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 256 code points, and an
1052 * inversion list for the others, but if there are code points that should
1053 * match only conditionally on the target string being UTF-8, those are
1054 * placed in the inversion list, and not the bitmap. Since there are
1055 * circumstances under which they could match, they are included in the
1056 * SSC. But if the ANYOF node is to be inverted, we have to exclude them
1057 * here, so that when we invert below, the end result actually does include
1058 * them. (Think about "\xe0" =~ /[^\xc0]/di;). We have to do this here
1059 * before we add the unconditionally matched code points */
1060 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1061 _invlist_intersection_complement_2nd(invlist,
1066 /* Add in the points from the bit map */
1067 for (i = 0; i < 256; i++) {
1068 if (ANYOF_BITMAP_TEST(node, i)) {
1069 invlist = add_cp_to_invlist(invlist, i);
1070 new_node_has_latin1 = TRUE;
1074 /* If this can match all upper Latin1 code points, have to add them
1076 if (ANYOF_FLAGS(node) & ANYOF_NON_UTF8_NON_ASCII_ALL) {
1077 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1080 /* Similarly for these */
1081 if (ANYOF_FLAGS(node) & ANYOF_ABOVE_LATIN1_ALL) {
1082 invlist = _add_range_to_invlist(invlist, 256, UV_MAX);
1085 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1086 _invlist_invert(invlist);
1088 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOF_LOC_FOLD) {
1090 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1091 * locale. We can skip this if there are no 0-255 at all. */
1092 _invlist_union(invlist, PL_Latin1, &invlist);
1095 /* Similarly add the UTF-8 locale possible matches. These have to be
1096 * deferred until after the non-UTF-8 locale ones are taken care of just
1097 * above, or it leads to wrong results under ANYOF_INVERT */
1098 if (only_utf8_locale_invlist) {
1099 _invlist_union_maybe_complement_2nd(invlist,
1100 only_utf8_locale_invlist,
1101 ANYOF_FLAGS(node) & ANYOF_INVERT,
1108 /* These two functions currently do the exact same thing */
1109 #define ssc_init_zero ssc_init
1111 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1112 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1114 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1115 * should not be inverted. 'and_with->flags & ANYOF_POSIXL' should be 0 if
1116 * 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1119 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1120 const regnode_charclass *and_with)
1122 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1123 * another SSC or a regular ANYOF class. Can create false positives. */
1128 PERL_ARGS_ASSERT_SSC_AND;
1130 assert(is_ANYOF_SYNTHETIC(ssc));
1132 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1133 * the code point inversion list and just the relevant flags */
1134 if (is_ANYOF_SYNTHETIC(and_with)) {
1135 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1136 anded_flags = ANYOF_FLAGS(and_with);
1138 /* XXX This is a kludge around what appears to be deficiencies in the
1139 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1140 * there are paths through the optimizer where it doesn't get weeded
1141 * out when it should. And if we don't make some extra provision for
1142 * it like the code just below, it doesn't get added when it should.
1143 * This solution is to add it only when AND'ing, which is here, and
1144 * only when what is being AND'ed is the pristine, original node
1145 * matching anything. Thus it is like adding it to ssc_anything() but
1146 * only when the result is to be AND'ed. Probably the same solution
1147 * could be adopted for the same problem we have with /l matching,
1148 * which is solved differently in S_ssc_init(), and that would lead to
1149 * fewer false positives than that solution has. But if this solution
1150 * creates bugs, the consequences are only that a warning isn't raised
1151 * that should be; while the consequences for having /l bugs is
1152 * incorrect matches */
1153 if (ssc_is_anything((regnode_ssc *)and_with)) {
1154 anded_flags |= ANYOF_WARN_SUPER;
1158 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1159 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1162 ANYOF_FLAGS(ssc) &= anded_flags;
1164 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1165 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1166 * 'and_with' may be inverted. When not inverted, we have the situation of
1168 * (C1 | P1) & (C2 | P2)
1169 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1170 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1171 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1172 * <= ((C1 & C2) | P1 | P2)
1173 * Alternatively, the last few steps could be:
1174 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1175 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1176 * <= (C1 | C2 | (P1 & P2))
1177 * We favor the second approach if either P1 or P2 is non-empty. This is
1178 * because these components are a barrier to doing optimizations, as what
1179 * they match cannot be known until the moment of matching as they are
1180 * dependent on the current locale, 'AND"ing them likely will reduce or
1182 * But we can do better if we know that C1,P1 are in their initial state (a
1183 * frequent occurrence), each matching everything:
1184 * (<everything>) & (C2 | P2) = C2 | P2
1185 * Similarly, if C2,P2 are in their initial state (again a frequent
1186 * occurrence), the result is a no-op
1187 * (C1 | P1) & (<everything>) = C1 | P1
1190 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1191 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1192 * <= (C1 & ~C2) | (P1 & ~P2)
1195 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1196 && ! is_ANYOF_SYNTHETIC(and_with))
1200 ssc_intersection(ssc,
1202 FALSE /* Has already been inverted */
1205 /* If either P1 or P2 is empty, the intersection will be also; can skip
1207 if (! (ANYOF_FLAGS(and_with) & ANYOF_POSIXL)) {
1208 ANYOF_POSIXL_ZERO(ssc);
1210 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1212 /* Note that the Posix class component P from 'and_with' actually
1214 * P = Pa | Pb | ... | Pn
1215 * where each component is one posix class, such as in [\w\s].
1217 * ~P = ~(Pa | Pb | ... | Pn)
1218 * = ~Pa & ~Pb & ... & ~Pn
1219 * <= ~Pa | ~Pb | ... | ~Pn
1220 * The last is something we can easily calculate, but unfortunately
1221 * is likely to have many false positives. We could do better
1222 * in some (but certainly not all) instances if two classes in
1223 * P have known relationships. For example
1224 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1226 * :lower: & :print: = :lower:
1227 * And similarly for classes that must be disjoint. For example,
1228 * since \s and \w can have no elements in common based on rules in
1229 * the POSIX standard,
1230 * \w & ^\S = nothing
1231 * Unfortunately, some vendor locales do not meet the Posix
1232 * standard, in particular almost everything by Microsoft.
1233 * The loop below just changes e.g., \w into \W and vice versa */
1235 regnode_charclass_posixl temp;
1236 int add = 1; /* To calculate the index of the complement */
1238 ANYOF_POSIXL_ZERO(&temp);
1239 for (i = 0; i < ANYOF_MAX; i++) {
1241 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1242 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1244 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1245 ANYOF_POSIXL_SET(&temp, i + add);
1247 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1249 ANYOF_POSIXL_AND(&temp, ssc);
1251 } /* else ssc already has no posixes */
1252 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1253 in its initial state */
1254 else if (! is_ANYOF_SYNTHETIC(and_with)
1255 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1257 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1258 * copy it over 'ssc' */
1259 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1260 if (is_ANYOF_SYNTHETIC(and_with)) {
1261 StructCopy(and_with, ssc, regnode_ssc);
1264 ssc->invlist = anded_cp_list;
1265 ANYOF_POSIXL_ZERO(ssc);
1266 if (ANYOF_FLAGS(and_with) & ANYOF_POSIXL) {
1267 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1271 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1272 || (ANYOF_FLAGS(and_with) & ANYOF_POSIXL))
1274 /* One or the other of P1, P2 is non-empty. */
1275 if (ANYOF_FLAGS(and_with) & ANYOF_POSIXL) {
1276 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1278 ssc_union(ssc, anded_cp_list, FALSE);
1280 else { /* P1 = P2 = empty */
1281 ssc_intersection(ssc, anded_cp_list, FALSE);
1287 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1288 const regnode_charclass *or_with)
1290 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1291 * another SSC or a regular ANYOF class. Can create false positives if
1292 * 'or_with' is to be inverted. */
1297 PERL_ARGS_ASSERT_SSC_OR;
1299 assert(is_ANYOF_SYNTHETIC(ssc));
1301 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1302 * the code point inversion list and just the relevant flags */
1303 if (is_ANYOF_SYNTHETIC(or_with)) {
1304 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1305 ored_flags = ANYOF_FLAGS(or_with);
1308 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1309 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1312 ANYOF_FLAGS(ssc) |= ored_flags;
1314 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1315 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1316 * 'or_with' may be inverted. When not inverted, we have the simple
1317 * situation of computing:
1318 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1319 * If P1|P2 yields a situation with both a class and its complement are
1320 * set, like having both \w and \W, this matches all code points, and we
1321 * can delete these from the P component of the ssc going forward. XXX We
1322 * might be able to delete all the P components, but I (khw) am not certain
1323 * about this, and it is better to be safe.
