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 #define PERL_IN_REGCOMP_C
77 #ifndef PERL_IN_XSUB_RE
82 #ifdef PERL_IN_XSUB_RE
84 EXTERN_C const struct regexp_engine my_reg_engine;
89 #include "dquote_inline.h"
90 #include "invlist_inline.h"
91 #include "unicode_constants.h"
93 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
94 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
96 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
97 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
98 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
101 #define STATIC static
104 /* this is a chain of data about sub patterns we are processing that
105 need to be handled separately/specially in study_chunk. Its so
106 we can simulate recursion without losing state. */
108 typedef struct scan_frame {
109 regnode *last_regnode; /* last node to process in this frame */
110 regnode *next_regnode; /* next node to process when last is reached */
111 U32 prev_recursed_depth;
112 I32 stopparen; /* what stopparen do we use */
114 struct scan_frame *this_prev_frame; /* this previous frame */
115 struct scan_frame *prev_frame; /* previous frame */
116 struct scan_frame *next_frame; /* next frame */
119 /* Certain characters are output as a sequence with the first being a
121 #define isBACKSLASHED_PUNCT(c) strchr("-[]\\^", c)
124 struct RExC_state_t {
125 U32 flags; /* RXf_* are we folding, multilining? */
126 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
127 char *precomp; /* uncompiled string. */
128 char *precomp_end; /* pointer to end of uncompiled string. */
129 REGEXP *rx_sv; /* The SV that is the regexp. */
130 regexp *rx; /* perl core regexp structure */
131 regexp_internal *rxi; /* internal data for regexp object
133 char *start; /* Start of input for compile */
134 char *end; /* End of input for compile */
135 char *parse; /* Input-scan pointer. */
136 char *adjusted_start; /* 'start', adjusted. See code use */
137 STRLEN precomp_adj; /* an offset beyond precomp. See code use */
138 SSize_t whilem_seen; /* number of WHILEM in this expr */
139 regnode *emit_start; /* Start of emitted-code area */
140 regnode *emit_bound; /* First regnode outside of the
142 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
143 implies compiling, so don't emit */
144 regnode_ssc emit_dummy; /* placeholder for emit to point to;
145 large enough for the largest
146 non-EXACTish node, so can use it as
148 I32 naughty; /* How bad is this pattern? */
149 I32 sawback; /* Did we see \1, ...? */
151 SSize_t size; /* Code size. */
152 I32 npar; /* Capture buffer count, (OPEN) plus
153 one. ("par" 0 is the whole
155 I32 nestroot; /* root parens we are in - used by
159 regnode **open_parens; /* pointers to open parens */
160 regnode **close_parens; /* pointers to close parens */
161 regnode *end_op; /* END node in program */
162 I32 utf8; /* whether the pattern is utf8 or not */
163 I32 orig_utf8; /* whether the pattern was originally in utf8 */
164 /* XXX use this for future optimisation of case
165 * where pattern must be upgraded to utf8. */
166 I32 uni_semantics; /* If a d charset modifier should use unicode
167 rules, even if the pattern is not in
169 HV *paren_names; /* Paren names */
171 regnode **recurse; /* Recurse regops */
172 I32 recurse_count; /* Number of recurse regops we have generated */
173 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
175 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
178 I32 override_recoding;
180 I32 recode_x_to_native;
182 I32 in_multi_char_class;
183 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
185 int code_index; /* next code_blocks[] slot */
186 SSize_t maxlen; /* mininum possible number of chars in string to match */
187 scan_frame *frame_head;
188 scan_frame *frame_last;
191 #ifdef ADD_TO_REGEXEC
192 char *starttry; /* -Dr: where regtry was called. */
193 #define RExC_starttry (pRExC_state->starttry)
195 SV *runtime_code_qr; /* qr with the runtime code blocks */
197 const char *lastparse;
199 AV *paren_name_list; /* idx -> name */
200 U32 study_chunk_recursed_count;
203 #define RExC_lastparse (pRExC_state->lastparse)
204 #define RExC_lastnum (pRExC_state->lastnum)
205 #define RExC_paren_name_list (pRExC_state->paren_name_list)
206 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
207 #define RExC_mysv (pRExC_state->mysv1)
208 #define RExC_mysv1 (pRExC_state->mysv1)
209 #define RExC_mysv2 (pRExC_state->mysv2)
212 bool seen_unfolded_sharp_s;
218 #define RExC_flags (pRExC_state->flags)
219 #define RExC_pm_flags (pRExC_state->pm_flags)
220 #define RExC_precomp (pRExC_state->precomp)
221 #define RExC_precomp_adj (pRExC_state->precomp_adj)
222 #define RExC_adjusted_start (pRExC_state->adjusted_start)
223 #define RExC_precomp_end (pRExC_state->precomp_end)
224 #define RExC_rx_sv (pRExC_state->rx_sv)
225 #define RExC_rx (pRExC_state->rx)
226 #define RExC_rxi (pRExC_state->rxi)
227 #define RExC_start (pRExC_state->start)
228 #define RExC_end (pRExC_state->end)
229 #define RExC_parse (pRExC_state->parse)
230 #define RExC_whilem_seen (pRExC_state->whilem_seen)
232 /* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
233 * EXACTF node, hence was parsed under /di rules. If later in the parse,
234 * something forces the pattern into using /ui rules, the sharp s should be
235 * folded into the sequence 'ss', which takes up more space than previously
236 * calculated. This means that the sizing pass needs to be restarted. (The
237 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
238 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
239 * so there is no need to resize [perl #125990]. */
240 #define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
242 #ifdef RE_TRACK_PATTERN_OFFSETS
243 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
246 #define RExC_emit (pRExC_state->emit)
247 #define RExC_emit_dummy (pRExC_state->emit_dummy)
248 #define RExC_emit_start (pRExC_state->emit_start)
249 #define RExC_emit_bound (pRExC_state->emit_bound)
250 #define RExC_sawback (pRExC_state->sawback)
251 #define RExC_seen (pRExC_state->seen)
252 #define RExC_size (pRExC_state->size)
253 #define RExC_maxlen (pRExC_state->maxlen)
254 #define RExC_npar (pRExC_state->npar)
255 #define RExC_nestroot (pRExC_state->nestroot)
256 #define RExC_extralen (pRExC_state->extralen)
257 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
258 #define RExC_utf8 (pRExC_state->utf8)
259 #define RExC_uni_semantics (pRExC_state->uni_semantics)
260 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
261 #define RExC_open_parens (pRExC_state->open_parens)
262 #define RExC_close_parens (pRExC_state->close_parens)
263 #define RExC_end_op (pRExC_state->end_op)
264 #define RExC_paren_names (pRExC_state->paren_names)
265 #define RExC_recurse (pRExC_state->recurse)
266 #define RExC_recurse_count (pRExC_state->recurse_count)
267 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
268 #define RExC_study_chunk_recursed_bytes \
269 (pRExC_state->study_chunk_recursed_bytes)
270 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
271 #define RExC_contains_locale (pRExC_state->contains_locale)
273 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
275 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
276 #define RExC_frame_head (pRExC_state->frame_head)
277 #define RExC_frame_last (pRExC_state->frame_last)
278 #define RExC_frame_count (pRExC_state->frame_count)
279 #define RExC_strict (pRExC_state->strict)
280 #define RExC_study_started (pRExC_state->study_started)
281 #define RExC_warn_text (pRExC_state->warn_text)
282 #define RExC_in_script_run (pRExC_state->in_script_run)
284 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
285 * a flag to disable back-off on the fixed/floating substrings - if it's
286 * a high complexity pattern we assume the benefit of avoiding a full match
287 * is worth the cost of checking for the substrings even if they rarely help.
289 #define RExC_naughty (pRExC_state->naughty)
290 #define TOO_NAUGHTY (10)
291 #define MARK_NAUGHTY(add) \
292 if (RExC_naughty < TOO_NAUGHTY) \
293 RExC_naughty += (add)
294 #define MARK_NAUGHTY_EXP(exp, add) \
295 if (RExC_naughty < TOO_NAUGHTY) \
296 RExC_naughty += RExC_naughty / (exp) + (add)
298 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
299 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
300 ((*s) == '{' && regcurly(s)))
303 * Flags to be passed up and down.
305 #define WORST 0 /* Worst case. */
306 #define HASWIDTH 0x01 /* Known to match non-null strings. */
308 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
309 * character. (There needs to be a case: in the switch statement in regexec.c
310 * for any node marked SIMPLE.) Note that this is not the same thing as
313 #define SPSTART 0x04 /* Starts with * or + */
314 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
315 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
316 #define RESTART_PASS1 0x20 /* Need to restart sizing pass */
317 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
318 calcuate sizes as UTF-8 */
320 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
322 /* whether trie related optimizations are enabled */
323 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
324 #define TRIE_STUDY_OPT
325 #define FULL_TRIE_STUDY
331 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
332 #define PBITVAL(paren) (1 << ((paren) & 7))
333 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
334 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
335 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
337 #define REQUIRE_UTF8(flagp) STMT_START { \
340 *flagp = RESTART_PASS1|NEED_UTF8; \
345 /* Change from /d into /u rules, and restart the parse if we've already seen
346 * something whose size would increase as a result, by setting *flagp and
347 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
348 * we've changed to /u during the parse. */
349 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
351 if (DEPENDS_SEMANTICS) { \
353 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
354 RExC_uni_semantics = 1; \
355 if (RExC_seen_unfolded_sharp_s) { \
356 *flagp |= RESTART_PASS1; \
357 return restart_retval; \
362 /* This converts the named class defined in regcomp.h to its equivalent class
363 * number defined in handy.h. */
364 #define namedclass_to_classnum(class) ((int) ((class) / 2))
365 #define classnum_to_namedclass(classnum) ((classnum) * 2)
367 #define _invlist_union_complement_2nd(a, b, output) \
368 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
369 #define _invlist_intersection_complement_2nd(a, b, output) \
370 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
372 /* About scan_data_t.
374 During optimisation we recurse through the regexp program performing
375 various inplace (keyhole style) optimisations. In addition study_chunk
376 and scan_commit populate this data structure with information about
377 what strings MUST appear in the pattern. We look for the longest
378 string that must appear at a fixed location, and we look for the
379 longest string that may appear at a floating location. So for instance
384 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
385 strings (because they follow a .* construct). study_chunk will identify
386 both FOO and BAR as being the longest fixed and floating strings respectively.
388 The strings can be composites, for instance
392 will result in a composite fixed substring 'foo'.
394 For each string some basic information is maintained:
397 This is the position the string must appear at, or not before.
398 It also implicitly (when combined with minlenp) tells us how many
399 characters must match before the string we are searching for.
400 Likewise when combined with minlenp and the length of the string it
401 tells us how many characters must appear after the string we have
405 Only used for floating strings. This is the rightmost point that
406 the string can appear at. If set to SSize_t_MAX it indicates that the
407 string can occur infinitely far to the right.
408 For fixed strings, it is equal to min_offset.
411 A pointer to the minimum number of characters of the pattern that the
412 string was found inside. This is important as in the case of positive
413 lookahead or positive lookbehind we can have multiple patterns
418 The minimum length of the pattern overall is 3, the minimum length
419 of the lookahead part is 3, but the minimum length of the part that
420 will actually match is 1. So 'FOO's minimum length is 3, but the
421 minimum length for the F is 1. This is important as the minimum length
422 is used to determine offsets in front of and behind the string being
423 looked for. Since strings can be composites this is the length of the
424 pattern at the time it was committed with a scan_commit. Note that
425 the length is calculated by study_chunk, so that the minimum lengths
426 are not known until the full pattern has been compiled, thus the
427 pointer to the value.
431 In the case of lookbehind the string being searched for can be
432 offset past the start point of the final matching string.
433 If this value was just blithely removed from the min_offset it would
434 invalidate some of the calculations for how many chars must match
435 before or after (as they are derived from min_offset and minlen and
436 the length of the string being searched for).
437 When the final pattern is compiled and the data is moved from the
438 scan_data_t structure into the regexp structure the information
439 about lookbehind is factored in, with the information that would
440 have been lost precalculated in the end_shift field for the
443 The fields pos_min and pos_delta are used to store the minimum offset
444 and the delta to the maximum offset at the current point in the pattern.
