value > 255) {
18481 if (value > MAX_LEGAL_CP && ( value != UV_MAX
18482 || prevvalue > MAX_LEGAL_CP))
18483 {
18484 vFAIL(form_cp_too_large_msg(16, NULL, 0, value));
18485 }
18486 REQUIRE_UNI_RULES(flagp, 0);
18487 if ( ! silence_non_portable
18488 && UNICODE_IS_PERL_EXTENDED(value)
18489 && TO_OUTPUT_WARNINGS(RExC_parse))
18490 {
18491 ckWARN2_non_literal_string(RExC_parse,
18492 packWARN(WARN_PORTABLE),
18493 PL_extended_cp_format,
18494 value);
18495 }
18496 }
18497
18498 /* Ready to process either the single value, or the completed range.
18499 * For single-valued non-inverted ranges, we consider the possibility
18500 * of multi-char folds. (We made a conscious decision to not do this
18501 * for the other cases because it can often lead to non-intuitive
18502 * results. For example, you have the peculiar case that:
18503 * "s s" =~ /^[^\xDF]+$/i => Y
18504 * "ss" =~ /^[^\xDF]+$/i => N
18505 *
18506 * See [perl #89750] */
18507 if (FOLD && allow_mutiple_chars && value == prevvalue) {
18508 if ( value == LATIN_SMALL_LETTER_SHARP_S
18509 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
18510 value)))
18511 {
18512 /* Here <value> is indeed a multi-char fold. Get what it is */
18513
18514 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
18515 STRLEN foldlen;
18516
18517 UV folded = _to_uni_fold_flags(
18518 value,
18519 foldbuf,
18520 &foldlen,
18521 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
18522 ? FOLD_FLAGS_NOMIX_ASCII
18523 : 0)
18524 );
18525
18526 /* Here, <folded> should be the first character of the
18527 * multi-char fold of <value>, with <foldbuf> containing the
18528 * whole thing. But, if this fold is not allowed (because of
18529 * the flags), <fold> will be the same as <value>, and should
18530 * be processed like any other character, so skip the special
18531 * handling */
18532 if (folded != value) {
18533
18534 /* Skip if we are recursed, currently parsing the class
18535 * again. Otherwise add this character to the list of
18536 * multi-char folds. */
18537 if (! RExC_in_multi_char_class) {
18538 STRLEN cp_count = utf8_length(foldbuf,
18539 foldbuf + foldlen);
18540 SV* multi_fold = sv_2mortal(newSVpvs(""));
18541
18542 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
18543
18544 multi_char_matches
18545 = add_multi_match(multi_char_matches,
18546 multi_fold,
18547 cp_count);
18548
18549 }
18550
18551 /* This element should not be processed further in this
18552 * class */
18553 element_count--;
18554 value = save_value;
18555 prevvalue = save_prevvalue;
18556 continue;
18557 }
18558 }
18559 }
18560
18561 if (strict && ckWARN(WARN_REGEXP)) {
18562 if (range) {
18563
18564 /* If the range starts above 255, everything is portable and
18565 * likely to be so for any forseeable character set, so don't
18566 * warn. */
18567 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
18568 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
18569 }
18570 else if (prevvalue != value) {
18571
18572 /* Under strict, ranges that stop and/or end in an ASCII
18573 * printable should have each end point be a portable value
18574 * for it (preferably like 'A', but we don't warn if it is
18575 * a (portable) Unicode name or code point), and the range
18576 * must be all digits or all letters of the same case.
18577 * Otherwise, the range is non-portable and unclear as to
18578 * what it contains */
18579 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
18580 && ( non_portable_endpoint
18581 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
18582 || (isLOWER_A(prevvalue) && isLOWER_A(value))
18583 || (isUPPER_A(prevvalue) && isUPPER_A(value))
18584 ))) {
18585 vWARN(RExC_parse, "Ranges of ASCII printables should"
18586 " be some subset of \"0-9\","
18587 " \"A-Z\", or \"a-z\"");
18588 }
18589 else if (prevvalue >= FIRST_NON_ASCII_DECIMAL_DIGIT) {
18590 SSize_t index_start;
18591 SSize_t index_final;
18592
18593 /* But the nature of Unicode and languages mean we
18594 * can't do the same checks for above-ASCII ranges,
18595 * except in the case of digit ones. These should
18596 * contain only digits from the same group of 10. The
18597 * ASCII case is handled just above. Hence here, the
18598 * range could be a range of digits. First some
18599 * unlikely special cases. Grandfather in that a range
18600 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
18601 * if its starting value is one of the 10 digits prior
18602 * to it. This is because it is an alternate way of
18603 * writing 19D1, and some people may expect it to be in
18604 * that group. But it is bad, because it won't give
18605 * the expected results. In Unicode 5.2 it was
18606 * considered to be in that group (of 11, hence), but
18607 * this was fixed in the next version */
18608
18609 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
18610 goto warn_bad_digit_range;
18611 }
18612 else if (UNLIKELY( prevvalue >= 0x1D7CE
18613 && value <= 0x1D7FF))
18614 {
18615 /* This is the only other case currently in Unicode
18616 * where the algorithm below fails. The code
18617 * points just above are the end points of a single
18618 * range containing only decimal digits. It is 5
18619 * different series of 0-9. All other ranges of
18620 * digits currently in Unicode are just a single
18621 * series. (And mktables will notify us if a later
18622 * Unicode version breaks this.)
18623 *
18624 * If the range being checked is at most 9 long,
18625 * and the digit values represented are in
18626 * numerical order, they are from the same series.
18627 * */
18628 if ( value - prevvalue > 9
18629 || ((( value - 0x1D7CE) % 10)
18630 <= (prevvalue - 0x1D7CE) % 10))
18631 {
18632 goto warn_bad_digit_range;
18633 }
18634 }
18635 else {
18636
18637 /* For all other ranges of digits in Unicode, the
18638 * algorithm is just to check if both end points
18639 * are in the same series, which is the same range.
18640 * */
18641 index_start = _invlist_search(
18642 PL_XPosix_ptrs[_CC_DIGIT],
18643 prevvalue);
18644
18645 /* Warn if the range starts and ends with a digit,
18646 * and they are not in the same group of 10. */
18647 if ( index_start >= 0
18648 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
18649 && (index_final =
18650 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
18651 value)) != index_start
18652 && index_final >= 0
18653 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
18654 {
18655 warn_bad_digit_range:
18656 vWARN(RExC_parse, "Ranges of digits should be"
18657 " from the same group of"
18658 " 10");
18659 }
18660 }
18661 }
18662 }
18663 }
18664 if ((! range || prevvalue == value) && non_portable_endpoint) {
18665 if (isPRINT_A(value)) {
18666 char literal[3];
18667 unsigned d = 0;
18668 if (isBACKSLASHED_PUNCT(value)) {
18669 literal[d++] = '\\';
18670 }
18671 literal[d++] = (char) value;
18672 literal[d++] = '\0';
18673
18674 vWARN4(RExC_parse,
18675 "\"%.*s\" is more clearly written simply as \"%s\"",
18676 (int) (RExC_parse - rangebegin),
18677 rangebegin,
18678 literal
18679 );
18680 }
18681 else if (isMNEMONIC_CNTRL(value)) {
18682 vWARN4(RExC_parse,
18683 "\"%.*s\" is more clearly written simply as \"%s\"",
18684 (int) (RExC_parse - rangebegin),
18685 rangebegin,
18686 cntrl_to_mnemonic((U8) value)
18687 );
18688 }
18689 }
18690 }
18691
18692 /* Deal with this element of the class */
18693
18694 #ifndef EBCDIC
18695 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18696 prevvalue, value);
18697 #else
18698 /* On non-ASCII platforms, for ranges that span all of 0..255, and ones
18699 * that don't require special handling, we can just add the range like
18700 * we do for ASCII platforms */
18701 if ((UNLIKELY(prevvalue == 0) && value >= 255)
18702 || ! (prevvalue < 256
18703 && (unicode_range
18704 || (! non_portable_endpoint
18705 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
18706 || (isUPPER_A(prevvalue)
18707 && isUPPER_A(value)))))))
18708 {
18709 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18710 prevvalue, value);
18711 }
18712 else {
18713 /* Here, requires special handling. This can be because it is a
18714 * range whose code points are considered to be Unicode, and so
18715 * must be individually translated into native, or because its a
18716 * subrange of 'A-Z' or 'a-z' which each aren't contiguous in
18717 * EBCDIC, but we have defined them to include only the "expected"
18718 * upper or lower case ASCII alphabetics. Subranges above 255 are
18719 * the same in native and Unicode, so can be added as a range */
18720 U8 start = NATIVE_TO_LATIN1(prevvalue);
18721 unsigned j;
18722 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
18723 for (j = start; j <= end; j++) {
18724 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
18725 }
18726 if (value > 255) {
18727 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18728 256, value);
18729 }
18730 }
18731 #endif
18732
18733 range = 0; /* this range (if it was one) is done now */
18734 } /* End of loop through all the text within the brackets */
18735
18736 if ( posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
18737 output_posix_warnings(pRExC_state, posix_warnings);
18738 }
18739
18740 /* If anything in the class expands to more than one character, we have to
18741 * deal with them by building up a substitute parse string, and recursively
18742 * calling reg() on it, instead of proceeding */
18743 if (multi_char_matches) {
18744 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
18745 I32 cp_count;
18746 STRLEN len;
18747 char *save_end = RExC_end;
18748 char *save_parse = RExC_parse;
18749 char *save_start = RExC_start;
18750 Size_t constructed_prefix_len = 0; /* This gives the length of the
18751 constructed portion of the
18752 substitute parse. */
18753 bool first_time = TRUE; /* First multi-char occurrence doesn't get
18754 a "|" */
18755 I32 reg_flags;
18756
18757 assert(! invert);
18758 /* Only one level of recursion allowed */
18759 assert(RExC_copy_start_in_constructed == RExC_precomp);
18760
18761 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
18762 because too confusing */
18763 if (invert) {
18764 sv_catpvs(substitute_parse, "(?:");
18765 }
18766 #endif
18767
18768 /* Look at the longest strings first */
18769 for (cp_count = av_tindex_skip_len_mg(multi_char_matches);
18770 cp_count > 0;
18771 cp_count--)
18772 {
18773
18774 if (av_exists(multi_char_matches, cp_count)) {
18775 AV** this_array_ptr;
18776 SV* this_sequence;
18777
18778 this_array_ptr = (AV**) av_fetch(multi_char_matches,
18779 cp_count, FALSE);
18780 while ((this_sequence = av_pop(*this_array_ptr)) !=
18781 &PL_sv_undef)
18782 {
18783 if (! first_time) {
18784 sv_catpvs(substitute_parse, "|");
18785 }
18786 first_time = FALSE;
18787
18788 sv_catpv(substitute_parse, SvPVX(this_sequence));
18789 }
18790 }
18791 }
18792
18793 /* If the character class contains anything else besides these
18794 * multi-character strings, have to include it in recursive parsing */
18795 if (element_count) {
18796 bool has_l_bracket = orig_parse > RExC_start && *(orig_parse - 1) == '[';
18797
18798 sv_catpvs(substitute_parse, "|");
18799 if (has_l_bracket) { /* Add an [ if the original had one */
18800 sv_catpvs(substitute_parse, "[");
18801 }
18802 constructed_prefix_len = SvCUR(substitute_parse);
18803 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
18804
18805 /* Put in a closing ']' to match any opening one, but not if going
18806 * off the end, as otherwise we are adding something that really
18807 * isn't there */
18808 if (has_l_bracket && RExC_parse < RExC_end) {
18809 sv_catpvs(substitute_parse, "]");
18810 }
18811 }
18812
18813 sv_catpvs(substitute_parse, ")");
18814 #if 0
18815 if (invert) {
18816 /* This is a way to get the parse to skip forward a whole named
18817 * sequence instead of matching the 2nd character when it fails the
18818 * first */
18819 sv_catpvs(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
18820 }
18821 #endif
18822
18823 /* Set up the data structure so that any errors will be properly
18824 * reported. See the comments at the definition of
18825 * REPORT_LOCATION_ARGS for details */
18826 RExC_copy_start_in_input = (char *) orig_parse;
18827 RExC_start = RExC_parse = SvPV(substitute_parse, len);
18828 RExC_copy_start_in_constructed = RExC_start + constructed_prefix_len;
18829 RExC_end = RExC_parse + len;
18830 RExC_in_multi_char_class = 1;
18831
18832 ret = reg(pRExC_state, 1, ®_flags, depth+1);
18833
18834 *flagp |= reg_flags & (HASWIDTH|SIMPLE|POSTPONED|RESTART_PARSE|NEED_UTF8);
18835
18836 /* And restore so can parse the rest of the pattern */
18837 RExC_parse = save_parse;
18838 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = save_start;
18839 RExC_end = save_end;
18840 RExC_in_multi_char_class = 0;
18841 SvREFCNT_dec_NN(multi_char_matches);
18842 SvREFCNT_dec(properties);
18843 SvREFCNT_dec(cp_list);
18844 SvREFCNT_dec(simple_posixes);
18845 SvREFCNT_dec(posixes);
18846 SvREFCNT_dec(nposixes);
18847 SvREFCNT_dec(cp_foldable_list);
18848 return ret;
18849 }
18850
18851 /* If folding, we calculate all characters that could fold to or from the
18852 * ones already on the list */
18853 if (cp_foldable_list) {
18854 if (FOLD) {
18855 UV start, end; /* End points of code point ranges */
18856
18857 SV* fold_intersection = NULL;
18858 SV** use_list;
18859
18860 /* Our calculated list will be for Unicode rules. For locale
18861 * matching, we have to keep a separate list that is consulted at
18862 * runtime only when the locale indicates Unicode rules (and we
18863 * don't include potential matches in the ASCII/Latin1 range, as
18864 * any code point could fold to any other, based on the run-time
18865 * locale). For non-locale, we just use the general list */
18866 if (LOC) {
18867 use_list = &only_utf8_locale_list;
18868 }
18869 else {
18870 use_list = &cp_list;
18871 }
18872
18873 /* Only the characters in this class that participate in folds need
18874 * be checked. Get the intersection of this class and all the
18875 * possible characters that are foldable. This can quickly narrow
18876 * down a large class */
18877 _invlist_intersection(PL_in_some_fold, cp_foldable_list,
18878 &fold_intersection);
18879
18880 /* Now look at the foldable characters in this class individually */
18881 invlist_iterinit(fold_intersection);
18882 while (invlist_iternext(fold_intersection, &start, &end)) {
18883 UV j;
18884 UV folded;
18885
18886 /* Look at every character in the range */
18887 for (j = start; j <= end; j++) {
18888 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
18889 STRLEN foldlen;
18890 unsigned int k;
18891 Size_t folds_count;
18892 U32 first_fold;
18893 const U32 * remaining_folds;
18894
18895 if (j < 256) {
18896
18897 /* Under /l, we don't know what code points below 256
18898 * fold to, except we do know the MICRO SIGN folds to
18899 * an above-255 character if the locale is UTF-8, so we
18900 * add it to the special list (in *use_list) Otherwise
18901 * we know now what things can match, though some folds
18902 * are valid under /d only if the target is UTF-8.
18903 * Those go in a separate list */
18904 if ( IS_IN_SOME_FOLD_L1(j)
18905 && ! (LOC && j != MICRO_SIGN))
18906 {
18907
18908 /* ASCII is always matched; non-ASCII is matched
18909 * only under Unicode rules (which could happen
18910 * under /l if the locale is a UTF-8 one */
18911 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
18912 *use_list = add_cp_to_invlist(*use_list,
18913 PL_fold_latin1[j]);
18914 }
18915 else if (j != PL_fold_latin1[j]) {
18916 upper_latin1_only_utf8_matches
18917 = add_cp_to_invlist(
18918 upper_latin1_only_utf8_matches,
18919 PL_fold_latin1[j]);
18920 }
18921 }
18922
18923 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
18924 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
18925 {
18926 add_above_Latin1_folds(pRExC_state,
18927 (U8) j,
18928 use_list);
18929 }
18930 continue;
18931 }
18932
18933 /* Here is an above Latin1 character. We don't have the
18934 * rules hard-coded for it. First, get its fold. This is
18935 * the simple fold, as the multi-character folds have been
18936 * handled earlier and separated out */
18937 folded = _to_uni_fold_flags(j, foldbuf, &foldlen,
18938 (ASCII_FOLD_RESTRICTED)
18939 ? FOLD_FLAGS_NOMIX_ASCII
18940 : 0);
18941
18942 /* Single character fold of above Latin1. Add everything
18943 * in its fold closure to the list that this node should
18944 * match. */
18945 folds_count = _inverse_folds(folded, &first_fold,
18946 &remaining_folds);
18947 for (k = 0; k <= folds_count; k++) {
18948 UV c = (k == 0) /* First time through use itself */
18949 ? folded
18950 : (k == 1) /* 2nd time use, the first fold */
18951 ? first_fold
18952
18953 /* Then the remaining ones */
18954 : remaining_folds[k-2];
18955
18956 /* /aa doesn't allow folds between ASCII and non- */
18957 if (( ASCII_FOLD_RESTRICTED
18958 && (isASCII(c) != isASCII(j))))
18959 {
18960 continue;
18961 }
18962
18963 /* Folds under /l which cross the 255/256 boundary are
18964 * added to a separate list. (These are valid only
18965 * when the locale is UTF-8.) */
18966 if (c < 256 && LOC) {
18967 *use_list = add_cp_to_invlist(*use_list, c);
18968 continue;
18969 }
18970
18971 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
18972 {
18973 cp_list = add_cp_to_invlist(cp_list, c);
18974 }
18975 else {
18976 /* Similarly folds involving non-ascii Latin1
18977 * characters under /d are added to their list */
18978 upper_latin1_only_utf8_matches
18979 = add_cp_to_invlist(
18980 upper_latin1_only_utf8_matches,
18981 c);
18982 }
18983 }
18984 }
18985 }
18986 SvREFCNT_dec_NN(fold_intersection);
18987 }
18988
18989 /* Now that we have finished adding all the folds, there is no reason
18990 * to keep the foldable list separate */
18991 _invlist_union(cp_list, cp_foldable_list, &cp_list);
18992 SvREFCNT_dec_NN(cp_foldable_list);
18993 }
18994
18995 /* And combine the result (if any) with any inversion lists from posix
18996 * classes. The lists are kept separate up to now because we don't want to
18997 * fold the classes */
18998 if (simple_posixes) { /* These are the classes known to be unaffected by
18999 /a, /aa, and /d */
19000 if (cp_list) {
19001 _invlist_union(cp_list, simple_posixes, &cp_list);
19002 SvREFCNT_dec_NN(simple_posixes);
19003 }
19004 else {
19005 cp_list = simple_posixes;
19006 }
19007 }
19008 if (posixes || nposixes) {
19009 if (! DEPENDS_SEMANTICS) {
19010
19011 /* For everything but /d, we can just add the current 'posixes' and
19012 * 'nposixes' to the main list */
19013 if (posixes) {
19014 if (cp_list) {
19015 _invlist_union(cp_list, posixes, &cp_list);
19016 SvREFCNT_dec_NN(posixes);
19017 }
19018 else {
19019 cp_list = posixes;
19020 }
19021 }
19022 if (nposixes) {
19023 if (cp_list) {
19024 _invlist_union(cp_list, nposixes, &cp_list);
19025 SvREFCNT_dec_NN(nposixes);
19026 }
19027 else {
19028 cp_list = nposixes;
19029 }
19030 }
19031 }
19032 else {
19033 /* Under /d, things like \w match upper Latin1 characters only if
19034 * the target string is in UTF-8. But things like \W match all the
19035 * upper Latin1 characters if the target string is not in UTF-8.
19036 *
19037 * Handle the case with something like \W separately */
19038 if (nposixes) {
19039 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1, NULL);
19040
19041 /* A complemented posix class matches all upper Latin1
19042 * characters if not in UTF-8. And it matches just certain
19043 * ones when in UTF-8. That means those certain ones are
19044 * matched regardless, so can just be added to the
19045 * unconditional list */
19046 if (cp_list) {
19047 _invlist_union(cp_list, nposixes, &cp_list);
19048 SvREFCNT_dec_NN(nposixes);
19049 nposixes = NULL;
19050 }
19051 else {
19052 cp_list = nposixes;
19053 }
19054
19055 /* Likewise for 'posixes' */
19056 _invlist_union(posixes, cp_list, &cp_list);
19057 SvREFCNT_dec(posixes);
19058
19059 /* Likewise for anything else in the range that matched only
19060 * under UTF-8 */
19061 if (upper_latin1_only_utf8_matches) {
19062 _invlist_union(cp_list,
19063 upper_latin1_only_utf8_matches,
19064 &cp_list);
19065 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
19066 upper_latin1_only_utf8_matches = NULL;
19067 }
19068
19069 /* If we don't match all the upper Latin1 characters regardless
19070 * of UTF-8ness, we have to set a flag to match the rest when
19071 * not in UTF-8 */
19072 _invlist_subtract(only_non_utf8_list, cp_list,
19073 &only_non_utf8_list);
19074 if (_invlist_len(only_non_utf8_list) != 0) {
19075 anyof_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
19076 }
19077 SvREFCNT_dec_NN(only_non_utf8_list);
19078 }
19079 else {
19080 /* Here there were no complemented posix classes. That means
19081 * the upper Latin1 characters in 'posixes' match only when the
19082 * target string is in UTF-8. So we have to add them to the
19083 * list of those types of code points, while adding the
19084 * remainder to the unconditional list.
19085 *
19086 * First calculate what they are */
19087 SV* nonascii_but_latin1_properties = NULL;
19088 _invlist_intersection(posixes, PL_UpperLatin1,
19089 &nonascii_but_latin1_properties);
19090
19091 /* And add them to the final list of such characters. */
19092 _invlist_union(upper_latin1_only_utf8_matches,
19093 nonascii_but_latin1_properties,
19094 &upper_latin1_only_utf8_matches);
19095
19096 /* Remove them from what now becomes the unconditional list */
19097 _invlist_subtract(posixes, nonascii_but_latin1_properties,
19098 &posixes);
19099
19100 /* And add those unconditional ones to the final list */
19101 if (cp_list) {
19102 _invlist_union(cp_list, posixes, &cp_list);
19103 SvREFCNT_dec_NN(posixes);
19104 posixes = NULL;
19105 }
19106 else {
19107 cp_list = posixes;
19108 }
19109
19110 SvREFCNT_dec(nonascii_but_latin1_properties);
19111
19112 /* Get rid of any characters from the conditional list that we
19113 * now know are matched unconditionally, which may make that
19114 * list empty */
19115 _invlist_subtract(upper_latin1_only_utf8_matches,
19116 cp_list,
19117 &upper_latin1_only_utf8_matches);
19118 if (_invlist_len(upper_latin1_only_utf8_matches) == 0) {
19119 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
19120 upper_latin1_only_utf8_matches = NULL;
19121 }
19122 }
19123 }
19124 }
19125
19126 /* And combine the result (if any) with any inversion list from properties.
19127 * The lists are kept separate up to now so that we can distinguish the two
19128 * in regards to matching above-Unicode. A run-time warning is generated
19129 * if a Unicode property is matched against a non-Unicode code point. But,
19130 * we allow user-defined properties to match anything, without any warning,
19131 * and we also suppress the warning if there is a portion of the character
19132 * class that isn't a Unicode property, and which matches above Unicode, \W
19133 * or [\x{110000}] for example.
19134 * (Note that in this case, unlike the Posix one above, there is no
19135 * <upper_latin1_only_utf8_matches>, because having a Unicode property
19136 * forces Unicode semantics */
19137 if (properties) {
19138 if (cp_list) {
19139
19140 /* If it matters to the final outcome, see if a non-property
19141 * component of the class matches above Unicode. If so, the
19142 * warning gets suppressed. This is true even if just a single
19143 * such code point is specified, as, though not strictly correct if
19144 * another such code point is matched against, the fact that they
19145 * are using above-Unicode code points indicates they should know
19146 * the issues involved */
19147 if (warn_super) {
19148 warn_super = ! (invert
19149 ^ (UNICODE_IS_SUPER(invlist_highest(cp_list))));
19150 }
19151
19152 _invlist_union(properties, cp_list, &cp_list);
19153 SvREFCNT_dec_NN(properties);
19154 }
19155 else {
19156 cp_list = properties;
19157 }
19158
19159 if (warn_super) {
19160 anyof_flags
19161 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
19162
19163 /* Because an ANYOF node is the only one that warns, this node
19164 * can't be optimized into something else */
19165 optimizable = FALSE;
19166 }
19167 }
19168
19169 /* Here, we have calculated what code points should be in the character
19170 * class.
19171 *
19172 * Now we can see about various optimizations. Fold calculation (which we
19173 * did above) needs to take place before inversion. Otherwise /[^k]/i
19174 * would invert to include K, which under /i would match k, which it
19175 * shouldn't. Therefore we can't invert folded locale now, as it won't be
19176 * folded until runtime */
19177
19178 /* If we didn't do folding, it's because some information isn't available
19179 * until runtime; set the run-time fold flag for these We know to set the
19180 * flag if we have a non-NULL list for UTF-8 locales, or the class matches
19181 * at least one 0-255 range code point */
19182 if (LOC && FOLD) {
19183
19184 /* Some things on the list might be unconditionally included because of
19185 * other components. Remove them, and clean up the list if it goes to
19186 * 0 elements */
19187 if (only_utf8_locale_list && cp_list) {
19188 _invlist_subtract(only_utf8_locale_list, cp_list,
19189 &only_utf8_locale_list);
19190
19191 if (_invlist_len(only_utf8_locale_list) == 0) {
19192 SvREFCNT_dec_NN(only_utf8_locale_list);
19193 only_utf8_locale_list = NULL;
19194 }
19195 }
19196 if ( only_utf8_locale_list
19197 || (cp_list && ( _invlist_contains_cp(cp_list, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE)
19198 || _invlist_contains_cp(cp_list, LATIN_SMALL_LETTER_DOTLESS_I))))
19199 {
19200 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
19201 anyof_flags
19202 |= ANYOFL_FOLD
19203 | ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
19204 }
19205 else if (cp_list && invlist_lowest(cp_list) < 256) {
19206 /* If nothing is below 256, has no locale dependency; otherwise it
19207 * does */
19208 anyof_flags |= ANYOFL_FOLD;
19209 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
19210 }
19211 }
19212 else if ( DEPENDS_SEMANTICS
19213 && ( upper_latin1_only_utf8_matches
19214 || (anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
19215 {
19216 RExC_seen_d_op = TRUE;
19217 has_runtime_dependency |= HAS_D_RUNTIME_DEPENDENCY;
19218 }
19219
19220 /* Optimize inverted patterns (e.g. [^a-z]) when everything is known at
19221 * compile time. */
19222 if ( cp_list
19223 && invert
19224 && ! has_runtime_dependency)
19225 {
19226 _invlist_invert(cp_list);
19227
19228 /* Clear the invert flag since have just done it here */
19229 invert = FALSE;
19230 }
19231
19232 /* All possible optimizations below still have these characteristics.
19233 * (Multi-char folds aren't SIMPLE, but they don't get this far in this
19234 * routine) */
19235 *flagp |= HASWIDTH|SIMPLE;
19236
19237 if (ret_invlist) {
19238 *ret_invlist = cp_list;
19239
19240 return (cp_list) ? RExC_emit : 0;
19241 }
19242
19243 if (anyof_flags & ANYOF_LOCALE_FLAGS) {
19244 RExC_contains_locale = 1;
19245 }
19246
19247 if (optimizable) {
19248
19249 /* Some character classes are equivalent to other nodes. Such nodes
19250 * take up less room, and some nodes require fewer operations to
19251 * execute, than ANYOF nodes. EXACTish nodes may be joinable with
19252 * adjacent nodes to improve efficiency. */
19253 op = optimize_regclass(pRExC_state, cp_list,
19254 only_utf8_locale_list,
19255 upper_latin1_only_utf8_matches,
19256 has_runtime_dependency,
19257 posixl,
19258 &anyof_flags, &invert, &ret, flagp);
19259 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
19260
19261 /* If optimized to something else and emitted, clean up and return */
19262 if (ret >= 0) {
19263 Set_Node_Offset_Length(REGNODE_p(ret), orig_parse - RExC_start,
19264 RExC_parse - orig_parse);;
19265 SvREFCNT_dec(cp_list);;
19266 SvREFCNT_dec(only_utf8_locale_list);
19267 SvREFCNT_dec(upper_latin1_only_utf8_matches);
19268 return ret;
19269 }
19270
19271 /* If no optimization was found, an END was returned and we will now
19272 * emit an ANYOF */
19273 if (op == END) {
19274 op = ANYOF;
19275 }
19276 }
19277
19278 /* Here are going to emit an ANYOF; set the particular type */
19279 if (op == ANYOF) {
19280 if (has_runtime_dependency & HAS_D_RUNTIME_DEPENDENCY) {
19281 op = ANYOFD;
19282 }
19283 else if (posixl) {
19284 op = ANYOFPOSIXL;
19285 }
19286 else if (LOC) {
19287 op = ANYOFL;
19288 }
19289 }
19290
19291 ret = regnode_guts(pRExC_state, op, regarglen[op], "anyof");
19292 FILL_NODE(ret, op); /* We set the argument later */
19293 RExC_emit += 1 + regarglen[op];
19294 ANYOF_FLAGS(REGNODE_p(ret)) = anyof_flags;
19295
19296 /* Here, <cp_list> contains all the code points we can determine at
19297 * compile time that match under all conditions. Go through it, and
19298 * for things that belong in the bitmap, put them there, and delete from
19299 * <cp_list>. While we are at it, see if everything above 255 is in the
19300 * list, and if so, set a flag to speed up execution */
19301
19302 populate_ANYOF_from_invlist(REGNODE_p(ret), &cp_list);
19303
19304 if (posixl) {
19305 ANYOF_POSIXL_SET_TO_BITMAP(REGNODE_p(ret), posixl);
19306 }
19307
19308 if (invert) {
19309 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_INVERT;
19310 }
19311
19312 /* Here, the bitmap has been populated with all the Latin1 code points that
19313 * always match. Can now add to the overall list those that match only
19314 * when the target string is UTF-8 (<upper_latin1_only_utf8_matches>).
19315 * */
19316 if (upper_latin1_only_utf8_matches) {
19317 if (cp_list) {
19318 _invlist_union(cp_list,
19319 upper_latin1_only_utf8_matches,
19320 &cp_list);
19321 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
19322 }
19323 else {
19324 cp_list = upper_latin1_only_utf8_matches;
19325 }
19326 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
19327 }
19328
19329 set_ANYOF_arg(pRExC_state, REGNODE_p(ret), cp_list,
19330 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
19331 ? listsv
19332 : NULL,
19333 only_utf8_locale_list);
19334
19335 SvREFCNT_dec(cp_list);;
19336 SvREFCNT_dec(only_utf8_locale_list);
19337 return ret;
19338 }
19339
19340 STATIC U8
S_optimize_regclass(pTHX_ RExC_state_t * pRExC_state,SV * cp_list,SV * only_utf8_locale_list,SV * upper_latin1_only_utf8_matches,const U32 has_runtime_dependency,const U32 posixl,U8 * anyof_flags,bool * invert,regnode_offset * ret,I32 * flagp)19341 S_optimize_regclass(pTHX_
19342 RExC_state_t *pRExC_state,
19343 SV * cp_list,
19344 SV* only_utf8_locale_list,
19345 SV* upper_latin1_only_utf8_matches,
19346 const U32 has_runtime_dependency,
19347 const U32 posixl,
19348 U8 * anyof_flags,
19349 bool * invert,
19350 regnode_offset * ret,
19351 I32 *flagp
19352 )
19353 {
19354 /* This function exists just to make S_regclass() smaller. It extracts out
19355 * the code that looks for potential optimizations away from a full generic
19356 * ANYOF node. The parameter names are the same as the corresponding
19357 * variables in S_regclass.
19358 *
19359 * It returns the new op (the impossible END one if no optimization found)
19360 * and sets *ret to any created regnode. If the new op is sufficiently
19361 * like plain ANYOF, it leaves *ret unchanged for allocation in S_regclass.
