5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
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.
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.
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!
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.
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.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
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:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
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
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.
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.
74 /* Note on debug output:
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.
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
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
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
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
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
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.
131 #define PERL_IN_REGCOMP_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;
140 # include "regcomp.h"
143 #include "invlist_inline.h"
144 #include "unicode_constants.h"
147 #define STATIC static
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. */
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 */
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 */
166 /* Certain characters are output as a sequence with the first being a
168 #define isBACKSLASHED_PUNCT(c) memCHRs("-[]\\^", c)
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
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
196 Size_t sets_depth; /* Counts recursion depth of already-
197 compiled regex set patterns */
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 */
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;
212 I32 npar; /* Capture buffer count so far in the
213 parse, (OPEN) plus one. ("par" 0 is
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
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
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
237 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
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 (?{})
246 SSize_t maxlen; /* mininum possible number of chars in string to match */
247 scan_frame *frame_head;
248 scan_frame *frame_last;
252 SV *runtime_code_qr; /* qr with the runtime code blocks */
254 const char *lastparse;
256 U32 study_chunk_recursed_count;
257 AV *paren_name_list; /* idx -> name */
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)
275 bool sWARN_EXPERIMENTAL__VLB;
276 bool sWARN_EXPERIMENTAL__REGEX_SETS;
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 ? */
297 #ifdef RE_TRACK_PATTERN_OFFSETS
298 # define RExC_offsets (RExC_rxi->u.offsets) /* I am not like the
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)
330 # define SET_recode_x_to_native(x) \
331 STMT_START { RExC_recode_x_to_native = (x); } STMT_END
333 # define SET_recode_x_to_native(x) NOOP
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)
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.
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)
363 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
364 #define ISMULT2(s) (ISMULT1(*s) || ((*s) == '{' && regcurly(s)))
367 * Flags to be passed up and down.
369 #define HASWIDTH 0x01 /* Known to not match null strings, could match
371 #define SIMPLE 0x02 /* Exactly one character wide */
372 /* (or LNBREAK as a special case) */
373 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
374 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
375 #define RESTART_PARSE 0x20 /* Need to redo the parse */
376 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PARSE, need to
377 calcuate sizes as UTF-8 */
379 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
381 /* whether trie related optimizations are enabled */
382 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
383 #define TRIE_STUDY_OPT
384 #define FULL_TRIE_STUDY
390 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
391 #define PBITVAL(paren) (1 << ((paren) & 7))
392 #define PAREN_OFFSET(depth) \
393 (RExC_study_chunk_recursed + (depth) * RExC_study_chunk_recursed_bytes)
394 #define PAREN_TEST(depth, paren) \
395 (PBYTE(PAREN_OFFSET(depth), paren) & PBITVAL(paren))
396 #define PAREN_SET(depth, paren) \
397 (PBYTE(PAREN_OFFSET(depth), paren) |= PBITVAL(paren))
398 #define PAREN_UNSET(depth, paren) \
399 (PBYTE(PAREN_OFFSET(depth), paren) &= ~PBITVAL(paren))
401 #define REQUIRE_UTF8(flagp) STMT_START { \
403 *flagp = RESTART_PARSE|NEED_UTF8; \
408 /* /u is to be chosen if we are supposed to use Unicode rules, or if the
409 * pattern is in UTF-8. This latter condition is in case the outermost rules
410 * are locale. See GH #17278 */
411 #define toUSE_UNI_CHARSET_NOT_DEPENDS (RExC_uni_semantics || UTF)
413 /* Change from /d into /u rules, and restart the parse. RExC_uni_semantics is
414 * a flag that indicates we need to override /d with /u as a result of
415 * something in the pattern. It should only be used in regards to calling
416 * set_regex_charset() or get_regex_charset() */
417 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
419 if (DEPENDS_SEMANTICS) { \
420 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
421 RExC_uni_semantics = 1; \
422 if (RExC_seen_d_op && LIKELY(! IN_PARENS_PASS)) { \
423 /* No need to restart the parse if we haven't seen \
424 * anything that differs between /u and /d, and no need \
425 * to restart immediately if we're going to reparse \
426 * anyway to count parens */ \
427 *flagp |= RESTART_PARSE; \
428 return restart_retval; \
433 #define REQUIRE_BRANCHJ(flagp, restart_retval) \
435 RExC_use_BRANCHJ = 1; \
436 *flagp |= RESTART_PARSE; \
437 return restart_retval; \
440 /* Until we have completed the parse, we leave RExC_total_parens at 0 or
441 * less. After that, it must always be positive, because the whole re is
442 * considered to be surrounded by virtual parens. Setting it to negative
443 * indicates there is some construct that needs to know the actual number of
444 * parens to be properly handled. And that means an extra pass will be
445 * required after we've counted them all */
446 #define ALL_PARENS_COUNTED (RExC_total_parens > 0)
447 #define REQUIRE_PARENS_PASS \
448 STMT_START { /* No-op if have completed a pass */ \
449 if (! ALL_PARENS_COUNTED) RExC_total_parens = -1; \
451 #define IN_PARENS_PASS (RExC_total_parens < 0)
454 /* This is used to return failure (zero) early from the calling function if
455 * various flags in 'flags' are set. Two flags always cause a return:
456 * 'RESTART_PARSE' and 'NEED_UTF8'. 'extra' can be used to specify any
457 * additional flags that should cause a return; 0 if none. If the return will
458 * be done, '*flagp' is first set to be all of the flags that caused the
460 #define RETURN_FAIL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
462 if ((flags) & (RESTART_PARSE|NEED_UTF8|(extra))) { \
463 *(flagp) = (flags) & (RESTART_PARSE|NEED_UTF8|(extra)); \
468 #define MUST_RESTART(flags) ((flags) & (RESTART_PARSE))
470 #define RETURN_FAIL_ON_RESTART(flags,flagp) \
471 RETURN_FAIL_ON_RESTART_OR_FLAGS( flags, flagp, 0)
472 #define RETURN_FAIL_ON_RESTART_FLAGP(flagp) \
473 if (MUST_RESTART(*(flagp))) return 0
475 /* This converts the named class defined in regcomp.h to its equivalent class
476 * number defined in handy.h. */
477 #define namedclass_to_classnum(class) ((int) ((class) / 2))
478 #define classnum_to_namedclass(classnum) ((classnum) * 2)
480 #define _invlist_union_complement_2nd(a, b, output) \
481 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
482 #define _invlist_intersection_complement_2nd(a, b, output) \
483 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
485 /* We add a marker if we are deferring expansion of a property that is both
486 * 1) potentiallly user-defined; and
487 * 2) could also be an official Unicode property.
489 * Without this marker, any deferred expansion can only be for a user-defined
490 * one. This marker shouldn't conflict with any that could be in a legal name,
491 * and is appended to its name to indicate this. There is a string and
493 #define DEFERRED_COULD_BE_OFFICIAL_MARKERs "~"
494 #define DEFERRED_COULD_BE_OFFICIAL_MARKERc '~'
496 /* What is infinity for optimization purposes */
497 #define OPTIMIZE_INFTY SSize_t_MAX
499 /* About scan_data_t.
501 During optimisation we recurse through the regexp program performing
502 various inplace (keyhole style) optimisations. In addition study_chunk
503 and scan_commit populate this data structure with information about
504 what strings MUST appear in the pattern. We look for the longest
505 string that must appear at a fixed location, and we look for the
506 longest string that may appear at a floating location. So for instance
511 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
512 strings (because they follow a .* construct). study_chunk will identify
513 both FOO and BAR as being the longest fixed and floating strings respectively.
515 The strings can be composites, for instance
519 will result in a composite fixed substring 'foo'.
521 For each string some basic information is maintained:
524 This is the position the string must appear at, or not before.
525 It also implicitly (when combined with minlenp) tells us how many
526 characters must match before the string we are searching for.
527 Likewise when combined with minlenp and the length of the string it
528 tells us how many characters must appear after the string we have
532 Only used for floating strings. This is the rightmost point that
533 the string can appear at. If set to OPTIMIZE_INFTY it indicates that the
534 string can occur infinitely far to the right.
535 For fixed strings, it is equal to min_offset.
538 A pointer to the minimum number of characters of the pattern that the
539 string was found inside. This is important as in the case of positive
540 lookahead or positive lookbehind we can have multiple patterns
545 The minimum length of the pattern overall is 3, the minimum length
546 of the lookahead part is 3, but the minimum length of the part that
547 will actually match is 1. So 'FOO's minimum length is 3, but the
548 minimum length for the F is 1. This is important as the minimum length
549 is used to determine offsets in front of and behind the string being
550 looked for. Since strings can be composites this is the length of the
551 pattern at the time it was committed with a scan_commit. Note that
552 the length is calculated by study_chunk, so that the minimum lengths
553 are not known until the full pattern has been compiled, thus the
554 pointer to the value.
558 In the case of lookbehind the string being searched for can be
559 offset past the start point of the final matching string.
560 If this value was just blithely removed from the min_offset it would
561 invalidate some of the calculations for how many chars must match
562 before or after (as they are derived from min_offset and minlen and
563 the length of the string being searched for).
564 When the final pattern is compiled and the data is moved from the
565 scan_data_t structure into the regexp structure the information
566 about lookbehind is factored in, with the information that would
567 have been lost precalculated in the end_shift field for the
570 The fields pos_min and pos_delta are used to store the minimum offset
571 and the delta to the maximum offset at the current point in the pattern.
575 struct scan_data_substrs {
576 SV *str; /* longest substring found in pattern */
577 SSize_t min_offset; /* earliest point in string it can appear */
578 SSize_t max_offset; /* latest point in string it can appear */
579 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
580 SSize_t lookbehind; /* is the pos of the string modified by LB */
581 I32 flags; /* per substring SF_* and SCF_* flags */
584 typedef struct scan_data_t {
585 /*I32 len_min; unused */
586 /*I32 len_delta; unused */
590 SSize_t last_end; /* min value, <0 unless valid. */
591 SSize_t last_start_min;
592 SSize_t last_start_max;
593 U8 cur_is_floating; /* whether the last_* values should be set as
594 * the next fixed (0) or floating (1)
597 /* [0] is longest fixed substring so far, [1] is longest float so far */
598 struct scan_data_substrs substrs[2];
600 I32 flags; /* common SF_* and SCF_* flags */
602 SSize_t *last_closep;
603 regnode_ssc *start_class;
607 * Forward declarations for pregcomp()'s friends.
610 static const scan_data_t zero_scan_data = {
611 0, 0, NULL, 0, 0, 0, 0,
613 { NULL, 0, 0, 0, 0, 0 },
614 { NULL, 0, 0, 0, 0, 0 },
621 #define SF_BEFORE_SEOL 0x0001
622 #define SF_BEFORE_MEOL 0x0002
623 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
625 #define SF_IS_INF 0x0040
626 #define SF_HAS_PAR 0x0080
627 #define SF_IN_PAR 0x0100
628 #define SF_HAS_EVAL 0x0200
631 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
632 * longest substring in the pattern. When it is not set the optimiser keeps
633 * track of position, but does not keep track of the actual strings seen,
635 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
638 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
639 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
640 * turned off because of the alternation (BRANCH). */
641 #define SCF_DO_SUBSTR 0x0400
643 #define SCF_DO_STCLASS_AND 0x0800
644 #define SCF_DO_STCLASS_OR 0x1000
645 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
646 #define SCF_WHILEM_VISITED_POS 0x2000
648 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
649 #define SCF_SEEN_ACCEPT 0x8000
650 #define SCF_TRIE_DOING_RESTUDY 0x10000
651 #define SCF_IN_DEFINE 0x20000
656 #define UTF cBOOL(RExC_utf8)
658 /* The enums for all these are ordered so things work out correctly */
659 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
660 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
661 == REGEX_DEPENDS_CHARSET)
662 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
663 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
664 >= REGEX_UNICODE_CHARSET)
665 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
666 == REGEX_ASCII_RESTRICTED_CHARSET)
667 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
668 >= REGEX_ASCII_RESTRICTED_CHARSET)
669 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
670 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
672 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
674 /* For programs that want to be strictly Unicode compatible by dying if any
675 * attempt is made to match a non-Unicode code point against a Unicode
677 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
679 #define OOB_NAMEDCLASS -1
681 /* There is no code point that is out-of-bounds, so this is problematic. But
682 * its only current use is to initialize a variable that is always set before
684 #define OOB_UNICODE 0xDEADBEEF
686 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
689 /* length of regex to show in messages that don't mark a position within */
690 #define RegexLengthToShowInErrorMessages 127
693 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
694 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
695 * op/pragma/warn/regcomp.
697 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
698 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
700 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
701 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
703 /* The code in this file in places uses one level of recursion with parsing
704 * rebased to an alternate string constructed by us in memory. This can take
705 * the form of something that is completely different from the input, or
706 * something that uses the input as part of the alternate. In the first case,
707 * there should be no possibility of an error, as we are in complete control of
708 * the alternate string. But in the second case we don't completely control
709 * the input portion, so there may be errors in that. Here's an example:
711 * is handled specially because \x{df} folds to a sequence of more than one
712 * character: 'ss'. What is done is to create and parse an alternate string,
713 * which looks like this:
714 * /(?:\x{DF}|[abc\x{DF}def])/ui
715 * where it uses the input unchanged in the middle of something it constructs,
716 * which is a branch for the DF outside the character class, and clustering
717 * parens around the whole thing. (It knows enough to skip the DF inside the
718 * class while in this substitute parse.) 'abc' and 'def' may have errors that
719 * need to be reported. The general situation looks like this:
721 * |<------- identical ------>|
723 * Input: ---------------------------------------------------------------
724 * Constructed: ---------------------------------------------------
726 * |<------- identical ------>|
728 * sI..eI is the portion of the input pattern we are concerned with here.
729 * sC..EC is the constructed substitute parse string.
730 * sC..tC is constructed by us
731 * tC..eC is an exact duplicate of the portion of the input pattern tI..eI.
732 * In the diagram, these are vertically aligned.
733 * eC..EC is also constructed by us.
734 * xC is the position in the substitute parse string where we found a
736 * xI is the position in the original pattern corresponding to xC.
