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_static.c"
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
105 #define MIN(a,b) ((a) < (b) ? (a) : (b))
108 /* this is a chain of data about sub patterns we are processing that
109 need to be handled separately/specially in study_chunk. Its so
110 we can simulate recursion without losing state. */
112 typedef struct scan_frame {
113 regnode *last_regnode; /* last node to process in this frame */
114 regnode *next_regnode; /* next node to process when last is reached */
115 U32 prev_recursed_depth;
116 I32 stopparen; /* what stopparen do we use */
117 U32 is_top_frame; /* what flags do we use? */
119 struct scan_frame *this_prev_frame; /* this previous frame */
120 struct scan_frame *prev_frame; /* previous frame */
121 struct scan_frame *next_frame; /* next frame */
124 /* Certain characters are output as a sequence with the first being a
126 #define isBACKSLASHED_PUNCT(c) \
127 ((c) == '-' || (c) == ']' || (c) == '\\' || (c) == '^')
130 struct RExC_state_t {
131 U32 flags; /* RXf_* are we folding, multilining? */
132 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
133 char *precomp; /* uncompiled string. */
134 REGEXP *rx_sv; /* The SV that is the regexp. */
135 regexp *rx; /* perl core regexp structure */
136 regexp_internal *rxi; /* internal data for regexp object
138 char *start; /* Start of input for compile */
139 char *end; /* End of input for compile */
140 char *parse; /* Input-scan pointer. */
141 SSize_t whilem_seen; /* number of WHILEM in this expr */
142 regnode *emit_start; /* Start of emitted-code area */
143 regnode *emit_bound; /* First regnode outside of the
145 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
146 implies compiling, so don't emit */
147 regnode_ssc emit_dummy; /* placeholder for emit to point to;
148 large enough for the largest
149 non-EXACTish node, so can use it as
151 I32 naughty; /* How bad is this pattern? */
152 I32 sawback; /* Did we see \1, ...? */
154 SSize_t size; /* Code size. */
155 I32 npar; /* Capture buffer count, (OPEN) plus
156 one. ("par" 0 is the whole
158 I32 nestroot; /* root parens we are in - used by
162 regnode **open_parens; /* pointers to open parens */
163 regnode **close_parens; /* pointers to close parens */
164 regnode *opend; /* END node in program */
165 I32 utf8; /* whether the pattern is utf8 or not */
166 I32 orig_utf8; /* whether the pattern was originally in utf8 */
167 /* XXX use this for future optimisation of case
168 * where pattern must be upgraded to utf8. */
169 I32 uni_semantics; /* If a d charset modifier should use unicode
170 rules, even if the pattern is not in
172 HV *paren_names; /* Paren names */
174 regnode **recurse; /* Recurse regops */
175 I32 recurse_count; /* Number of recurse regops */
176 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
178 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
182 I32 override_recoding;
184 I32 recode_x_to_native;
186 I32 in_multi_char_class;
187 struct reg_code_block *code_blocks; /* positions of literal (?{})
189 int num_code_blocks; /* size of code_blocks[] */
190 int code_index; /* next code_blocks[] slot */
191 SSize_t maxlen; /* mininum possible number of chars in string to match */
192 scan_frame *frame_head;
193 scan_frame *frame_last;
196 #ifdef ADD_TO_REGEXEC
197 char *starttry; /* -Dr: where regtry was called. */
198 #define RExC_starttry (pRExC_state->starttry)
200 SV *runtime_code_qr; /* qr with the runtime code blocks */
202 const char *lastparse;
204 AV *paren_name_list; /* idx -> name */
205 U32 study_chunk_recursed_count;
208 #define RExC_lastparse (pRExC_state->lastparse)
209 #define RExC_lastnum (pRExC_state->lastnum)
210 #define RExC_paren_name_list (pRExC_state->paren_name_list)
211 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
212 #define RExC_mysv (pRExC_state->mysv1)
213 #define RExC_mysv1 (pRExC_state->mysv1)
214 #define RExC_mysv2 (pRExC_state->mysv2)
219 #define RExC_flags (pRExC_state->flags)
220 #define RExC_pm_flags (pRExC_state->pm_flags)
221 #define RExC_precomp (pRExC_state->precomp)
222 #define RExC_rx_sv (pRExC_state->rx_sv)
223 #define RExC_rx (pRExC_state->rx)
224 #define RExC_rxi (pRExC_state->rxi)
225 #define RExC_start (pRExC_state->start)
226 #define RExC_end (pRExC_state->end)
227 #define RExC_parse (pRExC_state->parse)
228 #define RExC_whilem_seen (pRExC_state->whilem_seen)
229 #ifdef RE_TRACK_PATTERN_OFFSETS
230 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
233 #define RExC_emit (pRExC_state->emit)
234 #define RExC_emit_dummy (pRExC_state->emit_dummy)
235 #define RExC_emit_start (pRExC_state->emit_start)
236 #define RExC_emit_bound (pRExC_state->emit_bound)
237 #define RExC_sawback (pRExC_state->sawback)
238 #define RExC_seen (pRExC_state->seen)
239 #define RExC_size (pRExC_state->size)
240 #define RExC_maxlen (pRExC_state->maxlen)
241 #define RExC_npar (pRExC_state->npar)
242 #define RExC_nestroot (pRExC_state->nestroot)
243 #define RExC_extralen (pRExC_state->extralen)
244 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
245 #define RExC_utf8 (pRExC_state->utf8)
246 #define RExC_uni_semantics (pRExC_state->uni_semantics)
247 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
248 #define RExC_open_parens (pRExC_state->open_parens)
249 #define RExC_close_parens (pRExC_state->close_parens)
250 #define RExC_opend (pRExC_state->opend)
251 #define RExC_paren_names (pRExC_state->paren_names)
252 #define RExC_recurse (pRExC_state->recurse)
253 #define RExC_recurse_count (pRExC_state->recurse_count)
254 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
255 #define RExC_study_chunk_recursed_bytes \
256 (pRExC_state->study_chunk_recursed_bytes)
257 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
258 #define RExC_contains_locale (pRExC_state->contains_locale)
259 #define RExC_contains_i (pRExC_state->contains_i)
260 #define RExC_override_recoding (pRExC_state->override_recoding)
262 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
264 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
265 #define RExC_frame_head (pRExC_state->frame_head)
266 #define RExC_frame_last (pRExC_state->frame_last)
267 #define RExC_frame_count (pRExC_state->frame_count)
268 #define RExC_strict (pRExC_state->strict)
270 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
271 * a flag to disable back-off on the fixed/floating substrings - if it's
272 * a high complexity pattern we assume the benefit of avoiding a full match
273 * is worth the cost of checking for the substrings even if they rarely help.
275 #define RExC_naughty (pRExC_state->naughty)
276 #define TOO_NAUGHTY (10)
277 #define MARK_NAUGHTY(add) \
278 if (RExC_naughty < TOO_NAUGHTY) \
279 RExC_naughty += (add)
280 #define MARK_NAUGHTY_EXP(exp, add) \
281 if (RExC_naughty < TOO_NAUGHTY) \
282 RExC_naughty += RExC_naughty / (exp) + (add)
284 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
285 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
286 ((*s) == '{' && regcurly(s)))
289 * Flags to be passed up and down.
291 #define WORST 0 /* Worst case. */
292 #define HASWIDTH 0x01 /* Known to match non-null strings. */
294 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
295 * character. (There needs to be a case: in the switch statement in regexec.c
296 * for any node marked SIMPLE.) Note that this is not the same thing as
299 #define SPSTART 0x04 /* Starts with * or + */
300 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
301 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
302 #define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
304 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
306 /* whether trie related optimizations are enabled */
307 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
308 #define TRIE_STUDY_OPT
309 #define FULL_TRIE_STUDY
315 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
316 #define PBITVAL(paren) (1 << ((paren) & 7))
317 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
318 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
319 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
321 #define REQUIRE_UTF8 STMT_START { \
323 *flagp = RESTART_UTF8; \
328 /* This converts the named class defined in regcomp.h to its equivalent class
329 * number defined in handy.h. */
330 #define namedclass_to_classnum(class) ((int) ((class) / 2))
331 #define classnum_to_namedclass(classnum) ((classnum) * 2)
333 #define _invlist_union_complement_2nd(a, b, output) \
334 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
335 #define _invlist_intersection_complement_2nd(a, b, output) \
336 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
338 /* About scan_data_t.
340 During optimisation we recurse through the regexp program performing
341 various inplace (keyhole style) optimisations. In addition study_chunk
342 and scan_commit populate this data structure with information about
343 what strings MUST appear in the pattern. We look for the longest
344 string that must appear at a fixed location, and we look for the
345 longest string that may appear at a floating location. So for instance
350 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
351 strings (because they follow a .* construct). study_chunk will identify
352 both FOO and BAR as being the longest fixed and floating strings respectively.
354 The strings can be composites, for instance
358 will result in a composite fixed substring 'foo'.
360 For each string some basic information is maintained:
362 - offset or min_offset
363 This is the position the string must appear at, or not before.
364 It also implicitly (when combined with minlenp) tells us how many
365 characters must match before the string we are searching for.
366 Likewise when combined with minlenp and the length of the string it
367 tells us how many characters must appear after the string we have
371 Only used for floating strings. This is the rightmost point that
372 the string can appear at. If set to SSize_t_MAX it indicates that the
373 string can occur infinitely far to the right.
376 A pointer to the minimum number of characters of the pattern that the
377 string was found inside. This is important as in the case of positive
378 lookahead or positive lookbehind we can have multiple patterns
383 The minimum length of the pattern overall is 3, the minimum length
384 of the lookahead part is 3, but the minimum length of the part that
385 will actually match is 1. So 'FOO's minimum length is 3, but the
386 minimum length for the F is 1. This is important as the minimum length
387 is used to determine offsets in front of and behind the string being
388 looked for. Since strings can be composites this is the length of the
389 pattern at the time it was committed with a scan_commit. Note that
390 the length is calculated by study_chunk, so that the minimum lengths
391 are not known until the full pattern has been compiled, thus the
392 pointer to the value.
396 In the case of lookbehind the string being searched for can be
397 offset past the start point of the final matching string.
398 If this value was just blithely removed from the min_offset it would
399 invalidate some of the calculations for how many chars must match
400 before or after (as they are derived from min_offset and minlen and
401 the length of the string being searched for).
402 When the final pattern is compiled and the data is moved from the
403 scan_data_t structure into the regexp structure the information
404 about lookbehind is factored in, with the information that would
405 have been lost precalculated in the end_shift field for the
408 The fields pos_min and pos_delta are used to store the minimum offset
409 and the delta to the maximum offset at the current point in the pattern.
413 typedef struct scan_data_t {
414 /*I32 len_min; unused */
415 /*I32 len_delta; unused */
419 SSize_t last_end; /* min value, <0 unless valid. */
420 SSize_t last_start_min;
421 SSize_t last_start_max;
422 SV **longest; /* Either &l_fixed, or &l_float. */
423 SV *longest_fixed; /* longest fixed string found in pattern */
424 SSize_t offset_fixed; /* offset where it starts */
425 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
426 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
427 SV *longest_float; /* longest floating string found in pattern */
428 SSize_t offset_float_min; /* earliest point in string it can appear */
429 SSize_t offset_float_max; /* latest point in string it can appear */
430 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
431 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
434 SSize_t *last_closep;
435 regnode_ssc *start_class;
439 * Forward declarations for pregcomp()'s friends.
442 static const scan_data_t zero_scan_data =
443 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
445 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
446 #define SF_BEFORE_SEOL 0x0001
447 #define SF_BEFORE_MEOL 0x0002
448 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
449 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
451 #define SF_FIX_SHIFT_EOL (+2)
452 #define SF_FL_SHIFT_EOL (+4)
454 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
455 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
457 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
458 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
459 #define SF_IS_INF 0x0040
460 #define SF_HAS_PAR 0x0080
461 #define SF_IN_PAR 0x0100
462 #define SF_HAS_EVAL 0x0200
463 #define SCF_DO_SUBSTR 0x0400
464 #define SCF_DO_STCLASS_AND 0x0800
465 #define SCF_DO_STCLASS_OR 0x1000
466 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
467 #define SCF_WHILEM_VISITED_POS 0x2000
469 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
470 #define SCF_SEEN_ACCEPT 0x8000
471 #define SCF_TRIE_DOING_RESTUDY 0x10000
472 #define SCF_IN_DEFINE 0x20000
477 #define UTF cBOOL(RExC_utf8)
479 /* The enums for all these are ordered so things work out correctly */
480 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
481 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
482 == REGEX_DEPENDS_CHARSET)
483 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
484 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
485 >= REGEX_UNICODE_CHARSET)
486 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
487 == REGEX_ASCII_RESTRICTED_CHARSET)
488 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
489 >= REGEX_ASCII_RESTRICTED_CHARSET)
490 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
491 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
493 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
495 /* For programs that want to be strictly Unicode compatible by dying if any
496 * attempt is made to match a non-Unicode code point against a Unicode
498 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
500 #define OOB_NAMEDCLASS -1
502 /* There is no code point that is out-of-bounds, so this is problematic. But
503 * its only current use is to initialize a variable that is always set before
505 #define OOB_UNICODE 0xDEADBEEF
507 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
508 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
511 /* length of regex to show in messages that don't mark a position within */
512 #define RegexLengthToShowInErrorMessages 127
515 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
516 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
517 * op/pragma/warn/regcomp.