19362 *
19363 * Certain of the parameters may be updated as a result of the changes
19364 * herein */
19365
19366 U8 op = END; /* The returned node-type, initialized to an impossible
19367 one. */
19368 UV value = 0;
19369 PERL_UINT_FAST8_T i;
19370 UV partial_cp_count = 0;
19371 UV start[MAX_FOLD_FROMS+1] = { 0 }; /* +1 for the folded-to char */
19372 UV end[MAX_FOLD_FROMS+1] = { 0 };
19373 bool single_range = FALSE;
19374 UV lowest_cp = 0, highest_cp = 0;
19375
19376 PERL_ARGS_ASSERT_OPTIMIZE_REGCLASS;
19377
19378 if (cp_list) { /* Count the code points in enough ranges that we would see
19379 all the ones possible in any fold in this version of
19380 Unicode */
19381
19382 invlist_iterinit(cp_list);
19383 for (i = 0; i <= MAX_FOLD_FROMS; i++) {
19384 if (! invlist_iternext(cp_list, &start[i], &end[i])) {
19385 break;
19386 }
19387 partial_cp_count += end[i] - start[i] + 1;
19388 }
19389
19390 if (i == 1) {
19391 single_range = TRUE;
19392 }
19393 invlist_iterfinish(cp_list);
19394
19395 /* If we know at compile time that this matches every possible code
19396 * point, any run-time dependencies don't matter */
19397 if (start[0] == 0 && end[0] == UV_MAX) {
19398 if (*invert) {
19399 goto return_OPFAIL;
19400 }
19401 else {
19402 goto return_SANY;
19403 }
19404 }
19405
19406 /* Use a clearer mnemonic for below */
19407 lowest_cp = start[0];
19408
19409 highest_cp = invlist_highest(cp_list);
19410 }
19411
19412 /* Similarly, for /l posix classes, if both a class and its complement
19413 * match, any run-time dependencies don't matter */
19414 if (posixl) {
19415 int namedclass;
19416 for (namedclass = 0; namedclass < ANYOF_POSIXL_MAX; namedclass += 2) {
19417 if ( POSIXL_TEST(posixl, namedclass) /* class */
19418 && POSIXL_TEST(posixl, namedclass + 1)) /* its complement */
19419 {
19420 if (*invert) {
19421 goto return_OPFAIL;
19422 }
19423 else {
19424 goto return_SANY;
19425 }
19426 return op;
19427 }
19428 }
19429
19430 /* For well-behaved locales, some classes are subsets of others, so
19431 * complementing the subset and including the non-complemented superset
19432 * should match everything, like [\D[:alnum:]], and
19433 * [[:^alpha:][:alnum:]], but some implementations of locales are
19434 * buggy, and khw thinks its a bad idea to have optimization change
19435 * behavior, even if it avoids an OS bug in a given case */
19436
19437 #define isSINGLE_BIT_SET(n) isPOWER_OF_2(n)
19438
19439 /* If is a single posix /l class, can optimize to just that op. Such a
19440 * node will not match anything in the Latin1 range, as that is not
19441 * determinable until runtime, but will match whatever the class does
19442 * outside that range. (Note that some classes won't match anything
19443 * outside the range, like [:ascii:]) */
19444 if ( isSINGLE_BIT_SET(posixl)
19445 && (partial_cp_count == 0 || lowest_cp > 255))
19446 {
19447 U8 classnum;
19448 SV * class_above_latin1 = NULL;
19449 bool already_inverted;
19450 bool are_equivalent;
19451
19452
19453 namedclass = single_1bit_pos32(posixl);
19454 classnum = namedclass_to_classnum(namedclass);
19455
19456 /* The named classes are such that the inverted number is one
19457 * larger than the non-inverted one */
19458 already_inverted = namedclass - classnum_to_namedclass(classnum);
19459
19460 /* Create an inversion list of the official property, inverted if
19461 * the constructed node list is inverted, and restricted to only
19462 * the above latin1 code points, which are the only ones known at
19463 * compile time */
19464 _invlist_intersection_maybe_complement_2nd(
19465 PL_AboveLatin1,
19466 PL_XPosix_ptrs[classnum],
19467 already_inverted,
19468 &class_above_latin1);
19469 are_equivalent = _invlistEQ(class_above_latin1, cp_list, FALSE);
19470 SvREFCNT_dec_NN(class_above_latin1);
19471
19472 if (are_equivalent) {
19473
19474 /* Resolve the run-time inversion flag with this possibly
19475 * inverted class */
19476 *invert = *invert ^ already_inverted;
19477
19478 op = POSIXL + *invert * (NPOSIXL - POSIXL);
19479 *ret = reg_node(pRExC_state, op);
19480 FLAGS(REGNODE_p(*ret)) = classnum;
19481 return op;
19482 }
19483 }
19484 }
19485
19486 /* khw can't think of any other possible transformation involving these. */
19487 if (has_runtime_dependency & HAS_USER_DEFINED_PROPERTY) {
19488 return END;
19489 }
19490
19491 if (! has_runtime_dependency) {
19492
19493 /* If the list is empty, nothing matches. This happens, for example,
19494 * when a Unicode property that doesn't match anything is the only
19495 * element in the character class (perluniprops.pod notes such
19496 * properties). */
19497 if (partial_cp_count == 0) {
19498 if (*invert) {
19499 goto return_SANY;
19500 }
19501 else {
19502 goto return_OPFAIL;
19503 }
19504 }
19505
19506 /* If matches everything but \n */
19507 if ( start[0] == 0 && end[0] == '\n' - 1
19508 && start[1] == '\n' + 1 && end[1] == UV_MAX)
19509 {
19510 assert (! *invert);
19511 op = REG_ANY;
19512 *ret = reg_node(pRExC_state, op);
19513 MARK_NAUGHTY(1);
19514 return op;
19515 }
19516 }
19517
19518 /* Next see if can optimize classes that contain just a few code points
19519 * into an EXACTish node. The reason to do this is to let the optimizer
19520 * join this node with adjacent EXACTish ones, and ANYOF nodes require
19521 * runtime conversion to code point from UTF-8, which we'd like to avoid.
19522 *
19523 * An EXACTFish node can be generated even if not under /i, and vice versa.
19524 * But care must be taken. An EXACTFish node has to be such that it only
19525 * matches precisely the code points in the class, but we want to generate
19526 * the least restrictive one that does that, to increase the odds of being
19527 * able to join with an adjacent node. For example, if the class contains
19528 * [kK], we have to make it an EXACTFAA node to prevent the KELVIN SIGN
19529 * from matching. Whether we are under /i or not is irrelevant in this
19530 * case. Less obvious is the pattern qr/[\x{02BC}]n/i. U+02BC is MODIFIER
19531 * LETTER APOSTROPHE. That is supposed to match the single character U+0149
19532 * LATIN SMALL LETTER N PRECEDED BY APOSTROPHE. And so even though there
19533 * is no simple fold that includes \X{02BC}, there is a multi-char fold
19534 * that does, and so the node generated for it must be an EXACTFish one.
19535 * On the other hand qr/:/i should generate a plain EXACT node since the
19536 * colon participates in no fold whatsoever, and having it be EXACT tells
19537 * the optimizer the target string cannot match unless it has a colon in
19538 * it. */
19539 if ( ! posixl
19540 && ! *invert
19541
19542 /* Only try if there are no more code points in the class than in
19543 * the max possible fold */
19544 && inRANGE(partial_cp_count, 1, MAX_FOLD_FROMS + 1))
19545 {
19546 /* We can always make a single code point class into an EXACTish node.
19547 * */
19548 if (partial_cp_count == 1 && ! upper_latin1_only_utf8_matches) {
19549 if (LOC) {
19550
19551 /* Here is /l: Use EXACTL, except if there is a fold not known
19552 * until runtime so shows as only a single code point here.
19553 * For code points above 255, we know which can cause problems
19554 * by having a potential fold to the Latin1 range. */
19555 if ( ! FOLD
19556 || ( lowest_cp > 255
19557 && ! is_PROBLEMATIC_LOCALE_FOLD_cp(lowest_cp)))
19558 {
19559 op = EXACTL;
19560 }
19561 else {
19562 op = EXACTFL;
19563 }
19564 }
19565 else if (! FOLD) { /* Not /l and not /i */
19566 op = (lowest_cp < 256) ? EXACT : EXACT_REQ8;
19567 }
19568 else if (lowest_cp < 256) { /* /i, not /l, and the code point is
19569 small */
19570
19571 /* Under /i, it gets a little tricky. A code point that
19572 * doesn't participate in a fold should be an EXACT node. We
19573 * know this one isn't the result of a simple fold, or there'd
19574 * be more than one code point in the list, but it could be
19575 * part of a multi-character fold. In that case we better not
19576 * create an EXACT node, as we would wrongly be telling the
19577 * optimizer that this code point must be in the target string,
19578 * and that is wrong. This is because if the sequence around
19579 * this code point forms a multi-char fold, what needs to be in
19580 * the string could be the code point that folds to the
19581 * sequence.
19582 *
19583 * This handles the case of below-255 code points, as we have
19584 * an easy look up for those. The next clause handles the
19585 * above-256 one */
19586 op = IS_IN_SOME_FOLD_L1(lowest_cp)
19587 ? EXACTFU
19588 : EXACT;
19589 }
19590 else { /* /i, larger code point. Since we are under /i, and have
19591 just this code point, we know that it can't fold to
19592 something else, so PL_InMultiCharFold applies to it */
19593 op = (_invlist_contains_cp(PL_InMultiCharFold, lowest_cp))
19594 ? EXACTFU_REQ8
19595 : EXACT_REQ8;
19596 }
19597
19598 value = lowest_cp;
19599 }
19600 else if ( ! (has_runtime_dependency & ~HAS_D_RUNTIME_DEPENDENCY)
19601 && _invlist_contains_cp(PL_in_some_fold, lowest_cp))
19602 {
19603 /* Here, the only runtime dependency, if any, is from /d, and the
19604 * class matches more than one code point, and the lowest code
19605 * point participates in some fold. It might be that the other
19606 * code points are /i equivalent to this one, and hence they would
19607 * be representable by an EXACTFish node. Above, we eliminated
19608 * classes that contain too many code points to be EXACTFish, with
19609 * the test for MAX_FOLD_FROMS
19610 *
19611 * First, special case the ASCII fold pairs, like 'B' and 'b'. We
19612 * do this because we have EXACTFAA at our disposal for the ASCII
19613 * range */
19614 if (partial_cp_count == 2 && isASCII(lowest_cp)) {
19615
19616 /* The only ASCII characters that participate in folds are
19617 * alphabetics */
19618 assert(isALPHA(lowest_cp));
19619 if ( end[0] == start[0] /* First range is a single
19620 character, so 2nd exists */
19621 && isALPHA_FOLD_EQ(start[0], start[1]))
19622 {
19623 /* Here, is part of an ASCII fold pair */
19624
19625 if ( ASCII_FOLD_RESTRICTED
19626 || HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(lowest_cp))
19627 {
19628 /* If the second clause just above was true, it means
19629 * we can't be under /i, or else the list would have
19630 * included more than this fold pair. Therefore we
19631 * have to exclude the possibility of whatever else it
19632 * is that folds to these, by using EXACTFAA */
19633 op = EXACTFAA;
19634 }
19635 else if (HAS_NONLATIN1_FOLD_CLOSURE(lowest_cp)) {
19636
19637 /* Here, there's no simple fold that lowest_cp is part
19638 * of, but there is a multi-character one. If we are
19639 * not under /i, we want to exclude that possibility;
19640 * if under /i, we want to include it */
19641 op = (FOLD) ? EXACTFU : EXACTFAA;
19642 }
19643 else {
19644
19645 /* Here, the only possible fold lowest_cp particpates in
19646 * is with start[1]. /i or not isn't relevant */
19647 op = EXACTFU;
19648 }
19649
19650 value = toFOLD(lowest_cp);
19651 }
19652 }
19653 else if ( ! upper_latin1_only_utf8_matches
19654 || ( _invlist_len(upper_latin1_only_utf8_matches) == 2
19655 && PL_fold_latin1[
19656 invlist_highest(upper_latin1_only_utf8_matches)]
19657 == lowest_cp))
19658 {
19659 /* Here, the smallest character is non-ascii or there are more
19660 * than 2 code points matched by this node. Also, we either
19661 * don't have /d UTF-8 dependent matches, or if we do, they
19662 * look like they could be a single character that is the fold
19663 * of the lowest one is in the always-match list. This test
19664 * quickly excludes most of the false positives when there are
19665 * /d UTF-8 depdendent matches. These are like LATIN CAPITAL
19666 * LETTER A WITH GRAVE matching LATIN SMALL LETTER A WITH GRAVE
19667 * iff the target string is UTF-8. (We don't have to worry
19668 * above about exceeding the array bounds of PL_fold_latin1[]
19669 * because any code point in 'upper_latin1_only_utf8_matches'
19670 * is below 256.)
19671 *
19672 * EXACTFAA would apply only to pairs (hence exactly 2 code
19673 * points) in the ASCII range, so we can't use it here to
19674 * artificially restrict the fold domain, so we check if the
19675 * class does or does not match some EXACTFish node. Further,
19676 * if we aren't under /i, and and the folded-to character is
19677 * part of a multi-character fold, we can't do this
19678 * optimization, as the sequence around it could be that
19679 * multi-character fold, and we don't here know the context, so
19680 * we have to assume it is that multi-char fold, to prevent
19681 * potential bugs.
19682 *
19683 * To do the general case, we first find the fold of the lowest
19684 * code point (which may be higher than that lowest unfolded
19685 * one), then find everything that folds to it. (The data
19686 * structure we have only maps from the folded code points, so
19687 * we have to do the earlier step.) */
19688
19689 Size_t foldlen;
19690 U8 foldbuf[UTF8_MAXBYTES_CASE];
19691 UV folded = _to_uni_fold_flags(lowest_cp, foldbuf, &foldlen, 0);
19692 U32 first_fold;
19693 const U32 * remaining_folds;
19694 Size_t folds_to_this_cp_count = _inverse_folds(
19695 folded,
19696 &first_fold,
19697 &remaining_folds);
19698 Size_t folds_count = folds_to_this_cp_count + 1;
19699 SV * fold_list = _new_invlist(folds_count);
19700 unsigned int i;
19701
19702 /* If there are UTF-8 dependent matches, create a temporary
19703 * list of what this node matches, including them. */
19704 SV * all_cp_list = NULL;
19705 SV ** use_this_list = &cp_list;
19706
19707 if (upper_latin1_only_utf8_matches) {
19708 all_cp_list = _new_invlist(0);
19709 use_this_list = &all_cp_list;
19710 _invlist_union(cp_list,
19711 upper_latin1_only_utf8_matches,
19712 use_this_list);
19713 }
19714
19715 /* Having gotten everything that participates in the fold
19716 * containing the lowest code point, we turn that into an
19717 * inversion list, making sure everything is included. */
19718 fold_list = add_cp_to_invlist(fold_list, lowest_cp);
19719 fold_list = add_cp_to_invlist(fold_list, folded);
19720 if (folds_to_this_cp_count > 0) {
19721 fold_list = add_cp_to_invlist(fold_list, first_fold);
19722 for (i = 0; i + 1 < folds_to_this_cp_count; i++) {
19723 fold_list = add_cp_to_invlist(fold_list,
19724 remaining_folds[i]);
19725 }
19726 }
19727
19728 /* If the fold list is identical to what's in this ANYOF node,
19729 * the node can be represented by an EXACTFish one instead */
19730 if (_invlistEQ(*use_this_list, fold_list,
19731 0 /* Don't complement */ )
19732 ) {
19733
19734 /* But, we have to be careful, as mentioned above. Just
19735 * the right sequence of characters could match this if it
19736 * is part of a multi-character fold. That IS what we want
19737 * if we are under /i. But it ISN'T what we want if not
19738 * under /i, as it could match when it shouldn't. So, when
19739 * we aren't under /i and this character participates in a
19740 * multi-char fold, we don't optimize into an EXACTFish
19741 * node. So, for each case below we have to check if we
19742 * are folding, and if not, if it is not part of a
19743 * multi-char fold. */
19744 if (lowest_cp > 255) { /* Highish code point */
19745 if (FOLD || ! _invlist_contains_cp(
19746 PL_InMultiCharFold, folded))
19747 {
19748 op = (LOC)
19749 ? EXACTFLU8
19750 : (ASCII_FOLD_RESTRICTED)
19751 ? EXACTFAA
19752 : EXACTFU_REQ8;
19753 value = folded;
19754 }
19755 } /* Below, the lowest code point < 256 */
19756 else if ( FOLD
19757 && folded == 's'
19758 && DEPENDS_SEMANTICS)
19759 { /* An EXACTF node containing a single character 's',
19760 can be an EXACTFU if it doesn't get joined with an
19761 adjacent 's' */
19762 op = EXACTFU_S_EDGE;
19763 value = folded;
19764 }
19765 else if ( FOLD
19766 || ! HAS_NONLATIN1_FOLD_CLOSURE(lowest_cp))
19767 {
19768 if (upper_latin1_only_utf8_matches) {
19769 op = EXACTF;
19770
19771 /* We can't use the fold, as that only matches
19772 * under UTF-8 */
19773 value = lowest_cp;
19774 }
19775 else if ( UNLIKELY(lowest_cp == MICRO_SIGN)
19776 && ! UTF)
19777 { /* EXACTFUP is a special node for this character */
19778 op = (ASCII_FOLD_RESTRICTED)
19779 ? EXACTFAA
19780 : EXACTFUP;
19781 value = MICRO_SIGN;
19782 }
19783 else if ( ASCII_FOLD_RESTRICTED
19784 && ! isASCII(lowest_cp))
19785 { /* For ASCII under /iaa, we can use EXACTFU below
19786 */
19787 op = EXACTFAA;
19788 value = folded;
19789 }
19790 else {
19791 op = EXACTFU;
19792 value = folded;
19793 }
19794 }
19795 }
19796
19797 SvREFCNT_dec_NN(fold_list);
19798 SvREFCNT_dec(all_cp_list);
19799 }
19800 }
19801
19802 if (op != END) {
19803 U8 len;
19804
19805 /* Here, we have calculated what EXACTish node to use. Have to
19806 * convert to UTF-8 if not already there */
19807 if (value > 255) {
19808 if (! UTF) {
19809 SvREFCNT_dec(cp_list);;
19810 REQUIRE_UTF8(flagp);
19811 }
19812
19813 /* This is a kludge to the special casing issues with this
19814 * ligature under /aa. FB05 should fold to FB06, but the call
19815 * above to _to_uni_fold_flags() didn't find this, as it didn't
19816 * use the /aa restriction in order to not miss other folds
19817 * that would be affected. This is the only instance likely to
19818 * ever be a problem in all of Unicode. So special case it. */
19819 if ( value == LATIN_SMALL_LIGATURE_LONG_S_T
19820 && ASCII_FOLD_RESTRICTED)
19821 {
19822 value = LATIN_SMALL_LIGATURE_ST;
19823 }
19824 }
19825
19826 len = (UTF) ? UVCHR_SKIP(value) : 1;
19827
19828 *ret = regnode_guts(pRExC_state, op, len, "exact");
19829 FILL_NODE(*ret, op);
19830 RExC_emit += 1 + STR_SZ(len);
19831 setSTR_LEN(REGNODE_p(*ret), len);
19832 if (len == 1) {
19833 *STRINGs(REGNODE_p(*ret)) = (U8) value;
19834 }
19835 else {
19836 uvchr_to_utf8((U8 *) STRINGs(REGNODE_p(*ret)), value);
19837 }
19838 return op;
19839 }
19840 }
19841
19842 if (! has_runtime_dependency) {
19843
19844 /* See if this can be turned into an ANYOFM node. Think about the bit
19845 * patterns in two different bytes. In some positions, the bits in
19846 * each will be 1; and in other positions both will be 0; and in some
19847 * positions the bit will be 1 in one byte, and 0 in the other. Let
19848 * 'n' be the number of positions where the bits differ. We create a
19849 * mask which has exactly 'n' 0 bits, each in a position where the two
19850 * bytes differ. Now take the set of all bytes that when ANDed with
19851 * the mask yield the same result. That set has 2**n elements, and is
19852 * representable by just two 8 bit numbers: the result and the mask.
19853 * Importantly, matching the set can be vectorized by creating a word
19854 * full of the result bytes, and a word full of the mask bytes,
19855 * yielding a significant speed up. Here, see if this node matches
19856 * such a set. As a concrete example consider [01], and the byte
19857 * representing '0' which is 0x30 on ASCII machines. It has the bits
19858 * 0011 0000. Take the mask 1111 1110. If we AND 0x31 and 0x30 with
19859 * that mask we get 0x30. Any other bytes ANDed yield something else.
19860 * So [01], which is a common usage, is optimizable into ANYOFM, and
19861 * can benefit from the speed up. We can only do this on UTF-8
19862 * invariant bytes, because they have the same bit patterns under UTF-8
19863 * as not. */
19864 PERL_UINT_FAST8_T inverted = 0;
19865
19866 /* Highest possible UTF-8 invariant is 7F on ASCII platforms; FF on
19867 * EBCDIC */
19868 const PERL_UINT_FAST8_T max_permissible
19869 = nBIT_UMAX(7 + ONE_IF_EBCDIC_ZERO_IF_NOT);
19870
19871 /* If doesn't fit the criteria for ANYOFM, invert and try again. If
19872 * that works we will instead later generate an NANYOFM, and invert
19873 * back when through */
19874 if (highest_cp > max_permissible) {
19875 _invlist_invert(cp_list);
19876 inverted = 1;
19877 }
19878
19879 if (invlist_highest(cp_list) <= max_permissible) {
19880 UV this_start, this_end;
19881 UV lowest_cp = UV_MAX; /* init'ed to suppress compiler warn */
19882 U8 bits_differing = 0;
19883 Size_t full_cp_count = 0;
19884 bool first_time = TRUE;
19885
19886 /* Go through the bytes and find the bit positions that differ */
19887 invlist_iterinit(cp_list);
19888 while (invlist_iternext(cp_list, &this_start, &this_end)) {
19889 unsigned int i = this_start;
19890
19891 if (first_time) {
19892 if (! UVCHR_IS_INVARIANT(i)) {
19893 goto done_anyofm;
19894 }
19895
19896 first_time = FALSE;
19897 lowest_cp = this_start;
19898
19899 /* We have set up the code point to compare with. Don't
19900 * compare it with itself */
19901 i++;
19902 }
19903
19904 /* Find the bit positions that differ from the lowest code
19905 * point in the node. Keep track of all such positions by
19906 * OR'ing */
19907 for (; i <= this_end; i++) {
19908 if (! UVCHR_IS_INVARIANT(i)) {
19909 goto done_anyofm;
19910 }
19911
19912 bits_differing |= i ^ lowest_cp;
19913 }
19914
19915 full_cp_count += this_end - this_start + 1;
19916 }
19917
19918 /* At the end of the loop, we count how many bits differ from the
19919 * bits in lowest code point, call the count 'd'. If the set we
19920 * found contains 2**d elements, it is the closure of all code
19921 * points that differ only in those bit positions. To convince
19922 * yourself of that, first note that the number in the closure must
19923 * be a power of 2, which we test for. The only way we could have
19924 * that count and it be some differing set, is if we got some code
19925 * points that don't differ from the lowest code point in any
19926 * position, but do differ from each other in some other position.
19927 * That means one code point has a 1 in that position, and another
19928 * has a 0. But that would mean that one of them differs from the
19929 * lowest code point in that position, which possibility we've
19930 * already excluded. */
19931 if ( (inverted || full_cp_count > 1)
19932 && full_cp_count == 1U << PL_bitcount[bits_differing])
19933 {
19934 U8 ANYOFM_mask;
19935
19936 op = ANYOFM + inverted;;
19937
19938 /* We need to make the bits that differ be 0's */
19939 ANYOFM_mask = ~ bits_differing; /* This goes into FLAGS */
19940
19941 /* The argument is the lowest code point */
19942 *ret = reganode(pRExC_state, op, lowest_cp);
19943 FLAGS(REGNODE_p(*ret)) = ANYOFM_mask;
19944 }
19945
19946 done_anyofm:
19947 invlist_iterfinish(cp_list);
19948 }
19949
19950 if (inverted) {
19951 _invlist_invert(cp_list);
19952 }
19953
19954 if (op != END) {
19955 return op;
19956 }
19957
19958 /* XXX We could create an ANYOFR_LOW node here if we saved above if all
19959 * were invariants, it wasn't inverted, and there is a single range.
19960 * This would be faster than some of the posix nodes we create below
19961 * like /\d/a, but would be twice the size. Without having actually
19962 * measured the gain, khw doesn't think the tradeoff is really worth it
19963 * */
19964 }
19965
19966 if (! (*anyof_flags & ANYOF_LOCALE_FLAGS)) {
19967 PERL_UINT_FAST8_T type;
19968 SV * intersection = NULL;
19969 SV* d_invlist = NULL;
19970
19971 /* See if this matches any of the POSIX classes. The POSIXA and POSIXD
19972 * ones are about the same speed as ANYOF ops, but take less room; the
19973 * ones that have above-Latin1 code point matches are somewhat faster
19974 * than ANYOF. */
19975
19976 for (type = POSIXA; type >= POSIXD; type--) {
19977 int posix_class;
19978
19979 if (type == POSIXL) { /* But not /l posix classes */
19980 continue;
19981 }
19982
19983 for (posix_class = 0;
19984 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
19985 posix_class++)
19986 {
19987 SV** our_code_points = &cp_list;
19988 SV** official_code_points;
19989 int try_inverted;
19990
19991 if (type == POSIXA) {
19992 official_code_points = &PL_Posix_ptrs[posix_class];
19993 }
19994 else {
19995 official_code_points = &PL_XPosix_ptrs[posix_class];
19996 }
19997
19998 /* Skip non-existent classes of this type. e.g. \v only has an
19999 * entry in PL_XPosix_ptrs */
20000 if (! *official_code_points) {
20001 continue;
20002 }
20003
20004 /* Try both the regular class, and its inversion */
20005 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
20006 bool this_inverted = *invert ^ try_inverted;
20007
20008 if (type != POSIXD) {
20009
20010 /* This class that isn't /d can't match if we have /d
20011 * dependencies */
20012 if (has_runtime_dependency
20013 & HAS_D_RUNTIME_DEPENDENCY)
20014 {
20015 continue;
20016 }
20017 }
20018 else /* is /d */ if (! this_inverted) {
20019
20020 /* /d classes don't match anything non-ASCII below 256
20021 * unconditionally (which cp_list contains) */
20022 _invlist_intersection(cp_list, PL_UpperLatin1,
20023 &intersection);
20024 if (_invlist_len(intersection) != 0) {
20025 continue;
20026 }
20027
20028 SvREFCNT_dec(d_invlist);
20029 d_invlist = invlist_clone(cp_list, NULL);
20030
20031 /* But under UTF-8 it turns into using /u rules. Add
20032 * the things it matches under these conditions so that
20033 * we check below that these are identical to what the
20034 * tested class should match */
20035 if (upper_latin1_only_utf8_matches) {
20036 _invlist_union(
20037 d_invlist,
20038 upper_latin1_only_utf8_matches,
20039 &d_invlist);
20040 }
20041 our_code_points = &d_invlist;
20042 }
20043 else { /* POSIXD, inverted. If this doesn't have this
20044 flag set, it isn't /d. */
20045 if (! (*anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
20046 {
20047 continue;
20048 }
20049 our_code_points = &cp_list;
20050 }
20051
20052 /* Here, have weeded out some things. We want to see if
20053 * the list of characters this node contains
20054 * ('*our_code_points') precisely matches those of the
20055 * class we are currently checking against
20056 * ('*official_code_points'). */
20057 if (_invlistEQ(*our_code_points,
20058 *official_code_points,
20059 try_inverted))
20060 {
20061 /* Here, they precisely match. Optimize this ANYOF
20062 * node into its equivalent POSIX one of the correct
20063 * type, possibly inverted */
20064 op = (try_inverted)
20065 ? type + NPOSIXA - POSIXA
20066 : type;
20067 *ret = reg_node(pRExC_state, op);
20068 FLAGS(REGNODE_p(*ret)) = posix_class;
20069 SvREFCNT_dec(d_invlist);
20070 SvREFCNT_dec(intersection);
20071 return op;
20072 }
20073 }
20074 }
20075 }
20076 SvREFCNT_dec(d_invlist);
20077 SvREFCNT_dec(intersection);
20078 }
20079
20080 /* If it is a single contiguous range, ANYOFR is an efficient regnode, both
20081 * in size and speed. Currently, a 20 bit range base (smallest code point
20082 * in the range), and a 12 bit maximum delta are packed into a 32 bit word.
20083 * This allows for using it on all of the Unicode code points except for
20084 * the highest plane, which is only for private use code points. khw
20085 * doubts that a bigger delta is likely in real world applications */
20086 if ( single_range
20087 && ! has_runtime_dependency
20088 && *anyof_flags == 0
20089 && start[0] < (1 << ANYOFR_BASE_BITS)
20090 && end[0] - start[0]
20091 < ((1U << (sizeof(((struct regnode_1 *)NULL)->arg1)
20092 * CHARBITS - ANYOFR_BASE_BITS))))
20093
20094 {
20095 U8 low_utf8[UTF8_MAXBYTES+1];
20096 U8 high_utf8[UTF8_MAXBYTES+1];
20097
20098 op = ANYOFR;
20099 *ret = reganode(pRExC_state, op,
20100 (start[0] | (end[0] - start[0]) << ANYOFR_BASE_BITS));
20101
20102 /* Place the lowest UTF-8 start byte in the flags field, so as to allow
20103 * efficient ruling out at run time of many possible inputs. */
20104 (void) uvchr_to_utf8(low_utf8, start[0]);
20105 (void) uvchr_to_utf8(high_utf8, end[0]);
20106
20107 /* If all code points share the same first byte, this can be an
20108 * ANYOFRb. Otherwise store the lowest UTF-8 start byte which can
20109 * quickly rule out many inputs at run-time without having to compute
20110 * the code point from UTF-8. For EBCDIC, we use I8, as not doing that
20111 * transformation would not rule out nearly so many things */
20112 if (low_utf8[0] == high_utf8[0]) {
20113 op = ANYOFRb;
20114 OP(REGNODE_p(*ret)) = op;
20115 ANYOF_FLAGS(REGNODE_p(*ret)) = low_utf8[0];
20116 }
20117 else {
20118 ANYOF_FLAGS(REGNODE_p(*ret)) = NATIVE_UTF8_TO_I8(low_utf8[0]);
20119 }
20120
20121 return op;
20122 }
20123
20124 /* If didn't find an optimization and there is no need for a bitmap,
20125 * optimize to indicate that */
20126 if ( lowest_cp >= NUM_ANYOF_CODE_POINTS
20127 && ! LOC
20128 && ! upper_latin1_only_utf8_matches
20129 && *anyof_flags == 0)
20130 {
20131 U8 low_utf8[UTF8_MAXBYTES+1];
20132 UV highest_cp = invlist_highest(cp_list);
20133
20134 /* Currently the maximum allowed code point by the system is IV_MAX.
20135 * Higher ones are reserved for future internal use. This particular
20136 * regnode can be used for higher ones, but we can't calculate the code
20137 * point of those. IV_MAX suffices though, as it will be a large first
20138 * byte */
20139 Size_t low_len = uvchr_to_utf8(low_utf8, MIN(lowest_cp, IV_MAX))
20140 - low_utf8;
20141
20142 /* We store the lowest possible first byte of the UTF-8 representation,
20143 * using the flags field. This allows for quick ruling out of some
20144 * inputs without having to convert from UTF-8 to code point. For
20145 * EBCDIC, we use I8, as not doing that transformation would not rule
20146 * out nearly so many things */
20147 *anyof_flags = NATIVE_UTF8_TO_I8(low_utf8[0]);
20148
20149 op = ANYOFH;
20150
20151 /* If the first UTF-8 start byte for the highest code point in the
20152 * range is suitably small, we may be able to get an upper bound as
20153 * well */
20154 if (highest_cp <= IV_MAX) {
20155 U8 high_utf8[UTF8_MAXBYTES+1];
20156 Size_t high_len = uvchr_to_utf8(high_utf8, highest_cp) - high_utf8;
20157
20158 /* If the lowest and highest are the same, we can get an exact
20159 * first byte instead of a just minimum or even a sequence of exact
20160 * leading bytes. We signal these with different regnodes */
20161 if (low_utf8[0] == high_utf8[0]) {
20162 Size_t len = find_first_differing_byte_pos(low_utf8,
20163 high_utf8,
20164 MIN(low_len, high_len));
20165
20166 if (len == 1) {
20167
20168 /* No need to convert to I8 for EBCDIC as this is an exact
20169 * match */
20170 *anyof_flags = low_utf8[0];
20171 op = ANYOFHb;
20172 }
20173 else {
20174 op = ANYOFHs;
20175 *ret = regnode_guts(pRExC_state, op,
20176 regarglen[op] + STR_SZ(len),
20177 "anyofhs");
20178 FILL_NODE(*ret, op);
20179 ((struct regnode_anyofhs *) REGNODE_p(*ret))->str_len
20180 = len;
20181 Copy(low_utf8, /* Add the common bytes */
20182 ((struct regnode_anyofhs *) REGNODE_p(*ret))->string,
20183 len, U8);
20184 RExC_emit += NODE_SZ_STR(REGNODE_p(*ret));
20185 set_ANYOF_arg(pRExC_state, REGNODE_p(*ret), cp_list,
20186 NULL, only_utf8_locale_list);
20187 return op;
20188 }
20189 }
20190 else if (NATIVE_UTF8_TO_I8(high_utf8[0]) <= MAX_ANYOF_HRx_BYTE) {
20191
20192 /* Here, the high byte is not the same as the low, but is small
20193 * enough that its reasonable to have a loose upper bound,
20194 * which is packed in with the strict lower bound. See
20195 * comments at the definition of MAX_ANYOF_HRx_BYTE. On EBCDIC
20196 * platforms, I8 is used. On ASCII platforms I8 is the same
20197 * thing as UTF-8 */
20198
20199 U8 bits = 0;
20200 U8 max_range_diff = MAX_ANYOF_HRx_BYTE - *anyof_flags;
20201 U8 range_diff = NATIVE_UTF8_TO_I8(high_utf8[0])
20202 - *anyof_flags;
20203
20204 if (range_diff <= max_range_diff / 8) {
20205 bits = 3;
20206 }
20207 else if (range_diff <= max_range_diff / 4) {
20208 bits = 2;
20209 }
20210 else if (range_diff <= max_range_diff / 2) {
20211 bits = 1;
20212 }
20213 *anyof_flags = (*anyof_flags - 0xC0) << 2 | bits;
20214 op = ANYOFHr;
20215 }
20216 }
20217 }
20218
20219 return op;
20220
20221 return_OPFAIL:
20222 op = OPFAIL;
20223 *ret = reganode(pRExC_state, op, 0);
20224 return op;
20225
20226 return_SANY:
20227 op = SANY;
20228 *ret = reg_node(pRExC_state, op);
20229 MARK_NAUGHTY(1);
20230 return op;
20231 }
20232
20233 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
20234
20235 STATIC void
S_set_ANYOF_arg(pTHX_ RExC_state_t * const pRExC_state,regnode * const node,SV * const cp_list,SV * const runtime_defns,SV * const only_utf8_locale_list)20236 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
20237 regnode* const node,
20238 SV* const cp_list,
20239 SV* const runtime_defns,
20240 SV* const only_utf8_locale_list)
20241 {
20242 /* Sets the arg field of an ANYOF-type node 'node', using information about
20243 * the node passed-in. If there is nothing outside the node's bitmap, the
20244 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
20245 * the count returned by add_data(), having allocated and stored an array,
20246 * av, as follows:
20247 *
20248 * av[0] stores the inversion list defining this class as far as known at
20249 * this time, or PL_sv_undef if nothing definite is now known.
