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 "charclass_invlists.h"
91 #include "inline_invlist.c"
92 #include "unicode_constants.h"
94 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
95 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
96 #define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
97 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
98 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
99 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
102 #define STATIC static
106 struct RExC_state_t {
107 U32 flags; /* RXf_* are we folding, multilining? */
108 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
109 char *precomp; /* uncompiled string. */
110 REGEXP *rx_sv; /* The SV that is the regexp. */
111 regexp *rx; /* perl core regexp structure */
112 regexp_internal *rxi; /* internal data for regexp object
114 char *start; /* Start of input for compile */
115 char *end; /* End of input for compile */
116 char *parse; /* Input-scan pointer. */
117 SSize_t whilem_seen; /* number of WHILEM in this expr */
118 regnode *emit_start; /* Start of emitted-code area */
119 regnode *emit_bound; /* First regnode outside of the
121 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
122 implies compiling, so don't emit */
123 regnode_ssc emit_dummy; /* placeholder for emit to point to;
124 large enough for the largest
125 non-EXACTish node, so can use it as
127 I32 naughty; /* How bad is this pattern? */
128 I32 sawback; /* Did we see \1, ...? */
130 SSize_t size; /* Code size. */
131 I32 npar; /* Capture buffer count, (OPEN) plus
132 one. ("par" 0 is the whole
134 I32 nestroot; /* root parens we are in - used by
138 regnode **open_parens; /* pointers to open parens */
139 regnode **close_parens; /* pointers to close parens */
140 regnode *opend; /* END node in program */
141 I32 utf8; /* whether the pattern is utf8 or not */
142 I32 orig_utf8; /* whether the pattern was originally in utf8 */
143 /* XXX use this for future optimisation of case
144 * where pattern must be upgraded to utf8. */
145 I32 uni_semantics; /* If a d charset modifier should use unicode
146 rules, even if the pattern is not in
148 HV *paren_names; /* Paren names */
150 regnode **recurse; /* Recurse regops */
151 I32 recurse_count; /* Number of recurse regops */
152 U8 *study_chunk_recursed; /* bitmap of which parens we have moved
154 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
158 I32 override_recoding;
159 I32 in_multi_char_class;
160 struct reg_code_block *code_blocks; /* positions of literal (?{})
162 int num_code_blocks; /* size of code_blocks[] */
163 int code_index; /* next code_blocks[] slot */
164 SSize_t maxlen; /* mininum possible number of chars in string to match */
165 #ifdef ADD_TO_REGEXEC
166 char *starttry; /* -Dr: where regtry was called. */
167 #define RExC_starttry (pRExC_state->starttry)
169 SV *runtime_code_qr; /* qr with the runtime code blocks */
171 const char *lastparse;
173 AV *paren_name_list; /* idx -> name */
174 #define RExC_lastparse (pRExC_state->lastparse)
175 #define RExC_lastnum (pRExC_state->lastnum)
176 #define RExC_paren_name_list (pRExC_state->paren_name_list)
180 #define RExC_flags (pRExC_state->flags)
181 #define RExC_pm_flags (pRExC_state->pm_flags)
182 #define RExC_precomp (pRExC_state->precomp)
183 #define RExC_rx_sv (pRExC_state->rx_sv)
184 #define RExC_rx (pRExC_state->rx)
185 #define RExC_rxi (pRExC_state->rxi)
186 #define RExC_start (pRExC_state->start)
187 #define RExC_end (pRExC_state->end)
188 #define RExC_parse (pRExC_state->parse)
189 #define RExC_whilem_seen (pRExC_state->whilem_seen)
190 #ifdef RE_TRACK_PATTERN_OFFSETS
191 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
194 #define RExC_emit (pRExC_state->emit)
195 #define RExC_emit_dummy (pRExC_state->emit_dummy)
196 #define RExC_emit_start (pRExC_state->emit_start)
197 #define RExC_emit_bound (pRExC_state->emit_bound)
198 #define RExC_naughty (pRExC_state->naughty)
199 #define RExC_sawback (pRExC_state->sawback)
200 #define RExC_seen (pRExC_state->seen)
201 #define RExC_size (pRExC_state->size)
202 #define RExC_maxlen (pRExC_state->maxlen)
203 #define RExC_npar (pRExC_state->npar)
204 #define RExC_nestroot (pRExC_state->nestroot)
205 #define RExC_extralen (pRExC_state->extralen)
206 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
207 #define RExC_utf8 (pRExC_state->utf8)
208 #define RExC_uni_semantics (pRExC_state->uni_semantics)
209 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
210 #define RExC_open_parens (pRExC_state->open_parens)
211 #define RExC_close_parens (pRExC_state->close_parens)
212 #define RExC_opend (pRExC_state->opend)
213 #define RExC_paren_names (pRExC_state->paren_names)
214 #define RExC_recurse (pRExC_state->recurse)
215 #define RExC_recurse_count (pRExC_state->recurse_count)
216 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
217 #define RExC_study_chunk_recursed_bytes \
218 (pRExC_state->study_chunk_recursed_bytes)
219 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
220 #define RExC_contains_locale (pRExC_state->contains_locale)
221 #define RExC_contains_i (pRExC_state->contains_i)
222 #define RExC_override_recoding (pRExC_state->override_recoding)
223 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
226 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
227 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
228 ((*s) == '{' && regcurly(s)))
231 * Flags to be passed up and down.
233 #define WORST 0 /* Worst case. */
234 #define HASWIDTH 0x01 /* Known to match non-null strings. */
236 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
237 * character. (There needs to be a case: in the switch statement in regexec.c
238 * for any node marked SIMPLE.) Note that this is not the same thing as
241 #define SPSTART 0x04 /* Starts with * or + */
242 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
243 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
244 #define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
246 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
248 /* whether trie related optimizations are enabled */
249 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
250 #define TRIE_STUDY_OPT
251 #define FULL_TRIE_STUDY
257 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
258 #define PBITVAL(paren) (1 << ((paren) & 7))
259 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
260 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
261 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
263 #define REQUIRE_UTF8 STMT_START { \
265 *flagp = RESTART_UTF8; \
270 /* This converts the named class defined in regcomp.h to its equivalent class
271 * number defined in handy.h. */
272 #define namedclass_to_classnum(class) ((int) ((class) / 2))
273 #define classnum_to_namedclass(classnum) ((classnum) * 2)
275 #define _invlist_union_complement_2nd(a, b, output) \
276 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
277 #define _invlist_intersection_complement_2nd(a, b, output) \
278 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
280 /* About scan_data_t.
282 During optimisation we recurse through the regexp program performing
283 various inplace (keyhole style) optimisations. In addition study_chunk
284 and scan_commit populate this data structure with information about
285 what strings MUST appear in the pattern. We look for the longest
286 string that must appear at a fixed location, and we look for the
287 longest string that may appear at a floating location. So for instance
292 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
293 strings (because they follow a .* construct). study_chunk will identify
294 both FOO and BAR as being the longest fixed and floating strings respectively.
296 The strings can be composites, for instance
300 will result in a composite fixed substring 'foo'.
302 For each string some basic information is maintained:
304 - offset or min_offset
305 This is the position the string must appear at, or not before.
306 It also implicitly (when combined with minlenp) tells us how many
307 characters must match before the string we are searching for.
308 Likewise when combined with minlenp and the length of the string it
309 tells us how many characters must appear after the string we have
313 Only used for floating strings. This is the rightmost point that
314 the string can appear at. If set to SSize_t_MAX it indicates that the
315 string can occur infinitely far to the right.
318 A pointer to the minimum number of characters of the pattern that the
319 string was found inside. This is important as in the case of positive
320 lookahead or positive lookbehind we can have multiple patterns
325 The minimum length of the pattern overall is 3, the minimum length
326 of the lookahead part is 3, but the minimum length of the part that
327 will actually match is 1. So 'FOO's minimum length is 3, but the
328 minimum length for the F is 1. This is important as the minimum length
329 is used to determine offsets in front of and behind the string being
330 looked for. Since strings can be composites this is the length of the
331 pattern at the time it was committed with a scan_commit. Note that
332 the length is calculated by study_chunk, so that the minimum lengths
333 are not known until the full pattern has been compiled, thus the
334 pointer to the value.
338 In the case of lookbehind the string being searched for can be
339 offset past the start point of the final matching string.
340 If this value was just blithely removed from the min_offset it would
341 invalidate some of the calculations for how many chars must match
342 before or after (as they are derived from min_offset and minlen and
343 the length of the string being searched for).
344 When the final pattern is compiled and the data is moved from the
345 scan_data_t structure into the regexp structure the information
346 about lookbehind is factored in, with the information that would
347 have been lost precalculated in the end_shift field for the
350 The fields pos_min and pos_delta are used to store the minimum offset
351 and the delta to the maximum offset at the current point in the pattern.
355 typedef struct scan_data_t {
356 /*I32 len_min; unused */
357 /*I32 len_delta; unused */
361 SSize_t last_end; /* min value, <0 unless valid. */
362 SSize_t last_start_min;
363 SSize_t last_start_max;
364 SV **longest; /* Either &l_fixed, or &l_float. */
365 SV *longest_fixed; /* longest fixed string found in pattern */
366 SSize_t offset_fixed; /* offset where it starts */
367 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
368 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
369 SV *longest_float; /* longest floating string found in pattern */
370 SSize_t offset_float_min; /* earliest point in string it can appear */
371 SSize_t offset_float_max; /* latest point in string it can appear */
372 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
373 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
376 SSize_t *last_closep;
377 regnode_ssc *start_class;
380 /* The below is perhaps overboard, but this allows us to save a test at the
381 * expense of a mask. This is because on both EBCDIC and ASCII machines, 'A'
382 * and 'a' differ by a single bit; the same with the upper and lower case of
383 * all other ASCII-range alphabetics. On ASCII platforms, they are 32 apart;
384 * on EBCDIC, they are 64. This uses an exclusive 'or' to find that bit and
385 * then inverts it to form a mask, with just a single 0, in the bit position
386 * where the upper- and lowercase differ. XXX There are about 40 other
387 * instances in the Perl core where this micro-optimization could be used.
388 * Should decide if maintenance cost is worse, before changing those
390 * Returns a boolean as to whether or not 'v' is either a lowercase or
391 * uppercase instance of 'c', where 'c' is in [A-Za-z]. If 'c' is a
392 * compile-time constant, the generated code is better than some optimizing
393 * compilers figure out, amounting to a mask and test. The results are
394 * meaningless if 'c' is not one of [A-Za-z] */
395 #define isARG2_lower_or_UPPER_ARG1(c, v) \
396 (((v) & ~('A' ^ 'a')) == ((c) & ~('A' ^ 'a')))
399 * Forward declarations for pregcomp()'s friends.
402 static const scan_data_t zero_scan_data =
403 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
405 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
406 #define SF_BEFORE_SEOL 0x0001
407 #define SF_BEFORE_MEOL 0x0002
408 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
409 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
411 #define SF_FIX_SHIFT_EOL (+2)
412 #define SF_FL_SHIFT_EOL (+4)
414 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
415 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
417 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
418 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
419 #define SF_IS_INF 0x0040
420 #define SF_HAS_PAR 0x0080
421 #define SF_IN_PAR 0x0100
422 #define SF_HAS_EVAL 0x0200
423 #define SCF_DO_SUBSTR 0x0400
424 #define SCF_DO_STCLASS_AND 0x0800
425 #define SCF_DO_STCLASS_OR 0x1000
426 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
427 #define SCF_WHILEM_VISITED_POS 0x2000
429 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
430 #define SCF_SEEN_ACCEPT 0x8000
431 #define SCF_TRIE_DOING_RESTUDY 0x10000
433 #define UTF cBOOL(RExC_utf8)
435 /* The enums for all these are ordered so things work out correctly */
436 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
437 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
438 == REGEX_DEPENDS_CHARSET)
439 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
440 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
441 >= REGEX_UNICODE_CHARSET)
442 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
443 == REGEX_ASCII_RESTRICTED_CHARSET)
444 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
445 >= REGEX_ASCII_RESTRICTED_CHARSET)
446 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
447 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
449 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
451 /* For programs that want to be strictly Unicode compatible by dying if any
452 * attempt is made to match a non-Unicode code point against a Unicode
454 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
456 #define OOB_NAMEDCLASS -1
458 /* There is no code point that is out-of-bounds, so this is problematic. But
459 * its only current use is to initialize a variable that is always set before
461 #define OOB_UNICODE 0xDEADBEEF
463 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
464 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
467 /* length of regex to show in messages that don't mark a position within */
468 #define RegexLengthToShowInErrorMessages 127
471 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
472 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
473 * op/pragma/warn/regcomp.