519 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
520 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
522 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
523 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
525 #define REPORT_LOCATION_ARGS(offset) \
526 UTF8fARG(UTF, offset, RExC_precomp), \
527 UTF8fARG(UTF, RExC_end - RExC_precomp - offset, RExC_precomp + offset)
529 /* Used to point after bad bytes for an error message, but avoid skipping
530 * past a nul byte. */
531 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
534 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
535 * arg. Show regex, up to a maximum length. If it's too long, chop and add
538 #define _FAIL(code) STMT_START { \
539 const char *ellipses = ""; \
540 IV len = RExC_end - RExC_precomp; \
543 SAVEFREESV(RExC_rx_sv); \
544 if (len > RegexLengthToShowInErrorMessages) { \
545 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
546 len = RegexLengthToShowInErrorMessages - 10; \
552 #define FAIL(msg) _FAIL( \
553 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
554 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
556 #define FAIL2(msg,arg) _FAIL( \
557 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
558 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
561 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
563 #define Simple_vFAIL(m) STMT_START { \
565 (RExC_parse > RExC_end ? RExC_end : RExC_parse) - RExC_precomp; \
566 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
567 m, REPORT_LOCATION_ARGS(offset)); \
571 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
573 #define vFAIL(m) STMT_START { \
575 SAVEFREESV(RExC_rx_sv); \
580 * Like Simple_vFAIL(), but accepts two arguments.
582 #define Simple_vFAIL2(m,a1) STMT_START { \
583 const IV offset = RExC_parse - RExC_precomp; \
584 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
585 REPORT_LOCATION_ARGS(offset)); \
589 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
591 #define vFAIL2(m,a1) STMT_START { \
593 SAVEFREESV(RExC_rx_sv); \
594 Simple_vFAIL2(m, a1); \
599 * Like Simple_vFAIL(), but accepts three arguments.
601 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
602 const IV offset = RExC_parse - RExC_precomp; \
603 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
604 REPORT_LOCATION_ARGS(offset)); \
608 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
610 #define vFAIL3(m,a1,a2) STMT_START { \
612 SAVEFREESV(RExC_rx_sv); \
613 Simple_vFAIL3(m, a1, a2); \
617 * Like Simple_vFAIL(), but accepts four arguments.
619 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
620 const IV offset = RExC_parse - RExC_precomp; \
621 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
622 REPORT_LOCATION_ARGS(offset)); \
625 #define vFAIL4(m,a1,a2,a3) STMT_START { \
627 SAVEFREESV(RExC_rx_sv); \
628 Simple_vFAIL4(m, a1, a2, a3); \
631 /* A specialized version of vFAIL2 that works with UTF8f */
632 #define vFAIL2utf8f(m, a1) STMT_START { \
633 const IV offset = RExC_parse - RExC_precomp; \
635 SAVEFREESV(RExC_rx_sv); \
636 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
637 REPORT_LOCATION_ARGS(offset)); \
640 /* These have asserts in them because of [perl #122671] Many warnings in
641 * regcomp.c can occur twice. If they get output in pass1 and later in that
642 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
643 * would get output again. So they should be output in pass2, and these
644 * asserts make sure new warnings follow that paradigm. */
646 /* m is not necessarily a "literal string", in this macro */
647 #define reg_warn_non_literal_string(loc, m) STMT_START { \
648 const IV offset = loc - RExC_precomp; \
649 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
650 m, REPORT_LOCATION_ARGS(offset)); \
653 #define ckWARNreg(loc,m) STMT_START { \
654 const IV offset = loc - RExC_precomp; \
655 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
656 REPORT_LOCATION_ARGS(offset)); \
659 #define vWARN(loc, m) STMT_START { \
660 const IV offset = loc - RExC_precomp; \
661 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
662 REPORT_LOCATION_ARGS(offset)); \
665 #define vWARN_dep(loc, m) STMT_START { \
666 const IV offset = loc - RExC_precomp; \
667 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
668 REPORT_LOCATION_ARGS(offset)); \
671 #define ckWARNdep(loc,m) STMT_START { \
672 const IV offset = loc - RExC_precomp; \
673 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
675 REPORT_LOCATION_ARGS(offset)); \
678 #define ckWARNregdep(loc,m) STMT_START { \
679 const IV offset = loc - RExC_precomp; \
680 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
682 REPORT_LOCATION_ARGS(offset)); \
685 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
686 const IV offset = loc - RExC_precomp; \
687 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
689 a1, REPORT_LOCATION_ARGS(offset)); \
692 #define ckWARN2reg(loc, m, a1) STMT_START { \
693 const IV offset = loc - RExC_precomp; \
694 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
695 a1, REPORT_LOCATION_ARGS(offset)); \
698 #define vWARN3(loc, m, a1, a2) STMT_START { \
699 const IV offset = loc - RExC_precomp; \
700 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
701 a1, a2, REPORT_LOCATION_ARGS(offset)); \
704 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
705 const IV offset = loc - RExC_precomp; \
706 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
707 a1, a2, REPORT_LOCATION_ARGS(offset)); \
710 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
711 const IV offset = loc - RExC_precomp; \
712 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
713 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
716 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
717 const IV offset = loc - RExC_precomp; \
718 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
719 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
722 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
723 const IV offset = loc - RExC_precomp; \
724 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
725 a1, a2, a3, a4, REPORT_LOCATION_ARGS(offset)); \
728 /* Macros for recording node offsets. 20001227 mjd@plover.com
729 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
730 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
731 * Element 0 holds the number n.
732 * Position is 1 indexed.
734 #ifndef RE_TRACK_PATTERN_OFFSETS
735 #define Set_Node_Offset_To_R(node,byte)
736 #define Set_Node_Offset(node,byte)
737 #define Set_Cur_Node_Offset
738 #define Set_Node_Length_To_R(node,len)
739 #define Set_Node_Length(node,len)
740 #define Set_Node_Cur_Length(node,start)
741 #define Node_Offset(n)
742 #define Node_Length(n)
743 #define Set_Node_Offset_Length(node,offset,len)
744 #define ProgLen(ri) ri->u.proglen
745 #define SetProgLen(ri,x) ri->u.proglen = x
747 #define ProgLen(ri) ri->u.offsets[0]
748 #define SetProgLen(ri,x) ri->u.offsets[0] = x
749 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
751 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
752 __LINE__, (int)(node), (int)(byte))); \
754 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
757 RExC_offsets[2*(node)-1] = (byte); \
762 #define Set_Node_Offset(node,byte) \
763 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
764 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
766 #define Set_Node_Length_To_R(node,len) STMT_START { \
768 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
769 __LINE__, (int)(node), (int)(len))); \
771 Perl_croak(aTHX_ "value of node is %d in Length macro", \
774 RExC_offsets[2*(node)] = (len); \
779 #define Set_Node_Length(node,len) \
780 Set_Node_Length_To_R((node)-RExC_emit_start, len)
781 #define Set_Node_Cur_Length(node, start) \
782 Set_Node_Length(node, RExC_parse - start)
784 /* Get offsets and lengths */
785 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
786 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
788 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
789 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
790 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
794 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
795 #define EXPERIMENTAL_INPLACESCAN
796 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
798 #define DEBUG_RExC_seen() \
799 DEBUG_OPTIMISE_MORE_r({ \
800 PerlIO_printf(Perl_debug_log,"RExC_seen: "); \
802 if (RExC_seen & REG_ZERO_LEN_SEEN) \
803 PerlIO_printf(Perl_debug_log,"REG_ZERO_LEN_SEEN "); \
805 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
806 PerlIO_printf(Perl_debug_log,"REG_LOOKBEHIND_SEEN "); \
808 if (RExC_seen & REG_GPOS_SEEN) \
809 PerlIO_printf(Perl_debug_log,"REG_GPOS_SEEN "); \
811 if (RExC_seen & REG_RECURSE_SEEN) \
812 PerlIO_printf(Perl_debug_log,"REG_RECURSE_SEEN "); \
814 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
815 PerlIO_printf(Perl_debug_log,"REG_TOP_LEVEL_BRANCHES_SEEN "); \
817 if (RExC_seen & REG_VERBARG_SEEN) \
818 PerlIO_printf(Perl_debug_log,"REG_VERBARG_SEEN "); \
820 if (RExC_seen & REG_CUTGROUP_SEEN) \
821 PerlIO_printf(Perl_debug_log,"REG_CUTGROUP_SEEN "); \
823 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
824 PerlIO_printf(Perl_debug_log,"REG_RUN_ON_COMMENT_SEEN "); \
826 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
827 PerlIO_printf(Perl_debug_log,"REG_UNFOLDED_MULTI_SEEN "); \
829 if (RExC_seen & REG_GOSTART_SEEN) \
830 PerlIO_printf(Perl_debug_log,"REG_GOSTART_SEEN "); \
832 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
833 PerlIO_printf(Perl_debug_log,"REG_UNBOUNDED_QUANTIFIER_SEEN "); \
835 PerlIO_printf(Perl_debug_log,"\n"); \
838 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
839 if ((flags) & flag) PerlIO_printf(Perl_debug_log, "%s ", #flag)
841 #define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
843 PerlIO_printf(Perl_debug_log, "%s", open_str); \
844 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
845 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
846 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
847 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
848 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
849 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
850 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
851 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
852 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
853 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
854 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
855 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
856 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
857 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
858 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
859 PerlIO_printf(Perl_debug_log, "%s", close_str); \
863 #define DEBUG_STUDYDATA(str,data,depth) \
864 DEBUG_OPTIMISE_MORE_r(if(data){ \
865 PerlIO_printf(Perl_debug_log, \
866 "%*s" str "Pos:%"IVdf"/%"IVdf \
868 (int)(depth)*2, "", \
869 (IV)((data)->pos_min), \
870 (IV)((data)->pos_delta), \
871 (UV)((data)->flags) \
873 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
874 PerlIO_printf(Perl_debug_log, \
875 " Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
876 (IV)((data)->whilem_c), \
877 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
878 is_inf ? "INF " : "" \
880 if ((data)->last_found) \
881 PerlIO_printf(Perl_debug_log, \
882 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
883 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
884 SvPVX_const((data)->last_found), \
885 (IV)((data)->last_end), \
886 (IV)((data)->last_start_min), \
887 (IV)((data)->last_start_max), \
888 ((data)->longest && \
889 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
890 SvPVX_const((data)->longest_fixed), \
891 (IV)((data)->offset_fixed), \
892 ((data)->longest && \
893 (data)->longest==&((data)->longest_float)) ? "*" : "", \
894 SvPVX_const((data)->longest_float), \
895 (IV)((data)->offset_float_min), \
896 (IV)((data)->offset_float_max) \
898 PerlIO_printf(Perl_debug_log,"\n"); \
901 /* is c a control character for which we have a mnemonic? */
902 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
905 S_cntrl_to_mnemonic(const U8 c)
907 /* Returns the mnemonic string that represents character 'c', if one
908 * exists; NULL otherwise. The only ones that exist for the purposes of
909 * this routine are a few control characters */
912 case '\a': return "\\a";
913 case '\b': return "\\b";
914 case ESC_NATIVE: return "\\e";
915 case '\f': return "\\f";
916 case '\n': return "\\n";
917 case '\r': return "\\r";
918 case '\t': return "\\t";
924 /* Mark that we cannot extend a found fixed substring at this point.
925 Update the longest found anchored substring and the longest found
926 floating substrings if needed. */
929 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
930 SSize_t *minlenp, int is_inf)
932 const STRLEN l = CHR_SVLEN(data->last_found);
933 const STRLEN old_l = CHR_SVLEN(*data->longest);
934 GET_RE_DEBUG_FLAGS_DECL;
936 PERL_ARGS_ASSERT_SCAN_COMMIT;
938 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
939 SvSetMagicSV(*data->longest, data->last_found);
940 if (*data->longest == data->longest_fixed) {
941 data->offset_fixed = l ? data->last_start_min : data->pos_min;
942 if (data->flags & SF_BEFORE_EOL)
944 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
946 data->flags &= ~SF_FIX_BEFORE_EOL;
947 data->minlen_fixed=minlenp;
948 data->lookbehind_fixed=0;
950 else { /* *data->longest == data->longest_float */
951 data->offset_float_min = l ? data->last_start_min : data->pos_min;
952 data->offset_float_max = (l
953 ? data->last_start_max
954 : (data->pos_delta > SSize_t_MAX - data->pos_min
956 : data->pos_min + data->pos_delta));
958 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
959 data->offset_float_max = SSize_t_MAX;
960 if (data->flags & SF_BEFORE_EOL)
962 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
964 data->flags &= ~SF_FL_BEFORE_EOL;
965 data->minlen_float=minlenp;
966 data->lookbehind_float=0;
969 SvCUR_set(data->last_found, 0);
971 SV * const sv = data->last_found;
972 if (SvUTF8(sv) && SvMAGICAL(sv)) {
973 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
979 data->flags &= ~SF_BEFORE_EOL;
980 DEBUG_STUDYDATA("commit: ",data,0);
983 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
984 * list that describes which code points it matches */
987 S_ssc_anything(pTHX_ regnode_ssc *ssc)
989 /* Set the SSC 'ssc' to match an empty string or any code point */
991 PERL_ARGS_ASSERT_SSC_ANYTHING;
993 assert(is_ANYOF_SYNTHETIC(ssc));
995 ssc->invlist = sv_2mortal(_new_invlist(2)); /* mortalize so won't leak */
996 _append_range_to_invlist(ssc->invlist, 0, UV_MAX);
997 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1001 S_ssc_is_anything(const regnode_ssc *ssc)
1003 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1004 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1005 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1006 * in any way, so there's no point in using it */
1011 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1013 assert(is_ANYOF_SYNTHETIC(ssc));
1015 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1019 /* See if the list consists solely of the range 0 - Infinity */
1020 invlist_iterinit(ssc->invlist);
1021 ret = invlist_iternext(ssc->invlist, &start, &end)
1025 invlist_iterfinish(ssc->invlist);
1031 /* If e.g., both \w and \W are set, matches everything */
1032 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1034 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1035 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1045 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1047 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1048 * string, any code point, or any posix class under locale */
1050 PERL_ARGS_ASSERT_SSC_INIT;
1052 Zero(ssc, 1, regnode_ssc);
1053 set_ANYOF_SYNTHETIC(ssc);
1054 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1057 /* If any portion of the regex is to operate under locale rules that aren't
1058 * fully known at compile time, initialization includes it. The reason
1059 * this isn't done for all regexes is that the optimizer was written under
1060 * the assumption that locale was all-or-nothing. Given the complexity and
1061 * lack of documentation in the optimizer, and that there are inadequate
1062 * test cases for locale, many parts of it may not work properly, it is
1063 * safest to avoid locale unless necessary. */
1064 if (RExC_contains_locale) {
1065 ANYOF_POSIXL_SETALL(ssc);
1068 ANYOF_POSIXL_ZERO(ssc);
1073 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1074 const regnode_ssc *ssc)
1076 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1077 * to the list of code points matched, and locale posix classes; hence does
1078 * not check its flags) */
1083 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1085 assert(is_ANYOF_SYNTHETIC(ssc));
1087 invlist_iterinit(ssc->invlist);
1088 ret = invlist_iternext(ssc->invlist, &start, &end)
1092 invlist_iterfinish(ssc->invlist);
1098 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1106 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1107 const regnode_charclass* const node)
1109 /* Returns a mortal inversion list defining which code points are matched
1110 * by 'node', which is of type ANYOF. Handles complementing the result if
1111 * appropriate. If some code points aren't knowable at this time, the
1112 * returned list must, and will, contain every code point that is a
1115 SV* invlist = sv_2mortal(_new_invlist(0));
1116 SV* only_utf8_locale_invlist = NULL;
1118 const U32 n = ARG(node);
1119 bool new_node_has_latin1 = FALSE;
1121 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1123 /* Look at the data structure created by S_set_ANYOF_arg() */
1124 if (n != ANYOF_ONLY_HAS_BITMAP) {
1125 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1126 AV * const av = MUTABLE_AV(SvRV(rv));
1127 SV **const ary = AvARRAY(av);
1128 assert(RExC_rxi->data->what[n] == 's');
1130 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1131 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1133 else if (ary[0] && ary[0] != &PL_sv_undef) {
1135 /* Here, no compile-time swash, and there are things that won't be
1136 * known until runtime -- we have to assume it could be anything */
1137 return _add_range_to_invlist(invlist, 0, UV_MAX);
1139 else if (ary[3] && ary[3] != &PL_sv_undef) {
1141 /* Here no compile-time swash, and no run-time only data. Use the
1142 * node's inversion list */
1143 invlist = sv_2mortal(invlist_clone(ary[3]));
1146 /* Get the code points valid only under UTF-8 locales */
1147 if ((ANYOF_FLAGS(node) & ANYOF_LOC_FOLD)
1148 && ary[2] && ary[2] != &PL_sv_undef)
1150 only_utf8_locale_invlist = ary[2];
1154 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1155 * code points, and an inversion list for the others, but if there are code
1156 * points that should match only conditionally on the target string being
1157 * UTF-8, those are placed in the inversion list, and not the bitmap.