20250 * av[1] stores the inversion list of code points that match only if the
20251 * current locale is UTF-8, or if none, PL_sv_undef if there is an
20252 * av[2], or no entry otherwise.
20253 * av[2] stores the list of user-defined properties whose subroutine
20254 * definitions aren't known at this time, or no entry if none. */
20255
20256 UV n;
20257
20258 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
20259
20260 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
20261 assert(! (ANYOF_FLAGS(node)
20262 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
20263 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
20264 }
20265 else {
20266 AV * const av = newAV();
20267 SV *rv;
20268
20269 if (cp_list) {
20270 av_store(av, INVLIST_INDEX, SvREFCNT_inc_NN(cp_list));
20271 }
20272
20273 /* (Note that if any of this changes, the size calculations in
20274 * S_optimize_regclass() might need to be updated.) */
20275
20276 if (only_utf8_locale_list) {
20277 av_store(av, ONLY_LOCALE_MATCHES_INDEX,
20278 SvREFCNT_inc_NN(only_utf8_locale_list));
20279 }
20280
20281 if (runtime_defns) {
20282 av_store(av, DEFERRED_USER_DEFINED_INDEX,
20283 SvREFCNT_inc_NN(runtime_defns));
20284 }
20285
20286 rv = newRV_noinc(MUTABLE_SV(av));
20287 n = add_data(pRExC_state, STR_WITH_LEN("s"));
20288 RExC_rxi->data->data[n] = (void*)rv;
20289 ARG_SET(node, n);
20290 }
20291 }
20292
20293 SV *
20294
20295 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
Perl_get_regclass_nonbitmap_data(pTHX_ const regexp * prog,const regnode * node,bool doinit,SV ** listsvp,SV ** only_utf8_locale_ptr,SV ** output_invlist)20296 Perl_get_regclass_nonbitmap_data(pTHX_ const regexp *prog, const regnode* node, bool doinit, SV** listsvp, SV** only_utf8_locale_ptr, SV** output_invlist)
20297 #else
20298 Perl_get_re_gclass_nonbitmap_data(pTHX_ const regexp *prog, const regnode* node, bool doinit, SV** listsvp, SV** only_utf8_locale_ptr, SV** output_invlist)
20299 #endif
20300
20301 {
20302 /* For internal core use only.
20303 * Returns the inversion list for the input 'node' in the regex 'prog'.
20304 * If <doinit> is 'true', will attempt to create the inversion list if not
20305 * already done.
20306 * If <listsvp> is non-null, will return the printable contents of the
20307 * property definition. This can be used to get debugging information
20308 * even before the inversion list exists, by calling this function with
20309 * 'doinit' set to false, in which case the components that will be used
20310 * to eventually create the inversion list are returned (in a printable
20311 * form).
20312 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
20313 * store an inversion list of code points that should match only if the
20314 * execution-time locale is a UTF-8 one.
20315 * If <output_invlist> is not NULL, it is where this routine is to store an
20316 * inversion list of the code points that would be instead returned in
20317 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
20318 * when this parameter is used, is just the non-code point data that
20319 * will go into creating the inversion list. This currently should be just
20320 * user-defined properties whose definitions were not known at compile
20321 * time. Using this parameter allows for easier manipulation of the
20322 * inversion list's data by the caller. It is illegal to call this
20323 * function with this parameter set, but not <listsvp>
20324 *
20325 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
20326 * that, in spite of this function's name, the inversion list it returns
20327 * may include the bitmap data as well */
20328
20329 SV *si = NULL; /* Input initialization string */
20330 SV* invlist = NULL;
20331
20332 RXi_GET_DECL(prog, progi);
20333 const struct reg_data * const data = prog ? progi->data : NULL;
20334
20335 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
20336 PERL_ARGS_ASSERT_GET_REGCLASS_NONBITMAP_DATA;
20337 #else
20338 PERL_ARGS_ASSERT_GET_RE_GCLASS_NONBITMAP_DATA;
20339 #endif
20340 assert(! output_invlist || listsvp);
20341
20342 if (data && data->count) {
20343 const U32 n = ARG(node);
20344
20345 if (data->what[n] == 's') {
20346 SV * const rv = MUTABLE_SV(data->data[n]);
20347 AV * const av = MUTABLE_AV(SvRV(rv));
20348 SV **const ary = AvARRAY(av);
20349
20350 invlist = ary[INVLIST_INDEX];
20351
20352 if (av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX) {
20353 *only_utf8_locale_ptr = ary[ONLY_LOCALE_MATCHES_INDEX];
20354 }
20355
20356 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
20357 si = ary[DEFERRED_USER_DEFINED_INDEX];
20358 }
20359
20360 if (doinit && (si || invlist)) {
20361 if (si) {
20362 bool user_defined;
20363 SV * msg = newSVpvs_flags("", SVs_TEMP);
20364
20365 SV * prop_definition = handle_user_defined_property(
20366 "", 0, FALSE, /* There is no \p{}, \P{} */
20367 SvPVX_const(si)[1] - '0', /* /i or not has been
20368 stored here for just
20369 this occasion */
20370 TRUE, /* run time */
20371 FALSE, /* This call must find the defn */
20372 si, /* The property definition */
20373 &user_defined,
20374 msg,
20375 0 /* base level call */
20376 );
20377
20378 if (SvCUR(msg)) {
20379 assert(prop_definition == NULL);
20380
20381 Perl_croak(aTHX_ "%" UTF8f,
20382 UTF8fARG(SvUTF8(msg), SvCUR(msg), SvPVX(msg)));
20383 }
20384
20385 if (invlist) {
20386 _invlist_union(invlist, prop_definition, &invlist);
20387 SvREFCNT_dec_NN(prop_definition);
20388 }
20389 else {
20390 invlist = prop_definition;
20391 }
20392
20393 STATIC_ASSERT_STMT(ONLY_LOCALE_MATCHES_INDEX == 1 + INVLIST_INDEX);
20394 STATIC_ASSERT_STMT(DEFERRED_USER_DEFINED_INDEX == 1 + ONLY_LOCALE_MATCHES_INDEX);
20395
20396 ary[INVLIST_INDEX] = invlist;
20397 av_fill(av, (ary[ONLY_LOCALE_MATCHES_INDEX])
20398 ? ONLY_LOCALE_MATCHES_INDEX
20399 : INVLIST_INDEX);
20400 si = NULL;
20401 }
20402 }
20403 }
20404 }
20405
20406 /* If requested, return a printable version of what this ANYOF node matches
20407 * */
20408 if (listsvp) {
20409 SV* matches_string = NULL;
20410
20411 /* This function can be called at compile-time, before everything gets
20412 * resolved, in which case we return the currently best available
20413 * information, which is the string that will eventually be used to do
20414 * that resolving, 'si' */
20415 if (si) {
20416 /* Here, we only have 'si' (and possibly some passed-in data in
20417 * 'invlist', which is handled below) If the caller only wants
20418 * 'si', use that. */
20419 if (! output_invlist) {
20420 matches_string = newSVsv(si);
20421 }
20422 else {
20423 /* But if the caller wants an inversion list of the node, we
20424 * need to parse 'si' and place as much as possible in the
20425 * desired output inversion list, making 'matches_string' only
20426 * contain the currently unresolvable things */
20427 const char *si_string = SvPVX(si);
20428 STRLEN remaining = SvCUR(si);
20429 UV prev_cp = 0;
20430 U8 count = 0;
20431
20432 /* Ignore everything before and including the first new-line */
20433 si_string = (const char *) memchr(si_string, '\n', SvCUR(si));
20434 assert (si_string != NULL);
20435 si_string++;
20436 remaining = SvPVX(si) + SvCUR(si) - si_string;
20437
20438 while (remaining > 0) {
20439
20440 /* The data consists of just strings defining user-defined
20441 * property names, but in prior incarnations, and perhaps
20442 * somehow from pluggable regex engines, it could still
20443 * hold hex code point definitions, all of which should be
20444 * legal (or it wouldn't have gotten this far). Each
20445 * component of a range would be separated by a tab, and
20446 * each range by a new-line. If these are found, instead
20447 * add them to the inversion list */
20448 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
20449 |PERL_SCAN_SILENT_NON_PORTABLE;
20450 STRLEN len = remaining;
20451 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
20452
20453 /* If the hex decode routine found something, it should go
20454 * up to the next \n */
20455 if ( *(si_string + len) == '\n') {
20456 if (count) { /* 2nd code point on line */
20457 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
20458 }
20459 else {
20460 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
20461 }
20462 count = 0;
20463 goto prepare_for_next_iteration;
20464 }
20465
20466 /* If the hex decode was instead for the lower range limit,
20467 * save it, and go parse the upper range limit */
20468 if (*(si_string + len) == '\t') {
20469 assert(count == 0);
20470
20471 prev_cp = cp;
20472 count = 1;
20473 prepare_for_next_iteration:
20474 si_string += len + 1;
20475 remaining -= len + 1;
20476 continue;
20477 }
20478
20479 /* Here, didn't find a legal hex number. Just add the text
20480 * from here up to the next \n, omitting any trailing
20481 * markers. */
20482
20483 remaining -= len;
20484 len = strcspn(si_string,
20485 DEFERRED_COULD_BE_OFFICIAL_MARKERs "\n");
20486 remaining -= len;
20487 if (matches_string) {
20488 sv_catpvn(matches_string, si_string, len);
20489 }
20490 else {
20491 matches_string = newSVpvn(si_string, len);
20492 }
20493 sv_catpvs(matches_string, " ");
20494
20495 si_string += len;
20496 if ( remaining
20497 && UCHARAT(si_string)
20498 == DEFERRED_COULD_BE_OFFICIAL_MARKERc)
20499 {
20500 si_string++;
20501 remaining--;
20502 }
20503 if (remaining && UCHARAT(si_string) == '\n') {
20504 si_string++;
20505 remaining--;
20506 }
20507 } /* end of loop through the text */
20508
20509 assert(matches_string);
20510 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
20511 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
20512 }
20513 } /* end of has an 'si' */
20514 }
20515
20516 /* Add the stuff that's already known */
20517 if (invlist) {
20518
20519 /* Again, if the caller doesn't want the output inversion list, put
20520 * everything in 'matches-string' */
20521 if (! output_invlist) {
20522 if ( ! matches_string) {
20523 matches_string = newSVpvs("\n");
20524 }
20525 sv_catsv(matches_string, invlist_contents(invlist,
20526 TRUE /* traditional style */
20527 ));
20528 }
20529 else if (! *output_invlist) {
20530 *output_invlist = invlist_clone(invlist, NULL);
20531 }
20532 else {
20533 _invlist_union(*output_invlist, invlist, output_invlist);
20534 }
20535 }
20536
20537 *listsvp = matches_string;
20538 }
20539
20540 return invlist;
20541 }
20542
20543 /* reg_skipcomment()
20544
20545 Absorbs an /x style # comment from the input stream,
20546 returning a pointer to the first character beyond the comment, or if the
20547 comment terminates the pattern without anything following it, this returns
20548 one past the final character of the pattern (in other words, RExC_end) and
20549 sets the REG_RUN_ON_COMMENT_SEEN flag.
20550
20551 Note it's the callers responsibility to ensure that we are
20552 actually in /x mode
20553
20554 */
20555
20556 PERL_STATIC_INLINE char*
S_reg_skipcomment(RExC_state_t * pRExC_state,char * p)20557 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
20558 {
20559 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
20560
20561 assert(*p == '#');
20562
20563 while (p < RExC_end) {
20564 if (*(++p) == '\n') {
20565 return p+1;
20566 }
20567 }
20568
20569 /* we ran off the end of the pattern without ending the comment, so we have
20570 * to add an \n when wrapping */
20571 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
20572 return p;
20573 }
20574
20575 STATIC void
S_skip_to_be_ignored_text(pTHX_ RExC_state_t * pRExC_state,char ** p,const bool force_to_xmod)20576 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
20577 char ** p,
20578 const bool force_to_xmod
20579 )
20580 {
20581 /* If the text at the current parse position '*p' is a '(?#...)' comment,
20582 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
20583 * is /x whitespace, advance '*p' so that on exit it points to the first
20584 * byte past all such white space and comments */
20585
20586 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
20587
20588 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
20589
20590 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
20591
20592 for (;;) {
20593 if (RExC_end - (*p) >= 3
20594 && *(*p) == '('
20595 && *(*p + 1) == '?'
20596 && *(*p + 2) == '#')
20597 {
20598 while (*(*p) != ')') {
20599 if ((*p) == RExC_end)
20600 FAIL("Sequence (?#... not terminated");
20601 (*p)++;
20602 }
20603 (*p)++;
20604 continue;
20605 }
20606
20607 if (use_xmod) {
20608 const char * save_p = *p;
20609 while ((*p) < RExC_end) {
20610 STRLEN len;
20611 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
20612 (*p) += len;
20613 }
20614 else if (*(*p) == '#') {
20615 (*p) = reg_skipcomment(pRExC_state, (*p));
20616 }
20617 else {
20618 break;
20619 }
20620 }
20621 if (*p != save_p) {
20622 continue;
20623 }
20624 }
20625
20626 break;
20627 }
20628
20629 return;
20630 }
20631
20632 /* nextchar()
20633
20634 Advances the parse position by one byte, unless that byte is the beginning
20635 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
20636 those two cases, the parse position is advanced beyond all such comments and
20637 white space.
20638
20639 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
20640 */
20641
20642 STATIC void
S_nextchar(pTHX_ RExC_state_t * pRExC_state)20643 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
20644 {
20645 PERL_ARGS_ASSERT_NEXTCHAR;
20646
20647 if (RExC_parse < RExC_end) {
20648 assert( ! UTF
20649 || UTF8_IS_INVARIANT(*RExC_parse)
20650 || UTF8_IS_START(*RExC_parse));
20651
20652 RExC_parse += (UTF)
20653 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
20654 : 1;
20655
20656 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
20657 FALSE /* Don't force /x */ );
20658 }
20659 }
20660
20661 STATIC void
S_change_engine_size(pTHX_ RExC_state_t * pRExC_state,const Ptrdiff_t size)20662 S_change_engine_size(pTHX_ RExC_state_t *pRExC_state, const Ptrdiff_t size)
20663 {
20664 /* 'size' is the delta number of smallest regnode equivalents to add or
20665 * subtract from the current memory allocated to the regex engine being
20666 * constructed. */
20667
20668 PERL_ARGS_ASSERT_CHANGE_ENGINE_SIZE;
20669
20670 RExC_size += size;
20671
20672 Renewc(RExC_rxi,
20673 sizeof(regexp_internal) + (RExC_size + 1) * sizeof(regnode),
20674 /* +1 for REG_MAGIC */
20675 char,
20676 regexp_internal);
20677 if ( RExC_rxi == NULL )
20678 FAIL("Regexp out of space");
20679 RXi_SET(RExC_rx, RExC_rxi);
20680
20681 RExC_emit_start = RExC_rxi->program;
20682 if (size > 0) {
20683 Zero(REGNODE_p(RExC_emit), size, regnode);
20684 }
20685
20686 #ifdef RE_TRACK_PATTERN_OFFSETS
20687 Renew(RExC_offsets, 2*RExC_size+1, U32);
20688 if (size > 0) {
20689 Zero(RExC_offsets + 2*(RExC_size - size) + 1, 2 * size, U32);
20690 }
20691 RExC_offsets[0] = RExC_size;
20692 #endif
20693 }
20694
20695 STATIC regnode_offset
S_regnode_guts(pTHX_ RExC_state_t * pRExC_state,const U8 op,const STRLEN extra_size,const char * const name)20696 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
20697 {
20698 /* Allocate a regnode for 'op', with 'extra_size' extra (smallest) regnode
20699 * equivalents space. It aligns and increments RExC_size
20700 *
20701 * It returns the regnode's offset into the regex engine program */
20702
20703 const regnode_offset ret = RExC_emit;
20704
20705 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20706
20707 PERL_ARGS_ASSERT_REGNODE_GUTS;
20708
20709 SIZE_ALIGN(RExC_size);
20710 change_engine_size(pRExC_state, (Ptrdiff_t) 1 + extra_size);
20711 NODE_ALIGN_FILL(REGNODE_p(ret));
20712 #ifndef RE_TRACK_PATTERN_OFFSETS
20713 PERL_UNUSED_ARG(name);
20714 PERL_UNUSED_ARG(op);
20715 #else
20716 assert(extra_size >= regarglen[op] || PL_regkind[op] == ANYOF);
20717
20718 if (RExC_offsets) { /* MJD */
20719 MJD_OFFSET_DEBUG(
20720 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
20721 name, __LINE__,
20722 PL_reg_name[op],
20723 (UV)(RExC_emit) > RExC_offsets[0]
20724 ? "Overwriting end of array!\n" : "OK",
20725 (UV)(RExC_emit),
20726 (UV)(RExC_parse - RExC_start),
20727 (UV)RExC_offsets[0]));
20728 Set_Node_Offset(REGNODE_p(RExC_emit), RExC_parse + (op == END));
20729 }
20730 #endif
20731 return(ret);
20732 }
20733
20734 /*
20735 - reg_node - emit a node
20736 */
20737 STATIC regnode_offset /* Location. */
S_reg_node(pTHX_ RExC_state_t * pRExC_state,U8 op)20738 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
20739 {
20740 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
20741 regnode_offset ptr = ret;
20742
20743 PERL_ARGS_ASSERT_REG_NODE;
20744
20745 assert(regarglen[op] == 0);
20746
20747 FILL_ADVANCE_NODE(ptr, op);
20748 RExC_emit = ptr;
20749 return(ret);
20750 }
20751
20752 /*
20753 - reganode - emit a node with an argument
20754 */
20755 STATIC regnode_offset /* Location. */
S_reganode(pTHX_ RExC_state_t * pRExC_state,U8 op,U32 arg)20756 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
20757 {
20758 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
20759 regnode_offset ptr = ret;
20760
20761 PERL_ARGS_ASSERT_REGANODE;
20762
20763 /* ANYOF are special cased to allow non-length 1 args */
20764 assert(regarglen[op] == 1);
20765
20766 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
20767 RExC_emit = ptr;
20768 return(ret);
20769 }
20770
20771 /*
20772 - regpnode - emit a temporary node with a SV* argument
20773 */
20774 STATIC regnode_offset /* Location. */
S_regpnode(pTHX_ RExC_state_t * pRExC_state,U8 op,SV * arg)20775 S_regpnode(pTHX_ RExC_state_t *pRExC_state, U8 op, SV * arg)
20776 {
20777 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "regpnode");
20778 regnode_offset ptr = ret;
20779
20780 PERL_ARGS_ASSERT_REGPNODE;
20781
20782 FILL_ADVANCE_NODE_ARGp(ptr, op, arg);
20783 RExC_emit = ptr;
20784 return(ret);
20785 }
20786
20787 STATIC regnode_offset
S_reg2Lanode(pTHX_ RExC_state_t * pRExC_state,const U8 op,const U32 arg1,const I32 arg2)20788 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
20789 {
20790 /* emit a node with U32 and I32 arguments */
20791
20792 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
20793 regnode_offset ptr = ret;
20794
20795 PERL_ARGS_ASSERT_REG2LANODE;
20796
20797 assert(regarglen[op] == 2);
20798
20799 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
20800 RExC_emit = ptr;
20801 return(ret);
20802 }
20803
20804 /*
20805 - reginsert - insert an operator in front of already-emitted operand
20806 *
20807 * That means that on exit 'operand' is the offset of the newly inserted
20808 * operator, and the original operand has been relocated.
20809 *
20810 * IMPORTANT NOTE - it is the *callers* responsibility to correctly
20811 * set up NEXT_OFF() of the inserted node if needed. Something like this:
20812 *
20813 * reginsert(pRExC, OPFAIL, orig_emit, depth+1);
20814 * NEXT_OFF(orig_emit) = regarglen[OPFAIL] + NODE_STEP_REGNODE;
20815 *
20816 * ALSO NOTE - FLAGS(newly-inserted-operator) will be set to 0 as well.
20817 */
20818 STATIC void
S_reginsert(pTHX_ RExC_state_t * pRExC_state,const U8 op,const regnode_offset operand,const U32 depth)20819 S_reginsert(pTHX_ RExC_state_t *pRExC_state, const U8 op,
20820 const regnode_offset operand, const U32 depth)
20821 {
20822 regnode *src;
20823 regnode *dst;
20824 regnode *place;
20825 const int offset = regarglen[(U8)op];
20826 const int size = NODE_STEP_REGNODE + offset;
20827 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20828
20829 PERL_ARGS_ASSERT_REGINSERT;
20830 PERL_UNUSED_CONTEXT;
20831 PERL_UNUSED_ARG(depth);
20832 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
20833 DEBUG_PARSE_FMT("inst"," - %s", PL_reg_name[op]);
20834 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
20835 studying. If this is wrong then we need to adjust RExC_recurse
20836 below like we do with RExC_open_parens/RExC_close_parens. */
20837 change_engine_size(pRExC_state, (Ptrdiff_t) size);
20838 src = REGNODE_p(RExC_emit);
20839 RExC_emit += size;
20840 dst = REGNODE_p(RExC_emit);
20841
20842 /* If we are in a "count the parentheses" pass, the numbers are unreliable,
20843 * and [perl #133871] shows this can lead to problems, so skip this
20844 * realignment of parens until a later pass when they are reliable */
20845 if (! IN_PARENS_PASS && RExC_open_parens) {
20846 int paren;
20847 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
20848 /* remember that RExC_npar is rex->nparens + 1,
20849 * iow it is 1 more than the number of parens seen in
20850 * the pattern so far. */
20851 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
20852 /* note, RExC_open_parens[0] is the start of the
20853 * regex, it can't move. RExC_close_parens[0] is the end
20854 * of the regex, it *can* move. */
20855 if ( paren && RExC_open_parens[paren] >= operand ) {
20856 /*DEBUG_PARSE_FMT("open"," - %d", size);*/
20857 RExC_open_parens[paren] += size;
20858 } else {
20859 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
20860 }
20861 if ( RExC_close_parens[paren] >= operand ) {
20862 /*DEBUG_PARSE_FMT("close"," - %d", size);*/
20863 RExC_close_parens[paren] += size;
20864 } else {
20865 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
20866 }
20867 }
20868 }
20869 if (RExC_end_op)
20870 RExC_end_op += size;
20871
20872 while (src > REGNODE_p(operand)) {
20873 StructCopy(--src, --dst, regnode);
20874 #ifdef RE_TRACK_PATTERN_OFFSETS
20875 if (RExC_offsets) { /* MJD 20010112 */
20876 MJD_OFFSET_DEBUG(
20877 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
20878 "reginsert",
20879 __LINE__,
20880 PL_reg_name[op],
20881 (UV)(REGNODE_OFFSET(dst)) > RExC_offsets[0]
20882 ? "Overwriting end of array!\n" : "OK",
20883 (UV)REGNODE_OFFSET(src),
20884 (UV)REGNODE_OFFSET(dst),
20885 (UV)RExC_offsets[0]));
20886 Set_Node_Offset_To_R(REGNODE_OFFSET(dst), Node_Offset(src));
20887 Set_Node_Length_To_R(REGNODE_OFFSET(dst), Node_Length(src));
20888 }
20889 #endif
20890 }
20891
20892 place = REGNODE_p(operand); /* Op node, where operand used to be. */
20893 #ifdef RE_TRACK_PATTERN_OFFSETS
20894 if (RExC_offsets) { /* MJD */
20895 MJD_OFFSET_DEBUG(
20896 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
20897 "reginsert",
20898 __LINE__,
20899 PL_reg_name[op],
20900 (UV)REGNODE_OFFSET(place) > RExC_offsets[0]
20901 ? "Overwriting end of array!\n" : "OK",
20902 (UV)REGNODE_OFFSET(place),
20903 (UV)(RExC_parse - RExC_start),
20904 (UV)RExC_offsets[0]));
20905 Set_Node_Offset(place, RExC_parse);
20906 Set_Node_Length(place, 1);
20907 }
20908 #endif
20909 src = NEXTOPER(place);
20910 FLAGS(place) = 0;
20911 FILL_NODE(operand, op);
20912
20913 /* Zero out any arguments in the new node */
20914 Zero(src, offset, regnode);
20915 }
20916
20917 /*
20918 - regtail - set the next-pointer at the end of a node chain of p to val. If
20919 that value won't fit in the space available, instead returns FALSE.
20920 (Except asserts if we can't fit in the largest space the regex
20921 engine is designed for.)
20922 - SEE ALSO: regtail_study
20923 */
20924 STATIC bool
S_regtail(pTHX_ RExC_state_t * pRExC_state,const regnode_offset p,const regnode_offset val,const U32 depth)20925 S_regtail(pTHX_ RExC_state_t * pRExC_state,
20926 const regnode_offset p,
20927 const regnode_offset val,
20928 const U32 depth)
20929 {
20930 regnode_offset scan;
20931 DECLARE_AND_GET_RE_DEBUG_FLAGS;
20932
20933 PERL_ARGS_ASSERT_REGTAIL;
20934 #ifndef DEBUGGING
20935 PERL_UNUSED_ARG(depth);
20936 #endif
20937
20938 /* The final node in the chain is the first one with a nonzero next pointer
20939 * */
20940 scan = (regnode_offset) p;
20941 for (;;) {
20942 regnode * const temp = regnext(REGNODE_p(scan));
20943 DEBUG_PARSE_r({
20944 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
20945 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
20946 Perl_re_printf( aTHX_ "~ %s (%zu) %s %s\n",
20947 SvPV_nolen_const(RExC_mysv), scan,
20948 (temp == NULL ? "->" : ""),
20949 (temp == NULL ? PL_reg_name[OP(REGNODE_p(val))] : "")
20950 );
20951 });
20952 if (temp == NULL)
20953 break;
20954 scan = REGNODE_OFFSET(temp);
20955 }
20956
20957 /* Populate this node's next pointer */
20958 assert(val >= scan);
20959 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
20960 assert((UV) (val - scan) <= U32_MAX);
20961 ARG_SET(REGNODE_p(scan), val - scan);
20962 }
20963 else {
20964 if (val - scan > U16_MAX) {
20965 /* Populate this with something that won't loop and will likely
20966 * lead to a crash if the caller ignores the failure return, and
20967 * execution continues */
20968 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
20969 return FALSE;
20970 }
20971 NEXT_OFF(REGNODE_p(scan)) = val - scan;
20972 }
20973
20974 return TRUE;
20975 }
20976
20977 #ifdef DEBUGGING
20978 /*
20979 - regtail_study - set the next-pointer at the end of a node chain of p to val.
20980 - Look for optimizable sequences at the same time.
20981 - currently only looks for EXACT chains.
20982
20983 This is experimental code. The idea is to use this routine to perform
20984 in place optimizations on branches and groups as they are constructed,
20985 with the long term intention of removing optimization from study_chunk so
20986 that it is purely analytical.
20987
20988 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
20989 to control which is which.
20990
20991 This used to return a value that was ignored. It was a problem that it is
20992 #ifdef'd to be another function that didn't return a value. khw has changed it
20993 so both currently return a pass/fail return.