475 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
476 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
478 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
479 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
481 #define REPORT_LOCATION_ARGS(offset) \
482 UTF8fARG(UTF, offset, RExC_precomp), \
483 UTF8fARG(UTF, RExC_end - RExC_precomp - offset, RExC_precomp + offset)
486 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
487 * arg. Show regex, up to a maximum length. If it's too long, chop and add
490 #define _FAIL(code) STMT_START { \
491 const char *ellipses = ""; \
492 IV len = RExC_end - RExC_precomp; \
495 SAVEFREESV(RExC_rx_sv); \
496 if (len > RegexLengthToShowInErrorMessages) { \
497 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
498 len = RegexLengthToShowInErrorMessages - 10; \
504 #define FAIL(msg) _FAIL( \
505 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
506 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
508 #define FAIL2(msg,arg) _FAIL( \
509 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
510 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
513 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
515 #define Simple_vFAIL(m) STMT_START { \
516 const IV offset = RExC_parse - RExC_precomp; \
517 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
518 m, REPORT_LOCATION_ARGS(offset)); \
522 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
524 #define vFAIL(m) STMT_START { \
526 SAVEFREESV(RExC_rx_sv); \
531 * Like Simple_vFAIL(), but accepts two arguments.
533 #define Simple_vFAIL2(m,a1) STMT_START { \
534 const IV offset = RExC_parse - RExC_precomp; \
535 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
536 REPORT_LOCATION_ARGS(offset)); \
540 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
542 #define vFAIL2(m,a1) STMT_START { \
544 SAVEFREESV(RExC_rx_sv); \
545 Simple_vFAIL2(m, a1); \
550 * Like Simple_vFAIL(), but accepts three arguments.
552 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
553 const IV offset = RExC_parse - RExC_precomp; \
554 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
555 REPORT_LOCATION_ARGS(offset)); \
559 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
561 #define vFAIL3(m,a1,a2) STMT_START { \
563 SAVEFREESV(RExC_rx_sv); \
564 Simple_vFAIL3(m, a1, a2); \
568 * Like Simple_vFAIL(), but accepts four arguments.
570 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
571 const IV offset = RExC_parse - RExC_precomp; \
572 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
573 REPORT_LOCATION_ARGS(offset)); \
576 #define vFAIL4(m,a1,a2,a3) STMT_START { \
578 SAVEFREESV(RExC_rx_sv); \
579 Simple_vFAIL4(m, a1, a2, a3); \
582 /* A specialized version of vFAIL2 that works with UTF8f */
583 #define vFAIL2utf8f(m, a1) STMT_START { \
584 const IV offset = RExC_parse - RExC_precomp; \
586 SAVEFREESV(RExC_rx_sv); \
587 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
588 REPORT_LOCATION_ARGS(offset)); \
592 /* m is not necessarily a "literal string", in this macro */
593 #define reg_warn_non_literal_string(loc, m) STMT_START { \
594 const IV offset = loc - RExC_precomp; \
595 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
596 m, REPORT_LOCATION_ARGS(offset)); \
599 #define ckWARNreg(loc,m) STMT_START { \
600 const IV offset = loc - RExC_precomp; \
601 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
602 REPORT_LOCATION_ARGS(offset)); \
605 #define vWARN_dep(loc, m) STMT_START { \
606 const IV offset = loc - RExC_precomp; \
607 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
608 REPORT_LOCATION_ARGS(offset)); \
611 #define ckWARNdep(loc,m) STMT_START { \
612 const IV offset = loc - RExC_precomp; \
613 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
615 REPORT_LOCATION_ARGS(offset)); \
618 #define ckWARNregdep(loc,m) STMT_START { \
619 const IV offset = loc - RExC_precomp; \
620 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
622 REPORT_LOCATION_ARGS(offset)); \
625 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
626 const IV offset = loc - RExC_precomp; \
627 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
629 a1, REPORT_LOCATION_ARGS(offset)); \
632 #define ckWARN2reg(loc, m, a1) STMT_START { \
633 const IV offset = loc - RExC_precomp; \
634 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
635 a1, REPORT_LOCATION_ARGS(offset)); \
638 #define vWARN3(loc, m, a1, a2) STMT_START { \
639 const IV offset = loc - RExC_precomp; \
640 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
641 a1, a2, REPORT_LOCATION_ARGS(offset)); \
644 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
645 const IV offset = loc - RExC_precomp; \
646 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
647 a1, a2, REPORT_LOCATION_ARGS(offset)); \
650 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
651 const IV offset = loc - RExC_precomp; \
652 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
653 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
656 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
657 const IV offset = loc - RExC_precomp; \
658 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
659 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
662 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
663 const IV offset = loc - RExC_precomp; \
664 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
665 a1, a2, a3, a4, REPORT_LOCATION_ARGS(offset)); \
669 /* Allow for side effects in s */
670 #define REGC(c,s) STMT_START { \
671 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
674 /* Macros for recording node offsets. 20001227 mjd@plover.com
675 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
676 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
677 * Element 0 holds the number n.
678 * Position is 1 indexed.
680 #ifndef RE_TRACK_PATTERN_OFFSETS
681 #define Set_Node_Offset_To_R(node,byte)
682 #define Set_Node_Offset(node,byte)
683 #define Set_Cur_Node_Offset
684 #define Set_Node_Length_To_R(node,len)
685 #define Set_Node_Length(node,len)
686 #define Set_Node_Cur_Length(node,start)
687 #define Node_Offset(n)
688 #define Node_Length(n)
689 #define Set_Node_Offset_Length(node,offset,len)
690 #define ProgLen(ri) ri->u.proglen
691 #define SetProgLen(ri,x) ri->u.proglen = x
693 #define ProgLen(ri) ri->u.offsets[0]
694 #define SetProgLen(ri,x) ri->u.offsets[0] = x
695 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
697 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
698 __LINE__, (int)(node), (int)(byte))); \
700 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
703 RExC_offsets[2*(node)-1] = (byte); \
708 #define Set_Node_Offset(node,byte) \
709 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
710 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
712 #define Set_Node_Length_To_R(node,len) STMT_START { \
714 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
715 __LINE__, (int)(node), (int)(len))); \
717 Perl_croak(aTHX_ "value of node is %d in Length macro", \
720 RExC_offsets[2*(node)] = (len); \
725 #define Set_Node_Length(node,len) \
726 Set_Node_Length_To_R((node)-RExC_emit_start, len)
727 #define Set_Node_Cur_Length(node, start) \
728 Set_Node_Length(node, RExC_parse - start)
730 /* Get offsets and lengths */
731 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
732 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
734 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
735 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
736 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
740 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
741 #define EXPERIMENTAL_INPLACESCAN
742 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
744 #define DEBUG_RExC_seen() \
745 DEBUG_OPTIMISE_MORE_r({ \
746 PerlIO_printf(Perl_debug_log,"RExC_seen: "); \
748 if (RExC_seen & REG_ZERO_LEN_SEEN) \
749 PerlIO_printf(Perl_debug_log,"REG_ZERO_LEN_SEEN "); \
751 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
752 PerlIO_printf(Perl_debug_log,"REG_LOOKBEHIND_SEEN "); \
754 if (RExC_seen & REG_GPOS_SEEN) \
755 PerlIO_printf(Perl_debug_log,"REG_GPOS_SEEN "); \
757 if (RExC_seen & REG_CANY_SEEN) \
758 PerlIO_printf(Perl_debug_log,"REG_CANY_SEEN "); \
760 if (RExC_seen & REG_RECURSE_SEEN) \
761 PerlIO_printf(Perl_debug_log,"REG_RECURSE_SEEN "); \
763 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
764 PerlIO_printf(Perl_debug_log,"REG_TOP_LEVEL_BRANCHES_SEEN "); \
766 if (RExC_seen & REG_VERBARG_SEEN) \
767 PerlIO_printf(Perl_debug_log,"REG_VERBARG_SEEN "); \
769 if (RExC_seen & REG_CUTGROUP_SEEN) \
770 PerlIO_printf(Perl_debug_log,"REG_CUTGROUP_SEEN "); \
772 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
773 PerlIO_printf(Perl_debug_log,"REG_RUN_ON_COMMENT_SEEN "); \
775 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
776 PerlIO_printf(Perl_debug_log,"REG_UNFOLDED_MULTI_SEEN "); \
778 if (RExC_seen & REG_GOSTART_SEEN) \
779 PerlIO_printf(Perl_debug_log,"REG_GOSTART_SEEN "); \
781 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
782 PerlIO_printf(Perl_debug_log,"REG_UNBOUNDED_QUANTIFIER_SEEN "); \
784 PerlIO_printf(Perl_debug_log,"\n"); \
787 #define DEBUG_STUDYDATA(str,data,depth) \
788 DEBUG_OPTIMISE_MORE_r(if(data){ \
789 PerlIO_printf(Perl_debug_log, \
790 "%*s" str "Pos:%"IVdf"/%"IVdf \
791 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
792 (int)(depth)*2, "", \
793 (IV)((data)->pos_min), \
794 (IV)((data)->pos_delta), \
795 (UV)((data)->flags), \
796 (IV)((data)->whilem_c), \
797 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
798 is_inf ? "INF " : "" \
800 if ((data)->last_found) \
801 PerlIO_printf(Perl_debug_log, \
802 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
803 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
804 SvPVX_const((data)->last_found), \
805 (IV)((data)->last_end), \
806 (IV)((data)->last_start_min), \
807 (IV)((data)->last_start_max), \
808 ((data)->longest && \
809 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
810 SvPVX_const((data)->longest_fixed), \
811 (IV)((data)->offset_fixed), \
812 ((data)->longest && \
813 (data)->longest==&((data)->longest_float)) ? "*" : "", \
814 SvPVX_const((data)->longest_float), \
815 (IV)((data)->offset_float_min), \
816 (IV)((data)->offset_float_max) \
818 PerlIO_printf(Perl_debug_log,"\n"); \
821 /* Mark that we cannot extend a found fixed substring at this point.
822 Update the longest found anchored substring and the longest found
823 floating substrings if needed. */
826 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
827 SSize_t *minlenp, int is_inf)
829 const STRLEN l = CHR_SVLEN(data->last_found);
830 const STRLEN old_l = CHR_SVLEN(*data->longest);
831 GET_RE_DEBUG_FLAGS_DECL;
833 PERL_ARGS_ASSERT_SCAN_COMMIT;
835 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
836 SvSetMagicSV(*data->longest, data->last_found);
837 if (*data->longest == data->longest_fixed) {
838 data->offset_fixed = l ? data->last_start_min : data->pos_min;
839 if (data->flags & SF_BEFORE_EOL)
841 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
843 data->flags &= ~SF_FIX_BEFORE_EOL;
844 data->minlen_fixed=minlenp;
845 data->lookbehind_fixed=0;
847 else { /* *data->longest == data->longest_float */
848 data->offset_float_min = l ? data->last_start_min : data->pos_min;
849 data->offset_float_max = (l
850 ? data->last_start_max
851 : (data->pos_delta == SSize_t_MAX
853 : data->pos_min + data->pos_delta));
855 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
856 data->offset_float_max = SSize_t_MAX;
857 if (data->flags & SF_BEFORE_EOL)
859 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
861 data->flags &= ~SF_FL_BEFORE_EOL;
862 data->minlen_float=minlenp;
863 data->lookbehind_float=0;
866 SvCUR_set(data->last_found, 0);
868 SV * const sv = data->last_found;
869 if (SvUTF8(sv) && SvMAGICAL(sv)) {
870 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
876 data->flags &= ~SF_BEFORE_EOL;
877 DEBUG_STUDYDATA("commit: ",data,0);
880 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
881 * list that describes which code points it matches */
884 S_ssc_anything(pTHX_ regnode_ssc *ssc)
886 /* Set the SSC 'ssc' to match an empty string or any code point */
888 PERL_ARGS_ASSERT_SSC_ANYTHING;
890 assert(is_ANYOF_SYNTHETIC(ssc));
892 ssc->invlist = sv_2mortal(_new_invlist(2)); /* mortalize so won't leak */
893 _append_range_to_invlist(ssc->invlist, 0, UV_MAX);
894 ANYOF_FLAGS(ssc) |= ANYOF_EMPTY_STRING; /* Plus match empty string */
898 S_ssc_is_anything(const regnode_ssc *ssc)
900 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
901 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
902 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
903 * in any way, so there's no point in using it */
908 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
910 assert(is_ANYOF_SYNTHETIC(ssc));
912 if (! (ANYOF_FLAGS(ssc) & ANYOF_EMPTY_STRING)) {
916 /* See if the list consists solely of the range 0 - Infinity */
917 invlist_iterinit(ssc->invlist);
918 ret = invlist_iternext(ssc->invlist, &start, &end)
922 invlist_iterfinish(ssc->invlist);
928 /* If e.g., both \w and \W are set, matches everything */
929 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
931 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
932 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
942 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
944 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
945 * string, any code point, or any posix class under locale */
947 PERL_ARGS_ASSERT_SSC_INIT;
949 Zero(ssc, 1, regnode_ssc);
950 set_ANYOF_SYNTHETIC(ssc);
951 ARG_SET(ssc, ANYOF_NONBITMAP_EMPTY);
954 /* If any portion of the regex is to operate under locale rules,
955 * initialization includes it. The reason this isn't done for all regexes
956 * is that the optimizer was written under the assumption that locale was
957 * all-or-nothing. Given the complexity and lack of documentation in the
958 * optimizer, and that there are inadequate test cases for locale, many
959 * parts of it may not work properly, it is safest to avoid locale unless
961 if (RExC_contains_locale) {
962 ANYOF_POSIXL_SETALL(ssc);
965 ANYOF_POSIXL_ZERO(ssc);
970 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
971 const regnode_ssc *ssc)
973 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
974 * to the list of code points matched, and locale posix classes; hence does
975 * not check its flags) */
980 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
982 assert(is_ANYOF_SYNTHETIC(ssc));
984 invlist_iterinit(ssc->invlist);
985 ret = invlist_iternext(ssc->invlist, &start, &end)
989 invlist_iterfinish(ssc->invlist);
995 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1003 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1004 const regnode_charclass* const node)
1006 /* Returns a mortal inversion list defining which code points are matched
1007 * by 'node', which is of type ANYOF. Handles complementing the result if
1008 * appropriate. If some code points aren't knowable at this time, the
1009 * returned list must, and will, contain every code point that is a
1012 SV* invlist = sv_2mortal(_new_invlist(0));
1013 SV* only_utf8_locale_invlist = NULL;
1015 const U32 n = ARG(node);
1016 bool new_node_has_latin1 = FALSE;
1018 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1020 /* Look at the data structure created by S_set_ANYOF_arg() */
1021 if (n != ANYOF_NONBITMAP_EMPTY) {
1022 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1023 AV * const av = MUTABLE_AV(SvRV(rv));
1024 SV **const ary = AvARRAY(av);
1025 assert(RExC_rxi->data->what[n] == 's');
1027 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1028 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1030 else if (ary[0] && ary[0] != &PL_sv_undef) {
1032 /* Here, no compile-time swash, and there are things that won't be
1033 * known until runtime -- we have to assume it could be anything */
1034 return _add_range_to_invlist(invlist, 0, UV_MAX);
1036 else if (ary[3] && ary[3] != &PL_sv_undef) {
1038 /* Here no compile-time swash, and no run-time only data. Use the
1039 * node's inversion list */
1040 invlist = sv_2mortal(invlist_clone(ary[3]));
1043 /* Get the code points valid only under UTF-8 locales */
1044 if ((ANYOF_FLAGS(node) & ANYOF_LOC_FOLD)
1045 && ary[2] && ary[2] != &PL_sv_undef)
1047 only_utf8_locale_invlist = ary[2];
1051 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1052 * code points, and an inversion list for the others, but if there are code
1053 * points that should match only conditionally on the target string being
1054 * UTF-8, those are placed in the inversion list, and not the bitmap.