1158 * Since there are circumstances under which they could match, they are
1159 * included in the SSC. But if the ANYOF node is to be inverted, we have
1160 * to exclude them here, so that when we invert below, the end result
1161 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1162 * have to do this here before we add the unconditionally matched code
1164 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1165 _invlist_intersection_complement_2nd(invlist,
1170 /* Add in the points from the bit map */
1171 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1172 if (ANYOF_BITMAP_TEST(node, i)) {
1173 invlist = add_cp_to_invlist(invlist, i);
1174 new_node_has_latin1 = TRUE;
1178 /* If this can match all upper Latin1 code points, have to add them
1180 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_NON_UTF8_NON_ASCII) {
1181 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1184 /* Similarly for these */
1185 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1186 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1189 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1190 _invlist_invert(invlist);
1192 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOF_LOC_FOLD) {
1194 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1195 * locale. We can skip this if there are no 0-255 at all. */
1196 _invlist_union(invlist, PL_Latin1, &invlist);
1199 /* Similarly add the UTF-8 locale possible matches. These have to be
1200 * deferred until after the non-UTF-8 locale ones are taken care of just
1201 * above, or it leads to wrong results under ANYOF_INVERT */
1202 if (only_utf8_locale_invlist) {
1203 _invlist_union_maybe_complement_2nd(invlist,
1204 only_utf8_locale_invlist,
1205 ANYOF_FLAGS(node) & ANYOF_INVERT,
1212 /* These two functions currently do the exact same thing */
1213 #define ssc_init_zero ssc_init
1215 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1216 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1218 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1219 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1220 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1223 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1224 const regnode_charclass *and_with)
1226 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1227 * another SSC or a regular ANYOF class. Can create false positives. */
1232 PERL_ARGS_ASSERT_SSC_AND;
1234 assert(is_ANYOF_SYNTHETIC(ssc));
1236 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1237 * the code point inversion list and just the relevant flags */
1238 if (is_ANYOF_SYNTHETIC(and_with)) {
1239 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1240 anded_flags = ANYOF_FLAGS(and_with);
1242 /* XXX This is a kludge around what appears to be deficiencies in the
1243 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1244 * there are paths through the optimizer where it doesn't get weeded
1245 * out when it should. And if we don't make some extra provision for
1246 * it like the code just below, it doesn't get added when it should.
1247 * This solution is to add it only when AND'ing, which is here, and
1248 * only when what is being AND'ed is the pristine, original node
1249 * matching anything. Thus it is like adding it to ssc_anything() but
1250 * only when the result is to be AND'ed. Probably the same solution
1251 * could be adopted for the same problem we have with /l matching,
1252 * which is solved differently in S_ssc_init(), and that would lead to
1253 * fewer false positives than that solution has. But if this solution
1254 * creates bugs, the consequences are only that a warning isn't raised
1255 * that should be; while the consequences for having /l bugs is
1256 * incorrect matches */
1257 if (ssc_is_anything((regnode_ssc *)and_with)) {
1258 anded_flags |= ANYOF_WARN_SUPER;
1262 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1263 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1266 ANYOF_FLAGS(ssc) &= anded_flags;
1268 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1269 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1270 * 'and_with' may be inverted. When not inverted, we have the situation of
1272 * (C1 | P1) & (C2 | P2)
1273 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1274 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1275 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1276 * <= ((C1 & C2) | P1 | P2)
1277 * Alternatively, the last few steps could be:
1278 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1279 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1280 * <= (C1 | C2 | (P1 & P2))
1281 * We favor the second approach if either P1 or P2 is non-empty. This is
1282 * because these components are a barrier to doing optimizations, as what
1283 * they match cannot be known until the moment of matching as they are
1284 * dependent on the current locale, 'AND"ing them likely will reduce or
1286 * But we can do better if we know that C1,P1 are in their initial state (a
1287 * frequent occurrence), each matching everything:
1288 * (<everything>) & (C2 | P2) = C2 | P2
1289 * Similarly, if C2,P2 are in their initial state (again a frequent
1290 * occurrence), the result is a no-op
1291 * (C1 | P1) & (<everything>) = C1 | P1
1294 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1295 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1296 * <= (C1 & ~C2) | (P1 & ~P2)
1299 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1300 && ! is_ANYOF_SYNTHETIC(and_with))
1304 ssc_intersection(ssc,
1306 FALSE /* Has already been inverted */
1309 /* If either P1 or P2 is empty, the intersection will be also; can skip
1311 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1312 ANYOF_POSIXL_ZERO(ssc);
1314 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1316 /* Note that the Posix class component P from 'and_with' actually
1318 * P = Pa | Pb | ... | Pn
1319 * where each component is one posix class, such as in [\w\s].
1321 * ~P = ~(Pa | Pb | ... | Pn)
1322 * = ~Pa & ~Pb & ... & ~Pn
1323 * <= ~Pa | ~Pb | ... | ~Pn
1324 * The last is something we can easily calculate, but unfortunately
1325 * is likely to have many false positives. We could do better
1326 * in some (but certainly not all) instances if two classes in
1327 * P have known relationships. For example
1328 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1330 * :lower: & :print: = :lower:
1331 * And similarly for classes that must be disjoint. For example,
1332 * since \s and \w can have no elements in common based on rules in
1333 * the POSIX standard,
1334 * \w & ^\S = nothing
1335 * Unfortunately, some vendor locales do not meet the Posix
1336 * standard, in particular almost everything by Microsoft.
1337 * The loop below just changes e.g., \w into \W and vice versa */
1339 regnode_charclass_posixl temp;
1340 int add = 1; /* To calculate the index of the complement */
1342 ANYOF_POSIXL_ZERO(&temp);
1343 for (i = 0; i < ANYOF_MAX; i++) {
1345 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1346 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1348 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1349 ANYOF_POSIXL_SET(&temp, i + add);
1351 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1353 ANYOF_POSIXL_AND(&temp, ssc);
1355 } /* else ssc already has no posixes */
1356 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1357 in its initial state */
1358 else if (! is_ANYOF_SYNTHETIC(and_with)
1359 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1361 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1362 * copy it over 'ssc' */
1363 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1364 if (is_ANYOF_SYNTHETIC(and_with)) {
1365 StructCopy(and_with, ssc, regnode_ssc);
1368 ssc->invlist = anded_cp_list;
1369 ANYOF_POSIXL_ZERO(ssc);
1370 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1371 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1375 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1376 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1378 /* One or the other of P1, P2 is non-empty. */
1379 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1380 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1382 ssc_union(ssc, anded_cp_list, FALSE);
1384 else { /* P1 = P2 = empty */
1385 ssc_intersection(ssc, anded_cp_list, FALSE);
1391 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1392 const regnode_charclass *or_with)
1394 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1395 * another SSC or a regular ANYOF class. Can create false positives if
1396 * 'or_with' is to be inverted. */
1401 PERL_ARGS_ASSERT_SSC_OR;
1403 assert(is_ANYOF_SYNTHETIC(ssc));
1405 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1406 * the code point inversion list and just the relevant flags */
1407 if (is_ANYOF_SYNTHETIC(or_with)) {
1408 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1409 ored_flags = ANYOF_FLAGS(or_with);
1412 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1413 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1416 ANYOF_FLAGS(ssc) |= ored_flags;
1418 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1419 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1420 * 'or_with' may be inverted. When not inverted, we have the simple
1421 * situation of computing:
1422 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1423 * If P1|P2 yields a situation with both a class and its complement are
1424 * set, like having both \w and \W, this matches all code points, and we
1425 * can delete these from the P component of the ssc going forward. XXX We
1426 * might be able to delete all the P components, but I (khw) am not certain
1427 * about this, and it is better to be safe.
1430 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1431 * <= (C1 | P1) | ~C2
1432 * <= (C1 | ~C2) | P1
1433 * (which results in actually simpler code than the non-inverted case)
1436 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1437 && ! is_ANYOF_SYNTHETIC(or_with))
1439 /* We ignore P2, leaving P1 going forward */
1440 } /* else Not inverted */
1441 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1442 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1443 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1445 for (i = 0; i < ANYOF_MAX; i += 2) {
1446 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1448 ssc_match_all_cp(ssc);
1449 ANYOF_POSIXL_CLEAR(ssc, i);
1450 ANYOF_POSIXL_CLEAR(ssc, i+1);
1458 FALSE /* Already has been inverted */
1462 PERL_STATIC_INLINE void
1463 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1465 PERL_ARGS_ASSERT_SSC_UNION;
1467 assert(is_ANYOF_SYNTHETIC(ssc));
1469 _invlist_union_maybe_complement_2nd(ssc->invlist,
1475 PERL_STATIC_INLINE void
1476 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1478 const bool invert2nd)
1480 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1482 assert(is_ANYOF_SYNTHETIC(ssc));
1484 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1490 PERL_STATIC_INLINE void
1491 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1493 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1495 assert(is_ANYOF_SYNTHETIC(ssc));
1497 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1500 PERL_STATIC_INLINE void
1501 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1503 /* AND just the single code point 'cp' into the SSC 'ssc' */
1505 SV* cp_list = _new_invlist(2);
1507 PERL_ARGS_ASSERT_SSC_CP_AND;
1509 assert(is_ANYOF_SYNTHETIC(ssc));
1511 cp_list = add_cp_to_invlist(cp_list, cp);
1512 ssc_intersection(ssc, cp_list,
1513 FALSE /* Not inverted */
1515 SvREFCNT_dec_NN(cp_list);
1518 PERL_STATIC_INLINE void
1519 S_ssc_clear_locale(regnode_ssc *ssc)
1521 /* Set the SSC 'ssc' to not match any locale things */
1522 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1524 assert(is_ANYOF_SYNTHETIC(ssc));
1526 ANYOF_POSIXL_ZERO(ssc);
1527 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1530 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1533 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1535 /* The synthetic start class is used to hopefully quickly winnow down
1536 * places where a pattern could start a match in the target string. If it
1537 * doesn't really narrow things down that much, there isn't much point to
1538 * having the overhead of using it. This function uses some very crude
1539 * heuristics to decide if to use the ssc or not.
1541 * It returns TRUE if 'ssc' rules out more than half what it considers to
1542 * be the "likely" possible matches, but of course it doesn't know what the
1543 * actual things being matched are going to be; these are only guesses
1545 * For /l matches, it assumes that the only likely matches are going to be
1546 * in the 0-255 range, uniformly distributed, so half of that is 127
1547 * For /a and /d matches, it assumes that the likely matches will be just
1548 * the ASCII range, so half of that is 63
1549 * For /u and there isn't anything matching above the Latin1 range, it
1550 * assumes that that is the only range likely to be matched, and uses
1551 * half that as the cut-off: 127. If anything matches above Latin1,
1552 * it assumes that all of Unicode could match (uniformly), except for
1553 * non-Unicode code points and things in the General Category "Other"
1554 * (unassigned, private use, surrogates, controls and formats). This
1555 * is a much large number. */
1557 const U32 max_match = (LOC)
1561 : (invlist_highest(ssc->invlist) < 256)
1563 : ((NON_OTHER_COUNT + 1) / 2) - 1;
1564 U32 count = 0; /* Running total of number of code points matched by
1566 UV start, end; /* Start and end points of current range in inversion
1569 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1571 invlist_iterinit(ssc->invlist);
1572 while (invlist_iternext(ssc->invlist, &start, &end)) {
1574 /* /u is the only thing that we expect to match above 255; so if not /u
1575 * and even if there are matches above 255, ignore them. This catches
1576 * things like \d under /d which does match the digits above 255, but
1577 * since the pattern is /d, it is not likely to be expecting them */
1578 if (! UNI_SEMANTICS) {
1582 end = MIN(end, 255);
1584 count += end - start + 1;
1585 if (count > max_match) {
1586 invlist_iterfinish(ssc->invlist);
1596 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1598 /* The inversion list in the SSC is marked mortal; now we need a more
1599 * permanent copy, which is stored the same way that is done in a regular
1600 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1603 SV* invlist = invlist_clone(ssc->invlist);
1605 PERL_ARGS_ASSERT_SSC_FINALIZE;
1607 assert(is_ANYOF_SYNTHETIC(ssc));
1609 /* The code in this file assumes that all but these flags aren't relevant
1610 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1611 * by the time we reach here */
1612 assert(! (ANYOF_FLAGS(ssc) & ~ANYOF_COMMON_FLAGS));
1614 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1616 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1617 NULL, NULL, NULL, FALSE);
1619 /* Make sure is clone-safe */
1620 ssc->invlist = NULL;
1622 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1623 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
1626 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1629 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1630 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1631 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1632 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1633 ? (TRIE_LIST_CUR( idx ) - 1) \
1639 dump_trie(trie,widecharmap,revcharmap)
1640 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1641 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1643 These routines dump out a trie in a somewhat readable format.
