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 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) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
96 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
103 # if defined(BUGGY_MSC6)
104 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
105 # pragma optimize("a",off)
106 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
107 # pragma optimize("w",on )
108 # endif /* BUGGY_MSC6 */
112 #define STATIC static
116 typedef struct RExC_state_t {
117 U32 flags; /* RXf_* are we folding, multilining? */
118 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
119 char *precomp; /* uncompiled string. */
120 REGEXP *rx_sv; /* The SV that is the regexp. */
121 regexp *rx; /* perl core regexp structure */
122 regexp_internal *rxi; /* internal data for regexp object pprivate field */
123 char *start; /* Start of input for compile */
124 char *end; /* End of input for compile */
125 char *parse; /* Input-scan pointer. */
126 SSize_t whilem_seen; /* number of WHILEM in this expr */
127 regnode *emit_start; /* Start of emitted-code area */
128 regnode *emit_bound; /* First regnode outside of the allocated space */
129 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
130 implies compiling, so don't emit */
131 regnode emit_dummy; /* placeholder for emit to point to */
132 I32 naughty; /* How bad is this pattern? */
133 I32 sawback; /* Did we see \1, ...? */
135 SSize_t size; /* Code size. */
136 I32 npar; /* Capture buffer count, (OPEN). */
137 I32 cpar; /* Capture buffer count, (CLOSE). */
138 I32 nestroot; /* root parens we are in - used by accept */
141 regnode **open_parens; /* pointers to open parens */
142 regnode **close_parens; /* pointers to close parens */
143 regnode *opend; /* END node in program */
144 I32 utf8; /* whether the pattern is utf8 or not */
145 I32 orig_utf8; /* whether the pattern was originally in utf8 */
146 /* XXX use this for future optimisation of case
147 * where pattern must be upgraded to utf8. */
148 I32 uni_semantics; /* If a d charset modifier should use unicode
149 rules, even if the pattern is not in
151 HV *paren_names; /* Paren names */
153 regnode **recurse; /* Recurse regops */
154 I32 recurse_count; /* Number of recurse regops */
157 I32 override_recoding;
158 I32 in_multi_char_class;
159 struct reg_code_block *code_blocks; /* positions of literal (?{})
161 int num_code_blocks; /* size of code_blocks[] */
162 int code_index; /* next code_blocks[] slot */
164 char *starttry; /* -Dr: where regtry was called. */
165 #define RExC_starttry (pRExC_state->starttry)
167 SV *runtime_code_qr; /* qr with the runtime code blocks */
169 const char *lastparse;
171 AV *paren_name_list; /* idx -> name */
172 #define RExC_lastparse (pRExC_state->lastparse)
173 #define RExC_lastnum (pRExC_state->lastnum)
174 #define RExC_paren_name_list (pRExC_state->paren_name_list)
178 #define RExC_flags (pRExC_state->flags)
179 #define RExC_pm_flags (pRExC_state->pm_flags)
180 #define RExC_precomp (pRExC_state->precomp)
181 #define RExC_rx_sv (pRExC_state->rx_sv)
182 #define RExC_rx (pRExC_state->rx)
183 #define RExC_rxi (pRExC_state->rxi)
184 #define RExC_start (pRExC_state->start)
185 #define RExC_end (pRExC_state->end)
186 #define RExC_parse (pRExC_state->parse)
187 #define RExC_whilem_seen (pRExC_state->whilem_seen)
188 #ifdef RE_TRACK_PATTERN_OFFSETS
189 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
191 #define RExC_emit (pRExC_state->emit)
192 #define RExC_emit_dummy (pRExC_state->emit_dummy)
193 #define RExC_emit_start (pRExC_state->emit_start)
194 #define RExC_emit_bound (pRExC_state->emit_bound)
195 #define RExC_naughty (pRExC_state->naughty)
196 #define RExC_sawback (pRExC_state->sawback)
197 #define RExC_seen (pRExC_state->seen)
198 #define RExC_size (pRExC_state->size)
199 #define RExC_npar (pRExC_state->npar)
200 #define RExC_nestroot (pRExC_state->nestroot)
201 #define RExC_extralen (pRExC_state->extralen)
202 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
203 #define RExC_utf8 (pRExC_state->utf8)
204 #define RExC_uni_semantics (pRExC_state->uni_semantics)
205 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
206 #define RExC_open_parens (pRExC_state->open_parens)
207 #define RExC_close_parens (pRExC_state->close_parens)
208 #define RExC_opend (pRExC_state->opend)
209 #define RExC_paren_names (pRExC_state->paren_names)
210 #define RExC_recurse (pRExC_state->recurse)
211 #define RExC_recurse_count (pRExC_state->recurse_count)
212 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
213 #define RExC_contains_locale (pRExC_state->contains_locale)
214 #define RExC_override_recoding (pRExC_state->override_recoding)
215 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
218 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
219 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
220 ((*s) == '{' && regcurly(s, FALSE)))
223 #undef SPSTART /* dratted cpp namespace... */
226 * Flags to be passed up and down.