20994
20995 */
20996 /* TODO: All four parms should be const */
20997
20998 STATIC bool
S_regtail_study(pTHX_ RExC_state_t * pRExC_state,regnode_offset p,const regnode_offset val,U32 depth)20999 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode_offset p,
21000 const regnode_offset val, U32 depth)
21001 {
21002 regnode_offset scan;
21003 U8 exact = PSEUDO;
21004 #ifdef EXPERIMENTAL_INPLACESCAN
21005 I32 min = 0;
21006 #endif
21007 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21008
21009 PERL_ARGS_ASSERT_REGTAIL_STUDY;
21010
21011
21012 /* Find last node. */
21013
21014 scan = p;
21015 for (;;) {
21016 regnode * const temp = regnext(REGNODE_p(scan));
21017 #ifdef EXPERIMENTAL_INPLACESCAN
21018 if (PL_regkind[OP(REGNODE_p(scan))] == EXACT) {
21019 bool unfolded_multi_char; /* Unexamined in this routine */
21020 if (join_exact(pRExC_state, scan, &min,
21021 &unfolded_multi_char, 1, REGNODE_p(val), depth+1))
21022 return TRUE; /* Was return EXACT */
21023 }
21024 #endif
21025 if ( exact ) {
21026 if (PL_regkind[OP(REGNODE_p(scan))] == EXACT) {
21027 if (exact == PSEUDO )
21028 exact= OP(REGNODE_p(scan));
21029 else if (exact != OP(REGNODE_p(scan)) )
21030 exact= 0;
21031 }
21032 else if (OP(REGNODE_p(scan)) != NOTHING) {
21033 exact= 0;
21034 }
21035 }
21036 DEBUG_PARSE_r({
21037 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
21038 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
21039 Perl_re_printf( aTHX_ "~ %s (%zu) -> %s\n",
21040 SvPV_nolen_const(RExC_mysv),
21041 scan,
21042 PL_reg_name[exact]);
21043 });
21044 if (temp == NULL)
21045 break;
21046 scan = REGNODE_OFFSET(temp);
21047 }
21048 DEBUG_PARSE_r({
21049 DEBUG_PARSE_MSG("");
21050 regprop(RExC_rx, RExC_mysv, REGNODE_p(val), NULL, pRExC_state);
21051 Perl_re_printf( aTHX_
21052 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
21053 SvPV_nolen_const(RExC_mysv),
21054 (IV)val,
21055 (IV)(val - scan)
21056 );
21057 });
21058 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
21059 assert((UV) (val - scan) <= U32_MAX);
21060 ARG_SET(REGNODE_p(scan), val - scan);
21061 }
21062 else {
21063 if (val - scan > U16_MAX) {
21064 /* Populate this with something that won't loop and will likely
21065 * lead to a crash if the caller ignores the failure return, and
21066 * execution continues */
21067 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
21068 return FALSE;
21069 }
21070 NEXT_OFF(REGNODE_p(scan)) = val - scan;
21071 }
21072
21073 return TRUE; /* Was 'return exact' */
21074 }
21075 #endif
21076
21077 STATIC SV*
S_get_ANYOFM_contents(pTHX_ const regnode * n)21078 S_get_ANYOFM_contents(pTHX_ const regnode * n) {
21079
21080 /* Returns an inversion list of all the code points matched by the
21081 * ANYOFM/NANYOFM node 'n' */
21082
21083 SV * cp_list = _new_invlist(-1);
21084 const U8 lowest = (U8) ARG(n);
21085 unsigned int i;
21086 U8 count = 0;
21087 U8 needed = 1U << PL_bitcount[ (U8) ~ FLAGS(n)];
21088
21089 PERL_ARGS_ASSERT_GET_ANYOFM_CONTENTS;
21090
21091 /* Starting with the lowest code point, any code point that ANDed with the
21092 * mask yields the lowest code point is in the set */
21093 for (i = lowest; i <= 0xFF; i++) {
21094 if ((i & FLAGS(n)) == ARG(n)) {
21095 cp_list = add_cp_to_invlist(cp_list, i);
21096 count++;
21097
21098 /* We know how many code points (a power of two) that are in the
21099 * set. No use looking once we've got that number */
21100 if (count >= needed) break;
21101 }
21102 }
21103
21104 if (OP(n) == NANYOFM) {
21105 _invlist_invert(cp_list);
21106 }
21107 return cp_list;
21108 }
21109
21110 /*
21111 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
21112 */
21113 #ifdef DEBUGGING
21114
21115 static void
S_regdump_intflags(pTHX_ const char * lead,const U32 flags)21116 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
21117 {
21118 int bit;
21119 int set=0;
21120
21121 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
21122
21123 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
21124 if (flags & (1<<bit)) {
21125 if (!set++ && lead)
21126 Perl_re_printf( aTHX_ "%s", lead);
21127 Perl_re_printf( aTHX_ "%s ", PL_reg_intflags_name[bit]);
21128 }
21129 }
21130 if (lead) {
21131 if (set)
21132 Perl_re_printf( aTHX_ "\n");
21133 else
21134 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
21135 }
21136 }
21137
21138 static void
S_regdump_extflags(pTHX_ const char * lead,const U32 flags)21139 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
21140 {
21141 int bit;
21142 int set=0;
21143 regex_charset cs;
21144
21145 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
21146
21147 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
21148 if (flags & (1<<bit)) {
21149 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
21150 continue;
21151 }
21152 if (!set++ && lead)
21153 Perl_re_printf( aTHX_ "%s", lead);
21154 Perl_re_printf( aTHX_ "%s ", PL_reg_extflags_name[bit]);
21155 }
21156 }
21157 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
21158 if (!set++ && lead) {
21159 Perl_re_printf( aTHX_ "%s", lead);
21160 }
21161 switch (cs) {
21162 case REGEX_UNICODE_CHARSET:
21163 Perl_re_printf( aTHX_ "UNICODE");
21164 break;
21165 case REGEX_LOCALE_CHARSET:
21166 Perl_re_printf( aTHX_ "LOCALE");
21167 break;
21168 case REGEX_ASCII_RESTRICTED_CHARSET:
21169 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
21170 break;
21171 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
21172 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
21173 break;
21174 default:
21175 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
21176 break;
21177 }
21178 }
21179 if (lead) {
21180 if (set)
21181 Perl_re_printf( aTHX_ "\n");
21182 else
21183 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
21184 }
21185 }
21186 #endif
21187
21188 void
Perl_regdump(pTHX_ const regexp * r)21189 Perl_regdump(pTHX_ const regexp *r)
21190 {
21191 #ifdef DEBUGGING
21192 int i;
21193 SV * const sv = sv_newmortal();
21194 SV *dsv= sv_newmortal();
21195 RXi_GET_DECL(r, ri);
21196 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21197
21198 PERL_ARGS_ASSERT_REGDUMP;
21199
21200 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
21201
21202 /* Header fields of interest. */
21203 for (i = 0; i < 2; i++) {
21204 if (r->substrs->data[i].substr) {
21205 RE_PV_QUOTED_DECL(s, 0, dsv,
21206 SvPVX_const(r->substrs->data[i].substr),
21207 RE_SV_DUMPLEN(r->substrs->data[i].substr),
21208 PL_dump_re_max_len);
21209 Perl_re_printf( aTHX_
21210 "%s %s%s at %" IVdf "..%" UVuf " ",
21211 i ? "floating" : "anchored",
21212 s,
21213 RE_SV_TAIL(r->substrs->data[i].substr),
21214 (IV)r->substrs->data[i].min_offset,
21215 (UV)r->substrs->data[i].max_offset);
21216 }
21217 else if (r->substrs->data[i].utf8_substr) {
21218 RE_PV_QUOTED_DECL(s, 1, dsv,
21219 SvPVX_const(r->substrs->data[i].utf8_substr),
21220 RE_SV_DUMPLEN(r->substrs->data[i].utf8_substr),
21221 30);
21222 Perl_re_printf( aTHX_
21223 "%s utf8 %s%s at %" IVdf "..%" UVuf " ",
21224 i ? "floating" : "anchored",
21225 s,
21226 RE_SV_TAIL(r->substrs->data[i].utf8_substr),
21227 (IV)r->substrs->data[i].min_offset,
21228 (UV)r->substrs->data[i].max_offset);
21229 }
21230 }
21231
21232 if (r->check_substr || r->check_utf8)
21233 Perl_re_printf( aTHX_
21234 (const char *)
21235 ( r->check_substr == r->substrs->data[1].substr
21236 && r->check_utf8 == r->substrs->data[1].utf8_substr
21237 ? "(checking floating" : "(checking anchored"));
21238 if (r->intflags & PREGf_NOSCAN)
21239 Perl_re_printf( aTHX_ " noscan");
21240 if (r->extflags & RXf_CHECK_ALL)
21241 Perl_re_printf( aTHX_ " isall");
21242 if (r->check_substr || r->check_utf8)
21243 Perl_re_printf( aTHX_ ") ");
21244
21245 if (ri->regstclass) {
21246 regprop(r, sv, ri->regstclass, NULL, NULL);
21247 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
21248 }
21249 if (r->intflags & PREGf_ANCH) {
21250 Perl_re_printf( aTHX_ "anchored");
21251 if (r->intflags & PREGf_ANCH_MBOL)
21252 Perl_re_printf( aTHX_ "(MBOL)");
21253 if (r->intflags & PREGf_ANCH_SBOL)
21254 Perl_re_printf( aTHX_ "(SBOL)");
21255 if (r->intflags & PREGf_ANCH_GPOS)
21256 Perl_re_printf( aTHX_ "(GPOS)");
21257 Perl_re_printf( aTHX_ " ");
21258 }
21259 if (r->intflags & PREGf_GPOS_SEEN)
21260 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
21261 if (r->intflags & PREGf_SKIP)
21262 Perl_re_printf( aTHX_ "plus ");
21263 if (r->intflags & PREGf_IMPLICIT)
21264 Perl_re_printf( aTHX_ "implicit ");
21265 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
21266 if (r->extflags & RXf_EVAL_SEEN)
21267 Perl_re_printf( aTHX_ "with eval ");
21268 Perl_re_printf( aTHX_ "\n");
21269 DEBUG_FLAGS_r({
21270 regdump_extflags("r->extflags: ", r->extflags);
21271 regdump_intflags("r->intflags: ", r->intflags);
21272 });
21273 #else
21274 PERL_ARGS_ASSERT_REGDUMP;
21275 PERL_UNUSED_CONTEXT;
21276 PERL_UNUSED_ARG(r);
21277 #endif /* DEBUGGING */
21278 }
21279
21280 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
21281 #ifdef DEBUGGING
21282
21283 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
21284 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
21285 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
21286 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
21287 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
21288 || _CC_VERTSPACE != 15
21289 # error Need to adjust order of anyofs[]
21290 # endif
21291 static const char * const anyofs[] = {
21292 "\\w",
21293 "\\W",
21294 "\\d",
21295 "\\D",
21296 "[:alpha:]",
21297 "[:^alpha:]",
21298 "[:lower:]",
21299 "[:^lower:]",
21300 "[:upper:]",
21301 "[:^upper:]",
21302 "[:punct:]",
21303 "[:^punct:]",
21304 "[:print:]",
21305 "[:^print:]",
21306 "[:alnum:]",
21307 "[:^alnum:]",
21308 "[:graph:]",
21309 "[:^graph:]",
21310 "[:cased:]",
21311 "[:^cased:]",
21312 "\\s",
21313 "\\S",
21314 "[:blank:]",
21315 "[:^blank:]",
21316 "[:xdigit:]",
21317 "[:^xdigit:]",
21318 "[:cntrl:]",
21319 "[:^cntrl:]",
21320 "[:ascii:]",
21321 "[:^ascii:]",
21322 "\\v",
21323 "\\V"
21324 };
21325 #endif
21326
21327 /*
21328 - regprop - printable representation of opcode, with run time support
21329 */
21330
21331 void
Perl_regprop(pTHX_ const regexp * prog,SV * sv,const regnode * o,const regmatch_info * reginfo,const RExC_state_t * pRExC_state)21332 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
21333 {
21334 #ifdef DEBUGGING
21335 int k;
21336 RXi_GET_DECL(prog, progi);
21337 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21338
21339 PERL_ARGS_ASSERT_REGPROP;
21340
21341 SvPVCLEAR(sv);
21342
21343 if (OP(o) > REGNODE_MAX) { /* regnode.type is unsigned */
21344 if (pRExC_state) { /* This gives more info, if we have it */
21345 FAIL3("panic: corrupted regexp opcode %d > %d",
21346 (int)OP(o), (int)REGNODE_MAX);
21347 }
21348 else {
21349 Perl_croak(aTHX_ "panic: corrupted regexp opcode %d > %d",
21350 (int)OP(o), (int)REGNODE_MAX);
21351 }
21352 }
21353 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
21354
21355 k = PL_regkind[OP(o)];
21356
21357 if (k == EXACT) {
21358 sv_catpvs(sv, " ");
21359 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
21360 * is a crude hack but it may be the best for now since
21361 * we have no flag "this EXACTish node was UTF-8"
21362 * --jhi */
21363 pv_pretty(sv, STRING(o), STR_LEN(o), PL_dump_re_max_len,
21364 PL_colors[0], PL_colors[1],
21365 PERL_PV_ESCAPE_UNI_DETECT |
21366 PERL_PV_ESCAPE_NONASCII |
21367 PERL_PV_PRETTY_ELLIPSES |
21368 PERL_PV_PRETTY_LTGT |
21369 PERL_PV_PRETTY_NOCLEAR
21370 );
21371 } else if (k == TRIE) {
21372 /* print the details of the trie in dumpuntil instead, as
21373 * progi->data isn't available here */
21374 const char op = OP(o);
21375 const U32 n = ARG(o);
21376 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
21377 (reg_ac_data *)progi->data->data[n] :
21378 NULL;
21379 const reg_trie_data * const trie
21380 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
21381
21382 Perl_sv_catpvf(aTHX_ sv, "-%s", PL_reg_name[o->flags]);
21383 DEBUG_TRIE_COMPILE_r({
21384 if (trie->jump)
21385 sv_catpvs(sv, "(JUMP)");
21386 Perl_sv_catpvf(aTHX_ sv,
21387 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
21388 (UV)trie->startstate,
21389 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
21390 (UV)trie->wordcount,
21391 (UV)trie->minlen,
21392 (UV)trie->maxlen,
21393 (UV)TRIE_CHARCOUNT(trie),
21394 (UV)trie->uniquecharcount
21395 );
21396 });
21397 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
21398 sv_catpvs(sv, "[");
21399 (void) put_charclass_bitmap_innards(sv,
21400 ((IS_ANYOF_TRIE(op))
21401 ? ANYOF_BITMAP(o)
21402 : TRIE_BITMAP(trie)),
21403 NULL,
21404 NULL,
21405 NULL,
21406 0,
21407 FALSE
21408 );
21409 sv_catpvs(sv, "]");
21410 }
21411 } else if (k == CURLY) {
21412 U32 lo = ARG1(o), hi = ARG2(o);
21413 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
21414 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
21415 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
21416 if (hi == REG_INFTY)
21417 sv_catpvs(sv, "INFTY");
21418 else
21419 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
21420 sv_catpvs(sv, "}");
21421 }
21422 else if (k == WHILEM && o->flags) /* Ordinal/of */
21423 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
21424 else if (k == REF || k == OPEN || k == CLOSE
21425 || k == GROUPP || OP(o)==ACCEPT)
21426 {
21427 AV *name_list= NULL;
21428 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
21429 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
21430 if ( RXp_PAREN_NAMES(prog) ) {
21431 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
21432 } else if ( pRExC_state ) {
21433 name_list= RExC_paren_name_list;
21434 }
21435 if (name_list) {
21436 if ( k != REF || (OP(o) < REFN)) {
21437 SV **name= av_fetch(name_list, parno, 0 );
21438 if (name)
21439 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21440 }
21441 else {
21442 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
21443 I32 *nums=(I32*)SvPVX(sv_dat);
21444 SV **name= av_fetch(name_list, nums[0], 0 );
21445 I32 n;
21446 if (name) {
21447 for ( n=0; n<SvIVX(sv_dat); n++ ) {
21448 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
21449 (n ? "," : ""), (IV)nums[n]);
21450 }
21451 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21452 }
21453 }
21454 }
21455 if ( k == REF && reginfo) {
21456 U32 n = ARG(o); /* which paren pair */
21457 I32 ln = prog->offs[n].start;
21458 if (prog->lastparen < n || ln == -1 || prog->offs[n].end == -1)
21459 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
21460 else if (ln == prog->offs[n].end)
21461 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
21462 else {
21463 const char *s = reginfo->strbeg + ln;
21464 Perl_sv_catpvf(aTHX_ sv, ": ");
21465 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
21466 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
21467 }
21468 }
21469 } else if (k == GOSUB) {
21470 AV *name_list= NULL;
21471 if ( RXp_PAREN_NAMES(prog) ) {
21472 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
21473 } else if ( pRExC_state ) {
21474 name_list= RExC_paren_name_list;
21475 }
21476
21477 /* Paren and offset */
21478 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
21479 (int)((o + (int)ARG2L(o)) - progi->program) );
21480 if (name_list) {
21481 SV **name= av_fetch(name_list, ARG(o), 0 );
21482 if (name)
21483 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
21484 }
21485 }
21486 else if (k == LOGICAL)
21487 /* 2: embedded, otherwise 1 */
21488 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
21489 else if (k == ANYOF || k == ANYOFR) {
21490 U8 flags;
21491 char * bitmap;
21492 U32 arg;
21493 bool do_sep = FALSE; /* Do we need to separate various components of
21494 the output? */
21495 /* Set if there is still an unresolved user-defined property */
21496 SV *unresolved = NULL;
21497
21498 /* Things that are ignored except when the runtime locale is UTF-8 */
21499 SV *only_utf8_locale_invlist = NULL;
21500
21501 /* Code points that don't fit in the bitmap */
21502 SV *nonbitmap_invlist = NULL;
21503
21504 /* And things that aren't in the bitmap, but are small enough to be */
21505 SV* bitmap_range_not_in_bitmap = NULL;
21506
21507 bool inverted;
21508
21509 if (inRANGE(OP(o), ANYOFH, ANYOFRb)) {
21510 flags = 0;
21511 bitmap = NULL;
21512 arg = 0;
21513 }
21514 else {
21515 flags = ANYOF_FLAGS(o);
21516 bitmap = ANYOF_BITMAP(o);
21517 arg = ARG(o);
21518 }
21519
21520 if (OP(o) == ANYOFL || OP(o) == ANYOFPOSIXL) {
21521 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
21522 sv_catpvs(sv, "{utf8-locale-reqd}");
21523 }
21524 if (flags & ANYOFL_FOLD) {
21525 sv_catpvs(sv, "{i}");
21526 }
21527 }
21528
21529 inverted = flags & ANYOF_INVERT;
21530
21531 /* If there is stuff outside the bitmap, get it */
21532 if (arg != ANYOF_ONLY_HAS_BITMAP) {
21533 if (inRANGE(OP(o), ANYOFR, ANYOFRb)) {
21534 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
21535 ANYOFRbase(o),
21536 ANYOFRbase(o) + ANYOFRdelta(o));
21537 }
21538 else {
21539 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
21540 (void) get_regclass_nonbitmap_data(prog, o, FALSE,
21541 &unresolved,
21542 &only_utf8_locale_invlist,
21543 &nonbitmap_invlist);
21544 #else
21545 (void) get_re_gclass_nonbitmap_data(prog, o, FALSE,
21546 &unresolved,
21547 &only_utf8_locale_invlist,
21548 &nonbitmap_invlist);
21549 #endif
21550 }
21551
21552 /* The non-bitmap data may contain stuff that could fit in the
21553 * bitmap. This could come from a user-defined property being
21554 * finally resolved when this call was done; or much more likely
21555 * because there are matches that require UTF-8 to be valid, and so
21556 * aren't in the bitmap (or ANYOFR). This is teased apart later */
21557 _invlist_intersection(nonbitmap_invlist,
21558 PL_InBitmap,
21559 &bitmap_range_not_in_bitmap);
21560 /* Leave just the things that don't fit into the bitmap */
21561 _invlist_subtract(nonbitmap_invlist,
21562 PL_InBitmap,
21563 &nonbitmap_invlist);
21564 }
21565
21566 /* Obey this flag to add all above-the-bitmap code points */
21567 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
21568 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
21569 NUM_ANYOF_CODE_POINTS,
21570 UV_MAX);
21571 }
21572
21573 /* Ready to start outputting. First, the initial left bracket */
21574 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
21575
21576 /* ANYOFH by definition doesn't have anything that will fit inside the
21577 * bitmap; ANYOFR may or may not. */
21578 if ( ! inRANGE(OP(o), ANYOFH, ANYOFHr)
21579 && ( ! inRANGE(OP(o), ANYOFR, ANYOFRb)
21580 || ANYOFRbase(o) < NUM_ANYOF_CODE_POINTS))
21581 {
21582 /* Then all the things that could fit in the bitmap */
21583 do_sep = put_charclass_bitmap_innards(sv,
21584 bitmap,
21585 bitmap_range_not_in_bitmap,
21586 only_utf8_locale_invlist,
21587 o,
21588 flags,
21589
21590 /* Can't try inverting for a
21591 * better display if there
21592 * are things that haven't
21593 * been resolved */
21594 unresolved != NULL
21595 || inRANGE(OP(o), ANYOFR, ANYOFRb));
21596 SvREFCNT_dec(bitmap_range_not_in_bitmap);
21597
21598 /* If there are user-defined properties which haven't been defined
21599 * yet, output them. If the result is not to be inverted, it is
21600 * clearest to output them in a separate [] from the bitmap range
21601 * stuff. If the result is to be complemented, we have to show
21602 * everything in one [], as the inversion applies to the whole
21603 * thing. Use {braces} to separate them from anything in the
21604 * bitmap and anything above the bitmap. */
21605 if (unresolved) {
21606 if (inverted) {
21607 if (! do_sep) { /* If didn't output anything in the bitmap
21608 */
21609 sv_catpvs(sv, "^");
21610 }
21611 sv_catpvs(sv, "{");
21612 }
21613 else if (do_sep) {
21614 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1],
21615 PL_colors[0]);
21616 }
21617 sv_catsv(sv, unresolved);
21618 if (inverted) {
21619 sv_catpvs(sv, "}");
21620 }
21621 do_sep = ! inverted;
21622 }
21623 }
21624
21625 /* And, finally, add the above-the-bitmap stuff */
21626 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
21627 SV* contents;
21628
21629 /* See if truncation size is overridden */
21630 const STRLEN dump_len = (PL_dump_re_max_len > 256)
21631 ? PL_dump_re_max_len
21632 : 256;
21633
21634 /* This is output in a separate [] */
21635 if (do_sep) {
21636 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1], PL_colors[0]);
21637 }
21638
21639 /* And, for easy of understanding, it is shown in the
21640 * uncomplemented form if possible. The one exception being if
21641 * there are unresolved items, where the inversion has to be
21642 * delayed until runtime */
21643 if (inverted && ! unresolved) {
21644 _invlist_invert(nonbitmap_invlist);
21645 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
21646 }
21647
21648 contents = invlist_contents(nonbitmap_invlist,
21649 FALSE /* output suitable for catsv */
21650 );
21651
21652 /* If the output is shorter than the permissible maximum, just do it. */
21653 if (SvCUR(contents) <= dump_len) {
21654 sv_catsv(sv, contents);
21655 }
21656 else {
21657 const char * contents_string = SvPVX(contents);
21658 STRLEN i = dump_len;
21659
21660 /* Otherwise, start at the permissible max and work back to the
21661 * first break possibility */
21662 while (i > 0 && contents_string[i] != ' ') {
21663 i--;
21664 }
21665 if (i == 0) { /* Fail-safe. Use the max if we couldn't
21666 find a legal break */
21667 i = dump_len;
21668 }
21669
21670 sv_catpvn(sv, contents_string, i);
21671 sv_catpvs(sv, "...");
21672 }
21673
21674 SvREFCNT_dec_NN(contents);
21675 SvREFCNT_dec_NN(nonbitmap_invlist);
21676 }
21677
21678 /* And finally the matching, closing ']' */
21679 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
21680
21681 if (OP(o) == ANYOFHs) {
21682 Perl_sv_catpvf(aTHX_ sv, " (Leading UTF-8 bytes=%s", _byte_dump_string((U8 *) ((struct regnode_anyofhs *) o)->string, FLAGS(o), 1));
21683 }
21684 else if (inRANGE(OP(o), ANYOFH, ANYOFRb)) {
21685 U8 lowest = (OP(o) != ANYOFHr)
21686 ? FLAGS(o)
21687 : LOWEST_ANYOF_HRx_BYTE(FLAGS(o));
21688 U8 highest = (OP(o) == ANYOFHr)
21689 ? HIGHEST_ANYOF_HRx_BYTE(FLAGS(o))
21690 : (OP(o) == ANYOFH || OP(o) == ANYOFR)
21691 ? 0xFF
21692 : lowest;
21693 #ifndef EBCDIC
21694 if (OP(o) != ANYOFR || ! isASCII(ANYOFRbase(o) + ANYOFRdelta(o)))
21695 #endif
21696 {
21697 Perl_sv_catpvf(aTHX_ sv, " (First UTF-8 byte=%02X", lowest);
21698 if (lowest != highest) {
21699 Perl_sv_catpvf(aTHX_ sv, "-%02X", highest);
21700 }
21701 Perl_sv_catpvf(aTHX_ sv, ")");
21702 }
21703 }
21704
21705 SvREFCNT_dec(unresolved);
21706 }
21707 else if (k == ANYOFM) {
21708 SV * cp_list = get_ANYOFM_contents(o);
21709
21710 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
21711 if (OP(o) == NANYOFM) {
21712 _invlist_invert(cp_list);
21713 }
21714
21715 put_charclass_bitmap_innards(sv, NULL, cp_list, NULL, NULL, 0, TRUE);
21716 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
21717
21718 SvREFCNT_dec(cp_list);
21719 }
21720 else if (k == POSIXD || k == NPOSIXD) {
21721 U8 index = FLAGS(o) * 2;
21722 if (index < C_ARRAY_LENGTH(anyofs)) {
21723 if (*anyofs[index] != '[') {
21724 sv_catpvs(sv, "[");
21725 }
21726 sv_catpv(sv, anyofs[index]);
21727 if (*anyofs[index] != '[') {
21728 sv_catpvs(sv, "]");
21729 }
21730 }
21731 else {
21732 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
21733 }
21734 }
21735 else if (k == BOUND || k == NBOUND) {
21736 /* Must be synced with order of 'bound_type' in regcomp.h */
21737 const char * const bounds[] = {
21738 "", /* Traditional */
21739 "{gcb}",
21740 "{lb}",
21741 "{sb}",
21742 "{wb}"
21743 };
21744 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
21745 sv_catpv(sv, bounds[FLAGS(o)]);
21746 }
21747 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH)) {
21748 Perl_sv_catpvf(aTHX_ sv, "[%d", -(o->flags));
21749 if (o->next_off) {
21750 Perl_sv_catpvf(aTHX_ sv, "..-%d", o->flags - o->next_off);
21751 }
21752 Perl_sv_catpvf(aTHX_ sv, "]");
21753 }
21754 else if (OP(o) == SBOL)
21755 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
21756
21757 /* add on the verb argument if there is one */
21758 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
21759 if ( ARG(o) )
21760 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
21761 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
21762 else
21763 sv_catpvs(sv, ":NULL");
21764 }
21765 #else
21766 PERL_UNUSED_CONTEXT;
21767 PERL_UNUSED_ARG(sv);
21768 PERL_UNUSED_ARG(o);
21769 PERL_UNUSED_ARG(prog);
21770 PERL_UNUSED_ARG(reginfo);
21771 PERL_UNUSED_ARG(pRExC_state);
21772 #endif /* DEBUGGING */
21773 }
21774
21775
21776
21777 SV *
Perl_re_intuit_string(pTHX_ REGEXP * const r)21778 Perl_re_intuit_string(pTHX_ REGEXP * const r)
21779 { /* Assume that RE_INTUIT is set */
21780 /* Returns an SV containing a string that must appear in the target for it
21781 * to match, or NULL if nothing is known that must match.
21782 *
21783 * CAUTION: the SV can be freed during execution of the regex engine */
21784
21785 struct regexp *const prog = ReANY(r);
21786 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21787
21788 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
21789 PERL_UNUSED_CONTEXT;
21790
21791 DEBUG_COMPILE_r(
21792 {
21793 if (prog->maxlen > 0) {
21794 const char * const s = SvPV_nolen_const(RX_UTF8(r)
21795 ? prog->check_utf8 : prog->check_substr);
21796
21797 if (!PL_colorset) reginitcolors();
21798 Perl_re_printf( aTHX_
21799 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
21800 PL_colors[4],
21801 RX_UTF8(r) ? "utf8 " : "",
21802 PL_colors[5], PL_colors[0],
21803 s,
21804 PL_colors[1],
21805 (strlen(s) > PL_dump_re_max_len ? "..." : ""));
21806 }
21807 } );
21808
21809 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
21810 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
21811 }
21812
21813 /*
21814 pregfree()
21815
21816 handles refcounting and freeing the perl core regexp structure. When
21817 it is necessary to actually free the structure the first thing it
21818 does is call the 'free' method of the regexp_engine associated to
21819 the regexp, allowing the handling of the void *pprivate; member
21820 first. (This routine is not overridable by extensions, which is why
21821 the extensions free is called first.)
21822
21823 See regdupe and regdupe_internal if you change anything here.
21824 */
21825 #ifndef PERL_IN_XSUB_RE
21826 void
Perl_pregfree(pTHX_ REGEXP * r)21827 Perl_pregfree(pTHX_ REGEXP *r)
21828 {
21829 SvREFCNT_dec(r);
21830 }
21831
21832 void
Perl_pregfree2(pTHX_ REGEXP * rx)21833 Perl_pregfree2(pTHX_ REGEXP *rx)
21834 {
21835 struct regexp *const r = ReANY(rx);
21836 DECLARE_AND_GET_RE_DEBUG_FLAGS;
21837
21838 PERL_ARGS_ASSERT_PREGFREE2;
21839
21840 if (! r)
21841 return;
21842
21843 if (r->mother_re) {
21844 ReREFCNT_dec(r->mother_re);
21845 } else {
21846 CALLREGFREE_PVT(rx); /* free the private data */
21847 SvREFCNT_dec(RXp_PAREN_NAMES(r));
21848 }
21849 if (r->substrs) {
21850 int i;
21851 for (i = 0; i < 2; i++) {
21852 SvREFCNT_dec(r->substrs->data[i].substr);
21853 SvREFCNT_dec(r->substrs->data[i].utf8_substr);
21854 }
21855 Safefree(r->substrs);
21856 }
21857 RX_MATCH_COPY_FREE(rx);
21858 #ifdef PERL_ANY_COW
21859 SvREFCNT_dec(r->saved_copy);
21860 #endif
21861 Safefree(r->offs);
21862 SvREFCNT_dec(r->qr_anoncv);
21863 if (r->recurse_locinput)
21864 Safefree(r->recurse_locinput);
21865 }
21866
21867
21868 /* reg_temp_copy()
21869
21870 Copy ssv to dsv, both of which should of type SVt_REGEXP or SVt_PVLV,
21871 except that dsv will be created if NULL.
21872
21873 This function is used in two main ways. First to implement
21874 $r = qr/....; $s = $$r;
21875
21876 Secondly, it is used as a hacky workaround to the structural issue of
21877 match results
21878 being stored in the regexp structure which is in turn stored in
21879 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
21880 could be PL_curpm in multiple contexts, and could require multiple
21881 result sets being associated with the pattern simultaneously, such
21882 as when doing a recursive match with (??{$qr})
21883
21884 The solution is to make a lightweight copy of the regexp structure
21885 when a qr// is returned from the code executed by (??{$qr}) this
21886 lightweight copy doesn't actually own any of its data except for
21887 the starp/end and the actual regexp structure itself.
21888
21889 */
21890
21891
21892 REGEXP *
Perl_reg_temp_copy(pTHX_ REGEXP * dsv,REGEXP * ssv)21893 Perl_reg_temp_copy(pTHX_ REGEXP *dsv, REGEXP *ssv)
21894 {
21895 struct regexp *drx;
21896 struct regexp *const srx = ReANY(ssv);
21897 const bool islv = dsv && SvTYPE(dsv) == SVt_PVLV;
21898
21899 PERL_ARGS_ASSERT_REG_TEMP_COPY;
21900
21901 if (!dsv)
21902 dsv = (REGEXP*) newSV_type(SVt_REGEXP);
21903 else {
21904 assert(SvTYPE(dsv) == SVt_REGEXP || (SvTYPE(dsv) == SVt_PVLV));
21905
21906 /* our only valid caller, sv_setsv_flags(), should have done
21907 * a SV_CHECK_THINKFIRST_COW_DROP() by now */
21908 assert(!SvOOK(dsv));
21909 assert(!SvIsCOW(dsv));
21910 assert(!SvROK(dsv));
21911
21912 if (SvPVX_const(dsv)) {
21913 if (SvLEN(dsv))
21914 Safefree(SvPVX(dsv));
21915 SvPVX(dsv) = NULL;
21916 }
21917 SvLEN_set(dsv, 0);
21918 SvCUR_set(dsv, 0);
21919 SvOK_off((SV *)dsv);
21920
21921 if (islv) {
21922 /* For PVLVs, the head (sv_any) points to an XPVLV, while
21923 * the LV's xpvlenu_rx will point to a regexp body, which
21924 * we allocate here */
21925 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
21926 assert(!SvPVX(dsv));
21927 /* We "steal" the body from the newly allocated SV temp, changing
21928 * the pointer in its HEAD to NULL. We then change its type to
21929 * SVt_NULL so that when we immediately release its only reference,
21930 * no memory deallocation happens.
21931 *
21932 * The body will eventually be freed (from the PVLV) either in
21933 * Perl_sv_force_normal_flags() (if the PVLV is "downgraded" and
21934 * the regexp body needs to be removed)
21935 * or in Perl_sv_clear() (if the PVLV still holds the pointer until
21936 * the PVLV itself is deallocated). */
21937 ((XPV*)SvANY(dsv))->xpv_len_u.xpvlenu_rx = temp->sv_any;
21938 temp->sv_any = NULL;
21939 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
21940 SvREFCNT_dec_NN(temp);
21941 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
21942 ing below will not set it. */
21943 SvCUR_set(dsv, SvCUR(ssv));
21944 }
21945 }
21946 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
21947 sv_force_normal(sv) is called. */
21948 SvFAKE_on(dsv);
21949 drx = ReANY(dsv);
21950
21951 SvFLAGS(dsv) |= SvFLAGS(ssv) & (SVf_POK|SVp_POK|SVf_UTF8);
21952 SvPV_set(dsv, RX_WRAPPED(ssv));
21953 /* We share the same string buffer as the original regexp, on which we
21954 hold a reference count, incremented when mother_re is set below.
21955 The string pointer is copied here, being part of the regexp struct.
21956 */
21957 memcpy(&(drx->xpv_cur), &(srx->xpv_cur),
21958 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
21959 if (!islv)
21960 SvLEN_set(dsv, 0);
21961 if (srx->offs) {
21962 const I32 npar = srx->nparens+1;
21963 Newx(drx->offs, npar, regexp_paren_pair);
21964 Copy(srx->offs, drx->offs, npar, regexp_paren_pair);
21965 }
21966 if (srx->substrs) {
21967 int i;
21968 Newx(drx->substrs, 1, struct reg_substr_data);
21969 StructCopy(srx->substrs, drx->substrs, struct reg_substr_data);
21970
21971 for (i = 0; i < 2; i++) {
21972 SvREFCNT_inc_void(drx->substrs->data[i].substr);
21973 SvREFCNT_inc_void(drx->substrs->data[i].utf8_substr);
21974 }
21975
21976 /* check_substr and check_utf8, if non-NULL, point to either their
21977 anchored or float namesakes, and don't hold a second reference. */
21978 }
21979 RX_MATCH_COPIED_off(dsv);
21980 #ifdef PERL_ANY_COW
21981 drx->saved_copy = NULL;
21982 #endif
21983 drx->mother_re = ReREFCNT_inc(srx->mother_re ? srx->mother_re : ssv);
21984 SvREFCNT_inc_void(drx->qr_anoncv);
21985 if (srx->recurse_locinput)
21986 Newx(drx->recurse_locinput, srx->nparens + 1, char *);
21987
21988 return dsv;
21989 }
21990 #endif
21991
21992
21993 /* regfree_internal()
21994
21995 Free the private data in a regexp. This is overloadable by
21996 extensions. Perl takes care of the regexp structure in pregfree(),
21997 this covers the *pprivate pointer which technically perl doesn't
21998 know about, however of course we have to handle the
21999 regexp_internal structure when no extension is in use.
22000
22001 Note this is called before freeing anything in the regexp
22002 structure.
22003 */
22004
22005 void
Perl_regfree_internal(pTHX_ REGEXP * const rx)22006 Perl_regfree_internal(pTHX_ REGEXP * const rx)
22007 {
22008 struct regexp *const r = ReANY(rx);
22009 RXi_GET_DECL(r, ri);
22010 DECLARE_AND_GET_RE_DEBUG_FLAGS;
22011
22012 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
22013
22014 if (! ri) {
22015 return;
22016 }
22017
22018 DEBUG_COMPILE_r({
22019 if (!PL_colorset)
22020 reginitcolors();
22021 {
22022 SV *dsv= sv_newmortal();
22023 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
22024 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), PL_dump_re_max_len);
22025 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
22026 PL_colors[4], PL_colors[5], s);
22027 }
22028 });
22029
22030 #ifdef RE_TRACK_PATTERN_OFFSETS
22031 if (ri->u.offsets)
22032 Safefree(ri->u.offsets); /* 20010421 MJD */
22033 #endif
22034 if (ri->code_blocks)
22035 S_free_codeblocks(aTHX_ ri->code_blocks);
22036
22037 if (ri->data) {
22038 int n = ri->data->count;
22039
22040 while (--n >= 0) {
22041 /* If you add a ->what type here, update the comment in regcomp.h */
22042 switch (ri->data->what[n]) {
22043 case 'a':
22044 case 'r':
22045 case 's':
22046 case 'S':
22047 case 'u':
22048 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
22049 break;
22050 case 'f':
22051 Safefree(ri->data->data[n]);
22052 break;
22053 case 'l':
22054 case 'L':
22055 break;
22056 case 'T':
22057 { /* Aho Corasick add-on structure for a trie node.