738 * We want to display a message showing the real input string. Thus we need to
739 * translate from xC to xI. We know that xC >= tC, since the portion of the
740 * string sC..tC has been constructed by us, and so shouldn't have errors. We
742 * xI = tI + (xC - tC)
744 * When the substitute parse is constructed, the code needs to set:
747 * RExC_copy_start_in_input (tI)
748 * RExC_copy_start_in_constructed (tC)
749 * and restore them when done.
751 * During normal processing of the input pattern, both
752 * 'RExC_copy_start_in_input' and 'RExC_copy_start_in_constructed' are set to
753 * sI, so that xC equals xI.
756 #define sI RExC_precomp
757 #define eI RExC_precomp_end
758 #define sC RExC_start
760 #define tI RExC_copy_start_in_input
761 #define tC RExC_copy_start_in_constructed
762 #define xI(xC) (tI + (xC - tC))
763 #define xI_offset(xC) (xI(xC) - sI)
765 #define REPORT_LOCATION_ARGS(xC) \
767 (xI(xC) > eI) /* Don't run off end */ \
768 ? eI - sI /* Length before the <--HERE */ \
769 : ((xI_offset(xC) >= 0) \
771 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
772 IVdf " trying to output message for " \
774 __FILE__, __LINE__, (IV) xI_offset(xC), \
775 ((int) (eC - sC)), sC), 0)), \
776 sI), /* The input pattern printed up to the <--HERE */ \
778 (xI(xC) > eI) ? 0 : eI - xI(xC), /* Length after <--HERE */ \
779 (xI(xC) > eI) ? eI : xI(xC)) /* pattern after <--HERE */
781 /* Used to point after bad bytes for an error message, but avoid skipping
782 * past a nul byte. */
783 #define SKIP_IF_CHAR(s, e) (!*(s) ? 0 : UTF ? UTF8_SAFE_SKIP(s, e) : 1)
785 /* Set up to clean up after our imminent demise */
786 #define PREPARE_TO_DIE \
789 SAVEFREESV(RExC_rx_sv); \
790 if (RExC_open_parens) \
791 SAVEFREEPV(RExC_open_parens); \
792 if (RExC_close_parens) \
793 SAVEFREEPV(RExC_close_parens); \
797 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
798 * arg. Show regex, up to a maximum length. If it's too long, chop and add
801 #define _FAIL(code) STMT_START { \
802 const char *ellipses = ""; \
803 IV len = RExC_precomp_end - RExC_precomp; \
806 if (len > RegexLengthToShowInErrorMessages) { \
807 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
808 len = RegexLengthToShowInErrorMessages - 10; \
814 #define FAIL(msg) _FAIL( \
815 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
816 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
818 #define FAIL2(msg,arg) _FAIL( \
819 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
820 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
822 #define FAIL3(msg,arg1,arg2) _FAIL( \
823 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
824 arg1, arg2, UTF8fARG(UTF, len, RExC_precomp), ellipses))
827 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
829 #define Simple_vFAIL(m) STMT_START { \
830 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
831 m, REPORT_LOCATION_ARGS(RExC_parse)); \
835 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
837 #define vFAIL(m) STMT_START { \
843 * Like Simple_vFAIL(), but accepts two arguments.
845 #define Simple_vFAIL2(m,a1) STMT_START { \
846 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, \
847 REPORT_LOCATION_ARGS(RExC_parse)); \
851 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
853 #define vFAIL2(m,a1) STMT_START { \
855 Simple_vFAIL2(m, a1); \
860 * Like Simple_vFAIL(), but accepts three arguments.
862 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
863 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, \
864 REPORT_LOCATION_ARGS(RExC_parse)); \
868 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
870 #define vFAIL3(m,a1,a2) STMT_START { \
872 Simple_vFAIL3(m, a1, a2); \
876 * Like Simple_vFAIL(), but accepts four arguments.
878 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
879 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, a3, \
880 REPORT_LOCATION_ARGS(RExC_parse)); \
883 #define vFAIL4(m,a1,a2,a3) STMT_START { \
885 Simple_vFAIL4(m, a1, a2, a3); \
888 /* A specialized version of vFAIL2 that works with UTF8f */
889 #define vFAIL2utf8f(m, a1) STMT_START { \
891 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, \
892 REPORT_LOCATION_ARGS(RExC_parse)); \
895 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
897 S_re_croak(aTHX_ UTF, m REPORT_LOCATION, a1, a2, \
898 REPORT_LOCATION_ARGS(RExC_parse)); \
901 /* Setting this to NULL is a signal to not output warnings */
902 #define TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE \
904 RExC_save_copy_start_in_constructed = RExC_copy_start_in_constructed;\
905 RExC_copy_start_in_constructed = NULL; \
907 #define RESTORE_WARNINGS \
908 RExC_copy_start_in_constructed = RExC_save_copy_start_in_constructed
910 /* Since a warning can be generated multiple times as the input is reparsed, we
911 * output it the first time we come to that point in the parse, but suppress it
912 * otherwise. 'RExC_copy_start_in_constructed' being NULL is a flag to not
913 * generate any warnings */
914 #define TO_OUTPUT_WARNINGS(loc) \
915 ( RExC_copy_start_in_constructed \
916 && ((xI(loc)) - RExC_precomp) > (Ptrdiff_t) RExC_latest_warn_offset)
918 /* After we've emitted a warning, we save the position in the input so we don't
920 #define UPDATE_WARNINGS_LOC(loc) \
922 if (TO_OUTPUT_WARNINGS(loc)) { \
923 RExC_latest_warn_offset = MAX(sI, MIN(eI, xI(loc))) \
928 /* 'warns' is the output of the packWARNx macro used in 'code' */
929 #define _WARN_HELPER(loc, warns, code) \
931 if (! RExC_copy_start_in_constructed) { \
932 Perl_croak( aTHX_ "panic! %s: %d: Tried to warn when none" \
933 " expected at '%s'", \
934 __FILE__, __LINE__, loc); \
936 if (TO_OUTPUT_WARNINGS(loc)) { \
940 UPDATE_WARNINGS_LOC(loc); \
944 /* m is not necessarily a "literal string", in this macro */
945 #define warn_non_literal_string(loc, packed_warn, m) \
946 _WARN_HELPER(loc, packed_warn, \
947 Perl_warner(aTHX_ packed_warn, \
948 "%s" REPORT_LOCATION, \
949 m, REPORT_LOCATION_ARGS(loc)))
950 #define reg_warn_non_literal_string(loc, m) \
951 warn_non_literal_string(loc, packWARN(WARN_REGEXP), m)
953 #define ckWARN2_non_literal_string(loc, packwarn, m, a1) \
956 Size_t format_size = strlen(m) + strlen(REPORT_LOCATION)+ 1;\
957 Newx(format, format_size, char); \
958 my_strlcpy(format, m, format_size); \
959 my_strlcat(format, REPORT_LOCATION, format_size); \
960 SAVEFREEPV(format); \
961 _WARN_HELPER(loc, packwarn, \
962 Perl_ck_warner(aTHX_ packwarn, \
964 a1, REPORT_LOCATION_ARGS(loc))); \
967 #define ckWARNreg(loc,m) \
968 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
969 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
971 REPORT_LOCATION_ARGS(loc)))
973 #define vWARN(loc, m) \
974 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
975 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
977 REPORT_LOCATION_ARGS(loc))) \
979 #define vWARN_dep(loc, m) \
980 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
981 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
983 REPORT_LOCATION_ARGS(loc)))
985 #define ckWARNdep(loc,m) \
986 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
987 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
989 REPORT_LOCATION_ARGS(loc)))
991 #define ckWARNregdep(loc,m) \
992 _WARN_HELPER(loc, packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
993 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
996 REPORT_LOCATION_ARGS(loc)))
998 #define ckWARN2reg_d(loc,m, a1) \
999 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1000 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
1001 m REPORT_LOCATION, \
1002 a1, REPORT_LOCATION_ARGS(loc)))
1004 #define ckWARN2reg(loc, m, a1) \
1005 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1006 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1007 m REPORT_LOCATION, \
1008 a1, REPORT_LOCATION_ARGS(loc)))
1010 #define vWARN3(loc, m, a1, a2) \
1011 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1012 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1013 m REPORT_LOCATION, \
1014 a1, a2, REPORT_LOCATION_ARGS(loc)))
1016 #define ckWARN3reg(loc, m, a1, a2) \
1017 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1018 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1019 m REPORT_LOCATION, \
1021 REPORT_LOCATION_ARGS(loc)))
1023 #define vWARN4(loc, m, a1, a2, a3) \
1024 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1025 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1026 m REPORT_LOCATION, \
1028 REPORT_LOCATION_ARGS(loc)))
1030 #define ckWARN4reg(loc, m, a1, a2, a3) \
1031 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1032 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
1033 m REPORT_LOCATION, \
1035 REPORT_LOCATION_ARGS(loc)))
1037 #define vWARN5(loc, m, a1, a2, a3, a4) \
1038 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
1039 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
1040 m REPORT_LOCATION, \
1042 REPORT_LOCATION_ARGS(loc)))
1044 #define ckWARNexperimental(loc, class, m) \
1046 if (! RExC_warned_ ## class) { /* warn once per compilation */ \
1047 RExC_warned_ ## class = 1; \
1048 _WARN_HELPER(loc, packWARN(class), \
1049 Perl_ck_warner_d(aTHX_ packWARN(class), \
1050 m REPORT_LOCATION, \
1051 REPORT_LOCATION_ARGS(loc)));\
1055 /* Convert between a pointer to a node and its offset from the beginning of the
1057 #define REGNODE_p(offset) (RExC_emit_start + (offset))
1058 #define REGNODE_OFFSET(node) ((node) - RExC_emit_start)
1060 /* Macros for recording node offsets. 20001227 mjd@plover.com
1061 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
1062 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
1063 * Element 0 holds the number n.
1064 * Position is 1 indexed.
1066 #ifndef RE_TRACK_PATTERN_OFFSETS
1067 #define Set_Node_Offset_To_R(offset,byte)
1068 #define Set_Node_Offset(node,byte)
1069 #define Set_Cur_Node_Offset
1070 #define Set_Node_Length_To_R(node,len)
1071 #define Set_Node_Length(node,len)
1072 #define Set_Node_Cur_Length(node,start)
1073 #define Node_Offset(n)
1074 #define Node_Length(n)
1075 #define Set_Node_Offset_Length(node,offset,len)
1076 #define ProgLen(ri) ri->u.proglen
1077 #define SetProgLen(ri,x) ri->u.proglen = x
1078 #define Track_Code(code)
1080 #define ProgLen(ri) ri->u.offsets[0]
1081 #define SetProgLen(ri,x) ri->u.offsets[0] = x
1082 #define Set_Node_Offset_To_R(offset,byte) STMT_START { \
1083 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
1084 __LINE__, (int)(offset), (int)(byte))); \
1085 if((offset) < 0) { \
1086 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
1089 RExC_offsets[2*(offset)-1] = (byte); \
1093 #define Set_Node_Offset(node,byte) \
1094 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (byte)-RExC_start)
1095 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
1097 #define Set_Node_Length_To_R(node,len) STMT_START { \
1098 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
1099 __LINE__, (int)(node), (int)(len))); \
1101 Perl_croak(aTHX_ "value of node is %d in Length macro", \
1104 RExC_offsets[2*(node)] = (len); \
1108 #define Set_Node_Length(node,len) \
1109 Set_Node_Length_To_R(REGNODE_OFFSET(node), len)
1110 #define Set_Node_Cur_Length(node, start) \
1111 Set_Node_Length(node, RExC_parse - start)
1113 /* Get offsets and lengths */
1114 #define Node_Offset(n) (RExC_offsets[2*(REGNODE_OFFSET(n))-1])
1115 #define Node_Length(n) (RExC_offsets[2*(REGNODE_OFFSET(n))])
1117 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
1118 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (offset)); \
1119 Set_Node_Length_To_R(REGNODE_OFFSET(node), (len)); \
1122 #define Track_Code(code) STMT_START { code } STMT_END
1125 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
1126 #define EXPERIMENTAL_INPLACESCAN
1127 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
1131 Perl_re_printf(pTHX_ const char *fmt, ...)
1135 PerlIO *f= Perl_debug_log;
1136 PERL_ARGS_ASSERT_RE_PRINTF;
1138 result = PerlIO_vprintf(f, fmt, ap);
1144 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
1148 PerlIO *f= Perl_debug_log;
1149 PERL_ARGS_ASSERT_RE_INDENTF;
1150 va_start(ap, depth);
1151 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
1152 result = PerlIO_vprintf(f, fmt, ap);
1156 #endif /* DEBUGGING */
1158 #define DEBUG_RExC_seen() \
1159 DEBUG_OPTIMISE_MORE_r({ \
1160 Perl_re_printf( aTHX_ "RExC_seen: "); \
1162 if (RExC_seen & REG_ZERO_LEN_SEEN) \
1163 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
1165 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
1166 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
1168 if (RExC_seen & REG_GPOS_SEEN) \
1169 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
1171 if (RExC_seen & REG_RECURSE_SEEN) \
1172 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
1174 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
1175 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
1177 if (RExC_seen & REG_VERBARG_SEEN) \
1178 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
1180 if (RExC_seen & REG_CUTGROUP_SEEN) \
1181 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
1183 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
1184 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
1186 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
1187 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
1189 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
1190 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
1192 Perl_re_printf( aTHX_ "\n"); \
1195 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
1196 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
1201 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1202 const char *close_str)
1207 Perl_re_printf( aTHX_ "%s", open_str);
1208 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1209 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
1210 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1211 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1212 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1213 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1214 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1215 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1216 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1217 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1218 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1219 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1220 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1221 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1222 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1223 Perl_re_printf( aTHX_ "%s", close_str);
1228 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1229 U32 depth, int is_inf)
1231 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1233 DEBUG_OPTIMISE_MORE_r({
1236 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1240 (IV)data->pos_delta,
1244 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1246 Perl_re_printf( aTHX_
1247 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1249 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1250 is_inf ? "INF " : ""
1253 if (data->last_found) {
1255 Perl_re_printf(aTHX_
1256 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1257 SvPVX_const(data->last_found),
1259 (IV)data->last_start_min,
1260 (IV)data->last_start_max
1263 for (i = 0; i < 2; i++) {
1264 Perl_re_printf(aTHX_
1265 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1266 data->cur_is_floating == i ? "*" : "",
1267 i ? "Float" : "Fixed",
1268 SvPVX_const(data->substrs[i].str),
1269 (IV)data->substrs[i].min_offset,
1270 (IV)data->substrs[i].max_offset
1272 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1276 Perl_re_printf( aTHX_ "\n");
1282 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1283 regnode *scan, U32 depth, U32 flags)
1285 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1292 Next = regnext(scan);
1293 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1294 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1297 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1298 Next ? (REG_NODE_NUM(Next)) : 0 );
1299 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1300 Perl_re_printf( aTHX_ "\n");
1305 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1306 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1308 # define DEBUG_PEEP(str, scan, depth, flags) \
1309 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1312 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1313 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1317 /* =========================================================
1318 * BEGIN edit_distance stuff.