1326 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1327 * <= (C1 | P1) | ~C2
1328 * <= (C1 | ~C2) | P1
1329 * (which results in actually simpler code than the non-inverted case)
1332 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1333 && ! is_ANYOF_SYNTHETIC(or_with))
1335 /* We ignore P2, leaving P1 going forward */
1336 } /* else Not inverted */
1337 else if (ANYOF_FLAGS(or_with) & ANYOF_POSIXL) {
1338 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1339 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1341 for (i = 0; i < ANYOF_MAX; i += 2) {
1342 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1344 ssc_match_all_cp(ssc);
1345 ANYOF_POSIXL_CLEAR(ssc, i);
1346 ANYOF_POSIXL_CLEAR(ssc, i+1);
1354 FALSE /* Already has been inverted */
1358 PERL_STATIC_INLINE void
1359 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1361 PERL_ARGS_ASSERT_SSC_UNION;
1363 assert(is_ANYOF_SYNTHETIC(ssc));
1365 _invlist_union_maybe_complement_2nd(ssc->invlist,
1371 PERL_STATIC_INLINE void
1372 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1374 const bool invert2nd)
1376 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1378 assert(is_ANYOF_SYNTHETIC(ssc));
1380 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1386 PERL_STATIC_INLINE void
1387 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1389 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1391 assert(is_ANYOF_SYNTHETIC(ssc));
1393 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1396 PERL_STATIC_INLINE void
1397 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1399 /* AND just the single code point 'cp' into the SSC 'ssc' */
1401 SV* cp_list = _new_invlist(2);
1403 PERL_ARGS_ASSERT_SSC_CP_AND;
1405 assert(is_ANYOF_SYNTHETIC(ssc));
1407 cp_list = add_cp_to_invlist(cp_list, cp);
1408 ssc_intersection(ssc, cp_list,
1409 FALSE /* Not inverted */
1411 SvREFCNT_dec_NN(cp_list);
1414 PERL_STATIC_INLINE void
1415 S_ssc_clear_locale(regnode_ssc *ssc)
1417 /* Set the SSC 'ssc' to not match any locale things */
1418 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1420 assert(is_ANYOF_SYNTHETIC(ssc));
1422 ANYOF_POSIXL_ZERO(ssc);
1423 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1427 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1429 /* The inversion list in the SSC is marked mortal; now we need a more
1430 * permanent copy, which is stored the same way that is done in a regular
1431 * ANYOF node, with the first 256 code points in a bit map */
1433 SV* invlist = invlist_clone(ssc->invlist);
1435 PERL_ARGS_ASSERT_SSC_FINALIZE;
1437 assert(is_ANYOF_SYNTHETIC(ssc));
1439 /* The code in this file assumes that all but these flags aren't relevant
1440 * to the SSC, except ANYOF_EMPTY_STRING, which should be cleared by the
1441 * time we reach here */
1442 assert(! (ANYOF_FLAGS(ssc) & ~ANYOF_COMMON_FLAGS));
1444 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1446 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1447 NULL, NULL, NULL, FALSE);
1449 /* Make sure is clone-safe */
1450 ssc->invlist = NULL;
1452 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1453 ANYOF_FLAGS(ssc) |= ANYOF_POSIXL;
1456 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1459 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1460 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1461 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1462 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1463 ? (TRIE_LIST_CUR( idx ) - 1) \
1469 dump_trie(trie,widecharmap,revcharmap)
1470 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1471 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1473 These routines dump out a trie in a somewhat readable format.
1474 The _interim_ variants are used for debugging the interim
1475 tables that are used to generate the final compressed
1476 representation which is what dump_trie expects.
1478 Part of the reason for their existence is to provide a form
1479 of documentation as to how the different representations function.
1484 Dumps the final compressed table form of the trie to Perl_debug_log.
1485 Used for debugging make_trie().
1489 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1490 AV *revcharmap, U32 depth)
1493 SV *sv=sv_newmortal();
1494 int colwidth= widecharmap ? 6 : 4;
1496 GET_RE_DEBUG_FLAGS_DECL;
1498 PERL_ARGS_ASSERT_DUMP_TRIE;
1500 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1501 (int)depth * 2 + 2,"",
1502 "Match","Base","Ofs" );
1504 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1505 SV ** const tmp = av_fetch( revcharmap, state, 0);
1507 PerlIO_printf( Perl_debug_log, "%*s",
1509 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1510 PL_colors[0], PL_colors[1],
1511 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1512 PERL_PV_ESCAPE_FIRSTCHAR
1517 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1518 (int)depth * 2 + 2,"");
1520 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1521 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1522 PerlIO_printf( Perl_debug_log, "\n");
1524 for( state = 1 ; state < trie->statecount ; state++ ) {
1525 const U32 base = trie->states[ state ].trans.base;
1527 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1528 (int)depth * 2 + 2,"", (UV)state);
1530 if ( trie->states[ state ].wordnum ) {
1531 PerlIO_printf( Perl_debug_log, " W%4X",
1532 trie->states[ state ].wordnum );
1534 PerlIO_printf( Perl_debug_log, "%6s", "" );
1537 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1542 while( ( base + ofs < trie->uniquecharcount ) ||
1543 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1544 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1548 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1550 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1551 if ( ( base + ofs >= trie->uniquecharcount )
1552 && ( base + ofs - trie->uniquecharcount
1554 && trie->trans[ base + ofs
1555 - trie->uniquecharcount ].check == state )
1557 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1559 (UV)trie->trans[ base + ofs
1560 - trie->uniquecharcount ].next );
1562 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1566 PerlIO_printf( Perl_debug_log, "]");
1569 PerlIO_printf( Perl_debug_log, "\n" );
1571 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1573 for (word=1; word <= trie->wordcount; word++) {
1574 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1575 (int)word, (int)(trie->wordinfo[word].prev),
1576 (int)(trie->wordinfo[word].len));
1578 PerlIO_printf(Perl_debug_log, "\n" );
1581 Dumps a fully constructed but uncompressed trie in list form.
1582 List tries normally only are used for construction when the number of
1583 possible chars (trie->uniquecharcount) is very high.
1584 Used for debugging make_trie().
1587 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1588 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1592 SV *sv=sv_newmortal();
1593 int colwidth= widecharmap ? 6 : 4;
1594 GET_RE_DEBUG_FLAGS_DECL;
1596 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1598 /* print out the table precompression. */
1599 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1600 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1601 "------:-----+-----------------\n" );
1603 for( state=1 ; state < next_alloc ; state ++ ) {
1606 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1607 (int)depth * 2 + 2,"", (UV)state );
1608 if ( ! trie->states[ state ].wordnum ) {
1609 PerlIO_printf( Perl_debug_log, "%5s| ","");
1611 PerlIO_printf( Perl_debug_log, "W%4x| ",
1612 trie->states[ state ].wordnum
1615 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1616 SV ** const tmp = av_fetch( revcharmap,
1617 TRIE_LIST_ITEM(state,charid).forid, 0);
1619 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1621 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1623 PL_colors[0], PL_colors[1],
1624 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1625 | PERL_PV_ESCAPE_FIRSTCHAR
1627 TRIE_LIST_ITEM(state,charid).forid,
1628 (UV)TRIE_LIST_ITEM(state,charid).newstate
1631 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1632 (int)((depth * 2) + 14), "");
1635 PerlIO_printf( Perl_debug_log, "\n");
1640 Dumps a fully constructed but uncompressed trie in table form.
1641 This is the normal DFA style state transition table, with a few
1642 twists to facilitate compression later.
1643 Used for debugging make_trie().
1646 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1647 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1652 SV *sv=sv_newmortal();
1653 int colwidth= widecharmap ? 6 : 4;
1654 GET_RE_DEBUG_FLAGS_DECL;
1656 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1659 print out the table precompression so that we can do a visual check
1660 that they are identical.
1663 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1665 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1666 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1668 PerlIO_printf( Perl_debug_log, "%*s",
1670 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1671 PL_colors[0], PL_colors[1],
1672 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1673 PERL_PV_ESCAPE_FIRSTCHAR
1679 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1681 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1682 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1685 PerlIO_printf( Perl_debug_log, "\n" );
1687 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1689 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1690 (int)depth * 2 + 2,"",
1691 (UV)TRIE_NODENUM( state ) );
1693 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1694 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1696 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1698 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1700 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1701 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1702 (UV)trie->trans[ state ].check );
1704 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1705 (UV)trie->trans[ state ].check,
1706 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1714 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1715 startbranch: the first branch in the whole branch sequence
1716 first : start branch of sequence of branch-exact nodes.
1717 May be the same as startbranch
1718 last : Thing following the last branch.
1719 May be the same as tail.
1720 tail : item following the branch sequence
1721 count : words in the sequence
1722 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS)/
1723 depth : indent depth
1725 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1727 A trie is an N'ary tree where the branches are determined by digital
1728 decomposition of the key. IE, at the root node you look up the 1st character and
1729 follow that branch repeat until you find the end of the branches. Nodes can be
1730 marked as "accepting" meaning they represent a complete word. Eg:
1734 would convert into the following structure. Numbers represent states, letters
1735 following numbers represent valid transitions on the letter from that state, if
1736 the number is in square brackets it represents an accepting state, otherwise it
1737 will be in parenthesis.
1739 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1743 (1) +-i->(6)-+-s->[7]
1745 +-s->(3)-+-h->(4)-+-e->[5]
1747 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1749 This shows that when matching against the string 'hers' we will begin at state 1
1750 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1751 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1752 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1753 single traverse. We store a mapping from accepting to state to which word was
1754 matched, and then when we have multiple possibilities we try to complete the
1755 rest of the regex in the order in which they occured in the alternation.
1757 The only prior NFA like behaviour that would be changed by the TRIE support is
1758 the silent ignoring of duplicate alternations which are of the form:
1760 / (DUPE|DUPE) X? (?{ ... }) Y /x
1762 Thus EVAL blocks following a trie may be called a different number of times with
1763 and without the optimisation. With the optimisations dupes will be silently
1764 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1765 the following demonstrates:
1767 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1769 which prints out 'word' three times, but
1771 'words'=~/(word|word|word)(?{ print $1 })S/
1773 which doesnt print it out at all. This is due to other optimisations kicking in.