22058 Used in stclass optimization only */
22059 U32 refcount;
22060 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
22061 OP_REFCNT_LOCK;
22062 refcount = --aho->refcount;
22063 OP_REFCNT_UNLOCK;
22064 if ( !refcount ) {
22065 PerlMemShared_free(aho->states);
22066 PerlMemShared_free(aho->fail);
22067 /* do this last!!!! */
22068 PerlMemShared_free(ri->data->data[n]);
22069 /* we should only ever get called once, so
22070 * assert as much, and also guard the free
22071 * which /might/ happen twice. At the least
22072 * it will make code anlyzers happy and it
22073 * doesn't cost much. - Yves */
22074 assert(ri->regstclass);
22075 if (ri->regstclass) {
22076 PerlMemShared_free(ri->regstclass);
22077 ri->regstclass = 0;
22078 }
22079 }
22080 }
22081 break;
22082 case 't':
22083 {
22084 /* trie structure. */
22085 U32 refcount;
22086 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
22087 OP_REFCNT_LOCK;
22088 refcount = --trie->refcount;
22089 OP_REFCNT_UNLOCK;
22090 if ( !refcount ) {
22091 PerlMemShared_free(trie->charmap);
22092 PerlMemShared_free(trie->states);
22093 PerlMemShared_free(trie->trans);
22094 if (trie->bitmap)
22095 PerlMemShared_free(trie->bitmap);
22096 if (trie->jump)
22097 PerlMemShared_free(trie->jump);
22098 PerlMemShared_free(trie->wordinfo);
22099 /* do this last!!!! */
22100 PerlMemShared_free(ri->data->data[n]);
22101 }
22102 }
22103 break;
22104 default:
22105 Perl_croak(aTHX_ "panic: regfree data code '%c'",
22106 ri->data->what[n]);
22107 }
22108 }
22109 Safefree(ri->data->what);
22110 Safefree(ri->data);
22111 }
22112
22113 Safefree(ri);
22114 }
22115
22116 #define av_dup_inc(s, t) MUTABLE_AV(sv_dup_inc((const SV *)s, t))
22117 #define hv_dup_inc(s, t) MUTABLE_HV(sv_dup_inc((const SV *)s, t))
22118 #define SAVEPVN(p, n) ((p) ? savepvn(p, n) : NULL)
22119
22120 /*
22121 =for apidoc re_dup_guts
22122 Duplicate a regexp.
22123
22124 This routine is expected to clone a given regexp structure. It is only
22125 compiled under USE_ITHREADS.
22126
22127 After all of the core data stored in struct regexp is duplicated
22128 the C<regexp_engine.dupe> method is used to copy any private data
22129 stored in the *pprivate pointer. This allows extensions to handle
22130 any duplication they need to do.
22131
22132 =cut
22133
22134 See pregfree() and regfree_internal() if you change anything here.
22135 */
22136 #if defined(USE_ITHREADS)
22137 #ifndef PERL_IN_XSUB_RE
22138 void
Perl_re_dup_guts(pTHX_ const REGEXP * sstr,REGEXP * dstr,CLONE_PARAMS * param)22139 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
22140 {
22141 I32 npar;
22142 const struct regexp *r = ReANY(sstr);
22143 struct regexp *ret = ReANY(dstr);
22144
22145 PERL_ARGS_ASSERT_RE_DUP_GUTS;
22146
22147 npar = r->nparens+1;
22148 Newx(ret->offs, npar, regexp_paren_pair);
22149 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
22150
22151 if (ret->substrs) {
22152 /* Do it this way to avoid reading from *r after the StructCopy().
22153 That way, if any of the sv_dup_inc()s dislodge *r from the L1
22154 cache, it doesn't matter. */
22155 int i;
22156 const bool anchored = r->check_substr
22157 ? r->check_substr == r->substrs->data[0].substr
22158 : r->check_utf8 == r->substrs->data[0].utf8_substr;
22159 Newx(ret->substrs, 1, struct reg_substr_data);
22160 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
22161
22162 for (i = 0; i < 2; i++) {
22163 ret->substrs->data[i].substr =
22164 sv_dup_inc(ret->substrs->data[i].substr, param);
22165 ret->substrs->data[i].utf8_substr =
22166 sv_dup_inc(ret->substrs->data[i].utf8_substr, param);
22167 }
22168
22169 /* check_substr and check_utf8, if non-NULL, point to either their
22170 anchored or float namesakes, and don't hold a second reference. */
22171
22172 if (ret->check_substr) {
22173 if (anchored) {
22174 assert(r->check_utf8 == r->substrs->data[0].utf8_substr);
22175
22176 ret->check_substr = ret->substrs->data[0].substr;
22177 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
22178 } else {
22179 assert(r->check_substr == r->substrs->data[1].substr);
22180 assert(r->check_utf8 == r->substrs->data[1].utf8_substr);
22181
22182 ret->check_substr = ret->substrs->data[1].substr;
22183 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
22184 }
22185 } else if (ret->check_utf8) {
22186 if (anchored) {
22187 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
22188 } else {
22189 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
22190 }
22191 }
22192 }
22193
22194 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
22195 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
22196 if (r->recurse_locinput)
22197 Newx(ret->recurse_locinput, r->nparens + 1, char *);
22198
22199 if (ret->pprivate)
22200 RXi_SET(ret, CALLREGDUPE_PVT(dstr, param));
22201
22202 if (RX_MATCH_COPIED(dstr))
22203 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
22204 else
22205 ret->subbeg = NULL;
22206 #ifdef PERL_ANY_COW
22207 ret->saved_copy = NULL;
22208 #endif
22209
22210 /* Whether mother_re be set or no, we need to copy the string. We
22211 cannot refrain from copying it when the storage points directly to
22212 our mother regexp, because that's
22213 1: a buffer in a different thread
22214 2: something we no longer hold a reference on
22215 so we need to copy it locally. */
22216 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED_const(sstr), SvCUR(sstr)+1);
22217 /* set malloced length to a non-zero value so it will be freed
22218 * (otherwise in combination with SVf_FAKE it looks like an alien
22219 * buffer). It doesn't have to be the actual malloced size, since it
22220 * should never be grown */
22221 SvLEN_set(dstr, SvCUR(sstr)+1);
22222 ret->mother_re = NULL;
22223 }
22224 #endif /* PERL_IN_XSUB_RE */
22225
22226 /*
22227 regdupe_internal()
22228
22229 This is the internal complement to regdupe() which is used to copy
22230 the structure pointed to by the *pprivate pointer in the regexp.
22231 This is the core version of the extension overridable cloning hook.
22232 The regexp structure being duplicated will be copied by perl prior
22233 to this and will be provided as the regexp *r argument, however
22234 with the /old/ structures pprivate pointer value. Thus this routine
22235 may override any copying normally done by perl.
22236
22237 It returns a pointer to the new regexp_internal structure.
22238 */
22239
22240 void *
Perl_regdupe_internal(pTHX_ REGEXP * const rx,CLONE_PARAMS * param)22241 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
22242 {
22243 struct regexp *const r = ReANY(rx);
22244 regexp_internal *reti;
22245 int len;
22246 RXi_GET_DECL(r, ri);
22247
22248 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
22249
22250 len = ProgLen(ri);
22251
22252 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
22253 char, regexp_internal);
22254 Copy(ri->program, reti->program, len+1, regnode);
22255
22256
22257 if (ri->code_blocks) {
22258 int n;
22259 Newx(reti->code_blocks, 1, struct reg_code_blocks);
22260 Newx(reti->code_blocks->cb, ri->code_blocks->count,
22261 struct reg_code_block);
22262 Copy(ri->code_blocks->cb, reti->code_blocks->cb,
22263 ri->code_blocks->count, struct reg_code_block);
22264 for (n = 0; n < ri->code_blocks->count; n++)
22265 reti->code_blocks->cb[n].src_regex = (REGEXP*)
22266 sv_dup_inc((SV*)(ri->code_blocks->cb[n].src_regex), param);
22267 reti->code_blocks->count = ri->code_blocks->count;
22268 reti->code_blocks->refcnt = 1;
22269 }
22270 else
22271 reti->code_blocks = NULL;
22272
22273 reti->regstclass = NULL;
22274
22275 if (ri->data) {
22276 struct reg_data *d;
22277 const int count = ri->data->count;
22278 int i;
22279
22280 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
22281 char, struct reg_data);
22282 Newx(d->what, count, U8);
22283
22284 d->count = count;
22285 for (i = 0; i < count; i++) {
22286 d->what[i] = ri->data->what[i];
22287 switch (d->what[i]) {
22288 /* see also regcomp.h and regfree_internal() */
22289 case 'a': /* actually an AV, but the dup function is identical.
22290 values seem to be "plain sv's" generally. */
22291 case 'r': /* a compiled regex (but still just another SV) */
22292 case 's': /* an RV (currently only used for an RV to an AV by the ANYOF code)
22293 this use case should go away, the code could have used
22294 'a' instead - see S_set_ANYOF_arg() for array contents. */
22295 case 'S': /* actually an SV, but the dup function is identical. */
22296 case 'u': /* actually an HV, but the dup function is identical.
22297 values are "plain sv's" */
22298 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
22299 break;
22300 case 'f':
22301 /* Synthetic Start Class - "Fake" charclass we generate to optimize
22302 * patterns which could start with several different things. Pre-TRIE
22303 * this was more important than it is now, however this still helps
22304 * in some places, for instance /x?a+/ might produce a SSC equivalent
22305 * to [xa]. This is used by Perl_re_intuit_start() and S_find_byclass()
22306 * in regexec.c
22307 */
22308 /* This is cheating. */
22309 Newx(d->data[i], 1, regnode_ssc);
22310 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
22311 reti->regstclass = (regnode*)d->data[i];
22312 break;
22313 case 'T':
22314 /* AHO-CORASICK fail table */
22315 /* Trie stclasses are readonly and can thus be shared
22316 * without duplication. We free the stclass in pregfree
22317 * when the corresponding reg_ac_data struct is freed.
22318 */
22319 reti->regstclass= ri->regstclass;
22320 /* FALLTHROUGH */
22321 case 't':
22322 /* TRIE transition table */
22323 OP_REFCNT_LOCK;
22324 ((reg_trie_data*)ri->data->data[i])->refcount++;
22325 OP_REFCNT_UNLOCK;
22326 /* FALLTHROUGH */
22327 case 'l': /* (?{...}) or (??{ ... }) code (cb->block) */
22328 case 'L': /* same when RExC_pm_flags & PMf_HAS_CV and code
22329 is not from another regexp */
22330 d->data[i] = ri->data->data[i];
22331 break;
22332 default:
22333 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
22334 ri->data->what[i]);
22335 }
22336 }
22337
22338 reti->data = d;
22339 }
22340 else
22341 reti->data = NULL;
22342
22343 reti->name_list_idx = ri->name_list_idx;
22344
22345 #ifdef RE_TRACK_PATTERN_OFFSETS
22346 if (ri->u.offsets) {
22347 Newx(reti->u.offsets, 2*len+1, U32);
22348 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
22349 }
22350 #else
22351 SetProgLen(reti, len);
22352 #endif
22353
22354 return (void*)reti;
22355 }
22356
22357 #endif /* USE_ITHREADS */
22358
22359 #ifndef PERL_IN_XSUB_RE
22360
22361 /*
22362 - regnext - dig the "next" pointer out of a node
22363 */
22364 regnode *
Perl_regnext(pTHX_ regnode * p)22365 Perl_regnext(pTHX_ regnode *p)
22366 {
22367 I32 offset;
22368
22369 if (!p)
22370 return(NULL);
22371
22372 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
22373 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
22374 (int)OP(p), (int)REGNODE_MAX);
22375 }
22376
22377 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
22378 if (offset == 0)
22379 return(NULL);
22380
22381 return(p+offset);
22382 }
22383
22384 #endif
22385
22386 STATIC void
S_re_croak(pTHX_ bool utf8,const char * pat,...)22387 S_re_croak(pTHX_ bool utf8, const char* pat,...)
22388 {
22389 va_list args;
22390 STRLEN len = strlen(pat);
22391 char buf[512];
22392 SV *msv;
22393 const char *message;
22394
22395 PERL_ARGS_ASSERT_RE_CROAK;
22396
22397 if (len > 510)
22398 len = 510;
22399 Copy(pat, buf, len , char);
22400 buf[len] = '\n';
22401 buf[len + 1] = '\0';
22402 va_start(args, pat);
22403 msv = vmess(buf, &args);
22404 va_end(args);
22405 message = SvPV_const(msv, len);
22406 if (len > 512)
22407 len = 512;
22408 Copy(message, buf, len , char);
22409 /* len-1 to avoid \n */
22410 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, len-1, buf));
22411 }
22412
22413 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
22414
22415 #ifndef PERL_IN_XSUB_RE
22416 void
Perl_save_re_context(pTHX)22417 Perl_save_re_context(pTHX)
22418 {
22419 I32 nparens = -1;
22420 I32 i;
22421
22422 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
22423
22424 if (PL_curpm) {
22425 const REGEXP * const rx = PM_GETRE(PL_curpm);
22426 if (rx)
22427 nparens = RX_NPARENS(rx);
22428 }
22429
22430 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
22431 * that PL_curpm will be null, but that utf8.pm and the modules it
22432 * loads will only use $1..$3.
22433 * The t/porting/re_context.t test file checks this assumption.
22434 */
22435 if (nparens == -1)
22436 nparens = 3;
22437
22438 for (i = 1; i <= nparens; i++) {
22439 char digits[TYPE_CHARS(long)];
22440 const STRLEN len = my_snprintf(digits, sizeof(digits),
22441 "%lu", (long)i);
22442 GV *const *const gvp
22443 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
22444
22445 if (gvp) {
22446 GV * const gv = *gvp;
22447 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
22448 save_scalar(gv);
22449 }
22450 }
22451 }
22452 #endif
22453
22454 #ifdef DEBUGGING
22455
22456 STATIC void
S_put_code_point(pTHX_ SV * sv,UV c)22457 S_put_code_point(pTHX_ SV *sv, UV c)
22458 {
22459 PERL_ARGS_ASSERT_PUT_CODE_POINT;
22460
22461 if (c > 255) {
22462 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
22463 }
22464 else if (isPRINT(c)) {
22465 const char string = (char) c;
22466
22467 /* We use {phrase} as metanotation in the class, so also escape literal
22468 * braces */
22469 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
22470 sv_catpvs(sv, "\\");
22471 sv_catpvn(sv, &string, 1);
22472 }
22473 else if (isMNEMONIC_CNTRL(c)) {
22474 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
22475 }
22476 else {
22477 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
22478 }
22479 }
22480
22481 STATIC void
S_put_range(pTHX_ SV * sv,UV start,const UV end,const bool allow_literals)22482 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
22483 {
22484 /* Appends to 'sv' a displayable version of the range of code points from
22485 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
22486 * that have them, when they occur at the beginning or end of the range.
22487 * It uses hex to output the remaining code points, unless 'allow_literals'
22488 * is true, in which case the printable ASCII ones are output as-is (though
22489 * some of these will be escaped by put_code_point()).
22490 *
22491 * NOTE: This is designed only for printing ranges of code points that fit
22492 * inside an ANYOF bitmap. Higher code points are simply suppressed
22493 */
22494
22495 const unsigned int min_range_count = 3;
22496
22497 assert(start <= end);
22498
22499 PERL_ARGS_ASSERT_PUT_RANGE;
22500
22501 while (start <= end) {
22502 UV this_end;
22503 const char * format;
22504
22505 if ( end - start < min_range_count
22506 && (end - start <= 2 || (isPRINT_A(start) && isPRINT_A(end))))
22507 {
22508 /* Output a range of 1 or 2 chars individually, or longer ranges
22509 * when printable */
22510 for (; start <= end; start++) {
22511 put_code_point(sv, start);
22512 }
22513 break;
22514 }
22515
22516 /* If permitted by the input options, and there is a possibility that
22517 * this range contains a printable literal, look to see if there is
22518 * one. */
22519 if (allow_literals && start <= MAX_PRINT_A) {
22520
22521 /* If the character at the beginning of the range isn't an ASCII
22522 * printable, effectively split the range into two parts:
22523 * 1) the portion before the first such printable,
22524 * 2) the rest
22525 * and output them separately. */
22526 if (! isPRINT_A(start)) {
22527 UV temp_end = start + 1;
22528
22529 /* There is no point looking beyond the final possible
22530 * printable, in MAX_PRINT_A */
22531 UV max = MIN(end, MAX_PRINT_A);
22532
22533 while (temp_end <= max && ! isPRINT_A(temp_end)) {
22534 temp_end++;
22535 }
22536
22537 /* Here, temp_end points to one beyond the first printable if
22538 * found, or to one beyond 'max' if not. If none found, make
22539 * sure that we use the entire range */
22540 if (temp_end > MAX_PRINT_A) {
22541 temp_end = end + 1;
22542 }
22543
22544 /* Output the first part of the split range: the part that
22545 * doesn't have printables, with the parameter set to not look
22546 * for literals (otherwise we would infinitely recurse) */
22547 put_range(sv, start, temp_end - 1, FALSE);
22548
22549 /* The 2nd part of the range (if any) starts here. */
22550 start = temp_end;
22551
22552 /* We do a continue, instead of dropping down, because even if
22553 * the 2nd part is non-empty, it could be so short that we want
22554 * to output it as individual characters, as tested for at the
22555 * top of this loop. */
22556 continue;
22557 }
22558
22559 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
22560 * output a sub-range of just the digits or letters, then process
22561 * the remaining portion as usual. */
22562 if (isALPHANUMERIC_A(start)) {
22563 UV mask = (isDIGIT_A(start))
22564 ? _CC_DIGIT
22565 : isUPPER_A(start)
22566 ? _CC_UPPER
22567 : _CC_LOWER;
22568 UV temp_end = start + 1;
22569
22570 /* Find the end of the sub-range that includes just the
22571 * characters in the same class as the first character in it */
22572 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
22573 temp_end++;
22574 }
22575 temp_end--;
22576
22577 /* For short ranges, don't duplicate the code above to output
22578 * them; just call recursively */
22579 if (temp_end - start < min_range_count) {
22580 put_range(sv, start, temp_end, FALSE);
22581 }
22582 else { /* Output as a range */
22583 put_code_point(sv, start);
22584 sv_catpvs(sv, "-");
22585 put_code_point(sv, temp_end);
22586 }
22587 start = temp_end + 1;
22588 continue;
22589 }
22590
22591 /* We output any other printables as individual characters */
22592 if (isPUNCT_A(start) || isSPACE_A(start)) {
22593 while (start <= end && (isPUNCT_A(start)
22594 || isSPACE_A(start)))
22595 {
22596 put_code_point(sv, start);
22597 start++;
22598 }
22599 continue;
22600 }
22601 } /* End of looking for literals */
22602
22603 /* Here is not to output as a literal. Some control characters have
22604 * mnemonic names. Split off any of those at the beginning and end of
22605 * the range to print mnemonically. It isn't possible for many of
22606 * these to be in a row, so this won't overwhelm with output */
22607 if ( start <= end
22608 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
22609 {
22610 while (isMNEMONIC_CNTRL(start) && start <= end) {
22611 put_code_point(sv, start);
22612 start++;
22613 }
22614
22615 /* If this didn't take care of the whole range ... */
22616 if (start <= end) {
22617
22618 /* Look backwards from the end to find the final non-mnemonic
22619 * */
22620 UV temp_end = end;
22621 while (isMNEMONIC_CNTRL(temp_end)) {
22622 temp_end--;
22623 }
22624
22625 /* And separately output the interior range that doesn't start
22626 * or end with mnemonics */
22627 put_range(sv, start, temp_end, FALSE);
22628
22629 /* Then output the mnemonic trailing controls */
22630 start = temp_end + 1;
22631 while (start <= end) {
22632 put_code_point(sv, start);
22633 start++;
22634 }
22635 break;
22636 }
22637 }
22638
22639 /* As a final resort, output the range or subrange as hex. */
22640
22641 if (start >= NUM_ANYOF_CODE_POINTS) {
22642 this_end = end;
22643 }
22644 else { /* Have to split range at the bitmap boundary */
22645 this_end = (end < NUM_ANYOF_CODE_POINTS)
22646 ? end
22647 : NUM_ANYOF_CODE_POINTS - 1;
22648 }
22649 #if NUM_ANYOF_CODE_POINTS > 256
22650 format = (this_end < 256)
22651 ? "\\x%02" UVXf "-\\x%02" UVXf
22652 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
22653 #else
22654 format = "\\x%02" UVXf "-\\x%02" UVXf;
22655 #endif
22656 GCC_DIAG_IGNORE_STMT(-Wformat-nonliteral);
22657 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
22658 GCC_DIAG_RESTORE_STMT;
22659 break;
22660 }
22661 }
22662
22663 STATIC void
S_put_charclass_bitmap_innards_invlist(pTHX_ SV * sv,SV * invlist)22664 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
22665 {
22666 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
22667 * 'invlist' */
22668
22669 UV start, end;
22670 bool allow_literals = TRUE;
22671
22672 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
22673
22674 /* Generally, it is more readable if printable characters are output as
22675 * literals, but if a range (nearly) spans all of them, it's best to output
22676 * it as a single range. This code will use a single range if all but 2
22677 * ASCII printables are in it */
22678 invlist_iterinit(invlist);
22679 while (invlist_iternext(invlist, &start, &end)) {
22680
22681 /* If the range starts beyond the final printable, it doesn't have any
22682 * in it */
22683 if (start > MAX_PRINT_A) {
22684 break;
22685 }
22686
22687 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
22688 * all but two, the range must start and end no later than 2 from
22689 * either end */
22690 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
22691 if (end > MAX_PRINT_A) {
22692 end = MAX_PRINT_A;
22693 }
22694 if (start < ' ') {
22695 start = ' ';
22696 }
22697 if (end - start >= MAX_PRINT_A - ' ' - 2) {
22698 allow_literals = FALSE;
22699 }
22700 break;
22701 }
22702 }
22703 invlist_iterfinish(invlist);
22704
22705 /* Here we have figured things out. Output each range */
22706 invlist_iterinit(invlist);
22707 while (invlist_iternext(invlist, &start, &end)) {
22708 if (start >= NUM_ANYOF_CODE_POINTS) {
22709 break;
22710 }
22711 put_range(sv, start, end, allow_literals);
22712 }
22713 invlist_iterfinish(invlist);
22714
22715 return;
22716 }
22717
22718 STATIC SV*
S_put_charclass_bitmap_innards_common(pTHX_ SV * invlist,SV * posixes,SV * only_utf8,SV * not_utf8,SV * only_utf8_locale,const bool invert)22719 S_put_charclass_bitmap_innards_common(pTHX_
22720 SV* invlist, /* The bitmap */
22721 SV* posixes, /* Under /l, things like [:word:], \S */
22722 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
22723 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
22724 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
22725 const bool invert /* Is the result to be inverted? */
22726 )
22727 {
22728 /* Create and return an SV containing a displayable version of the bitmap
22729 * and associated information determined by the input parameters. If the
22730 * output would have been only the inversion indicator '^', NULL is instead
22731 * returned. */
22732
22733 SV * output;
22734
22735 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
22736
22737 if (invert) {
22738 output = newSVpvs("^");
22739 }
22740 else {
22741 output = newSVpvs("");
22742 }
22743
22744 /* First, the code points in the bitmap that are unconditionally there */
22745 put_charclass_bitmap_innards_invlist(output, invlist);
22746
22747 /* Traditionally, these have been placed after the main code points */
22748 if (posixes) {
22749 sv_catsv(output, posixes);
22750 }
22751
22752 if (only_utf8 && _invlist_len(only_utf8)) {
22753 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
22754 put_charclass_bitmap_innards_invlist(output, only_utf8);
22755 }
22756
22757 if (not_utf8 && _invlist_len(not_utf8)) {
22758 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
22759 put_charclass_bitmap_innards_invlist(output, not_utf8);
22760 }
22761
22762 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
22763 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
22764 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
22765
22766 /* This is the only list in this routine that can legally contain code
22767 * points outside the bitmap range. The call just above to
22768 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
22769 * output them here. There's about a half-dozen possible, and none in
22770 * contiguous ranges longer than 2 */
22771 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
22772 UV start, end;
22773 SV* above_bitmap = NULL;
22774
22775 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
22776
22777 invlist_iterinit(above_bitmap);
22778 while (invlist_iternext(above_bitmap, &start, &end)) {
22779 UV i;
22780
22781 for (i = start; i <= end; i++) {
22782 put_code_point(output, i);
22783 }
22784 }
22785 invlist_iterfinish(above_bitmap);
22786 SvREFCNT_dec_NN(above_bitmap);
22787 }
22788 }
22789
22790 if (invert && SvCUR(output) == 1) {
22791 return NULL;
22792 }
22793
22794 return output;
22795 }
22796
22797 STATIC bool
S_put_charclass_bitmap_innards(pTHX_ SV * sv,char * bitmap,SV * nonbitmap_invlist,SV * only_utf8_locale_invlist,const regnode * const node,const U8 flags,const bool force_as_is_display)22798 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
22799 char *bitmap,
22800 SV *nonbitmap_invlist,
22801 SV *only_utf8_locale_invlist,
22802 const regnode * const node,
22803 const U8 flags,
22804 const bool force_as_is_display)
22805 {
22806 /* Appends to 'sv' a displayable version of the innards of the bracketed
22807 * character class defined by the other arguments:
22808 * 'bitmap' points to the bitmap, or NULL if to ignore that.
22809 * 'nonbitmap_invlist' is an inversion list of the code points that are in
22810 * the bitmap range, but for some reason aren't in the bitmap; NULL if
22811 * none. The reasons for this could be that they require some
22812 * condition such as the target string being or not being in UTF-8
22813 * (under /d), or because they came from a user-defined property that
22814 * was not resolved at the time of the regex compilation (under /u)
22815 * 'only_utf8_locale_invlist' is an inversion list of the code points that
22816 * are valid only if the runtime locale is a UTF-8 one; NULL if none
22817 * 'node' is the regex pattern ANYOF node. It is needed only when the
22818 * above two parameters are not null, and is passed so that this
22819 * routine can tease apart the various reasons for them.
22820 * 'flags' is the flags field of 'node'
22821 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
22822 * to invert things to see if that leads to a cleaner display. If
22823 * FALSE, this routine is free to use its judgment about doing this.
22824 *
22825 * It returns TRUE if there was actually something output. (It may be that
22826 * the bitmap, etc is empty.)
22827 *
22828 * When called for outputting the bitmap of a non-ANYOF node, just pass the
22829 * bitmap, with the succeeding parameters set to NULL, and the final one to
22830 * FALSE.
22831 */
22832
22833 /* In general, it tries to display the 'cleanest' representation of the
22834 * innards, choosing whether to display them inverted or not, regardless of
22835 * whether the class itself is to be inverted. However, there are some
22836 * cases where it can't try inverting, as what actually matches isn't known
22837 * until runtime, and hence the inversion isn't either. */
22838
22839 bool inverting_allowed = ! force_as_is_display;
22840
22841 int i;
22842 STRLEN orig_sv_cur = SvCUR(sv);
22843
22844 SV* invlist; /* Inversion list we accumulate of code points that
22845 are unconditionally matched */
22846 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
22847 UTF-8 */
22848 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
22849 */
22850 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
22851 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
22852 is UTF-8 */
22853
22854 SV* as_is_display; /* The output string when we take the inputs
22855 literally */
22856 SV* inverted_display; /* The output string when we invert the inputs */
22857
22858 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
22859 to match? */
22860 /* We are biased in favor of displaying things without them being inverted,
22861 * as that is generally easier to understand */
22862 const int bias = 5;
22863
22864 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
22865
22866 /* Start off with whatever code points are passed in. (We clone, so we
22867 * don't change the caller's list) */
22868 if (nonbitmap_invlist) {
22869 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
22870 invlist = invlist_clone(nonbitmap_invlist, NULL);
22871 }
22872 else { /* Worst case size is every other code point is matched */
22873 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
22874 }
22875
22876 if (flags) {
22877 if (OP(node) == ANYOFD) {
22878
22879 /* This flag indicates that the code points below 0x100 in the
22880 * nonbitmap list are precisely the ones that match only when the
22881 * target is UTF-8 (they should all be non-ASCII). */
22882 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
22883 {
22884 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
22885 _invlist_subtract(invlist, only_utf8, &invlist);
22886 }
22887
22888 /* And this flag for matching all non-ASCII 0xFF and below */
22889 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
22890 {
22891 not_utf8 = invlist_clone(PL_UpperLatin1, NULL);
22892 }
22893 }
22894 else if (OP(node) == ANYOFL || OP(node) == ANYOFPOSIXL) {
22895
22896 /* If either of these flags are set, what matches isn't
22897 * determinable except during execution, so don't know enough here
22898 * to invert */
22899 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
22900 inverting_allowed = FALSE;
22901 }
22902
22903 /* What the posix classes match also varies at runtime, so these
22904 * will be output symbolically. */
22905 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
22906 int i;
22907
22908 posixes = newSVpvs("");
22909 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
22910 if (ANYOF_POSIXL_TEST(node, i)) {
22911 sv_catpv(posixes, anyofs[i]);
22912 }
22913 }
22914 }
22915 }
22916 }
22917
22918 /* Accumulate the bit map into the unconditional match list */
22919 if (bitmap) {
22920 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
22921 if (BITMAP_TEST(bitmap, i)) {
22922 int start = i++;
22923 for (;
22924 i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i);
22925 i++)
22926 { /* empty */ }
22927 invlist = _add_range_to_invlist(invlist, start, i-1);
22928 }
22929 }
22930 }
22931
22932 /* Make sure that the conditional match lists don't have anything in them
22933 * that match unconditionally; otherwise the output is quite confusing.
22934 * This could happen if the code that populates these misses some
22935 * duplication. */
22936 if (only_utf8) {
22937 _invlist_subtract(only_utf8, invlist, &only_utf8);
22938 }
22939 if (not_utf8) {
22940 _invlist_subtract(not_utf8, invlist, ¬_utf8);
22941 }
22942
22943 if (only_utf8_locale_invlist) {
22944
22945 /* Since this list is passed in, we have to make a copy before
22946 * modifying it */
22947 only_utf8_locale = invlist_clone(only_utf8_locale_invlist, NULL);
22948
22949 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
22950
22951 /* And, it can get really weird for us to try outputting an inverted
22952 * form of this list when it has things above the bitmap, so don't even
22953 * try */
22954 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
22955 inverting_allowed = FALSE;
22956 }
22957 }
22958
22959 /* Calculate what the output would be if we take the input as-is */
22960 as_is_display = put_charclass_bitmap_innards_common(invlist,
22961 posixes,
22962 only_utf8,
22963 not_utf8,
22964 only_utf8_locale,
22965 invert);
22966
22967 /* If have to take the output as-is, just do that */
22968 if (! inverting_allowed) {
22969 if (as_is_display) {
22970 sv_catsv(sv, as_is_display);
22971 SvREFCNT_dec_NN(as_is_display);
22972 }
22973 }
22974 else { /* But otherwise, create the output again on the inverted input, and
22975 use whichever version is shorter */
22976
22977 int inverted_bias, as_is_bias;
22978
22979 /* We will apply our bias to whichever of the results doesn't have
22980 * the '^' */
22981 if (invert) {
22982 invert = FALSE;
22983 as_is_bias = bias;
22984 inverted_bias = 0;
22985 }
22986 else {
22987 invert = TRUE;
22988 as_is_bias = 0;
22989 inverted_bias = bias;
22990 }
22991
22992 /* Now invert each of the lists that contribute to the output,
22993 * excluding from the result things outside the possible range */
22994
22995 /* For the unconditional inversion list, we have to add in all the
22996 * conditional code points, so that when inverted, they will be gone
22997 * from it */
22998 _invlist_union(only_utf8, invlist, &invlist);
22999 _invlist_union(not_utf8, invlist, &invlist);
23000 _invlist_union(only_utf8_locale, invlist, &invlist);
23001 _invlist_invert(invlist);
23002 _invlist_intersection(invlist, PL_InBitmap, &invlist);
23003
23004 if (only_utf8) {
23005 _invlist_invert(only_utf8);
23006 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
23007 }
23008 else if (not_utf8) {
23009
23010 /* If a code point matches iff the target string is not in UTF-8,
23011 * then complementing the result has it not match iff not in UTF-8,
23012 * which is the same thing as matching iff it is UTF-8. */
23013 only_utf8 = not_utf8;
23014 not_utf8 = NULL;
23015 }
23016
23017 if (only_utf8_locale) {
23018 _invlist_invert(only_utf8_locale);
23019 _invlist_intersection(only_utf8_locale,
23020 PL_InBitmap,
23021 &only_utf8_locale);
23022 }
23023
23024 inverted_display = put_charclass_bitmap_innards_common(
23025 invlist,
23026 posixes,
23027 only_utf8,
23028 not_utf8,
23029 only_utf8_locale, invert);
23030
23031 /* Use the shortest representation, taking into account our bias
23032 * against showing it inverted */
23033 if ( inverted_display
23034 && ( ! as_is_display
23035 || ( SvCUR(inverted_display) + inverted_bias
23036 < SvCUR(as_is_display) + as_is_bias)))
23037 {
23038 sv_catsv(sv, inverted_display);
23039 }
23040 else if (as_is_display) {
23041 sv_catsv(sv, as_is_display);
23042 }
23043
23044 SvREFCNT_dec(as_is_display);
23045 SvREFCNT_dec(inverted_display);
23046 }
23047
23048 SvREFCNT_dec_NN(invlist);
23049 SvREFCNT_dec(only_utf8);
23050 SvREFCNT_dec(not_utf8);
23051 SvREFCNT_dec(posixes);
23052 SvREFCNT_dec(only_utf8_locale);
23053
23054 return SvCUR(sv) > orig_sv_cur;
23055 }
23056
23057 #define CLEAR_OPTSTART \
23058 if (optstart) STMT_START { \
23059 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
23060 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
23061 optstart=NULL; \
23062 } STMT_END
23063
23064 #define DUMPUNTIL(b,e) \
23065 CLEAR_OPTSTART; \
23066 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
23067
23068 STATIC const regnode *
S_dumpuntil(pTHX_ const regexp * r,const regnode * start,const regnode * node,const regnode * last,const regnode * plast,SV * sv,I32 indent,U32 depth)23069 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
23070 const regnode *last, const regnode *plast,
23071 SV* sv, I32 indent, U32 depth)
23072 {
23073 U8 op = PSEUDO; /* Arbitrary non-END op. */
23074 const regnode *next;
23075 const regnode *optstart= NULL;
23076
23077 RXi_GET_DECL(r, ri);
23078 DECLARE_AND_GET_RE_DEBUG_FLAGS;
23079
23080 PERL_ARGS_ASSERT_DUMPUNTIL;
23081
23082 #ifdef DEBUG_DUMPUNTIL
23083 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n", indent, node-start,
23084 last ? last-start : 0, plast ? plast-start : 0);
23085 #endif
23086
23087 if (plast && plast < last)
23088 last= plast;
23089
23090 while (PL_regkind[op] != END && (!last || node < last)) {
23091 assert(node);
23092 /* While that wasn't END last time... */
23093 NODE_ALIGN(node);
23094 op = OP(node);
23095 if (op == CLOSE || op == SRCLOSE || op == WHILEM)
23096 indent--;
23097 next = regnext((regnode *)node);
23098
23099 /* Where, what. */
23100 if (OP(node) == OPTIMIZED) {
23101 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
23102 optstart = node;
23103 else
23104 goto after_print;
23105 } else
23106 CLEAR_OPTSTART;
23107
23108 regprop(r, sv, node, NULL, NULL);
23109 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
23110 (int)(2*indent + 1), "", SvPVX_const(sv));
23111
23112 if (OP(node) != OPTIMIZED) {
23113 if (next == NULL) /* Next ptr. */
23114 Perl_re_printf( aTHX_ " (0)");
23115 else if (PL_regkind[(U8)op] == BRANCH
23116 && PL_regkind[OP(next)] != BRANCH )
23117 Perl_re_printf( aTHX_ " (FAIL)");
23118 else
23119 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
23120 Perl_re_printf( aTHX_ "\n");
23121 }
23122
23123 after_print:
23124 if (PL_regkind[(U8)op] == BRANCHJ) {
23125 assert(next);
23126 {
23127 const regnode *nnode = (OP(next) == LONGJMP
23128 ? regnext((regnode *)next)
23129 : next);
23130 if (last && nnode > last)
23131 nnode = last;
23132 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
23133 }
23134 }
23135 else if (PL_regkind[(U8)op] == BRANCH) {
23136 assert(next);
23137 DUMPUNTIL(NEXTOPER(node), next);
23138 }
23139 else if ( PL_regkind[(U8)op] == TRIE ) {
23140 const regnode *this_trie = node;
23141 const char op = OP(node);
23142 const U32 n = ARG(node);
23143 const reg_ac_data * const ac = op>=AHOCORASICK ?