1320 * This calculates how many single character changes of any type are needed to
1321 * transform a string into another one. It is taken from version 3.1 of
1323 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1326 /* Our unsorted dictionary linked list. */
1327 /* Note we use UVs, not chars. */
1332 struct dictionary* next;
1334 typedef struct dictionary item;
1337 PERL_STATIC_INLINE item*
1338 push(UV key, item* curr)
1341 Newx(head, 1, item);
1349 PERL_STATIC_INLINE item*
1350 find(item* head, UV key)
1352 item* iterator = head;
1354 if (iterator->key == key){
1357 iterator = iterator->next;
1363 PERL_STATIC_INLINE item*
1364 uniquePush(item* head, UV key)
1366 item* iterator = head;
1369 if (iterator->key == key) {
1372 iterator = iterator->next;
1375 return push(key, head);
1378 PERL_STATIC_INLINE void
1379 dict_free(item* head)
1381 item* iterator = head;
1384 item* temp = iterator;
1385 iterator = iterator->next;
1392 /* End of Dictionary Stuff */
1394 /* All calculations/work are done here */
1396 S_edit_distance(const UV* src,
1398 const STRLEN x, /* length of src[] */
1399 const STRLEN y, /* length of tgt[] */
1400 const SSize_t maxDistance
1404 UV swapCount, swapScore, targetCharCount, i, j;
1406 UV score_ceil = x + y;
1408 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1410 /* intialize matrix start values */
1411 Newx(scores, ( (x + 2) * (y + 2)), UV);
1412 scores[0] = score_ceil;
1413 scores[1 * (y + 2) + 0] = score_ceil;
1414 scores[0 * (y + 2) + 1] = score_ceil;
1415 scores[1 * (y + 2) + 1] = 0;
1416 head = uniquePush(uniquePush(head, src[0]), tgt[0]);
1421 for (i=1;i<=x;i++) {
1423 head = uniquePush(head, src[i]);
1424 scores[(i+1) * (y + 2) + 1] = i;
1425 scores[(i+1) * (y + 2) + 0] = score_ceil;
1428 for (j=1;j<=y;j++) {
1431 head = uniquePush(head, tgt[j]);
1432 scores[1 * (y + 2) + (j + 1)] = j;
1433 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1436 targetCharCount = find(head, tgt[j-1])->value;
1437 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1439 if (src[i-1] != tgt[j-1]){
1440 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));
1444 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1448 find(head, src[i-1])->value = i;
1452 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1455 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1459 /* END of edit_distance() stuff
1460 * ========================================================= */
1462 /* Mark that we cannot extend a found fixed substring at this point.
1463 Update the longest found anchored substring or the longest found
1464 floating substrings if needed. */
1467 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1468 SSize_t *minlenp, int is_inf)
1470 const STRLEN l = CHR_SVLEN(data->last_found);
1471 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1472 const STRLEN old_l = CHR_SVLEN(longest_sv);
1473 DECLARE_AND_GET_RE_DEBUG_FLAGS;
1475 PERL_ARGS_ASSERT_SCAN_COMMIT;
1477 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1478 const U8 i = data->cur_is_floating;
1479 SvSetMagicSV(longest_sv, data->last_found);
1480 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1483 data->substrs[0].max_offset = data->substrs[0].min_offset;
1485 data->substrs[1].max_offset =
1489 ? data->last_start_max
1490 /* temporary underflow guard for 5.32 */
1491 : data->pos_delta < 0 ? OPTIMIZE_INFTY
1492 : (data->pos_delta > OPTIMIZE_INFTY - data->pos_min
1494 : data->pos_min + data->pos_delta));
1497 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1498 data->substrs[i].flags |= data->flags & SF_BEFORE_EOL;
1499 data->substrs[i].minlenp = minlenp;
1500 data->substrs[i].lookbehind = 0;
1503 SvCUR_set(data->last_found, 0);
1505 SV * const sv = data->last_found;
1506 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1507 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1512 data->last_end = -1;
1513 data->flags &= ~SF_BEFORE_EOL;
1514 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1517 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1518 * list that describes which code points it matches */
1521 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1523 /* Set the SSC 'ssc' to match an empty string or any code point */
1525 PERL_ARGS_ASSERT_SSC_ANYTHING;
1527 assert(is_ANYOF_SYNTHETIC(ssc));
1529 /* mortalize so won't leak */
1530 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1531 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1535 S_ssc_is_anything(const regnode_ssc *ssc)
1537 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1538 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1539 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1540 * in any way, so there's no point in using it */
1545 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1547 assert(is_ANYOF_SYNTHETIC(ssc));
1549 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1553 /* See if the list consists solely of the range 0 - Infinity */
1554 invlist_iterinit(ssc->invlist);
1555 ret = invlist_iternext(ssc->invlist, &start, &end)
1559 invlist_iterfinish(ssc->invlist);
1565 /* If e.g., both \w and \W are set, matches everything */
1566 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1568 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1569 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1579 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1581 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1582 * string, any code point, or any posix class under locale */
1584 PERL_ARGS_ASSERT_SSC_INIT;
1586 Zero(ssc, 1, regnode_ssc);
1587 set_ANYOF_SYNTHETIC(ssc);
1588 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1591 /* If any portion of the regex is to operate under locale rules that aren't
1592 * fully known at compile time, initialization includes it. The reason
1593 * this isn't done for all regexes is that the optimizer was written under
1594 * the assumption that locale was all-or-nothing. Given the complexity and
1595 * lack of documentation in the optimizer, and that there are inadequate
1596 * test cases for locale, many parts of it may not work properly, it is
1597 * safest to avoid locale unless necessary. */
1598 if (RExC_contains_locale) {
1599 ANYOF_POSIXL_SETALL(ssc);
1602 ANYOF_POSIXL_ZERO(ssc);
1607 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1608 const regnode_ssc *ssc)
1610 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1611 * to the list of code points matched, and locale posix classes; hence does
1612 * not check its flags) */
1617 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1619 assert(is_ANYOF_SYNTHETIC(ssc));
1621 invlist_iterinit(ssc->invlist);
1622 ret = invlist_iternext(ssc->invlist, &start, &end)
1626 invlist_iterfinish(ssc->invlist);
1632 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1639 #define INVLIST_INDEX 0
1640 #define ONLY_LOCALE_MATCHES_INDEX 1
1641 #define DEFERRED_USER_DEFINED_INDEX 2
1644 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1645 const regnode_charclass* const node)
1647 /* Returns a mortal inversion list defining which code points are matched
1648 * by 'node', which is of type ANYOF. Handles complementing the result if
1649 * appropriate. If some code points aren't knowable at this time, the
1650 * returned list must, and will, contain every code point that is a
1654 SV* only_utf8_locale_invlist = NULL;
1656 const U32 n = ARG(node);
1657 bool new_node_has_latin1 = FALSE;
1658 const U8 flags = (inRANGE(OP(node), ANYOFH, ANYOFRb))
1660 : ANYOF_FLAGS(node);
1662 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1664 /* Look at the data structure created by S_set_ANYOF_arg() */
1665 if (n != ANYOF_ONLY_HAS_BITMAP) {
1666 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1667 AV * const av = MUTABLE_AV(SvRV(rv));
1668 SV **const ary = AvARRAY(av);
1669 assert(RExC_rxi->data->what[n] == 's');
1671 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
1673 /* Here there are things that won't be known until runtime -- we
1674 * have to assume it could be anything */
1675 invlist = sv_2mortal(_new_invlist(1));
1676 return _add_range_to_invlist(invlist, 0, UV_MAX);
1678 else if (ary[INVLIST_INDEX]) {
1680 /* Use the node's inversion list */
1681 invlist = sv_2mortal(invlist_clone(ary[INVLIST_INDEX], NULL));
1684 /* Get the code points valid only under UTF-8 locales */
1685 if ( (flags & ANYOFL_FOLD)
1686 && av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX)
1688 only_utf8_locale_invlist = ary[ONLY_LOCALE_MATCHES_INDEX];
1693 invlist = sv_2mortal(_new_invlist(0));
1696 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1697 * code points, and an inversion list for the others, but if there are code
1698 * points that should match only conditionally on the target string being
1699 * UTF-8, those are placed in the inversion list, and not the bitmap.
1700 * Since there are circumstances under which they could match, they are
1701 * included in the SSC. But if the ANYOF node is to be inverted, we have
1702 * to exclude them here, so that when we invert below, the end result
1703 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1704 * have to do this here before we add the unconditionally matched code
1706 if (flags & ANYOF_INVERT) {
1707 _invlist_intersection_complement_2nd(invlist,
1712 /* Add in the points from the bit map */
1713 if (! inRANGE(OP(node), ANYOFH, ANYOFRb)) {
1714 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1715 if (ANYOF_BITMAP_TEST(node, i)) {
1716 unsigned int start = i++;
1718 for (; i < NUM_ANYOF_CODE_POINTS
1719 && ANYOF_BITMAP_TEST(node, i); ++i)
1723 invlist = _add_range_to_invlist(invlist, start, i-1);
1724 new_node_has_latin1 = TRUE;
1729 /* If this can match all upper Latin1 code points, have to add them
1730 * as well. But don't add them if inverting, as when that gets done below,
1731 * it would exclude all these characters, including the ones it shouldn't
1732 * that were added just above */
1733 if (! (flags & ANYOF_INVERT) && OP(node) == ANYOFD
1734 && (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1736 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1739 /* Similarly for these */
1740 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1741 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1744 if (flags & ANYOF_INVERT) {
1745 _invlist_invert(invlist);
1747 else if (flags & ANYOFL_FOLD) {
1748 if (new_node_has_latin1) {
1750 /* Under /li, any 0-255 could fold to any other 0-255, depending on
1751 * the locale. We can skip this if there are no 0-255 at all. */
1752 _invlist_union(invlist, PL_Latin1, &invlist);
1754 invlist = add_cp_to_invlist(invlist, LATIN_SMALL_LETTER_DOTLESS_I);
1755 invlist = add_cp_to_invlist(invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
1758 if (_invlist_contains_cp(invlist, LATIN_SMALL_LETTER_DOTLESS_I)) {
1759 invlist = add_cp_to_invlist(invlist, 'I');
1761 if (_invlist_contains_cp(invlist,
1762 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE))
1764 invlist = add_cp_to_invlist(invlist, 'i');
1769 /* Similarly add the UTF-8 locale possible matches. These have to be
1770 * deferred until after the non-UTF-8 locale ones are taken care of just
1771 * above, or it leads to wrong results under ANYOF_INVERT */
1772 if (only_utf8_locale_invlist) {
1773 _invlist_union_maybe_complement_2nd(invlist,
1774 only_utf8_locale_invlist,
1775 flags & ANYOF_INVERT,
1782 /* These two functions currently do the exact same thing */
1783 #define ssc_init_zero ssc_init
1785 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1786 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1788 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1789 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1790 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1793 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1794 const regnode_charclass *and_with)
1796 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1797 * another SSC or a regular ANYOF class. Can create false positives. */
1800 U8 and_with_flags = inRANGE(OP(and_with), ANYOFH, ANYOFRb)
1802 : ANYOF_FLAGS(and_with);
1805 PERL_ARGS_ASSERT_SSC_AND;
1807 assert(is_ANYOF_SYNTHETIC(ssc));
1809 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1810 * the code point inversion list and just the relevant flags */
1811 if (is_ANYOF_SYNTHETIC(and_with)) {
1812 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1813 anded_flags = and_with_flags;
1815 /* XXX This is a kludge around what appears to be deficiencies in the
1816 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1817 * there are paths through the optimizer where it doesn't get weeded
1818 * out when it should. And if we don't make some extra provision for
1819 * it like the code just below, it doesn't get added when it should.
1820 * This solution is to add it only when AND'ing, which is here, and
1821 * only when what is being AND'ed is the pristine, original node
1822 * matching anything. Thus it is like adding it to ssc_anything() but
1823 * only when the result is to be AND'ed. Probably the same solution
1824 * could be adopted for the same problem we have with /l matching,
1825 * which is solved differently in S_ssc_init(), and that would lead to
1826 * fewer false positives than that solution has. But if this solution
1827 * creates bugs, the consequences are only that a warning isn't raised
1828 * that should be; while the consequences for having /l bugs is
1829 * incorrect matches */
1830 if (ssc_is_anything((regnode_ssc *)and_with)) {
1831 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1835 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1836 if (OP(and_with) == ANYOFD) {
1837 anded_flags = and_with_flags & ANYOF_COMMON_FLAGS;
1840 anded_flags = and_with_flags
1841 &( ANYOF_COMMON_FLAGS
1842 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1843 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1844 if (ANYOFL_UTF8_LOCALE_REQD(and_with_flags)) {
1846 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1851 ANYOF_FLAGS(ssc) &= anded_flags;
1853 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1854 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1855 * 'and_with' may be inverted. When not inverted, we have the situation of
1857 * (C1 | P1) & (C2 | P2)
1858 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1859 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1860 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1861 * <= ((C1 & C2) | P1 | P2)
1862 * Alternatively, the last few steps could be:
1863 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1864 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1865 * <= (C1 | C2 | (P1 & P2))
1866 * We favor the second approach if either P1 or P2 is non-empty. This is
1867 * because these components are a barrier to doing optimizations, as what
1868 * they match cannot be known until the moment of matching as they are
1869 * dependent on the current locale, 'AND"ing them likely will reduce or
1871 * But we can do better if we know that C1,P1 are in their initial state (a
1872 * frequent occurrence), each matching everything:
1873 * (<everything>) & (C2 | P2) = C2 | P2
1874 * Similarly, if C2,P2 are in their initial state (again a frequent
1875 * occurrence), the result is a no-op
1876 * (C1 | P1) & (<everything>) = C1 | P1
1879 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1880 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1881 * <= (C1 & ~C2) | (P1 & ~P2)
1884 if ((and_with_flags & ANYOF_INVERT)
1885 && ! is_ANYOF_SYNTHETIC(and_with))
1889 ssc_intersection(ssc,
1891 FALSE /* Has already been inverted */
1894 /* If either P1 or P2 is empty, the intersection will be also; can skip
1896 if (! (and_with_flags & ANYOF_MATCHES_POSIXL)) {
1897 ANYOF_POSIXL_ZERO(ssc);
1899 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1901 /* Note that the Posix class component P from 'and_with' actually
1903 * P = Pa | Pb | ... | Pn
1904 * where each component is one posix class, such as in [\w\s].