1055 * Since there are circumstances under which they could match, they are
1056 * included in the SSC. But if the ANYOF node is to be inverted, we have
1057 * to exclude them here, so that when we invert below, the end result
1058 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1059 * have to do this here before we add the unconditionally matched code
1061 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1062 _invlist_intersection_complement_2nd(invlist,
1067 /* Add in the points from the bit map */
1068 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1069 if (ANYOF_BITMAP_TEST(node, i)) {
1070 invlist = add_cp_to_invlist(invlist, i);
1071 new_node_has_latin1 = TRUE;
1075 /* If this can match all upper Latin1 code points, have to add them
1077 if (ANYOF_FLAGS(node) & ANYOF_NON_UTF8_NON_ASCII_ALL) {
1078 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1081 /* Similarly for these */
1082 if (ANYOF_FLAGS(node) & ANYOF_ABOVE_LATIN1_ALL) {
1083 invlist = _add_range_to_invlist(invlist, 256, UV_MAX);
1086 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1087 _invlist_invert(invlist);
1089 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOF_LOC_FOLD) {
1091 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1092 * locale. We can skip this if there are no 0-255 at all. */
1093 _invlist_union(invlist, PL_Latin1, &invlist);
1096 /* Similarly add the UTF-8 locale possible matches. These have to be
1097 * deferred until after the non-UTF-8 locale ones are taken care of just
1098 * above, or it leads to wrong results under ANYOF_INVERT */
1099 if (only_utf8_locale_invlist) {
1100 _invlist_union_maybe_complement_2nd(invlist,
1101 only_utf8_locale_invlist,
1102 ANYOF_FLAGS(node) & ANYOF_INVERT,
1109 /* These two functions currently do the exact same thing */
1110 #define ssc_init_zero ssc_init
1112 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1113 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1115 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1116 * should not be inverted. 'and_with->flags & ANYOF_POSIXL' should be 0 if
1117 * 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1120 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1121 const regnode_charclass *and_with)
1123 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1124 * another SSC or a regular ANYOF class. Can create false positives. */
1129 PERL_ARGS_ASSERT_SSC_AND;
1131 assert(is_ANYOF_SYNTHETIC(ssc));
1133 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1134 * the code point inversion list and just the relevant flags */
1135 if (is_ANYOF_SYNTHETIC(and_with)) {
1136 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1137 anded_flags = ANYOF_FLAGS(and_with);
1139 /* XXX This is a kludge around what appears to be deficiencies in the
1140 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1141 * there are paths through the optimizer where it doesn't get weeded
1142 * out when it should. And if we don't make some extra provision for
1143 * it like the code just below, it doesn't get added when it should.
1144 * This solution is to add it only when AND'ing, which is here, and
1145 * only when what is being AND'ed is the pristine, original node
1146 * matching anything. Thus it is like adding it to ssc_anything() but
1147 * only when the result is to be AND'ed. Probably the same solution
1148 * could be adopted for the same problem we have with /l matching,
1149 * which is solved differently in S_ssc_init(), and that would lead to
1150 * fewer false positives than that solution has. But if this solution
1151 * creates bugs, the consequences are only that a warning isn't raised
1152 * that should be; while the consequences for having /l bugs is
1153 * incorrect matches */
1154 if (ssc_is_anything((regnode_ssc *)and_with)) {
1155 anded_flags |= ANYOF_WARN_SUPER;
1159 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1160 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1163 ANYOF_FLAGS(ssc) &= anded_flags;
1165 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1166 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1167 * 'and_with' may be inverted. When not inverted, we have the situation of
1169 * (C1 | P1) & (C2 | P2)
1170 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1171 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1172 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1173 * <= ((C1 & C2) | P1 | P2)
1174 * Alternatively, the last few steps could be:
1175 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1176 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1177 * <= (C1 | C2 | (P1 & P2))
1178 * We favor the second approach if either P1 or P2 is non-empty. This is
1179 * because these components are a barrier to doing optimizations, as what
1180 * they match cannot be known until the moment of matching as they are
1181 * dependent on the current locale, 'AND"ing them likely will reduce or
1183 * But we can do better if we know that C1,P1 are in their initial state (a
1184 * frequent occurrence), each matching everything:
1185 * (<everything>) & (C2 | P2) = C2 | P2
1186 * Similarly, if C2,P2 are in their initial state (again a frequent
1187 * occurrence), the result is a no-op
1188 * (C1 | P1) & (<everything>) = C1 | P1
1191 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1192 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1193 * <= (C1 & ~C2) | (P1 & ~P2)
1196 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1197 && ! is_ANYOF_SYNTHETIC(and_with))
1201 ssc_intersection(ssc,
1203 FALSE /* Has already been inverted */
1206 /* If either P1 or P2 is empty, the intersection will be also; can skip
1208 if (! (ANYOF_FLAGS(and_with) & ANYOF_POSIXL)) {
1209 ANYOF_POSIXL_ZERO(ssc);
1211 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1213 /* Note that the Posix class component P from 'and_with' actually
1215 * P = Pa | Pb | ... | Pn
1216 * where each component is one posix class, such as in [\w\s].
1218 * ~P = ~(Pa | Pb | ... | Pn)
1219 * = ~Pa & ~Pb & ... & ~Pn
1220 * <= ~Pa | ~Pb | ... | ~Pn
1221 * The last is something we can easily calculate, but unfortunately
1222 * is likely to have many false positives. We could do better
1223 * in some (but certainly not all) instances if two classes in
1224 * P have known relationships. For example
1225 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1227 * :lower: & :print: = :lower:
1228 * And similarly for classes that must be disjoint. For example,
1229 * since \s and \w can have no elements in common based on rules in
1230 * the POSIX standard,
1231 * \w & ^\S = nothing
1232 * Unfortunately, some vendor locales do not meet the Posix
1233 * standard, in particular almost everything by Microsoft.
1234 * The loop below just changes e.g., \w into \W and vice versa */
1236 regnode_charclass_posixl temp;
1237 int add = 1; /* To calculate the index of the complement */
1239 ANYOF_POSIXL_ZERO(&temp);
1240 for (i = 0; i < ANYOF_MAX; i++) {
1242 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1243 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1245 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1246 ANYOF_POSIXL_SET(&temp, i + add);
1248 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1250 ANYOF_POSIXL_AND(&temp, ssc);
1252 } /* else ssc already has no posixes */
1253 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1254 in its initial state */
1255 else if (! is_ANYOF_SYNTHETIC(and_with)
1256 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1258 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1259 * copy it over 'ssc' */
1260 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1261 if (is_ANYOF_SYNTHETIC(and_with)) {
1262 StructCopy(and_with, ssc, regnode_ssc);
1265 ssc->invlist = anded_cp_list;
1266 ANYOF_POSIXL_ZERO(ssc);
1267 if (ANYOF_FLAGS(and_with) & ANYOF_POSIXL) {
1268 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1272 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1273 || (ANYOF_FLAGS(and_with) & ANYOF_POSIXL))
1275 /* One or the other of P1, P2 is non-empty. */
1276 if (ANYOF_FLAGS(and_with) & ANYOF_POSIXL) {
1277 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1279 ssc_union(ssc, anded_cp_list, FALSE);
1281 else { /* P1 = P2 = empty */
1282 ssc_intersection(ssc, anded_cp_list, FALSE);
1288 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1289 const regnode_charclass *or_with)
1291 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1292 * another SSC or a regular ANYOF class. Can create false positives if
1293 * 'or_with' is to be inverted. */
1298 PERL_ARGS_ASSERT_SSC_OR;
1300 assert(is_ANYOF_SYNTHETIC(ssc));
1302 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1303 * the code point inversion list and just the relevant flags */
1304 if (is_ANYOF_SYNTHETIC(or_with)) {
1305 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1306 ored_flags = ANYOF_FLAGS(or_with);
1309 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1310 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1313 ANYOF_FLAGS(ssc) |= ored_flags;
1315 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1316 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1317 * 'or_with' may be inverted. When not inverted, we have the simple
1318 * situation of computing:
1319 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1320 * If P1|P2 yields a situation with both a class and its complement are
1321 * set, like having both \w and \W, this matches all code points, and we
1322 * can delete these from the P component of the ssc going forward. XXX We
1323 * might be able to delete all the P components, but I (khw) am not certain
1324 * about this, and it is better to be safe.
1327 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1328 * <= (C1 | P1) | ~C2
1329 * <= (C1 | ~C2) | P1
1330 * (which results in actually simpler code than the non-inverted case)
1333 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1334 && ! is_ANYOF_SYNTHETIC(or_with))
1336 /* We ignore P2, leaving P1 going forward */
1337 } /* else Not inverted */
1338 else if (ANYOF_FLAGS(or_with) & ANYOF_POSIXL) {
1339 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1340 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1342 for (i = 0; i < ANYOF_MAX; i += 2) {
1343 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1345 ssc_match_all_cp(ssc);
1346 ANYOF_POSIXL_CLEAR(ssc, i);
1347 ANYOF_POSIXL_CLEAR(ssc, i+1);
1355 FALSE /* Already has been inverted */
1359 PERL_STATIC_INLINE void
1360 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1362 PERL_ARGS_ASSERT_SSC_UNION;
1364 assert(is_ANYOF_SYNTHETIC(ssc));
1366 _invlist_union_maybe_complement_2nd(ssc->invlist,
1372 PERL_STATIC_INLINE void
1373 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1375 const bool invert2nd)
1377 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1379 assert(is_ANYOF_SYNTHETIC(ssc));
1381 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1387 PERL_STATIC_INLINE void
1388 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1390 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1392 assert(is_ANYOF_SYNTHETIC(ssc));
1394 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1397 PERL_STATIC_INLINE void
1398 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1400 /* AND just the single code point 'cp' into the SSC 'ssc' */
1402 SV* cp_list = _new_invlist(2);
1404 PERL_ARGS_ASSERT_SSC_CP_AND;
1406 assert(is_ANYOF_SYNTHETIC(ssc));
1408 cp_list = add_cp_to_invlist(cp_list, cp);
1409 ssc_intersection(ssc, cp_list,
1410 FALSE /* Not inverted */
1412 SvREFCNT_dec_NN(cp_list);
1415 PERL_STATIC_INLINE void
1416 S_ssc_clear_locale(regnode_ssc *ssc)
1418 /* Set the SSC 'ssc' to not match any locale things */
1419 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1421 assert(is_ANYOF_SYNTHETIC(ssc));
1423 ANYOF_POSIXL_ZERO(ssc);
1424 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1428 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1430 /* The inversion list in the SSC is marked mortal; now we need a more
1431 * permanent copy, which is stored the same way that is done in a regular
1432 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1435 SV* invlist = invlist_clone(ssc->invlist);
1437 PERL_ARGS_ASSERT_SSC_FINALIZE;
1439 assert(is_ANYOF_SYNTHETIC(ssc));
1441 /* The code in this file assumes that all but these flags aren't relevant
1442 * to the SSC, except ANYOF_EMPTY_STRING, which should be cleared by the
1443 * time we reach here */
1444 assert(! (ANYOF_FLAGS(ssc) & ~ANYOF_COMMON_FLAGS));
1446 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1448 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1449 NULL, NULL, NULL, FALSE);
1451 /* Make sure is clone-safe */
1452 ssc->invlist = NULL;
1454 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1455 ANYOF_FLAGS(ssc) |= ANYOF_POSIXL;
1458 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1461 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1462 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1463 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1464 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1465 ? (TRIE_LIST_CUR( idx ) - 1) \
1471 dump_trie(trie,widecharmap,revcharmap)
1472 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1473 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1475 These routines dump out a trie in a somewhat readable format.