1644 The _interim_ variants are used for debugging the interim
1645 tables that are used to generate the final compressed
1646 representation which is what dump_trie expects.
1648 Part of the reason for their existence is to provide a form
1649 of documentation as to how the different representations function.
1654 Dumps the final compressed table form of the trie to Perl_debug_log.
1655 Used for debugging make_trie().
1659 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1660 AV *revcharmap, U32 depth)
1663 SV *sv=sv_newmortal();
1664 int colwidth= widecharmap ? 6 : 4;
1666 GET_RE_DEBUG_FLAGS_DECL;
1668 PERL_ARGS_ASSERT_DUMP_TRIE;
1670 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1671 (int)depth * 2 + 2,"",
1672 "Match","Base","Ofs" );
1674 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1675 SV ** const tmp = av_fetch( revcharmap, state, 0);
1677 PerlIO_printf( Perl_debug_log, "%*s",
1679 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1680 PL_colors[0], PL_colors[1],
1681 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1682 PERL_PV_ESCAPE_FIRSTCHAR
1687 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1688 (int)depth * 2 + 2,"");
1690 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1691 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1692 PerlIO_printf( Perl_debug_log, "\n");
1694 for( state = 1 ; state < trie->statecount ; state++ ) {
1695 const U32 base = trie->states[ state ].trans.base;
1697 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1698 (int)depth * 2 + 2,"", (UV)state);
1700 if ( trie->states[ state ].wordnum ) {
1701 PerlIO_printf( Perl_debug_log, " W%4X",
1702 trie->states[ state ].wordnum );
1704 PerlIO_printf( Perl_debug_log, "%6s", "" );
1707 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1712 while( ( base + ofs < trie->uniquecharcount ) ||
1713 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1714 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1718 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1720 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1721 if ( ( base + ofs >= trie->uniquecharcount )
1722 && ( base + ofs - trie->uniquecharcount
1724 && trie->trans[ base + ofs
1725 - trie->uniquecharcount ].check == state )
1727 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1729 (UV)trie->trans[ base + ofs
1730 - trie->uniquecharcount ].next );
1732 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1736 PerlIO_printf( Perl_debug_log, "]");
1739 PerlIO_printf( Perl_debug_log, "\n" );
1741 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1743 for (word=1; word <= trie->wordcount; word++) {
1744 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1745 (int)word, (int)(trie->wordinfo[word].prev),
1746 (int)(trie->wordinfo[word].len));
1748 PerlIO_printf(Perl_debug_log, "\n" );
1751 Dumps a fully constructed but uncompressed trie in list form.
1752 List tries normally only are used for construction when the number of
1753 possible chars (trie->uniquecharcount) is very high.
1754 Used for debugging make_trie().
1757 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1758 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1762 SV *sv=sv_newmortal();
1763 int colwidth= widecharmap ? 6 : 4;
1764 GET_RE_DEBUG_FLAGS_DECL;
1766 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1768 /* print out the table precompression. */
1769 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1770 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1771 "------:-----+-----------------\n" );
1773 for( state=1 ; state < next_alloc ; state ++ ) {
1776 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1777 (int)depth * 2 + 2,"", (UV)state );
1778 if ( ! trie->states[ state ].wordnum ) {
1779 PerlIO_printf( Perl_debug_log, "%5s| ","");
1781 PerlIO_printf( Perl_debug_log, "W%4x| ",
1782 trie->states[ state ].wordnum
1785 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1786 SV ** const tmp = av_fetch( revcharmap,
1787 TRIE_LIST_ITEM(state,charid).forid, 0);
1789 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1791 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1793 PL_colors[0], PL_colors[1],
1794 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1795 | PERL_PV_ESCAPE_FIRSTCHAR
1797 TRIE_LIST_ITEM(state,charid).forid,
1798 (UV)TRIE_LIST_ITEM(state,charid).newstate
1801 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1802 (int)((depth * 2) + 14), "");
1805 PerlIO_printf( Perl_debug_log, "\n");
1810 Dumps a fully constructed but uncompressed trie in table form.
1811 This is the normal DFA style state transition table, with a few
1812 twists to facilitate compression later.
1813 Used for debugging make_trie().
1816 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1817 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1822 SV *sv=sv_newmortal();
1823 int colwidth= widecharmap ? 6 : 4;
1824 GET_RE_DEBUG_FLAGS_DECL;
1826 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1829 print out the table precompression so that we can do a visual check
1830 that they are identical.
1833 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1835 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1836 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1838 PerlIO_printf( Perl_debug_log, "%*s",
1840 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1841 PL_colors[0], PL_colors[1],
1842 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1843 PERL_PV_ESCAPE_FIRSTCHAR
1849 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1851 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1852 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1855 PerlIO_printf( Perl_debug_log, "\n" );
1857 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1859 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1860 (int)depth * 2 + 2,"",
1861 (UV)TRIE_NODENUM( state ) );
1863 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1864 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1866 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1868 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1870 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1871 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1872 (UV)trie->trans[ state ].check );
1874 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1875 (UV)trie->trans[ state ].check,
1876 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1884 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1885 startbranch: the first branch in the whole branch sequence
1886 first : start branch of sequence of branch-exact nodes.
1887 May be the same as startbranch
1888 last : Thing following the last branch.
1889 May be the same as tail.
1890 tail : item following the branch sequence
1891 count : words in the sequence
1892 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
1893 depth : indent depth
1895 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1897 A trie is an N'ary tree where the branches are determined by digital
1898 decomposition of the key. IE, at the root node you look up the 1st character and
1899 follow that branch repeat until you find the end of the branches. Nodes can be
1900 marked as "accepting" meaning they represent a complete word. Eg:
1904 would convert into the following structure. Numbers represent states, letters
1905 following numbers represent valid transitions on the letter from that state, if
1906 the number is in square brackets it represents an accepting state, otherwise it
1907 will be in parenthesis.
1909 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1913 (1) +-i->(6)-+-s->[7]
1915 +-s->(3)-+-h->(4)-+-e->[5]
1917 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1919 This shows that when matching against the string 'hers' we will begin at state 1
1920 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1921 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1922 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1923 single traverse. We store a mapping from accepting to state to which word was
1924 matched, and then when we have multiple possibilities we try to complete the
1925 rest of the regex in the order in which they occurred in the alternation.
1927 The only prior NFA like behaviour that would be changed by the TRIE support is
1928 the silent ignoring of duplicate alternations which are of the form:
1930 / (DUPE|DUPE) X? (?{ ... }) Y /x
1932 Thus EVAL blocks following a trie may be called a different number of times with
1933 and without the optimisation. With the optimisations dupes will be silently
1934 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1935 the following demonstrates:
1937 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1939 which prints out 'word' three times, but
1941 'words'=~/(word|word|word)(?{ print $1 })S/
1943 which doesnt print it out at all. This is due to other optimisations kicking in.
1945 Example of what happens on a structural level:
1947 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1949 1: CURLYM[1] {1,32767}(18)
1960 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1961 and should turn into:
1963 1: CURLYM[1] {1,32767}(18)
1965 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1973 Cases where tail != last would be like /(?foo|bar)baz/:
1983 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1984 and would end up looking like:
1987 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1994 d = uvchr_to_utf8_flags(d, uv, 0);
1996 is the recommended Unicode-aware way of saying
2001 #define TRIE_STORE_REVCHAR(val) \
2004 SV *zlopp = newSV(7); /* XXX: optimize me */ \
2005 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2006 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2007 SvCUR_set(zlopp, kapow - flrbbbbb); \
2010 av_push(revcharmap, zlopp); \
2012 char ooooff = (char)val; \
2013 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2017 /* This gets the next character from the input, folding it if not already
2019 #define TRIE_READ_CHAR STMT_START { \
2022 /* if it is UTF then it is either already folded, or does not need \
2024 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2026 else if (folder == PL_fold_latin1) { \
2027 /* This folder implies Unicode rules, which in the range expressible \
2028 * by not UTF is the lower case, with the two exceptions, one of \
2029 * which should have been taken care of before calling this */ \
2030 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2031 uvc = toLOWER_L1(*uc); \
2032 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2035 /* raw data, will be folded later if needed */ \
2043 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2044 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2045 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2046 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2048 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2049 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2050 TRIE_LIST_CUR( state )++; \
2053 #define TRIE_LIST_NEW(state) STMT_START { \
2054 Newxz( trie->states[ state ].trans.list, \
2055 4, reg_trie_trans_le ); \
2056 TRIE_LIST_CUR( state ) = 1; \
2057 TRIE_LIST_LEN( state ) = 4; \
2060 #define TRIE_HANDLE_WORD(state) STMT_START { \
2061 U16 dupe= trie->states[ state ].wordnum; \
2062 regnode * const noper_next = regnext( noper ); \
2065 /* store the word for dumping */ \
2067 if (OP(noper) != NOTHING) \
2068 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2070 tmp = newSVpvn_utf8( "", 0, UTF ); \
2071 av_push( trie_words, tmp ); \
2075 trie->wordinfo[curword].prev = 0; \
2076 trie->wordinfo[curword].len = wordlen; \
2077 trie->wordinfo[curword].accept = state; \
2079 if ( noper_next < tail ) { \
2081 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2083 trie->jump[curword] = (U16)(noper_next - convert); \
2085 jumper = noper_next; \
2087 nextbranch= regnext(cur); \
2091 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2092 /* chain, so that when the bits of chain are later */\
2093 /* linked together, the dups appear in the chain */\
2094 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2095 trie->wordinfo[dupe].prev = curword; \
2097 /* we haven't inserted this word yet. */ \
2098 trie->states[ state ].wordnum = curword; \
2103 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2104 ( ( base + charid >= ucharcount \
2105 && base + charid < ubound \
2106 && state == trie->trans[ base - ucharcount + charid ].check \
2107 && trie->trans[ base - ucharcount + charid ].next ) \
2108 ? trie->trans[ base - ucharcount + charid ].next \
2109 : ( state==1 ? special : 0 ) \
2113 #define MADE_JUMP_TRIE 2
2114 #define MADE_EXACT_TRIE 4
2117 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2118 regnode *first, regnode *last, regnode *tail,
2119 U32 word_count, U32 flags, U32 depth)
2121 /* first pass, loop through and scan words */
2122 reg_trie_data *trie;
2123 HV *widecharmap = NULL;
2124 AV *revcharmap = newAV();
2130 regnode *jumper = NULL;
2131 regnode *nextbranch = NULL;
2132 regnode *convert = NULL;
2133 U32 *prev_states; /* temp array mapping each state to previous one */
2134 /* we just use folder as a flag in utf8 */
2135 const U8 * folder = NULL;
2138 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2139 AV *trie_words = NULL;
2140 /* along with revcharmap, this only used during construction but both are
2141 * useful during debugging so we store them in the struct when debugging.
2144 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2145 STRLEN trie_charcount=0;
2147 SV *re_trie_maxbuff;
2148 GET_RE_DEBUG_FLAGS_DECL;
2150 PERL_ARGS_ASSERT_MAKE_TRIE;
2152 PERL_UNUSED_ARG(depth);
2156 case EXACT: case EXACTL: break;
2160 case EXACTFLU8: folder = PL_fold_latin1; break;
2161 case EXACTF: folder = PL_fold; break;
2162 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2165 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2167 trie->startstate = 1;
2168 trie->wordcount = word_count;
2169 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2170 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2171 if (flags == EXACT || flags == EXACTL)
2172 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2173 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2174 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2177 trie_words = newAV();
2180 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2181 assert(re_trie_maxbuff);
2182 if (!SvIOK(re_trie_maxbuff)) {
2183 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2185 DEBUG_TRIE_COMPILE_r({
2186 PerlIO_printf( Perl_debug_log,
2187 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2188 (int)depth * 2 + 2, "",
2189 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2190 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2193 /* Find the node we are going to overwrite */
2194 if ( first == startbranch && OP( last ) != BRANCH ) {
2195 /* whole branch chain */
2198 /* branch sub-chain */
2199 convert = NEXTOPER( first );
2202 /* -- First loop and Setup --
2204 We first traverse the branches and scan each word to determine if it
2205 contains widechars, and how many unique chars there are, this is
2206 important as we have to build a table with at least as many columns as we
2209 We use an array of integers to represent the character codes 0..255
2210 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2211 the native representation of the character value as the key and IV's for
2214 *TODO* If we keep track of how many times each character is used we can
2215 remap the columns so that the table compression later on is more
2216 efficient in terms of memory by ensuring the most common value is in the
2217 middle and the least common are on the outside. IMO this would be better
2218 than a most to least common mapping as theres a decent chance the most
2219 common letter will share a node with the least common, meaning the node
2220 will not be compressible. With a middle is most common approach the worst
2221 case is when we have the least common nodes twice.