1906 * ~P = ~(Pa | Pb | ... | Pn)
1907 * = ~Pa & ~Pb & ... & ~Pn
1908 * <= ~Pa | ~Pb | ... | ~Pn
1909 * The last is something we can easily calculate, but unfortunately
1910 * is likely to have many false positives. We could do better
1911 * in some (but certainly not all) instances if two classes in
1912 * P have known relationships. For example
1913 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1915 * :lower: & :print: = :lower:
1916 * And similarly for classes that must be disjoint. For example,
1917 * since \s and \w can have no elements in common based on rules in
1918 * the POSIX standard,
1919 * \w & ^\S = nothing
1920 * Unfortunately, some vendor locales do not meet the Posix
1921 * standard, in particular almost everything by Microsoft.
1922 * The loop below just changes e.g., \w into \W and vice versa */
1924 regnode_charclass_posixl temp;
1925 int add = 1; /* To calculate the index of the complement */
1927 Zero(&temp, 1, regnode_charclass_posixl);
1928 ANYOF_POSIXL_ZERO(&temp);
1929 for (i = 0; i < ANYOF_MAX; i++) {
1931 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1932 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1934 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1935 ANYOF_POSIXL_SET(&temp, i + add);
1937 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1939 ANYOF_POSIXL_AND(&temp, ssc);
1941 } /* else ssc already has no posixes */
1942 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1943 in its initial state */
1944 else if (! is_ANYOF_SYNTHETIC(and_with)
1945 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1947 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1948 * copy it over 'ssc' */
1949 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1950 if (is_ANYOF_SYNTHETIC(and_with)) {
1951 StructCopy(and_with, ssc, regnode_ssc);
1954 ssc->invlist = anded_cp_list;
1955 ANYOF_POSIXL_ZERO(ssc);
1956 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1957 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1961 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1962 || (and_with_flags & ANYOF_MATCHES_POSIXL))
1964 /* One or the other of P1, P2 is non-empty. */
1965 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1966 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1968 ssc_union(ssc, anded_cp_list, FALSE);
1970 else { /* P1 = P2 = empty */
1971 ssc_intersection(ssc, anded_cp_list, FALSE);
1977 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1978 const regnode_charclass *or_with)
1980 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1981 * another SSC or a regular ANYOF class. Can create false positives if
1982 * 'or_with' is to be inverted. */
1986 U8 or_with_flags = inRANGE(OP(or_with), ANYOFH, ANYOFRb)
1988 : ANYOF_FLAGS(or_with);
1990 PERL_ARGS_ASSERT_SSC_OR;
1992 assert(is_ANYOF_SYNTHETIC(ssc));
1994 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1995 * the code point inversion list and just the relevant flags */
1996 if (is_ANYOF_SYNTHETIC(or_with)) {
1997 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1998 ored_flags = or_with_flags;
2001 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
2002 ored_flags = or_with_flags & ANYOF_COMMON_FLAGS;
2003 if (OP(or_with) != ANYOFD) {
2006 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2007 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
2008 if (ANYOFL_UTF8_LOCALE_REQD(or_with_flags)) {
2010 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
2015 ANYOF_FLAGS(ssc) |= ored_flags;
2017 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
2018 * C2 is the list of code points in 'or-with'; P2, its posix classes.
2019 * 'or_with' may be inverted. When not inverted, we have the simple
2020 * situation of computing:
2021 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
2022 * If P1|P2 yields a situation with both a class and its complement are
2023 * set, like having both \w and \W, this matches all code points, and we
2024 * can delete these from the P component of the ssc going forward. XXX We
2025 * might be able to delete all the P components, but I (khw) am not certain
2026 * about this, and it is better to be safe.
2029 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
2030 * <= (C1 | P1) | ~C2
2031 * <= (C1 | ~C2) | P1
2032 * (which results in actually simpler code than the non-inverted case)
2035 if ((or_with_flags & ANYOF_INVERT)
2036 && ! is_ANYOF_SYNTHETIC(or_with))
2038 /* We ignore P2, leaving P1 going forward */
2039 } /* else Not inverted */
2040 else if (or_with_flags & ANYOF_MATCHES_POSIXL) {
2041 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
2042 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2044 for (i = 0; i < ANYOF_MAX; i += 2) {
2045 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
2047 ssc_match_all_cp(ssc);
2048 ANYOF_POSIXL_CLEAR(ssc, i);
2049 ANYOF_POSIXL_CLEAR(ssc, i+1);
2057 FALSE /* Already has been inverted */
2062 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
2064 PERL_ARGS_ASSERT_SSC_UNION;
2066 assert(is_ANYOF_SYNTHETIC(ssc));
2068 _invlist_union_maybe_complement_2nd(ssc->invlist,
2075 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
2077 const bool invert2nd)
2079 PERL_ARGS_ASSERT_SSC_INTERSECTION;
2081 assert(is_ANYOF_SYNTHETIC(ssc));
2083 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
2090 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
2092 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
2094 assert(is_ANYOF_SYNTHETIC(ssc));
2096 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
2100 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
2102 /* AND just the single code point 'cp' into the SSC 'ssc' */
2104 SV* cp_list = _new_invlist(2);
2106 PERL_ARGS_ASSERT_SSC_CP_AND;
2108 assert(is_ANYOF_SYNTHETIC(ssc));
2110 cp_list = add_cp_to_invlist(cp_list, cp);
2111 ssc_intersection(ssc, cp_list,
2112 FALSE /* Not inverted */
2114 SvREFCNT_dec_NN(cp_list);
2118 S_ssc_clear_locale(regnode_ssc *ssc)
2120 /* Set the SSC 'ssc' to not match any locale things */
2121 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
2123 assert(is_ANYOF_SYNTHETIC(ssc));
2125 ANYOF_POSIXL_ZERO(ssc);
2126 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
2130 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
2132 /* The synthetic start class is used to hopefully quickly winnow down
2133 * places where a pattern could start a match in the target string. If it
2134 * doesn't really narrow things down that much, there isn't much point to
2135 * having the overhead of using it. This function uses some very crude
2136 * heuristics to decide if to use the ssc or not.
2138 * It returns TRUE if 'ssc' rules out more than half what it considers to
2139 * be the "likely" possible matches, but of course it doesn't know what the
2140 * actual things being matched are going to be; these are only guesses
2142 * For /l matches, it assumes that the only likely matches are going to be
2143 * in the 0-255 range, uniformly distributed, so half of that is 127
2144 * For /a and /d matches, it assumes that the likely matches will be just
2145 * the ASCII range, so half of that is 63
2146 * For /u and there isn't anything matching above the Latin1 range, it
2147 * assumes that that is the only range likely to be matched, and uses
2148 * half that as the cut-off: 127. If anything matches above Latin1,
2149 * it assumes that all of Unicode could match (uniformly), except for
2150 * non-Unicode code points and things in the General Category "Other"
2151 * (unassigned, private use, surrogates, controls and formats). This
2152 * is a much large number. */
2154 U32 count = 0; /* Running total of number of code points matched by
2156 UV start, end; /* Start and end points of current range in inversion
2157 XXX outdated. UTF-8 locales are common, what about invert? list */
2158 const U32 max_code_points = (LOC)
2160 : (( ! UNI_SEMANTICS
2161 || invlist_highest(ssc->invlist) < 256)
2164 const U32 max_match = max_code_points / 2;
2166 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
2168 invlist_iterinit(ssc->invlist);
2169 while (invlist_iternext(ssc->invlist, &start, &end)) {
2170 if (start >= max_code_points) {
2173 end = MIN(end, max_code_points - 1);
2174 count += end - start + 1;
2175 if (count >= max_match) {
2176 invlist_iterfinish(ssc->invlist);
2186 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2188 /* The inversion list in the SSC is marked mortal; now we need a more
2189 * permanent copy, which is stored the same way that is done in a regular
2190 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2193 SV* invlist = invlist_clone(ssc->invlist, NULL);
2195 PERL_ARGS_ASSERT_SSC_FINALIZE;
2197 assert(is_ANYOF_SYNTHETIC(ssc));
2199 /* The code in this file assumes that all but these flags aren't relevant
2200 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2201 * by the time we reach here */
2202 assert(! (ANYOF_FLAGS(ssc)
2203 & ~( ANYOF_COMMON_FLAGS
2204 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2205 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
2207 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2209 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist, NULL, NULL);
2210 SvREFCNT_dec(invlist);
2212 /* Make sure is clone-safe */
2213 ssc->invlist = NULL;
2215 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2216 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2217 OP(ssc) = ANYOFPOSIXL;
2219 else if (RExC_contains_locale) {
2223 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2226 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2227 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2228 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2229 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2230 ? (TRIE_LIST_CUR( idx ) - 1) \
2236 dump_trie(trie,widecharmap,revcharmap)
2237 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2238 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2240 These routines dump out a trie in a somewhat readable format.
2241 The _interim_ variants are used for debugging the interim
2242 tables that are used to generate the final compressed
2243 representation which is what dump_trie expects.
2245 Part of the reason for their existence is to provide a form
2246 of documentation as to how the different representations function.
2251 Dumps the final compressed table form of the trie to Perl_debug_log.
2252 Used for debugging make_trie().
2256 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2257 AV *revcharmap, U32 depth)
2260 SV *sv=sv_newmortal();
2261 int colwidth= widecharmap ? 6 : 4;
2263 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2265 PERL_ARGS_ASSERT_DUMP_TRIE;
2267 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2268 depth+1, "Match","Base","Ofs" );
2270 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2271 SV ** const tmp = av_fetch( revcharmap, state, 0);
2273 Perl_re_printf( aTHX_ "%*s",
2275 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2276 PL_colors[0], PL_colors[1],
2277 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2278 PERL_PV_ESCAPE_FIRSTCHAR
2283 Perl_re_printf( aTHX_ "\n");
2284 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2286 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2287 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2288 Perl_re_printf( aTHX_ "\n");
2290 for( state = 1 ; state < trie->statecount ; state++ ) {
2291 const U32 base = trie->states[ state ].trans.base;
2293 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2295 if ( trie->states[ state ].wordnum ) {
2296 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2298 Perl_re_printf( aTHX_ "%6s", "" );
2301 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2306 while( ( base + ofs < trie->uniquecharcount ) ||
2307 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2308 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2312 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2314 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2315 if ( ( base + ofs >= trie->uniquecharcount )
2316 && ( base + ofs - trie->uniquecharcount
2318 && trie->trans[ base + ofs
2319 - trie->uniquecharcount ].check == state )
2321 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2322 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2325 Perl_re_printf( aTHX_ "%*s", colwidth," ." );
2329 Perl_re_printf( aTHX_ "]");
2332 Perl_re_printf( aTHX_ "\n" );
2334 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2336 for (word=1; word <= trie->wordcount; word++) {
2337 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2338 (int)word, (int)(trie->wordinfo[word].prev),
2339 (int)(trie->wordinfo[word].len));
2341 Perl_re_printf( aTHX_ "\n" );
2344 Dumps a fully constructed but uncompressed trie in list form.
2345 List tries normally only are used for construction when the number of
2346 possible chars (trie->uniquecharcount) is very high.
2347 Used for debugging make_trie().
2350 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2351 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2355 SV *sv=sv_newmortal();
2356 int colwidth= widecharmap ? 6 : 4;
2357 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2359 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2361 /* print out the table precompression. */
2362 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2364 Perl_re_indentf( aTHX_ "%s",
2365 depth+1, "------:-----+-----------------\n" );
2367 for( state=1 ; state < next_alloc ; state ++ ) {
2370 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2371 depth+1, (UV)state );
2372 if ( ! trie->states[ state ].wordnum ) {
2373 Perl_re_printf( aTHX_ "%5s| ","");
2375 Perl_re_printf( aTHX_ "W%4x| ",
2376 trie->states[ state ].wordnum
2379 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2380 SV ** const tmp = av_fetch( revcharmap,
2381 TRIE_LIST_ITEM(state, charid).forid, 0);
2383 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2385 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2387 PL_colors[0], PL_colors[1],
2388 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2389 | PERL_PV_ESCAPE_FIRSTCHAR
2391 TRIE_LIST_ITEM(state, charid).forid,
2392 (UV)TRIE_LIST_ITEM(state, charid).newstate
2395 Perl_re_printf( aTHX_ "\n%*s| ",
2396 (int)((depth * 2) + 14), "");
2399 Perl_re_printf( aTHX_ "\n");
2404 Dumps a fully constructed but uncompressed trie in table form.
2405 This is the normal DFA style state transition table, with a few
2406 twists to facilitate compression later.
2407 Used for debugging make_trie().
2410 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2411 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2416 SV *sv=sv_newmortal();
2417 int colwidth= widecharmap ? 6 : 4;
2418 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2420 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2423 print out the table precompression so that we can do a visual check
2424 that they are identical.
2427 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2429 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2430 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2432 Perl_re_printf( aTHX_ "%*s",
2434 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2435 PL_colors[0], PL_colors[1],
2436 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2437 PERL_PV_ESCAPE_FIRSTCHAR
2443 Perl_re_printf( aTHX_ "\n");
2444 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2446 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2447 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2450 Perl_re_printf( aTHX_ "\n" );
2452 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2454 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2456 (UV)TRIE_NODENUM( state ) );
2458 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2459 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2461 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2463 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2465 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2466 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2467 (UV)trie->trans[ state ].check );
2469 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2470 (UV)trie->trans[ state ].check,
2471 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2479 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2480 startbranch: the first branch in the whole branch sequence
2481 first : start branch of sequence of branch-exact nodes.