1476 The _interim_ variants are used for debugging the interim
1477 tables that are used to generate the final compressed
1478 representation which is what dump_trie expects.
1480 Part of the reason for their existence is to provide a form
1481 of documentation as to how the different representations function.
1486 Dumps the final compressed table form of the trie to Perl_debug_log.
1487 Used for debugging make_trie().
1491 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1492 AV *revcharmap, U32 depth)
1495 SV *sv=sv_newmortal();
1496 int colwidth= widecharmap ? 6 : 4;
1498 GET_RE_DEBUG_FLAGS_DECL;
1500 PERL_ARGS_ASSERT_DUMP_TRIE;
1502 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1503 (int)depth * 2 + 2,"",
1504 "Match","Base","Ofs" );
1506 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1507 SV ** const tmp = av_fetch( revcharmap, state, 0);
1509 PerlIO_printf( Perl_debug_log, "%*s",
1511 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1512 PL_colors[0], PL_colors[1],
1513 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1514 PERL_PV_ESCAPE_FIRSTCHAR
1519 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1520 (int)depth * 2 + 2,"");
1522 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1523 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1524 PerlIO_printf( Perl_debug_log, "\n");
1526 for( state = 1 ; state < trie->statecount ; state++ ) {
1527 const U32 base = trie->states[ state ].trans.base;
1529 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1530 (int)depth * 2 + 2,"", (UV)state);
1532 if ( trie->states[ state ].wordnum ) {
1533 PerlIO_printf( Perl_debug_log, " W%4X",
1534 trie->states[ state ].wordnum );
1536 PerlIO_printf( Perl_debug_log, "%6s", "" );
1539 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1544 while( ( base + ofs < trie->uniquecharcount ) ||
1545 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1546 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1550 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1552 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1553 if ( ( base + ofs >= trie->uniquecharcount )
1554 && ( base + ofs - trie->uniquecharcount
1556 && trie->trans[ base + ofs
1557 - trie->uniquecharcount ].check == state )
1559 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1561 (UV)trie->trans[ base + ofs
1562 - trie->uniquecharcount ].next );
1564 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1568 PerlIO_printf( Perl_debug_log, "]");
1571 PerlIO_printf( Perl_debug_log, "\n" );
1573 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1575 for (word=1; word <= trie->wordcount; word++) {
1576 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1577 (int)word, (int)(trie->wordinfo[word].prev),
1578 (int)(trie->wordinfo[word].len));
1580 PerlIO_printf(Perl_debug_log, "\n" );
1583 Dumps a fully constructed but uncompressed trie in list form.
1584 List tries normally only are used for construction when the number of
1585 possible chars (trie->uniquecharcount) is very high.
1586 Used for debugging make_trie().
1589 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1590 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1594 SV *sv=sv_newmortal();
1595 int colwidth= widecharmap ? 6 : 4;
1596 GET_RE_DEBUG_FLAGS_DECL;
1598 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1600 /* print out the table precompression. */
1601 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1602 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1603 "------:-----+-----------------\n" );
1605 for( state=1 ; state < next_alloc ; state ++ ) {
1608 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1609 (int)depth * 2 + 2,"", (UV)state );
1610 if ( ! trie->states[ state ].wordnum ) {
1611 PerlIO_printf( Perl_debug_log, "%5s| ","");
1613 PerlIO_printf( Perl_debug_log, "W%4x| ",
1614 trie->states[ state ].wordnum
1617 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1618 SV ** const tmp = av_fetch( revcharmap,
1619 TRIE_LIST_ITEM(state,charid).forid, 0);
1621 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1623 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1625 PL_colors[0], PL_colors[1],
1626 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1627 | PERL_PV_ESCAPE_FIRSTCHAR
1629 TRIE_LIST_ITEM(state,charid).forid,
1630 (UV)TRIE_LIST_ITEM(state,charid).newstate
1633 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1634 (int)((depth * 2) + 14), "");
1637 PerlIO_printf( Perl_debug_log, "\n");
1642 Dumps a fully constructed but uncompressed trie in table form.
1643 This is the normal DFA style state transition table, with a few
1644 twists to facilitate compression later.
1645 Used for debugging make_trie().
1648 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1649 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1654 SV *sv=sv_newmortal();
1655 int colwidth= widecharmap ? 6 : 4;
1656 GET_RE_DEBUG_FLAGS_DECL;
1658 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1661 print out the table precompression so that we can do a visual check
1662 that they are identical.
1665 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1667 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1668 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1670 PerlIO_printf( Perl_debug_log, "%*s",
1672 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1673 PL_colors[0], PL_colors[1],
1674 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1675 PERL_PV_ESCAPE_FIRSTCHAR
1681 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1683 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1684 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1687 PerlIO_printf( Perl_debug_log, "\n" );
1689 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1691 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1692 (int)depth * 2 + 2,"",
1693 (UV)TRIE_NODENUM( state ) );
1695 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1696 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1698 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1700 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1702 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1703 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1704 (UV)trie->trans[ state ].check );
1706 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1707 (UV)trie->trans[ state ].check,
1708 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1716 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1717 startbranch: the first branch in the whole branch sequence
1718 first : start branch of sequence of branch-exact nodes.
1719 May be the same as startbranch
1720 last : Thing following the last branch.
1721 May be the same as tail.
1722 tail : item following the branch sequence
1723 count : words in the sequence
1724 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS)/
1725 depth : indent depth
1727 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1729 A trie is an N'ary tree where the branches are determined by digital
1730 decomposition of the key. IE, at the root node you look up the 1st character and
1731 follow that branch repeat until you find the end of the branches. Nodes can be
1732 marked as "accepting" meaning they represent a complete word. Eg:
1736 would convert into the following structure. Numbers represent states, letters
1737 following numbers represent valid transitions on the letter from that state, if
1738 the number is in square brackets it represents an accepting state, otherwise it
1739 will be in parenthesis.
1741 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1745 (1) +-i->(6)-+-s->[7]
1747 +-s->(3)-+-h->(4)-+-e->[5]
1749 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1751 This shows that when matching against the string 'hers' we will begin at state 1
1752 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1753 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1754 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1755 single traverse. We store a mapping from accepting to state to which word was
1756 matched, and then when we have multiple possibilities we try to complete the
1757 rest of the regex in the order in which they occured in the alternation.
1759 The only prior NFA like behaviour that would be changed by the TRIE support is
1760 the silent ignoring of duplicate alternations which are of the form:
1762 / (DUPE|DUPE) X? (?{ ... }) Y /x
1764 Thus EVAL blocks following a trie may be called a different number of times with
1765 and without the optimisation. With the optimisations dupes will be silently
1766 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1767 the following demonstrates:
1769 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1771 which prints out 'word' three times, but
1773 'words'=~/(word|word|word)(?{ print $1 })S/
1775 which doesnt print it out at all. This is due to other optimisations kicking in.
1777 Example of what happens on a structural level:
1779 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1781 1: CURLYM[1] {1,32767}(18)
1792 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1793 and should turn into:
1795 1: CURLYM[1] {1,32767}(18)
1797 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1805 Cases where tail != last would be like /(?foo|bar)baz/:
1815 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1816 and would end up looking like:
1819 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1826 d = uvchr_to_utf8_flags(d, uv, 0);
1828 is the recommended Unicode-aware way of saying
1833 #define TRIE_STORE_REVCHAR(val) \
1836 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1837 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1838 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
1839 SvCUR_set(zlopp, kapow - flrbbbbb); \
1842 av_push(revcharmap, zlopp); \
1844 char ooooff = (char)val; \
1845 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1849 /* This gets the next character from the input, folding it if not already
1851 #define TRIE_READ_CHAR STMT_START { \
1854 /* if it is UTF then it is either already folded, or does not need \
1856 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
1858 else if (folder == PL_fold_latin1) { \
1859 /* This folder implies Unicode rules, which in the range expressible \
1860 * by not UTF is the lower case, with the two exceptions, one of \
1861 * which should have been taken care of before calling this */ \
1862 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
1863 uvc = toLOWER_L1(*uc); \
1864 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
1867 /* raw data, will be folded later if needed */ \
1875 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1876 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1877 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1878 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1880 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1881 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1882 TRIE_LIST_CUR( state )++; \
1885 #define TRIE_LIST_NEW(state) STMT_START { \
1886 Newxz( trie->states[ state ].trans.list, \
1887 4, reg_trie_trans_le ); \
1888 TRIE_LIST_CUR( state ) = 1; \
1889 TRIE_LIST_LEN( state ) = 4; \
1892 #define TRIE_HANDLE_WORD(state) STMT_START { \
1893 U16 dupe= trie->states[ state ].wordnum; \
1894 regnode * const noper_next = regnext( noper ); \
1897 /* store the word for dumping */ \
1899 if (OP(noper) != NOTHING) \
1900 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1902 tmp = newSVpvn_utf8( "", 0, UTF ); \
1903 av_push( trie_words, tmp ); \
1907 trie->wordinfo[curword].prev = 0; \
1908 trie->wordinfo[curword].len = wordlen; \
1909 trie->wordinfo[curword].accept = state; \
1911 if ( noper_next < tail ) { \
1913 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
1915 trie->jump[curword] = (U16)(noper_next - convert); \
1917 jumper = noper_next; \
1919 nextbranch= regnext(cur); \
1923 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1924 /* chain, so that when the bits of chain are later */\
1925 /* linked together, the dups appear in the chain */\
1926 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1927 trie->wordinfo[dupe].prev = curword; \
1929 /* we haven't inserted this word yet. */ \
1930 trie->states[ state ].wordnum = curword; \
1935 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1936 ( ( base + charid >= ucharcount \
1937 && base + charid < ubound \
1938 && state == trie->trans[ base - ucharcount + charid ].check \
1939 && trie->trans[ base - ucharcount + charid ].next ) \
1940 ? trie->trans[ base - ucharcount + charid ].next \
1941 : ( state==1 ? special : 0 ) \
1945 #define MADE_JUMP_TRIE 2
1946 #define MADE_EXACT_TRIE 4
1949 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
1950 regnode *first, regnode *last, regnode *tail,
1951 U32 word_count, U32 flags, U32 depth)
1953 /* first pass, loop through and scan words */
1954 reg_trie_data *trie;
1955 HV *widecharmap = NULL;
1956 AV *revcharmap = newAV();
1962 regnode *jumper = NULL;
1963 regnode *nextbranch = NULL;
1964 regnode *convert = NULL;
1965 U32 *prev_states; /* temp array mapping each state to previous one */
1966 /* we just use folder as a flag in utf8 */
1967 const U8 * folder = NULL;
1970 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
1971 AV *trie_words = NULL;
1972 /* along with revcharmap, this only used during construction but both are
1973 * useful during debugging so we store them in the struct when debugging.