2225 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2226 regnode *noper = NEXTOPER( cur );
2227 const U8 *uc = (U8*)STRING( noper );
2228 const U8 *e = uc + STR_LEN( noper );
2230 U32 wordlen = 0; /* required init */
2231 STRLEN minchars = 0;
2232 STRLEN maxchars = 0;
2233 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2236 if (OP(noper) == NOTHING) {
2237 regnode *noper_next= regnext(noper);
2238 if (noper_next != tail && OP(noper_next) == flags) {
2240 uc= (U8*)STRING(noper);
2241 e= uc + STR_LEN(noper);
2242 trie->minlen= STR_LEN(noper);
2249 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2250 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2251 regardless of encoding */
2252 if (OP( noper ) == EXACTFU_SS) {
2253 /* false positives are ok, so just set this */
2254 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2257 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2259 TRIE_CHARCOUNT(trie)++;
2262 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2263 * is in effect. Under /i, this character can match itself, or
2264 * anything that folds to it. If not under /i, it can match just
2265 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2266 * all fold to k, and all are single characters. But some folds
2267 * expand to more than one character, so for example LATIN SMALL
2268 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2269 * the string beginning at 'uc' is 'ffi', it could be matched by
2270 * three characters, or just by the one ligature character. (It
2271 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2272 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2273 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2274 * match.) The trie needs to know the minimum and maximum number
2275 * of characters that could match so that it can use size alone to
2276 * quickly reject many match attempts. The max is simple: it is
2277 * the number of folded characters in this branch (since a fold is
2278 * never shorter than what folds to it. */
2282 /* And the min is equal to the max if not under /i (indicated by
2283 * 'folder' being NULL), or there are no multi-character folds. If
2284 * there is a multi-character fold, the min is incremented just
2285 * once, for the character that folds to the sequence. Each
2286 * character in the sequence needs to be added to the list below of
2287 * characters in the trie, but we count only the first towards the
2288 * min number of characters needed. This is done through the
2289 * variable 'foldlen', which is returned by the macros that look
2290 * for these sequences as the number of bytes the sequence
2291 * occupies. Each time through the loop, we decrement 'foldlen' by
2292 * how many bytes the current char occupies. Only when it reaches
2293 * 0 do we increment 'minchars' or look for another multi-character
2295 if (folder == NULL) {
2298 else if (foldlen > 0) {
2299 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2304 /* See if *uc is the beginning of a multi-character fold. If
2305 * so, we decrement the length remaining to look at, to account
2306 * for the current character this iteration. (We can use 'uc'
2307 * instead of the fold returned by TRIE_READ_CHAR because for
2308 * non-UTF, the latin1_safe macro is smart enough to account
2309 * for all the unfolded characters, and because for UTF, the
2310 * string will already have been folded earlier in the
2311 * compilation process */
2313 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2314 foldlen -= UTF8SKIP(uc);
2317 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2322 /* The current character (and any potential folds) should be added
2323 * to the possible matching characters for this position in this
2327 U8 folded= folder[ (U8) uvc ];
2328 if ( !trie->charmap[ folded ] ) {
2329 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2330 TRIE_STORE_REVCHAR( folded );
2333 if ( !trie->charmap[ uvc ] ) {
2334 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2335 TRIE_STORE_REVCHAR( uvc );
2338 /* store the codepoint in the bitmap, and its folded
2340 TRIE_BITMAP_SET(trie, uvc);
2342 /* store the folded codepoint */
2343 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
2346 /* store first byte of utf8 representation of
2347 variant codepoints */
2348 if (! UVCHR_IS_INVARIANT(uvc)) {
2349 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
2352 set_bit = 0; /* We've done our bit :-) */
2356 /* XXX We could come up with the list of code points that fold
2357 * to this using PL_utf8_foldclosures, except not for
2358 * multi-char folds, as there may be multiple combinations
2359 * there that could work, which needs to wait until runtime to
2360 * resolve (The comment about LIGATURE FFI above is such an
2365 widecharmap = newHV();
2367 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2370 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
2372 if ( !SvTRUE( *svpp ) ) {
2373 sv_setiv( *svpp, ++trie->uniquecharcount );
2374 TRIE_STORE_REVCHAR(uvc);
2377 } /* end loop through characters in this branch of the trie */
2379 /* We take the min and max for this branch and combine to find the min
2380 * and max for all branches processed so far */
2381 if( cur == first ) {
2382 trie->minlen = minchars;
2383 trie->maxlen = maxchars;
2384 } else if (minchars < trie->minlen) {
2385 trie->minlen = minchars;
2386 } else if (maxchars > trie->maxlen) {
2387 trie->maxlen = maxchars;
2389 } /* end first pass */
2390 DEBUG_TRIE_COMPILE_r(
2391 PerlIO_printf( Perl_debug_log,
2392 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2393 (int)depth * 2 + 2,"",
2394 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2395 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2396 (int)trie->minlen, (int)trie->maxlen )
2400 We now know what we are dealing with in terms of unique chars and
2401 string sizes so we can calculate how much memory a naive
2402 representation using a flat table will take. If it's over a reasonable
2403 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2404 conservative but potentially much slower representation using an array
2407 At the end we convert both representations into the same compressed
2408 form that will be used in regexec.c for matching with. The latter
2409 is a form that cannot be used to construct with but has memory
2410 properties similar to the list form and access properties similar
2411 to the table form making it both suitable for fast searches and
2412 small enough that its feasable to store for the duration of a program.
2414 See the comment in the code where the compressed table is produced
2415 inplace from the flat tabe representation for an explanation of how
2416 the compression works.
2421 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2424 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2425 > SvIV(re_trie_maxbuff) )
2428 Second Pass -- Array Of Lists Representation
2430 Each state will be represented by a list of charid:state records
2431 (reg_trie_trans_le) the first such element holds the CUR and LEN
2432 points of the allocated array. (See defines above).
2434 We build the initial structure using the lists, and then convert
2435 it into the compressed table form which allows faster lookups
2436 (but cant be modified once converted).
2439 STRLEN transcount = 1;
2441 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2442 "%*sCompiling trie using list compiler\n",
2443 (int)depth * 2 + 2, ""));
2445 trie->states = (reg_trie_state *)
2446 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2447 sizeof(reg_trie_state) );
2451 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2453 regnode *noper = NEXTOPER( cur );
2454 U8 *uc = (U8*)STRING( noper );
2455 const U8 *e = uc + STR_LEN( noper );
2456 U32 state = 1; /* required init */
2457 U16 charid = 0; /* sanity init */
2458 U32 wordlen = 0; /* required init */
2460 if (OP(noper) == NOTHING) {
2461 regnode *noper_next= regnext(noper);
2462 if (noper_next != tail && OP(noper_next) == flags) {
2464 uc= (U8*)STRING(noper);
2465 e= uc + STR_LEN(noper);
2469 if (OP(noper) != NOTHING) {
2470 for ( ; uc < e ; uc += len ) {
2475 charid = trie->charmap[ uvc ];
2477 SV** const svpp = hv_fetch( widecharmap,
2484 charid=(U16)SvIV( *svpp );
2487 /* charid is now 0 if we dont know the char read, or
2488 * nonzero if we do */
2495 if ( !trie->states[ state ].trans.list ) {
2496 TRIE_LIST_NEW( state );
2499 check <= TRIE_LIST_USED( state );
2502 if ( TRIE_LIST_ITEM( state, check ).forid
2505 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2510 newstate = next_alloc++;
2511 prev_states[newstate] = state;
2512 TRIE_LIST_PUSH( state, charid, newstate );
2517 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2521 TRIE_HANDLE_WORD(state);
2523 } /* end second pass */
2525 /* next alloc is the NEXT state to be allocated */
2526 trie->statecount = next_alloc;
2527 trie->states = (reg_trie_state *)
2528 PerlMemShared_realloc( trie->states,
2530 * sizeof(reg_trie_state) );
2532 /* and now dump it out before we compress it */
2533 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2534 revcharmap, next_alloc,
2538 trie->trans = (reg_trie_trans *)
2539 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2546 for( state=1 ; state < next_alloc ; state ++ ) {
2550 DEBUG_TRIE_COMPILE_MORE_r(
2551 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2555 if (trie->states[state].trans.list) {
2556 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2560 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2561 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2562 if ( forid < minid ) {
2564 } else if ( forid > maxid ) {
2568 if ( transcount < tp + maxid - minid + 1) {
2570 trie->trans = (reg_trie_trans *)
2571 PerlMemShared_realloc( trie->trans,
2573 * sizeof(reg_trie_trans) );
2574 Zero( trie->trans + (transcount / 2),
2578 base = trie->uniquecharcount + tp - minid;
2579 if ( maxid == minid ) {
2581 for ( ; zp < tp ; zp++ ) {
2582 if ( ! trie->trans[ zp ].next ) {
2583 base = trie->uniquecharcount + zp - minid;
2584 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2586 trie->trans[ zp ].check = state;
2592 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2594 trie->trans[ tp ].check = state;
2599 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2600 const U32 tid = base
2601 - trie->uniquecharcount
2602 + TRIE_LIST_ITEM( state, idx ).forid;
2603 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2605 trie->trans[ tid ].check = state;
2607 tp += ( maxid - minid + 1 );
2609 Safefree(trie->states[ state ].trans.list);
2612 DEBUG_TRIE_COMPILE_MORE_r(
2613 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2616 trie->states[ state ].trans.base=base;
2618 trie->lasttrans = tp + 1;
2622 Second Pass -- Flat Table Representation.
2624 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2625 each. We know that we will need Charcount+1 trans at most to store
2626 the data (one row per char at worst case) So we preallocate both
2627 structures assuming worst case.
2629 We then construct the trie using only the .next slots of the entry
2632 We use the .check field of the first entry of the node temporarily
2633 to make compression both faster and easier by keeping track of how
2634 many non zero fields are in the node.
2636 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2639 There are two terms at use here: state as a TRIE_NODEIDX() which is
2640 a number representing the first entry of the node, and state as a
2641 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2642 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2643 if there are 2 entrys per node. eg:
2651 The table is internally in the right hand, idx form. However as we
2652 also have to deal with the states array which is indexed by nodenum
2653 we have to use TRIE_NODENUM() to convert.
2656 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2657 "%*sCompiling trie using table compiler\n",
2658 (int)depth * 2 + 2, ""));
2660 trie->trans = (reg_trie_trans *)
2661 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2662 * trie->uniquecharcount + 1,
2663 sizeof(reg_trie_trans) );
2664 trie->states = (reg_trie_state *)
2665 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2666 sizeof(reg_trie_state) );
2667 next_alloc = trie->uniquecharcount + 1;
2670 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2672 regnode *noper = NEXTOPER( cur );
2673 const U8 *uc = (U8*)STRING( noper );
2674 const U8 *e = uc + STR_LEN( noper );
2676 U32 state = 1; /* required init */
2678 U16 charid = 0; /* sanity init */
2679 U32 accept_state = 0; /* sanity init */
2681 U32 wordlen = 0; /* required init */
2683 if (OP(noper) == NOTHING) {
2684 regnode *noper_next= regnext(noper);
2685 if (noper_next != tail && OP(noper_next) == flags) {
2687 uc= (U8*)STRING(noper);
2688 e= uc + STR_LEN(noper);
2692 if ( OP(noper) != NOTHING ) {
2693 for ( ; uc < e ; uc += len ) {
2698 charid = trie->charmap[ uvc ];
2700 SV* const * const svpp = hv_fetch( widecharmap,
2704 charid = svpp ? (U16)SvIV(*svpp) : 0;
2708 if ( !trie->trans[ state + charid ].next ) {
2709 trie->trans[ state + charid ].next = next_alloc;
2710 trie->trans[ state ].check++;
2711 prev_states[TRIE_NODENUM(next_alloc)]
2712 = TRIE_NODENUM(state);
2713 next_alloc += trie->uniquecharcount;
2715 state = trie->trans[ state + charid ].next;
2717 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2719 /* charid is now 0 if we dont know the char read, or
2720 * nonzero if we do */
2723 accept_state = TRIE_NODENUM( state );
2724 TRIE_HANDLE_WORD(accept_state);
2726 } /* end second pass */
2728 /* and now dump it out before we compress it */
2729 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2731 next_alloc, depth+1));
2735 * Inplace compress the table.*
2737 For sparse data sets the table constructed by the trie algorithm will
2738 be mostly 0/FAIL transitions or to put it another way mostly empty.
2739 (Note that leaf nodes will not contain any transitions.)
2741 This algorithm compresses the tables by eliminating most such
2742 transitions, at the cost of a modest bit of extra work during lookup:
2744 - Each states[] entry contains a .base field which indicates the
2745 index in the state[] array wheres its transition data is stored.
2747 - If .base is 0 there are no valid transitions from that node.
2749 - If .base is nonzero then charid is added to it to find an entry in
2752 -If trans[states[state].base+charid].check!=state then the
2753 transition is taken to be a 0/Fail transition. Thus if there are fail
2754 transitions at the front of the node then the .base offset will point
2755 somewhere inside the previous nodes data (or maybe even into a node
2756 even earlier), but the .check field determines if the transition is
2760 The following process inplace converts the table to the compressed
2761 table: We first do not compress the root node 1,and mark all its
2762 .check pointers as 1 and set its .base pointer as 1 as well. This
2763 allows us to do a DFA construction from the compressed table later,
2764 and ensures that any .base pointers we calculate later are greater
2767 - We set 'pos' to indicate the first entry of the second node.
2769 - We then iterate over the columns of the node, finding the first and
2770 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2771 and set the .check pointers accordingly, and advance pos
2772 appropriately and repreat for the next node. Note that when we copy
2773 the next pointers we have to convert them from the original
2774 NODEIDX form to NODENUM form as the former is not valid post
2777 - If a node has no transitions used we mark its base as 0 and do not
2778 advance the pos pointer.
2780 - If a node only has one transition we use a second pointer into the
2781 structure to fill in allocated fail transitions from other states.
2782 This pointer is independent of the main pointer and scans forward
2783 looking for null transitions that are allocated to a state. When it
2784 finds one it writes the single transition into the "hole". If the
2785 pointer doesnt find one the single transition is appended as normal.