2482 May be the same as startbranch
2483 last : Thing following the last branch.
2484 May be the same as tail.
2485 tail : item following the branch sequence
2486 count : words in the sequence
2487 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2488 depth : indent depth
2490 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2492 A trie is an N'ary tree where the branches are determined by digital
2493 decomposition of the key. IE, at the root node you look up the 1st character and
2494 follow that branch repeat until you find the end of the branches. Nodes can be
2495 marked as "accepting" meaning they represent a complete word. Eg:
2499 would convert into the following structure. Numbers represent states, letters
2500 following numbers represent valid transitions on the letter from that state, if
2501 the number is in square brackets it represents an accepting state, otherwise it
2502 will be in parenthesis.
2504 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2508 (1) +-i->(6)-+-s->[7]
2510 +-s->(3)-+-h->(4)-+-e->[5]
2512 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2514 This shows that when matching against the string 'hers' we will begin at state 1
2515 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2516 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2517 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2518 single traverse. We store a mapping from accepting to state to which word was
2519 matched, and then when we have multiple possibilities we try to complete the
2520 rest of the regex in the order in which they occurred in the alternation.
2522 The only prior NFA like behaviour that would be changed by the TRIE support is
2523 the silent ignoring of duplicate alternations which are of the form:
2525 / (DUPE|DUPE) X? (?{ ... }) Y /x
2527 Thus EVAL blocks following a trie may be called a different number of times with
2528 and without the optimisation. With the optimisations dupes will be silently
2529 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2530 the following demonstrates:
2532 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2534 which prints out 'word' three times, but
2536 'words'=~/(word|word|word)(?{ print $1 })S/
2538 which doesnt print it out at all. This is due to other optimisations kicking in.
2540 Example of what happens on a structural level:
2542 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2544 1: CURLYM[1] {1,32767}(18)
2555 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2556 and should turn into:
2558 1: CURLYM[1] {1,32767}(18)
2560 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2568 Cases where tail != last would be like /(?foo|bar)baz/:
2578 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2579 and would end up looking like:
2582 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2589 d = uvchr_to_utf8_flags(d, uv, 0);
2591 is the recommended Unicode-aware way of saying
2596 #define TRIE_STORE_REVCHAR(val) \
2599 SV *zlopp = newSV(UTF8_MAXBYTES); \
2600 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2601 unsigned char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2603 SvCUR_set(zlopp, kapow - flrbbbbb); \
2606 av_push(revcharmap, zlopp); \
2608 char ooooff = (char)val; \
2609 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2613 /* This gets the next character from the input, folding it if not already
2615 #define TRIE_READ_CHAR STMT_START { \
2618 /* if it is UTF then it is either already folded, or does not need \
2620 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2622 else if (folder == PL_fold_latin1) { \
2623 /* This folder implies Unicode rules, which in the range expressible \
2624 * by not UTF is the lower case, with the two exceptions, one of \
2625 * which should have been taken care of before calling this */ \
2626 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2627 uvc = toLOWER_L1(*uc); \
2628 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2631 /* raw data, will be folded later if needed */ \
2639 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2640 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2641 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2642 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2643 TRIE_LIST_LEN( state ) = ging; \
2645 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2646 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2647 TRIE_LIST_CUR( state )++; \
2650 #define TRIE_LIST_NEW(state) STMT_START { \
2651 Newx( trie->states[ state ].trans.list, \
2652 4, reg_trie_trans_le ); \
2653 TRIE_LIST_CUR( state ) = 1; \
2654 TRIE_LIST_LEN( state ) = 4; \
2657 #define TRIE_HANDLE_WORD(state) STMT_START { \
2658 U16 dupe= trie->states[ state ].wordnum; \
2659 regnode * const noper_next = regnext( noper ); \
2662 /* store the word for dumping */ \
2664 if (OP(noper) != NOTHING) \
2665 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2667 tmp = newSVpvn_utf8( "", 0, UTF ); \
2668 av_push( trie_words, tmp ); \
2672 trie->wordinfo[curword].prev = 0; \
2673 trie->wordinfo[curword].len = wordlen; \
2674 trie->wordinfo[curword].accept = state; \
2676 if ( noper_next < tail ) { \
2678 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2680 trie->jump[curword] = (U16)(noper_next - convert); \
2682 jumper = noper_next; \
2684 nextbranch= regnext(cur); \
2688 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2689 /* chain, so that when the bits of chain are later */\
2690 /* linked together, the dups appear in the chain */\
2691 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2692 trie->wordinfo[dupe].prev = curword; \
2694 /* we haven't inserted this word yet. */ \
2695 trie->states[ state ].wordnum = curword; \
2700 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2701 ( ( base + charid >= ucharcount \
2702 && base + charid < ubound \
2703 && state == trie->trans[ base - ucharcount + charid ].check \
2704 && trie->trans[ base - ucharcount + charid ].next ) \
2705 ? trie->trans[ base - ucharcount + charid ].next \
2706 : ( state==1 ? special : 0 ) \
2709 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2711 TRIE_BITMAP_SET(trie, uvc); \
2712 /* store the folded codepoint */ \
2714 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2717 /* store first byte of utf8 representation of */ \
2718 /* variant codepoints */ \
2719 if (! UVCHR_IS_INVARIANT(uvc)) { \
2720 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2725 #define MADE_JUMP_TRIE 2
2726 #define MADE_EXACT_TRIE 4
2729 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2730 regnode *first, regnode *last, regnode *tail,
2731 U32 word_count, U32 flags, U32 depth)
2733 /* first pass, loop through and scan words */
2734 reg_trie_data *trie;
2735 HV *widecharmap = NULL;
2736 AV *revcharmap = newAV();
2742 regnode *jumper = NULL;
2743 regnode *nextbranch = NULL;
2744 regnode *convert = NULL;
2745 U32 *prev_states; /* temp array mapping each state to previous one */
2746 /* we just use folder as a flag in utf8 */
2747 const U8 * folder = NULL;
2749 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2750 * which stands for one trie structure, one hash, optionally followed
2753 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2754 AV *trie_words = NULL;
2755 /* along with revcharmap, this only used during construction but both are
2756 * useful during debugging so we store them in the struct when debugging.
2759 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2760 STRLEN trie_charcount=0;
2762 SV *re_trie_maxbuff;
2763 DECLARE_AND_GET_RE_DEBUG_FLAGS;
2765 PERL_ARGS_ASSERT_MAKE_TRIE;
2767 PERL_UNUSED_ARG(depth);
2771 case EXACT: case EXACT_REQ8: case EXACTL: break;
2775 case EXACTFLU8: folder = PL_fold_latin1; break;
2776 case EXACTF: folder = PL_fold; break;
2777 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2780 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2782 trie->startstate = 1;
2783 trie->wordcount = word_count;
2784 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2785 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2786 if (flags == EXACT || flags == EXACT_REQ8 || flags == EXACTL)
2787 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2788 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2789 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2792 trie_words = newAV();
2795 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, GV_ADD);
2796 assert(re_trie_maxbuff);
2797 if (!SvIOK(re_trie_maxbuff)) {
2798 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2800 DEBUG_TRIE_COMPILE_r({
2801 Perl_re_indentf( aTHX_
2802 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2804 REG_NODE_NUM(startbranch), REG_NODE_NUM(first),
2805 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2808 /* Find the node we are going to overwrite */
2809 if ( first == startbranch && OP( last ) != BRANCH ) {
2810 /* whole branch chain */
2813 /* branch sub-chain */
2814 convert = NEXTOPER( first );
2817 /* -- First loop and Setup --
2819 We first traverse the branches and scan each word to determine if it
2820 contains widechars, and how many unique chars there are, this is
2821 important as we have to build a table with at least as many columns as we
2824 We use an array of integers to represent the character codes 0..255
2825 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2826 the native representation of the character value as the key and IV's for
2829 *TODO* If we keep track of how many times each character is used we can
2830 remap the columns so that the table compression later on is more
2831 efficient in terms of memory by ensuring the most common value is in the
2832 middle and the least common are on the outside. IMO this would be better
2833 than a most to least common mapping as theres a decent chance the most
2834 common letter will share a node with the least common, meaning the node
2835 will not be compressible. With a middle is most common approach the worst
2836 case is when we have the least common nodes twice.
2840 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2841 regnode *noper = NEXTOPER( cur );
2845 U32 wordlen = 0; /* required init */
2846 STRLEN minchars = 0;
2847 STRLEN maxchars = 0;
2848 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2851 if (OP(noper) == NOTHING) {
2852 /* skip past a NOTHING at the start of an alternation
2853 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2855 * If the next node is not something we are supposed to process
2856 * we will just ignore it due to the condition guarding the
2860 regnode *noper_next= regnext(noper);
2861 if (noper_next < tail)
2866 && ( OP(noper) == flags
2867 || (flags == EXACT && OP(noper) == EXACT_REQ8)
2868 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
2869 || OP(noper) == EXACTFUP))))
2871 uc= (U8*)STRING(noper);
2872 e= uc + STR_LEN(noper);
2879 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2880 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2881 regardless of encoding */
2882 if (OP( noper ) == EXACTFUP) {
2883 /* false positives are ok, so just set this */
2884 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2888 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2890 TRIE_CHARCOUNT(trie)++;
2893 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2894 * is in effect. Under /i, this character can match itself, or
2895 * anything that folds to it. If not under /i, it can match just
2896 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2897 * all fold to k, and all are single characters. But some folds
2898 * expand to more than one character, so for example LATIN SMALL
2899 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2900 * the string beginning at 'uc' is 'ffi', it could be matched by
2901 * three characters, or just by the one ligature character. (It
2902 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2903 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2904 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2905 * match.) The trie needs to know the minimum and maximum number
2906 * of characters that could match so that it can use size alone to
2907 * quickly reject many match attempts. The max is simple: it is
2908 * the number of folded characters in this branch (since a fold is
2909 * never shorter than what folds to it. */
2913 /* And the min is equal to the max if not under /i (indicated by
2914 * 'folder' being NULL), or there are no multi-character folds. If
2915 * there is a multi-character fold, the min is incremented just
2916 * once, for the character that folds to the sequence. Each
2917 * character in the sequence needs to be added to the list below of
2918 * characters in the trie, but we count only the first towards the
2919 * min number of characters needed. This is done through the
2920 * variable 'foldlen', which is returned by the macros that look
2921 * for these sequences as the number of bytes the sequence
2922 * occupies. Each time through the loop, we decrement 'foldlen' by
2923 * how many bytes the current char occupies. Only when it reaches
2924 * 0 do we increment 'minchars' or look for another multi-character
2926 if (folder == NULL) {
2929 else if (foldlen > 0) {
2930 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2935 /* See if *uc is the beginning of a multi-character fold. If
2936 * so, we decrement the length remaining to look at, to account
2937 * for the current character this iteration. (We can use 'uc'
2938 * instead of the fold returned by TRIE_READ_CHAR because the
2939 * macro is smart enough to account for any unfolded
2942 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2943 foldlen -= UTF8SKIP(uc);
2946 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2951 /* The current character (and any potential folds) should be added
2952 * to the possible matching characters for this position in this
2956 U8 folded= folder[ (U8) uvc ];
2957 if ( !trie->charmap[ folded ] ) {
2958 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2959 TRIE_STORE_REVCHAR( folded );
2962 if ( !trie->charmap[ uvc ] ) {
2963 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2964 TRIE_STORE_REVCHAR( uvc );
2967 /* store the codepoint in the bitmap, and its folded
2969 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2970 set_bit = 0; /* We've done our bit :-) */
2974 /* XXX We could come up with the list of code points that fold
2975 * to this using PL_utf8_foldclosures, except not for
2976 * multi-char folds, as there may be multiple combinations
2977 * there that could work, which needs to wait until runtime to
2978 * resolve (The comment about LIGATURE FFI above is such an
2983 widecharmap = newHV();
2985 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2988 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2990 if ( !SvTRUE( *svpp ) ) {
2991 sv_setiv( *svpp, ++trie->uniquecharcount );
2992 TRIE_STORE_REVCHAR(uvc);
2995 } /* end loop through characters in this branch of the trie */
2997 /* We take the min and max for this branch and combine to find the min
2998 * and max for all branches processed so far */
2999 if( cur == first ) {
3000 trie->minlen = minchars;
3001 trie->maxlen = maxchars;
3002 } else if (minchars < trie->minlen) {
3003 trie->minlen = minchars;
3004 } else if (maxchars > trie->maxlen) {
3005 trie->maxlen = maxchars;
3007 } /* end first pass */
3008 DEBUG_TRIE_COMPILE_r(
3009 Perl_re_indentf( aTHX_
3010 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
3012 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
3013 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
3014 (int)trie->minlen, (int)trie->maxlen )
3018 We now know what we are dealing with in terms of unique chars and
3019 string sizes so we can calculate how much memory a naive
3020 representation using a flat table will take. If it's over a reasonable
3021 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
3022 conservative but potentially much slower representation using an array
3025 At the end we convert both representations into the same compressed
3026 form that will be used in regexec.c for matching with. The latter
3027 is a form that cannot be used to construct with but has memory
3028 properties similar to the list form and access properties similar
3029 to the table form making it both suitable for fast searches and
3030 small enough that its feasable to store for the duration of a program.
3032 See the comment in the code where the compressed table is produced
3033 inplace from the flat tabe representation for an explanation of how
3034 the compression works.
3039 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
3042 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
3043 > SvIV(re_trie_maxbuff) )
3046 Second Pass -- Array Of Lists Representation
3048 Each state will be represented by a list of charid:state records
3049 (reg_trie_trans_le) the first such element holds the CUR and LEN
3050 points of the allocated array. (See defines above).
3052 We build the initial structure using the lists, and then convert
3053 it into the compressed table form which allows faster lookups
3054 (but cant be modified once converted).