1976 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
1977 STRLEN trie_charcount=0;
1979 SV *re_trie_maxbuff;
1980 GET_RE_DEBUG_FLAGS_DECL;
1982 PERL_ARGS_ASSERT_MAKE_TRIE;
1984 PERL_UNUSED_ARG(depth);
1991 case EXACTFU: folder = PL_fold_latin1; break;
1992 case EXACTF: folder = PL_fold; break;
1993 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1996 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1998 trie->startstate = 1;
1999 trie->wordcount = word_count;
2000 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2001 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2003 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2004 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2005 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2008 trie_words = newAV();
2011 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2012 assert(re_trie_maxbuff);
2013 if (!SvIOK(re_trie_maxbuff)) {
2014 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2016 DEBUG_TRIE_COMPILE_r({
2017 PerlIO_printf( Perl_debug_log,
2018 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2019 (int)depth * 2 + 2, "",
2020 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2021 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2024 /* Find the node we are going to overwrite */
2025 if ( first == startbranch && OP( last ) != BRANCH ) {
2026 /* whole branch chain */
2029 /* branch sub-chain */
2030 convert = NEXTOPER( first );
2033 /* -- First loop and Setup --
2035 We first traverse the branches and scan each word to determine if it
2036 contains widechars, and how many unique chars there are, this is
2037 important as we have to build a table with at least as many columns as we
2040 We use an array of integers to represent the character codes 0..255
2041 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2042 the native representation of the character value as the key and IV's for
2045 *TODO* If we keep track of how many times each character is used we can
2046 remap the columns so that the table compression later on is more
2047 efficient in terms of memory by ensuring the most common value is in the
2048 middle and the least common are on the outside. IMO this would be better
2049 than a most to least common mapping as theres a decent chance the most
2050 common letter will share a node with the least common, meaning the node
2051 will not be compressible. With a middle is most common approach the worst
2052 case is when we have the least common nodes twice.
2056 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2057 regnode *noper = NEXTOPER( cur );
2058 const U8 *uc = (U8*)STRING( noper );
2059 const U8 *e = uc + STR_LEN( noper );
2061 U32 wordlen = 0; /* required init */
2062 STRLEN minchars = 0;
2063 STRLEN maxchars = 0;
2064 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2067 if (OP(noper) == NOTHING) {
2068 regnode *noper_next= regnext(noper);
2069 if (noper_next != tail && OP(noper_next) == flags) {
2071 uc= (U8*)STRING(noper);
2072 e= uc + STR_LEN(noper);
2073 trie->minlen= STR_LEN(noper);
2080 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2081 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2082 regardless of encoding */
2083 if (OP( noper ) == EXACTFU_SS) {
2084 /* false positives are ok, so just set this */
2085 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2088 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2090 TRIE_CHARCOUNT(trie)++;
2093 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2094 * is in effect. Under /i, this character can match itself, or
2095 * anything that folds to it. If not under /i, it can match just
2096 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2097 * all fold to k, and all are single characters. But some folds
2098 * expand to more than one character, so for example LATIN SMALL
2099 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2100 * the string beginning at 'uc' is 'ffi', it could be matched by
2101 * three characters, or just by the one ligature character. (It
2102 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2103 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2104 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2105 * match.) The trie needs to know the minimum and maximum number
2106 * of characters that could match so that it can use size alone to
2107 * quickly reject many match attempts. The max is simple: it is
2108 * the number of folded characters in this branch (since a fold is
2109 * never shorter than what folds to it. */
2113 /* And the min is equal to the max if not under /i (indicated by
2114 * 'folder' being NULL), or there are no multi-character folds. If
2115 * there is a multi-character fold, the min is incremented just
2116 * once, for the character that folds to the sequence. Each
2117 * character in the sequence needs to be added to the list below of
2118 * characters in the trie, but we count only the first towards the
2119 * min number of characters needed. This is done through the
2120 * variable 'foldlen', which is returned by the macros that look
2121 * for these sequences as the number of bytes the sequence
2122 * occupies. Each time through the loop, we decrement 'foldlen' by
2123 * how many bytes the current char occupies. Only when it reaches
2124 * 0 do we increment 'minchars' or look for another multi-character
2126 if (folder == NULL) {
2129 else if (foldlen > 0) {
2130 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2135 /* See if *uc is the beginning of a multi-character fold. If
2136 * so, we decrement the length remaining to look at, to account
2137 * for the current character this iteration. (We can use 'uc'
2138 * instead of the fold returned by TRIE_READ_CHAR because for
2139 * non-UTF, the latin1_safe macro is smart enough to account
2140 * for all the unfolded characters, and because for UTF, the
2141 * string will already have been folded earlier in the
2142 * compilation process */
2144 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2145 foldlen -= UTF8SKIP(uc);
2148 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2153 /* The current character (and any potential folds) should be added
2154 * to the possible matching characters for this position in this
2158 U8 folded= folder[ (U8) uvc ];
2159 if ( !trie->charmap[ folded ] ) {
2160 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2161 TRIE_STORE_REVCHAR( folded );
2164 if ( !trie->charmap[ uvc ] ) {
2165 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2166 TRIE_STORE_REVCHAR( uvc );
2169 /* store the codepoint in the bitmap, and its folded
2171 TRIE_BITMAP_SET(trie, uvc);
2173 /* store the folded codepoint */
2174 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
2177 /* store first byte of utf8 representation of
2178 variant codepoints */
2179 if (! UVCHR_IS_INVARIANT(uvc)) {
2180 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
2183 set_bit = 0; /* We've done our bit :-) */
2187 /* XXX We could come up with the list of code points that fold
2188 * to this using PL_utf8_foldclosures, except not for
2189 * multi-char folds, as there may be multiple combinations
2190 * there that could work, which needs to wait until runtime to
2191 * resolve (The comment about LIGATURE FFI above is such an
2196 widecharmap = newHV();
2198 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2201 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
2203 if ( !SvTRUE( *svpp ) ) {
2204 sv_setiv( *svpp, ++trie->uniquecharcount );
2205 TRIE_STORE_REVCHAR(uvc);
2208 } /* end loop through characters in this branch of the trie */
2210 /* We take the min and max for this branch and combine to find the min
2211 * and max for all branches processed so far */
2212 if( cur == first ) {
2213 trie->minlen = minchars;
2214 trie->maxlen = maxchars;
2215 } else if (minchars < trie->minlen) {
2216 trie->minlen = minchars;
2217 } else if (maxchars > trie->maxlen) {
2218 trie->maxlen = maxchars;
2220 } /* end first pass */
2221 DEBUG_TRIE_COMPILE_r(
2222 PerlIO_printf( Perl_debug_log,
2223 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2224 (int)depth * 2 + 2,"",
2225 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2226 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2227 (int)trie->minlen, (int)trie->maxlen )
2231 We now know what we are dealing with in terms of unique chars and
2232 string sizes so we can calculate how much memory a naive
2233 representation using a flat table will take. If it's over a reasonable
2234 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2235 conservative but potentially much slower representation using an array
2238 At the end we convert both representations into the same compressed
2239 form that will be used in regexec.c for matching with. The latter
2240 is a form that cannot be used to construct with but has memory
2241 properties similar to the list form and access properties similar
2242 to the table form making it both suitable for fast searches and
2243 small enough that its feasable to store for the duration of a program.
2245 See the comment in the code where the compressed table is produced
2246 inplace from the flat tabe representation for an explanation of how
2247 the compression works.
2252 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2255 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2256 > SvIV(re_trie_maxbuff) )
2259 Second Pass -- Array Of Lists Representation
2261 Each state will be represented by a list of charid:state records
2262 (reg_trie_trans_le) the first such element holds the CUR and LEN
2263 points of the allocated array. (See defines above).
2265 We build the initial structure using the lists, and then convert
2266 it into the compressed table form which allows faster lookups
2267 (but cant be modified once converted).
2270 STRLEN transcount = 1;
2272 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2273 "%*sCompiling trie using list compiler\n",
2274 (int)depth * 2 + 2, ""));
2276 trie->states = (reg_trie_state *)
2277 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2278 sizeof(reg_trie_state) );
2282 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2284 regnode *noper = NEXTOPER( cur );
2285 U8 *uc = (U8*)STRING( noper );
2286 const U8 *e = uc + STR_LEN( noper );
2287 U32 state = 1; /* required init */
2288 U16 charid = 0; /* sanity init */
2289 U32 wordlen = 0; /* required init */
2291 if (OP(noper) == NOTHING) {
2292 regnode *noper_next= regnext(noper);
2293 if (noper_next != tail && OP(noper_next) == flags) {
2295 uc= (U8*)STRING(noper);
2296 e= uc + STR_LEN(noper);
2300 if (OP(noper) != NOTHING) {
2301 for ( ; uc < e ; uc += len ) {
2306 charid = trie->charmap[ uvc ];
2308 SV** const svpp = hv_fetch( widecharmap,
2315 charid=(U16)SvIV( *svpp );
2318 /* charid is now 0 if we dont know the char read, or
2319 * nonzero if we do */
2326 if ( !trie->states[ state ].trans.list ) {
2327 TRIE_LIST_NEW( state );
2330 check <= TRIE_LIST_USED( state );
2333 if ( TRIE_LIST_ITEM( state, check ).forid
2336 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2341 newstate = next_alloc++;
2342 prev_states[newstate] = state;
2343 TRIE_LIST_PUSH( state, charid, newstate );
2348 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2352 TRIE_HANDLE_WORD(state);
2354 } /* end second pass */
2356 /* next alloc is the NEXT state to be allocated */
2357 trie->statecount = next_alloc;
2358 trie->states = (reg_trie_state *)
2359 PerlMemShared_realloc( trie->states,
2361 * sizeof(reg_trie_state) );
2363 /* and now dump it out before we compress it */
2364 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2365 revcharmap, next_alloc,
2369 trie->trans = (reg_trie_trans *)
2370 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2377 for( state=1 ; state < next_alloc ; state ++ ) {
2381 DEBUG_TRIE_COMPILE_MORE_r(
2382 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2386 if (trie->states[state].trans.list) {
2387 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2391 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2392 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2393 if ( forid < minid ) {
2395 } else if ( forid > maxid ) {
2399 if ( transcount < tp + maxid - minid + 1) {
2401 trie->trans = (reg_trie_trans *)
2402 PerlMemShared_realloc( trie->trans,
2404 * sizeof(reg_trie_trans) );
2405 Zero( trie->trans + (transcount / 2),
2409 base = trie->uniquecharcount + tp - minid;
2410 if ( maxid == minid ) {
2412 for ( ; zp < tp ; zp++ ) {
2413 if ( ! trie->trans[ zp ].next ) {
2414 base = trie->uniquecharcount + zp - minid;
2415 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2417 trie->trans[ zp ].check = state;
2423 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2425 trie->trans[ tp ].check = state;
2430 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2431 const U32 tid = base
2432 - trie->uniquecharcount
2433 + TRIE_LIST_ITEM( state, idx ).forid;
2434 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2436 trie->trans[ tid ].check = state;
2438 tp += ( maxid - minid + 1 );
2440 Safefree(trie->states[ state ].trans.list);
2443 DEBUG_TRIE_COMPILE_MORE_r(
2444 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2447 trie->states[ state ].trans.base=base;
2449 trie->lasttrans = tp + 1;
2453 Second Pass -- Flat Table Representation.
2455 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2456 each. We know that we will need Charcount+1 trans at most to store
2457 the data (one row per char at worst case) So we preallocate both
2458 structures assuming worst case.
2460 We then construct the trie using only the .next slots of the entry
2463 We use the .check field of the first entry of the node temporarily
2464 to make compression both faster and easier by keeping track of how
2465 many non zero fields are in the node.
2467 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2470 There are two terms at use here: state as a TRIE_NODEIDX() which is
2471 a number representing the first entry of the node, and state as a
2472 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2473 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2474 if there are 2 entrys per node. eg:
2482 The table is internally in the right hand, idx form. However as we
2483 also have to deal with the states array which is indexed by nodenum
2484 we have to use TRIE_NODENUM() to convert.
2487 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2488 "%*sCompiling trie using table compiler\n",
2489 (int)depth * 2 + 2, ""));
2491 trie->trans = (reg_trie_trans *)
2492 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2493 * trie->uniquecharcount + 1,
2494 sizeof(reg_trie_trans) );
2495 trie->states = (reg_trie_state *)
2496 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2497 sizeof(reg_trie_state) );
2498 next_alloc = trie->uniquecharcount + 1;
2501 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2503 regnode *noper = NEXTOPER( cur );
2504 const U8 *uc = (U8*)STRING( noper );
2505 const U8 *e = uc + STR_LEN( noper );
2507 U32 state = 1; /* required init */
2509 U16 charid = 0; /* sanity init */
2510 U32 accept_state = 0; /* sanity init */
2512 U32 wordlen = 0; /* required init */
2514 if (OP(noper) == NOTHING) {
2515 regnode *noper_next= regnext(noper);
2516 if (noper_next != tail && OP(noper_next) == flags) {
2518 uc= (U8*)STRING(noper);
2519 e= uc + STR_LEN(noper);
2523 if ( OP(noper) != NOTHING ) {
2524 for ( ; uc < e ; uc += len ) {
2529 charid = trie->charmap[ uvc ];
2531 SV* const * const svpp = hv_fetch( widecharmap,
2535 charid = svpp ? (U16)SvIV(*svpp) : 0;
2539 if ( !trie->trans[ state + charid ].next ) {
2540 trie->trans[ state + charid ].next = next_alloc;
2541 trie->trans[ state ].check++;
2542 prev_states[TRIE_NODENUM(next_alloc)]
2543 = TRIE_NODENUM(state);
2544 next_alloc += trie->uniquecharcount;
2546 state = trie->trans[ state + charid ].next;
2548 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2550 /* charid is now 0 if we dont know the char read, or
2551 * nonzero if we do */
2554 accept_state = TRIE_NODENUM( state );
2555 TRIE_HANDLE_WORD(accept_state);
2557 } /* end second pass */
2559 /* and now dump it out before we compress it */
2560 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2562 next_alloc, depth+1));
2566 * Inplace compress the table.*
2568 For sparse data sets the table constructed by the trie algorithm will
2569 be mostly 0/FAIL transitions or to put it another way mostly empty.