23144 (reg_ac_data *)ri->data->data[n] :
23145 NULL;
23146 const reg_trie_data * const trie =
23147 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
23148 #ifdef DEBUGGING
23149 AV *const trie_words
23150 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
23151 #endif
23152 const regnode *nextbranch= NULL;
23153 I32 word_idx;
23154 SvPVCLEAR(sv);
23155 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
23156 SV ** const elem_ptr = av_fetch(trie_words, word_idx, 0);
23157
23158 Perl_re_indentf( aTHX_ "%s ",
23159 indent+3,
23160 elem_ptr
23161 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
23162 SvCUR(*elem_ptr), PL_dump_re_max_len,
23163 PL_colors[0], PL_colors[1],
23164 (SvUTF8(*elem_ptr)
23165 ? PERL_PV_ESCAPE_UNI
23166 : 0)
23167 | PERL_PV_PRETTY_ELLIPSES
23168 | PERL_PV_PRETTY_LTGT
23169 )
23170 : "???"
23171 );
23172 if (trie->jump) {
23173 U16 dist= trie->jump[word_idx+1];
23174 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
23175 (UV)((dist ? this_trie + dist : next) - start));
23176 if (dist) {
23177 if (!nextbranch)
23178 nextbranch= this_trie + trie->jump[0];
23179 DUMPUNTIL(this_trie + dist, nextbranch);
23180 }
23181 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
23182 nextbranch= regnext((regnode *)nextbranch);
23183 } else {
23184 Perl_re_printf( aTHX_ "\n");
23185 }
23186 }
23187 if (last && next > last)
23188 node= last;
23189 else
23190 node= next;
23191 }
23192 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
23193 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
23194 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
23195 }
23196 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
23197 assert(next);
23198 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
23199 }
23200 else if ( op == PLUS || op == STAR) {
23201 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
23202 }
23203 else if (PL_regkind[(U8)op] == EXACT || op == ANYOFHs) {
23204 /* Literal string, where present. */
23205 node += NODE_SZ_STR(node) - 1;
23206 node = NEXTOPER(node);
23207 }
23208 else {
23209 node = NEXTOPER(node);
23210 node += regarglen[(U8)op];
23211 }
23212 if (op == CURLYX || op == OPEN || op == SROPEN)
23213 indent++;
23214 }
23215 CLEAR_OPTSTART;
23216 #ifdef DEBUG_DUMPUNTIL
23217 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
23218 #endif
23219 return node;
23220 }
23221
23222 #endif /* DEBUGGING */
23223
23224 #ifndef PERL_IN_XSUB_RE
23225
23226 # include "uni_keywords.h"
23227
23228 void
Perl_init_uniprops(pTHX)23229 Perl_init_uniprops(pTHX)
23230 {
23231
23232 # ifdef DEBUGGING
23233 char * dump_len_string;
23234
23235 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
23236 if ( ! dump_len_string
23237 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
23238 {
23239 PL_dump_re_max_len = 60; /* A reasonable default */
23240 }
23241 # endif
23242
23243 PL_user_def_props = newHV();
23244
23245 # ifdef USE_ITHREADS
23246
23247 HvSHAREKEYS_off(PL_user_def_props);
23248 PL_user_def_props_aTHX = aTHX;
23249
23250 # endif
23251
23252 /* Set up the inversion list interpreter-level variables */
23253
23254 PL_XPosix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
23255 PL_XPosix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALNUM]);
23256 PL_XPosix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALPHA]);
23257 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXBLANK]);
23258 PL_XPosix_ptrs[_CC_CASED] = _new_invlist_C_array(uni_prop_ptrs[UNI_CASED]);
23259 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXCNTRL]);
23260 PL_XPosix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXDIGIT]);
23261 PL_XPosix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXGRAPH]);
23262 PL_XPosix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXLOWER]);
23263 PL_XPosix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPRINT]);
23264 PL_XPosix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPUNCT]);
23265 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXSPACE]);
23266 PL_XPosix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXUPPER]);
23267 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_VERTSPACE]);
23268 PL_XPosix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXWORD]);
23269 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXXDIGIT]);
23270
23271 PL_Posix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
23272 PL_Posix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALNUM]);
23273 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALPHA]);
23274 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXBLANK]);
23275 PL_Posix_ptrs[_CC_CASED] = PL_Posix_ptrs[_CC_ALPHA];
23276 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXCNTRL]);
23277 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXDIGIT]);
23278 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXGRAPH]);
23279 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXLOWER]);
23280 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPRINT]);
23281 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPUNCT]);
23282 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXSPACE]);
23283 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXUPPER]);
23284 PL_Posix_ptrs[_CC_VERTSPACE] = NULL;
23285 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXWORD]);
23286 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXXDIGIT]);
23287
23288 PL_GCB_invlist = _new_invlist_C_array(_Perl_GCB_invlist);
23289 PL_SB_invlist = _new_invlist_C_array(_Perl_SB_invlist);
23290 PL_WB_invlist = _new_invlist_C_array(_Perl_WB_invlist);
23291 PL_LB_invlist = _new_invlist_C_array(_Perl_LB_invlist);
23292 PL_SCX_invlist = _new_invlist_C_array(_Perl_SCX_invlist);
23293
23294 PL_InBitmap = _new_invlist_C_array(InBitmap_invlist);
23295 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
23296 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
23297 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
23298
23299 PL_Assigned_invlist = _new_invlist_C_array(uni_prop_ptrs[UNI_ASSIGNED]);
23300
23301 PL_utf8_perl_idstart = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDSTART]);
23302 PL_utf8_perl_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDCONT]);
23303
23304 PL_utf8_charname_begin = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_BEGIN]);
23305 PL_utf8_charname_continue = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_CONTINUE]);
23306
23307 PL_in_some_fold = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_ANY_FOLDS]);
23308 PL_HasMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
23309 UNI__PERL_FOLDS_TO_MULTI_CHAR]);
23310 PL_InMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
23311 UNI__PERL_IS_IN_MULTI_CHAR_FOLD]);
23312 PL_utf8_toupper = _new_invlist_C_array(Uppercase_Mapping_invlist);
23313 PL_utf8_tolower = _new_invlist_C_array(Lowercase_Mapping_invlist);
23314 PL_utf8_totitle = _new_invlist_C_array(Titlecase_Mapping_invlist);
23315 PL_utf8_tofold = _new_invlist_C_array(Case_Folding_invlist);
23316 PL_utf8_tosimplefold = _new_invlist_C_array(Simple_Case_Folding_invlist);
23317 PL_utf8_foldclosures = _new_invlist_C_array(_Perl_IVCF_invlist);
23318 PL_utf8_mark = _new_invlist_C_array(uni_prop_ptrs[UNI_M]);
23319 PL_CCC_non0_non230 = _new_invlist_C_array(_Perl_CCC_non0_non230_invlist);
23320 PL_Private_Use = _new_invlist_C_array(uni_prop_ptrs[UNI_CO]);
23321
23322 # ifdef UNI_XIDC
23323 /* The below are used only by deprecated functions. They could be removed */
23324 PL_utf8_xidcont = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDC]);
23325 PL_utf8_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI_IDC]);
23326 PL_utf8_xidstart = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDS]);
23327 # endif
23328 }
23329
23330 /* These four functions are compiled only in regcomp.c, where they have access
23331 * to the data they return. They are a way for re_comp.c to get access to that
23332 * data without having to compile the whole data structures. */
23333
23334 I16
Perl_do_uniprop_match(const char * const key,const U16 key_len)23335 Perl_do_uniprop_match(const char * const key, const U16 key_len)
23336 {
23337 PERL_ARGS_ASSERT_DO_UNIPROP_MATCH;
23338
23339 return match_uniprop((U8 *) key, key_len);
23340 }
23341
23342 SV *
Perl_get_prop_definition(pTHX_ const int table_index)23343 Perl_get_prop_definition(pTHX_ const int table_index)
23344 {
23345 PERL_ARGS_ASSERT_GET_PROP_DEFINITION;
23346
23347 /* Create and return the inversion list */
23348 return _new_invlist_C_array(uni_prop_ptrs[table_index]);
23349 }
23350
23351 const char * const *
Perl_get_prop_values(const int table_index)23352 Perl_get_prop_values(const int table_index)
23353 {
23354 PERL_ARGS_ASSERT_GET_PROP_VALUES;
23355
23356 return UNI_prop_value_ptrs[table_index];
23357 }
23358
23359 const char *
Perl_get_deprecated_property_msg(const Size_t warning_offset)23360 Perl_get_deprecated_property_msg(const Size_t warning_offset)
23361 {
23362 PERL_ARGS_ASSERT_GET_DEPRECATED_PROPERTY_MSG;
23363
23364 return deprecated_property_msgs[warning_offset];
23365 }
23366
23367 # if 0
23368
23369 This code was mainly added for backcompat to give a warning for non-portable
23370 code points in user-defined properties. But experiments showed that the
23371 warning in earlier perls were only omitted on overflow, which should be an
23372 error, so there really isnt a backcompat issue, and actually adding the
23373 warning when none was present before might cause breakage, for little gain. So
23374 khw left this code in, but not enabled. Tests were never added.
23375
23376 embed.fnc entry:
23377 Ei |const char *|get_extended_utf8_msg|const UV cp
23378
23379 PERL_STATIC_INLINE const char *
23380 S_get_extended_utf8_msg(pTHX_ const UV cp)
23381 {
23382 U8 dummy[UTF8_MAXBYTES + 1];
23383 HV *msgs;
23384 SV **msg;
23385
23386 uvchr_to_utf8_flags_msgs(dummy, cp, UNICODE_WARN_PERL_EXTENDED,
23387 &msgs);
23388
23389 msg = hv_fetchs(msgs, "text", 0);
23390 assert(msg);
23391
23392 (void) sv_2mortal((SV *) msgs);
23393
23394 return SvPVX(*msg);
23395 }
23396
23397 # endif
23398 #endif /* end of ! PERL_IN_XSUB_RE */
23399
23400 STATIC REGEXP *
S_compile_wildcard(pTHX_ const char * subpattern,const STRLEN len,const bool ignore_case)23401 S_compile_wildcard(pTHX_ const char * subpattern, const STRLEN len,
23402 const bool ignore_case)
23403 {
23404 /* Pretends that the input subpattern is qr/subpattern/aam, compiling it
23405 * possibly with /i if the 'ignore_case' parameter is true. Use /aa
23406 * because nothing outside of ASCII will match. Use /m because the input
23407 * string may be a bunch of lines strung together.
23408 *
23409 * Also sets up the debugging info */
23410
23411 U32 flags = PMf_MULTILINE|PMf_WILDCARD;
23412 U32 rx_flags;
23413 SV * subpattern_sv = sv_2mortal(newSVpvn(subpattern, len));
23414 REGEXP * subpattern_re;
23415 DECLARE_AND_GET_RE_DEBUG_FLAGS;
23416
23417 PERL_ARGS_ASSERT_COMPILE_WILDCARD;
23418
23419 if (ignore_case) {
23420 flags |= PMf_FOLD;
23421 }
23422 set_regex_charset(&flags, REGEX_ASCII_MORE_RESTRICTED_CHARSET);
23423
23424 /* Like in op.c, we copy the compile time pm flags to the rx ones */
23425 rx_flags = flags & RXf_PMf_COMPILETIME;
23426
23427 #ifndef PERL_IN_XSUB_RE
23428 /* Use the core engine if this file is regcomp.c. That means no
23429 * 'use re "Debug ..." is in effect, so the core engine is sufficient */
23430 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23431 &PL_core_reg_engine,
23432 NULL, NULL,
23433 rx_flags, flags);
23434 #else
23435 if (isDEBUG_WILDCARD) {
23436 /* Use the special debugging engine if this file is re_comp.c and wants
23437 * to output the wildcard matching. This uses whatever
23438 * 'use re "Debug ..." is in effect */
23439 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23440 &my_reg_engine,
23441 NULL, NULL,
23442 rx_flags, flags);
23443 }
23444 else {
23445 /* Use the special wildcard engine if this file is re_comp.c and
23446 * doesn't want to output the wildcard matching. This uses whatever
23447 * 'use re "Debug ..." is in effect for compilation, but this engine
23448 * structure has been set up so that it uses the core engine for
23449 * execution, so no execution debugging as a result of re.pm will be
23450 * displayed. */
23451 subpattern_re = Perl_re_op_compile(aTHX_ &subpattern_sv, 1, NULL,
23452 &wild_reg_engine,
23453 NULL, NULL,
23454 rx_flags, flags);
23455 /* XXX The above has the effect that any user-supplied regex engine
23456 * won't be called for matching wildcards. That might be good, or bad.
23457 * It could be changed in several ways. The reason it is done the
23458 * current way is to avoid having to save and restore
23459 * ^{^RE_DEBUG_FLAGS} around the execution. save_scalar() perhaps
23460 * could be used. Another suggestion is to keep the authoritative
23461 * value of the debug flags in a thread-local variable and add set/get
23462 * magic to ${^RE_DEBUG_FLAGS} to keep the C level variable up to date.
23463 * Still another is to pass a flag, say in the engine's intflags that
23464 * would be checked each time before doing the debug output */
23465 }
23466 #endif
23467
23468 assert(subpattern_re); /* Should have died if didn't compile successfully */
23469 return subpattern_re;
23470 }
23471
23472 STATIC I32
S_execute_wildcard(pTHX_ REGEXP * const prog,char * stringarg,char * strend,char * strbeg,SSize_t minend,SV * screamer,U32 nosave)23473 S_execute_wildcard(pTHX_ REGEXP * const prog, char* stringarg, char *strend,
23474 char *strbeg, SSize_t minend, SV *screamer, U32 nosave)
23475 {
23476 I32 result;
23477 DECLARE_AND_GET_RE_DEBUG_FLAGS;
23478
23479 PERL_ARGS_ASSERT_EXECUTE_WILDCARD;
23480
23481 ENTER;
23482
23483 /* The compilation has set things up so that if the program doesn't want to
23484 * see the wildcard matching procedure, it will get the core execution
23485 * engine, which is subject only to -Dr. So we have to turn that off
23486 * around this procedure */
23487 if (! isDEBUG_WILDCARD) {
23488 /* Note! Casts away 'volatile' */
23489 SAVEI32(PL_debug);
23490 PL_debug &= ~ DEBUG_r_FLAG;
23491 }
23492
23493 result = CALLREGEXEC(prog, stringarg, strend, strbeg, minend, screamer,
23494 NULL, nosave);
23495 LEAVE;
23496
23497 return result;
23498 }
23499
23500 SV *
S_handle_user_defined_property(pTHX_ const char * name,const STRLEN name_len,const bool is_utf8,const bool to_fold,const bool runtime,const bool deferrable,SV * contents,bool * user_defined_ptr,SV * msg,const STRLEN level)23501 S_handle_user_defined_property(pTHX_
23502
23503 /* Parses the contents of a user-defined property definition; returning the
23504 * expanded definition if possible. If so, the return is an inversion
23505 * list.
23506 *
23507 * If there are subroutines that are part of the expansion and which aren't
23508 * known at the time of the call to this function, this returns what
23509 * parse_uniprop_string() returned for the first one encountered.
23510 *
23511 * If an error was found, NULL is returned, and 'msg' gets a suitable
23512 * message appended to it. (Appending allows the back trace of how we got
23513 * to the faulty definition to be displayed through nested calls of
23514 * user-defined subs.)
23515 *
23516 * The caller IS responsible for freeing any returned SV.
23517 *
23518 * The syntax of the contents is pretty much described in perlunicode.pod,
23519 * but we also allow comments on each line */
23520
23521 const char * name, /* Name of property */
23522 const STRLEN name_len, /* The name's length in bytes */
23523 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23524 const bool to_fold, /* ? Is this under /i */
23525 const bool runtime, /* ? Are we in compile- or run-time */
23526 const bool deferrable, /* Is it ok for this property's full definition
23527 to be deferred until later? */
23528 SV* contents, /* The property's definition */
23529 bool *user_defined_ptr, /* This will be set TRUE as we wouldn't be
23530 getting called unless this is thought to be
23531 a user-defined property */
23532 SV * msg, /* Any error or warning msg(s) are appended to
23533 this */
23534 const STRLEN level) /* Recursion level of this call */
23535 {
23536 STRLEN len;
23537 const char * string = SvPV_const(contents, len);
23538 const char * const e = string + len;
23539 const bool is_contents_utf8 = cBOOL(SvUTF8(contents));
23540 const STRLEN msgs_length_on_entry = SvCUR(msg);
23541
23542 const char * s0 = string; /* Points to first byte in the current line
23543 being parsed in 'string' */
23544 const char overflow_msg[] = "Code point too large in \"";
23545 SV* running_definition = NULL;
23546
23547 PERL_ARGS_ASSERT_HANDLE_USER_DEFINED_PROPERTY;
23548
23549 *user_defined_ptr = TRUE;
23550
23551 /* Look at each line */
23552 while (s0 < e) {
23553 const char * s; /* Current byte */
23554 char op = '+'; /* Default operation is 'union' */
23555 IV min = 0; /* range begin code point */
23556 IV max = -1; /* and range end */
23557 SV* this_definition;
23558
23559 /* Skip comment lines */
23560 if (*s0 == '#') {
23561 s0 = strchr(s0, '\n');
23562 if (s0 == NULL) {
23563 break;
23564 }
23565 s0++;
23566 continue;
23567 }
23568
23569 /* For backcompat, allow an empty first line */
23570 if (*s0 == '\n') {
23571 s0++;
23572 continue;
23573 }
23574
23575 /* First character in the line may optionally be the operation */
23576 if ( *s0 == '+'
23577 || *s0 == '!'
23578 || *s0 == '-'
23579 || *s0 == '&')
23580 {
23581 op = *s0++;
23582 }
23583
23584 /* If the line is one or two hex digits separated by blank space, its
23585 * a range; otherwise it is either another user-defined property or an
23586 * error */
23587
23588 s = s0;
23589
23590 if (! isXDIGIT(*s)) {
23591 goto check_if_property;
23592 }
23593
23594 do { /* Each new hex digit will add 4 bits. */
23595 if (min > ( (IV) MAX_LEGAL_CP >> 4)) {
23596 s = strchr(s, '\n');
23597 if (s == NULL) {
23598 s = e;
23599 }
23600 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23601 sv_catpv(msg, overflow_msg);
23602 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23603 UTF8fARG(is_contents_utf8, s - s0, s0));
23604 sv_catpvs(msg, "\"");
23605 goto return_failure;
23606 }
23607
23608 /* Accumulate this digit into the value */
23609 min = (min << 4) + READ_XDIGIT(s);
23610 } while (isXDIGIT(*s));
23611
23612 while (isBLANK(*s)) { s++; }
23613
23614 /* We allow comments at the end of the line */
23615 if (*s == '#') {
23616 s = strchr(s, '\n');
23617 if (s == NULL) {
23618 s = e;
23619 }
23620 s++;
23621 }
23622 else if (s < e && *s != '\n') {
23623 if (! isXDIGIT(*s)) {
23624 goto check_if_property;
23625 }
23626
23627 /* Look for the high point of the range */
23628 max = 0;
23629 do {
23630 if (max > ( (IV) MAX_LEGAL_CP >> 4)) {
23631 s = strchr(s, '\n');
23632 if (s == NULL) {
23633 s = e;
23634 }
23635 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23636 sv_catpv(msg, overflow_msg);
23637 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23638 UTF8fARG(is_contents_utf8, s - s0, s0));
23639 sv_catpvs(msg, "\"");
23640 goto return_failure;
23641 }
23642
23643 max = (max << 4) + READ_XDIGIT(s);
23644 } while (isXDIGIT(*s));
23645
23646 while (isBLANK(*s)) { s++; }
23647
23648 if (*s == '#') {
23649 s = strchr(s, '\n');
23650 if (s == NULL) {
23651 s = e;
23652 }
23653 }
23654 else if (s < e && *s != '\n') {
23655 goto check_if_property;
23656 }
23657 }
23658
23659 if (max == -1) { /* The line only had one entry */
23660 max = min;
23661 }
23662 else if (max < min) {
23663 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23664 sv_catpvs(msg, "Illegal range in \"");
23665 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23666 UTF8fARG(is_contents_utf8, s - s0, s0));
23667 sv_catpvs(msg, "\"");
23668 goto return_failure;
23669 }
23670
23671 # if 0 /* See explanation at definition above of get_extended_utf8_msg() */
23672
23673 if ( UNICODE_IS_PERL_EXTENDED(min)
23674 || UNICODE_IS_PERL_EXTENDED(max))
23675 {
23676 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23677
23678 /* If both code points are non-portable, warn only on the lower
23679 * one. */
23680 sv_catpv(msg, get_extended_utf8_msg(
23681 (UNICODE_IS_PERL_EXTENDED(min))
23682 ? min : max));
23683 sv_catpvs(msg, " in \"");
23684 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
23685 UTF8fARG(is_contents_utf8, s - s0, s0));
23686 sv_catpvs(msg, "\"");
23687 }
23688
23689 # endif
23690
23691 /* Here, this line contains a legal range */
23692 this_definition = sv_2mortal(_new_invlist(2));
23693 this_definition = _add_range_to_invlist(this_definition, min, max);
23694 goto calculate;
23695
23696 check_if_property:
23697
23698 /* Here it isn't a legal range line. See if it is a legal property
23699 * line. First find the end of the meat of the line */
23700 s = strpbrk(s, "#\n");
23701 if (s == NULL) {
23702 s = e;
23703 }
23704
23705 /* Ignore trailing blanks in keeping with the requirements of
23706 * parse_uniprop_string() */
23707 s--;
23708 while (s > s0 && isBLANK_A(*s)) {
23709 s--;
23710 }
23711 s++;
23712
23713 this_definition = parse_uniprop_string(s0, s - s0,
23714 is_utf8, to_fold, runtime,
23715 deferrable,
23716 NULL,
23717 user_defined_ptr, msg,
23718 (name_len == 0)
23719 ? level /* Don't increase level
23720 if input is empty */
23721 : level + 1
23722 );
23723 if (this_definition == NULL) {
23724 goto return_failure; /* 'msg' should have had the reason
23725 appended to it by the above call */
23726 }
23727
23728 if (! is_invlist(this_definition)) { /* Unknown at this time */
23729 return newSVsv(this_definition);
23730 }
23731
23732 if (*s != '\n') {
23733 s = strchr(s, '\n');
23734 if (s == NULL) {
23735 s = e;
23736 }
23737 }
23738
23739 calculate:
23740
23741 switch (op) {
23742 case '+':
23743 _invlist_union(running_definition, this_definition,
23744 &running_definition);
23745 break;
23746 case '-':
23747 _invlist_subtract(running_definition, this_definition,
23748 &running_definition);
23749 break;
23750 case '&':
23751 _invlist_intersection(running_definition, this_definition,
23752 &running_definition);
23753 break;
23754 case '!':
23755 _invlist_union_complement_2nd(running_definition,
23756 this_definition, &running_definition);
23757 break;
23758 default:
23759 Perl_croak(aTHX_ "panic: %s: %d: Unexpected operation %d",
23760 __FILE__, __LINE__, op);
23761 break;
23762 }
23763
23764 /* Position past the '\n' */
23765 s0 = s + 1;
23766 } /* End of loop through the lines of 'contents' */
23767
23768 /* Here, we processed all the lines in 'contents' without error. If we
23769 * didn't add any warnings, simply return success */
23770 if (msgs_length_on_entry == SvCUR(msg)) {
23771
23772 /* If the expansion was empty, the answer isn't nothing: its an empty
23773 * inversion list */
23774 if (running_definition == NULL) {
23775 running_definition = _new_invlist(1);
23776 }
23777
23778 return running_definition;
23779 }
23780
23781 /* Otherwise, add some explanatory text, but we will return success */
23782 goto return_msg;
23783
23784 return_failure:
23785 running_definition = NULL;
23786
23787 return_msg:
23788
23789 if (name_len > 0) {
23790 sv_catpvs(msg, " in expansion of ");
23791 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
23792 }
23793
23794 return running_definition;
23795 }
23796
23797 /* As explained below, certain operations need to take place in the first
23798 * thread created. These macros switch contexts */
23799 # ifdef USE_ITHREADS
23800 # define DECLARATION_FOR_GLOBAL_CONTEXT \
23801 PerlInterpreter * save_aTHX = aTHX;
23802 # define SWITCH_TO_GLOBAL_CONTEXT \
23803 PERL_SET_CONTEXT((aTHX = PL_user_def_props_aTHX))
23804 # define RESTORE_CONTEXT PERL_SET_CONTEXT((aTHX = save_aTHX));
23805 # define CUR_CONTEXT aTHX
23806 # define ORIGINAL_CONTEXT save_aTHX
23807 # else
23808 # define DECLARATION_FOR_GLOBAL_CONTEXT dNOOP
23809 # define SWITCH_TO_GLOBAL_CONTEXT NOOP
23810 # define RESTORE_CONTEXT NOOP
23811 # define CUR_CONTEXT NULL
23812 # define ORIGINAL_CONTEXT NULL
23813 # endif
23814
23815 STATIC void
S_delete_recursion_entry(pTHX_ void * key)23816 S_delete_recursion_entry(pTHX_ void *key)
23817 {
23818 /* Deletes the entry used to detect recursion when expanding user-defined
23819 * properties. This is a function so it can be set up to be called even if
23820 * the program unexpectedly quits */
23821
23822 SV ** current_entry;
23823 const STRLEN key_len = strlen((const char *) key);
23824 DECLARATION_FOR_GLOBAL_CONTEXT;
23825
23826 SWITCH_TO_GLOBAL_CONTEXT;
23827
23828 /* If the entry is one of these types, it is a permanent entry, and not the
23829 * one used to detect recursions. This function should delete only the
23830 * recursion entry */
23831 current_entry = hv_fetch(PL_user_def_props, (const char *) key, key_len, 0);
23832 if ( current_entry
23833 && ! is_invlist(*current_entry)
23834 && ! SvPOK(*current_entry))
23835 {
23836 (void) hv_delete(PL_user_def_props, (const char *) key, key_len,
23837 G_DISCARD);
23838 }
23839
23840 RESTORE_CONTEXT;
23841 }
23842
23843 STATIC SV *
S_get_fq_name(pTHX_ const char * const name,const Size_t name_len,const bool is_utf8,const bool has_colon_colon)23844 S_get_fq_name(pTHX_
23845 const char * const name, /* The first non-blank in the \p{}, \P{} */
23846 const Size_t name_len, /* Its length in bytes, not including any trailing space */
23847 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23848 const bool has_colon_colon
23849 )
23850 {
23851 /* Returns a mortal SV containing the fully qualified version of the input
23852 * name */
23853
23854 SV * fq_name;
23855
23856 fq_name = newSVpvs_flags("", SVs_TEMP);
23857
23858 /* Use the current package if it wasn't included in our input */
23859 if (! has_colon_colon) {
23860 const HV * pkg = (IN_PERL_COMPILETIME)
23861 ? PL_curstash
23862 : CopSTASH(PL_curcop);
23863 const char* pkgname = HvNAME(pkg);
23864
23865 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
23866 UTF8fARG(is_utf8, strlen(pkgname), pkgname));
23867 sv_catpvs(fq_name, "::");
23868 }
23869
23870 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
23871 UTF8fARG(is_utf8, name_len, name));
23872 return fq_name;
23873 }
23874
23875 STATIC SV *
S_parse_uniprop_string(pTHX_ const char * const name,Size_t name_len,const bool is_utf8,const bool to_fold,const bool runtime,const bool deferrable,AV ** strings,bool * user_defined_ptr,SV * msg,const STRLEN level)23876 S_parse_uniprop_string(pTHX_
23877
23878 /* Parse the interior of a \p{}, \P{}. Returns its definition if knowable
23879 * now. If so, the return is an inversion list.
23880 *
23881 * If the property is user-defined, it is a subroutine, which in turn
23882 * may call other subroutines. This function will call the whole nest of
23883 * them to get the definition they return; if some aren't known at the time
23884 * of the call to this function, the fully qualified name of the highest
23885 * level sub is returned. It is an error to call this function at runtime
23886 * without every sub defined.
23887 *
23888 * If an error was found, NULL is returned, and 'msg' gets a suitable
23889 * message appended to it. (Appending allows the back trace of how we got
23890 * to the faulty definition to be displayed through nested calls of
23891 * user-defined subs.)
23892 *
23893 * The caller should NOT try to free any returned inversion list.