3057 STRLEN transcount = 1;
3059 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
3062 trie->states = (reg_trie_state *)
3063 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3064 sizeof(reg_trie_state) );
3068 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3070 regnode *noper = NEXTOPER( cur );
3071 U32 state = 1; /* required init */
3072 U16 charid = 0; /* sanity init */
3073 U32 wordlen = 0; /* required init */
3075 if (OP(noper) == NOTHING) {
3076 regnode *noper_next= regnext(noper);
3077 if (noper_next < tail)
3079 /* we will undo this assignment if noper does not
3080 * point at a trieable type in the else clause of
3081 * the following statement. */
3085 && ( OP(noper) == flags
3086 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3087 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
3088 || OP(noper) == EXACTFUP))))
3090 const U8 *uc= (U8*)STRING(noper);
3091 const U8 *e= uc + STR_LEN(noper);
3093 for ( ; uc < e ; uc += len ) {
3098 charid = trie->charmap[ uvc ];
3100 SV** const svpp = hv_fetch( widecharmap,
3107 charid=(U16)SvIV( *svpp );
3110 /* charid is now 0 if we dont know the char read, or
3111 * nonzero if we do */
3118 if ( !trie->states[ state ].trans.list ) {
3119 TRIE_LIST_NEW( state );
3122 check <= TRIE_LIST_USED( state );
3125 if ( TRIE_LIST_ITEM( state, check ).forid
3128 newstate = TRIE_LIST_ITEM( state, check ).newstate;
3133 newstate = next_alloc++;
3134 prev_states[newstate] = state;
3135 TRIE_LIST_PUSH( state, charid, newstate );
3140 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3144 /* If we end up here it is because we skipped past a NOTHING, but did not end up
3145 * on a trieable type. So we need to reset noper back to point at the first regop
3146 * in the branch before we call TRIE_HANDLE_WORD()
3148 noper= NEXTOPER(cur);
3150 TRIE_HANDLE_WORD(state);
3152 } /* end second pass */
3154 /* next alloc is the NEXT state to be allocated */
3155 trie->statecount = next_alloc;
3156 trie->states = (reg_trie_state *)
3157 PerlMemShared_realloc( trie->states,
3159 * sizeof(reg_trie_state) );
3161 /* and now dump it out before we compress it */
3162 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
3163 revcharmap, next_alloc,
3167 trie->trans = (reg_trie_trans *)
3168 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
3175 for( state=1 ; state < next_alloc ; state ++ ) {
3179 DEBUG_TRIE_COMPILE_MORE_r(
3180 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
3184 if (trie->states[state].trans.list) {
3185 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
3189 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3190 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3191 if ( forid < minid ) {
3193 } else if ( forid > maxid ) {
3197 if ( transcount < tp + maxid - minid + 1) {
3199 trie->trans = (reg_trie_trans *)
3200 PerlMemShared_realloc( trie->trans,
3202 * sizeof(reg_trie_trans) );
3203 Zero( trie->trans + (transcount / 2),
3207 base = trie->uniquecharcount + tp - minid;
3208 if ( maxid == minid ) {
3210 for ( ; zp < tp ; zp++ ) {
3211 if ( ! trie->trans[ zp ].next ) {
3212 base = trie->uniquecharcount + zp - minid;
3213 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3215 trie->trans[ zp ].check = state;
3221 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3223 trie->trans[ tp ].check = state;
3228 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3229 const U32 tid = base
3230 - trie->uniquecharcount
3231 + TRIE_LIST_ITEM( state, idx ).forid;
3232 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3234 trie->trans[ tid ].check = state;
3236 tp += ( maxid - minid + 1 );
3238 Safefree(trie->states[ state ].trans.list);
3241 DEBUG_TRIE_COMPILE_MORE_r(
3242 Perl_re_printf( aTHX_ " base: %d\n",base);
3245 trie->states[ state ].trans.base=base;
3247 trie->lasttrans = tp + 1;
3251 Second Pass -- Flat Table Representation.
3253 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3254 each. We know that we will need Charcount+1 trans at most to store
3255 the data (one row per char at worst case) So we preallocate both
3256 structures assuming worst case.
3258 We then construct the trie using only the .next slots of the entry
3261 We use the .check field of the first entry of the node temporarily
3262 to make compression both faster and easier by keeping track of how
3263 many non zero fields are in the node.
3265 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3268 There are two terms at use here: state as a TRIE_NODEIDX() which is
3269 a number representing the first entry of the node, and state as a
3270 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3271 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3272 if there are 2 entrys per node. eg:
3280 The table is internally in the right hand, idx form. However as we
3281 also have to deal with the states array which is indexed by nodenum
3282 we have to use TRIE_NODENUM() to convert.
3285 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3288 trie->trans = (reg_trie_trans *)
3289 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3290 * trie->uniquecharcount + 1,
3291 sizeof(reg_trie_trans) );
3292 trie->states = (reg_trie_state *)
3293 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3294 sizeof(reg_trie_state) );
3295 next_alloc = trie->uniquecharcount + 1;
3298 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3300 regnode *noper = NEXTOPER( cur );
3302 U32 state = 1; /* required init */
3304 U16 charid = 0; /* sanity init */
3305 U32 accept_state = 0; /* sanity init */
3307 U32 wordlen = 0; /* required init */
3309 if (OP(noper) == NOTHING) {
3310 regnode *noper_next= regnext(noper);
3311 if (noper_next < tail)
3313 /* we will undo this assignment if noper does not
3314 * point at a trieable type in the else clause of
3315 * the following statement. */
3319 && ( OP(noper) == flags
3320 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3321 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
3322 || OP(noper) == EXACTFUP))))
3324 const U8 *uc= (U8*)STRING(noper);
3325 const U8 *e= uc + STR_LEN(noper);
3327 for ( ; uc < e ; uc += len ) {
3332 charid = trie->charmap[ uvc ];
3334 SV* const * const svpp = hv_fetch( widecharmap,
3338 charid = svpp ? (U16)SvIV(*svpp) : 0;
3342 if ( !trie->trans[ state + charid ].next ) {
3343 trie->trans[ state + charid ].next = next_alloc;
3344 trie->trans[ state ].check++;
3345 prev_states[TRIE_NODENUM(next_alloc)]
3346 = TRIE_NODENUM(state);
3347 next_alloc += trie->uniquecharcount;
3349 state = trie->trans[ state + charid ].next;
3351 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3353 /* charid is now 0 if we dont know the char read, or
3354 * nonzero if we do */
3357 /* If we end up here it is because we skipped past a NOTHING, but did not end up
3358 * on a trieable type. So we need to reset noper back to point at the first regop
3359 * in the branch before we call TRIE_HANDLE_WORD().
3361 noper= NEXTOPER(cur);
3363 accept_state = TRIE_NODENUM( state );
3364 TRIE_HANDLE_WORD(accept_state);
3366 } /* end second pass */
3368 /* and now dump it out before we compress it */
3369 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3371 next_alloc, depth+1));
3375 * Inplace compress the table.*
3377 For sparse data sets the table constructed by the trie algorithm will
3378 be mostly 0/FAIL transitions or to put it another way mostly empty.
3379 (Note that leaf nodes will not contain any transitions.)
3381 This algorithm compresses the tables by eliminating most such
3382 transitions, at the cost of a modest bit of extra work during lookup:
3384 - Each states[] entry contains a .base field which indicates the
3385 index in the state[] array wheres its transition data is stored.
3387 - If .base is 0 there are no valid transitions from that node.
3389 - If .base is nonzero then charid is added to it to find an entry in
3392 -If trans[states[state].base+charid].check!=state then the
3393 transition is taken to be a 0/Fail transition. Thus if there are fail
3394 transitions at the front of the node then the .base offset will point
3395 somewhere inside the previous nodes data (or maybe even into a node
3396 even earlier), but the .check field determines if the transition is
3400 The following process inplace converts the table to the compressed
3401 table: We first do not compress the root node 1,and mark all its
3402 .check pointers as 1 and set its .base pointer as 1 as well. This
3403 allows us to do a DFA construction from the compressed table later,
3404 and ensures that any .base pointers we calculate later are greater
3407 - We set 'pos' to indicate the first entry of the second node.
3409 - We then iterate over the columns of the node, finding the first and
3410 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3411 and set the .check pointers accordingly, and advance pos
3412 appropriately and repreat for the next node. Note that when we copy
3413 the next pointers we have to convert them from the original
3414 NODEIDX form to NODENUM form as the former is not valid post
3417 - If a node has no transitions used we mark its base as 0 and do not
3418 advance the pos pointer.
3420 - If a node only has one transition we use a second pointer into the
3421 structure to fill in allocated fail transitions from other states.
3422 This pointer is independent of the main pointer and scans forward
3423 looking for null transitions that are allocated to a state. When it
3424 finds one it writes the single transition into the "hole". If the
3425 pointer doesnt find one the single transition is appended as normal.
3427 - Once compressed we can Renew/realloc the structures to release the
3430 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3431 specifically Fig 3.47 and the associated pseudocode.
3435 const U32 laststate = TRIE_NODENUM( next_alloc );
3438 trie->statecount = laststate;
3440 for ( state = 1 ; state < laststate ; state++ ) {
3442 const U32 stateidx = TRIE_NODEIDX( state );
3443 const U32 o_used = trie->trans[ stateidx ].check;
3444 U32 used = trie->trans[ stateidx ].check;
3445 trie->trans[ stateidx ].check = 0;
3448 used && charid < trie->uniquecharcount;
3451 if ( flag || trie->trans[ stateidx + charid ].next ) {
3452 if ( trie->trans[ stateidx + charid ].next ) {
3454 for ( ; zp < pos ; zp++ ) {
3455 if ( ! trie->trans[ zp ].next ) {
3459 trie->states[ state ].trans.base
3461 + trie->uniquecharcount
3463 trie->trans[ zp ].next
3464 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3466 trie->trans[ zp ].check = state;
3467 if ( ++zp > pos ) pos = zp;
3474 trie->states[ state ].trans.base
3475 = pos + trie->uniquecharcount - charid ;
3477 trie->trans[ pos ].next
3478 = SAFE_TRIE_NODENUM(
3479 trie->trans[ stateidx + charid ].next );
3480 trie->trans[ pos ].check = state;
3485 trie->lasttrans = pos + 1;
3486 trie->states = (reg_trie_state *)
3487 PerlMemShared_realloc( trie->states, laststate
3488 * sizeof(reg_trie_state) );
3489 DEBUG_TRIE_COMPILE_MORE_r(
3490 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3492 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3496 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3499 } /* end table compress */
3501 DEBUG_TRIE_COMPILE_MORE_r(
3502 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3504 (UV)trie->statecount,
3505 (UV)trie->lasttrans)
3507 /* resize the trans array to remove unused space */
3508 trie->trans = (reg_trie_trans *)
3509 PerlMemShared_realloc( trie->trans, trie->lasttrans
3510 * sizeof(reg_trie_trans) );
3512 { /* Modify the program and insert the new TRIE node */
3513 U8 nodetype =(U8)(flags & 0xFF);
3517 regnode *optimize = NULL;
3518 #ifdef RE_TRACK_PATTERN_OFFSETS
3521 U32 mjd_nodelen = 0;
3522 #endif /* RE_TRACK_PATTERN_OFFSETS */
3523 #endif /* DEBUGGING */
3525 This means we convert either the first branch or the first Exact,
3526 depending on whether the thing following (in 'last') is a branch
3527 or not and whther first is the startbranch (ie is it a sub part of
3528 the alternation or is it the whole thing.)
3529 Assuming its a sub part we convert the EXACT otherwise we convert
3530 the whole branch sequence, including the first.
3532 /* Find the node we are going to overwrite */
3533 if ( first != startbranch || OP( last ) == BRANCH ) {
3534 /* branch sub-chain */
3535 NEXT_OFF( first ) = (U16)(last - first);
3536 #ifdef RE_TRACK_PATTERN_OFFSETS
3538 mjd_offset= Node_Offset((convert));
3539 mjd_nodelen= Node_Length((convert));
3542 /* whole branch chain */
3544 #ifdef RE_TRACK_PATTERN_OFFSETS
3547 const regnode *nop = NEXTOPER( convert );
3548 mjd_offset= Node_Offset((nop));
3549 mjd_nodelen= Node_Length((nop));
3553 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3555 (UV)mjd_offset, (UV)mjd_nodelen)
3558 /* But first we check to see if there is a common prefix we can
3559 split out as an EXACT and put in front of the TRIE node. */
3560 trie->startstate= 1;
3561 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3562 /* we want to find the first state that has more than
3563 * one transition, if that state is not the first state
3564 * then we have a common prefix which we can remove.
3567 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3569 I32 first_ofs = -1; /* keeps track of the ofs of the first
3570 transition, -1 means none */
3572 const U32 base = trie->states[ state ].trans.base;
3574 /* does this state terminate an alternation? */
3575 if ( trie->states[state].wordnum )
3578 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3579 if ( ( base + ofs >= trie->uniquecharcount ) &&
3580 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3581 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3583 if ( ++count > 1 ) {
3584 /* we have more than one transition */
3587 /* if this is the first state there is no common prefix
3588 * to extract, so we can exit */
3589 if ( state == 1 ) break;
3590 tmp = av_fetch( revcharmap, ofs, 0);
3591 ch = (U8*)SvPV_nolen_const( *tmp );
3593 /* if we are on count 2 then we need to initialize the
3594 * bitmap, and store the previous char if there was one
3597 /* clear the bitmap */
3598 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3600 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3603 if (first_ofs >= 0) {
3604 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3605 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3607 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3609 Perl_re_printf( aTHX_ "%s", (char*)ch)
3613 /* store the current firstchar in the bitmap */
3614 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3615 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3621 /* This state has only one transition, its transition is part
3622 * of a common prefix - we need to concatenate the char it
3623 * represents to what we have so far. */
3624 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3626 char *ch = SvPV( *tmp, len );
3628 SV *sv=sv_newmortal();
3629 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3631 (UV)state, (UV)first_ofs,
3632 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3633 PL_colors[0], PL_colors[1],
3634 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3635 PERL_PV_ESCAPE_FIRSTCHAR
3640 OP( convert ) = nodetype;
3641 str=STRING(convert);
3642 setSTR_LEN(convert, 0);
3644 assert( ( STR_LEN(convert) + len ) < 256 );
3645 setSTR_LEN(convert, (U8)(STR_LEN(convert) + len));
3651 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3656 trie->prefixlen = (state-1);
3658 regnode *n = convert+NODE_SZ_STR(convert);
3659 assert( NODE_SZ_STR(convert) <= U16_MAX );
3660 NEXT_OFF(convert) = (U16)(NODE_SZ_STR(convert));
3661 trie->startstate = state;
3662 trie->minlen -= (state - 1);
3663 trie->maxlen -= (state - 1);
3665 /* At least the UNICOS C compiler choked on this
3666 * being argument to DEBUG_r(), so let's just have
3669 #ifdef PERL_EXT_RE_BUILD
3675 regnode *fix = convert;
3676 U32 word = trie->wordcount;
3677 #ifdef RE_TRACK_PATTERN_OFFSETS
3680 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3681 while( ++fix < n ) {
3682 Set_Node_Offset_Length(fix, 0, 0);
3685 SV ** const tmp = av_fetch( trie_words, word, 0 );
3687 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3688 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3690 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3698 NEXT_OFF(convert) = (U16)(tail - convert);
3699 DEBUG_r(optimize= n);
3705 if ( trie->maxlen ) {
3706 NEXT_OFF( convert ) = (U16)(tail - convert);
3707 ARG_SET( convert, data_slot );
3708 /* Store the offset to the first unabsorbed branch in
3709 jump[0], which is otherwise unused by the jump logic.