2570 (Note that leaf nodes will not contain any transitions.)
2572 This algorithm compresses the tables by eliminating most such
2573 transitions, at the cost of a modest bit of extra work during lookup:
2575 - Each states[] entry contains a .base field which indicates the
2576 index in the state[] array wheres its transition data is stored.
2578 - If .base is 0 there are no valid transitions from that node.
2580 - If .base is nonzero then charid is added to it to find an entry in
2583 -If trans[states[state].base+charid].check!=state then the
2584 transition is taken to be a 0/Fail transition. Thus if there are fail
2585 transitions at the front of the node then the .base offset will point
2586 somewhere inside the previous nodes data (or maybe even into a node
2587 even earlier), but the .check field determines if the transition is
2591 The following process inplace converts the table to the compressed
2592 table: We first do not compress the root node 1,and mark all its
2593 .check pointers as 1 and set its .base pointer as 1 as well. This
2594 allows us to do a DFA construction from the compressed table later,
2595 and ensures that any .base pointers we calculate later are greater
2598 - We set 'pos' to indicate the first entry of the second node.
2600 - We then iterate over the columns of the node, finding the first and
2601 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2602 and set the .check pointers accordingly, and advance pos
2603 appropriately and repreat for the next node. Note that when we copy
2604 the next pointers we have to convert them from the original
2605 NODEIDX form to NODENUM form as the former is not valid post
2608 - If a node has no transitions used we mark its base as 0 and do not
2609 advance the pos pointer.
2611 - If a node only has one transition we use a second pointer into the
2612 structure to fill in allocated fail transitions from other states.
2613 This pointer is independent of the main pointer and scans forward
2614 looking for null transitions that are allocated to a state. When it
2615 finds one it writes the single transition into the "hole". If the
2616 pointer doesnt find one the single transition is appended as normal.
2618 - Once compressed we can Renew/realloc the structures to release the
2621 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2622 specifically Fig 3.47 and the associated pseudocode.
2626 const U32 laststate = TRIE_NODENUM( next_alloc );
2629 trie->statecount = laststate;
2631 for ( state = 1 ; state < laststate ; state++ ) {
2633 const U32 stateidx = TRIE_NODEIDX( state );
2634 const U32 o_used = trie->trans[ stateidx ].check;
2635 U32 used = trie->trans[ stateidx ].check;
2636 trie->trans[ stateidx ].check = 0;
2639 used && charid < trie->uniquecharcount;
2642 if ( flag || trie->trans[ stateidx + charid ].next ) {
2643 if ( trie->trans[ stateidx + charid ].next ) {
2645 for ( ; zp < pos ; zp++ ) {
2646 if ( ! trie->trans[ zp ].next ) {
2650 trie->states[ state ].trans.base
2652 + trie->uniquecharcount
2654 trie->trans[ zp ].next
2655 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2657 trie->trans[ zp ].check = state;
2658 if ( ++zp > pos ) pos = zp;
2665 trie->states[ state ].trans.base
2666 = pos + trie->uniquecharcount - charid ;
2668 trie->trans[ pos ].next
2669 = SAFE_TRIE_NODENUM(
2670 trie->trans[ stateidx + charid ].next );
2671 trie->trans[ pos ].check = state;
2676 trie->lasttrans = pos + 1;
2677 trie->states = (reg_trie_state *)
2678 PerlMemShared_realloc( trie->states, laststate
2679 * sizeof(reg_trie_state) );
2680 DEBUG_TRIE_COMPILE_MORE_r(
2681 PerlIO_printf( Perl_debug_log,
2682 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2683 (int)depth * 2 + 2,"",
2684 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2688 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2691 } /* end table compress */
2693 DEBUG_TRIE_COMPILE_MORE_r(
2694 PerlIO_printf(Perl_debug_log,
2695 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2696 (int)depth * 2 + 2, "",
2697 (UV)trie->statecount,
2698 (UV)trie->lasttrans)
2700 /* resize the trans array to remove unused space */
2701 trie->trans = (reg_trie_trans *)
2702 PerlMemShared_realloc( trie->trans, trie->lasttrans
2703 * sizeof(reg_trie_trans) );
2705 { /* Modify the program and insert the new TRIE node */
2706 U8 nodetype =(U8)(flags & 0xFF);
2710 regnode *optimize = NULL;
2711 #ifdef RE_TRACK_PATTERN_OFFSETS
2714 U32 mjd_nodelen = 0;
2715 #endif /* RE_TRACK_PATTERN_OFFSETS */
2716 #endif /* DEBUGGING */
2718 This means we convert either the first branch or the first Exact,
2719 depending on whether the thing following (in 'last') is a branch
2720 or not and whther first is the startbranch (ie is it a sub part of
2721 the alternation or is it the whole thing.)
2722 Assuming its a sub part we convert the EXACT otherwise we convert
2723 the whole branch sequence, including the first.
2725 /* Find the node we are going to overwrite */
2726 if ( first != startbranch || OP( last ) == BRANCH ) {
2727 /* branch sub-chain */
2728 NEXT_OFF( first ) = (U16)(last - first);
2729 #ifdef RE_TRACK_PATTERN_OFFSETS
2731 mjd_offset= Node_Offset((convert));
2732 mjd_nodelen= Node_Length((convert));
2735 /* whole branch chain */
2737 #ifdef RE_TRACK_PATTERN_OFFSETS
2740 const regnode *nop = NEXTOPER( convert );
2741 mjd_offset= Node_Offset((nop));
2742 mjd_nodelen= Node_Length((nop));
2746 PerlIO_printf(Perl_debug_log,
2747 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2748 (int)depth * 2 + 2, "",
2749 (UV)mjd_offset, (UV)mjd_nodelen)
2752 /* But first we check to see if there is a common prefix we can
2753 split out as an EXACT and put in front of the TRIE node. */
2754 trie->startstate= 1;
2755 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2757 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2761 const U32 base = trie->states[ state ].trans.base;
2763 if ( trie->states[state].wordnum )
2766 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2767 if ( ( base + ofs >= trie->uniquecharcount ) &&
2768 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2769 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2771 if ( ++count > 1 ) {
2772 SV **tmp = av_fetch( revcharmap, ofs, 0);
2773 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2774 if ( state == 1 ) break;
2776 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2778 PerlIO_printf(Perl_debug_log,
2779 "%*sNew Start State=%"UVuf" Class: [",
2780 (int)depth * 2 + 2, "",
2783 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2784 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2786 TRIE_BITMAP_SET(trie,*ch);
2788 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2790 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2794 TRIE_BITMAP_SET(trie,*ch);
2796 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2797 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2803 SV **tmp = av_fetch( revcharmap, idx, 0);
2805 char *ch = SvPV( *tmp, len );
2807 SV *sv=sv_newmortal();
2808 PerlIO_printf( Perl_debug_log,
2809 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2810 (int)depth * 2 + 2, "",
2812 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2813 PL_colors[0], PL_colors[1],
2814 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2815 PERL_PV_ESCAPE_FIRSTCHAR
2820 OP( convert ) = nodetype;
2821 str=STRING(convert);
2824 STR_LEN(convert) += len;
2830 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2835 trie->prefixlen = (state-1);
2837 regnode *n = convert+NODE_SZ_STR(convert);
2838 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2839 trie->startstate = state;
2840 trie->minlen -= (state - 1);
2841 trie->maxlen -= (state - 1);
2843 /* At least the UNICOS C compiler choked on this
2844 * being argument to DEBUG_r(), so let's just have
2847 #ifdef PERL_EXT_RE_BUILD
2853 regnode *fix = convert;
2854 U32 word = trie->wordcount;
2856 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2857 while( ++fix < n ) {
2858 Set_Node_Offset_Length(fix, 0, 0);
2861 SV ** const tmp = av_fetch( trie_words, word, 0 );
2863 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2864 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2866 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2874 NEXT_OFF(convert) = (U16)(tail - convert);
2875 DEBUG_r(optimize= n);
2881 if ( trie->maxlen ) {
2882 NEXT_OFF( convert ) = (U16)(tail - convert);
2883 ARG_SET( convert, data_slot );
2884 /* Store the offset to the first unabsorbed branch in
2885 jump[0], which is otherwise unused by the jump logic.
2886 We use this when dumping a trie and during optimisation. */
2888 trie->jump[0] = (U16)(nextbranch - convert);
2890 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2891 * and there is a bitmap
2892 * and the first "jump target" node we found leaves enough room
2893 * then convert the TRIE node into a TRIEC node, with the bitmap
2894 * embedded inline in the opcode - this is hypothetically faster.
2896 if ( !trie->states[trie->startstate].wordnum
2898 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2900 OP( convert ) = TRIEC;
2901 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2902 PerlMemShared_free(trie->bitmap);
2905 OP( convert ) = TRIE;
2907 /* store the type in the flags */
2908 convert->flags = nodetype;
2912 + regarglen[ OP( convert ) ];
2914 /* XXX We really should free up the resource in trie now,
2915 as we won't use them - (which resources?) dmq */
2917 /* needed for dumping*/
2918 DEBUG_r(if (optimize) {
2919 regnode *opt = convert;
2921 while ( ++opt < optimize) {
2922 Set_Node_Offset_Length(opt,0,0);
2925 Try to clean up some of the debris left after the
2928 while( optimize < jumper ) {
2929 mjd_nodelen += Node_Length((optimize));
2930 OP( optimize ) = OPTIMIZED;
2931 Set_Node_Offset_Length(optimize,0,0);
2934 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2936 } /* end node insert */
2938 /* Finish populating the prev field of the wordinfo array. Walk back
2939 * from each accept state until we find another accept state, and if
2940 * so, point the first word's .prev field at the second word. If the
2941 * second already has a .prev field set, stop now. This will be the
2942 * case either if we've already processed that word's accept state,
2943 * or that state had multiple words, and the overspill words were
2944 * already linked up earlier.
2951 for (word=1; word <= trie->wordcount; word++) {
2953 if (trie->wordinfo[word].prev)
2955 state = trie->wordinfo[word].accept;
2957 state = prev_states[state];
2960 prev = trie->states[state].wordnum;
2964 trie->wordinfo[word].prev = prev;
2966 Safefree(prev_states);
2970 /* and now dump out the compressed format */
2971 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2973 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2975 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2976 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2978 SvREFCNT_dec_NN(revcharmap);
2982 : trie->startstate>1
2988 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
2990 /* The Trie is constructed and compressed now so we can build a fail array if
2993 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
2995 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
2999 We find the fail state for each state in the trie, this state is the longest
3000 proper suffix of the current state's 'word' that is also a proper prefix of
3001 another word in our trie. State 1 represents the word '' and is thus the
3002 default fail state. This allows the DFA not to have to restart after its
3003 tried and failed a word at a given point, it simply continues as though it
3004 had been matching the other word in the first place.