1775 Example of what happens on a structural level:
1777 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1779 1: CURLYM[1] {1,32767}(18)
1790 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1791 and should turn into:
1793 1: CURLYM[1] {1,32767}(18)
1795 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1803 Cases where tail != last would be like /(?foo|bar)baz/:
1813 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1814 and would end up looking like:
1817 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1824 d = uvchr_to_utf8_flags(d, uv, 0);
1826 is the recommended Unicode-aware way of saying
1831 #define TRIE_STORE_REVCHAR(val) \
1834 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1835 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1836 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
1837 SvCUR_set(zlopp, kapow - flrbbbbb); \
1840 av_push(revcharmap, zlopp); \
1842 char ooooff = (char)val; \
1843 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1847 /* This gets the next character from the input, folding it if not already
1849 #define TRIE_READ_CHAR STMT_START { \
1852 /* if it is UTF then it is either already folded, or does not need \
1854 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
1856 else if (folder == PL_fold_latin1) { \
1857 /* This folder implies Unicode rules, which in the range expressible \
1858 * by not UTF is the lower case, with the two exceptions, one of \
1859 * which should have been taken care of before calling this */ \
1860 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
1861 uvc = toLOWER_L1(*uc); \
1862 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
1865 /* raw data, will be folded later if needed */ \
1873 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1874 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1875 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1876 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1878 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1879 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1880 TRIE_LIST_CUR( state )++; \
1883 #define TRIE_LIST_NEW(state) STMT_START { \
1884 Newxz( trie->states[ state ].trans.list, \
1885 4, reg_trie_trans_le ); \
1886 TRIE_LIST_CUR( state ) = 1; \
1887 TRIE_LIST_LEN( state ) = 4; \
1890 #define TRIE_HANDLE_WORD(state) STMT_START { \
1891 U16 dupe= trie->states[ state ].wordnum; \
1892 regnode * const noper_next = regnext( noper ); \
1895 /* store the word for dumping */ \
1897 if (OP(noper) != NOTHING) \
1898 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1900 tmp = newSVpvn_utf8( "", 0, UTF ); \
1901 av_push( trie_words, tmp ); \
1905 trie->wordinfo[curword].prev = 0; \
1906 trie->wordinfo[curword].len = wordlen; \
1907 trie->wordinfo[curword].accept = state; \
1909 if ( noper_next < tail ) { \
1911 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
1913 trie->jump[curword] = (U16)(noper_next - convert); \
1915 jumper = noper_next; \
1917 nextbranch= regnext(cur); \
1921 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1922 /* chain, so that when the bits of chain are later */\
1923 /* linked together, the dups appear in the chain */\
1924 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1925 trie->wordinfo[dupe].prev = curword; \
1927 /* we haven't inserted this word yet. */ \
1928 trie->states[ state ].wordnum = curword; \
1933 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1934 ( ( base + charid >= ucharcount \
1935 && base + charid < ubound \
1936 && state == trie->trans[ base - ucharcount + charid ].check \
1937 && trie->trans[ base - ucharcount + charid ].next ) \
1938 ? trie->trans[ base - ucharcount + charid ].next \
1939 : ( state==1 ? special : 0 ) \
1943 #define MADE_JUMP_TRIE 2
1944 #define MADE_EXACT_TRIE 4
1947 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
1948 regnode *first, regnode *last, regnode *tail,
1949 U32 word_count, U32 flags, U32 depth)
1951 /* first pass, loop through and scan words */
1952 reg_trie_data *trie;
1953 HV *widecharmap = NULL;
1954 AV *revcharmap = newAV();
1960 regnode *jumper = NULL;
1961 regnode *nextbranch = NULL;
1962 regnode *convert = NULL;
1963 U32 *prev_states; /* temp array mapping each state to previous one */
1964 /* we just use folder as a flag in utf8 */
1965 const U8 * folder = NULL;
1968 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
1969 AV *trie_words = NULL;
1970 /* along with revcharmap, this only used during construction but both are
1971 * useful during debugging so we store them in the struct when debugging.
1974 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
1975 STRLEN trie_charcount=0;
1977 SV *re_trie_maxbuff;
1978 GET_RE_DEBUG_FLAGS_DECL;
1980 PERL_ARGS_ASSERT_MAKE_TRIE;
1982 PERL_UNUSED_ARG(depth);
1989 case EXACTFU: folder = PL_fold_latin1; break;
1990 case EXACTF: folder = PL_fold; break;
1991 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1994 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1996 trie->startstate = 1;
1997 trie->wordcount = word_count;
1998 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1999 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2001 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2002 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2003 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2006 trie_words = newAV();
2009 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2010 assert(re_trie_maxbuff);
2011 if (!SvIOK(re_trie_maxbuff)) {
2012 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2014 DEBUG_TRIE_COMPILE_r({
2015 PerlIO_printf( Perl_debug_log,
2016 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2017 (int)depth * 2 + 2, "",
2018 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2019 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2022 /* Find the node we are going to overwrite */
2023 if ( first == startbranch && OP( last ) != BRANCH ) {
2024 /* whole branch chain */
2027 /* branch sub-chain */
2028 convert = NEXTOPER( first );
2031 /* -- First loop and Setup --
2033 We first traverse the branches and scan each word to determine if it
2034 contains widechars, and how many unique chars there are, this is
2035 important as we have to build a table with at least as many columns as we
2038 We use an array of integers to represent the character codes 0..255
2039 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2040 the native representation of the character value as the key and IV's for
2043 *TODO* If we keep track of how many times each character is used we can
2044 remap the columns so that the table compression later on is more
2045 efficient in terms of memory by ensuring the most common value is in the
2046 middle and the least common are on the outside. IMO this would be better
2047 than a most to least common mapping as theres a decent chance the most
2048 common letter will share a node with the least common, meaning the node
2049 will not be compressible. With a middle is most common approach the worst
2050 case is when we have the least common nodes twice.
2054 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2055 regnode *noper = NEXTOPER( cur );
2056 const U8 *uc = (U8*)STRING( noper );
2057 const U8 *e = uc + STR_LEN( noper );
2059 U32 wordlen = 0; /* required init */
2060 STRLEN minchars = 0;
2061 STRLEN maxchars = 0;
2062 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2065 if (OP(noper) == NOTHING) {
2066 regnode *noper_next= regnext(noper);
2067 if (noper_next != tail && OP(noper_next) == flags) {
2069 uc= (U8*)STRING(noper);
2070 e= uc + STR_LEN(noper);
2071 trie->minlen= STR_LEN(noper);
2078 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2079 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2080 regardless of encoding */
2081 if (OP( noper ) == EXACTFU_SS) {
2082 /* false positives are ok, so just set this */
2083 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2086 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2088 TRIE_CHARCOUNT(trie)++;
2091 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2092 * is in effect. Under /i, this character can match itself, or
2093 * anything that folds to it. If not under /i, it can match just
2094 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2095 * all fold to k, and all are single characters. But some folds
2096 * expand to more than one character, so for example LATIN SMALL
2097 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2098 * the string beginning at 'uc' is 'ffi', it could be matched by
2099 * three characters, or just by the one ligature character. (It
2100 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2101 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2102 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2103 * match.) The trie needs to know the minimum and maximum number
2104 * of characters that could match so that it can use size alone to
2105 * quickly reject many match attempts. The max is simple: it is
2106 * the number of folded characters in this branch (since a fold is
2107 * never shorter than what folds to it. */
2111 /* And the min is equal to the max if not under /i (indicated by
2112 * 'folder' being NULL), or there are no multi-character folds. If
2113 * there is a multi-character fold, the min is incremented just
2114 * once, for the character that folds to the sequence. Each
2115 * character in the sequence needs to be added to the list below of
2116 * characters in the trie, but we count only the first towards the
2117 * min number of characters needed. This is done through the
2118 * variable 'foldlen', which is returned by the macros that look
2119 * for these sequences as the number of bytes the sequence
2120 * occupies. Each time through the loop, we decrement 'foldlen' by
2121 * how many bytes the current char occupies. Only when it reaches
2122 * 0 do we increment 'minchars' or look for another multi-character
2124 if (folder == NULL) {
2127 else if (foldlen > 0) {
2128 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2133 /* See if *uc is the beginning of a multi-character fold. If
2134 * so, we decrement the length remaining to look at, to account
2135 * for the current character this iteration. (We can use 'uc'
2136 * instead of the fold returned by TRIE_READ_CHAR because for
2137 * non-UTF, the latin1_safe macro is smart enough to account
2138 * for all the unfolded characters, and because for UTF, the
2139 * string will already have been folded earlier in the
2140 * compilation process */
2142 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2143 foldlen -= UTF8SKIP(uc);
2146 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2151 /* The current character (and any potential folds) should be added
2152 * to the possible matching characters for this position in this
2156 U8 folded= folder[ (U8) uvc ];
2157 if ( !trie->charmap[ folded ] ) {
2158 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2159 TRIE_STORE_REVCHAR( folded );
2162 if ( !trie->charmap[ uvc ] ) {
2163 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2164 TRIE_STORE_REVCHAR( uvc );
2167 /* store the codepoint in the bitmap, and its folded
2169 TRIE_BITMAP_SET(trie, uvc);
2171 /* store the folded codepoint */
2172 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
2175 /* store first byte of utf8 representation of
2176 variant codepoints */
2177 if (! UVCHR_IS_INVARIANT(uvc)) {
2178 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
2181 set_bit = 0; /* We've done our bit :-) */
2185 /* XXX We could come up with the list of code points that fold
2186 * to this using PL_utf8_foldclosures, except not for
2187 * multi-char folds, as there may be multiple combinations
2188 * there that could work, which needs to wait until runtime to
2189 * resolve (The comment about LIGATURE FFI above is such an
2194 widecharmap = newHV();
2196 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2199 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
2201 if ( !SvTRUE( *svpp ) ) {
2202 sv_setiv( *svpp, ++trie->uniquecharcount );
2203 TRIE_STORE_REVCHAR(uvc);
2206 } /* end loop through characters in this branch of the trie */
2208 /* We take the min and max for this branch and combine to find the min
2209 * and max for all branches processed so far */
2210 if( cur == first ) {
2211 trie->minlen = minchars;
2212 trie->maxlen = maxchars;
2213 } else if (minchars < trie->minlen) {
2214 trie->minlen = minchars;
2215 } else if (maxchars > trie->maxlen) {
2216 trie->maxlen = maxchars;
2218 } /* end first pass */
2219 DEBUG_TRIE_COMPILE_r(
2220 PerlIO_printf( Perl_debug_log,
2221 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2222 (int)depth * 2 + 2,"",
2223 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2224 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2225 (int)trie->minlen, (int)trie->maxlen )
2229 We now know what we are dealing with in terms of unique chars and
2230 string sizes so we can calculate how much memory a naive
2231 representation using a flat table will take. If it's over a reasonable
2232 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2233 conservative but potentially much slower representation using an array
2236 At the end we convert both representations into the same compressed
2237 form that will be used in regexec.c for matching with. The latter
2238 is a form that cannot be used to construct with but has memory
2239 properties similar to the list form and access properties similar
2240 to the table form making it both suitable for fast searches and
2241 small enough that its feasable to store for the duration of a program.