3710 We use this when dumping a trie and during optimisation. */
3712 trie->jump[0] = (U16)(nextbranch - convert);
3714 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3715 * and there is a bitmap
3716 * and the first "jump target" node we found leaves enough room
3717 * then convert the TRIE node into a TRIEC node, with the bitmap
3718 * embedded inline in the opcode - this is hypothetically faster.
3720 if ( !trie->states[trie->startstate].wordnum
3722 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3724 OP( convert ) = TRIEC;
3725 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3726 PerlMemShared_free(trie->bitmap);
3729 OP( convert ) = TRIE;
3731 /* store the type in the flags */
3732 convert->flags = nodetype;
3736 + regarglen[ OP( convert ) ];
3738 /* XXX We really should free up the resource in trie now,
3739 as we won't use them - (which resources?) dmq */
3741 /* needed for dumping*/
3742 DEBUG_r(if (optimize) {
3743 regnode *opt = convert;
3745 while ( ++opt < optimize) {
3746 Set_Node_Offset_Length(opt, 0, 0);
3749 Try to clean up some of the debris left after the
3752 while( optimize < jumper ) {
3753 Track_Code( mjd_nodelen += Node_Length((optimize)); );
3754 OP( optimize ) = OPTIMIZED;
3755 Set_Node_Offset_Length(optimize, 0, 0);
3758 Set_Node_Offset_Length(convert, mjd_offset, mjd_nodelen);
3760 } /* end node insert */
3762 /* Finish populating the prev field of the wordinfo array. Walk back
3763 * from each accept state until we find another accept state, and if
3764 * so, point the first word's .prev field at the second word. If the
3765 * second already has a .prev field set, stop now. This will be the
3766 * case either if we've already processed that word's accept state,
3767 * or that state had multiple words, and the overspill words were
3768 * already linked up earlier.
3775 for (word=1; word <= trie->wordcount; word++) {
3777 if (trie->wordinfo[word].prev)
3779 state = trie->wordinfo[word].accept;
3781 state = prev_states[state];
3784 prev = trie->states[state].wordnum;
3788 trie->wordinfo[word].prev = prev;
3790 Safefree(prev_states);
3794 /* and now dump out the compressed format */
3795 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3797 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3799 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3800 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3802 SvREFCNT_dec_NN(revcharmap);
3806 : trie->startstate>1
3812 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3814 /* The Trie is constructed and compressed now so we can build a fail array if
3817 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3819 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3823 We find the fail state for each state in the trie, this state is the longest
3824 proper suffix of the current state's 'word' that is also a proper prefix of
3825 another word in our trie. State 1 represents the word '' and is thus the
3826 default fail state. This allows the DFA not to have to restart after its
3827 tried and failed a word at a given point, it simply continues as though it
3828 had been matching the other word in the first place.
3830 'abcdgu'=~/abcdefg|cdgu/
3831 When we get to 'd' we are still matching the first word, we would encounter
3832 'g' which would fail, which would bring us to the state representing 'd' in
3833 the second word where we would try 'g' and succeed, proceeding to match
3836 /* add a fail transition */
3837 const U32 trie_offset = ARG(source);
3838 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3840 const U32 ucharcount = trie->uniquecharcount;
3841 const U32 numstates = trie->statecount;
3842 const U32 ubound = trie->lasttrans + ucharcount;
3846 U32 base = trie->states[ 1 ].trans.base;
3849 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3851 DECLARE_AND_GET_RE_DEBUG_FLAGS;
3853 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3854 PERL_UNUSED_CONTEXT;
3856 PERL_UNUSED_ARG(depth);
3859 if ( OP(source) == TRIE ) {
3860 struct regnode_1 *op = (struct regnode_1 *)
3861 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3862 StructCopy(source, op, struct regnode_1);
3863 stclass = (regnode *)op;
3865 struct regnode_charclass *op = (struct regnode_charclass *)
3866 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3867 StructCopy(source, op, struct regnode_charclass);
3868 stclass = (regnode *)op;
3870 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3872 ARG_SET( stclass, data_slot );
3873 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3874 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3875 aho->trie=trie_offset;
3876 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3877 Copy( trie->states, aho->states, numstates, reg_trie_state );
3878 Newx( q, numstates, U32);
3879 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3882 /* initialize fail[0..1] to be 1 so that we always have
3883 a valid final fail state */
3884 fail[ 0 ] = fail[ 1 ] = 1;
3886 for ( charid = 0; charid < ucharcount ; charid++ ) {
3887 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3889 q[ q_write ] = newstate;
3890 /* set to point at the root */
3891 fail[ q[ q_write++ ] ]=1;
3894 while ( q_read < q_write) {
3895 const U32 cur = q[ q_read++ % numstates ];
3896 base = trie->states[ cur ].trans.base;
3898 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3899 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3901 U32 fail_state = cur;
3904 fail_state = fail[ fail_state ];
3905 fail_base = aho->states[ fail_state ].trans.base;
3906 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3908 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3909 fail[ ch_state ] = fail_state;
3910 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3912 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3914 q[ q_write++ % numstates] = ch_state;
3918 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3919 when we fail in state 1, this allows us to use the
3920 charclass scan to find a valid start char. This is based on the principle
3921 that theres a good chance the string being searched contains lots of stuff
3922 that cant be a start char.
3924 fail[ 0 ] = fail[ 1 ] = 0;
3925 DEBUG_TRIE_COMPILE_r({
3926 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3927 depth, (UV)numstates
3929 for( q_read=1; q_read<numstates; q_read++ ) {
3930 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3932 Perl_re_printf( aTHX_ "\n");
3935 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3940 /* The below joins as many adjacent EXACTish nodes as possible into a single
3941 * one. The regop may be changed if the node(s) contain certain sequences that
3942 * require special handling. The joining is only done if:
3943 * 1) there is room in the current conglomerated node to entirely contain the
3945 * 2) they are compatible node types
3947 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3948 * these get optimized out
3950 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3951 * as possible, even if that means splitting an existing node so that its first
3952 * part is moved to the preceeding node. This would maximise the efficiency of
3953 * memEQ during matching.
3955 * If a node is to match under /i (folded), the number of characters it matches
3956 * can be different than its character length if it contains a multi-character
3957 * fold. *min_subtract is set to the total delta number of characters of the
3960 * And *unfolded_multi_char is set to indicate whether or not the node contains
3961 * an unfolded multi-char fold. This happens when it won't be known until
3962 * runtime whether the fold is valid or not; namely
3963 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3964 * target string being matched against turns out to be UTF-8 is that fold
3966 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3968 * (Multi-char folds whose components are all above the Latin1 range are not
3969 * run-time locale dependent, and have already been folded by the time this
3970 * function is called.)
3972 * This is as good a place as any to discuss the design of handling these
3973 * multi-character fold sequences. It's been wrong in Perl for a very long
3974 * time. There are three code points in Unicode whose multi-character folds
3975 * were long ago discovered to mess things up. The previous designs for
3976 * dealing with these involved assigning a special node for them. This
3977 * approach doesn't always work, as evidenced by this example:
3978 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3979 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3980 * would match just the \xDF, it won't be able to handle the case where a
3981 * successful match would have to cross the node's boundary. The new approach
3982 * that hopefully generally solves the problem generates an EXACTFUP node
3983 * that is "sss" in this case.
3985 * It turns out that there are problems with all multi-character folds, and not
3986 * just these three. Now the code is general, for all such cases. The
3987 * approach taken is:
3988 * 1) This routine examines each EXACTFish node that could contain multi-
3989 * character folded sequences. Since a single character can fold into
3990 * such a sequence, the minimum match length for this node is less than
3991 * the number of characters in the node. This routine returns in
3992 * *min_subtract how many characters to subtract from the actual
3993 * length of the string to get a real minimum match length; it is 0 if
3994 * there are no multi-char foldeds. This delta is used by the caller to
3995 * adjust the min length of the match, and the delta between min and max,
3996 * so that the optimizer doesn't reject these possibilities based on size
3999 * 2) For the sequence involving the LATIN SMALL LETTER SHARP S (U+00DF)
4000 * under /u, we fold it to 'ss' in regatom(), and in this routine, after
4001 * joining, we scan for occurrences of the sequence 'ss' in non-UTF-8
4002 * EXACTFU nodes. The node type of such nodes is then changed to
4003 * EXACTFUP, indicating it is problematic, and needs careful handling.
4004 * (The procedures in step 1) above are sufficient to handle this case in
4005 * UTF-8 encoded nodes.) The reason this is problematic is that this is
4006 * the only case where there is a possible fold length change in non-UTF-8
4007 * patterns. By reserving a special node type for problematic cases, the
4008 * far more common regular EXACTFU nodes can be processed faster.
4009 * regexec.c takes advantage of this.
4011 * EXACTFUP has been created as a grab-bag for (hopefully uncommon)
4012 * problematic cases. These all only occur when the pattern is not
4013 * UTF-8. In addition to the 'ss' sequence where there is a possible fold
4014 * length change, it handles the situation where the string cannot be
4015 * entirely folded. The strings in an EXACTFish node are folded as much
4016 * as possible during compilation in regcomp.c. This saves effort in
4017 * regex matching. By using an EXACTFUP node when it is not possible to
4018 * fully fold at compile time, regexec.c can know that everything in an
4019 * EXACTFU node is folded, so folding can be skipped at runtime. The only
4020 * case where folding in EXACTFU nodes can't be done at compile time is
4021 * the presumably uncommon MICRO SIGN, when the pattern isn't UTF-8. This
4022 * is because its fold requires UTF-8 to represent. Thus EXACTFUP nodes
4023 * handle two very different cases. Alternatively, there could have been
4024 * a node type where there are length changes, one for unfolded, and one
4025 * for both. If yet another special case needed to be created, the number
4026 * of required node types would have to go to 7. khw figures that even
4027 * though there are plenty of node types to spare, that the maintenance
4028 * cost wasn't worth the small speedup of doing it that way, especially
4029 * since he thinks the MICRO SIGN is rarely encountered in practice.
4031 * There are other cases where folding isn't done at compile time, but
4032 * none of them are under /u, and hence not for EXACTFU nodes. The folds
4033 * in EXACTFL nodes aren't known until runtime, and vary as the locale
4034 * changes. Some folds in EXACTF depend on if the runtime target string
4035 * is UTF-8 or not. (regatom() will create an EXACTFU node even under /di
4036 * when no fold in it depends on the UTF-8ness of the target string.)
4038 * 3) A problem remains for unfolded multi-char folds. (These occur when the
4039 * validity of the fold won't be known until runtime, and so must remain
4040 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
4041 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
4042 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
4043 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
4044 * The reason this is a problem is that the optimizer part of regexec.c
4045 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
4046 * that a character in the pattern corresponds to at most a single
4047 * character in the target string. (And I do mean character, and not byte
4048 * here, unlike other parts of the documentation that have never been
4049 * updated to account for multibyte Unicode.) Sharp s in EXACTF and
4050 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
4051 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
4052 * EXACTFL nodes, violate the assumption, and they are the only instances
4053 * where it is violated. I'm reluctant to try to change the assumption,
4054 * as the code involved is impenetrable to me (khw), so instead the code
4055 * here punts. This routine examines EXACTFL nodes, and (when the pattern
4056 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
4057 * boolean indicating whether or not the node contains such a fold. When
4058 * it is true, the caller sets a flag that later causes the optimizer in
4059 * this file to not set values for the floating and fixed string lengths,
4060 * and thus avoids the optimizer code in regexec.c that makes the invalid
4061 * assumption. Thus, there is no optimization based on string lengths for
4062 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
4063 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
4064 * assumption is wrong only in these cases is that all other non-UTF-8
4065 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
4066 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
4067 * EXACTF nodes because we don't know at compile time if it actually
4068 * matches 'ss' or not. For EXACTF nodes it will match iff the target
4069 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
4070 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
4071 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
4072 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
4073 * string would require the pattern to be forced into UTF-8, the overhead
4074 * of which we want to avoid. Similarly the unfolded multi-char folds in
4075 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
4078 * Similarly, the code that generates tries doesn't currently handle
4079 * not-already-folded multi-char folds, and it looks like a pain to change
4080 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
4081 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
4082 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
4083 * using /iaa matching will be doing so almost entirely with ASCII
4084 * strings, so this should rarely be encountered in practice */
4087 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
4088 UV *min_subtract, bool *unfolded_multi_char,
4089 U32 flags, regnode *val, U32 depth)
4091 /* Merge several consecutive EXACTish nodes into one. */
4093 regnode *n = regnext(scan);
4095 regnode *next = scan + NODE_SZ_STR(scan);
4099 regnode *stop = scan;
4100 DECLARE_AND_GET_RE_DEBUG_FLAGS;
4102 PERL_UNUSED_ARG(depth);
4105 PERL_ARGS_ASSERT_JOIN_EXACT;
4106 #ifndef EXPERIMENTAL_INPLACESCAN
4107 PERL_UNUSED_ARG(flags);
4108 PERL_UNUSED_ARG(val);
4110 DEBUG_PEEP("join", scan, depth, 0);
4112 assert(PL_regkind[OP(scan)] == EXACT);
4114 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
4115 * EXACT ones that are mergeable to the current one. */
4117 && ( PL_regkind[OP(n)] == NOTHING
4118 || (stringok && PL_regkind[OP(n)] == EXACT))
4120 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
4123 if (OP(n) == TAIL || n > next)
4125 if (PL_regkind[OP(n)] == NOTHING) {
4126 DEBUG_PEEP("skip:", n, depth, 0);
4127 NEXT_OFF(scan) += NEXT_OFF(n);
4128 next = n + NODE_STEP_REGNODE;
4135 else if (stringok) {
4136 const unsigned int oldl = STR_LEN(scan);
4137 regnode * const nnext = regnext(n);
4139 /* XXX I (khw) kind of doubt that this works on platforms (should
4140 * Perl ever run on one) where U8_MAX is above 255 because of lots
4141 * of other assumptions */
4142 /* Don't join if the sum can't fit into a single node */
4143 if (oldl + STR_LEN(n) > U8_MAX)
4146 /* Joining something that requires UTF-8 with something that
4147 * doesn't, means the result requires UTF-8. */
4148 if (OP(scan) == EXACT && (OP(n) == EXACT_REQ8)) {
4149 OP(scan) = EXACT_REQ8;
4151 else if (OP(scan) == EXACT_REQ8 && (OP(n) == EXACT)) {
4152 ; /* join is compatible, no need to change OP */
4154 else if ((OP(scan) == EXACTFU) && (OP(n) == EXACTFU_REQ8)) {
4155 OP(scan) = EXACTFU_REQ8;
4157 else if ((OP(scan) == EXACTFU_REQ8) && (OP(n) == EXACTFU)) {
4158 ; /* join is compatible, no need to change OP */
4160 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU) {
4161 ; /* join is compatible, no need to change OP */
4163 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU_S_EDGE) {
4165 /* Under /di, temporary EXACTFU_S_EDGE nodes are generated,
4166 * which can join with EXACTFU ones. We check for this case
4167 * here. These need to be resolved to either EXACTFU or
4168 * EXACTF at joining time. They have nothing in them that
4169 * would forbid them from being the more desirable EXACTFU
4170 * nodes except that they begin and/or end with a single [Ss].