2787 - Once compressed we can Renew/realloc the structures to release the
2790 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2791 specifically Fig 3.47 and the associated pseudocode.
2795 const U32 laststate = TRIE_NODENUM( next_alloc );
2798 trie->statecount = laststate;
2800 for ( state = 1 ; state < laststate ; state++ ) {
2802 const U32 stateidx = TRIE_NODEIDX( state );
2803 const U32 o_used = trie->trans[ stateidx ].check;
2804 U32 used = trie->trans[ stateidx ].check;
2805 trie->trans[ stateidx ].check = 0;
2808 used && charid < trie->uniquecharcount;
2811 if ( flag || trie->trans[ stateidx + charid ].next ) {
2812 if ( trie->trans[ stateidx + charid ].next ) {
2814 for ( ; zp < pos ; zp++ ) {
2815 if ( ! trie->trans[ zp ].next ) {
2819 trie->states[ state ].trans.base
2821 + trie->uniquecharcount
2823 trie->trans[ zp ].next
2824 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2826 trie->trans[ zp ].check = state;
2827 if ( ++zp > pos ) pos = zp;
2834 trie->states[ state ].trans.base
2835 = pos + trie->uniquecharcount - charid ;
2837 trie->trans[ pos ].next
2838 = SAFE_TRIE_NODENUM(
2839 trie->trans[ stateidx + charid ].next );
2840 trie->trans[ pos ].check = state;
2845 trie->lasttrans = pos + 1;
2846 trie->states = (reg_trie_state *)
2847 PerlMemShared_realloc( trie->states, laststate
2848 * sizeof(reg_trie_state) );
2849 DEBUG_TRIE_COMPILE_MORE_r(
2850 PerlIO_printf( Perl_debug_log,
2851 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2852 (int)depth * 2 + 2,"",
2853 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2857 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2860 } /* end table compress */
2862 DEBUG_TRIE_COMPILE_MORE_r(
2863 PerlIO_printf(Perl_debug_log,
2864 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2865 (int)depth * 2 + 2, "",
2866 (UV)trie->statecount,
2867 (UV)trie->lasttrans)
2869 /* resize the trans array to remove unused space */
2870 trie->trans = (reg_trie_trans *)
2871 PerlMemShared_realloc( trie->trans, trie->lasttrans
2872 * sizeof(reg_trie_trans) );
2874 { /* Modify the program and insert the new TRIE node */
2875 U8 nodetype =(U8)(flags & 0xFF);
2879 regnode *optimize = NULL;
2880 #ifdef RE_TRACK_PATTERN_OFFSETS
2883 U32 mjd_nodelen = 0;
2884 #endif /* RE_TRACK_PATTERN_OFFSETS */
2885 #endif /* DEBUGGING */
2887 This means we convert either the first branch or the first Exact,
2888 depending on whether the thing following (in 'last') is a branch
2889 or not and whther first is the startbranch (ie is it a sub part of
2890 the alternation or is it the whole thing.)
2891 Assuming its a sub part we convert the EXACT otherwise we convert
2892 the whole branch sequence, including the first.
2894 /* Find the node we are going to overwrite */
2895 if ( first != startbranch || OP( last ) == BRANCH ) {
2896 /* branch sub-chain */
2897 NEXT_OFF( first ) = (U16)(last - first);
2898 #ifdef RE_TRACK_PATTERN_OFFSETS
2900 mjd_offset= Node_Offset((convert));
2901 mjd_nodelen= Node_Length((convert));
2904 /* whole branch chain */
2906 #ifdef RE_TRACK_PATTERN_OFFSETS
2909 const regnode *nop = NEXTOPER( convert );
2910 mjd_offset= Node_Offset((nop));
2911 mjd_nodelen= Node_Length((nop));
2915 PerlIO_printf(Perl_debug_log,
2916 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2917 (int)depth * 2 + 2, "",
2918 (UV)mjd_offset, (UV)mjd_nodelen)
2921 /* But first we check to see if there is a common prefix we can
2922 split out as an EXACT and put in front of the TRIE node. */
2923 trie->startstate= 1;
2924 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2926 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2930 const U32 base = trie->states[ state ].trans.base;
2932 if ( trie->states[state].wordnum )
2935 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2936 if ( ( base + ofs >= trie->uniquecharcount ) &&
2937 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2938 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2940 if ( ++count > 1 ) {
2941 SV **tmp = av_fetch( revcharmap, ofs, 0);
2942 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2943 if ( state == 1 ) break;
2945 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2947 PerlIO_printf(Perl_debug_log,
2948 "%*sNew Start State=%"UVuf" Class: [",
2949 (int)depth * 2 + 2, "",
2952 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2953 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2955 TRIE_BITMAP_SET(trie,*ch);
2957 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2959 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2963 TRIE_BITMAP_SET(trie,*ch);
2965 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2966 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2972 SV **tmp = av_fetch( revcharmap, idx, 0);
2974 char *ch = SvPV( *tmp, len );
2976 SV *sv=sv_newmortal();
2977 PerlIO_printf( Perl_debug_log,
2978 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2979 (int)depth * 2 + 2, "",
2981 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2982 PL_colors[0], PL_colors[1],
2983 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2984 PERL_PV_ESCAPE_FIRSTCHAR
2989 OP( convert ) = nodetype;
2990 str=STRING(convert);
2993 STR_LEN(convert) += len;
2999 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
3004 trie->prefixlen = (state-1);
3006 regnode *n = convert+NODE_SZ_STR(convert);
3007 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3008 trie->startstate = state;
3009 trie->minlen -= (state - 1);
3010 trie->maxlen -= (state - 1);
3012 /* At least the UNICOS C compiler choked on this
3013 * being argument to DEBUG_r(), so let's just have
3016 #ifdef PERL_EXT_RE_BUILD
3022 regnode *fix = convert;
3023 U32 word = trie->wordcount;
3025 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3026 while( ++fix < n ) {
3027 Set_Node_Offset_Length(fix, 0, 0);
3030 SV ** const tmp = av_fetch( trie_words, word, 0 );
3032 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3033 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3035 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3043 NEXT_OFF(convert) = (U16)(tail - convert);
3044 DEBUG_r(optimize= n);
3050 if ( trie->maxlen ) {
3051 NEXT_OFF( convert ) = (U16)(tail - convert);
3052 ARG_SET( convert, data_slot );
3053 /* Store the offset to the first unabsorbed branch in
3054 jump[0], which is otherwise unused by the jump logic.
3055 We use this when dumping a trie and during optimisation. */
3057 trie->jump[0] = (U16)(nextbranch - convert);
3059 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3060 * and there is a bitmap
3061 * and the first "jump target" node we found leaves enough room
3062 * then convert the TRIE node into a TRIEC node, with the bitmap
3063 * embedded inline in the opcode - this is hypothetically faster.
3065 if ( !trie->states[trie->startstate].wordnum
3067 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3069 OP( convert ) = TRIEC;
3070 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3071 PerlMemShared_free(trie->bitmap);
3074 OP( convert ) = TRIE;
3076 /* store the type in the flags */
3077 convert->flags = nodetype;
3081 + regarglen[ OP( convert ) ];
3083 /* XXX We really should free up the resource in trie now,
3084 as we won't use them - (which resources?) dmq */
3086 /* needed for dumping*/
3087 DEBUG_r(if (optimize) {
3088 regnode *opt = convert;
3090 while ( ++opt < optimize) {
3091 Set_Node_Offset_Length(opt,0,0);
3094 Try to clean up some of the debris left after the
3097 while( optimize < jumper ) {
3098 mjd_nodelen += Node_Length((optimize));
3099 OP( optimize ) = OPTIMIZED;
3100 Set_Node_Offset_Length(optimize,0,0);
3103 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3105 } /* end node insert */
3107 /* Finish populating the prev field of the wordinfo array. Walk back
3108 * from each accept state until we find another accept state, and if
3109 * so, point the first word's .prev field at the second word. If the
3110 * second already has a .prev field set, stop now. This will be the
3111 * case either if we've already processed that word's accept state,
3112 * or that state had multiple words, and the overspill words were
3113 * already linked up earlier.
3120 for (word=1; word <= trie->wordcount; word++) {
3122 if (trie->wordinfo[word].prev)
3124 state = trie->wordinfo[word].accept;
3126 state = prev_states[state];
3129 prev = trie->states[state].wordnum;
3133 trie->wordinfo[word].prev = prev;
3135 Safefree(prev_states);
3139 /* and now dump out the compressed format */
3140 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3142 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3144 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3145 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3147 SvREFCNT_dec_NN(revcharmap);
3151 : trie->startstate>1
3157 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3159 /* The Trie is constructed and compressed now so we can build a fail array if
3162 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3164 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3168 We find the fail state for each state in the trie, this state is the longest
3169 proper suffix of the current state's 'word' that is also a proper prefix of
3170 another word in our trie. State 1 represents the word '' and is thus the
3171 default fail state. This allows the DFA not to have to restart after its
3172 tried and failed a word at a given point, it simply continues as though it
3173 had been matching the other word in the first place.
3175 'abcdgu'=~/abcdefg|cdgu/
3176 When we get to 'd' we are still matching the first word, we would encounter
3177 'g' which would fail, which would bring us to the state representing 'd' in
3178 the second word where we would try 'g' and succeed, proceeding to match
3181 /* add a fail transition */
3182 const U32 trie_offset = ARG(source);
3183 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3185 const U32 ucharcount = trie->uniquecharcount;
3186 const U32 numstates = trie->statecount;
3187 const U32 ubound = trie->lasttrans + ucharcount;
3191 U32 base = trie->states[ 1 ].trans.base;
3194 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3196 GET_RE_DEBUG_FLAGS_DECL;
3198 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3199 PERL_UNUSED_CONTEXT;
3201 PERL_UNUSED_ARG(depth);
3204 if ( OP(source) == TRIE ) {
3205 struct regnode_1 *op = (struct regnode_1 *)
3206 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3207 StructCopy(source,op,struct regnode_1);
3208 stclass = (regnode *)op;
3210 struct regnode_charclass *op = (struct regnode_charclass *)
3211 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3212 StructCopy(source,op,struct regnode_charclass);
3213 stclass = (regnode *)op;
3215 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3217 ARG_SET( stclass, data_slot );
3218 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3219 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3220 aho->trie=trie_offset;
3221 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3222 Copy( trie->states, aho->states, numstates, reg_trie_state );
3223 Newxz( q, numstates, U32);
3224 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3227 /* initialize fail[0..1] to be 1 so that we always have
3228 a valid final fail state */
3229 fail[ 0 ] = fail[ 1 ] = 1;
3231 for ( charid = 0; charid < ucharcount ; charid++ ) {
3232 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3234 q[ q_write ] = newstate;
3235 /* set to point at the root */
3236 fail[ q[ q_write++ ] ]=1;
3239 while ( q_read < q_write) {
3240 const U32 cur = q[ q_read++ % numstates ];
3241 base = trie->states[ cur ].trans.base;
3243 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3244 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3246 U32 fail_state = cur;
3249 fail_state = fail[ fail_state ];
3250 fail_base = aho->states[ fail_state ].trans.base;
3251 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3253 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3254 fail[ ch_state ] = fail_state;
3255 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3257 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3259 q[ q_write++ % numstates] = ch_state;
3263 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3264 when we fail in state 1, this allows us to use the
3265 charclass scan to find a valid start char. This is based on the principle
3266 that theres a good chance the string being searched contains lots of stuff
3267 that cant be a start char.
3269 fail[ 0 ] = fail[ 1 ] = 0;
3270 DEBUG_TRIE_COMPILE_r({
3271 PerlIO_printf(Perl_debug_log,
3272 "%*sStclass Failtable (%"UVuf" states): 0",
3273 (int)(depth * 2), "", (UV)numstates
3275 for( q_read=1; q_read<numstates; q_read++ ) {
3276 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3278 PerlIO_printf(Perl_debug_log, "\n");
3281 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3286 #define DEBUG_PEEP(str,scan,depth) \
3287 DEBUG_OPTIMISE_r({if (scan){ \
3288 regnode *Next = regnext(scan); \
3289 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state); \
3290 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)", \
3291 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),\
3292 Next ? (REG_NODE_NUM(Next)) : 0 ); \
3293 DEBUG_SHOW_STUDY_FLAGS(flags," [ ","]");\
3294 PerlIO_printf(Perl_debug_log, "\n"); \
3297 /* The below joins as many adjacent EXACTish nodes as possible into a single
3298 * one. The regop may be changed if the node(s) contain certain sequences that
3299 * require special handling. The joining is only done if:
3300 * 1) there is room in the current conglomerated node to entirely contain the
3302 * 2) they are the exact same node type
3304 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3305 * these get optimized out
3307 * If a node is to match under /i (folded), the number of characters it matches
3308 * can be different than its character length if it contains a multi-character
3309 * fold. *min_subtract is set to the total delta number of characters of the
3312 * And *unfolded_multi_char is set to indicate whether or not the node contains
3313 * an unfolded multi-char fold. This happens when whether the fold is valid or
3314 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3315 * SMALL LETTER SHARP S, as only if the target string being matched against
3316 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3317 * folding rules depend on the locale in force at runtime. (Multi-char folds
3318 * whose components are all above the Latin1 range are not run-time locale
3319 * dependent, and have already been folded by the time this function is
3322 * This is as good a place as any to discuss the design of handling these
3323 * multi-character fold sequences. It's been wrong in Perl for a very long
3324 * time. There are three code points in Unicode whose multi-character folds
3325 * were long ago discovered to mess things up. The previous designs for
3326 * dealing with these involved assigning a special node for them. This
3327 * approach doesn't always work, as evidenced by this example:
3328 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3329 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3330 * would match just the \xDF, it won't be able to handle the case where a
3331 * successful match would have to cross the node's boundary. The new approach
3332 * that hopefully generally solves the problem generates an EXACTFU_SS node
3333 * that is "sss" in this case.