448 struct scan_data_substrs {
449 SV *str; /* longest substring found in pattern */
450 SSize_t min_offset; /* earliest point in string it can appear */
451 SSize_t max_offset; /* latest point in string it can appear */
452 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
453 SSize_t lookbehind; /* is the pos of the string modified by LB */
454 I32 flags; /* per substring SF_* and SCF_* flags */
457 typedef struct scan_data_t {
458 /*I32 len_min; unused */
459 /*I32 len_delta; unused */
463 SSize_t last_end; /* min value, <0 unless valid. */
464 SSize_t last_start_min;
465 SSize_t last_start_max;
466 U8 cur_is_floating; /* whether the last_* values should be set as
467 * the next fixed (0) or floating (1)
470 /* [0] is longest fixed substring so far, [1] is longest float so far */
471 struct scan_data_substrs substrs[2];
473 I32 flags; /* common SF_* and SCF_* flags */
475 SSize_t *last_closep;
476 regnode_ssc *start_class;
480 * Forward declarations for pregcomp()'s friends.
483 static const scan_data_t zero_scan_data = {
484 0, 0, NULL, 0, 0, 0, 0,
486 { NULL, 0, 0, 0, 0, 0 },
487 { NULL, 0, 0, 0, 0, 0 },
494 #define SF_BEFORE_SEOL 0x0001
495 #define SF_BEFORE_MEOL 0x0002
496 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
498 #define SF_IS_INF 0x0040
499 #define SF_HAS_PAR 0x0080
500 #define SF_IN_PAR 0x0100
501 #define SF_HAS_EVAL 0x0200
504 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
505 * longest substring in the pattern. When it is not set the optimiser keeps
506 * track of position, but does not keep track of the actual strings seen,
508 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
511 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
512 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
513 * turned off because of the alternation (BRANCH). */
514 #define SCF_DO_SUBSTR 0x0400
516 #define SCF_DO_STCLASS_AND 0x0800
517 #define SCF_DO_STCLASS_OR 0x1000
518 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
519 #define SCF_WHILEM_VISITED_POS 0x2000
521 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
522 #define SCF_SEEN_ACCEPT 0x8000
523 #define SCF_TRIE_DOING_RESTUDY 0x10000
524 #define SCF_IN_DEFINE 0x20000
529 #define UTF cBOOL(RExC_utf8)
531 /* The enums for all these are ordered so things work out correctly */
532 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
533 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
534 == REGEX_DEPENDS_CHARSET)
535 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
536 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
537 >= REGEX_UNICODE_CHARSET)
538 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
539 == REGEX_ASCII_RESTRICTED_CHARSET)
540 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
541 >= REGEX_ASCII_RESTRICTED_CHARSET)
542 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
543 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
545 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
547 /* For programs that want to be strictly Unicode compatible by dying if any
548 * attempt is made to match a non-Unicode code point against a Unicode
550 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
552 #define OOB_NAMEDCLASS -1
554 /* There is no code point that is out-of-bounds, so this is problematic. But
555 * its only current use is to initialize a variable that is always set before
557 #define OOB_UNICODE 0xDEADBEEF
559 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
562 /* length of regex to show in messages that don't mark a position within */
563 #define RegexLengthToShowInErrorMessages 127
566 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
567 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
568 * op/pragma/warn/regcomp.
570 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
571 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
573 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
574 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
576 /* The code in this file in places uses one level of recursion with parsing
577 * rebased to an alternate string constructed by us in memory. This can take
578 * the form of something that is completely different from the input, or
579 * something that uses the input as part of the alternate. In the first case,
580 * there should be no possibility of an error, as we are in complete control of
581 * the alternate string. But in the second case we don't control the input
582 * portion, so there may be errors in that. Here's an example:
584 * is handled specially because \x{df} folds to a sequence of more than one
585 * character, 'ss'. What is done is to create and parse an alternate string,
586 * which looks like this:
587 * /(?:\x{DF}|[abc\x{DF}def])/ui
588 * where it uses the input unchanged in the middle of something it constructs,
589 * which is a branch for the DF outside the character class, and clustering
590 * parens around the whole thing. (It knows enough to skip the DF inside the
591 * class while in this substitute parse.) 'abc' and 'def' may have errors that
592 * need to be reported. The general situation looks like this:
595 * Input: ----------------------------------------------------
596 * Constructed: ---------------------------------------------------
599 * The input string sI..eI is the input pattern. The string sC..EC is the
600 * constructed substitute parse string. The portions sC..tC and eC..EC are
601 * constructed by us. The portion tC..eC is an exact duplicate of the input
602 * pattern tI..eI. In the diagram, these are vertically aligned. Suppose that
603 * while parsing, we find an error at xC. We want to display a message showing
604 * the real input string. Thus we need to find the point xI in it which
605 * corresponds to xC. xC >= tC, since the portion of the string sC..tC has
606 * been constructed by us, and so shouldn't have errors. We get:
608 * xI = sI + (tI - sI) + (xC - tC)
610 * and, the offset into sI is:
612 * (xI - sI) = (tI - sI) + (xC - tC)
614 * When the substitute is constructed, we save (tI -sI) as RExC_precomp_adj,
615 * and we save tC as RExC_adjusted_start.
617 * During normal processing of the input pattern, everything points to that,
618 * with RExC_precomp_adj set to 0, and RExC_adjusted_start set to sI.
621 #define tI_sI RExC_precomp_adj
622 #define tC RExC_adjusted_start
623 #define sC RExC_precomp
624 #define xI_offset(xC) ((IV) (tI_sI + (xC - tC)))
625 #define xI(xC) (sC + xI_offset(xC))
626 #define eC RExC_precomp_end
628 #define REPORT_LOCATION_ARGS(xC) \
630 (xI(xC) > eC) /* Don't run off end */ \
631 ? eC - sC /* Length before the <--HERE */ \
632 : ( __ASSERT_(xI_offset(xC) >= 0) xI_offset(xC) ), \
633 sC), /* The input pattern printed up to the <--HERE */ \
635 (xI(xC) > eC) ? 0 : eC - xI(xC), /* Length after <--HERE */ \
636 (xI(xC) > eC) ? eC : xI(xC)) /* pattern after <--HERE */
638 /* Used to point after bad bytes for an error message, but avoid skipping
639 * past a nul byte. */
640 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
643 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
644 * arg. Show regex, up to a maximum length. If it's too long, chop and add
647 #define _FAIL(code) STMT_START { \
648 const char *ellipses = ""; \
649 IV len = RExC_precomp_end - RExC_precomp; \
652 SAVEFREESV(RExC_rx_sv); \
653 if (len > RegexLengthToShowInErrorMessages) { \
654 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
655 len = RegexLengthToShowInErrorMessages - 10; \
661 #define FAIL(msg) _FAIL( \
662 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
663 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
665 #define FAIL2(msg,arg) _FAIL( \
666 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
667 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
670 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
672 #define Simple_vFAIL(m) STMT_START { \
673 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
674 m, REPORT_LOCATION_ARGS(RExC_parse)); \
678 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
680 #define vFAIL(m) STMT_START { \
682 SAVEFREESV(RExC_rx_sv); \
687 * Like Simple_vFAIL(), but accepts two arguments.
689 #define Simple_vFAIL2(m,a1) STMT_START { \
690 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
691 REPORT_LOCATION_ARGS(RExC_parse)); \
695 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
697 #define vFAIL2(m,a1) STMT_START { \
699 SAVEFREESV(RExC_rx_sv); \
700 Simple_vFAIL2(m, a1); \
705 * Like Simple_vFAIL(), but accepts three arguments.
707 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
708 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
709 REPORT_LOCATION_ARGS(RExC_parse)); \
713 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
715 #define vFAIL3(m,a1,a2) STMT_START { \
717 SAVEFREESV(RExC_rx_sv); \
718 Simple_vFAIL3(m, a1, a2); \
722 * Like Simple_vFAIL(), but accepts four arguments.
724 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
725 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
726 REPORT_LOCATION_ARGS(RExC_parse)); \
729 #define vFAIL4(m,a1,a2,a3) STMT_START { \
731 SAVEFREESV(RExC_rx_sv); \
732 Simple_vFAIL4(m, a1, a2, a3); \
735 /* A specialized version of vFAIL2 that works with UTF8f */
736 #define vFAIL2utf8f(m, a1) STMT_START { \
738 SAVEFREESV(RExC_rx_sv); \
739 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
740 REPORT_LOCATION_ARGS(RExC_parse)); \
743 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
745 SAVEFREESV(RExC_rx_sv); \
746 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
747 REPORT_LOCATION_ARGS(RExC_parse)); \
750 /* These have asserts in them because of [perl #122671] Many warnings in
751 * regcomp.c can occur twice. If they get output in pass1 and later in that
752 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
753 * would get output again. So they should be output in pass2, and these
754 * asserts make sure new warnings follow that paradigm. */
756 /* m is not necessarily a "literal string", in this macro */
757 #define reg_warn_non_literal_string(loc, m) STMT_START { \
758 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
759 "%s" REPORT_LOCATION, \
760 m, REPORT_LOCATION_ARGS(loc)); \
763 #define ckWARNreg(loc,m) STMT_START { \
764 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
766 REPORT_LOCATION_ARGS(loc)); \
769 #define vWARN(loc, m) STMT_START { \
770 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
772 REPORT_LOCATION_ARGS(loc)); \
775 #define vWARN_dep(loc, m) STMT_START { \
776 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
778 REPORT_LOCATION_ARGS(loc)); \
781 #define ckWARNdep(loc,m) STMT_START { \
782 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
784 REPORT_LOCATION_ARGS(loc)); \
787 #define ckWARNregdep(loc,m) STMT_START { \
788 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
791 REPORT_LOCATION_ARGS(loc)); \
794 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
795 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
797 a1, REPORT_LOCATION_ARGS(loc)); \
800 #define ckWARN2reg(loc, m, a1) STMT_START { \
801 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
803 a1, REPORT_LOCATION_ARGS(loc)); \
806 #define vWARN3(loc, m, a1, a2) STMT_START { \
807 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
809 a1, a2, REPORT_LOCATION_ARGS(loc)); \
812 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
813 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
816 REPORT_LOCATION_ARGS(loc)); \
819 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
820 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
823 REPORT_LOCATION_ARGS(loc)); \
826 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
827 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
830 REPORT_LOCATION_ARGS(loc)); \
833 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
834 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
837 REPORT_LOCATION_ARGS(loc)); \
840 /* Macros for recording node offsets. 20001227 mjd@plover.com
841 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
842 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
843 * Element 0 holds the number n.
844 * Position is 1 indexed.