4171 * The reason this is problematic is because they could be
4172 * joined in this loop with an adjacent node that ends and/or
4173 * begins with [Ss] which would then form the sequence 'ss',
4174 * which matches differently under /di than /ui, in which case
4175 * EXACTFU can't be used. If the 'ss' sequence doesn't get
4176 * formed, the nodes get absorbed into any adjacent EXACTFU
4177 * node. And if the only adjacent node is EXACTF, they get
4178 * absorbed into that, under the theory that a longer node is
4179 * better than two shorter ones, even if one is EXACTFU. Note
4180 * that EXACTFU_REQ8 is generated only for UTF-8 patterns,
4181 * and the EXACTFU_S_EDGE ones only for non-UTF-8. */
4183 if (STRING(n)[STR_LEN(n)-1] == 's') {
4185 /* Here the joined node would end with 's'. If the node
4186 * following the combination is an EXACTF one, it's better to
4187 * join this trailing edge 's' node with that one, leaving the
4188 * current one in 'scan' be the more desirable EXACTFU */
4189 if (OP(nnext) == EXACTF) {
4193 OP(scan) = EXACTFU_S_EDGE;
4195 } /* Otherwise, the beginning 's' of the 2nd node just
4196 becomes an interior 's' in 'scan' */
4198 else if (OP(scan) == EXACTF && OP(n) == EXACTF) {
4199 ; /* join is compatible, no need to change OP */
4201 else if (OP(scan) == EXACTF && OP(n) == EXACTFU_S_EDGE) {
4203 /* EXACTF nodes are compatible for joining with EXACTFU_S_EDGE
4204 * nodes. But the latter nodes can be also joined with EXACTFU
4205 * ones, and that is a better outcome, so if the node following
4206 * 'n' is EXACTFU, quit now so that those two can be joined
4208 if (OP(nnext) == EXACTFU) {
4212 /* The join is compatible, and the combined node will be
4213 * EXACTF. (These don't care if they begin or end with 's' */
4215 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU_S_EDGE) {
4216 if ( STRING(scan)[STR_LEN(scan)-1] == 's'
4217 && STRING(n)[0] == 's')
4219 /* When combined, we have the sequence 'ss', which means we
4220 * have to remain /di */
4224 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU) {
4225 if (STRING(n)[0] == 's') {
4226 ; /* Here the join is compatible and the combined node
4227 starts with 's', no need to change OP */
4229 else { /* Now the trailing 's' is in the interior */
4233 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTF) {
4235 /* The join is compatible, and the combined node will be
4236 * EXACTF. (These don't care if they begin or end with 's' */
4239 else if (OP(scan) != OP(n)) {
4241 /* The only other compatible joinings are the same node type */
4245 DEBUG_PEEP("merg", n, depth, 0);
4248 NEXT_OFF(scan) += NEXT_OFF(n);
4249 assert( ( STR_LEN(scan) + STR_LEN(n) ) < 256 );
4250 setSTR_LEN(scan, (U8)(STR_LEN(scan) + STR_LEN(n)));
4251 next = n + NODE_SZ_STR(n);
4252 /* Now we can overwrite *n : */
4253 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
4261 #ifdef EXPERIMENTAL_INPLACESCAN
4262 if (flags && !NEXT_OFF(n)) {
4263 DEBUG_PEEP("atch", val, depth, 0);
4264 if (reg_off_by_arg[OP(n)]) {
4265 ARG_SET(n, val - n);
4268 NEXT_OFF(n) = val - n;
4275 /* This temporary node can now be turned into EXACTFU, and must, as
4276 * regexec.c doesn't handle it */
4277 if (OP(scan) == EXACTFU_S_EDGE) {
4282 *unfolded_multi_char = FALSE;
4284 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
4285 * can now analyze for sequences of problematic code points. (Prior to
4286 * this final joining, sequences could have been split over boundaries, and
4287 * hence missed). The sequences only happen in folding, hence for any
4288 * non-EXACT EXACTish node */
4289 if (OP(scan) != EXACT && OP(scan) != EXACT_REQ8 && OP(scan) != EXACTL) {
4290 U8* s0 = (U8*) STRING(scan);
4292 U8* s_end = s0 + STR_LEN(scan);
4294 int total_count_delta = 0; /* Total delta number of characters that
4295 multi-char folds expand to */
4297 /* One pass is made over the node's string looking for all the
4298 * possibilities. To avoid some tests in the loop, there are two main
4299 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
4304 if (OP(scan) == EXACTFL) {
4307 /* An EXACTFL node would already have been changed to another
4308 * node type unless there is at least one character in it that
4309 * is problematic; likely a character whose fold definition
4310 * won't be known until runtime, and so has yet to be folded.
4311 * For all but the UTF-8 locale, folds are 1-1 in length, but
4312 * to handle the UTF-8 case, we need to create a temporary
4313 * folded copy using UTF-8 locale rules in order to analyze it.
4314 * This is because our macros that look to see if a sequence is
4315 * a multi-char fold assume everything is folded (otherwise the
4316 * tests in those macros would be too complicated and slow).
4317 * Note that here, the non-problematic folds will have already
4318 * been done, so we can just copy such characters. We actually
4319 * don't completely fold the EXACTFL string. We skip the
4320 * unfolded multi-char folds, as that would just create work
4321 * below to figure out the size they already are */
4323 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
4326 STRLEN s_len = UTF8SKIP(s);
4327 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
4328 Copy(s, d, s_len, U8);
4331 else if (is_FOLDS_TO_MULTI_utf8(s)) {
4332 *unfolded_multi_char = TRUE;
4333 Copy(s, d, s_len, U8);
4336 else if (isASCII(*s)) {
4337 *(d++) = toFOLD(*s);
4341 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
4347 /* Point the remainder of the routine to look at our temporary
4351 } /* End of creating folded copy of EXACTFL string */
4353 /* Examine the string for a multi-character fold sequence. UTF-8
4354 * patterns have all characters pre-folded by the time this code is
4356 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
4357 length sequence we are looking for is 2 */
4359 int count = 0; /* How many characters in a multi-char fold */
4360 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
4361 if (! len) { /* Not a multi-char fold: get next char */
4366 { /* Here is a generic multi-char fold. */
4367 U8* multi_end = s + len;
4369 /* Count how many characters are in it. In the case of
4370 * /aa, no folds which contain ASCII code points are
4371 * allowed, so check for those, and skip if found. */
4372 if (OP(scan) != EXACTFAA && OP(scan) != EXACTFAA_NO_TRIE) {
4373 count = utf8_length(s, multi_end);
4377 while (s < multi_end) {
4380 goto next_iteration;
4390 /* The delta is how long the sequence is minus 1 (1 is how long
4391 * the character that folds to the sequence is) */
4392 total_count_delta += count - 1;
4396 /* We created a temporary folded copy of the string in EXACTFL
4397 * nodes. Therefore we need to be sure it doesn't go below zero,
4398 * as the real string could be shorter */
4399 if (OP(scan) == EXACTFL) {
4400 int total_chars = utf8_length((U8*) STRING(scan),
4401 (U8*) STRING(scan) + STR_LEN(scan));
4402 if (total_count_delta > total_chars) {
4403 total_count_delta = total_chars;
4407 *min_subtract += total_count_delta;
4410 else if (OP(scan) == EXACTFAA) {
4412 /* Non-UTF-8 pattern, EXACTFAA node. There can't be a multi-char
4413 * fold to the ASCII range (and there are no existing ones in the
4414 * upper latin1 range). But, as outlined in the comments preceding
4415 * this function, we need to flag any occurrences of the sharp s.
4416 * This character forbids trie formation (because of added
4418 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
4419 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
4420 || UNICODE_DOT_DOT_VERSION > 0)
4422 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
4423 OP(scan) = EXACTFAA_NO_TRIE;
4424 *unfolded_multi_char = TRUE;
4430 else if (OP(scan) != EXACTFAA_NO_TRIE) {
4432 /* Non-UTF-8 pattern, not EXACTFAA node. Look for the multi-char
4433 * folds that are all Latin1. As explained in the comments
4434 * preceding this function, we look also for the sharp s in EXACTF
4435 * and EXACTFL nodes; it can be in the final position. Otherwise
4436 * we can stop looking 1 byte earlier because have to find at least
4437 * two characters for a multi-fold */
4438 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
4443 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
4444 if (! len) { /* Not a multi-char fold. */
4445 if (*s == LATIN_SMALL_LETTER_SHARP_S
4446 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4448 *unfolded_multi_char = TRUE;
4455 && isALPHA_FOLD_EQ(*s, 's')
4456 && isALPHA_FOLD_EQ(*(s+1), 's'))
4459 /* EXACTF nodes need to know that the minimum length
4460 * changed so that a sharp s in the string can match this
4461 * ss in the pattern, but they remain EXACTF nodes, as they
4462 * won't match this unless the target string is in UTF-8,
4463 * which we don't know until runtime. EXACTFL nodes can't
4464 * transform into EXACTFU nodes */
4465 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4466 OP(scan) = EXACTFUP;
4470 *min_subtract += len - 1;
4478 /* Allow dumping but overwriting the collection of skipped
4479 * ops and/or strings with fake optimized ops */
4480 n = scan + NODE_SZ_STR(scan);
4488 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl", scan, depth, 0);});
4492 /* REx optimizer. Converts nodes into quicker variants "in place".
4493 Finds fixed substrings. */
4495 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4496 to the position after last scanned or to NULL. */
4498 #define INIT_AND_WITHP \
4499 assert(!and_withp); \
4500 Newx(and_withp, 1, regnode_ssc); \
4501 SAVEFREEPV(and_withp)
4505 S_unwind_scan_frames(pTHX_ const void *p)
4507 scan_frame *f= (scan_frame *)p;
4509 scan_frame *n= f->next_frame;
4515 /* Follow the next-chain of the current node and optimize away
4516 all the NOTHINGs from it.
4519 S_rck_elide_nothing(pTHX_ regnode *node)
4521 PERL_ARGS_ASSERT_RCK_ELIDE_NOTHING;
4523 if (OP(node) != CURLYX) {
4524 const int max = (reg_off_by_arg[OP(node)]
4526 /* I32 may be smaller than U16 on CRAYs! */
4527 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4528 int off = (reg_off_by_arg[OP(node)] ? ARG(node) : NEXT_OFF(node));
4532 /* Skip NOTHING and LONGJMP. */
4536 (PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4537 || ((OP(n) == LONGJMP) && (noff = ARG(n)))
4543 if (reg_off_by_arg[OP(node)])
4546 NEXT_OFF(node) = off;
4551 /* the return from this sub is the minimum length that could possibly match */
4553 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4554 SSize_t *minlenp, SSize_t *deltap,
4559 regnode_ssc *and_withp,
4560 U32 flags, U32 depth, bool was_mutate_ok)
4561 /* scanp: Start here (read-write). */
4562 /* deltap: Write maxlen-minlen here. */
4563 /* last: Stop before this one. */
4564 /* data: string data about the pattern */
4565 /* stopparen: treat close N as END */
4566 /* recursed: which subroutines have we recursed into */
4567 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4569 SSize_t final_minlen;
4570 /* There must be at least this number of characters to match */
4573 regnode *scan = *scanp, *next;
4575 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4576 int is_inf_internal = 0; /* The studied chunk is infinite */
4577 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4578 scan_data_t data_fake;
4579 SV *re_trie_maxbuff = NULL;
4580 regnode *first_non_open = scan;
4581 SSize_t stopmin = OPTIMIZE_INFTY;
4582 scan_frame *frame = NULL;
4583 DECLARE_AND_GET_RE_DEBUG_FLAGS;
4585 PERL_ARGS_ASSERT_STUDY_CHUNK;
4586 RExC_study_started= 1;
4588 Zero(&data_fake, 1, scan_data_t);
4591 while (first_non_open && OP(first_non_open) == OPEN)
4592 first_non_open=regnext(first_non_open);
4598 RExC_study_chunk_recursed_count++;
https://perl5.git.perl.org / perl5.git / blob