2243 See the comment in the code where the compressed table is produced
2244 inplace from the flat tabe representation for an explanation of how
2245 the compression works.
2250 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2253 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2254 > SvIV(re_trie_maxbuff) )
2257 Second Pass -- Array Of Lists Representation
2259 Each state will be represented by a list of charid:state records
2260 (reg_trie_trans_le) the first such element holds the CUR and LEN
2261 points of the allocated array. (See defines above).
2263 We build the initial structure using the lists, and then convert
2264 it into the compressed table form which allows faster lookups
2265 (but cant be modified once converted).
2268 STRLEN transcount = 1;
2270 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2271 "%*sCompiling trie using list compiler\n",
2272 (int)depth * 2 + 2, ""));
2274 trie->states = (reg_trie_state *)
2275 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2276 sizeof(reg_trie_state) );
2280 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2282 regnode *noper = NEXTOPER( cur );
2283 U8 *uc = (U8*)STRING( noper );
2284 const U8 *e = uc + STR_LEN( noper );
2285 U32 state = 1; /* required init */
2286 U16 charid = 0; /* sanity init */
2287 U32 wordlen = 0; /* required init */
2289 if (OP(noper) == NOTHING) {
2290 regnode *noper_next= regnext(noper);
2291 if (noper_next != tail && OP(noper_next) == flags) {
2293 uc= (U8*)STRING(noper);
2294 e= uc + STR_LEN(noper);
2298 if (OP(noper) != NOTHING) {
2299 for ( ; uc < e ; uc += len ) {
2304 charid = trie->charmap[ uvc ];
2306 SV** const svpp = hv_fetch( widecharmap,
2313 charid=(U16)SvIV( *svpp );
2316 /* charid is now 0 if we dont know the char read, or
2317 * nonzero if we do */
2324 if ( !trie->states[ state ].trans.list ) {
2325 TRIE_LIST_NEW( state );
2328 check <= TRIE_LIST_USED( state );
2331 if ( TRIE_LIST_ITEM( state, check ).forid
2334 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2339 newstate = next_alloc++;
2340 prev_states[newstate] = state;
2341 TRIE_LIST_PUSH( state, charid, newstate );
2346 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2350 TRIE_HANDLE_WORD(state);
2352 } /* end second pass */
2354 /* next alloc is the NEXT state to be allocated */
2355 trie->statecount = next_alloc;
2356 trie->states = (reg_trie_state *)
2357 PerlMemShared_realloc( trie->states,
2359 * sizeof(reg_trie_state) );
2361 /* and now dump it out before we compress it */
2362 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2363 revcharmap, next_alloc,
2367 trie->trans = (reg_trie_trans *)
2368 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2375 for( state=1 ; state < next_alloc ; state ++ ) {
2379 DEBUG_TRIE_COMPILE_MORE_r(
2380 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2384 if (trie->states[state].trans.list) {
2385 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2389 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2390 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2391 if ( forid < minid ) {
2393 } else if ( forid > maxid ) {
2397 if ( transcount < tp + maxid - minid + 1) {
2399 trie->trans = (reg_trie_trans *)
2400 PerlMemShared_realloc( trie->trans,
2402 * sizeof(reg_trie_trans) );
2403 Zero( trie->trans + (transcount / 2),
2407 base = trie->uniquecharcount + tp - minid;
2408 if ( maxid == minid ) {
2410 for ( ; zp < tp ; zp++ ) {
2411 if ( ! trie->trans[ zp ].next ) {
2412 base = trie->uniquecharcount + zp - minid;
2413 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2415 trie->trans[ zp ].check = state;
2421 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2423 trie->trans[ tp ].check = state;
2428 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2429 const U32 tid = base
2430 - trie->uniquecharcount
2431 + TRIE_LIST_ITEM( state, idx ).forid;
2432 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2434 trie->trans[ tid ].check = state;
2436 tp += ( maxid - minid + 1 );
2438 Safefree(trie->states[ state ].trans.list);
2441 DEBUG_TRIE_COMPILE_MORE_r(
2442 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2445 trie->states[ state ].trans.base=base;
2447 trie->lasttrans = tp + 1;
2451 Second Pass -- Flat Table Representation.
2453 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2454 each. We know that we will need Charcount+1 trans at most to store
2455 the data (one row per char at worst case) So we preallocate both
2456 structures assuming worst case.
2458 We then construct the trie using only the .next slots of the entry
2461 We use the .check field of the first entry of the node temporarily
2462 to make compression both faster and easier by keeping track of how
2463 many non zero fields are in the node.
2465 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2468 There are two terms at use here: state as a TRIE_NODEIDX() which is
2469 a number representing the first entry of the node, and state as a
2470 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2471 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2472 if there are 2 entrys per node. eg:
2480 The table is internally in the right hand, idx form. However as we
2481 also have to deal with the states array which is indexed by nodenum
2482 we have to use TRIE_NODENUM() to convert.
2485 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2486 "%*sCompiling trie using table compiler\n",
2487 (int)depth * 2 + 2, ""));
2489 trie->trans = (reg_trie_trans *)
2490 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2491 * trie->uniquecharcount + 1,
2492 sizeof(reg_trie_trans) );
2493 trie->states = (reg_trie_state *)
2494 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2495 sizeof(reg_trie_state) );
2496 next_alloc = trie->uniquecharcount + 1;
2499 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2501 regnode *noper = NEXTOPER( cur );
2502 const U8 *uc = (U8*)STRING( noper );
2503 const U8 *e = uc + STR_LEN( noper );
2505 U32 state = 1; /* required init */
2507 U16 charid = 0; /* sanity init */
2508 U32 accept_state = 0; /* sanity init */
2510 U32 wordlen = 0; /* required init */
2512 if (OP(noper) == NOTHING) {
2513 regnode *noper_next= regnext(noper);
2514 if (noper_next != tail && OP(noper_next) == flags) {
2516 uc= (U8*)STRING(noper);
2517 e= uc + STR_LEN(noper);
2521 if ( OP(noper) != NOTHING ) {
2522 for ( ; uc < e ; uc += len ) {
2527 charid = trie->charmap[ uvc ];
2529 SV* const * const svpp = hv_fetch( widecharmap,
2533 charid = svpp ? (U16)SvIV(*svpp) : 0;
2537 if ( !trie->trans[ state + charid ].next ) {
2538 trie->trans[ state + charid ].next = next_alloc;
2539 trie->trans[ state ].check++;
2540 prev_states[TRIE_NODENUM(next_alloc)]
2541 = TRIE_NODENUM(state);
2542 next_alloc += trie->uniquecharcount;
2544 state = trie->trans[ state + charid ].next;
2546 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2548 /* charid is now 0 if we dont know the char read, or
2549 * nonzero if we do */
2552 accept_state = TRIE_NODENUM( state );
2553 TRIE_HANDLE_WORD(accept_state);
2555 } /* end second pass */
2557 /* and now dump it out before we compress it */
2558 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2560 next_alloc, depth+1));
2564 * Inplace compress the table.*
2566 For sparse data sets the table constructed by the trie algorithm will
2567 be mostly 0/FAIL transitions or to put it another way mostly empty.
2568 (Note that leaf nodes will not contain any transitions.)
2570 This algorithm compresses the tables by eliminating most such
2571 transitions, at the cost of a modest bit of extra work during lookup:
2573 - Each states[] entry contains a .base field which indicates the
2574 index in the state[] array wheres its transition data is stored.