846 #ifndef RE_TRACK_PATTERN_OFFSETS
847 #define Set_Node_Offset_To_R(node,byte)
848 #define Set_Node_Offset(node,byte)
849 #define Set_Cur_Node_Offset
850 #define Set_Node_Length_To_R(node,len)
851 #define Set_Node_Length(node,len)
852 #define Set_Node_Cur_Length(node,start)
853 #define Node_Offset(n)
854 #define Node_Length(n)
855 #define Set_Node_Offset_Length(node,offset,len)
856 #define ProgLen(ri) ri->u.proglen
857 #define SetProgLen(ri,x) ri->u.proglen = x
859 #define ProgLen(ri) ri->u.offsets[0]
860 #define SetProgLen(ri,x) ri->u.offsets[0] = x
861 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
863 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
864 __LINE__, (int)(node), (int)(byte))); \
866 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
869 RExC_offsets[2*(node)-1] = (byte); \
874 #define Set_Node_Offset(node,byte) \
875 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
876 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
878 #define Set_Node_Length_To_R(node,len) STMT_START { \
880 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
881 __LINE__, (int)(node), (int)(len))); \
883 Perl_croak(aTHX_ "value of node is %d in Length macro", \
886 RExC_offsets[2*(node)] = (len); \
891 #define Set_Node_Length(node,len) \
892 Set_Node_Length_To_R((node)-RExC_emit_start, len)
893 #define Set_Node_Cur_Length(node, start) \
894 Set_Node_Length(node, RExC_parse - start)
896 /* Get offsets and lengths */
897 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
898 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
900 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
901 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
902 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
906 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
907 #define EXPERIMENTAL_INPLACESCAN
908 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
912 Perl_re_printf(pTHX_ const char *fmt, ...)
916 PerlIO *f= Perl_debug_log;
917 PERL_ARGS_ASSERT_RE_PRINTF;
919 result = PerlIO_vprintf(f, fmt, ap);
925 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
929 PerlIO *f= Perl_debug_log;
930 PERL_ARGS_ASSERT_RE_INDENTF;
932 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
933 result = PerlIO_vprintf(f, fmt, ap);
937 #endif /* DEBUGGING */
939 #define DEBUG_RExC_seen() \
940 DEBUG_OPTIMISE_MORE_r({ \
941 Perl_re_printf( aTHX_ "RExC_seen: "); \
943 if (RExC_seen & REG_ZERO_LEN_SEEN) \
944 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
946 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
947 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
949 if (RExC_seen & REG_GPOS_SEEN) \
950 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
952 if (RExC_seen & REG_RECURSE_SEEN) \
953 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
955 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
956 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
958 if (RExC_seen & REG_VERBARG_SEEN) \
959 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
961 if (RExC_seen & REG_CUTGROUP_SEEN) \
962 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
964 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
965 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
967 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
968 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
970 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
971 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
973 Perl_re_printf( aTHX_ "\n"); \
976 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
977 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
982 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
983 const char *close_str)
988 Perl_re_printf( aTHX_ "%s", open_str);
989 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
990 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
991 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
992 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
993 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
994 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
995 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
996 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
997 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
998 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
999 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1000 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1001 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1002 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1003 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1004 Perl_re_printf( aTHX_ "%s", close_str);
1009 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1010 U32 depth, int is_inf)
1012 GET_RE_DEBUG_FLAGS_DECL;
1014 DEBUG_OPTIMISE_MORE_r({
1017 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1021 (IV)data->pos_delta,
1025 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1027 Perl_re_printf( aTHX_
1028 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1030 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1031 is_inf ? "INF " : ""
1034 if (data->last_found) {
1036 Perl_re_printf(aTHX_
1037 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1038 SvPVX_const(data->last_found),
1040 (IV)data->last_start_min,
1041 (IV)data->last_start_max
1044 for (i = 0; i < 2; i++) {
1045 Perl_re_printf(aTHX_
1046 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1047 data->cur_is_floating == i ? "*" : "",
1048 i ? "Float" : "Fixed",
1049 SvPVX_const(data->substrs[i].str),
1050 (IV)data->substrs[i].min_offset,
1051 (IV)data->substrs[i].max_offset
1053 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1057 Perl_re_printf( aTHX_ "\n");
1063 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1064 regnode *scan, U32 depth, U32 flags)
1066 GET_RE_DEBUG_FLAGS_DECL;
1073 Next = regnext(scan);
1074 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1075 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1078 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1079 Next ? (REG_NODE_NUM(Next)) : 0 );
1080 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1081 Perl_re_printf( aTHX_ "\n");
1086 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1087 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1089 # define DEBUG_PEEP(str, scan, depth, flags) \
1090 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1093 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1094 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1098 /* =========================================================
1099 * BEGIN edit_distance stuff.
1101 * This calculates how many single character changes of any type are needed to
1102 * transform a string into another one. It is taken from version 3.1 of
1104 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1107 /* Our unsorted dictionary linked list. */
1108 /* Note we use UVs, not chars. */
1113 struct dictionary* next;
1115 typedef struct dictionary item;
1118 PERL_STATIC_INLINE item*
1119 push(UV key,item* curr)
1122 Newx(head, 1, item);
1130 PERL_STATIC_INLINE item*
1131 find(item* head, UV key)
1133 item* iterator = head;
1135 if (iterator->key == key){
1138 iterator = iterator->next;
1144 PERL_STATIC_INLINE item*
1145 uniquePush(item* head,UV key)
1147 item* iterator = head;
1150 if (iterator->key == key) {
1153 iterator = iterator->next;
1156 return push(key,head);
1159 PERL_STATIC_INLINE void
1160 dict_free(item* head)
1162 item* iterator = head;
1165 item* temp = iterator;
1166 iterator = iterator->next;
1173 /* End of Dictionary Stuff */
1175 /* All calculations/work are done here */
1177 S_edit_distance(const UV* src,
1179 const STRLEN x, /* length of src[] */
1180 const STRLEN y, /* length of tgt[] */
1181 const SSize_t maxDistance
1185 UV swapCount,swapScore,targetCharCount,i,j;
1187 UV score_ceil = x + y;
1189 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1191 /* intialize matrix start values */
1192 Newx(scores, ( (x + 2) * (y + 2)), UV);
1193 scores[0] = score_ceil;
1194 scores[1 * (y + 2) + 0] = score_ceil;
1195 scores[0 * (y + 2) + 1] = score_ceil;
1196 scores[1 * (y + 2) + 1] = 0;
1197 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1202 for (i=1;i<=x;i++) {
1204 head = uniquePush(head,src[i]);
1205 scores[(i+1) * (y + 2) + 1] = i;
1206 scores[(i+1) * (y + 2) + 0] = score_ceil;
1209 for (j=1;j<=y;j++) {
1212 head = uniquePush(head,tgt[j]);
1213 scores[1 * (y + 2) + (j + 1)] = j;
1214 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1217 targetCharCount = find(head,tgt[j-1])->value;
1218 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1220 if (src[i-1] != tgt[j-1]){
1221 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));
1225 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1229 find(head,src[i-1])->value = i;
1233 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1236 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1240 /* END of edit_distance() stuff
1241 * ========================================================= */
1243 /* is c a control character for which we have a mnemonic? */
1244 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1247 S_cntrl_to_mnemonic(const U8 c)
1249 /* Returns the mnemonic string that represents character 'c', if one
1250 * exists; NULL otherwise. The only ones that exist for the purposes of
1251 * this routine are a few control characters */
1254 case '\a': return "\\a";
1255 case '\b': return "\\b";
1256 case ESC_NATIVE: return "\\e";
1257 case '\f': return "\\f";
1258 case '\n': return "\\n";
1259 case '\r': return "\\r";
1260 case '\t': return "\\t";
1266 /* Mark that we cannot extend a found fixed substring at this point.
1267 Update the longest found anchored substring or the longest found
1268 floating substrings if needed. */
1271 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1272 SSize_t *minlenp, int is_inf)
1274 const STRLEN l = CHR_SVLEN(data->last_found);
1275 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1276 const STRLEN old_l = CHR_SVLEN(longest_sv);
1277 GET_RE_DEBUG_FLAGS_DECL;
1279 PERL_ARGS_ASSERT_SCAN_COMMIT;
1281 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1282 const U8 i = data->cur_is_floating;
1283 SvSetMagicSV(longest_sv, data->last_found);
1284 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1287 data->substrs[0].max_offset = data->substrs[0].min_offset;
1289 data->substrs[1].max_offset = (l
1290 ? data->last_start_max
1291 : (data->pos_delta > SSize_t_MAX - data->pos_min
1293 : data->pos_min + data->pos_delta));
1295 || (STRLEN)data->substrs[1].max_offset > (STRLEN)SSize_t_MAX)
1296 data->substrs[1].max_offset = SSize_t_MAX;
1299 if (data->flags & SF_BEFORE_EOL)
1300 data->substrs[i].flags |= (data->flags & SF_BEFORE_EOL);
1302 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1303 data->substrs[i].minlenp = minlenp;
1304 data->substrs[i].lookbehind = 0;
1307 SvCUR_set(data->last_found, 0);
1309 SV * const sv = data->last_found;
1310 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1311 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1316 data->last_end = -1;
1317 data->flags &= ~SF_BEFORE_EOL;
1318 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1321 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1322 * list that describes which code points it matches */
1325 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1327 /* Set the SSC 'ssc' to match an empty string or any code point */
1329 PERL_ARGS_ASSERT_SSC_ANYTHING;
1331 assert(is_ANYOF_SYNTHETIC(ssc));
1333 /* mortalize so won't leak */
1334 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1335 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1339 S_ssc_is_anything(const regnode_ssc *ssc)
1341 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1342 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1343 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1344 * in any way, so there's no point in using it */
1349 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1351 assert(is_ANYOF_SYNTHETIC(ssc));
1353 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1357 /* See if the list consists solely of the range 0 - Infinity */
1358 invlist_iterinit(ssc->invlist);
1359 ret = invlist_iternext(ssc->invlist, &start, &end)
1363 invlist_iterfinish(ssc->invlist);
1369 /* If e.g., both \w and \W are set, matches everything */
1370 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1372 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1373 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1383 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1385 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1386 * string, any code point, or any posix class under locale */
1388 PERL_ARGS_ASSERT_SSC_INIT;
1390 Zero(ssc, 1, regnode_ssc);
1391 set_ANYOF_SYNTHETIC(ssc);
1392 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1395 /* If any portion of the regex is to operate under locale rules that aren't
1396 * fully known at compile time, initialization includes it. The reason
1397 * this isn't done for all regexes is that the optimizer was written under
1398 * the assumption that locale was all-or-nothing. Given the complexity and
1399 * lack of documentation in the optimizer, and that there are inadequate
1400 * test cases for locale, many parts of it may not work properly, it is
1401 * safest to avoid locale unless necessary. */
1402 if (RExC_contains_locale) {
1403 ANYOF_POSIXL_SETALL(ssc);
1406 ANYOF_POSIXL_ZERO(ssc);
1411 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1412 const regnode_ssc *ssc)
1414 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1415 * to the list of code points matched, and locale posix classes; hence does
1416 * not check its flags) */
1421 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1423 assert(is_ANYOF_SYNTHETIC(ssc));
1425 invlist_iterinit(ssc->invlist);
1426 ret = invlist_iternext(ssc->invlist, &start, &end)
1430 invlist_iterfinish(ssc->invlist);
1436 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1444 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1445 const regnode_charclass* const node)
1447 /* Returns a mortal inversion list defining which code points are matched
1448 * by 'node', which is of type ANYOF. Handles complementing the result if
1449 * appropriate. If some code points aren't knowable at this time, the
1450 * returned list must, and will, contain every code point that is a
1454 SV* only_utf8_locale_invlist = NULL;
1456 const U32 n = ARG(node);
1457 bool new_node_has_latin1 = FALSE;
1459 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1461 /* Look at the data structure created by S_set_ANYOF_arg() */
1462 if (n != ANYOF_ONLY_HAS_BITMAP) {
1463 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1464 AV * const av = MUTABLE_AV(SvRV(rv));
1465 SV **const ary = AvARRAY(av);
1466 assert(RExC_rxi->data->what[n] == 's');
1468 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1469 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1471 else if (ary[0] && ary[0] != &PL_sv_undef) {
1473 /* Here, no compile-time swash, and there are things that won't be
1474 * known until runtime -- we have to assume it could be anything */
1475 invlist = sv_2mortal(_new_invlist(1));
1476 return _add_range_to_invlist(invlist, 0, UV_MAX);
1478 else if (ary[3] && ary[3] != &PL_sv_undef) {
1480 /* Here no compile-time swash, and no run-time only data. Use the
1481 * node's inversion list */
1482 invlist = sv_2mortal(invlist_clone(ary[3]));
1485 /* Get the code points valid only under UTF-8 locales */
1486 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1487 && ary[2] && ary[2] != &PL_sv_undef)
1489 only_utf8_locale_invlist = ary[2];
1494 invlist = sv_2mortal(_new_invlist(0));
1497 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1498 * code points, and an inversion list for the others, but if there are code
1499 * points that should match only conditionally on the target string being
1500 * UTF-8, those are placed in the inversion list, and not the bitmap.