3006 'abcdgu'=~/abcdefg|cdgu/
3007 When we get to 'd' we are still matching the first word, we would encounter
3008 'g' which would fail, which would bring us to the state representing 'd' in
3009 the second word where we would try 'g' and succeed, proceeding to match
3012 /* add a fail transition */
3013 const U32 trie_offset = ARG(source);
3014 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3016 const U32 ucharcount = trie->uniquecharcount;
3017 const U32 numstates = trie->statecount;
3018 const U32 ubound = trie->lasttrans + ucharcount;
3022 U32 base = trie->states[ 1 ].trans.base;
3025 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3027 GET_RE_DEBUG_FLAGS_DECL;
3029 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3030 PERL_UNUSED_CONTEXT;
3032 PERL_UNUSED_ARG(depth);
3035 if ( OP(source) == TRIE ) {
3036 struct regnode_1 *op = (struct regnode_1 *)
3037 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3038 StructCopy(source,op,struct regnode_1);
3039 stclass = (regnode *)op;
3041 struct regnode_charclass *op = (struct regnode_charclass *)
3042 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3043 StructCopy(source,op,struct regnode_charclass);
3044 stclass = (regnode *)op;
3046 OP(stclass)+=2; /* covert the TRIE type to its AHO-CORASICK equivalent */
3048 ARG_SET( stclass, data_slot );
3049 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3050 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3051 aho->trie=trie_offset;
3052 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3053 Copy( trie->states, aho->states, numstates, reg_trie_state );
3054 Newxz( q, numstates, U32);
3055 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3058 /* initialize fail[0..1] to be 1 so that we always have
3059 a valid final fail state */
3060 fail[ 0 ] = fail[ 1 ] = 1;
3062 for ( charid = 0; charid < ucharcount ; charid++ ) {
3063 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3065 q[ q_write ] = newstate;
3066 /* set to point at the root */
3067 fail[ q[ q_write++ ] ]=1;
3070 while ( q_read < q_write) {
3071 const U32 cur = q[ q_read++ % numstates ];
3072 base = trie->states[ cur ].trans.base;
3074 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3075 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3077 U32 fail_state = cur;
3080 fail_state = fail[ fail_state ];
3081 fail_base = aho->states[ fail_state ].trans.base;
3082 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3084 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3085 fail[ ch_state ] = fail_state;
3086 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3088 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3090 q[ q_write++ % numstates] = ch_state;
3094 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3095 when we fail in state 1, this allows us to use the
3096 charclass scan to find a valid start char. This is based on the principle
3097 that theres a good chance the string being searched contains lots of stuff
3098 that cant be a start char.
3100 fail[ 0 ] = fail[ 1 ] = 0;
3101 DEBUG_TRIE_COMPILE_r({
3102 PerlIO_printf(Perl_debug_log,
3103 "%*sStclass Failtable (%"UVuf" states): 0",
3104 (int)(depth * 2), "", (UV)numstates
3106 for( q_read=1; q_read<numstates; q_read++ ) {
3107 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3109 PerlIO_printf(Perl_debug_log, "\n");
3112 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3117 #define DEBUG_PEEP(str,scan,depth) \
3118 DEBUG_OPTIMISE_r({if (scan){ \
3119 SV * const mysv=sv_newmortal(); \
3120 regnode *Next = regnext(scan); \
3121 regprop(RExC_rx, mysv, scan, NULL); \
3122 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
3123 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
3124 Next ? (REG_NODE_NUM(Next)) : 0 ); \
3128 /* The below joins as many adjacent EXACTish nodes as possible into a single
3129 * one. The regop may be changed if the node(s) contain certain sequences that
3130 * require special handling. The joining is only done if:
3131 * 1) there is room in the current conglomerated node to entirely contain the
3133 * 2) they are the exact same node type
3135 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3136 * these get optimized out
3138 * If a node is to match under /i (folded), the number of characters it matches
3139 * can be different than its character length if it contains a multi-character
3140 * fold. *min_subtract is set to the total delta number of characters of the
3143 * And *unfolded_multi_char is set to indicate whether or not the node contains
3144 * an unfolded multi-char fold. This happens when whether the fold is valid or
3145 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3146 * SMALL LETTER SHARP S, as only if the target string being matched against
3147 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3148 * folding rules depend on the locale in force at runtime. (Multi-char folds
3149 * whose components are all above the Latin1 range are not run-time locale
3150 * dependent, and have already been folded by the time this function is
3153 * This is as good a place as any to discuss the design of handling these
3154 * multi-character fold sequences. It's been wrong in Perl for a very long
3155 * time. There are three code points in Unicode whose multi-character folds
3156 * were long ago discovered to mess things up. The previous designs for
3157 * dealing with these involved assigning a special node for them. This
3158 * approach doesn't always work, as evidenced by this example:
3159 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3160 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3161 * would match just the \xDF, it won't be able to handle the case where a
3162 * successful match would have to cross the node's boundary. The new approach
3163 * that hopefully generally solves the problem generates an EXACTFU_SS node
3164 * that is "sss" in this case.
3166 * It turns out that there are problems with all multi-character folds, and not
3167 * just these three. Now the code is general, for all such cases. The
3168 * approach taken is:
3169 * 1) This routine examines each EXACTFish node that could contain multi-
3170 * character folded sequences. Since a single character can fold into
3171 * such a sequence, the minimum match length for this node is less than
3172 * the number of characters in the node. This routine returns in
3173 * *min_subtract how many characters to subtract from the the actual
3174 * length of the string to get a real minimum match length; it is 0 if
3175 * there are no multi-char foldeds. This delta is used by the caller to
3176 * adjust the min length of the match, and the delta between min and max,
3177 * so that the optimizer doesn't reject these possibilities based on size
3179 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3180 * is used for an EXACTFU node that contains at least one "ss" sequence in
3181 * it. For non-UTF-8 patterns and strings, this is the only case where
3182 * there is a possible fold length change. That means that a regular
3183 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3184 * with length changes, and so can be processed faster. regexec.c takes
3185 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3186 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3187 * known until runtime). This saves effort in regex matching. However,
3188 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3189 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3190 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3191 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3192 * possibilities for the non-UTF8 patterns are quite simple, except for
3193 * the sharp s. All the ones that don't involve a UTF-8 target string are
3194 * members of a fold-pair, and arrays are set up for all of them so that
3195 * the other member of the pair can be found quickly. Code elsewhere in
3196 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3197 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3198 * described in the next item.
3199 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3200 * validity of the fold won't be known until runtime, and so must remain
3201 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3202 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3203 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3204 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3205 * The reason this is a problem is that the optimizer part of regexec.c
3206 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3207 * that a character in the pattern corresponds to at most a single
3208 * character in the target string. (And I do mean character, and not byte
3209 * here, unlike other parts of the documentation that have never been
3210 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3211 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3212 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3213 * nodes, violate the assumption, and they are the only instances where it
3214 * is violated. I'm reluctant to try to change the assumption, as the
3215 * code involved is impenetrable to me (khw), so instead the code here
3216 * punts. This routine examines EXACTFL nodes, and (when the pattern
3217 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3218 * boolean indicating whether or not the node contains such a fold. When
3219 * it is true, the caller sets a flag that later causes the optimizer in
3220 * this file to not set values for the floating and fixed string lengths,
3221 * and thus avoids the optimizer code in regexec.c that makes the invalid
3222 * assumption. Thus, there is no optimization based on string lengths for
3223 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3224 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3225 * assumption is wrong only in these cases is that all other non-UTF-8
3226 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3227 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3228 * EXACTF nodes because we don't know at compile time if it actually
3229 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3230 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3231 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3232 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3233 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3234 * string would require the pattern to be forced into UTF-8, the overhead
3235 * of which we want to avoid. Similarly the unfolded multi-char folds in
3236 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3239 * Similarly, the code that generates tries doesn't currently handle
3240 * not-already-folded multi-char folds, and it looks like a pain to change
3241 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3242 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3243 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3244 * using /iaa matching will be doing so almost entirely with ASCII
3245 * strings, so this should rarely be encountered in practice */
3247 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3248 if (PL_regkind[OP(scan)] == EXACT) \
3249 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3252 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3253 UV *min_subtract, bool *unfolded_multi_char,
3254 U32 flags,regnode *val, U32 depth)
3256 /* Merge several consecutive EXACTish nodes into one. */
3257 regnode *n = regnext(scan);
3259 regnode *next = scan + NODE_SZ_STR(scan);
3263 regnode *stop = scan;
3264 GET_RE_DEBUG_FLAGS_DECL;
3266 PERL_UNUSED_ARG(depth);
3269 PERL_ARGS_ASSERT_JOIN_EXACT;
3270 #ifndef EXPERIMENTAL_INPLACESCAN
3271 PERL_UNUSED_ARG(flags);
3272 PERL_UNUSED_ARG(val);
3274 DEBUG_PEEP("join",scan,depth);
3276 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3277 * EXACT ones that are mergeable to the current one. */
3279 && (PL_regkind[OP(n)] == NOTHING
3280 || (stringok && OP(n) == OP(scan)))
3282 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3285 if (OP(n) == TAIL || n > next)
3287 if (PL_regkind[OP(n)] == NOTHING) {
3288 DEBUG_PEEP("skip:",n,depth);
3289 NEXT_OFF(scan) += NEXT_OFF(n);
3290 next = n + NODE_STEP_REGNODE;
3297 else if (stringok) {
3298 const unsigned int oldl = STR_LEN(scan);
3299 regnode * const nnext = regnext(n);
3301 /* XXX I (khw) kind of doubt that this works on platforms (should
3302 * Perl ever run on one) where U8_MAX is above 255 because of lots
3303 * of other assumptions */
3304 /* Don't join if the sum can't fit into a single node */
3305 if (oldl + STR_LEN(n) > U8_MAX)
3308 DEBUG_PEEP("merg",n,depth);
3311 NEXT_OFF(scan) += NEXT_OFF(n);
3312 STR_LEN(scan) += STR_LEN(n);
3313 next = n + NODE_SZ_STR(n);
3314 /* Now we can overwrite *n : */
3315 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3323 #ifdef EXPERIMENTAL_INPLACESCAN
3324 if (flags && !NEXT_OFF(n)) {
3325 DEBUG_PEEP("atch", val, depth);
3326 if (reg_off_by_arg[OP(n)]) {
3327 ARG_SET(n, val - n);
3330 NEXT_OFF(n) = val - n;
3338 *unfolded_multi_char = FALSE;
3340 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3341 * can now analyze for sequences of problematic code points. (Prior to
3342 * this final joining, sequences could have been split over boundaries, and
3343 * hence missed). The sequences only happen in folding, hence for any
3344 * non-EXACT EXACTish node */
3345 if (OP(scan) != EXACT) {
3346 U8* s0 = (U8*) STRING(scan);
3348 U8* s_end = s0 + STR_LEN(scan);
3350 int total_count_delta = 0; /* Total delta number of characters that
3351 multi-char folds expand to */
3353 /* One pass is made over the node's string looking for all the
3354 * possibilities. To avoid some tests in the loop, there are two main
3355 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3360 if (OP(scan) == EXACTFL) {
3363 /* An EXACTFL node would already have been changed to another
3364 * node type unless there is at least one character in it that
3365 * is problematic; likely a character whose fold definition
3366 * won't be known until runtime, and so has yet to be folded.
3367 * For all but the UTF-8 locale, folds are 1-1 in length, but
3368 * to handle the UTF-8 case, we need to create a temporary
3369 * folded copy using UTF-8 locale rules in order to analyze it.
3370 * This is because our macros that look to see if a sequence is
3371 * a multi-char fold assume everything is folded (otherwise the
3372 * tests in those macros would be too complicated and slow).
3373 * Note that here, the non-problematic folds will have already
3374 * been done, so we can just copy such characters. We actually
3375 * don't completely fold the EXACTFL string. We skip the
3376 * unfolded multi-char folds, as that would just create work
3377 * below to figure out the size they already are */
3379 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3382 STRLEN s_len = UTF8SKIP(s);
3383 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3384 Copy(s, d, s_len, U8);
3387 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3388 *unfolded_multi_char = TRUE;
3389 Copy(s, d, s_len, U8);
3392 else if (isASCII(*s)) {
3393 *(d++) = toFOLD(*s);
3397 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3403 /* Point the remainder of the routine to look at our temporary
3407 } /* End of creating folded copy of EXACTFL string */
3409 /* Examine the string for a multi-character fold sequence. UTF-8
3410 * patterns have all characters pre-folded by the time this code is
3412 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3413 length sequence we are looking for is 2 */
3415 int count = 0; /* How many characters in a multi-char fold */
3416 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3417 if (! len) { /* Not a multi-char fold: get next char */
3422 /* Nodes with 'ss' require special handling, except for
3423 * EXACTFA-ish for which there is no multi-char fold to this */
3424 if (len == 2 && *s == 's' && *(s+1) == 's'
3425 && OP(scan) != EXACTFA
3426 && OP(scan) != EXACTFA_NO_TRIE)
3429 if (OP(scan) != EXACTFL) {
3430 OP(scan) = EXACTFU_SS;
3434 else { /* Here is a generic multi-char fold. */
3435 U8* multi_end = s + len;
3437 /* Count how many characters are in it. In the case of
3438 * /aa, no folds which contain ASCII code points are
3439 * allowed, so check for those, and skip if found. */
3440 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3441 count = utf8_length(s, multi_end);
3445 while (s < multi_end) {
3448 goto next_iteration;
3458 /* The delta is how long the sequence is minus 1 (1 is how long
3459 * the character that folds to the sequence is) */
3460 total_count_delta += count - 1;
3464 /* We created a temporary folded copy of the string in EXACTFL
3465 * nodes. Therefore we need to be sure it doesn't go below zero,
3466 * as the real string could be shorter */
3467 if (OP(scan) == EXACTFL) {
3468 int total_chars = utf8_length((U8*) STRING(scan),
3469 (U8*) STRING(scan) + STR_LEN(scan));
3470 if (total_count_delta > total_chars) {
3471 total_count_delta = total_chars;
3475 *min_subtract += total_count_delta;
3478 else if (OP(scan) == EXACTFA) {
3480 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3481 * fold to the ASCII range (and there are no existing ones in the
3482 * upper latin1 range). But, as outlined in the comments preceding
3483 * this function, we need to flag any occurrences of the sharp s.