228 #define WORST 0 /* Worst case. */
229 #define HASWIDTH 0x01 /* Known to match non-null strings. */
231 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
232 * character. (There needs to be a case: in the switch statement in regexec.c
233 * for any node marked SIMPLE.) Note that this is not the same thing as
236 #define SPSTART 0x04 /* Starts with * or + */
237 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
238 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
239 #define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
241 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
243 /* whether trie related optimizations are enabled */
244 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
245 #define TRIE_STUDY_OPT
246 #define FULL_TRIE_STUDY
252 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
253 #define PBITVAL(paren) (1 << ((paren) & 7))
254 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
255 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
256 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
258 #define REQUIRE_UTF8 STMT_START { \
260 *flagp = RESTART_UTF8; \
265 /* This converts the named class defined in regcomp.h to its equivalent class
266 * number defined in handy.h. */
267 #define namedclass_to_classnum(class) ((int) ((class) / 2))
268 #define classnum_to_namedclass(classnum) ((classnum) * 2)
270 /* About scan_data_t.
272 During optimisation we recurse through the regexp program performing
273 various inplace (keyhole style) optimisations. In addition study_chunk
274 and scan_commit populate this data structure with information about
275 what strings MUST appear in the pattern. We look for the longest
276 string that must appear at a fixed location, and we look for the
277 longest string that may appear at a floating location. So for instance
282 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
283 strings (because they follow a .* construct). study_chunk will identify
284 both FOO and BAR as being the longest fixed and floating strings respectively.
286 The strings can be composites, for instance
290 will result in a composite fixed substring 'foo'.
292 For each string some basic information is maintained:
294 - offset or min_offset
295 This is the position the string must appear at, or not before.
296 It also implicitly (when combined with minlenp) tells us how many
297 characters must match before the string we are searching for.
298 Likewise when combined with minlenp and the length of the string it
299 tells us how many characters must appear after the string we have
303 Only used for floating strings. This is the rightmost point that
304 the string can appear at. If set to SSize_t_MAX it indicates that the
305 string can occur infinitely far to the right.
308 A pointer to the minimum number of characters of the pattern that the
309 string was found inside. This is important as in the case of positive
310 lookahead or positive lookbehind we can have multiple patterns
315 The minimum length of the pattern overall is 3, the minimum length
316 of the lookahead part is 3, but the minimum length of the part that
317 will actually match is 1. So 'FOO's minimum length is 3, but the
318 minimum length for the F is 1. This is important as the minimum length
319 is used to determine offsets in front of and behind the string being
320 looked for. Since strings can be composites this is the length of the
321 pattern at the time it was committed with a scan_commit. Note that
322 the length is calculated by study_chunk, so that the minimum lengths
323 are not known until the full pattern has been compiled, thus the
324 pointer to the value.
328 In the case of lookbehind the string being searched for can be
329 offset past the start point of the final matching string.
330 If this value was just blithely removed from the min_offset it would
331 invalidate some of the calculations for how many chars must match
332 before or after (as they are derived from min_offset and minlen and
333 the length of the string being searched for).
334 When the final pattern is compiled and the data is moved from the
335 scan_data_t structure into the regexp structure the information
336 about lookbehind is factored in, with the information that would
337 have been lost precalculated in the end_shift field for the
340 The fields pos_min and pos_delta are used to store the minimum offset
341 and the delta to the maximum offset at the current point in the pattern.
345 typedef struct scan_data_t {
346 /*I32 len_min; unused */
347 /*I32 len_delta; unused */
351 SSize_t last_end; /* min value, <0 unless valid. */
352 SSize_t last_start_min;
353 SSize_t last_start_max;
354 SV **longest; /* Either &l_fixed, or &l_float. */
355 SV *longest_fixed; /* longest fixed string found in pattern */
356 SSize_t offset_fixed; /* offset where it starts */
357 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
358 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
359 SV *longest_float; /* longest floating string found in pattern */
360 SSize_t offset_float_min; /* earliest point in string it can appear */
361 SSize_t offset_float_max; /* latest point in string it can appear */
362 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
363 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
366 SSize_t *last_closep;
367 struct regnode_charclass_class *start_class;
370 /* The below is perhaps overboard, but this allows us to save a test at the
371 * expense of a mask. This is because on both EBCDIC and ASCII machines, 'A'
372 * and 'a' differ by a single bit; the same with the upper and lower case of
373 * all other ASCII-range alphabetics. On ASCII platforms, they are 32 apart;
374 * on EBCDIC, they are 64. This uses an exclusive 'or' to find that bit and
375 * then inverts it to form a mask, with just a single 0, in the bit position
376 * where the upper- and lowercase differ. XXX There are about 40 other
377 * instances in the Perl core where this micro-optimization could be used.