1501 * Since there are circumstances under which they could match, they are
1502 * included in the SSC. But if the ANYOF node is to be inverted, we have
1503 * to exclude them here, so that when we invert below, the end result
1504 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1505 * have to do this here before we add the unconditionally matched code
1507 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1508 _invlist_intersection_complement_2nd(invlist,
1513 /* Add in the points from the bit map */
1514 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1515 if (ANYOF_BITMAP_TEST(node, i)) {
1516 unsigned int start = i++;
1518 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1521 invlist = _add_range_to_invlist(invlist, start, i-1);
1522 new_node_has_latin1 = TRUE;
1526 /* If this can match all upper Latin1 code points, have to add them
1527 * as well. But don't add them if inverting, as when that gets done below,
1528 * it would exclude all these characters, including the ones it shouldn't
1529 * that were added just above */
1530 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
1531 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1533 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1536 /* Similarly for these */
1537 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1538 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1541 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1542 _invlist_invert(invlist);
1544 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
1546 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1547 * locale. We can skip this if there are no 0-255 at all. */
1548 _invlist_union(invlist, PL_Latin1, &invlist);
1551 /* Similarly add the UTF-8 locale possible matches. These have to be
1552 * deferred until after the non-UTF-8 locale ones are taken care of just
1553 * above, or it leads to wrong results under ANYOF_INVERT */
1554 if (only_utf8_locale_invlist) {
1555 _invlist_union_maybe_complement_2nd(invlist,
1556 only_utf8_locale_invlist,
1557 ANYOF_FLAGS(node) & ANYOF_INVERT,
1564 /* These two functions currently do the exact same thing */
1565 #define ssc_init_zero ssc_init
1567 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1568 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1570 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1571 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1572 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1575 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1576 const regnode_charclass *and_with)
1578 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1579 * another SSC or a regular ANYOF class. Can create false positives. */
1584 PERL_ARGS_ASSERT_SSC_AND;
1586 assert(is_ANYOF_SYNTHETIC(ssc));
1588 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1589 * the code point inversion list and just the relevant flags */
1590 if (is_ANYOF_SYNTHETIC(and_with)) {
1591 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1592 anded_flags = ANYOF_FLAGS(and_with);
1594 /* XXX This is a kludge around what appears to be deficiencies in the
1595 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1596 * there are paths through the optimizer where it doesn't get weeded
1597 * out when it should. And if we don't make some extra provision for
1598 * it like the code just below, it doesn't get added when it should.
1599 * This solution is to add it only when AND'ing, which is here, and
1600 * only when what is being AND'ed is the pristine, original node
1601 * matching anything. Thus it is like adding it to ssc_anything() but
1602 * only when the result is to be AND'ed. Probably the same solution
1603 * could be adopted for the same problem we have with /l matching,
1604 * which is solved differently in S_ssc_init(), and that would lead to
1605 * fewer false positives than that solution has. But if this solution
1606 * creates bugs, the consequences are only that a warning isn't raised
1607 * that should be; while the consequences for having /l bugs is
1608 * incorrect matches */
1609 if (ssc_is_anything((regnode_ssc *)and_with)) {
1610 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1614 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1615 if (OP(and_with) == ANYOFD) {
1616 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1619 anded_flags = ANYOF_FLAGS(and_with)
1620 &( ANYOF_COMMON_FLAGS
1621 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1622 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1623 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1625 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1630 ANYOF_FLAGS(ssc) &= anded_flags;
1632 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1633 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1634 * 'and_with' may be inverted. When not inverted, we have the situation of
1636 * (C1 | P1) & (C2 | P2)
1637 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1638 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1639 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1640 * <= ((C1 & C2) | P1 | P2)
1641 * Alternatively, the last few steps could be:
1642 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1643 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1644 * <= (C1 | C2 | (P1 & P2))
1645 * We favor the second approach if either P1 or P2 is non-empty. This is
1646 * because these components are a barrier to doing optimizations, as what
1647 * they match cannot be known until the moment of matching as they are
1648 * dependent on the current locale, 'AND"ing them likely will reduce or
1650 * But we can do better if we know that C1,P1 are in their initial state (a
1651 * frequent occurrence), each matching everything:
1652 * (<everything>) & (C2 | P2) = C2 | P2
1653 * Similarly, if C2,P2 are in their initial state (again a frequent
1654 * occurrence), the result is a no-op
1655 * (C1 | P1) & (<everything>) = C1 | P1
1658 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1659 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1660 * <= (C1 & ~C2) | (P1 & ~P2)
1663 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1664 && ! is_ANYOF_SYNTHETIC(and_with))
1668 ssc_intersection(ssc,
1670 FALSE /* Has already been inverted */
1673 /* If either P1 or P2 is empty, the intersection will be also; can skip
1675 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1676 ANYOF_POSIXL_ZERO(ssc);
1678 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1680 /* Note that the Posix class component P from 'and_with' actually
1682 * P = Pa | Pb | ... | Pn
1683 * where each component is one posix class, such as in [\w\s].
1685 * ~P = ~(Pa | Pb | ... | Pn)
1686 * = ~Pa & ~Pb & ... & ~Pn
1687 * <= ~Pa | ~Pb | ... | ~Pn
1688 * The last is something we can easily calculate, but unfortunately
1689 * is likely to have many false positives. We could do better
1690 * in some (but certainly not all) instances if two classes in
1691 * P have known relationships. For example
1692 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1694 * :lower: & :print: = :lower:
1695 * And similarly for classes that must be disjoint. For example,
1696 * since \s and \w can have no elements in common based on rules in
1697 * the POSIX standard,
1698 * \w & ^\S = nothing
1699 * Unfortunately, some vendor locales do not meet the Posix
1700 * standard, in particular almost everything by Microsoft.
1701 * The loop below just changes e.g., \w into \W and vice versa */
1703 regnode_charclass_posixl temp;
1704 int add = 1; /* To calculate the index of the complement */
1706 Zero(&temp, 1, regnode_charclass_posixl);
1707 ANYOF_POSIXL_ZERO(&temp);
1708 for (i = 0; i < ANYOF_MAX; i++) {
1710 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1711 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1713 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1714 ANYOF_POSIXL_SET(&temp, i + add);
1716 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1718 ANYOF_POSIXL_AND(&temp, ssc);
1720 } /* else ssc already has no posixes */
1721 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1722 in its initial state */
1723 else if (! is_ANYOF_SYNTHETIC(and_with)
1724 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1726 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1727 * copy it over 'ssc' */
1728 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1729 if (is_ANYOF_SYNTHETIC(and_with)) {
1730 StructCopy(and_with, ssc, regnode_ssc);
1733 ssc->invlist = anded_cp_list;
1734 ANYOF_POSIXL_ZERO(ssc);
1735 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1736 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1740 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1741 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1743 /* One or the other of P1, P2 is non-empty. */
1744 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1745 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1747 ssc_union(ssc, anded_cp_list, FALSE);
1749 else { /* P1 = P2 = empty */
1750 ssc_intersection(ssc, anded_cp_list, FALSE);
1756 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1757 const regnode_charclass *or_with)
1759 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1760 * another SSC or a regular ANYOF class. Can create false positives if
1761 * 'or_with' is to be inverted. */
1766 PERL_ARGS_ASSERT_SSC_OR;
1768 assert(is_ANYOF_SYNTHETIC(ssc));
1770 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1771 * the code point inversion list and just the relevant flags */
1772 if (is_ANYOF_SYNTHETIC(or_with)) {
1773 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1774 ored_flags = ANYOF_FLAGS(or_with);
1777 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1778 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1779 if (OP(or_with) != ANYOFD) {
1781 |= ANYOF_FLAGS(or_with)
1782 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1783 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1784 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1786 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1791 ANYOF_FLAGS(ssc) |= ored_flags;
1793 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1794 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1795 * 'or_with' may be inverted. When not inverted, we have the simple
1796 * situation of computing:
1797 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1798 * If P1|P2 yields a situation with both a class and its complement are
1799 * set, like having both \w and \W, this matches all code points, and we
1800 * can delete these from the P component of the ssc going forward. XXX We
1801 * might be able to delete all the P components, but I (khw) am not certain
1802 * about this, and it is better to be safe.
1805 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1806 * <= (C1 | P1) | ~C2
1807 * <= (C1 | ~C2) | P1
1808 * (which results in actually simpler code than the non-inverted case)
1811 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1812 && ! is_ANYOF_SYNTHETIC(or_with))
1814 /* We ignore P2, leaving P1 going forward */
1815 } /* else Not inverted */
1816 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1817 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1818 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1820 for (i = 0; i < ANYOF_MAX; i += 2) {
1821 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1823 ssc_match_all_cp(ssc);
1824 ANYOF_POSIXL_CLEAR(ssc, i);
1825 ANYOF_POSIXL_CLEAR(ssc, i+1);
1833 FALSE /* Already has been inverted */
1837 PERL_STATIC_INLINE void
1838 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1840 PERL_ARGS_ASSERT_SSC_UNION;
1842 assert(is_ANYOF_SYNTHETIC(ssc));
1844 _invlist_union_maybe_complement_2nd(ssc->invlist,
1850 PERL_STATIC_INLINE void
1851 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1853 const bool invert2nd)
1855 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1857 assert(is_ANYOF_SYNTHETIC(ssc));
1859 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1865 PERL_STATIC_INLINE void
1866 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1868 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1870 assert(is_ANYOF_SYNTHETIC(ssc));
1872 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1875 PERL_STATIC_INLINE void
1876 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1878 /* AND just the single code point 'cp' into the SSC 'ssc' */
1880 SV* cp_list = _new_invlist(2);
1882 PERL_ARGS_ASSERT_SSC_CP_AND;
1884 assert(is_ANYOF_SYNTHETIC(ssc));
1886 cp_list = add_cp_to_invlist(cp_list, cp);
1887 ssc_intersection(ssc, cp_list,
1888 FALSE /* Not inverted */
1890 SvREFCNT_dec_NN(cp_list);
1893 PERL_STATIC_INLINE void
1894 S_ssc_clear_locale(regnode_ssc *ssc)
1896 /* Set the SSC 'ssc' to not match any locale things */
1897 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1899 assert(is_ANYOF_SYNTHETIC(ssc));
1901 ANYOF_POSIXL_ZERO(ssc);
1902 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1905 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1908 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1910 /* The synthetic start class is used to hopefully quickly winnow down
1911 * places where a pattern could start a match in the target string. If it
1912 * doesn't really narrow things down that much, there isn't much point to
1913 * having the overhead of using it. This function uses some very crude
1914 * heuristics to decide if to use the ssc or not.
1916 * It returns TRUE if 'ssc' rules out more than half what it considers to
1917 * be the "likely" possible matches, but of course it doesn't know what the
1918 * actual things being matched are going to be; these are only guesses
1920 * For /l matches, it assumes that the only likely matches are going to be
1921 * in the 0-255 range, uniformly distributed, so half of that is 127
1922 * For /a and /d matches, it assumes that the likely matches will be just
1923 * the ASCII range, so half of that is 63
1924 * For /u and there isn't anything matching above the Latin1 range, it
1925 * assumes that that is the only range likely to be matched, and uses
1926 * half that as the cut-off: 127. If anything matches above Latin1,
1927 * it assumes that all of Unicode could match (uniformly), except for
1928 * non-Unicode code points and things in the General Category "Other"
1929 * (unassigned, private use, surrogates, controls and formats). This
1930 * is a much large number. */
1932 U32 count = 0; /* Running total of number of code points matched by
1934 UV start, end; /* Start and end points of current range in inversion
1936 const U32 max_code_points = (LOC)
1938 : (( ! UNI_SEMANTICS
1939 || invlist_highest(ssc->invlist) < 256)
1942 const U32 max_match = max_code_points / 2;
1944 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1946 invlist_iterinit(ssc->invlist);
1947 while (invlist_iternext(ssc->invlist, &start, &end)) {
1948 if (start >= max_code_points) {
1951 end = MIN(end, max_code_points - 1);
1952 count += end - start + 1;
1953 if (count >= max_match) {
1954 invlist_iterfinish(ssc->invlist);
1964 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1966 /* The inversion list in the SSC is marked mortal; now we need a more
1967 * permanent copy, which is stored the same way that is done in a regular
1968 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1971 SV* invlist = invlist_clone(ssc->invlist);
1973 PERL_ARGS_ASSERT_SSC_FINALIZE;
1975 assert(is_ANYOF_SYNTHETIC(ssc));
1977 /* The code in this file assumes that all but these flags aren't relevant
1978 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1979 * by the time we reach here */
1980 assert(! (ANYOF_FLAGS(ssc)
1981 & ~( ANYOF_COMMON_FLAGS
1982 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1983 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
1985 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1987 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1988 NULL, NULL, NULL, FALSE);
1990 /* Make sure is clone-safe */
1991 ssc->invlist = NULL;
1993 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1994 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
1997 if (RExC_contains_locale) {
2001 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2004 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2005 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2006 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2007 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2008 ? (TRIE_LIST_CUR( idx ) - 1) \
2014 dump_trie(trie,widecharmap,revcharmap)
2015 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2016 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2018 These routines dump out a trie in a somewhat readable format.