3335 * It turns out that there are problems with all multi-character folds, and not
3336 * just these three. Now the code is general, for all such cases. The
3337 * approach taken is:
3338 * 1) This routine examines each EXACTFish node that could contain multi-
3339 * character folded sequences. Since a single character can fold into
3340 * such a sequence, the minimum match length for this node is less than
3341 * the number of characters in the node. This routine returns in
3342 * *min_subtract how many characters to subtract from the the actual
3343 * length of the string to get a real minimum match length; it is 0 if
3344 * there are no multi-char foldeds. This delta is used by the caller to
3345 * adjust the min length of the match, and the delta between min and max,
3346 * so that the optimizer doesn't reject these possibilities based on size
3348 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3349 * is used for an EXACTFU node that contains at least one "ss" sequence in
3350 * it. For non-UTF-8 patterns and strings, this is the only case where
3351 * there is a possible fold length change. That means that a regular
3352 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3353 * with length changes, and so can be processed faster. regexec.c takes
3354 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3355 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3356 * known until runtime). This saves effort in regex matching. However,
3357 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3358 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3359 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3360 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3361 * possibilities for the non-UTF8 patterns are quite simple, except for
3362 * the sharp s. All the ones that don't involve a UTF-8 target string are
3363 * members of a fold-pair, and arrays are set up for all of them so that
3364 * the other member of the pair can be found quickly. Code elsewhere in
3365 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3366 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3367 * described in the next item.
3368 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3369 * validity of the fold won't be known until runtime, and so must remain
3370 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3371 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3372 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3373 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3374 * The reason this is a problem is that the optimizer part of regexec.c
3375 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3376 * that a character in the pattern corresponds to at most a single
3377 * character in the target string. (And I do mean character, and not byte
3378 * here, unlike other parts of the documentation that have never been
3379 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3380 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3381 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3382 * nodes, violate the assumption, and they are the only instances where it
3383 * is violated. I'm reluctant to try to change the assumption, as the
3384 * code involved is impenetrable to me (khw), so instead the code here
3385 * punts. This routine examines EXACTFL nodes, and (when the pattern
3386 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3387 * boolean indicating whether or not the node contains such a fold. When
3388 * it is true, the caller sets a flag that later causes the optimizer in
3389 * this file to not set values for the floating and fixed string lengths,
3390 * and thus avoids the optimizer code in regexec.c that makes the invalid
3391 * assumption. Thus, there is no optimization based on string lengths for
3392 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3393 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3394 * assumption is wrong only in these cases is that all other non-UTF-8
3395 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3396 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3397 * EXACTF nodes because we don't know at compile time if it actually
3398 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3399 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3400 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3401 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3402 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3403 * string would require the pattern to be forced into UTF-8, the overhead
3404 * of which we want to avoid. Similarly the unfolded multi-char folds in
3405 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3408 * Similarly, the code that generates tries doesn't currently handle
3409 * not-already-folded multi-char folds, and it looks like a pain to change
3410 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3411 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3412 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3413 * using /iaa matching will be doing so almost entirely with ASCII
3414 * strings, so this should rarely be encountered in practice */
3416 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3417 if (PL_regkind[OP(scan)] == EXACT) \
3418 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3421 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3422 UV *min_subtract, bool *unfolded_multi_char,
3423 U32 flags,regnode *val, U32 depth)
3425 /* Merge several consecutive EXACTish nodes into one. */
3426 regnode *n = regnext(scan);
3428 regnode *next = scan + NODE_SZ_STR(scan);
3432 regnode *stop = scan;
3433 GET_RE_DEBUG_FLAGS_DECL;
3435 PERL_UNUSED_ARG(depth);
3438 PERL_ARGS_ASSERT_JOIN_EXACT;
3439 #ifndef EXPERIMENTAL_INPLACESCAN
3440 PERL_UNUSED_ARG(flags);
3441 PERL_UNUSED_ARG(val);
3443 DEBUG_PEEP("join",scan,depth);
3445 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3446 * EXACT ones that are mergeable to the current one. */
3448 && (PL_regkind[OP(n)] == NOTHING
3449 || (stringok && OP(n) == OP(scan)))
3451 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3454 if (OP(n) == TAIL || n > next)
3456 if (PL_regkind[OP(n)] == NOTHING) {
3457 DEBUG_PEEP("skip:",n,depth);
3458 NEXT_OFF(scan) += NEXT_OFF(n);
3459 next = n + NODE_STEP_REGNODE;
3466 else if (stringok) {
3467 const unsigned int oldl = STR_LEN(scan);
3468 regnode * const nnext = regnext(n);
3470 /* XXX I (khw) kind of doubt that this works on platforms (should
3471 * Perl ever run on one) where U8_MAX is above 255 because of lots
3472 * of other assumptions */
3473 /* Don't join if the sum can't fit into a single node */
3474 if (oldl + STR_LEN(n) > U8_MAX)
3477 DEBUG_PEEP("merg",n,depth);
3480 NEXT_OFF(scan) += NEXT_OFF(n);
3481 STR_LEN(scan) += STR_LEN(n);
3482 next = n + NODE_SZ_STR(n);
3483 /* Now we can overwrite *n : */
3484 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3492 #ifdef EXPERIMENTAL_INPLACESCAN
3493 if (flags && !NEXT_OFF(n)) {
3494 DEBUG_PEEP("atch", val, depth);
3495 if (reg_off_by_arg[OP(n)]) {
3496 ARG_SET(n, val - n);
3499 NEXT_OFF(n) = val - n;
3507 *unfolded_multi_char = FALSE;
3509 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3510 * can now analyze for sequences of problematic code points. (Prior to
3511 * this final joining, sequences could have been split over boundaries, and
3512 * hence missed). The sequences only happen in folding, hence for any
3513 * non-EXACT EXACTish node */
3514 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3515 U8* s0 = (U8*) STRING(scan);
3517 U8* s_end = s0 + STR_LEN(scan);
3519 int total_count_delta = 0; /* Total delta number of characters that
3520 multi-char folds expand to */
3522 /* One pass is made over the node's string looking for all the
3523 * possibilities. To avoid some tests in the loop, there are two main
3524 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3529 if (OP(scan) == EXACTFL) {
3532 /* An EXACTFL node would already have been changed to another
3533 * node type unless there is at least one character in it that
3534 * is problematic; likely a character whose fold definition
3535 * won't be known until runtime, and so has yet to be folded.
3536 * For all but the UTF-8 locale, folds are 1-1 in length, but
3537 * to handle the UTF-8 case, we need to create a temporary
3538 * folded copy using UTF-8 locale rules in order to analyze it.
3539 * This is because our macros that look to see if a sequence is
3540 * a multi-char fold assume everything is folded (otherwise the
3541 * tests in those macros would be too complicated and slow).
3542 * Note that here, the non-problematic folds will have already
3543 * been done, so we can just copy such characters. We actually
3544 * don't completely fold the EXACTFL string. We skip the
3545 * unfolded multi-char folds, as that would just create work
3546 * below to figure out the size they already are */
3548 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3551 STRLEN s_len = UTF8SKIP(s);
3552 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3553 Copy(s, d, s_len, U8);
3556 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3557 *unfolded_multi_char = TRUE;
3558 Copy(s, d, s_len, U8);
3561 else if (isASCII(*s)) {
3562 *(d++) = toFOLD(*s);
3566 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3572 /* Point the remainder of the routine to look at our temporary
3576 } /* End of creating folded copy of EXACTFL string */
3578 /* Examine the string for a multi-character fold sequence. UTF-8
3579 * patterns have all characters pre-folded by the time this code is
3581 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3582 length sequence we are looking for is 2 */
3584 int count = 0; /* How many characters in a multi-char fold */
3585 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3586 if (! len) { /* Not a multi-char fold: get next char */
3591 /* Nodes with 'ss' require special handling, except for
3592 * EXACTFA-ish for which there is no multi-char fold to this */
3593 if (len == 2 && *s == 's' && *(s+1) == 's'
3594 && OP(scan) != EXACTFA
3595 && OP(scan) != EXACTFA_NO_TRIE)
3598 if (OP(scan) != EXACTFL) {
3599 OP(scan) = EXACTFU_SS;
3603 else { /* Here is a generic multi-char fold. */
3604 U8* multi_end = s + len;
3606 /* Count how many characters are in it. In the case of
3607 * /aa, no folds which contain ASCII code points are
3608 * allowed, so check for those, and skip if found. */
3609 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3610 count = utf8_length(s, multi_end);
3614 while (s < multi_end) {
3617 goto next_iteration;
3627 /* The delta is how long the sequence is minus 1 (1 is how long
3628 * the character that folds to the sequence is) */
3629 total_count_delta += count - 1;
3633 /* We created a temporary folded copy of the string in EXACTFL
3634 * nodes. Therefore we need to be sure it doesn't go below zero,
3635 * as the real string could be shorter */
3636 if (OP(scan) == EXACTFL) {
3637 int total_chars = utf8_length((U8*) STRING(scan),
3638 (U8*) STRING(scan) + STR_LEN(scan));
3639 if (total_count_delta > total_chars) {
3640 total_count_delta = total_chars;
3644 *min_subtract += total_count_delta;
3647 else if (OP(scan) == EXACTFA) {
3649 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3650 * fold to the ASCII range (and there are no existing ones in the
3651 * upper latin1 range). But, as outlined in the comments preceding
3652 * this function, we need to flag any occurrences of the sharp s.
3653 * This character forbids trie formation (because of added
3656 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
3657 OP(scan) = EXACTFA_NO_TRIE;
3658 *unfolded_multi_char = TRUE;
3667 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3668 * folds that are all Latin1. As explained in the comments
3669 * preceding this function, we look also for the sharp s in EXACTF
3670 * and EXACTFL nodes; it can be in the final position. Otherwise
3671 * we can stop looking 1 byte earlier because have to find at least
3672 * two characters for a multi-fold */
3673 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
3678 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
3679 if (! len) { /* Not a multi-char fold. */
3680 if (*s == LATIN_SMALL_LETTER_SHARP_S
3681 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
3683 *unfolded_multi_char = TRUE;
3690 && isALPHA_FOLD_EQ(*s, 's')
3691 && isALPHA_FOLD_EQ(*(s+1), 's'))
3694 /* EXACTF nodes need to know that the minimum length
3695 * changed so that a sharp s in the string can match this
3696 * ss in the pattern, but they remain EXACTF nodes, as they
3697 * won't match this unless the target string is is UTF-8,
3698 * which we don't know until runtime. EXACTFL nodes can't
3699 * transform into EXACTFU nodes */
3700 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
3701 OP(scan) = EXACTFU_SS;
3705 *min_subtract += len - 1;
3712 /* Allow dumping but overwriting the collection of skipped
3713 * ops and/or strings with fake optimized ops */
3714 n = scan + NODE_SZ_STR(scan);
3722 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3726 /* REx optimizer. Converts nodes into quicker variants "in place".
3727 Finds fixed substrings. */
3729 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3730 to the position after last scanned or to NULL. */
3732 #define INIT_AND_WITHP \
3733 assert(!and_withp); \
3734 Newx(and_withp,1, regnode_ssc); \
3735 SAVEFREEPV(and_withp)
3739 S_unwind_scan_frames(pTHX_ const void *p)
3741 scan_frame *f= (scan_frame *)p;
3743 scan_frame *n= f->next_frame;
3751 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3752 SSize_t *minlenp, SSize_t *deltap,
3757 regnode_ssc *and_withp,
3758 U32 flags, U32 depth)
3759 /* scanp: Start here (read-write). */
3760 /* deltap: Write maxlen-minlen here. */
3761 /* last: Stop before this one. */
3762 /* data: string data about the pattern */
3763 /* stopparen: treat close N as END */
3764 /* recursed: which subroutines have we recursed into */
3765 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3767 /* There must be at least this number of characters to match */
3770 regnode *scan = *scanp, *next;
3772 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3773 int is_inf_internal = 0; /* The studied chunk is infinite */
3774 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3775 scan_data_t data_fake;
3776 SV *re_trie_maxbuff = NULL;
3777 regnode *first_non_open = scan;
3778 SSize_t stopmin = SSize_t_MAX;
3779 scan_frame *frame = NULL;
3780 GET_RE_DEBUG_FLAGS_DECL;
3782 PERL_ARGS_ASSERT_STUDY_CHUNK;
3786 while (first_non_open && OP(first_non_open) == OPEN)
3787 first_non_open=regnext(first_non_open);
3793 RExC_study_chunk_recursed_count++;
3795 DEBUG_OPTIMISE_MORE_r(
3797 PerlIO_printf(Perl_debug_log,
3798 "%*sstudy_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
3799 (int)(depth*2), "", (long)stopparen,
3800 (unsigned long)RExC_study_chunk_recursed_count,
3801 (unsigned long)depth, (unsigned long)recursed_depth,
3804 if (recursed_depth) {
3807 for ( j = 0 ; j < recursed_depth ; j++ ) {
3808 for ( i = 0 ; i < (U32)RExC_npar ; i++ ) {
3810 PAREN_TEST(RExC_study_chunk_recursed +
3811 ( j * RExC_study_chunk_recursed_bytes), i )
3814 !PAREN_TEST(RExC_study_chunk_recursed +
3815 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
3818 PerlIO_printf(Perl_debug_log," %d",(int)i);
3822 if ( j + 1 < recursed_depth ) {
3823 PerlIO_printf(Perl_debug_log, ",");
3827 PerlIO_printf(Perl_debug_log,"\n");
3830 while ( scan && OP(scan) != END && scan < last ){
3831 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3832 node length to get a real minimum (because
3833 the folded version may be shorter) */
3834 bool unfolded_multi_char = FALSE;
3835 /* Peephole optimizer: */
3836 DEBUG_STUDYDATA("Peep:", data, depth);
3837 DEBUG_PEEP("Peep", scan, depth);
3840 /* The reason we do this here we need to deal with things like /(?:f)(?:o)(?:o)/
3841 * which cant be dealt with by the normal EXACT parsing code, as each (?:..) is handled
3842 * by a different invocation of reg() -- Yves
3844 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
3846 /* Follow the next-chain of the current node and optimize
3847 away all the NOTHINGs from it. */
3848 if (OP(scan) != CURLYX) {
3849 const int max = (reg_off_by_arg[OP(scan)]
3851 /* I32 may be smaller than U16 on CRAYs! */
3852 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3853 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3857 /* Skip NOTHING and LONGJMP. */
3858 while ((n = regnext(n))
3859 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3860 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3861 && off + noff < max)
3863 if (reg_off_by_arg[OP(scan)])
3866 NEXT_OFF(scan) = off;
3869 /* The principal pseudo-switch. Cannot be a switch, since we
3870 look into several different things. */
3871 if ( OP(scan) == DEFINEP ) {
3873 SSize_t deltanext = 0;
3874 SSize_t fake_last_close = 0;
3875 I32 f = SCF_IN_DEFINE;
3877 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3878 scan = regnext(scan);
3879 assert( OP(scan) == IFTHEN );
3880 DEBUG_PEEP("expect IFTHEN", scan, depth);
3882 data_fake.last_closep= &fake_last_close;
3884 next = regnext(scan);
3885 scan = NEXTOPER(NEXTOPER(scan));
3886 DEBUG_PEEP("scan", scan, depth);
3887 DEBUG_PEEP("next", next, depth);
3889 /* we suppose the run is continuous, last=next...