378 * Should decide if maintenance cost is worse, before changing those
380 * Returns a boolean as to whether or not 'v' is either a lowercase or
381 * uppercase instance of 'c', where 'c' is in [A-Za-z]. If 'c' is a
382 * compile-time constant, the generated code is better than some optimizing
383 * compilers figure out, amounting to a mask and test. The results are
384 * meaningless if 'c' is not one of [A-Za-z] */
385 #define isARG2_lower_or_UPPER_ARG1(c, v) \
386 (((v) & ~('A' ^ 'a')) == ((c) & ~('A' ^ 'a')))
389 * Forward declarations for pregcomp()'s friends.
392 static const scan_data_t zero_scan_data =
393 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
395 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
396 #define SF_BEFORE_SEOL 0x0001
397 #define SF_BEFORE_MEOL 0x0002
398 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
399 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
402 # define SF_FIX_SHIFT_EOL (0+2)
403 # define SF_FL_SHIFT_EOL (0+4)
405 # define SF_FIX_SHIFT_EOL (+2)
406 # define SF_FL_SHIFT_EOL (+4)
409 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
410 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
412 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
413 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
414 #define SF_IS_INF 0x0040
415 #define SF_HAS_PAR 0x0080
416 #define SF_IN_PAR 0x0100
417 #define SF_HAS_EVAL 0x0200
418 #define SCF_DO_SUBSTR 0x0400
419 #define SCF_DO_STCLASS_AND 0x0800
420 #define SCF_DO_STCLASS_OR 0x1000
421 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
422 #define SCF_WHILEM_VISITED_POS 0x2000
424 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
425 #define SCF_SEEN_ACCEPT 0x8000
426 #define SCF_TRIE_DOING_RESTUDY 0x10000
428 #define UTF cBOOL(RExC_utf8)
430 /* The enums for all these are ordered so things work out correctly */
431 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
432 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
433 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
434 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
435 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
436 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
437 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
439 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
441 #define OOB_NAMEDCLASS -1
443 /* There is no code point that is out-of-bounds, so this is problematic. But
444 * its only current use is to initialize a variable that is always set before
446 #define OOB_UNICODE 0xDEADBEEF
448 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
449 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
452 /* length of regex to show in messages that don't mark a position within */
453 #define RegexLengthToShowInErrorMessages 127
456 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
457 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
458 * op/pragma/warn/regcomp.
460 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
461 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
463 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
465 #define REPORT_LOCATION_ARGS(offset) \
466 UTF8fARG(UTF, offset, RExC_precomp), \
467 UTF8fARG(UTF, RExC_end - RExC_precomp - offset, RExC_precomp + offset)
470 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
471 * arg. Show regex, up to a maximum length. If it's too long, chop and add
474 #define _FAIL(code) STMT_START { \
475 const char *ellipses = ""; \
476 IV len = RExC_end - RExC_precomp; \
479 SAVEFREESV(RExC_rx_sv); \
480 if (len > RegexLengthToShowInErrorMessages) { \
481 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
482 len = RegexLengthToShowInErrorMessages - 10; \
488 #define FAIL(msg) _FAIL( \
489 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
490 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
492 #define FAIL2(msg,arg) _FAIL( \
493 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
494 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
497 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
499 #define Simple_vFAIL(m) STMT_START { \
500 const IV offset = RExC_parse - RExC_precomp; \
501 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
502 m, REPORT_LOCATION_ARGS(offset)); \
506 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
508 #define vFAIL(m) STMT_START { \
510 SAVEFREESV(RExC_rx_sv); \
515 * Like Simple_vFAIL(), but accepts two arguments.
517 #define Simple_vFAIL2(m,a1) STMT_START { \
518 const IV offset = RExC_parse - RExC_precomp; \
519 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
520 REPORT_LOCATION_ARGS(offset)); \
524 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
526 #define vFAIL2(m,a1) STMT_START { \
528 SAVEFREESV(RExC_rx_sv); \
529 Simple_vFAIL2(m, a1); \
534 * Like Simple_vFAIL(), but accepts three arguments.