2576 - If .base is 0 there are no valid transitions from that node.
2578 - If .base is nonzero then charid is added to it to find an entry in
2581 -If trans[states[state].base+charid].check!=state then the
2582 transition is taken to be a 0/Fail transition. Thus if there are fail
2583 transitions at the front of the node then the .base offset will point
2584 somewhere inside the previous nodes data (or maybe even into a node
2585 even earlier), but the .check field determines if the transition is
2589 The following process inplace converts the table to the compressed
2590 table: We first do not compress the root node 1,and mark all its
2591 .check pointers as 1 and set its .base pointer as 1 as well. This
2592 allows us to do a DFA construction from the compressed table later,
2593 and ensures that any .base pointers we calculate later are greater
2596 - We set 'pos' to indicate the first entry of the second node.
2598 - We then iterate over the columns of the node, finding the first and
2599 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2600 and set the .check pointers accordingly, and advance pos
2601 appropriately and repreat for the next node. Note that when we copy
2602 the next pointers we have to convert them from the original
2603 NODEIDX form to NODENUM form as the former is not valid post
2606 - If a node has no transitions used we mark its base as 0 and do not
2607 advance the pos pointer.
2609 - If a node only has one transition we use a second pointer into the
2610 structure to fill in allocated fail transitions from other states.
2611 This pointer is independent of the main pointer and scans forward
2612 looking for null transitions that are allocated to a state. When it
2613 finds one it writes the single transition into the "hole". If the
2614 pointer doesnt find one the single transition is appended as normal.
2616 - Once compressed we can Renew/realloc the structures to release the
2619 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2620 specifically Fig 3.47 and the associated pseudocode.
2624 const U32 laststate = TRIE_NODENUM( next_alloc );
2627 trie->statecount = laststate;
2629 for ( state = 1 ; state < laststate ; state++ ) {
2631 const U32 stateidx = TRIE_NODEIDX( state );
2632 const U32 o_used = trie->trans[ stateidx ].check;
2633 U32 used = trie->trans[ stateidx ].check;
2634 trie->trans[ stateidx ].check = 0;
2637 used && charid < trie->uniquecharcount;
2640 if ( flag || trie->trans[ stateidx + charid ].next ) {
2641 if ( trie->trans[ stateidx + charid ].next ) {
2643 for ( ; zp < pos ; zp++ ) {
2644 if ( ! trie->trans[ zp ].next ) {
2648 trie->states[ state ].trans.base
2650 + trie->uniquecharcount
2652 trie->trans[ zp ].next
2653 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2655 trie->trans[ zp ].check = state;
2656 if ( ++zp > pos ) pos = zp;
2663 trie->states[ state ].trans.base
2664 = pos + trie->uniquecharcount - charid ;
2666 trie->trans[ pos ].next
2667 = SAFE_TRIE_NODENUM(
2668 trie->trans[ stateidx + charid ].next );
2669 trie->trans[ pos ].check = state;
2674 trie->lasttrans = pos + 1;
2675 trie->states = (reg_trie_state *)
2676 PerlMemShared_realloc( trie->states, laststate
2677 * sizeof(reg_trie_state) );
2678 DEBUG_TRIE_COMPILE_MORE_r(
2679 PerlIO_printf( Perl_debug_log,
2680 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2681 (int)depth * 2 + 2,"",
2682 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2686 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2689 } /* end table compress */
2691 DEBUG_TRIE_COMPILE_MORE_r(
2692 PerlIO_printf(Perl_debug_log,
2693 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2694 (int)depth * 2 + 2, "",
2695 (UV)trie->statecount,
2696 (UV)trie->lasttrans)
2698 /* resize the trans array to remove unused space */
2699 trie->trans = (reg_trie_trans *)
2700 PerlMemShared_realloc( trie->trans, trie->lasttrans
2701 * sizeof(reg_trie_trans) );
2703 { /* Modify the program and insert the new TRIE node */
2704 U8 nodetype =(U8)(flags & 0xFF);
2708 regnode *optimize = NULL;
2709 #ifdef RE_TRACK_PATTERN_OFFSETS
2712 U32 mjd_nodelen = 0;
2713 #endif /* RE_TRACK_PATTERN_OFFSETS */
2714 #endif /* DEBUGGING */
2716 This means we convert either the first branch or the first Exact,
2717 depending on whether the thing following (in 'last') is a branch
2718 or not and whther first is the startbranch (ie is it a sub part of
2719 the alternation or is it the whole thing.)
2720 Assuming its a sub part we convert the EXACT otherwise we convert
2721 the whole branch sequence, including the first.
2723 /* Find the node we are going to overwrite */
2724 if ( first != startbranch || OP( last ) == BRANCH ) {
2725 /* branch sub-chain */
2726 NEXT_OFF( first ) = (U16)(last - first);
2727 #ifdef RE_TRACK_PATTERN_OFFSETS
2729 mjd_offset= Node_Offset((convert));
2730 mjd_nodelen= Node_Length((convert));
2733 /* whole branch chain */
2735 #ifdef RE_TRACK_PATTERN_OFFSETS
2738 const regnode *nop = NEXTOPER( convert );
2739 mjd_offset= Node_Offset((nop));
2740 mjd_nodelen= Node_Length((nop));
2744 PerlIO_printf(Perl_debug_log,
2745 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2746 (int)depth * 2 + 2, "",
2747 (UV)mjd_offset, (UV)mjd_nodelen)
2750 /* But first we check to see if there is a common prefix we can
2751 split out as an EXACT and put in front of the TRIE node. */
2752 trie->startstate= 1;
2753 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2755 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2759 const U32 base = trie->states[ state ].trans.base;
2761 if ( trie->states[state].wordnum )
2764 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2765 if ( ( base + ofs >= trie->uniquecharcount ) &&
2766 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2767 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2769 if ( ++count > 1 ) {
2770 SV **tmp = av_fetch( revcharmap, ofs, 0);
2771 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2772 if ( state == 1 ) break;
2774 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2776 PerlIO_printf(Perl_debug_log,
2777 "%*sNew Start State=%"UVuf" Class: [",
2778 (int)depth * 2 + 2, "",
2781 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2782 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2784 TRIE_BITMAP_SET(trie,*ch);
2786 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2788 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2792 TRIE_BITMAP_SET(trie,*ch);
2794 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2795 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2801 SV **tmp = av_fetch( revcharmap, idx, 0);
2803 char *ch = SvPV( *tmp, len );
2805 SV *sv=sv_newmortal();
2806 PerlIO_printf( Perl_debug_log,
2807 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2808 (int)depth * 2 + 2, "",
2810 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2811 PL_colors[0], PL_colors[1],
2812 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2813 PERL_PV_ESCAPE_FIRSTCHAR
2818 OP( convert ) = nodetype;
2819 str=STRING(convert);
2822 STR_LEN(convert) += len;
2828 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2833 trie->prefixlen = (state-1);
2835 regnode *n = convert+NODE_SZ_STR(convert);
2836 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2837 trie->startstate = state;
2838 trie->minlen -= (state - 1);
2839 trie->maxlen -= (state - 1);
2841 /* At least the UNICOS C compiler choked on this
2842 * being argument to DEBUG_r(), so let's just have
2845 #ifdef PERL_EXT_RE_BUILD
2851 regnode *fix = convert;
2852 U32 word = trie->wordcount;
2854 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2855 while( ++fix < n ) {
2856 Set_Node_Offset_Length(fix, 0, 0);
2859 SV ** const tmp = av_fetch( trie_words, word, 0 );
2861 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2862 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2864 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2872 NEXT_OFF(convert) = (U16)(tail - convert);
2873 DEBUG_r(optimize= n);
2879 if ( trie->maxlen ) {
2880 NEXT_OFF( convert ) = (U16)(tail - convert);
2881 ARG_SET( convert, data_slot );
2882 /* Store the offset to the first unabsorbed branch in
2883 jump[0], which is otherwise unused by the jump logic.
2884 We use this when dumping a trie and during optimisation. */
2886 trie->jump[0] = (U16)(nextbranch - convert);
2888 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2889 * and there is a bitmap
2890 * and the first "jump target" node we found leaves enough room
2891 * then convert the TRIE node into a TRIEC node, with the bitmap
2892 * embedded inline in the opcode - this is hypothetically faster.
2894 if ( !trie->states[trie->startstate].wordnum
2896 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2898 OP( convert ) = TRIEC;
2899 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2900 PerlMemShared_free(trie->bitmap);
2903 OP( convert ) = TRIE;
2905 /* store the type in the flags */
2906 convert->flags = nodetype;
2910 + regarglen[ OP( convert ) ];
2912 /* XXX We really should free up the resource in trie now,
2913 as we won't use them - (which resources?) dmq */
2915 /* needed for dumping*/
2916 DEBUG_r(if (optimize) {
2917 regnode *opt = convert;
2919 while ( ++opt < optimize) {
2920 Set_Node_Offset_Length(opt,0,0);
2923 Try to clean up some of the debris left after the
2926 while( optimize < jumper ) {
2927 mjd_nodelen += Node_Length((optimize));
2928 OP( optimize ) = OPTIMIZED;
2929 Set_Node_Offset_Length(optimize,0,0);
2932 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2934 } /* end node insert */
2936 /* Finish populating the prev field of the wordinfo array. Walk back
2937 * from each accept state until we find another accept state, and if
2938 * so, point the first word's .prev field at the second word. If the
2939 * second already has a .prev field set, stop now. This will be the
2940 * case either if we've already processed that word's accept state,
2941 * or that state had multiple words, and the overspill words were
2942 * already linked up earlier.