3890 * NOTE we dont use the return here! */
3891 (void)study_chunk(pRExC_state, &scan, &minlen,
3892 &deltanext, next, &data_fake, stopparen,
3893 recursed_depth, NULL, f, depth+1);
3898 OP(scan) == BRANCH ||
3899 OP(scan) == BRANCHJ ||
3902 next = regnext(scan);
3905 /* The op(next)==code check below is to see if we
3906 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
3907 * IFTHEN is special as it might not appear in pairs.
3908 * Not sure whether BRANCH-BRANCHJ is possible, regardless
3909 * we dont handle it cleanly. */
3910 if (OP(next) == code || code == IFTHEN) {
3911 /* NOTE - There is similar code to this block below for
3912 * handling TRIE nodes on a re-study. If you change stuff here
3913 * check there too. */
3914 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
3916 regnode * const startbranch=scan;
3918 if (flags & SCF_DO_SUBSTR) {
3919 /* Cannot merge strings after this. */
3920 scan_commit(pRExC_state, data, minlenp, is_inf);
3923 if (flags & SCF_DO_STCLASS)
3924 ssc_init_zero(pRExC_state, &accum);
3926 while (OP(scan) == code) {
3927 SSize_t deltanext, minnext, fake;
3929 regnode_ssc this_class;
3931 DEBUG_PEEP("Branch", scan, depth);
3934 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3936 data_fake.whilem_c = data->whilem_c;
3937 data_fake.last_closep = data->last_closep;
3940 data_fake.last_closep = &fake;
3942 data_fake.pos_delta = delta;
3943 next = regnext(scan);
3945 scan = NEXTOPER(scan); /* everything */
3946 if (code != BRANCH) /* everything but BRANCH */
3947 scan = NEXTOPER(scan);
3949 if (flags & SCF_DO_STCLASS) {
3950 ssc_init(pRExC_state, &this_class);
3951 data_fake.start_class = &this_class;
3952 f = SCF_DO_STCLASS_AND;
3954 if (flags & SCF_WHILEM_VISITED_POS)
3955 f |= SCF_WHILEM_VISITED_POS;
3957 /* we suppose the run is continuous, last=next...*/
3958 minnext = study_chunk(pRExC_state, &scan, minlenp,
3959 &deltanext, next, &data_fake, stopparen,
3960 recursed_depth, NULL, f,depth+1);
3964 if (deltanext == SSize_t_MAX) {
3965 is_inf = is_inf_internal = 1;
3967 } else if (max1 < minnext + deltanext)
3968 max1 = minnext + deltanext;
3970 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3972 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3973 if ( stopmin > minnext)
3974 stopmin = min + min1;
3975 flags &= ~SCF_DO_SUBSTR;
3977 data->flags |= SCF_SEEN_ACCEPT;
3980 if (data_fake.flags & SF_HAS_EVAL)
3981 data->flags |= SF_HAS_EVAL;
3982 data->whilem_c = data_fake.whilem_c;
3984 if (flags & SCF_DO_STCLASS)
3985 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
3987 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3989 if (flags & SCF_DO_SUBSTR) {
3990 data->pos_min += min1;
3991 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3992 data->pos_delta = SSize_t_MAX;
3994 data->pos_delta += max1 - min1;
3995 if (max1 != min1 || is_inf)
3996 data->longest = &(data->longest_float);
3999 if (delta == SSize_t_MAX
4000 || SSize_t_MAX - delta - (max1 - min1) < 0)
4001 delta = SSize_t_MAX;
4003 delta += max1 - min1;
4004 if (flags & SCF_DO_STCLASS_OR) {
4005 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4007 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4008 flags &= ~SCF_DO_STCLASS;
4011 else if (flags & SCF_DO_STCLASS_AND) {
4013 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4014 flags &= ~SCF_DO_STCLASS;
4017 /* Switch to OR mode: cache the old value of
4018 * data->start_class */
4020 StructCopy(data->start_class, and_withp, regnode_ssc);
4021 flags &= ~SCF_DO_STCLASS_AND;
4022 StructCopy(&accum, data->start_class, regnode_ssc);
4023 flags |= SCF_DO_STCLASS_OR;
4027 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4028 OP( startbranch ) == BRANCH )
4032 Assuming this was/is a branch we are dealing with: 'scan'
4033 now points at the item that follows the branch sequence,
4034 whatever it is. We now start at the beginning of the
4035 sequence and look for subsequences of
4041 which would be constructed from a pattern like
4044 If we can find such a subsequence we need to turn the first
4045 element into a trie and then add the subsequent branch exact
4046 strings to the trie.
4050 1. patterns where the whole set of branches can be
4053 2. patterns where only a subset can be converted.
4055 In case 1 we can replace the whole set with a single regop
4056 for the trie. In case 2 we need to keep the start and end
4059 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4060 becomes BRANCH TRIE; BRANCH X;
4062 There is an additional case, that being where there is a
4063 common prefix, which gets split out into an EXACT like node
4064 preceding the TRIE node.
4066 If x(1..n)==tail then we can do a simple trie, if not we make
4067 a "jump" trie, such that when we match the appropriate word
4068 we "jump" to the appropriate tail node. Essentially we turn
4069 a nested if into a case structure of sorts.
4074 if (!re_trie_maxbuff) {
4075 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4076 if (!SvIOK(re_trie_maxbuff))
4077 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4079 if ( SvIV(re_trie_maxbuff)>=0 ) {
4081 regnode *first = (regnode *)NULL;
4082 regnode *last = (regnode *)NULL;
4083 regnode *tail = scan;
4087 /* var tail is used because there may be a TAIL
4088 regop in the way. Ie, the exacts will point to the
4089 thing following the TAIL, but the last branch will
4090 point at the TAIL. So we advance tail. If we
4091 have nested (?:) we may have to move through several
4095 while ( OP( tail ) == TAIL ) {
4096 /* this is the TAIL generated by (?:) */
4097 tail = regnext( tail );
4101 DEBUG_TRIE_COMPILE_r({
4102 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4103 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
4104 (int)depth * 2 + 2, "",
4105 "Looking for TRIE'able sequences. Tail node is: ",
4106 SvPV_nolen_const( RExC_mysv )
4112 Step through the branches
4113 cur represents each branch,
4114 noper is the first thing to be matched as part
4116 noper_next is the regnext() of that node.
4118 We normally handle a case like this
4119 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4120 support building with NOJUMPTRIE, which restricts
4121 the trie logic to structures like /FOO|BAR/.
4123 If noper is a trieable nodetype then the branch is
4124 a possible optimization target. If we are building
4125 under NOJUMPTRIE then we require that noper_next is
4126 the same as scan (our current position in the regex
4129 Once we have two or more consecutive such branches
4130 we can create a trie of the EXACT's contents and
4131 stitch it in place into the program.
4133 If the sequence represents all of the branches in
4134 the alternation we replace the entire thing with a
4137 Otherwise when it is a subsequence we need to
4138 stitch it in place and replace only the relevant
4139 branches. This means the first branch has to remain
4140 as it is used by the alternation logic, and its
4141 next pointer, and needs to be repointed at the item
4142 on the branch chain following the last branch we
4143 have optimized away.
4145 This could be either a BRANCH, in which case the
4146 subsequence is internal, or it could be the item
4147 following the branch sequence in which case the
4148 subsequence is at the end (which does not
4149 necessarily mean the first node is the start of the
4152 TRIE_TYPE(X) is a define which maps the optype to a
4156 ----------------+-----------
4160 EXACTFU_SS | EXACTFU
4163 EXACTFLU8 | EXACTFLU8
4167 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4169 : ( EXACT == (X) ) \
4171 : ( EXACTFU == (X) || EXACTFU_SS == (X) ) \
4173 : ( EXACTFA == (X) ) \
4175 : ( EXACTL == (X) ) \
4177 : ( EXACTFLU8 == (X) ) \
4181 /* dont use tail as the end marker for this traverse */
4182 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4183 regnode * const noper = NEXTOPER( cur );
4184 U8 noper_type = OP( noper );
4185 U8 noper_trietype = TRIE_TYPE( noper_type );
4186 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4187 regnode * const noper_next = regnext( noper );
4188 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
4189 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
4192 DEBUG_TRIE_COMPILE_r({
4193 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4194 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
4195 (int)depth * 2 + 2,"", SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4197 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4198 PerlIO_printf( Perl_debug_log, " -> %s",
4199 SvPV_nolen_const(RExC_mysv));
4202 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4203 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
4204 SvPV_nolen_const(RExC_mysv));
4206 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
4207 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4208 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
4212 /* Is noper a trieable nodetype that can be merged
4213 * with the current trie (if there is one)? */
4217 ( noper_trietype == NOTHING)
4218 || ( trietype == NOTHING )
4219 || ( trietype == noper_trietype )
4222 && noper_next == tail
4226 /* Handle mergable triable node Either we are
4227 * the first node in a new trieable sequence,
4228 * in which case we do some bookkeeping,
4229 * otherwise we update the end pointer. */
4232 if ( noper_trietype == NOTHING ) {
4233 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4234 regnode * const noper_next = regnext( noper );
4235 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
4236 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4239 if ( noper_next_trietype ) {
4240 trietype = noper_next_trietype;
4241 } else if (noper_next_type) {
4242 /* a NOTHING regop is 1 regop wide.
4243 * We need at least two for a trie
4244 * so we can't merge this in */
4248 trietype = noper_trietype;
4251 if ( trietype == NOTHING )
4252 trietype = noper_trietype;
4257 } /* end handle mergable triable node */
4259 /* handle unmergable node -
4260 * noper may either be a triable node which can
4261 * not be tried together with the current trie,
4262 * or a non triable node */
4264 /* If last is set and trietype is not
4265 * NOTHING then we have found at least two
4266 * triable branch sequences in a row of a
4267 * similar trietype so we can turn them
4268 * into a trie. If/when we allow NOTHING to
4269 * start a trie sequence this condition
4270 * will be required, and it isn't expensive
4271 * so we leave it in for now. */
4272 if ( trietype && trietype != NOTHING )
4273 make_trie( pRExC_state,
4274 startbranch, first, cur, tail,
4275 count, trietype, depth+1 );
4276 last = NULL; /* note: we clear/update
4277 first, trietype etc below,
4278 so we dont do it here */
4282 && noper_next == tail
4285 /* noper is triable, so we can start a new
4289 trietype = noper_trietype;
4291 /* if we already saw a first but the
4292 * current node is not triable then we have
4293 * to reset the first information. */
4298 } /* end handle unmergable node */
4299 } /* loop over branches */
4300 DEBUG_TRIE_COMPILE_r({
4301 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4302 PerlIO_printf( Perl_debug_log,
4303 "%*s- %s (%d) <SCAN FINISHED>\n",
4305 "", SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4308 if ( last && trietype ) {
4309 if ( trietype != NOTHING ) {
4310 /* the last branch of the sequence was part of
4311 * a trie, so we have to construct it here
4312 * outside of the loop */
4313 made= make_trie( pRExC_state, startbranch,
4314 first, scan, tail, count,
4315 trietype, depth+1 );
4316 #ifdef TRIE_STUDY_OPT
4317 if ( ((made == MADE_EXACT_TRIE &&
4318 startbranch == first)
4319 || ( first_non_open == first )) &&
4321 flags |= SCF_TRIE_RESTUDY;
4322 if ( startbranch == first
4325 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4330 /* at this point we know whatever we have is a
4331 * NOTHING sequence/branch AND if 'startbranch'
4332 * is 'first' then we can turn the whole thing
4335 if ( startbranch == first ) {
4337 /* the entire thing is a NOTHING sequence,
4338 * something like this: (?:|) So we can
4339 * turn it into a plain NOTHING op. */
4340 DEBUG_TRIE_COMPILE_r({
4341 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4342 PerlIO_printf( Perl_debug_log,
4343 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
4344 "", SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4347 OP(startbranch)= NOTHING;
4348 NEXT_OFF(startbranch)= tail - startbranch;
4349 for ( opt= startbranch + 1; opt < tail ; opt++ )
4353 } /* end if ( last) */
4354 } /* TRIE_MAXBUF is non zero */
4359 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4360 scan = NEXTOPER(NEXTOPER(scan));
4361 } else /* single branch is optimized. */
4362 scan = NEXTOPER(scan);
4364 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
4366 regnode *start = NULL;
4367 regnode *end = NULL;
4368 U32 my_recursed_depth= recursed_depth;
4371 if (OP(scan) != SUSPEND) { /* GOSUB/GOSTART */
4372 /* Do setup, note this code has side effects beyond
4373 * the rest of this block. Specifically setting
4374 * RExC_recurse[] must happen at least once during
4376 if (OP(scan) == GOSUB) {
4378 RExC_recurse[ARG2L(scan)] = scan;
4379 start = RExC_open_parens[paren-1];
4380 end = RExC_close_parens[paren-1];
4382 start = RExC_rxi->program + 1;
https://perl5.git.perl.org / perl5.git / blob