2019 The _interim_ variants are used for debugging the interim
2020 tables that are used to generate the final compressed
2021 representation which is what dump_trie expects.
2023 Part of the reason for their existence is to provide a form
2024 of documentation as to how the different representations function.
2029 Dumps the final compressed table form of the trie to Perl_debug_log.
2030 Used for debugging make_trie().
2034 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2035 AV *revcharmap, U32 depth)
2038 SV *sv=sv_newmortal();
2039 int colwidth= widecharmap ? 6 : 4;
2041 GET_RE_DEBUG_FLAGS_DECL;
2043 PERL_ARGS_ASSERT_DUMP_TRIE;
2045 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2046 depth+1, "Match","Base","Ofs" );
2048 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2049 SV ** const tmp = av_fetch( revcharmap, state, 0);
2051 Perl_re_printf( aTHX_ "%*s",
2053 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2054 PL_colors[0], PL_colors[1],
2055 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2056 PERL_PV_ESCAPE_FIRSTCHAR
2061 Perl_re_printf( aTHX_ "\n");
2062 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2064 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2065 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2066 Perl_re_printf( aTHX_ "\n");
2068 for( state = 1 ; state < trie->statecount ; state++ ) {
2069 const U32 base = trie->states[ state ].trans.base;
2071 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2073 if ( trie->states[ state ].wordnum ) {
2074 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2076 Perl_re_printf( aTHX_ "%6s", "" );
2079 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2084 while( ( base + ofs < trie->uniquecharcount ) ||
2085 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2086 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2090 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2092 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2093 if ( ( base + ofs >= trie->uniquecharcount )
2094 && ( base + ofs - trie->uniquecharcount
2096 && trie->trans[ base + ofs
2097 - trie->uniquecharcount ].check == state )
2099 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2100 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2103 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
2107 Perl_re_printf( aTHX_ "]");
2110 Perl_re_printf( aTHX_ "\n" );
2112 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2114 for (word=1; word <= trie->wordcount; word++) {
2115 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2116 (int)word, (int)(trie->wordinfo[word].prev),
2117 (int)(trie->wordinfo[word].len));
2119 Perl_re_printf( aTHX_ "\n" );
2122 Dumps a fully constructed but uncompressed trie in list form.
2123 List tries normally only are used for construction when the number of
2124 possible chars (trie->uniquecharcount) is very high.
2125 Used for debugging make_trie().
2128 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2129 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2133 SV *sv=sv_newmortal();
2134 int colwidth= widecharmap ? 6 : 4;
2135 GET_RE_DEBUG_FLAGS_DECL;
2137 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2139 /* print out the table precompression. */
2140 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2142 Perl_re_indentf( aTHX_ "%s",
2143 depth+1, "------:-----+-----------------\n" );
2145 for( state=1 ; state < next_alloc ; state ++ ) {
2148 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2149 depth+1, (UV)state );
2150 if ( ! trie->states[ state ].wordnum ) {
2151 Perl_re_printf( aTHX_ "%5s| ","");
2153 Perl_re_printf( aTHX_ "W%4x| ",
2154 trie->states[ state ].wordnum
2157 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2158 SV ** const tmp = av_fetch( revcharmap,
2159 TRIE_LIST_ITEM(state,charid).forid, 0);
2161 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2163 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2165 PL_colors[0], PL_colors[1],
2166 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2167 | PERL_PV_ESCAPE_FIRSTCHAR
2169 TRIE_LIST_ITEM(state,charid).forid,
2170 (UV)TRIE_LIST_ITEM(state,charid).newstate
2173 Perl_re_printf( aTHX_ "\n%*s| ",
2174 (int)((depth * 2) + 14), "");
2177 Perl_re_printf( aTHX_ "\n");
2182 Dumps a fully constructed but uncompressed trie in table form.
2183 This is the normal DFA style state transition table, with a few
2184 twists to facilitate compression later.
2185 Used for debugging make_trie().
2188 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2189 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2194 SV *sv=sv_newmortal();
2195 int colwidth= widecharmap ? 6 : 4;
2196 GET_RE_DEBUG_FLAGS_DECL;
2198 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2201 print out the table precompression so that we can do a visual check
2202 that they are identical.
2205 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2207 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2208 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2210 Perl_re_printf( aTHX_ "%*s",
2212 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2213 PL_colors[0], PL_colors[1],
2214 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2215 PERL_PV_ESCAPE_FIRSTCHAR
2221 Perl_re_printf( aTHX_ "\n");
2222 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2224 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2225 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2228 Perl_re_printf( aTHX_ "\n" );
2230 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2232 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2234 (UV)TRIE_NODENUM( state ) );
2236 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2237 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2239 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2241 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2243 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2244 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2245 (UV)trie->trans[ state ].check );
2247 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2248 (UV)trie->trans[ state ].check,
2249 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2257 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2258 startbranch: the first branch in the whole branch sequence
2259 first : start branch of sequence of branch-exact nodes.
2260 May be the same as startbranch
2261 last : Thing following the last branch.
2262 May be the same as tail.
2263 tail : item following the branch sequence
2264 count : words in the sequence
2265 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2266 depth : indent depth
2268 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2270 A trie is an N'ary tree where the branches are determined by digital
2271 decomposition of the key. IE, at the root node you look up the 1st character and
2272 follow that branch repeat until you find the end of the branches. Nodes can be
2273 marked as "accepting" meaning they represent a complete word. Eg:
2277 would convert into the following structure. Numbers represent states, letters
2278 following numbers represent valid transitions on the letter from that state, if
2279 the number is in square brackets it represents an accepting state, otherwise it
2280 will be in parenthesis.
2282 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2286 (1) +-i->(6)-+-s->[7]
2288 +-s->(3)-+-h->(4)-+-e->[5]
2290 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2292 This shows that when matching against the string 'hers' we will begin at state 1
2293 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2294 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2295 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2296 single traverse. We store a mapping from accepting to state to which word was
2297 matched, and then when we have multiple possibilities we try to complete the
2298 rest of the regex in the order in which they occurred in the alternation.
2300 The only prior NFA like behaviour that would be changed by the TRIE support is
2301 the silent ignoring of duplicate alternations which are of the form:
2303 / (DUPE|DUPE) X? (?{ ... }) Y /x
2305 Thus EVAL blocks following a trie may be called a different number of times with
2306 and without the optimisation. With the optimisations dupes will be silently
2307 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2308 the following demonstrates:
2310 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2312 which prints out 'word' three times, but
2314 'words'=~/(word|word|word)(?{ print $1 })S/
2316 which doesnt print it out at all. This is due to other optimisations kicking in.
2318 Example of what happens on a structural level:
2320 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2322 1: CURLYM[1] {1,32767}(18)
2333 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2334 and should turn into:
2336 1: CURLYM[1] {1,32767}(18)
2338 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2346 Cases where tail != last would be like /(?foo|bar)baz/:
2356 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2357 and would end up looking like:
2360 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2367 d = uvchr_to_utf8_flags(d, uv, 0);
2369 is the recommended Unicode-aware way of saying
2374 #define TRIE_STORE_REVCHAR(val) \
2377 SV *zlopp = newSV(UTF8_MAXBYTES); \
2378 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2379 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2380 SvCUR_set(zlopp, kapow - flrbbbbb); \
2383 av_push(revcharmap, zlopp); \
2385 char ooooff = (char)val; \
2386 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2390 /* This gets the next character from the input, folding it if not already
2392 #define TRIE_READ_CHAR STMT_START { \
2395 /* if it is UTF then it is either already folded, or does not need \
2397 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2399 else if (folder == PL_fold_latin1) { \
2400 /* This folder implies Unicode rules, which in the range expressible \
2401 * by not UTF is the lower case, with the two exceptions, one of \
2402 * which should have been taken care of before calling this */ \
2403 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2404 uvc = toLOWER_L1(*uc); \
2405 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2408 /* raw data, will be folded later if needed */ \
2416 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2417 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2418 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2419 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2420 TRIE_LIST_LEN( state ) = ging; \
2422 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2423 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2424 TRIE_LIST_CUR( state )++; \
2427 #define TRIE_LIST_NEW(state) STMT_START { \
2428 Newx( trie->states[ state ].trans.list, \
2429 4, reg_trie_trans_le ); \
2430 TRIE_LIST_CUR( state ) = 1; \
2431 TRIE_LIST_LEN( state ) = 4; \
2434 #define TRIE_HANDLE_WORD(state) STMT_START { \
2435 U16 dupe= trie->states[ state ].wordnum; \
2436 regnode * const noper_next = regnext( noper ); \
2439 /* store the word for dumping */ \
2441 if (OP(noper) != NOTHING) \
2442 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2444 tmp = newSVpvn_utf8( "", 0, UTF ); \
2445 av_push( trie_words, tmp ); \
2449 trie->wordinfo[curword].prev = 0; \
2450 trie->wordinfo[curword].len = wordlen; \
2451 trie->wordinfo[curword].accept = state; \
2453 if ( noper_next < tail ) { \
2455 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2457 trie->jump[curword] = (U16)(noper_next - convert); \
2459 jumper = noper_next; \
2461 nextbranch= regnext(cur); \
2465 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2466 /* chain, so that when the bits of chain are later */\
2467 /* linked together, the dups appear in the chain */\
2468 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2469 trie->wordinfo[dupe].prev = curword; \
2471 /* we haven't inserted this word yet. */ \
2472 trie->states[ state ].wordnum = curword; \
2477 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2478 ( ( base + charid >= ucharcount \
2479 && base + charid < ubound \
2480 && state == trie->trans[ base - ucharcount + charid ].check \
2481 && trie->trans[ base - ucharcount + charid ].next ) \
2482 ? trie->trans[ base - ucharcount + charid ].next \
2483 : ( state==1 ? special : 0 ) \
2486 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2488 TRIE_BITMAP_SET(trie, uvc); \
2489 /* store the folded codepoint */ \
2491 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2494 /* store first byte of utf8 representation of */ \
2495 /* variant codepoints */ \
2496 if (! UVCHR_IS_INVARIANT(uvc)) { \
2497 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2502 #define MADE_JUMP_TRIE 2
2503 #define MADE_EXACT_TRIE 4
2506 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2507 regnode *first, regnode *last, regnode *tail,
2508 U32 word_count, U32 flags, U32 depth)
2510 /* first pass, loop through and scan words */
2511 reg_trie_data *trie;
2512 HV *widecharmap = NULL;
2513 AV *revcharmap = newAV();
2519 regnode *jumper = NULL;
2520 regnode *nextbranch = NULL;
2521 regnode *convert = NULL;
2522 U32 *prev_states; /* temp array mapping each state to previous one */
2523 /* we just use folder as a flag in utf8 */
2524 const U8 * folder = NULL;
2526 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2527 * which stands for one trie structure, one hash, optionally followed
2530 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2531 AV *trie_words = NULL;
2532 /* along with revcharmap, this only used during construction but both are
2533 * useful during debugging so we store them in the struct when debugging.
2536 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2537 STRLEN trie_charcount=0;
2539 SV *re_trie_maxbuff;
2540 GET_RE_DEBUG_FLAGS_DECL;
2542 PERL_ARGS_ASSERT_MAKE_TRIE;
2544 PERL_UNUSED_ARG(depth);
2548 case EXACT: case EXACTL: break;
2552 case EXACTFLU8: folder = PL_fold_latin1; break;
2553 case EXACTF: folder = PL_fold; break;
2554 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2557 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2559 trie->startstate = 1;
2560 trie->wordcount = word_count;
2561 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2562 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2563 if (flags == EXACT || flags == EXACTL)
2564 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2565 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2566 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2569 trie_words = newAV();
2572 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2573 assert(re_trie_maxbuff);
2574 if (!SvIOK(re_trie_maxbuff)) {
2575 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2577 DEBUG_TRIE_COMPILE_r({
2578 Perl_re_indentf( aTHX_
2579 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2581 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2582 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2585 /* Find the node we are going to overwrite */
2586 if ( first == startbranch && OP( last ) != BRANCH ) {
2587 /* whole branch chain */
2590 /* branch sub-chain */
2591 convert = NEXTOPER( first );
2594 /* -- First loop and Setup --
2596 We first traverse the branches and scan each word to determine if it
2597 contains widechars, and how many unique chars there are, this is
2598 important as we have to build a table with at least as many columns as we
2601 We use an array of integers to represent the character codes 0..255
2602 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2603 the native representation of the character value as the key and IV's for
2606 *TODO* If we keep track of how many times each character is used we can
2607 remap the columns so that the table compression later on is more
2608 efficient in terms of memory by ensuring the most common value is in the
2609 middle and the least common are on the outside. IMO this would be better
2610 than a most to least common mapping as theres a decent chance the most
2611 common letter will share a node with the least common, meaning the node
2612 will not be compressible. With a middle is most common approach the worst
2613 case is when we have the least common nodes twice.