3484 * This character forbids trie formation (because of added
3487 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
3488 OP(scan) = EXACTFA_NO_TRIE;
3489 *unfolded_multi_char = TRUE;
3498 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3499 * folds that are all Latin1. As explained in the comments
3500 * preceding this function, we look also for the sharp s in EXACTF
3501 * and EXACTFL nodes; it can be in the final position. Otherwise
3502 * we can stop looking 1 byte earlier because have to find at least
3503 * two characters for a multi-fold */
3504 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
3509 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
3510 if (! len) { /* Not a multi-char fold. */
3511 if (*s == LATIN_SMALL_LETTER_SHARP_S
3512 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
3514 *unfolded_multi_char = TRUE;
3521 && isARG2_lower_or_UPPER_ARG1('s', *s)
3522 && isARG2_lower_or_UPPER_ARG1('s', *(s+1)))
3525 /* EXACTF nodes need to know that the minimum length
3526 * changed so that a sharp s in the string can match this
3527 * ss in the pattern, but they remain EXACTF nodes, as they
3528 * won't match this unless the target string is is UTF-8,
3529 * which we don't know until runtime. EXACTFL nodes can't
3530 * transform into EXACTFU nodes */
3531 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
3532 OP(scan) = EXACTFU_SS;
3536 *min_subtract += len - 1;
3543 /* Allow dumping but overwriting the collection of skipped
3544 * ops and/or strings with fake optimized ops */
3545 n = scan + NODE_SZ_STR(scan);
3553 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3557 /* REx optimizer. Converts nodes into quicker variants "in place".
3558 Finds fixed substrings. */
3560 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3561 to the position after last scanned or to NULL. */
3563 #define INIT_AND_WITHP \
3564 assert(!and_withp); \
3565 Newx(and_withp,1, regnode_ssc); \
3566 SAVEFREEPV(and_withp)
3568 /* this is a chain of data about sub patterns we are processing that
3569 need to be handled separately/specially in study_chunk. Its so
3570 we can simulate recursion without losing state. */
3572 typedef struct scan_frame {
3573 regnode *last; /* last node to process in this frame */
3574 regnode *next; /* next node to process when last is reached */
3575 struct scan_frame *prev; /*previous frame*/
3576 U32 prev_recursed_depth;
3577 I32 stop; /* what stopparen do we use */
3582 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3583 SSize_t *minlenp, SSize_t *deltap,
3588 regnode_ssc *and_withp,
3589 U32 flags, U32 depth)
3590 /* scanp: Start here (read-write). */
3591 /* deltap: Write maxlen-minlen here. */
3592 /* last: Stop before this one. */
3593 /* data: string data about the pattern */
3594 /* stopparen: treat close N as END */
3595 /* recursed: which subroutines have we recursed into */
3596 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3598 /* There must be at least this number of characters to match */
3601 regnode *scan = *scanp, *next;
3603 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3604 int is_inf_internal = 0; /* The studied chunk is infinite */
3605 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3606 scan_data_t data_fake;
3607 SV *re_trie_maxbuff = NULL;
3608 regnode *first_non_open = scan;
3609 SSize_t stopmin = SSize_t_MAX;
3610 scan_frame *frame = NULL;
3611 GET_RE_DEBUG_FLAGS_DECL;
3613 PERL_ARGS_ASSERT_STUDY_CHUNK;
3616 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3619 while (first_non_open && OP(first_non_open) == OPEN)
3620 first_non_open=regnext(first_non_open);
3625 while ( scan && OP(scan) != END && scan < last ){
3626 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3627 node length to get a real minimum (because
3628 the folded version may be shorter) */
3629 bool unfolded_multi_char = FALSE;
3630 /* Peephole optimizer: */
3631 DEBUG_OPTIMISE_MORE_r(
3633 PerlIO_printf(Perl_debug_log,
3634 "%*sstudy_chunk stopparen=%ld depth=%lu recursed_depth=%lu ",
3635 ((int) depth*2), "", (long)stopparen,
3636 (unsigned long)depth, (unsigned long)recursed_depth);
3637 if (recursed_depth) {
3640 for ( j = 0 ; j < recursed_depth ; j++ ) {
3641 PerlIO_printf(Perl_debug_log,"[");
3642 for ( i = 0 ; i < (U32)RExC_npar ; i++ )
3643 PerlIO_printf(Perl_debug_log,"%d",
3644 PAREN_TEST(RExC_study_chunk_recursed +
3645 (j * RExC_study_chunk_recursed_bytes), i)
3648 PerlIO_printf(Perl_debug_log,"]");
3651 PerlIO_printf(Perl_debug_log,"\n");
3654 DEBUG_STUDYDATA("Peep:", data, depth);
3655 DEBUG_PEEP("Peep", scan, depth);
3658 /* The reason we do this here we need to deal with things like /(?:f)(?:o)(?:o)/
3659 * which cant be dealt with by the normal EXACT parsing code, as each (?:..) is handled
3660 * by a different invocation of reg() -- Yves
3662 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
3664 /* Follow the next-chain of the current node and optimize
3665 away all the NOTHINGs from it. */
3666 if (OP(scan) != CURLYX) {
3667 const int max = (reg_off_by_arg[OP(scan)]
3669 /* I32 may be smaller than U16 on CRAYs! */
3670 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3671 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3675 /* Skip NOTHING and LONGJMP. */
3676 while ((n = regnext(n))
3677 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3678 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3679 && off + noff < max)
3681 if (reg_off_by_arg[OP(scan)])
3684 NEXT_OFF(scan) = off;
3689 /* The principal pseudo-switch. Cannot be a switch, since we
3690 look into several different things. */
3691 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3692 || OP(scan) == IFTHEN) {
3693 next = regnext(scan);
3695 /* demq: the op(next)==code check is to see if we have
3696 * "branch-branch" AFAICT */
3698 if (OP(next) == code || code == IFTHEN) {
3699 /* NOTE - There is similar code to this block below for
3700 * handling TRIE nodes on a re-study. If you change stuff here
3701 * check there too. */
3702 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
3704 regnode * const startbranch=scan;
3706 if (flags & SCF_DO_SUBSTR) {
3707 /* Cannot merge strings after this. */
3708 scan_commit(pRExC_state, data, minlenp, is_inf);
3711 if (flags & SCF_DO_STCLASS)
3712 ssc_init_zero(pRExC_state, &accum);
3714 while (OP(scan) == code) {
3715 SSize_t deltanext, minnext, fake;
3717 regnode_ssc this_class;
3720 data_fake.flags = 0;
3722 data_fake.whilem_c = data->whilem_c;
3723 data_fake.last_closep = data->last_closep;
3726 data_fake.last_closep = &fake;
3728 data_fake.pos_delta = delta;
3729 next = regnext(scan);
3730 scan = NEXTOPER(scan);
3732 scan = NEXTOPER(scan);
3733 if (flags & SCF_DO_STCLASS) {
3734 ssc_init(pRExC_state, &this_class);
3735 data_fake.start_class = &this_class;
3736 f = SCF_DO_STCLASS_AND;
3738 if (flags & SCF_WHILEM_VISITED_POS)
3739 f |= SCF_WHILEM_VISITED_POS;
3741 /* we suppose the run is continuous, last=next...*/
3742 minnext = study_chunk(pRExC_state, &scan, minlenp,
3743 &deltanext, next, &data_fake, stopparen,
3744 recursed_depth, NULL, f,depth+1);
3747 if (deltanext == SSize_t_MAX) {
3748 is_inf = is_inf_internal = 1;
3750 } else if (max1 < minnext + deltanext)
3751 max1 = minnext + deltanext;
3753 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3755 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3756 if ( stopmin > minnext)
3757 stopmin = min + min1;
3758 flags &= ~SCF_DO_SUBSTR;
3760 data->flags |= SCF_SEEN_ACCEPT;
3763 if (data_fake.flags & SF_HAS_EVAL)
3764 data->flags |= SF_HAS_EVAL;
3765 data->whilem_c = data_fake.whilem_c;
3767 if (flags & SCF_DO_STCLASS)
3768 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
3770 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3772 if (flags & SCF_DO_SUBSTR) {
3773 data->pos_min += min1;
3774 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3775 data->pos_delta = SSize_t_MAX;
3777 data->pos_delta += max1 - min1;
3778 if (max1 != min1 || is_inf)
3779 data->longest = &(data->longest_float);
3782 if (delta == SSize_t_MAX
3783 || SSize_t_MAX - delta - (max1 - min1) < 0)
3784 delta = SSize_t_MAX;
3786 delta += max1 - min1;
3787 if (flags & SCF_DO_STCLASS_OR) {
3788 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
3790 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
3791 flags &= ~SCF_DO_STCLASS;
3794 else if (flags & SCF_DO_STCLASS_AND) {
3796 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
3797 flags &= ~SCF_DO_STCLASS;
3800 /* Switch to OR mode: cache the old value of
3801 * data->start_class */
3803 StructCopy(data->start_class, and_withp, regnode_ssc);
3804 flags &= ~SCF_DO_STCLASS_AND;
3805 StructCopy(&accum, data->start_class, regnode_ssc);
3806 flags |= SCF_DO_STCLASS_OR;
3810 if (PERL_ENABLE_TRIE_OPTIMISATION &&
3811 OP( startbranch ) == BRANCH )
3815 Assuming this was/is a branch we are dealing with: 'scan'
3816 now points at the item that follows the branch sequence,
3817 whatever it is. We now start at the beginning of the
3818 sequence and look for subsequences of
3824 which would be constructed from a pattern like
3827 If we can find such a subsequence we need to turn the first
3828 element into a trie and then add the subsequent branch exact
3829 strings to the trie.
3833 1. patterns where the whole set of branches can be
3836 2. patterns where only a subset can be converted.
3838 In case 1 we can replace the whole set with a single regop
3839 for the trie. In case 2 we need to keep the start and end
3842 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3843 becomes BRANCH TRIE; BRANCH X;
3845 There is an additional case, that being where there is a
3846 common prefix, which gets split out into an EXACT like node
3847 preceding the TRIE node.
3849 If x(1..n)==tail then we can do a simple trie, if not we make
3850 a "jump" trie, such that when we match the appropriate word
3851 we "jump" to the appropriate tail node. Essentially we turn
3852 a nested if into a case structure of sorts.
3857 if (!re_trie_maxbuff) {
3858 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3859 if (!SvIOK(re_trie_maxbuff))
3860 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3862 if ( SvIV(re_trie_maxbuff)>=0 ) {
3864 regnode *first = (regnode *)NULL;
3865 regnode *last = (regnode *)NULL;
3866 regnode *tail = scan;
3871 SV * const mysv = sv_newmortal(); /* for dumping */
3873 /* var tail is used because there may be a TAIL
3874 regop in the way. Ie, the exacts will point to the
3875 thing following the TAIL, but the last branch will
3876 point at the TAIL. So we advance tail. If we
3877 have nested (?:) we may have to move through several
3881 while ( OP( tail ) == TAIL ) {
3882 /* this is the TAIL generated by (?:) */
3883 tail = regnext( tail );
3887 DEBUG_TRIE_COMPILE_r({
3888 regprop(RExC_rx, mysv, tail, NULL);
3889 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3890 (int)depth * 2 + 2, "",
3891 "Looking for TRIE'able sequences. Tail node is: ",
3892 SvPV_nolen_const( mysv )
3898 Step through the branches
3899 cur represents each branch,
3900 noper is the first thing to be matched as part
3902 noper_next is the regnext() of that node.
3904 We normally handle a case like this
3905 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
3906 support building with NOJUMPTRIE, which restricts
3907 the trie logic to structures like /FOO|BAR/.
3909 If noper is a trieable nodetype then the branch is
3910 a possible optimization target. If we are building
3911 under NOJUMPTRIE then we require that noper_next is
3912 the same as scan (our current position in the regex
3915 Once we have two or more consecutive such branches
3916 we can create a trie of the EXACT's contents and
3917 stitch it in place into the program.
3919 If the sequence represents all of the branches in
3920 the alternation we replace the entire thing with a
3923 Otherwise when it is a subsequence we need to
3924 stitch it in place and replace only the relevant
3925 branches. This means the first branch has to remain
3926 as it is used by the alternation logic, and its
3927 next pointer, and needs to be repointed at the item
3928 on the branch chain following the last branch we
3929 have optimized away.
3931 This could be either a BRANCH, in which case the
3932 subsequence is internal, or it could be the item
3933 following the branch sequence in which case the
3934 subsequence is at the end (which does not
https://perl5.git.perl.org / perl5.git / blob