23894 *
23895 * Other parameters will be set on return as described below */
23896
23897 const char * const name, /* The first non-blank in the \p{}, \P{} */
23898 Size_t name_len, /* Its length in bytes, not including any
23899 trailing space */
23900 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23901 const bool to_fold, /* ? Is this under /i */
23902 const bool runtime, /* TRUE if this is being called at run time */
23903 const bool deferrable, /* TRUE if it's ok for the definition to not be
23904 known at this call */
23905 AV ** strings, /* To return string property values, like named
23906 sequences */
23907 bool *user_defined_ptr, /* Upon return from this function it will be
23908 set to TRUE if any component is a
23909 user-defined property */
23910 SV * msg, /* Any error or warning msg(s) are appended to
23911 this */
23912 const STRLEN level) /* Recursion level of this call */
23913 {
23914 char* lookup_name; /* normalized name for lookup in our tables */
23915 unsigned lookup_len; /* Its length */
23916 enum { Not_Strict = 0, /* Some properties have stricter name */
23917 Strict, /* normalization rules, which we decide */
23918 As_Is /* upon based on parsing */
23919 } stricter = Not_Strict;
23920
23921 /* nv= or numeric_value=, or possibly one of the cjk numeric properties
23922 * (though it requires extra effort to download them from Unicode and
23923 * compile perl to know about them) */
23924 bool is_nv_type = FALSE;
23925
23926 unsigned int i, j = 0;
23927 int equals_pos = -1; /* Where the '=' is found, or negative if none */
23928 int slash_pos = -1; /* Where the '/' is found, or negative if none */
23929 int table_index = 0; /* The entry number for this property in the table
23930 of all Unicode property names */
23931 bool starts_with_Is = FALSE; /* ? Does the name start with 'Is' */
23932 Size_t lookup_offset = 0; /* Used to ignore the first few characters of
23933 the normalized name in certain situations */
23934 Size_t non_pkg_begin = 0; /* Offset of first byte in 'name' that isn't
23935 part of a package name */
23936 Size_t lun_non_pkg_begin = 0; /* Similarly for 'lookup_name' */
23937 bool could_be_user_defined = TRUE; /* ? Could this be a user-defined
23938 property rather than a Unicode
23939 one. */
23940 SV * prop_definition = NULL; /* The returned definition of 'name' or NULL
23941 if an error. If it is an inversion list,
23942 it is the definition. Otherwise it is a
23943 string containing the fully qualified sub
23944 name of 'name' */
23945 SV * fq_name = NULL; /* For user-defined properties, the fully
23946 qualified name */
23947 bool invert_return = FALSE; /* ? Do we need to complement the result before
23948 returning it */
23949 bool stripped_utf8_pkg = FALSE; /* Set TRUE if the input includes an
23950 explicit utf8:: package that we strip
23951 off */
23952 /* The expansion of properties that could be either user-defined or
23953 * official unicode ones is deferred until runtime, including a marker for
23954 * those that might be in the latter category. This boolean indicates if
23955 * we've seen that marker. If not, what we're parsing can't be such an
23956 * official Unicode property whose expansion was deferred */
23957 bool could_be_deferred_official = FALSE;
23958
23959 PERL_ARGS_ASSERT_PARSE_UNIPROP_STRING;
23960
23961 /* The input will be normalized into 'lookup_name' */
23962 Newx(lookup_name, name_len, char);
23963 SAVEFREEPV(lookup_name);
23964
23965 /* Parse the input. */
23966 for (i = 0; i < name_len; i++) {
23967 char cur = name[i];
23968
23969 /* Most of the characters in the input will be of this ilk, being parts
23970 * of a name */
23971 if (isIDCONT_A(cur)) {
23972
23973 /* Case differences are ignored. Our lookup routine assumes
23974 * everything is lowercase, so normalize to that */
23975 if (isUPPER_A(cur)) {
23976 lookup_name[j++] = toLOWER_A(cur);
23977 continue;
23978 }
23979
23980 if (cur == '_') { /* Don't include these in the normalized name */
23981 continue;
23982 }
23983
23984 lookup_name[j++] = cur;
23985
23986 /* The first character in a user-defined name must be of this type.
23987 * */
23988 if (i - non_pkg_begin == 0 && ! isIDFIRST_A(cur)) {
23989 could_be_user_defined = FALSE;
23990 }
23991
23992 continue;
23993 }
23994
23995 /* Here, the character is not something typically in a name, But these
23996 * two types of characters (and the '_' above) can be freely ignored in
23997 * most situations. Later it may turn out we shouldn't have ignored
23998 * them, and we have to reparse, but we don't have enough information
23999 * yet to make that decision */
24000 if (cur == '-' || isSPACE_A(cur)) {
24001 could_be_user_defined = FALSE;
24002 continue;
24003 }
24004
24005 /* An equals sign or single colon mark the end of the first part of
24006 * the property name */
24007 if ( cur == '='
24008 || (cur == ':' && (i >= name_len - 1 || name[i+1] != ':')))
24009 {
24010 lookup_name[j++] = '='; /* Treat the colon as an '=' */
24011 equals_pos = j; /* Note where it occurred in the input */
24012 could_be_user_defined = FALSE;
24013 break;
24014 }
24015
24016 /* If this looks like it is a marker we inserted at compile time,
24017 * set a flag and otherwise ignore it. If it isn't in the final
24018 * position, keep it as it would have been user input. */
24019 if ( UNLIKELY(cur == DEFERRED_COULD_BE_OFFICIAL_MARKERc)
24020 && ! deferrable
24021 && could_be_user_defined
24022 && i == name_len - 1)
24023 {
24024 name_len--;
24025 could_be_deferred_official = TRUE;
24026 continue;
24027 }
24028
24029 /* Otherwise, this character is part of the name. */
24030 lookup_name[j++] = cur;
24031
24032 /* Here it isn't a single colon, so if it is a colon, it must be a
24033 * double colon */
24034 if (cur == ':') {
24035
24036 /* A double colon should be a package qualifier. We note its
24037 * position and continue. Note that one could have
24038 * pkg1::pkg2::...::foo
24039 * so that the position at the end of the loop will be just after
24040 * the final qualifier */
24041
24042 i++;
24043 non_pkg_begin = i + 1;
24044 lookup_name[j++] = ':';
24045 lun_non_pkg_begin = j;
24046 }
24047 else { /* Only word chars (and '::') can be in a user-defined name */
24048 could_be_user_defined = FALSE;
24049 }
24050 } /* End of parsing through the lhs of the property name (or all of it if
24051 no rhs) */
24052
24053 # define STRLENs(s) (sizeof("" s "") - 1)
24054
24055 /* If there is a single package name 'utf8::', it is ambiguous. It could
24056 * be for a user-defined property, or it could be a Unicode property, as
24057 * all of them are considered to be for that package. For the purposes of
24058 * parsing the rest of the property, strip it off */
24059 if (non_pkg_begin == STRLENs("utf8::") && memBEGINPs(name, name_len, "utf8::")) {
24060 lookup_name += STRLENs("utf8::");
24061 j -= STRLENs("utf8::");
24062 equals_pos -= STRLENs("utf8::");
24063 stripped_utf8_pkg = TRUE;
24064 }
24065
24066 /* Here, we are either done with the whole property name, if it was simple;
24067 * or are positioned just after the '=' if it is compound. */
24068
24069 if (equals_pos >= 0) {
24070 assert(stricter == Not_Strict); /* We shouldn't have set this yet */
24071
24072 /* Space immediately after the '=' is ignored */
24073 i++;
24074 for (; i < name_len; i++) {
24075 if (! isSPACE_A(name[i])) {
24076 break;
24077 }
24078 }
24079
24080 /* Most punctuation after the equals indicates a subpattern, like
24081 * \p{foo=/bar/} */
24082 if ( isPUNCT_A(name[i])
24083 && name[i] != '-'
24084 && name[i] != '+'
24085 && name[i] != '_'
24086 && name[i] != '{'
24087 /* A backslash means the real delimitter is the next character,
24088 * but it must be punctuation */
24089 && (name[i] != '\\' || (i < name_len && isPUNCT_A(name[i+1]))))
24090 {
24091 bool special_property = memEQs(lookup_name, j - 1, "name")
24092 || memEQs(lookup_name, j - 1, "na");
24093 if (! special_property) {
24094 /* Find the property. The table includes the equals sign, so
24095 * we use 'j' as-is */
24096 table_index = do_uniprop_match(lookup_name, j);
24097 }
24098 if (special_property || table_index) {
24099 REGEXP * subpattern_re;
24100 char open = name[i++];
24101 char close;
24102 const char * pos_in_brackets;
24103 const char * const * prop_values;
24104 bool escaped = 0;
24105
24106 /* Backslash => delimitter is the character following. We
24107 * already checked that it is punctuation */
24108 if (open == '\\') {
24109 open = name[i++];
24110 escaped = 1;
24111 }
24112
24113 /* This data structure is constructed so that the matching
24114 * closing bracket is 3 past its matching opening. The second
24115 * set of closing is so that if the opening is something like
24116 * ']', the closing will be that as well. Something similar is
24117 * done in toke.c */
24118 pos_in_brackets = memCHRs("([<)]>)]>", open);
24119 close = (pos_in_brackets) ? pos_in_brackets[3] : open;
24120
24121 if ( i >= name_len
24122 || name[name_len-1] != close
24123 || (escaped && name[name_len-2] != '\\')
24124 /* Also make sure that there are enough characters.
24125 * e.g., '\\\' would show up incorrectly as legal even
24126 * though it is too short */
24127 || (SSize_t) (name_len - i - 1 - escaped) < 0)
24128 {
24129 sv_catpvs(msg, "Unicode property wildcard not terminated");
24130 goto append_name_to_msg;
24131 }
24132
24133 Perl_ck_warner_d(aTHX_
24134 packWARN(WARN_EXPERIMENTAL__UNIPROP_WILDCARDS),
24135 "The Unicode property wildcards feature is experimental");
24136
24137 if (special_property) {
24138 const char * error_msg;
24139 const char * revised_name = name + i;
24140 Size_t revised_name_len = name_len - (i + 1 + escaped);
24141
24142 /* Currently, the only 'special_property' is name, which we
24143 * lookup in _charnames.pm */
24144
24145 if (! load_charnames(newSVpvs("placeholder"),
24146 revised_name, revised_name_len,
24147 &error_msg))
24148 {
24149 sv_catpv(msg, error_msg);
24150 goto append_name_to_msg;
24151 }
24152
24153 /* Farm this out to a function just to make the current
24154 * function less unwieldy */
24155 if (handle_names_wildcard(revised_name, revised_name_len,
24156 &prop_definition,
24157 strings))
24158 {
24159 return prop_definition;
24160 }
24161
24162 goto failed;
24163 }
24164
24165 prop_values = get_prop_values(table_index);
24166
24167 /* Now create and compile the wildcard subpattern. Use /i
24168 * because the property values are supposed to match with case
24169 * ignored. */
24170 subpattern_re = compile_wildcard(name + i,
24171 name_len - i - 1 - escaped,
24172 TRUE /* /i */
24173 );
24174
24175 /* For each legal property value, see if the supplied pattern
24176 * matches it. */
24177 while (*prop_values) {
24178 const char * const entry = *prop_values;
24179 const Size_t len = strlen(entry);
24180 SV* entry_sv = newSVpvn_flags(entry, len, SVs_TEMP);
24181
24182 if (execute_wildcard(subpattern_re,
24183 (char *) entry,
24184 (char *) entry + len,
24185 (char *) entry, 0,
24186 entry_sv,
24187 0))
24188 { /* Here, matched. Add to the returned list */
24189 Size_t total_len = j + len;
24190 SV * sub_invlist = NULL;
24191 char * this_string;
24192
24193 /* We know this is a legal \p{property=value}. Call
24194 * the function to return the list of code points that
24195 * match it */
24196 Newxz(this_string, total_len + 1, char);
24197 Copy(lookup_name, this_string, j, char);
24198 my_strlcat(this_string, entry, total_len + 1);
24199 SAVEFREEPV(this_string);
24200 sub_invlist = parse_uniprop_string(this_string,
24201 total_len,
24202 is_utf8,
24203 to_fold,
24204 runtime,
24205 deferrable,
24206 NULL,
24207 user_defined_ptr,
24208 msg,
24209 level + 1);
24210 _invlist_union(prop_definition, sub_invlist,
24211 &prop_definition);
24212 }
24213
24214 prop_values++; /* Next iteration, look at next propvalue */
24215 } /* End of looking through property values; (the data
24216 structure is terminated by a NULL ptr) */
24217
24218 SvREFCNT_dec_NN(subpattern_re);
24219
24220 if (prop_definition) {
24221 return prop_definition;
24222 }
24223
24224 sv_catpvs(msg, "No Unicode property value wildcard matches:");
24225 goto append_name_to_msg;
24226 }
24227
24228 /* Here's how khw thinks we should proceed to handle the properties
24229 * not yet done: Bidi Mirroring Glyph can map to ""
24230 Bidi Paired Bracket can map to ""
24231 Case Folding (both full and simple)
24232 Shouldn't /i be good enough for Full
24233 Decomposition Mapping
24234 Equivalent Unified Ideograph can map to ""
24235 Lowercase Mapping (both full and simple)
24236 NFKC Case Fold can map to ""
24237 Titlecase Mapping (both full and simple)
24238 Uppercase Mapping (both full and simple)
24239 * Handle these the same way Name is done, using say, _wild.pm, but
24240 * having both loose and full, like in charclass_invlists.h.
24241 * Perhaps move block and script to that as they are somewhat large
24242 * in charclass_invlists.h.
24243 * For properties where the default is the code point itself, such
24244 * as any of the case changing mappings, the string would otherwise
24245 * consist of all Unicode code points in UTF-8 strung together.
24246 * This would be impractical. So instead, examine their compiled
24247 * pattern, looking at the ssc. If none, reject the pattern as an
24248 * error. Otherwise run the pattern against every code point in
24249 * the ssc. The ssc is kind of like tr18's 3.9 Possible Match Sets
24250 * And it might be good to create an API to return the ssc.
24251 * Or handle them like the algorithmic names are done
24252 */
24253 } /* End of is a wildcard subppattern */
24254
24255 /* \p{name=...} is handled specially. Instead of using the normal
24256 * mechanism involving charclass_invlists.h, it uses _charnames.pm
24257 * which has the necessary (huge) data accessible to it, and which
24258 * doesn't get loaded unless necessary. The legal syntax for names is
24259 * somewhat different than other properties due both to the vagaries of
24260 * a few outlier official names, and the fact that only a few ASCII
24261 * characters are permitted in them */
24262 if ( memEQs(lookup_name, j - 1, "name")
24263 || memEQs(lookup_name, j - 1, "na"))
24264 {
24265 dSP;
24266 HV * table;
24267 SV * character;
24268 const char * error_msg;
24269 CV* lookup_loose;
24270 SV * character_name;
24271 STRLEN character_len;
24272 UV cp;
24273
24274 stricter = As_Is;
24275
24276 /* Since the RHS (after skipping initial space) is passed unchanged
24277 * to charnames, and there are different criteria for what are
24278 * legal characters in the name, just parse it here. A character
24279 * name must begin with an ASCII alphabetic */
24280 if (! isALPHA(name[i])) {
24281 goto failed;
24282 }
24283 lookup_name[j++] = name[i];
24284
24285 for (++i; i < name_len; i++) {
24286 /* Official names can only be in the ASCII range, and only
24287 * certain characters */
24288 if (! isASCII(name[i]) || ! isCHARNAME_CONT(name[i])) {
24289 goto failed;
24290 }
24291 lookup_name[j++] = name[i];
24292 }
24293
24294 /* Finished parsing, save the name into an SV */
24295 character_name = newSVpvn(lookup_name + equals_pos, j - equals_pos);
24296
24297 /* Make sure _charnames is loaded. (The parameters give context
24298 * for any errors generated */
24299 table = load_charnames(character_name, name, name_len, &error_msg);
24300 if (table == NULL) {
24301 sv_catpv(msg, error_msg);
24302 goto append_name_to_msg;
24303 }
24304
24305 lookup_loose = get_cvs("_charnames::_loose_regcomp_lookup", 0);
24306 if (! lookup_loose) {
24307 Perl_croak(aTHX_
24308 "panic: Can't find '_charnames::_loose_regcomp_lookup");
24309 }
24310
24311 PUSHSTACKi(PERLSI_REGCOMP);
24312 ENTER ;
24313 SAVETMPS;
24314 save_re_context();
24315
24316 PUSHMARK(SP) ;
24317 XPUSHs(character_name);
24318 PUTBACK;
24319 call_sv(MUTABLE_SV(lookup_loose), G_SCALAR);
24320
24321 SPAGAIN ;
24322
24323 character = POPs;
24324 SvREFCNT_inc_simple_void_NN(character);
24325
24326 PUTBACK ;
24327 FREETMPS ;
24328 LEAVE ;
24329 POPSTACK;
24330
24331 if (! SvOK(character)) {
24332 goto failed;
24333 }
24334
24335 cp = valid_utf8_to_uvchr((U8 *) SvPVX(character), &character_len);
24336 if (character_len == SvCUR(character)) {
24337 prop_definition = add_cp_to_invlist(NULL, cp);
24338 }
24339 else {
24340 AV * this_string;
24341
24342 /* First of the remaining characters in the string. */
24343 char * remaining = SvPVX(character) + character_len;
24344
24345 if (strings == NULL) {
24346 goto failed; /* XXX Perhaps a specific msg instead, like
24347 'not available here' */
24348 }
24349
24350 if (*strings == NULL) {
24351 *strings = newAV();
24352 }
24353
24354 this_string = newAV();
24355 av_push(this_string, newSVuv(cp));
24356
24357 do {
24358 cp = valid_utf8_to_uvchr((U8 *) remaining, &character_len);
24359 av_push(this_string, newSVuv(cp));
24360 remaining += character_len;
24361 } while (remaining < SvEND(character));
24362
24363 av_push(*strings, (SV *) this_string);
24364 }
24365
24366 return prop_definition;
24367 }
24368
24369 /* Certain properties whose values are numeric need special handling.
24370 * They may optionally be prefixed by 'is'. Ignore that prefix for the
24371 * purposes of checking if this is one of those properties */
24372 if (memBEGINPs(lookup_name, j, "is")) {
24373 lookup_offset = 2;
24374 }
24375
24376 /* Then check if it is one of these specially-handled properties. The
24377 * possibilities are hard-coded because easier this way, and the list
24378 * is unlikely to change.
24379 *
24380 * All numeric value type properties are of this ilk, and are also
24381 * special in a different way later on. So find those first. There
24382 * are several numeric value type properties in the Unihan DB (which is
24383 * unlikely to be compiled with perl, but we handle it here in case it
24384 * does get compiled). They all end with 'numeric'. The interiors
24385 * aren't checked for the precise property. This would stop working if
24386 * a cjk property were to be created that ended with 'numeric' and
24387 * wasn't a numeric type */
24388 is_nv_type = memEQs(lookup_name + lookup_offset,
24389 j - 1 - lookup_offset, "numericvalue")
24390 || memEQs(lookup_name + lookup_offset,
24391 j - 1 - lookup_offset, "nv")
24392 || ( memENDPs(lookup_name + lookup_offset,
24393 j - 1 - lookup_offset, "numeric")
24394 && ( memBEGINPs(lookup_name + lookup_offset,
24395 j - 1 - lookup_offset, "cjk")
24396 || memBEGINPs(lookup_name + lookup_offset,
24397 j - 1 - lookup_offset, "k")));
24398 if ( is_nv_type
24399 || memEQs(lookup_name + lookup_offset,
24400 j - 1 - lookup_offset, "canonicalcombiningclass")
24401 || memEQs(lookup_name + lookup_offset,
24402 j - 1 - lookup_offset, "ccc")
24403 || memEQs(lookup_name + lookup_offset,
24404 j - 1 - lookup_offset, "age")
24405 || memEQs(lookup_name + lookup_offset,
24406 j - 1 - lookup_offset, "in")
24407 || memEQs(lookup_name + lookup_offset,
24408 j - 1 - lookup_offset, "presentin"))
24409 {
24410 unsigned int k;
24411
24412 /* Since the stuff after the '=' is a number, we can't throw away
24413 * '-' willy-nilly, as those could be a minus sign. Other stricter
24414 * rules also apply. However, these properties all can have the
24415 * rhs not be a number, in which case they contain at least one
24416 * alphabetic. In those cases, the stricter rules don't apply.
24417 * But the numeric type properties can have the alphas [Ee] to
24418 * signify an exponent, and it is still a number with stricter
24419 * rules. So look for an alpha that signifies not-strict */
24420 stricter = Strict;
24421 for (k = i; k < name_len; k++) {
24422 if ( isALPHA_A(name[k])
24423 && (! is_nv_type || ! isALPHA_FOLD_EQ(name[k], 'E')))
24424 {
24425 stricter = Not_Strict;
24426 break;
24427 }
24428 }
24429 }
24430
24431 if (stricter) {
24432
24433 /* A number may have a leading '+' or '-'. The latter is retained
24434 * */
24435 if (name[i] == '+') {
24436 i++;
24437 }
24438 else if (name[i] == '-') {
24439 lookup_name[j++] = '-';
24440 i++;
24441 }
24442
24443 /* Skip leading zeros including single underscores separating the
24444 * zeros, or between the final leading zero and the first other
24445 * digit */
24446 for (; i < name_len - 1; i++) {
24447 if ( name[i] != '0'
24448 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
24449 {
24450 break;
24451 }
24452 }
24453 }
24454 }
24455 else { /* No '=' */
24456
24457 /* Only a few properties without an '=' should be parsed with stricter
24458 * rules. The list is unlikely to change. */
24459 if ( memBEGINPs(lookup_name, j, "perl")
24460 && memNEs(lookup_name + 4, j - 4, "space")
24461 && memNEs(lookup_name + 4, j - 4, "word"))
24462 {
24463 stricter = Strict;
24464
24465 /* We set the inputs back to 0 and the code below will reparse,
24466 * using strict */
24467 i = j = 0;
24468 }
24469 }
24470
24471 /* Here, we have either finished the property, or are positioned to parse
24472 * the remainder, and we know if stricter rules apply. Finish out, if not
24473 * already done */
24474 for (; i < name_len; i++) {
24475 char cur = name[i];
24476
24477 /* In all instances, case differences are ignored, and we normalize to
24478 * lowercase */
24479 if (isUPPER_A(cur)) {
24480 lookup_name[j++] = toLOWER(cur);
24481 continue;
24482 }
24483
24484 /* An underscore is skipped, but not under strict rules unless it
24485 * separates two digits */
24486 if (cur == '_') {
24487 if ( stricter
24488 && ( i == 0 || (int) i == equals_pos || i == name_len- 1
24489 || ! isDIGIT_A(name[i-1]) || ! isDIGIT_A(name[i+1])))
24490 {
24491 lookup_name[j++] = '_';
24492 }
24493 continue;
24494 }
24495
24496 /* Hyphens are skipped except under strict */
24497 if (cur == '-' && ! stricter) {
24498 continue;
24499 }
24500
24501 /* XXX Bug in documentation. It says white space skipped adjacent to
24502 * non-word char. Maybe we should, but shouldn't skip it next to a dot
24503 * in a number */
24504 if (isSPACE_A(cur) && ! stricter) {
24505 continue;
24506 }
24507
24508 lookup_name[j++] = cur;
24509
24510 /* Unless this is a non-trailing slash, we are done with it */
24511 if (i >= name_len - 1 || cur != '/') {
24512 continue;
24513 }
24514
24515 slash_pos = j;
24516
24517 /* A slash in the 'numeric value' property indicates that what follows
24518 * is a denominator. It can have a leading '+' and '0's that should be
24519 * skipped. But we have never allowed a negative denominator, so treat
24520 * a minus like every other character. (No need to rule out a second
24521 * '/', as that won't match anything anyway */
24522 if (is_nv_type) {
24523 i++;
24524 if (i < name_len && name[i] == '+') {
24525 i++;
24526 }
24527
24528 /* Skip leading zeros including underscores separating digits */
24529 for (; i < name_len - 1; i++) {
24530 if ( name[i] != '0'
24531 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
24532 {
24533 break;
24534 }
24535 }
24536
24537 /* Store the first real character in the denominator */
24538 if (i < name_len) {
24539 lookup_name[j++] = name[i];
24540 }
24541 }
24542 }
24543
24544 /* Here are completely done parsing the input 'name', and 'lookup_name'
24545 * contains a copy, normalized.
24546 *
24547 * This special case is grandfathered in: 'L_' and 'GC=L_' are accepted and
24548 * different from without the underscores. */
24549 if ( ( UNLIKELY(memEQs(lookup_name, j, "l"))
24550 || UNLIKELY(memEQs(lookup_name, j, "gc=l")))
24551 && UNLIKELY(name[name_len-1] == '_'))
24552 {
24553 lookup_name[j++] = '&';
24554 }
24555
24556 /* If the original input began with 'In' or 'Is', it could be a subroutine
24557 * call to a user-defined property instead of a Unicode property name. */
24558 if ( name_len - non_pkg_begin > 2
24559 && name[non_pkg_begin+0] == 'I'
24560 && (name[non_pkg_begin+1] == 'n' || name[non_pkg_begin+1] == 's'))
24561 {
24562 /* Names that start with In have different characterstics than those
24563 * that start with Is */
24564 if (name[non_pkg_begin+1] == 's') {
24565 starts_with_Is = TRUE;
24566 }
24567 }
24568 else {
24569 could_be_user_defined = FALSE;
24570 }
24571
24572 if (could_be_user_defined) {
24573 CV* user_sub;
24574
24575 /* If the user defined property returns the empty string, it could
24576 * easily be because the pattern is being compiled before the data it
24577 * actually needs to compile is available. This could be argued to be
24578 * a bug in the perl code, but this is a change of behavior for Perl,
24579 * so we handle it. This means that intentionally returning nothing
24580 * will not be resolved until runtime */
24581 bool empty_return = FALSE;
24582
24583 /* Here, the name could be for a user defined property, which are
24584 * implemented as subs. */
24585 user_sub = get_cvn_flags(name, name_len, 0);
24586 if (! user_sub) {
24587
24588 /* Here, the property name could be a user-defined one, but there
24589 * is no subroutine to handle it (as of now). Defer handling it
24590 * until runtime. Otherwise, a block defined by Unicode in a later
24591 * release would get the synonym InFoo added for it, and existing
24592 * code that used that name would suddenly break if it referred to
24593 * the property before the sub was declared. See [perl #134146] */
24594 if (deferrable) {
24595 goto definition_deferred;
24596 }
24597
24598 /* Here, we are at runtime, and didn't find the user property. It
24599 * could be an official property, but only if no package was
24600 * specified, or just the utf8:: package. */
24601 if (could_be_deferred_official) {
24602 lookup_name += lun_non_pkg_begin;
24603 j -= lun_non_pkg_begin;
24604 }
24605 else if (! stripped_utf8_pkg) {
24606 goto unknown_user_defined;
24607 }
24608
24609 /* Drop down to look up in the official properties */
24610 }
24611 else {
24612 const char insecure[] = "Insecure user-defined property";
24613
24614 /* Here, there is a sub by the correct name. Normally we call it
24615 * to get the property definition */
24616 dSP;
24617 SV * user_sub_sv = MUTABLE_SV(user_sub);
24618 SV * error; /* Any error returned by calling 'user_sub' */
24619 SV * key; /* The key into the hash of user defined sub names
24620 */
24621 SV * placeholder;
24622 SV ** saved_user_prop_ptr; /* Hash entry for this property */
24623
24624 /* How many times to retry when another thread is in the middle of
24625 * expanding the same definition we want */
24626 PERL_INT_FAST8_T retry_countdown = 10;
24627
24628 DECLARATION_FOR_GLOBAL_CONTEXT;
24629
24630 /* If we get here, we know this property is user-defined */
24631 *user_defined_ptr = TRUE;
24632
24633 /* We refuse to call a potentially tainted subroutine; returning an
24634 * error instead */
24635 if (TAINT_get) {
24636 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24637 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
24638 goto append_name_to_msg;
24639 }
24640
24641 /* In principal, we only call each subroutine property definition
24642 * once during the life of the program. This guarantees that the
24643 * property definition never changes. The results of the single
24644 * sub call are stored in a hash, which is used instead for future
24645 * references to this property. The property definition is thus
24646 * immutable. But, to allow the user to have a /i-dependent
24647 * definition, we call the sub once for non-/i, and once for /i,
24648 * should the need arise, passing the /i status as a parameter.
24649 *
24650 * We start by constructing the hash key name, consisting of the
24651 * fully qualified subroutine name, preceded by the /i status, so
24652 * that there is a key for /i and a different key for non-/i */
24653 key = newSVpvn(((to_fold) ? "1" : "0"), 1);
24654 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
24655 non_pkg_begin != 0);
24656 sv_catsv(key, fq_name);
24657 sv_2mortal(key);
24658
24659 /* We only call the sub once throughout the life of the program
24660 * (with the /i, non-/i exception noted above). That means the
24661 * hash must be global and accessible to all threads. It is
24662 * created at program start-up, before any threads are created, so
24663 * is accessible to all children. But this creates some
24664 * complications.
24665 *
24666 * 1) The keys can't be shared, or else problems arise; sharing is
24667 * turned off at hash creation time
24668 * 2) All SVs in it are there for the remainder of the life of the
24669 * program, and must be created in the same interpreter context
24670 * as the hash, or else they will be freed from the wrong pool
24671 * at global destruction time. This is handled by switching to
24672 * the hash's context to create each SV going into it, and then
24673 * immediately switching back
24674 * 3) All accesses to the hash must be controlled by a mutex, to
24675 * prevent two threads from getting an unstable state should
24676 * they simultaneously be accessing it. The code below is
24677 * crafted so that the mutex is locked whenever there is an
24678 * access and unlocked only when the next stable state is
24679 * achieved.
24680 *
24681 * The hash stores either the definition of the property if it was
24682 * valid, or, if invalid, the error message that was raised. We
24683 * use the type of SV to distinguish.
24684 *
24685 * There's also the need to guard against the definition expansion
24686 * from infinitely recursing. This is handled by storing the aTHX
24687 * of the expanding thread during the expansion. Again the SV type
24688 * is used to distinguish this from the other two cases. If we
24689 * come to here and the hash entry for this property is our aTHX,
24690 * it means we have recursed, and the code assumes that we would
24691 * infinitely recurse, so instead stops and raises an error.
24692 * (Any recursion has always been treated as infinite recursion in
24693 * this feature.)
24694 *
24695 * If instead, the entry is for a different aTHX, it means that
24696 * that thread has gotten here first, and hasn't finished expanding
24697 * the definition yet. We just have to wait until it is done. We
24698 * sleep and retry a few times, returning an error if the other
24699 * thread doesn't complete. */
24700
24701 re_fetch:
24702 USER_PROP_MUTEX_LOCK;
24703
24704 /* If we have an entry for this key, the subroutine has already
24705 * been called once with this /i status. */
24706 saved_user_prop_ptr = hv_fetch(PL_user_def_props,
24707 SvPVX(key), SvCUR(key), 0);
24708 if (saved_user_prop_ptr) {
24709
24710 /* If the saved result is an inversion list, it is the valid
24711 * definition of this property */
24712 if (is_invlist(*saved_user_prop_ptr)) {
24713 prop_definition = *saved_user_prop_ptr;
24714
24715 /* The SV in the hash won't be removed until global
24716 * destruction, so it is stable and we can unlock */
24717 USER_PROP_MUTEX_UNLOCK;
24718
24719 /* The caller shouldn't try to free this SV */
24720 return prop_definition;
24721 }
24722
24723 /* Otherwise, if it is a string, it is the error message
24724 * that was returned when we first tried to evaluate this
24725 * property. Fail, and append the message */
24726 if (SvPOK(*saved_user_prop_ptr)) {
24727 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24728 sv_catsv(msg, *saved_user_prop_ptr);
24729
24730 /* The SV in the hash won't be removed until global
24731 * destruction, so it is stable and we can unlock */
24732 USER_PROP_MUTEX_UNLOCK;
24733
24734 return NULL;
24735 }
24736
24737 assert(SvIOK(*saved_user_prop_ptr));
24738
24739 /* Here, we have an unstable entry in the hash. Either another
24740 * thread is in the middle of expanding the property's
24741 * definition, or we are ourselves recursing. We use the aTHX
24742 * in it to distinguish */
24743 if (SvIV(*saved_user_prop_ptr) != PTR2IV(CUR_CONTEXT)) {
24744
24745 /* Here, it's another thread doing the expanding. We've
24746 * looked as much as we are going to at the contents of the
24747 * hash entry. It's safe to unlock. */
24748 USER_PROP_MUTEX_UNLOCK;
24749
24750 /* Retry a few times */
24751 if (retry_countdown-- > 0) {
24752 PerlProc_sleep(1);
24753 goto re_fetch;
24754 }
24755
24756 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24757 sv_catpvs(msg, "Timeout waiting for another thread to "
24758 "define");
24759 goto append_name_to_msg;
24760 }
24761
24762 /* Here, we are recursing; don't dig any deeper */
24763 USER_PROP_MUTEX_UNLOCK;
24764
24765 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24766 sv_catpvs(msg,
24767 "Infinite recursion in user-defined property");
24768 goto append_name_to_msg;
24769 }
24770
24771 /* Here, this thread has exclusive control, and there is no entry
24772 * for this property in the hash. So we have the go ahead to
24773 * expand the definition ourselves. */
24774
24775 PUSHSTACKi(PERLSI_REGCOMP);
24776 ENTER;
24777
24778 /* Create a temporary placeholder in the hash to detect recursion
24779 * */
24780 SWITCH_TO_GLOBAL_CONTEXT;
24781 placeholder= newSVuv(PTR2IV(ORIGINAL_CONTEXT));
24782 (void) hv_store_ent(PL_user_def_props, key, placeholder, 0);
24783 RESTORE_CONTEXT;
24784
24785 /* Now that we have a placeholder, we can let other threads
24786 * continue */
24787 USER_PROP_MUTEX_UNLOCK;
24788
24789 /* Make sure the placeholder always gets destroyed */
24790 SAVEDESTRUCTOR_X(S_delete_recursion_entry, SvPVX(key));
24791
24792 PUSHMARK(SP);
24793 SAVETMPS;
24794
24795 /* Call the user's function, with the /i status as a parameter.