2617 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2618 regnode *noper = NEXTOPER( cur );
2622 U32 wordlen = 0; /* required init */
2623 STRLEN minchars = 0;
2624 STRLEN maxchars = 0;
2625 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2628 if (OP(noper) == NOTHING) {
2629 /* skip past a NOTHING at the start of an alternation
2630 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2632 regnode *noper_next= regnext(noper);
2633 if (noper_next < tail)
2637 if ( noper < tail &&
2639 OP(noper) == flags ||
2642 OP(noper) == EXACTFU_SS
2646 uc= (U8*)STRING(noper);
2647 e= uc + STR_LEN(noper);
2654 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2655 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2656 regardless of encoding */
2657 if (OP( noper ) == EXACTFU_SS) {
2658 /* false positives are ok, so just set this */
2659 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2663 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2665 TRIE_CHARCOUNT(trie)++;
2668 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2669 * is in effect. Under /i, this character can match itself, or
2670 * anything that folds to it. If not under /i, it can match just
2671 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2672 * all fold to k, and all are single characters. But some folds
2673 * expand to more than one character, so for example LATIN SMALL
2674 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2675 * the string beginning at 'uc' is 'ffi', it could be matched by
2676 * three characters, or just by the one ligature character. (It
2677 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2678 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2679 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2680 * match.) The trie needs to know the minimum and maximum number
2681 * of characters that could match so that it can use size alone to
2682 * quickly reject many match attempts. The max is simple: it is
2683 * the number of folded characters in this branch (since a fold is
2684 * never shorter than what folds to it. */
2688 /* And the min is equal to the max if not under /i (indicated by
2689 * 'folder' being NULL), or there are no multi-character folds. If
2690 * there is a multi-character fold, the min is incremented just
2691 * once, for the character that folds to the sequence. Each
2692 * character in the sequence needs to be added to the list below of
2693 * characters in the trie, but we count only the first towards the
2694 * min number of characters needed. This is done through the
2695 * variable 'foldlen', which is returned by the macros that look
2696 * for these sequences as the number of bytes the sequence
2697 * occupies. Each time through the loop, we decrement 'foldlen' by
2698 * how many bytes the current char occupies. Only when it reaches
2699 * 0 do we increment 'minchars' or look for another multi-character
2701 if (folder == NULL) {
2704 else if (foldlen > 0) {
2705 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2710 /* See if *uc is the beginning of a multi-character fold. If
2711 * so, we decrement the length remaining to look at, to account
2712 * for the current character this iteration. (We can use 'uc'
2713 * instead of the fold returned by TRIE_READ_CHAR because for
2714 * non-UTF, the latin1_safe macro is smart enough to account
2715 * for all the unfolded characters, and because for UTF, the
2716 * string will already have been folded earlier in the
2717 * compilation process */
2719 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2720 foldlen -= UTF8SKIP(uc);
2723 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2728 /* The current character (and any potential folds) should be added
2729 * to the possible matching characters for this position in this
2733 U8 folded= folder[ (U8) uvc ];
2734 if ( !trie->charmap[ folded ] ) {
2735 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2736 TRIE_STORE_REVCHAR( folded );
2739 if ( !trie->charmap[ uvc ] ) {
2740 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2741 TRIE_STORE_REVCHAR( uvc );
2744 /* store the codepoint in the bitmap, and its folded
2746 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2747 set_bit = 0; /* We've done our bit :-) */
2751 /* XXX We could come up with the list of code points that fold
2752 * to this using PL_utf8_foldclosures, except not for
2753 * multi-char folds, as there may be multiple combinations
2754 * there that could work, which needs to wait until runtime to
2755 * resolve (The comment about LIGATURE FFI above is such an
2760 widecharmap = newHV();
2762 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2765 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2767 if ( !SvTRUE( *svpp ) ) {
2768 sv_setiv( *svpp, ++trie->uniquecharcount );
2769 TRIE_STORE_REVCHAR(uvc);
2772 } /* end loop through characters in this branch of the trie */
2774 /* We take the min and max for this branch and combine to find the min
2775 * and max for all branches processed so far */
2776 if( cur == first ) {
2777 trie->minlen = minchars;
2778 trie->maxlen = maxchars;
2779 } else if (minchars < trie->minlen) {
2780 trie->minlen = minchars;
2781 } else if (maxchars > trie->maxlen) {
2782 trie->maxlen = maxchars;
2784 } /* end first pass */
2785 DEBUG_TRIE_COMPILE_r(
2786 Perl_re_indentf( aTHX_
2787 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2789 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2790 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2791 (int)trie->minlen, (int)trie->maxlen )
2795 We now know what we are dealing with in terms of unique chars and
2796 string sizes so we can calculate how much memory a naive
2797 representation using a flat table will take. If it's over a reasonable
2798 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2799 conservative but potentially much slower representation using an array
2802 At the end we convert both representations into the same compressed
2803 form that will be used in regexec.c for matching with. The latter
2804 is a form that cannot be used to construct with but has memory
2805 properties similar to the list form and access properties similar
2806 to the table form making it both suitable for fast searches and
2807 small enough that its feasable to store for the duration of a program.
2809 See the comment in the code where the compressed table is produced
2810 inplace from the flat tabe representation for an explanation of how
2811 the compression works.
2816 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2819 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2820 > SvIV(re_trie_maxbuff) )
2823 Second Pass -- Array Of Lists Representation
2825 Each state will be represented by a list of charid:state records
2826 (reg_trie_trans_le) the first such element holds the CUR and LEN
2827 points of the allocated array. (See defines above).
2829 We build the initial structure using the lists, and then convert
2830 it into the compressed table form which allows faster lookups
2831 (but cant be modified once converted).
2834 STRLEN transcount = 1;
2836 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2839 trie->states = (reg_trie_state *)
2840 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2841 sizeof(reg_trie_state) );
2845 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2847 regnode *noper = NEXTOPER( cur );
2848 U32 state = 1; /* required init */
2849 U16 charid = 0; /* sanity init */
2850 U32 wordlen = 0; /* required init */
2852 if (OP(noper) == NOTHING) {
2853 regnode *noper_next= regnext(noper);
2854 if (noper_next < tail)
2858 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2859 const U8 *uc= (U8*)STRING(noper);
2860 const U8 *e= uc + STR_LEN(noper);
2862 for ( ; uc < e ; uc += len ) {
2867 charid = trie->charmap[ uvc ];
2869 SV** const svpp = hv_fetch( widecharmap,
2876 charid=(U16)SvIV( *svpp );
2879 /* charid is now 0 if we dont know the char read, or
2880 * nonzero if we do */
2887 if ( !trie->states[ state ].trans.list ) {
2888 TRIE_LIST_NEW( state );
2891 check <= TRIE_LIST_USED( state );
2894 if ( TRIE_LIST_ITEM( state, check ).forid
2897 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2902 newstate = next_alloc++;
2903 prev_states[newstate] = state;
2904 TRIE_LIST_PUSH( state, charid, newstate );
2909 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
2913 TRIE_HANDLE_WORD(state);
2915 } /* end second pass */
2917 /* next alloc is the NEXT state to be allocated */
2918 trie->statecount = next_alloc;
2919 trie->states = (reg_trie_state *)
2920 PerlMemShared_realloc( trie->states,
2922 * sizeof(reg_trie_state) );
2924 /* and now dump it out before we compress it */
2925 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2926 revcharmap, next_alloc,
2930 trie->trans = (reg_trie_trans *)
2931 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2938 for( state=1 ; state < next_alloc ; state ++ ) {
2942 DEBUG_TRIE_COMPILE_MORE_r(
2943 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
2947 if (trie->states[state].trans.list) {
2948 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2952 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2953 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2954 if ( forid < minid ) {
2956 } else if ( forid > maxid ) {
2960 if ( transcount < tp + maxid - minid + 1) {
2962 trie->trans = (reg_trie_trans *)
2963 PerlMemShared_realloc( trie->trans,
2965 * sizeof(reg_trie_trans) );
2966 Zero( trie->trans + (transcount / 2),
2970 base = trie->uniquecharcount + tp - minid;
2971 if ( maxid == minid ) {
2973 for ( ; zp < tp ; zp++ ) {
2974 if ( ! trie->trans[ zp ].next ) {
2975 base = trie->uniquecharcount + zp - minid;
2976 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2978 trie->trans[ zp ].check = state;
2984 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2986 trie->trans[ tp ].check = state;
2991 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2992 const U32 tid = base
2993 - trie->uniquecharcount
2994 + TRIE_LIST_ITEM( state, idx ).forid;
2995 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2997 trie->trans[ tid ].check = state;
2999 tp += ( maxid - minid + 1 );
3001 Safefree(trie->states[ state ].trans.list);
3004 DEBUG_TRIE_COMPILE_MORE_r(
3005 Perl_re_printf( aTHX_ " base: %d\n",base);
3008 trie->states[ state ].trans.base=base;
3010 trie->lasttrans = tp + 1;
3014 Second Pass -- Flat Table Representation.
3016 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3017 each. We know that we will need Charcount+1 trans at most to store
3018 the data (one row per char at worst case) So we preallocate both
3019 structures assuming worst case.
3021 We then construct the trie using only the .next slots of the entry
3024 We use the .check field of the first entry of the node temporarily
3025 to make compression both faster and easier by keeping track of how
3026 many non zero fields are in the node.
3028 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3031 There are two terms at use here: state as a TRIE_NODEIDX() which is
3032 a number representing the first entry of the node, and state as a
3033 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3034 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3035 if there are 2 entrys per node. eg:
3043 The table is internally in the right hand, idx form. However as we
3044 also have to deal with the states array which is indexed by nodenum
3045 we have to use TRIE_NODENUM() to convert.
3048 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3051 trie->trans = (reg_trie_trans *)
3052 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3053 * trie->uniquecharcount + 1,
3054 sizeof(reg_trie_trans) );
3055 trie->states = (reg_trie_state *)
3056 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3057 sizeof(reg_trie_state) );
3058 next_alloc = trie->uniquecharcount + 1;
3061 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3063 regnode *noper = NEXTOPER( cur );
3065 U32 state = 1; /* required init */
3067 U16 charid = 0; /* sanity init */
3068 U32 accept_state = 0; /* sanity init */
3070 U32 wordlen = 0; /* required init */
3072 if (OP(noper) == NOTHING) {
3073 regnode *noper_next= regnext(noper);
3074 if (noper_next < tail)
3078 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3079 const U8 *uc= (U8*)STRING(noper);
3080 const U8 *e= uc + STR_LEN(noper);
3082 for ( ; uc < e ; uc += len ) {
3087 charid = trie->charmap[ uvc ];
3089 SV* const * const svpp = hv_fetch( widecharmap,
3093 charid = svpp ? (U16)SvIV(*svpp) : 0;
3097 if ( !trie->trans[ state + charid ].next ) {
3098 trie->trans[ state + charid ].next = next_alloc;
3099 trie->trans[ state ].check++;
3100 prev_states[TRIE_NODENUM(next_alloc)]
3101 = TRIE_NODENUM(state);
3102 next_alloc += trie->uniquecharcount;
3104 state = trie->trans[ state + charid ].next;
3106 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3108 /* charid is now 0 if we dont know the char read, or
3109 * nonzero if we do */
3112 accept_state = TRIE_NODENUM( state );
3113 TRIE_HANDLE_WORD(accept_state);
3115 } /* end second pass */
3117 /* and now dump it out before we compress it */
3118 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3120 next_alloc, depth+1));
3124 * Inplace compress the table.*
3126 For sparse data sets the table constructed by the trie algorithm will
3127 be mostly 0/FAIL transitions or to put it another way mostly empty.