2949 for (word=1; word <= trie->wordcount; word++) {
2951 if (trie->wordinfo[word].prev)
2953 state = trie->wordinfo[word].accept;
2955 state = prev_states[state];
2958 prev = trie->states[state].wordnum;
2962 trie->wordinfo[word].prev = prev;
2964 Safefree(prev_states);
2968 /* and now dump out the compressed format */
2969 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2971 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2973 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2974 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2976 SvREFCNT_dec_NN(revcharmap);
2980 : trie->startstate>1
2986 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
2988 /* The Trie is constructed and compressed now so we can build a fail array if
2991 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
2993 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
2997 We find the fail state for each state in the trie, this state is the longest
2998 proper suffix of the current state's 'word' that is also a proper prefix of
2999 another word in our trie. State 1 represents the word '' and is thus the
3000 default fail state. This allows the DFA not to have to restart after its
3001 tried and failed a word at a given point, it simply continues as though it
3002 had been matching the other word in the first place.
3004 'abcdgu'=~/abcdefg|cdgu/
3005 When we get to 'd' we are still matching the first word, we would encounter
3006 'g' which would fail, which would bring us to the state representing 'd' in
3007 the second word where we would try 'g' and succeed, proceeding to match
3010 /* add a fail transition */
3011 const U32 trie_offset = ARG(source);
3012 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3014 const U32 ucharcount = trie->uniquecharcount;
3015 const U32 numstates = trie->statecount;
3016 const U32 ubound = trie->lasttrans + ucharcount;
3020 U32 base = trie->states[ 1 ].trans.base;
3023 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3025 GET_RE_DEBUG_FLAGS_DECL;
3027 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3028 PERL_UNUSED_CONTEXT;
3030 PERL_UNUSED_ARG(depth);
3033 if ( OP(source) == TRIE ) {
3034 struct regnode_1 *op = (struct regnode_1 *)
3035 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3036 StructCopy(source,op,struct regnode_1);
3037 stclass = (regnode *)op;
3039 struct regnode_charclass *op = (struct regnode_charclass *)
3040 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3041 StructCopy(source,op,struct regnode_charclass);
3042 stclass = (regnode *)op;
3044 OP(stclass)+=2; /* covert the TRIE type to its AHO-CORASICK equivalent */
3046 ARG_SET( stclass, data_slot );
3047 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3048 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3049 aho->trie=trie_offset;
3050 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3051 Copy( trie->states, aho->states, numstates, reg_trie_state );
3052 Newxz( q, numstates, U32);
3053 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3056 /* initialize fail[0..1] to be 1 so that we always have
3057 a valid final fail state */
3058 fail[ 0 ] = fail[ 1 ] = 1;
3060 for ( charid = 0; charid < ucharcount ; charid++ ) {
3061 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3063 q[ q_write ] = newstate;
3064 /* set to point at the root */
3065 fail[ q[ q_write++ ] ]=1;
3068 while ( q_read < q_write) {
3069 const U32 cur = q[ q_read++ % numstates ];
3070 base = trie->states[ cur ].trans.base;
3072 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3073 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3075 U32 fail_state = cur;
3078 fail_state = fail[ fail_state ];
3079 fail_base = aho->states[ fail_state ].trans.base;
3080 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3082 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3083 fail[ ch_state ] = fail_state;
3084 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3086 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3088 q[ q_write++ % numstates] = ch_state;
3092 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3093 when we fail in state 1, this allows us to use the
3094 charclass scan to find a valid start char. This is based on the principle
3095 that theres a good chance the string being searched contains lots of stuff
3096 that cant be a start char.
3098 fail[ 0 ] = fail[ 1 ] = 0;
3099 DEBUG_TRIE_COMPILE_r({
3100 PerlIO_printf(Perl_debug_log,
3101 "%*sStclass Failtable (%"UVuf" states): 0",
3102 (int)(depth * 2), "", (UV)numstates
3104 for( q_read=1; q_read<numstates; q_read++ ) {
3105 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3107 PerlIO_printf(Perl_debug_log, "\n");
3110 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3115 #define DEBUG_PEEP(str,scan,depth) \
3116 DEBUG_OPTIMISE_r({if (scan){ \
3117 SV * const mysv=sv_newmortal(); \
3118 regnode *Next = regnext(scan); \
3119 regprop(RExC_rx, mysv, scan, NULL); \
3120 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
3121 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
3122 Next ? (REG_NODE_NUM(Next)) : 0 ); \
3126 /* The below joins as many adjacent EXACTish nodes as possible into a single
3127 * one. The regop may be changed if the node(s) contain certain sequences that
3128 * require special handling. The joining is only done if:
3129 * 1) there is room in the current conglomerated node to entirely contain the
3131 * 2) they are the exact same node type
3133 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3134 * these get optimized out
3136 * If a node is to match under /i (folded), the number of characters it matches
3137 * can be different than its character length if it contains a multi-character
3138 * fold. *min_subtract is set to the total delta number of characters of the
3141 * And *unfolded_multi_char is set to indicate whether or not the node contains
3142 * an unfolded multi-char fold. This happens when whether the fold is valid or
3143 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3144 * SMALL LETTER SHARP S, as only if the target string being matched against
3145 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3146 * folding rules depend on the locale in force at runtime. (Multi-char folds
3147 * whose components are all above the Latin1 range are not run-time locale
3148 * dependent, and have already been folded by the time this function is
3151 * This is as good a place as any to discuss the design of handling these
3152 * multi-character fold sequences. It's been wrong in Perl for a very long
3153 * time. There are three code points in Unicode whose multi-character folds
3154 * were long ago discovered to mess things up. The previous designs for
3155 * dealing with these involved assigning a special node for them. This
3156 * approach doesn't always work, as evidenced by this example:
3157 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3158 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3159 * would match just the \xDF, it won't be able to handle the case where a
3160 * successful match would have to cross the node's boundary. The new approach
3161 * that hopefully generally solves the problem generates an EXACTFU_SS node
3162 * that is "sss" in this case.
3164 * It turns out that there are problems with all multi-character folds, and not
3165 * just these three. Now the code is general, for all such cases. The
3166 * approach taken is:
3167 * 1) This routine examines each EXACTFish node that could contain multi-
3168 * character folded sequences. Since a single character can fold into
3169 * such a sequence, the minimum match length for this node is less than
3170 * the number of characters in the node. This routine returns in
3171 * *min_subtract how many characters to subtract from the the actual
3172 * length of the string to get a real minimum match length; it is 0 if
3173 * there are no multi-char foldeds. This delta is used by the caller to
3174 * adjust the min length of the match, and the delta between min and max,
3175 * so that the optimizer doesn't reject these possibilities based on size
3177 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3178 * is used for an EXACTFU node that contains at least one "ss" sequence in
3179 * it. For non-UTF-8 patterns and strings, this is the only case where
3180 * there is a possible fold length change. That means that a regular
3181 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3182 * with length changes, and so can be processed faster. regexec.c takes
3183 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3184 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3185 * known until runtime). This saves effort in regex matching. However,
3186 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3187 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3188 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3189 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3190 * possibilities for the non-UTF8 patterns are quite simple, except for
3191 * the sharp s. All the ones that don't involve a UTF-8 target string are
3192 * members of a fold-pair, and arrays are set up for all of them so that
3193 * the other member of the pair can be found quickly. Code elsewhere in
3194 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3195 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3196 * described in the next item.