24796 * Note that we have gone to a lot of trouble to keep this call
24797 * from being within the locked mutex region. */
24798 XPUSHs(boolSV(to_fold));
24799 PUTBACK;
24800
24801 /* The following block was taken from swash_init(). Presumably
24802 * they apply to here as well, though we no longer use a swash --
24803 * khw */
24804 SAVEHINTS();
24805 save_re_context();
24806 /* We might get here via a subroutine signature which uses a utf8
24807 * parameter name, at which point PL_subname will have been set
24808 * but not yet used. */
24809 save_item(PL_subname);
24810
24811 /* G_SCALAR guarantees a single return value */
24812 (void) call_sv(user_sub_sv, G_EVAL|G_SCALAR);
24813
24814 SPAGAIN;
24815
24816 error = ERRSV;
24817 if (TAINT_get || SvTRUE(error)) {
24818 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24819 if (SvTRUE(error)) {
24820 sv_catpvs(msg, "Error \"");
24821 sv_catsv(msg, error);
24822 sv_catpvs(msg, "\"");
24823 }
24824 if (TAINT_get) {
24825 if (SvTRUE(error)) sv_catpvs(msg, "; ");
24826 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
24827 }
24828
24829 if (name_len > 0) {
24830 sv_catpvs(msg, " in expansion of ");
24831 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8,
24832 name_len,
24833 name));
24834 }
24835
24836 (void) POPs;
24837 prop_definition = NULL;
24838 }
24839 else {
24840 SV * contents = POPs;
24841
24842 /* The contents is supposed to be the expansion of the property
24843 * definition. If the definition is deferrable, and we got an
24844 * empty string back, set a flag to later defer it (after clean
24845 * up below). */
24846 if ( deferrable
24847 && (! SvPOK(contents) || SvCUR(contents) == 0))
24848 {
24849 empty_return = TRUE;
24850 }
24851 else { /* Otherwise, call a function to check for valid syntax,
24852 and handle it */
24853
24854 prop_definition = handle_user_defined_property(
24855 name, name_len,
24856 is_utf8, to_fold, runtime,
24857 deferrable,
24858 contents, user_defined_ptr,
24859 msg,
24860 level);
24861 }
24862 }
24863
24864 /* Here, we have the results of the expansion. Delete the
24865 * placeholder, and if the definition is now known, replace it with
24866 * that definition. We need exclusive access to the hash, and we
24867 * can't let anyone else in, between when we delete the placeholder
24868 * and add the permanent entry */
24869 USER_PROP_MUTEX_LOCK;
24870
24871 S_delete_recursion_entry(aTHX_ SvPVX(key));
24872
24873 if ( ! empty_return
24874 && (! prop_definition || is_invlist(prop_definition)))
24875 {
24876 /* If we got success we use the inversion list defining the
24877 * property; otherwise use the error message */
24878 SWITCH_TO_GLOBAL_CONTEXT;
24879 (void) hv_store_ent(PL_user_def_props,
24880 key,
24881 ((prop_definition)
24882 ? newSVsv(prop_definition)
24883 : newSVsv(msg)),
24884 0);
24885 RESTORE_CONTEXT;
24886 }
24887
24888 /* All done, and the hash now has a permanent entry for this
24889 * property. Give up exclusive control */
24890 USER_PROP_MUTEX_UNLOCK;
24891
24892 FREETMPS;
24893 LEAVE;
24894 POPSTACK;
24895
24896 if (empty_return) {
24897 goto definition_deferred;
24898 }
24899
24900 if (prop_definition) {
24901
24902 /* If the definition is for something not known at this time,
24903 * we toss it, and go return the main property name, as that's
24904 * the one the user will be aware of */
24905 if (! is_invlist(prop_definition)) {
24906 SvREFCNT_dec_NN(prop_definition);
24907 goto definition_deferred;
24908 }
24909
24910 sv_2mortal(prop_definition);
24911 }
24912
24913 /* And return */
24914 return prop_definition;
24915
24916 } /* End of calling the subroutine for the user-defined property */
24917 } /* End of it could be a user-defined property */
24918
24919 /* Here it wasn't a user-defined property that is known at this time. See
24920 * if it is a Unicode property */
24921
24922 lookup_len = j; /* This is a more mnemonic name than 'j' */
24923
24924 /* Get the index into our pointer table of the inversion list corresponding
24925 * to the property */
24926 table_index = do_uniprop_match(lookup_name, lookup_len);
24927
24928 /* If it didn't find the property ... */
24929 if (table_index == 0) {
24930
24931 /* Try again stripping off any initial 'Is'. This is because we
24932 * promise that an initial Is is optional. The same isn't true of
24933 * names that start with 'In'. Those can match only blocks, and the
24934 * lookup table already has those accounted for. The lookup table also
24935 * has already accounted for Perl extensions (without and = sign)
24936 * starting with 'i's'. */
24937 if (starts_with_Is && equals_pos >= 0) {
24938 lookup_name += 2;
24939 lookup_len -= 2;
24940 equals_pos -= 2;
24941 slash_pos -= 2;
24942
24943 table_index = do_uniprop_match(lookup_name, lookup_len);
24944 }
24945
24946 if (table_index == 0) {
24947 char * canonical;
24948
24949 /* Here, we didn't find it. If not a numeric type property, and
24950 * can't be a user-defined one, it isn't a legal property */
24951 if (! is_nv_type) {
24952 if (! could_be_user_defined) {
24953 goto failed;
24954 }
24955
24956 /* Here, the property name is legal as a user-defined one. At
24957 * compile time, it might just be that the subroutine for that
24958 * property hasn't been encountered yet, but at runtime, it's
24959 * an error to try to use an undefined one */
24960 if (! deferrable) {
24961 goto unknown_user_defined;;
24962 }
24963
24964 goto definition_deferred;
24965 } /* End of isn't a numeric type property */
24966
24967 /* The numeric type properties need more work to decide. What we
24968 * do is make sure we have the number in canonical form and look
24969 * that up. */
24970
24971 if (slash_pos < 0) { /* No slash */
24972
24973 /* When it isn't a rational, take the input, convert it to a
24974 * NV, then create a canonical string representation of that
24975 * NV. */
24976
24977 NV value;
24978 SSize_t value_len = lookup_len - equals_pos;
24979
24980 /* Get the value */
24981 if ( value_len <= 0
24982 || my_atof3(lookup_name + equals_pos, &value,
24983 value_len)
24984 != lookup_name + lookup_len)
24985 {
24986 goto failed;
24987 }
24988
24989 /* If the value is an integer, the canonical value is integral
24990 * */
24991 if (Perl_ceil(value) == value) {
24992 canonical = Perl_form(aTHX_ "%.*s%.0" NVff,
24993 equals_pos, lookup_name, value);
24994 }
24995 else { /* Otherwise, it is %e with a known precision */
24996 char * exp_ptr;
24997
24998 canonical = Perl_form(aTHX_ "%.*s%.*" NVef,
24999 equals_pos, lookup_name,
25000 PL_E_FORMAT_PRECISION, value);
25001
25002 /* The exponent generated is expecting two digits, whereas
25003 * %e on some systems will generate three. Remove leading
25004 * zeros in excess of 2 from the exponent. We start
25005 * looking for them after the '=' */
25006 exp_ptr = strchr(canonical + equals_pos, 'e');
25007 if (exp_ptr) {
25008 char * cur_ptr = exp_ptr + 2; /* past the 'e[+-]' */
25009 SSize_t excess_exponent_len = strlen(cur_ptr) - 2;
25010
25011 assert(*(cur_ptr - 1) == '-' || *(cur_ptr - 1) == '+');
25012
25013 if (excess_exponent_len > 0) {
25014 SSize_t leading_zeros = strspn(cur_ptr, "0");
25015 SSize_t excess_leading_zeros
25016 = MIN(leading_zeros, excess_exponent_len);
25017 if (excess_leading_zeros > 0) {
25018 Move(cur_ptr + excess_leading_zeros,
25019 cur_ptr,
25020 strlen(cur_ptr) - excess_leading_zeros
25021 + 1, /* Copy the NUL as well */
25022 char);
25023 }
25024 }
25025 }
25026 }
25027 }
25028 else { /* Has a slash. Create a rational in canonical form */
25029 UV numerator, denominator, gcd, trial;
25030 const char * end_ptr;
25031 const char * sign = "";
25032
25033 /* We can't just find the numerator, denominator, and do the
25034 * division, then use the method above, because that is
25035 * inexact. And the input could be a rational that is within
25036 * epsilon (given our precision) of a valid rational, and would
25037 * then incorrectly compare valid.
25038 *
25039 * We're only interested in the part after the '=' */
25040 const char * this_lookup_name = lookup_name + equals_pos;
25041 lookup_len -= equals_pos;
25042 slash_pos -= equals_pos;
25043
25044 /* Handle any leading minus */
25045 if (this_lookup_name[0] == '-') {
25046 sign = "-";
25047 this_lookup_name++;
25048 lookup_len--;
25049 slash_pos--;
25050 }
25051
25052 /* Convert the numerator to numeric */
25053 end_ptr = this_lookup_name + slash_pos;
25054 if (! grok_atoUV(this_lookup_name, &numerator, &end_ptr)) {
25055 goto failed;
25056 }
25057
25058 /* It better have included all characters before the slash */
25059 if (*end_ptr != '/') {
25060 goto failed;
25061 }
25062
25063 /* Set to look at just the denominator */
25064 this_lookup_name += slash_pos;
25065 lookup_len -= slash_pos;
25066 end_ptr = this_lookup_name + lookup_len;
25067
25068 /* Convert the denominator to numeric */
25069 if (! grok_atoUV(this_lookup_name, &denominator, &end_ptr)) {
25070 goto failed;
25071 }
25072
25073 /* It better be the rest of the characters, and don't divide by
25074 * 0 */
25075 if ( end_ptr != this_lookup_name + lookup_len
25076 || denominator == 0)
25077 {
25078 goto failed;
25079 }
25080
25081 /* Get the greatest common denominator using
25082 http://en.wikipedia.org/wiki/Euclidean_algorithm */
25083 gcd = numerator;
25084 trial = denominator;
25085 while (trial != 0) {
25086 UV temp = trial;
25087 trial = gcd % trial;
25088 gcd = temp;
25089 }
25090
25091 /* If already in lowest possible terms, we have already tried
25092 * looking this up */
25093 if (gcd == 1) {
25094 goto failed;
25095 }
25096
25097 /* Reduce the rational, which should put it in canonical form
25098 * */
25099 numerator /= gcd;
25100 denominator /= gcd;
25101
25102 canonical = Perl_form(aTHX_ "%.*s%s%" UVuf "/%" UVuf,
25103 equals_pos, lookup_name, sign, numerator, denominator);
25104 }
25105
25106 /* Here, we have the number in canonical form. Try that */
25107 table_index = do_uniprop_match(canonical, strlen(canonical));
25108 if (table_index == 0) {
25109 goto failed;
25110 }
25111 } /* End of still didn't find the property in our table */
25112 } /* End of didn't find the property in our table */
25113
25114 /* Here, we have a non-zero return, which is an index into a table of ptrs.
25115 * A negative return signifies that the real index is the absolute value,
25116 * but the result needs to be inverted */
25117 if (table_index < 0) {
25118 invert_return = TRUE;
25119 table_index = -table_index;
25120 }
25121
25122 /* Out-of band indices indicate a deprecated property. The proper index is
25123 * modulo it with the table size. And dividing by the table size yields
25124 * an offset into a table constructed by regen/mk_invlists.pl to contain
25125 * the corresponding warning message */
25126 if (table_index > MAX_UNI_KEYWORD_INDEX) {
25127 Size_t warning_offset = table_index / MAX_UNI_KEYWORD_INDEX;
25128 table_index %= MAX_UNI_KEYWORD_INDEX;
25129 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED),
25130 "Use of '%.*s' in \\p{} or \\P{} is deprecated because: %s",
25131 (int) name_len, name,
25132 get_deprecated_property_msg(warning_offset));
25133 }
25134
25135 /* In a few properties, a different property is used under /i. These are
25136 * unlikely to change, so are hard-coded here. */
25137 if (to_fold) {
25138 if ( table_index == UNI_XPOSIXUPPER
25139 || table_index == UNI_XPOSIXLOWER
25140 || table_index == UNI_TITLE)
25141 {
25142 table_index = UNI_CASED;
25143 }
25144 else if ( table_index == UNI_UPPERCASELETTER
25145 || table_index == UNI_LOWERCASELETTER
25146 # ifdef UNI_TITLECASELETTER /* Missing from early Unicodes */
25147 || table_index == UNI_TITLECASELETTER
25148 # endif
25149 ) {
25150 table_index = UNI_CASEDLETTER;
25151 }
25152 else if ( table_index == UNI_POSIXUPPER
25153 || table_index == UNI_POSIXLOWER)
25154 {
25155 table_index = UNI_POSIXALPHA;
25156 }
25157 }
25158
25159 /* Create and return the inversion list */
25160 prop_definition = get_prop_definition(table_index);
25161 sv_2mortal(prop_definition);
25162
25163 /* See if there is a private use override to add to this definition */
25164 {
25165 COPHH * hinthash = (IN_PERL_COMPILETIME)
25166 ? CopHINTHASH_get(&PL_compiling)
25167 : CopHINTHASH_get(PL_curcop);
25168 SV * pu_overrides = cophh_fetch_pv(hinthash, "private_use", 0, 0);
25169
25170 if (UNLIKELY(pu_overrides && SvPOK(pu_overrides))) {
25171
25172 /* See if there is an element in the hints hash for this table */
25173 SV * pu_lookup = Perl_newSVpvf(aTHX_ "%d=", table_index);
25174 const char * pos = strstr(SvPVX(pu_overrides), SvPVX(pu_lookup));
25175
25176 if (pos) {
25177 bool dummy;
25178 SV * pu_definition;
25179 SV * pu_invlist;
25180 SV * expanded_prop_definition =
25181 sv_2mortal(invlist_clone(prop_definition, NULL));
25182
25183 /* If so, it's definition is the string from here to the next
25184 * \a character. And its format is the same as a user-defined
25185 * property */
25186 pos += SvCUR(pu_lookup);
25187 pu_definition = newSVpvn(pos, strchr(pos, '\a') - pos);
25188 pu_invlist = handle_user_defined_property(lookup_name,
25189 lookup_len,
25190 0, /* Not UTF-8 */
25191 0, /* Not folded */
25192 runtime,
25193 deferrable,
25194 pu_definition,
25195 &dummy,
25196 msg,
25197 level);
25198 if (TAINT_get) {
25199 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
25200 sv_catpvs(msg, "Insecure private-use override");
25201 goto append_name_to_msg;
25202 }
25203
25204 /* For now, as a safety measure, make sure that it doesn't
25205 * override non-private use code points */
25206 _invlist_intersection(pu_invlist, PL_Private_Use, &pu_invlist);
25207
25208 /* Add it to the list to be returned */
25209 _invlist_union(prop_definition, pu_invlist,
25210 &expanded_prop_definition);
25211 prop_definition = expanded_prop_definition;
25212 Perl_ck_warner_d(aTHX_ packWARN(WARN_EXPERIMENTAL__PRIVATE_USE), "The private_use feature is experimental");
25213 }
25214 }
25215 }
25216
25217 if (invert_return) {
25218 _invlist_invert(prop_definition);
25219 }
25220 return prop_definition;
25221
25222 unknown_user_defined:
25223 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
25224 sv_catpvs(msg, "Unknown user-defined property name");
25225 goto append_name_to_msg;
25226
25227 failed:
25228 if (non_pkg_begin != 0) {
25229 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
25230 sv_catpvs(msg, "Illegal user-defined property name");
25231 }
25232 else {
25233 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
25234 sv_catpvs(msg, "Can't find Unicode property definition");
25235 }
25236 /* FALLTHROUGH */
25237
25238 append_name_to_msg:
25239 {
25240 const char * prefix = (runtime && level == 0) ? " \\p{" : " \"";
25241 const char * suffix = (runtime && level == 0) ? "}" : "\"";
25242
25243 sv_catpv(msg, prefix);
25244 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
25245 sv_catpv(msg, suffix);
25246 }
25247
25248 return NULL;
25249
25250 definition_deferred:
25251
25252 {
25253 bool is_qualified = non_pkg_begin != 0; /* If has "::" */
25254
25255 /* Here it could yet to be defined, so defer evaluation of this until
25256 * its needed at runtime. We need the fully qualified property name to
25257 * avoid ambiguity */
25258 if (! fq_name) {
25259 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
25260 is_qualified);
25261 }
25262
25263 /* If it didn't come with a package, or the package is utf8::, this
25264 * actually could be an official Unicode property whose inclusion we
25265 * are deferring until runtime to make sure that it isn't overridden by
25266 * a user-defined property of the same name (which we haven't
25267 * encountered yet). Add a marker to indicate this possibility, for
25268 * use at such time when we first need the definition during pattern
25269 * matching execution */
25270 if (! is_qualified || memBEGINPs(name, non_pkg_begin, "utf8::")) {
25271 sv_catpvs(fq_name, DEFERRED_COULD_BE_OFFICIAL_MARKERs);
25272 }
25273
25274 /* We also need a trailing newline */
25275 sv_catpvs(fq_name, "\n");
25276
25277 *user_defined_ptr = TRUE;
25278 return fq_name;
25279 }
25280 }
25281
25282 STATIC bool
S_handle_names_wildcard(pTHX_ const char * wname,const STRLEN wname_len,SV ** prop_definition,AV ** strings)25283 S_handle_names_wildcard(pTHX_ const char * wname, /* wildcard name to match */
25284 const STRLEN wname_len, /* Its length */
25285 SV ** prop_definition,
25286 AV ** strings)
25287 {
25288 /* Deal with Name property wildcard subpatterns; returns TRUE if there were
25289 * any matches, adding them to prop_definition */
25290
25291 dSP;
25292
25293 CV * get_names_info; /* entry to charnames.pm to get info we need */
25294 SV * names_string; /* Contains all character names, except algo */
25295 SV * algorithmic_names; /* Contains info about algorithmically
25296 generated character names */
25297 REGEXP * subpattern_re; /* The user's pattern to match with */
25298 struct regexp * prog; /* The compiled pattern */
25299 char * all_names_start; /* lib/unicore/Name.pl string of every
25300 (non-algorithmic) character name */
25301 char * cur_pos; /* We match, effectively using /gc; this is
25302 where we are now */
25303 bool found_matches = FALSE; /* Did any name match so far? */
25304 SV * empty; /* For matching zero length names */
25305 SV * must_sv; /* Contains the substring, if any, that must be
25306 in a name for the subpattern to match */
25307 const char * must; /* The PV of 'must' */
25308 STRLEN must_len; /* And its length */
25309 SV * syllable_name = NULL; /* For Hangul syllables */
25310 const char hangul_prefix[] = "HANGUL SYLLABLE ";
25311 const STRLEN hangul_prefix_len = sizeof(hangul_prefix) - 1;
25312
25313 /* By inspection, there are a maximum of 7 bytes in the suffix of a hangul
25314 * syllable name, and these are immutable and guaranteed by the Unicode
25315 * standard to never be extended */
25316 const STRLEN syl_max_len = hangul_prefix_len + 7;
25317
25318 IV i;
25319
25320 PERL_ARGS_ASSERT_HANDLE_NAMES_WILDCARD;
25321
25322 /* Make sure _charnames is loaded. (The parameters give context
25323 * for any errors generated */
25324 get_names_info = get_cv("_charnames::_get_names_info", 0);
25325 if (! get_names_info) {
25326 Perl_croak(aTHX_ "panic: Can't find '_charnames::_get_names_info");
25327 }
25328
25329 /* Get the charnames data */
25330 PUSHSTACKi(PERLSI_REGCOMP);
25331 ENTER ;
25332 SAVETMPS;
25333 save_re_context();
25334
25335 PUSHMARK(SP) ;
25336 PUTBACK;
25337
25338 /* Special _charnames entry point that returns the info this routine
25339 * requires */
25340 call_sv(MUTABLE_SV(get_names_info), G_LIST);
25341
25342 SPAGAIN ;
25343
25344 /* Data structure for names which end in their very own code points */
25345 algorithmic_names = POPs;
25346 SvREFCNT_inc_simple_void_NN(algorithmic_names);
25347
25348 /* The lib/unicore/Name.pl string */
25349 names_string = POPs;
25350 SvREFCNT_inc_simple_void_NN(names_string);
25351
25352 PUTBACK ;
25353 FREETMPS ;
25354 LEAVE ;
25355 POPSTACK;
25356
25357 if ( ! SvROK(names_string)
25358 || ! SvROK(algorithmic_names))
25359 { /* Perhaps should panic instead XXX */
25360 SvREFCNT_dec(names_string);
25361 SvREFCNT_dec(algorithmic_names);
25362 return FALSE;
25363 }
25364
25365 names_string = sv_2mortal(SvRV(names_string));
25366 all_names_start = SvPVX(names_string);
25367 cur_pos = all_names_start;
25368
25369 algorithmic_names= sv_2mortal(SvRV(algorithmic_names));
25370
25371 /* Compile the subpattern consisting of the name being looked for */
25372 subpattern_re = compile_wildcard(wname, wname_len, FALSE /* /-i */ );
25373
25374 must_sv = re_intuit_string(subpattern_re);
25375 if (must_sv) {
25376 /* regexec.c can free the re_intuit_string() return. GH #17734 */
25377 must_sv = sv_2mortal(newSVsv(must_sv));
25378 must = SvPV(must_sv, must_len);
25379 }
25380 else {
25381 must = "";
25382 must_len = 0;
25383 }
25384
25385 /* (Note: 'must' could contain a NUL. And yet we use strspn() below on it.
25386 * This works because the NUL causes the function to return early, thus
25387 * showing that there are characters in it other than the acceptable ones,
25388 * which is our desired result.) */
25389
25390 prog = ReANY(subpattern_re);
25391
25392 /* If only nothing is matched, skip to where empty names are looked for */
25393 if (prog->maxlen == 0) {
25394 goto check_empty;
25395 }
25396
25397 /* And match against the string of all names /gc. Don't even try if it
25398 * must match a character not found in any name. */
25399 if (strspn(must, "\n -0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ()") == must_len)
25400 {
25401 while (execute_wildcard(subpattern_re,
25402 cur_pos,
25403 SvEND(names_string),
25404 all_names_start, 0,
25405 names_string,
25406 0))
25407 { /* Here, matched. */
25408
25409 /* Note the string entries look like
25410 * 00001\nSTART OF HEADING\n\n
25411 * so we could match anywhere in that string. We have to rule out
25412 * matching a code point line */
25413 char * this_name_start = all_names_start
25414 + RX_OFFS(subpattern_re)->start;
25415 char * this_name_end = all_names_start
25416 + RX_OFFS(subpattern_re)->end;
25417 char * cp_start;
25418 char * cp_end;
25419 UV cp = 0; /* Silences some compilers */
25420 AV * this_string = NULL;
25421 bool is_multi = FALSE;
25422
25423 /* If matched nothing, advance to next possible match */
25424 if (this_name_start == this_name_end) {
25425 cur_pos = (char *) memchr(this_name_end + 1, '\n',
25426 SvEND(names_string) - this_name_end);
25427 if (cur_pos == NULL) {
25428 break;
25429 }
25430 }
25431 else {
25432 /* Position the next match to start beyond the current returned
25433 * entry */
25434 cur_pos = (char *) memchr(this_name_end, '\n',
25435 SvEND(names_string) - this_name_end);
25436 }
25437
25438 /* Back up to the \n just before the beginning of the character. */
25439 cp_end = (char *) my_memrchr(all_names_start,
25440 '\n',
25441 this_name_start - all_names_start);
25442
25443 /* If we didn't find a \n, it means it matched somewhere in the
25444 * initial '00000' in the string, so isn't a real match */
25445 if (cp_end == NULL) {
25446 continue;
25447 }
25448
25449 this_name_start = cp_end + 1; /* The name starts just after */
25450 cp_end--; /* the \n, and the code point */
25451 /* ends just before it */
25452
25453 /* All code points are 5 digits long */
25454 cp_start = cp_end - 4;
25455
25456 /* This shouldn't happen, as we found a \n, and the first \n is
25457 * further along than what we subtracted */
25458 assert(cp_start >= all_names_start);
25459
25460 if (cp_start == all_names_start) {
25461 *prop_definition = add_cp_to_invlist(*prop_definition, 0);
25462 continue;
25463 }
25464
25465 /* If the character is a blank, we either have a named sequence, or
25466 * something is wrong */
25467 if (*(cp_start - 1) == ' ') {
25468 cp_start = (char *) my_memrchr(all_names_start,
25469 '\n',
25470 cp_start - all_names_start);
25471 cp_start++;
25472 }
25473
25474 assert(cp_start != NULL && cp_start >= all_names_start + 2);
25475
25476 /* Except for the first line in the string, the sequence before the
25477 * code point is \n\n. If that isn't the case here, we didn't
25478 * match the name of a character. (We could have matched a named
25479 * sequence, not currently handled */
25480 if (*(cp_start - 1) != '\n' || *(cp_start - 2) != '\n') {
25481 continue;
25482 }
25483
25484 /* We matched! Add this to the list */
25485 found_matches = TRUE;
25486
25487 /* Loop through all the code points in the sequence */
25488 while (cp_start < cp_end) {
25489
25490 /* Calculate this code point from its 5 digits */
25491 cp = (XDIGIT_VALUE(cp_start[0]) << 16)
25492 + (XDIGIT_VALUE(cp_start[1]) << 12)
25493 + (XDIGIT_VALUE(cp_start[2]) << 8)
25494 + (XDIGIT_VALUE(cp_start[3]) << 4)
25495 + XDIGIT_VALUE(cp_start[4]);
25496
25497 cp_start += 6; /* Go past any blank */
25498
25499 if (cp_start < cp_end || is_multi) {
25500 if (this_string == NULL) {
25501 this_string = newAV();
25502 }
25503
25504 is_multi = TRUE;
25505 av_push(this_string, newSVuv(cp));
25506 }
25507 }
25508
25509 if (is_multi) { /* Was more than one code point */
25510 if (*strings == NULL) {
25511 *strings = newAV();
25512 }
25513
25514 av_push(*strings, (SV *) this_string);
25515 }
25516 else { /* Only a single code point */
25517 *prop_definition = add_cp_to_invlist(*prop_definition, cp);
25518 }
25519 } /* End of loop through the non-algorithmic names string */
25520 }
25521
25522 /* There are also character names not in 'names_string'. These are
25523 * algorithmically generatable. Try this pattern on each possible one.
25524 * (khw originally planned to leave this out given the large number of
25525 * matches attempted; but the speed turned out to be quite acceptable
25526 *
25527 * There are plenty of opportunities to optimize to skip many of the tests.
25528 * beyond the rudimentary ones already here */
25529
25530 /* First see if the subpattern matches any of the algorithmic generatable
25531 * Hangul syllable names.
25532 *
25533 * We know none of these syllable names will match if the input pattern
25534 * requires more bytes than any syllable has, or if the input pattern only
25535 * matches an empty name, or if the pattern has something it must match and
25536 * one of the characters in that isn't in any Hangul syllable. */
25537 if ( prog->minlen <= (SSize_t) syl_max_len
25538 && prog->maxlen > 0
25539 && (strspn(must, "\n ABCDEGHIJKLMNOPRSTUWY") == must_len))
25540 {
25541 /* These constants, names, values, and algorithm are adapted from the
25542 * Unicode standard, version 5.1, section 3.12, and should never
25543 * change. */
25544 const char * JamoL[] = {
25545 "G", "GG", "N", "D", "DD", "R", "M", "B", "BB",
25546 "S", "SS", "", "J", "JJ", "C", "K", "T", "P", "H"
25547 };
25548 const int LCount = C_ARRAY_LENGTH(JamoL);
25549
25550 const char * JamoV[] = {
25551 "A", "AE", "YA", "YAE", "EO", "E", "YEO", "YE", "O", "WA",
25552 "WAE", "OE", "YO", "U", "WEO", "WE", "WI", "YU", "EU", "YI",
25553 "I"
25554 };
25555 const int VCount = C_ARRAY_LENGTH(JamoV);
25556
25557 const char * JamoT[] = {
25558 "", "G", "GG", "GS", "N", "NJ", "NH", "D", "L",
25559 "LG", "LM", "LB", "LS", "LT", "LP", "LH", "M", "B",
25560 "BS", "S", "SS", "NG", "J", "C", "K", "T", "P", "H"
25561 };
25562 const int TCount = C_ARRAY_LENGTH(JamoT);
25563
25564 int L, V, T;
25565
25566 /* This is the initial Hangul syllable code point; each time through the
25567 * inner loop, it maps to the next higher code point. For more info,
25568 * see the Hangul syllable section of the Unicode standard. */
25569 int cp = 0xAC00;
25570
25571 syllable_name = sv_2mortal(newSV(syl_max_len));
25572 sv_setpvn(syllable_name, hangul_prefix, hangul_prefix_len);
25573
25574 for (L = 0; L < LCount; L++) {
25575 for (V = 0; V < VCount; V++) {
25576 for (T = 0; T < TCount; T++) {
25577
25578 /* Truncate back to the prefix, which is unvarying */
25579 SvCUR_set(syllable_name, hangul_prefix_len);
25580
25581 sv_catpv(syllable_name, JamoL[L]);
25582 sv_catpv(syllable_name, JamoV[V]);
25583 sv_catpv(syllable_name, JamoT[T]);
25584
25585 if (execute_wildcard(subpattern_re,
25586 SvPVX(syllable_name),
25587 SvEND(syllable_name),
25588 SvPVX(syllable_name), 0,
25589 syllable_name,
25590 0))
25591 {
25592 *prop_definition = add_cp_to_invlist(*prop_definition,
25593 cp);
25594 found_matches = TRUE;
25595 }
25596
25597 cp++;
25598 }
25599 }
25600 }
25601 }
25602
25603 /* The rest of the algorithmically generatable names are of the form
25604 * "PREFIX-code_point". The prefixes and the code point limits of each
25605 * were returned to us in the array 'algorithmic_names' from data in
25606 * lib/unicore/Name.pm. 'code_point' in the name is expressed in hex. */
25607 for (i = 0; i <= av_top_index((AV *) algorithmic_names); i++) {
25608 IV j;
25609
25610 /* Each element of the array is a hash, giving the details for the
25611 * series of names it covers. There is the base name of the characters
25612 * in the series, and the low and high code points in the series. And,
25613 * for optimization purposes a string containing all the legal
25614 * characters that could possibly be in a name in this series. */
25615 HV * this_series = (HV *) SvRV(* av_fetch((AV *) algorithmic_names, i, 0));
25616 SV * prefix = * hv_fetchs(this_series, "name", 0);
25617 IV low = SvIV(* hv_fetchs(this_series, "low", 0));
25618 IV high = SvIV(* hv_fetchs(this_series, "high", 0));
25619 char * legal = SvPVX(* hv_fetchs(this_series, "legal", 0));
25620
25621 /* Pre-allocate an SV with enough space */
25622 SV * algo_name = sv_2mortal(Perl_newSVpvf(aTHX_ "%s-0000",
25623 SvPVX(prefix)));
25624 if (high >= 0x10000) {
25625 sv_catpvs(algo_name, "0");
25626 }
25627
25628 /* This series can be skipped entirely if the pattern requires
25629 * something longer than any name in the series, or can only match an
25630 * empty name, or contains a character not found in any name in the
25631 * series */
25632 if ( prog->minlen <= (SSize_t) SvCUR(algo_name)
25633 && prog->maxlen > 0
25634 && (strspn(must, legal) == must_len))
25635 {
25636 for (j = low; j <= high; j++) { /* For each code point in the series */
25637
25638 /* Get its name, and see if it matches the subpattern */
25639 Perl_sv_setpvf(aTHX_ algo_name, "%s-%X", SvPVX(prefix),
25640 (unsigned) j);
25641
25642 if (execute_wildcard(subpattern_re,
25643 SvPVX(algo_name),
25644 SvEND(algo_name),
25645 SvPVX(algo_name), 0,
25646 algo_name,
25647 0))
25648 {
25649 *prop_definition = add_cp_to_invlist(*prop_definition, j);
25650 found_matches = TRUE;
25651 }
25652 }
25653 }
25654 }
25655
25656 check_empty:
25657 /* Finally, see if the subpattern matches an empty string */
25658 empty = newSVpvs("");
25659 if (execute_wildcard(subpattern_re,
25660 SvPVX(empty),
25661 SvEND(empty),
25662 SvPVX(empty), 0,
25663 empty,
25664 0))
25665 {
25666 /* Many code points have empty names. Currently these are the \p{GC=C}
25667 * ones, minus CC and CF */
25668
25669 SV * empty_names_ref = get_prop_definition(UNI_C);
25670 SV * empty_names = invlist_clone(empty_names_ref, NULL);
25671
25672 SV * subtract = get_prop_definition(UNI_CC);
25673
25674 _invlist_subtract(empty_names, subtract, &empty_names);
25675 SvREFCNT_dec_NN(empty_names_ref);
25676 SvREFCNT_dec_NN(subtract);
25677
25678 subtract = get_prop_definition(UNI_CF);
25679 _invlist_subtract(empty_names, subtract, &empty_names);
25680 SvREFCNT_dec_NN(subtract);
25681
25682 _invlist_union(*prop_definition, empty_names, prop_definition);
25683 found_matches = TRUE;
25684 SvREFCNT_dec_NN(empty_names);
25685 }
25686 SvREFCNT_dec_NN(empty);
25687
25688 #if 0
25689 /* If we ever were to accept aliases for, say private use names, we would
25690 * need to do something fancier to find empty names. The code below works
25691 * (at the time it was written), and is slower than the above */
25692 const char empties_pat[] = "^.";
25693 if (strNE(name, empties_pat)) {
25694 SV * empty = newSVpvs("");
25695 if (execute_wildcard(subpattern_re,
25696 SvPVX(empty),
25697 SvEND(empty),
25698 SvPVX(empty), 0,
25699 empty,
25700 0))
25701 {
25702 SV * empties = NULL;
25703
25704 (void) handle_names_wildcard(empties_pat, strlen(empties_pat), &empties);
25705
25706 _invlist_union_complement_2nd(*prop_definition, empties, prop_definition);
25707 SvREFCNT_dec_NN(empties);
25708
25709 found_matches = TRUE;
25710 }
25711 SvREFCNT_dec_NN(empty);
25712 }
25713 #endif
25714
25715 SvREFCNT_dec_NN(subpattern_re);
25716 return found_matches;
25717 }
25718
25719 /*
25720 * ex: set ts=8 sts=4 sw=4 et:
25721 */
25722