3128 (Note that leaf nodes will not contain any transitions.)
3130 This algorithm compresses the tables by eliminating most such
3131 transitions, at the cost of a modest bit of extra work during lookup:
3133 - Each states[] entry contains a .base field which indicates the
3134 index in the state[] array wheres its transition data is stored.
3136 - If .base is 0 there are no valid transitions from that node.
3138 - If .base is nonzero then charid is added to it to find an entry in
3141 -If trans[states[state].base+charid].check!=state then the
3142 transition is taken to be a 0/Fail transition. Thus if there are fail
3143 transitions at the front of the node then the .base offset will point
3144 somewhere inside the previous nodes data (or maybe even into a node
3145 even earlier), but the .check field determines if the transition is
3149 The following process inplace converts the table to the compressed
3150 table: We first do not compress the root node 1,and mark all its
3151 .check pointers as 1 and set its .base pointer as 1 as well. This
3152 allows us to do a DFA construction from the compressed table later,
3153 and ensures that any .base pointers we calculate later are greater
3156 - We set 'pos' to indicate the first entry of the second node.
3158 - We then iterate over the columns of the node, finding the first and
3159 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3160 and set the .check pointers accordingly, and advance pos
3161 appropriately and repreat for the next node. Note that when we copy
3162 the next pointers we have to convert them from the original
3163 NODEIDX form to NODENUM form as the former is not valid post
3166 - If a node has no transitions used we mark its base as 0 and do not
3167 advance the pos pointer.
3169 - If a node only has one transition we use a second pointer into the
3170 structure to fill in allocated fail transitions from other states.
3171 This pointer is independent of the main pointer and scans forward
3172 looking for null transitions that are allocated to a state. When it
3173 finds one it writes the single transition into the "hole". If the
3174 pointer doesnt find one the single transition is appended as normal.
3176 - Once compressed we can Renew/realloc the structures to release the
3179 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3180 specifically Fig 3.47 and the associated pseudocode.
3184 const U32 laststate = TRIE_NODENUM( next_alloc );
3187 trie->statecount = laststate;
3189 for ( state = 1 ; state < laststate ; state++ ) {
3191 const U32 stateidx = TRIE_NODEIDX( state );
3192 const U32 o_used = trie->trans[ stateidx ].check;
3193 U32 used = trie->trans[ stateidx ].check;
3194 trie->trans[ stateidx ].check = 0;
3197 used && charid < trie->uniquecharcount;
3200 if ( flag || trie->trans[ stateidx + charid ].next ) {
3201 if ( trie->trans[ stateidx + charid ].next ) {
3203 for ( ; zp < pos ; zp++ ) {
3204 if ( ! trie->trans[ zp ].next ) {
3208 trie->states[ state ].trans.base
3210 + trie->uniquecharcount
3212 trie->trans[ zp ].next
3213 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3215 trie->trans[ zp ].check = state;
3216 if ( ++zp > pos ) pos = zp;
3223 trie->states[ state ].trans.base
3224 = pos + trie->uniquecharcount - charid ;
3226 trie->trans[ pos ].next
3227 = SAFE_TRIE_NODENUM(
3228 trie->trans[ stateidx + charid ].next );
3229 trie->trans[ pos ].check = state;
3234 trie->lasttrans = pos + 1;
3235 trie->states = (reg_trie_state *)
3236 PerlMemShared_realloc( trie->states, laststate
3237 * sizeof(reg_trie_state) );
3238 DEBUG_TRIE_COMPILE_MORE_r(
3239 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3241 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3245 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3248 } /* end table compress */
3250 DEBUG_TRIE_COMPILE_MORE_r(
3251 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3253 (UV)trie->statecount,
3254 (UV)trie->lasttrans)
3256 /* resize the trans array to remove unused space */
3257 trie->trans = (reg_trie_trans *)
3258 PerlMemShared_realloc( trie->trans, trie->lasttrans
3259 * sizeof(reg_trie_trans) );
3261 { /* Modify the program and insert the new TRIE node */
3262 U8 nodetype =(U8)(flags & 0xFF);
3266 regnode *optimize = NULL;
3267 #ifdef RE_TRACK_PATTERN_OFFSETS
3270 U32 mjd_nodelen = 0;
3271 #endif /* RE_TRACK_PATTERN_OFFSETS */
3272 #endif /* DEBUGGING */
3274 This means we convert either the first branch or the first Exact,
3275 depending on whether the thing following (in 'last') is a branch
3276 or not and whther first is the startbranch (ie is it a sub part of
3277 the alternation or is it the whole thing.)
3278 Assuming its a sub part we convert the EXACT otherwise we convert
3279 the whole branch sequence, including the first.
3281 /* Find the node we are going to overwrite */
3282 if ( first != startbranch || OP( last ) == BRANCH ) {
3283 /* branch sub-chain */
3284 NEXT_OFF( first ) = (U16)(last - first);
3285 #ifdef RE_TRACK_PATTERN_OFFSETS
3287 mjd_offset= Node_Offset((convert));
3288 mjd_nodelen= Node_Length((convert));
3291 /* whole branch chain */
3293 #ifdef RE_TRACK_PATTERN_OFFSETS
3296 const regnode *nop = NEXTOPER( convert );
3297 mjd_offset= Node_Offset((nop));
3298 mjd_nodelen= Node_Length((nop));
3302 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3304 (UV)mjd_offset, (UV)mjd_nodelen)
3307 /* But first we check to see if there is a common prefix we can
3308 split out as an EXACT and put in front of the TRIE node. */
3309 trie->startstate= 1;
3310 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3311 /* we want to find the first state that has more than
3312 * one transition, if that state is not the first state
3313 * then we have a common prefix which we can remove.
3316 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3318 I32 first_ofs = -1; /* keeps track of the ofs of the first
3319 transition, -1 means none */
3321 const U32 base = trie->states[ state ].trans.base;
3323 /* does this state terminate an alternation? */
3324 if ( trie->states[state].wordnum )
3327 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3328 if ( ( base + ofs >= trie->uniquecharcount ) &&
3329 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3330 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3332 if ( ++count > 1 ) {
3333 /* we have more than one transition */
3336 /* if this is the first state there is no common prefix
3337 * to extract, so we can exit */
3338 if ( state == 1 ) break;
3339 tmp = av_fetch( revcharmap, ofs, 0);
3340 ch = (U8*)SvPV_nolen_const( *tmp );
3342 /* if we are on count 2 then we need to initialize the
3343 * bitmap, and store the previous char if there was one
3346 /* clear the bitmap */
3347 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3349 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3352 if (first_ofs >= 0) {
3353 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3354 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3356 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3358 Perl_re_printf( aTHX_ "%s", (char*)ch)
3362 /* store the current firstchar in the bitmap */
3363 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3364 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3370 /* This state has only one transition, its transition is part
3371 * of a common prefix - we need to concatenate the char it
3372 * represents to what we have so far. */
3373 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3375 char *ch = SvPV( *tmp, len );
3377 SV *sv=sv_newmortal();
3378 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3380 (UV)state, (UV)first_ofs,
3381 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3382 PL_colors[0], PL_colors[1],
3383 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3384 PERL_PV_ESCAPE_FIRSTCHAR
3389 OP( convert ) = nodetype;
3390 str=STRING(convert);
3393 STR_LEN(convert) += len;
3399 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3404 trie->prefixlen = (state-1);
3406 regnode *n = convert+NODE_SZ_STR(convert);
3407 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3408 trie->startstate = state;
3409 trie->minlen -= (state - 1);
3410 trie->maxlen -= (state - 1);
3412 /* At least the UNICOS C compiler choked on this
3413 * being argument to DEBUG_r(), so let's just have
3416 #ifdef PERL_EXT_RE_BUILD
3422 regnode *fix = convert;
3423 U32 word = trie->wordcount;
3425 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3426 while( ++fix < n ) {
3427 Set_Node_Offset_Length(fix, 0, 0);
3430 SV ** const tmp = av_fetch( trie_words, word, 0 );
3432 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3433 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3435 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3443 NEXT_OFF(convert) = (U16)(tail - convert);
3444 DEBUG_r(optimize= n);
3450 if ( trie->maxlen ) {
3451 NEXT_OFF( convert ) = (U16)(tail - convert);
3452 ARG_SET( convert, data_slot );
3453 /* Store the offset to the first unabsorbed branch in
3454 jump[0], which is otherwise unused by the jump logic.
3455 We use this when dumping a trie and during optimisation. */
3457 trie->jump[0] = (U16)(nextbranch - convert);
3459 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3460 * and there is a bitmap
3461 * and the first "jump target" node we found leaves enough room
3462 * then convert the TRIE node into a TRIEC node, with the bitmap
3463 * embedded inline in the opcode - this is hypothetically faster.
3465 if ( !trie->states[trie->startstate].wordnum
3467 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3469 OP( convert ) = TRIEC;
3470 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3471 PerlMemShared_free(trie->bitmap);
3474 OP( convert ) = TRIE;
3476 /* store the type in the flags */
3477 convert->flags = nodetype;
3481 + regarglen[ OP( convert ) ];
3483 /* XXX We really should free up the resource in trie now,
3484 as we won't use them - (which resources?) dmq */
3486 /* needed for dumping*/
3487 DEBUG_r(if (optimize) {
3488 regnode *opt = convert;
3490 while ( ++opt < optimize) {
3491 Set_Node_Offset_Length(opt,0,0);
3494 Try to clean up some of the debris left after the
3497 while( optimize < jumper ) {
3498 mjd_nodelen += Node_Length((optimize));
3499 OP( optimize ) = OPTIMIZED;
3500 Set_Node_Offset_Length(optimize,0,0);
3503 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3505 } /* end node insert */
3507 /* Finish populating the prev field of the wordinfo array. Walk back
3508 * from each accept state until we find another accept state, and if
3509 * so, point the first word's .prev field at the second word. If the
3510 * second already has a .prev field set, stop now. This will be the
3511 * case either if we've already processed that word's accept state,
3512 * or that state had multiple words, and the overspill words were
3513 * already linked up earlier.
3520 for (word=1; word <= trie->wordcount; word++) {
3522 if (trie->wordinfo[word].prev)
3524 state = trie->wordinfo[word].accept;
3526 state = prev_states[state];
3529 prev = trie->states[state].wordnum;
3533 trie->wordinfo[word].prev = prev;
3535 Safefree(prev_states);
3539 /* and now dump out the compressed format */
3540 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3542 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3544 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3545 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3547 SvREFCNT_dec_NN(revcharmap);
3551 : trie->startstate>1
3557 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3559 /* The Trie is constructed and compressed now so we can build a fail array if
3562 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3564 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3568 We find the fail state for each state in the trie, this state is the longest
3569 proper suffix of the current state's 'word' that is also a proper prefix of
3570 another word in our trie. State 1 represents the word '' and is thus the
3571 default fail state. This allows the DFA not to have to restart after its
3572 tried and failed a word at a given point, it simply continues as though it
3573 had been matching the other word in the first place.