3197 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3198 * validity of the fold won't be known until runtime, and so must remain
3199 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3200 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3201 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3202 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3203 * The reason this is a problem is that the optimizer part of regexec.c
3204 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3205 * that a character in the pattern corresponds to at most a single
3206 * character in the target string. (And I do mean character, and not byte
3207 * here, unlike other parts of the documentation that have never been
3208 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3209 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3210 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3211 * nodes, violate the assumption, and they are the only instances where it
3212 * is violated. I'm reluctant to try to change the assumption, as the
3213 * code involved is impenetrable to me (khw), so instead the code here
3214 * punts. This routine examines EXACTFL nodes, and (when the pattern
3215 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3216 * boolean indicating whether or not the node contains such a fold. When
3217 * it is true, the caller sets a flag that later causes the optimizer in
3218 * this file to not set values for the floating and fixed string lengths,
3219 * and thus avoids the optimizer code in regexec.c that makes the invalid
3220 * assumption. Thus, there is no optimization based on string lengths for
3221 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3222 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3223 * assumption is wrong only in these cases is that all other non-UTF-8
3224 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3225 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3226 * EXACTF nodes because we don't know at compile time if it actually
3227 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3228 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3229 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3230 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3231 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3232 * string would require the pattern to be forced into UTF-8, the overhead
3233 * of which we want to avoid. Similarly the unfolded multi-char folds in
3234 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3237 * Similarly, the code that generates tries doesn't currently handle
3238 * not-already-folded multi-char folds, and it looks like a pain to change
3239 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3240 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3241 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3242 * using /iaa matching will be doing so almost entirely with ASCII
3243 * strings, so this should rarely be encountered in practice */
3245 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3246 if (PL_regkind[OP(scan)] == EXACT) \
3247 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3250 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3251 UV *min_subtract, bool *unfolded_multi_char,
3252 U32 flags,regnode *val, U32 depth)
3254 /* Merge several consecutive EXACTish nodes into one. */
3255 regnode *n = regnext(scan);
3257 regnode *next = scan + NODE_SZ_STR(scan);
3261 regnode *stop = scan;
3262 GET_RE_DEBUG_FLAGS_DECL;
3264 PERL_UNUSED_ARG(depth);
3267 PERL_ARGS_ASSERT_JOIN_EXACT;
3268 #ifndef EXPERIMENTAL_INPLACESCAN
3269 PERL_UNUSED_ARG(flags);
3270 PERL_UNUSED_ARG(val);
3272 DEBUG_PEEP("join",scan,depth);
3274 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3275 * EXACT ones that are mergeable to the current one. */
3277 && (PL_regkind[OP(n)] == NOTHING
3278 || (stringok && OP(n) == OP(scan)))
3280 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3283 if (OP(n) == TAIL || n > next)
3285 if (PL_regkind[OP(n)] == NOTHING) {
3286 DEBUG_PEEP("skip:",n,depth);
3287 NEXT_OFF(scan) += NEXT_OFF(n);
3288 next = n + NODE_STEP_REGNODE;
3295 else if (stringok) {
3296 const unsigned int oldl = STR_LEN(scan);
3297 regnode * const nnext = regnext(n);
3299 /* XXX I (khw) kind of doubt that this works on platforms (should
3300 * Perl ever run on one) where U8_MAX is above 255 because of lots
3301 * of other assumptions */
3302 /* Don't join if the sum can't fit into a single node */
3303 if (oldl + STR_LEN(n) > U8_MAX)
3306 DEBUG_PEEP("merg",n,depth);
3309 NEXT_OFF(scan) += NEXT_OFF(n);
3310 STR_LEN(scan) += STR_LEN(n);
3311 next = n + NODE_SZ_STR(n);
3312 /* Now we can overwrite *n : */
3313 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3321 #ifdef EXPERIMENTAL_INPLACESCAN
3322 if (flags && !NEXT_OFF(n)) {
3323 DEBUG_PEEP("atch", val, depth);
3324 if (reg_off_by_arg[OP(n)]) {
3325 ARG_SET(n, val - n);
3328 NEXT_OFF(n) = val - n;
3336 *unfolded_multi_char = FALSE;
3338 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3339 * can now analyze for sequences of problematic code points. (Prior to
3340 * this final joining, sequences could have been split over boundaries, and
3341 * hence missed). The sequences only happen in folding, hence for any
3342 * non-EXACT EXACTish node */
3343 if (OP(scan) != EXACT) {
3344 U8* s0 = (U8*) STRING(scan);
3346 U8* s_end = s0 + STR_LEN(scan);
3348 int total_count_delta = 0; /* Total delta number of characters that
3349 multi-char folds expand to */
3351 /* One pass is made over the node's string looking for all the
3352 * possibilities. To avoid some tests in the loop, there are two main
3353 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3358 if (OP(scan) == EXACTFL) {
3361 /* An EXACTFL node would already have been changed to another
3362 * node type unless there is at least one character in it that
3363 * is problematic; likely a character whose fold definition
3364 * won't be known until runtime, and so has yet to be folded.
3365 * For all but the UTF-8 locale, folds are 1-1 in length, but
3366 * to handle the UTF-8 case, we need to create a temporary
3367 * folded copy using UTF-8 locale rules in order to analyze it.
3368 * This is because our macros that look to see if a sequence is
3369 * a multi-char fold assume everything is folded (otherwise the
3370 * tests in those macros would be too complicated and slow).
3371 * Note that here, the non-problematic folds will have already
3372 * been done, so we can just copy such characters. We actually
3373 * don't completely fold the EXACTFL string. We skip the
3374 * unfolded multi-char folds, as that would just create work
3375 * below to figure out the size they already are */
3377 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3380 STRLEN s_len = UTF8SKIP(s);
3381 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3382 Copy(s, d, s_len, U8);
3385 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3386 *unfolded_multi_char = TRUE;
3387 Copy(s, d, s_len, U8);
3390 else if (isASCII(*s)) {
3391 *(d++) = toFOLD(*s);
3395 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3401 /* Point the remainder of the routine to look at our temporary
3405 } /* End of creating folded copy of EXACTFL string */
3407 /* Examine the string for a multi-character fold sequence. UTF-8
3408 * patterns have all characters pre-folded by the time this code is
3410 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3411 length sequence we are looking for is 2 */
3413 int count = 0; /* How many characters in a multi-char fold */
3414 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3415 if (! len) { /* Not a multi-char fold: get next char */
3420 /* Nodes with 'ss' require special handling, except for
3421 * EXACTFA-ish for which there is no multi-char fold to this */
3422 if (len == 2 && *s == 's' && *(s+1) == 's'
3423 && OP(scan) != EXACTFA
3424 && OP(scan) != EXACTFA_NO_TRIE)
3427 if (OP(scan) != EXACTFL) {
3428 OP(scan) = EXACTFU_SS;
3432 else { /* Here is a generic multi-char fold. */
3433 U8* multi_end = s + len;
3435 /* Count how many characters are in it. In the case of
3436 * /aa, no folds which contain ASCII code points are
3437 * allowed, so check for those, and skip if found. */
3438 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3439 count = utf8_length(s, multi_end);
3443 while (s < multi_end) {
3446 goto next_iteration;
3456 /* The delta is how long the sequence is minus 1 (1 is how long
3457 * the character that folds to the sequence is) */
3458 total_count_delta += count - 1;
3462 /* We created a temporary folded copy of the string in EXACTFL
3463 * nodes. Therefore we need to be sure it doesn't go below zero,
3464 * as the real string could be shorter */
3465 if (OP(scan) == EXACTFL) {
3466 int total_chars = utf8_length((U8*) STRING(scan),
3467 (U8*) STRING(scan) + STR_LEN(scan));
3468 if (total_count_delta > total_chars) {
3469 total_count_delta = total_chars;
3473 *min_subtract += total_count_delta;
3476 else if (OP(scan) == EXACTFA) {
3478 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3479 * fold to the ASCII range (and there are no existing ones in the
3480 * upper latin1 range). But, as outlined in the comments preceding
3481 * this function, we need to flag any occurrences of the sharp s.
3482 * This character forbids trie formation (because of added
3485 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
3486 OP(scan) = EXACTFA_NO_TRIE;
3487 *unfolded_multi_char = TRUE;
3496 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3497 * folds that are all Latin1. As explained in the comments
3498 * preceding this function, we look also for the sharp s in EXACTF
3499 * and EXACTFL nodes; it can be in the final position. Otherwise
3500 * we can stop looking 1 byte earlier because have to find at least
3501 * two characters for a multi-fold */
3502 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
3507 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
3508 if (! len) { /* Not a multi-char fold. */
3509 if (*s == LATIN_SMALL_LETTER_SHARP_S
3510 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
3512 *unfolded_multi_char = TRUE;
3519 && isARG2_lower_or_UPPER_ARG1('s', *s)
3520 && isARG2_lower_or_UPPER_ARG1('s', *(s+1)))
3523 /* EXACTF nodes need to know that the minimum length
3524 * changed so that a sharp s in the string can match this
3525 * ss in the pattern, but they remain EXACTF nodes, as they
3526 * won't match this unless the target string is is UTF-8,
3527 * which we don't know until runtime. EXACTFL nodes can't
3528 * transform into EXACTFU nodes */
3529 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
3530 OP(scan) = EXACTFU_SS;
3534 *min_subtract += len - 1;
3541 /* Allow dumping but overwriting the collection of skipped
3542 * ops and/or strings with fake optimized ops */
3543 n = scan + NODE_SZ_STR(scan);
3551 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3555 /* REx optimizer. Converts nodes into quicker variants "in place".