3575 'abcdgu'=~/abcdefg|cdgu/
3576 When we get to 'd' we are still matching the first word, we would encounter
3577 'g' which would fail, which would bring us to the state representing 'd' in
3578 the second word where we would try 'g' and succeed, proceeding to match
3581 /* add a fail transition */
3582 const U32 trie_offset = ARG(source);
3583 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3585 const U32 ucharcount = trie->uniquecharcount;
3586 const U32 numstates = trie->statecount;
3587 const U32 ubound = trie->lasttrans + ucharcount;
3591 U32 base = trie->states[ 1 ].trans.base;
3594 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3596 GET_RE_DEBUG_FLAGS_DECL;
3598 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3599 PERL_UNUSED_CONTEXT;
3601 PERL_UNUSED_ARG(depth);
3604 if ( OP(source) == TRIE ) {
3605 struct regnode_1 *op = (struct regnode_1 *)
3606 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3607 StructCopy(source,op,struct regnode_1);
3608 stclass = (regnode *)op;
3610 struct regnode_charclass *op = (struct regnode_charclass *)
3611 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3612 StructCopy(source,op,struct regnode_charclass);
3613 stclass = (regnode *)op;
3615 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3617 ARG_SET( stclass, data_slot );
3618 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3619 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3620 aho->trie=trie_offset;
3621 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3622 Copy( trie->states, aho->states, numstates, reg_trie_state );
3623 Newx( q, numstates, U32);
3624 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3627 /* initialize fail[0..1] to be 1 so that we always have
3628 a valid final fail state */
3629 fail[ 0 ] = fail[ 1 ] = 1;
3631 for ( charid = 0; charid < ucharcount ; charid++ ) {
3632 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3634 q[ q_write ] = newstate;
3635 /* set to point at the root */
3636 fail[ q[ q_write++ ] ]=1;
3639 while ( q_read < q_write) {
3640 const U32 cur = q[ q_read++ % numstates ];
3641 base = trie->states[ cur ].trans.base;
3643 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3644 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3646 U32 fail_state = cur;
3649 fail_state = fail[ fail_state ];
3650 fail_base = aho->states[ fail_state ].trans.base;
3651 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3653 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3654 fail[ ch_state ] = fail_state;
3655 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3657 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3659 q[ q_write++ % numstates] = ch_state;
3663 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3664 when we fail in state 1, this allows us to use the
3665 charclass scan to find a valid start char. This is based on the principle
3666 that theres a good chance the string being searched contains lots of stuff
3667 that cant be a start char.
3669 fail[ 0 ] = fail[ 1 ] = 0;
3670 DEBUG_TRIE_COMPILE_r({
3671 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3672 depth, (UV)numstates
3674 for( q_read=1; q_read<numstates; q_read++ ) {
3675 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3677 Perl_re_printf( aTHX_ "\n");
3680 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3685 /* The below joins as many adjacent EXACTish nodes as possible into a single
3686 * one. The regop may be changed if the node(s) contain certain sequences that
3687 * require special handling. The joining is only done if:
3688 * 1) there is room in the current conglomerated node to entirely contain the
3690 * 2) they are the exact same node type
3692 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3693 * these get optimized out
3695 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3696 * as possible, even if that means splitting an existing node so that its first
3697 * part is moved to the preceeding node. This would maximise the efficiency of
3698 * memEQ during matching.
3700 * If a node is to match under /i (folded), the number of characters it matches
3701 * can be different than its character length if it contains a multi-character
3702 * fold. *min_subtract is set to the total delta number of characters of the
3705 * And *unfolded_multi_char is set to indicate whether or not the node contains
3706 * an unfolded multi-char fold. This happens when it won't be known until
3707 * runtime whether the fold is valid or not; namely
3708 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3709 * target string being matched against turns out to be UTF-8 is that fold
3711 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3713 * (Multi-char folds whose components are all above the Latin1 range are not
3714 * run-time locale dependent, and have already been folded by the time this
3715 * function is called.)
3717 * This is as good a place as any to discuss the design of handling these
3718 * multi-character fold sequences. It's been wrong in Perl for a very long
3719 * time. There are three code points in Unicode whose multi-character folds
3720 * were long ago discovered to mess things up. The previous designs for
3721 * dealing with these involved assigning a special node for them. This
3722 * approach doesn't always work, as evidenced by this example:
3723 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3724 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3725 * would match just the \xDF, it won't be able to handle the case where a
3726 * successful match would have to cross the node's boundary. The new approach
3727 * that hopefully generally solves the problem generates an EXACTFU_SS node
3728 * that is "sss" in this case.
3730 * It turns out that there are problems with all multi-character folds, and not
3731 * just these three. Now the code is general, for all such cases. The
3732 * approach taken is:
3733 * 1) This routine examines each EXACTFish node that could contain multi-
3734 * character folded sequences. Since a single character can fold into
3735 * such a sequence, the minimum match length for this node is less than
3736 * the number of characters in the node. This routine returns in
3737 * *min_subtract how many characters to subtract from the the actual
3738 * length of the string to get a real minimum match length; it is 0 if
3739 * there are no multi-char foldeds. This delta is used by the caller to
3740 * adjust the min length of the match, and the delta between min and max,
3741 * so that the optimizer doesn't reject these possibilities based on size
3743 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3744 * is used for an EXACTFU node that contains at least one "ss" sequence in
3745 * it. For non-UTF-8 patterns and strings, this is the only case where
3746 * there is a possible fold length change. That means that a regular
3747 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3748 * with length changes, and so can be processed faster. regexec.c takes
3749 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3750 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3751 * known until runtime). This saves effort in regex matching. However,
3752 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3753 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3754 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3755 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3756 * possibilities for the non-UTF8 patterns are quite simple, except for
3757 * the sharp s. All the ones that don't involve a UTF-8 target string are
3758 * members of a fold-pair, and arrays are set up for all of them so that
3759 * the other member of the pair can be found quickly. Code elsewhere in
3760 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3761 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3762 * described in the next item.
3763 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3764 * validity of the fold won't be known until runtime, and so must remain
3765 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3766 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3767 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3768 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3769 * The reason this is a problem is that the optimizer part of regexec.c
3770 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3771 * that a character in the pattern corresponds to at most a single
3772 * character in the target string. (And I do mean character, and not byte
3773 * here, unlike other parts of the documentation that have never been
3774 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3775 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3776 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3777 * nodes, violate the assumption, and they are the only instances where it
3778 * is violated. I'm reluctant to try to change the assumption, as the
3779 * code involved is impenetrable to me (khw), so instead the code here
3780 * punts. This routine examines EXACTFL nodes, and (when the pattern
3781 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3782 * boolean indicating whether or not the node contains such a fold. When
3783 * it is true, the caller sets a flag that later causes the optimizer in
3784 * this file to not set values for the floating and fixed string lengths,
3785 * and thus avoids the optimizer code in regexec.c that makes the invalid
3786 * assumption. Thus, there is no optimization based on string lengths for
3787 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3788 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3789 * assumption is wrong only in these cases is that all other non-UTF-8
3790 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3791 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3792 * EXACTF nodes because we don't know at compile time if it actually
3793 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3794 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3795 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3796 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3797 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3798 * string would require the pattern to be forced into UTF-8, the overhead
3799 * of which we want to avoid. Similarly the unfolded multi-char folds in
3800 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3803 * Similarly, the code that generates tries doesn't currently handle
3804 * not-already-folded multi-char folds, and it looks like a pain to change
3805 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3806 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3807 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3808 * using /iaa matching will be doing so almost entirely with ASCII
3809 * strings, so this should rarely be encountered in practice */
3811 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3812 if (PL_regkind[OP(scan)] == EXACT) \
3813 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3816 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3817 UV *min_subtract, bool *unfolded_multi_char,
3818 U32 flags,regnode *val, U32 depth)
3820 /* Merge several consecutive EXACTish nodes into one. */
3821 regnode *n = regnext(scan);
3823 regnode *next = scan + NODE_SZ_STR(scan);
3827 regnode *stop = scan;
3828 GET_RE_DEBUG_FLAGS_DECL;
3830 PERL_UNUSED_ARG(depth);
3833 PERL_ARGS_ASSERT_JOIN_EXACT;
3834 #ifndef EXPERIMENTAL_INPLACESCAN
3835 PERL_UNUSED_ARG(flags);
3836 PERL_UNUSED_ARG(val);
3838 DEBUG_PEEP("join", scan, depth, 0);
3840 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3841 * EXACT ones that are mergeable to the current one. */
3843 && (PL_regkind[OP(n)] == NOTHING
3844 || (stringok && OP(n) == OP(scan)))
3846 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3849 if (OP(n) == TAIL || n > next)
3851 if (PL_regkind[OP(n)] == NOTHING) {
3852 DEBUG_PEEP("skip:", n, depth, 0);
3853 NEXT_OFF(scan) += NEXT_OFF(n);
3854 next = n + NODE_STEP_REGNODE;
3861 else if (stringok) {
3862 const unsigned int oldl = STR_LEN(scan);
3863 regnode * const nnext = regnext(n);
3865 /* XXX I (khw) kind of doubt that this works on platforms (should
3866 * Perl ever run on one) where U8_MAX is above 255 because of lots
3867 * of other assumptions */
3868 /* Don't join if the sum can't fit into a single node */
3869 if (oldl + STR_LEN(n) > U8_MAX)
3872 DEBUG_PEEP("merg", n, depth, 0);
3875 NEXT_OFF(scan) += NEXT_OFF(n);
3876 STR_LEN(scan) += STR_LEN(n);
3877 next = n + NODE_SZ_STR(n);
3878 /* Now we can overwrite *n : */
3879 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3887 #ifdef EXPERIMENTAL_INPLACESCAN
3888 if (flags && !NEXT_OFF(n)) {
3889 DEBUG_PEEP("atch", val, depth, 0);
3890 if (reg_off_by_arg[OP(n)]) {
3891 ARG_SET(n, val - n);
3894 NEXT_OFF(n) = val - n;
3902 *unfolded_multi_char = FALSE;
3904 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3905 * can now analyze for sequences of problematic code points. (Prior to
3906 * this final joining, sequences could have been split over boundaries, and
3907 * hence missed). The sequences only happen in folding, hence for any
3908 * non-EXACT EXACTish node */
3909 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3910 U8* s0 = (U8*) STRING(scan);
3912 U8* s_end = s0 + STR_LEN(scan);
3914 int total_count_delta = 0; /* Total delta number of characters that
3915 multi-char folds expand to */
3917 /* One pass is made over the node's string looking for all the
3918 * possibilities. To avoid some tests in the loop, there are two main
3919 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3924 if (OP(scan) == EXACTFL) {
3927 /* An EXACTFL node would already have been changed to another
3928 * node type unless there is at least one character in it that
3929 * is problematic; likely a character whose fold definition
3930 * won't be known until runtime, and so has yet to be folded.
3931 * For all but the UTF-8 locale, folds are 1-1 in length, but
3932 * to handle the UTF-8 case, we need to create a temporary
3933 * folded copy using UTF-8 locale rules in order to analyze it.
3934 * This is because our macros that look to see if a sequence is
3935 * a multi-char fold assume everything is folded (otherwise the
3936 * tests in those macros would be too complicated and slow).
3937 * Note that here, the non-problematic folds will have already
3938 * been done, so we can just copy such characters. We actually
3939 * don't completely fold the EXACTFL string. We skip the
3940 * unfolded multi-char folds, as that would just create work
3941 * below to figure out the size they already are */
3943 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3946 STRLEN s_len = UTF8SKIP(s);
3947 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3948 Copy(s, d, s_len, U8);
3951 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3952 *unfolded_multi_char = TRUE;
3953 Copy(s, d, s_len, U8);
3956 else if (isASCII(*s)) {
3957 *(d++) = toFOLD(*s);
3961 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
3967 /* Point the remainder of the routine to look at our temporary
3971 } /* End of creating folded copy of EXACTFL string */
3973 /* Examine the string for a multi-character fold sequence. UTF-8
3974 * patterns have all characters pre-folded by the time this code is
3976 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3977 length sequence we are looking for is 2 */
3979 int count = 0; /* How many characters in a multi-char fold */
3980 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3981 if (! len) { /* Not a multi-char fold: get next char */
3986 /* Nodes with 'ss' require special handling, except for
3987 * EXACTFA-ish for which there is no multi-char fold to this */
3988 if (len == 2 && *s == 's' && *(s+1) == 's'
3989 && OP(scan) != EXACTFA
3990 && OP(scan) != EXACTFA_NO_TRIE)
3993 if (OP(scan) != EXACTFL) {
3994 OP(scan) = EXACTFU_SS;
3998 else { /* Here is a generic multi-char fold. */
3999 U8* multi_end = s + len;
4001 /* Count how many characters are in it. In the case of
4002 * /aa, no folds which contain ASCII code points are
4003 * allowed, so check for those, and skip if found. */
4004 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
4005 count = utf8_length(s, multi_end);
4009 while (s < multi_end) {
4012 goto next_iteration;
4022 /* The delta is how long the sequence is minus 1 (1 is how long
4023 * the character that folds to the sequence is) */
https://perl5.git.perl.org / perl5.git / blob