3556 Finds fixed substrings. */
3558 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3559 to the position after last scanned or to NULL. */
3561 #define INIT_AND_WITHP \
3562 assert(!and_withp); \
3563 Newx(and_withp,1, regnode_ssc); \
3564 SAVEFREEPV(and_withp)
3566 /* this is a chain of data about sub patterns we are processing that
3567 need to be handled separately/specially in study_chunk. Its so
3568 we can simulate recursion without losing state. */
3570 typedef struct scan_frame {
3571 regnode *last; /* last node to process in this frame */
3572 regnode *next; /* next node to process when last is reached */
3573 struct scan_frame *prev; /*previous frame*/
3574 U32 prev_recursed_depth;
3575 I32 stop; /* what stopparen do we use */
3580 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3581 SSize_t *minlenp, SSize_t *deltap,
3586 regnode_ssc *and_withp,
3587 U32 flags, U32 depth)
3588 /* scanp: Start here (read-write). */
3589 /* deltap: Write maxlen-minlen here. */
3590 /* last: Stop before this one. */
3591 /* data: string data about the pattern */
3592 /* stopparen: treat close N as END */
3593 /* recursed: which subroutines have we recursed into */
3594 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3596 /* There must be at least this number of characters to match */
3599 regnode *scan = *scanp, *next;
3601 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3602 int is_inf_internal = 0; /* The studied chunk is infinite */
3603 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3604 scan_data_t data_fake;
3605 SV *re_trie_maxbuff = NULL;
3606 regnode *first_non_open = scan;
3607 SSize_t stopmin = SSize_t_MAX;
3608 scan_frame *frame = NULL;
3609 GET_RE_DEBUG_FLAGS_DECL;
3611 PERL_ARGS_ASSERT_STUDY_CHUNK;
3614 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3617 while (first_non_open && OP(first_non_open) == OPEN)
3618 first_non_open=regnext(first_non_open);
3623 while ( scan && OP(scan) != END && scan < last ){
3624 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3625 node length to get a real minimum (because
3626 the folded version may be shorter) */
3627 bool unfolded_multi_char = FALSE;
3628 /* Peephole optimizer: */
3629 DEBUG_OPTIMISE_MORE_r(
3631 PerlIO_printf(Perl_debug_log,
3632 "%*sstudy_chunk stopparen=%ld depth=%lu recursed_depth=%lu ",
3633 ((int) depth*2), "", (long)stopparen,
3634 (unsigned long)depth, (unsigned long)recursed_depth);
3635 if (recursed_depth) {
3638 for ( j = 0 ; j < recursed_depth ; j++ ) {
3639 PerlIO_printf(Perl_debug_log,"[");
3640 for ( i = 0 ; i < (U32)RExC_npar ; i++ )
3641 PerlIO_printf(Perl_debug_log,"%d",
3642 PAREN_TEST(RExC_study_chunk_recursed +
3643 (j * RExC_study_chunk_recursed_bytes), i)
3646 PerlIO_printf(Perl_debug_log,"]");
3649 PerlIO_printf(Perl_debug_log,"\n");
3652 DEBUG_STUDYDATA("Peep:", data, depth);
3653 DEBUG_PEEP("Peep", scan, depth);
3656 /* The reason we do this here we need to deal with things like /(?:f)(?:o)(?:o)/
3657 * which cant be dealt with by the normal EXACT parsing code, as each (?:..) is handled
3658 * by a different invocation of reg() -- Yves
3660 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
3662 /* Follow the next-chain of the current node and optimize
3663 away all the NOTHINGs from it. */
3664 if (OP(scan) != CURLYX) {
3665 const int max = (reg_off_by_arg[OP(scan)]
3667 /* I32 may be smaller than U16 on CRAYs! */
3668 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3669 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3673 /* Skip NOTHING and LONGJMP. */
3674 while ((n = regnext(n))
3675 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3676 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3677 && off + noff < max)
3679 if (reg_off_by_arg[OP(scan)])
3682 NEXT_OFF(scan) = off;
3687 /* The principal pseudo-switch. Cannot be a switch, since we
3688 look into several different things. */
3689 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3690 || OP(scan) == IFTHEN) {
3691 next = regnext(scan);
3693 /* demq: the op(next)==code check is to see if we have
3694 * "branch-branch" AFAICT */
3696 if (OP(next) == code || code == IFTHEN) {
3697 /* NOTE - There is similar code to this block below for
3698 * handling TRIE nodes on a re-study. If you change stuff here
3699 * check there too. */
3700 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
3702 regnode * const startbranch=scan;
3704 if (flags & SCF_DO_SUBSTR) {
3705 /* Cannot merge strings after this. */
3706 scan_commit(pRExC_state, data, minlenp, is_inf);
3709 if (flags & SCF_DO_STCLASS)
3710 ssc_init_zero(pRExC_state, &accum);
3712 while (OP(scan) == code) {
3713 SSize_t deltanext, minnext, fake;
3715 regnode_ssc this_class;
3718 data_fake.flags = 0;
3720 data_fake.whilem_c = data->whilem_c;
3721 data_fake.last_closep = data->last_closep;
3724 data_fake.last_closep = &fake;
3726 data_fake.pos_delta = delta;
3727 next = regnext(scan);
3728 scan = NEXTOPER(scan);
3730 scan = NEXTOPER(scan);
3731 if (flags & SCF_DO_STCLASS) {
3732 ssc_init(pRExC_state, &this_class);
3733 data_fake.start_class = &this_class;
3734 f = SCF_DO_STCLASS_AND;
3736 if (flags & SCF_WHILEM_VISITED_POS)
3737 f |= SCF_WHILEM_VISITED_POS;
3739 /* we suppose the run is continuous, last=next...*/
3740 minnext = study_chunk(pRExC_state, &scan, minlenp,
3741 &deltanext, next, &data_fake, stopparen,
3742 recursed_depth, NULL, f,depth+1);
3745 if (deltanext == SSize_t_MAX) {
3746 is_inf = is_inf_internal = 1;
3748 } else if (max1 < minnext + deltanext)
3749 max1 = minnext + deltanext;
3751 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3753 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3754 if ( stopmin > minnext)
3755 stopmin = min + min1;
3756 flags &= ~SCF_DO_SUBSTR;
3758 data->flags |= SCF_SEEN_ACCEPT;
3761 if (data_fake.flags & SF_HAS_EVAL)
3762 data->flags |= SF_HAS_EVAL;
3763 data->whilem_c = data_fake.whilem_c;
3765 if (flags & SCF_DO_STCLASS)
3766 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
3768 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3770 if (flags & SCF_DO_SUBSTR) {
3771 data->pos_min += min1;
3772 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3773 data->pos_delta = SSize_t_MAX;
3775 data->pos_delta += max1 - min1;
3776 if (max1 != min1 || is_inf)
3777 data->longest = &(data->longest_float);
3780 if (delta == SSize_t_MAX
3781 || SSize_t_MAX - delta - (max1 - min1) < 0)
3782 delta = SSize_t_MAX;
3784 delta += max1 - min1;
3785 if (flags & SCF_DO_STCLASS_OR) {
3786 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
3788 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
3789 flags &= ~SCF_DO_STCLASS;
3792 else if (flags & SCF_DO_STCLASS_AND) {
3794 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
3795 flags &= ~SCF_DO_STCLASS;
3798 /* Switch to OR mode: cache the old value of
3799 * data->start_class */
3801 StructCopy(data->start_class, and_withp, regnode_ssc);
3802 flags &= ~SCF_DO_STCLASS_AND;
3803 StructCopy(&accum, data->start_class, regnode_ssc);
3804 flags |= SCF_DO_STCLASS_OR;
3808 if (PERL_ENABLE_TRIE_OPTIMISATION &&
3809 OP( startbranch ) == BRANCH )
3813 Assuming this was/is a branch we are dealing with: 'scan'
3814 now points at the item that follows the branch sequence,
3815 whatever it is. We now start at the beginning of the
3816 sequence and look for subsequences of
3822 which would be constructed from a pattern like
3825 If we can find such a subsequence we need to turn the first
3826 element into a trie and then add the subsequent branch exact
3827 strings to the trie.
3831 1. patterns where the whole set of branches can be
3834 2. patterns where only a subset can be converted.
3836 In case 1 we can replace the whole set with a single regop
3837 for the trie. In case 2 we need to keep the start and end
3840 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3841 becomes BRANCH TRIE; BRANCH X;
3843 There is an additional case, that being where there is a
3844 common prefix, which gets split out into an EXACT like node
3845 preceding the TRIE node.
3847 If x(1..n)==tail then we can do a simple trie, if not we make
3848 a "jump" trie, such that when we match the appropriate word
3849 we "jump" to the appropriate tail node. Essentially we turn
3850 a nested if into a case structure of sorts.
3855 if (!re_trie_maxbuff) {
3856 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3857 if (!SvIOK(re_trie_maxbuff))
3858 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3860 if ( SvIV(re_trie_maxbuff)>=0 ) {
3862 regnode *first = (regnode *)NULL;
3863 regnode *last = (regnode *)NULL;
3864 regnode *tail = scan;
3869 SV * const mysv = sv_newmortal(); /* for dumping */
3871 /* var tail is used because there may be a TAIL
3872 regop in the way. Ie, the exacts will point to the
3873 thing following the TAIL, but the last branch will
3874 point at the TAIL. So we advance tail. If we
3875 have nested (?:) we may have to move through several
3879 while ( OP( tail ) == TAIL ) {
3880 /* this is the TAIL generated by (?:) */
3881 tail = regnext( tail );
3885 DEBUG_TRIE_COMPILE_r({
3886 regprop(RExC_rx, mysv, tail, NULL);
3887 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3888 (int)depth * 2 + 2, "",
3889 "Looking for TRIE'able sequences. Tail node is: ",
3890 SvPV_nolen_const( mysv )
3896 Step through the branches
3897 cur represents each branch,
3898 noper is the first thing to be matched as part
3900 noper_next is the regnext() of that node.
3902 We normally handle a case like this
3903 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
3904 support building with NOJUMPTRIE, which restricts
3905 the trie logic to structures like /FOO|BAR/.
3907 If noper is a trieable nodetype then the branch is
3908 a possible optimization target. If we are building
3909 under NOJUMPTRIE then we require that noper_next is
3910 the same as scan (our current position in the regex
3913 Once we have two or more consecutive such branches
3914 we can create a trie of the EXACT's contents and
3915 stitch it in place into the program.
3917 If the sequence represents all of the branches in
3918 the alternation we replace the entire thing with a
3921 Otherwise when it is a subsequence we need to
3922 stitch it in place and replace only the relevant
3923 branches. This means the first branch has to remain
3924 as it is used by the alternation logic, and its
3925 next pointer, and needs to be repointed at the item
https://perl5.git.perl.org / perl5.git / blob