536 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
537 const IV offset = RExC_parse - RExC_precomp; \
538 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
539 REPORT_LOCATION_ARGS(offset)); \
543 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
545 #define vFAIL3(m,a1,a2) STMT_START { \
547 SAVEFREESV(RExC_rx_sv); \
548 Simple_vFAIL3(m, a1, a2); \
552 * Like Simple_vFAIL(), but accepts four arguments.
554 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
555 const IV offset = RExC_parse - RExC_precomp; \
556 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
557 REPORT_LOCATION_ARGS(offset)); \
560 #define vFAIL4(m,a1,a2,a3) STMT_START { \
562 SAVEFREESV(RExC_rx_sv); \
563 Simple_vFAIL4(m, a1, a2, a3); \
566 /* A specialized version of vFAIL2 that works with UTF8f */
567 #define vFAIL2utf8f(m, a1) STMT_START { \
569 SAVEFREESV(RExC_rx_sv); \
570 Simple_vFAIL4(m, a1); \
574 /* m is not necessarily a "literal string", in this macro */
575 #define reg_warn_non_literal_string(loc, m) STMT_START { \
576 const IV offset = loc - RExC_precomp; \
577 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
578 m, REPORT_LOCATION_ARGS(offset)); \
581 #define ckWARNreg(loc,m) STMT_START { \
582 const IV offset = loc - RExC_precomp; \
583 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
584 REPORT_LOCATION_ARGS(offset)); \
587 #define vWARN_dep(loc, m) STMT_START { \
588 const IV offset = loc - RExC_precomp; \
589 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
590 REPORT_LOCATION_ARGS(offset)); \
593 #define ckWARNdep(loc,m) STMT_START { \
594 const IV offset = loc - RExC_precomp; \
595 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
597 REPORT_LOCATION_ARGS(offset)); \
600 #define ckWARNregdep(loc,m) STMT_START { \
601 const IV offset = loc - RExC_precomp; \
602 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
604 REPORT_LOCATION_ARGS(offset)); \
607 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
608 const IV offset = loc - RExC_precomp; \
609 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
611 a1, REPORT_LOCATION_ARGS(offset)); \
614 #define ckWARN2reg(loc, m, a1) STMT_START { \
615 const IV offset = loc - RExC_precomp; \
616 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
617 a1, REPORT_LOCATION_ARGS(offset)); \
620 #define vWARN3(loc, m, a1, a2) STMT_START { \
621 const IV offset = loc - RExC_precomp; \
622 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
623 a1, a2, REPORT_LOCATION_ARGS(offset)); \
626 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
627 const IV offset = loc - RExC_precomp; \
628 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
629 a1, a2, REPORT_LOCATION_ARGS(offset)); \
632 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
633 const IV offset = loc - RExC_precomp; \
634 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
635 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
638 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
639 const IV offset = loc - RExC_precomp; \
640 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
641 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
644 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
645 const IV offset = loc - RExC_precomp; \
646 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
647 a1, a2, a3, a4, REPORT_LOCATION_ARGS(offset)); \
651 /* Allow for side effects in s */
652 #define REGC(c,s) STMT_START { \
653 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
656 /* Macros for recording node offsets. 20001227 mjd@plover.com
657 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
658 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
659 * Element 0 holds the number n.
660 * Position is 1 indexed.
662 #ifndef RE_TRACK_PATTERN_OFFSETS
663 #define Set_Node_Offset_To_R(node,byte)
664 #define Set_Node_Offset(node,byte)
665 #define Set_Cur_Node_Offset
666 #define Set_Node_Length_To_R(node,len)
667 #define Set_Node_Length(node,len)
668 #define Set_Node_Cur_Length(node,start)
669 #define Node_Offset(n)
670 #define Node_Length(n)
671 #define Set_Node_Offset_Length(node,offset,len)
672 #define ProgLen(ri) ri->u.proglen
673 #define SetProgLen(ri,x) ri->u.proglen = x
675 #define ProgLen(ri) ri->u.offsets[0]
676 #define SetProgLen(ri,x) ri->u.offsets[0] = x
677 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
679 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
680 __LINE__, (int)(node), (int)(byte))); \
682 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
684 RExC_offsets[2*(node)-1] = (byte); \
689 #define Set_Node_Offset(node,byte) \
690 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
691 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
693 #define Set_Node_Length_To_R(node,len) STMT_START { \
695 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
696 __LINE__, (int)(node), (int)(len))); \
698 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
700 RExC_offsets[2*(node)] = (len); \
705 #define Set_Node_Length(node,len) \
706 Set_Node_Length_To_R((node)-RExC_emit_start, len)
707 #define Set_Node_Cur_Length(node, start) \
708 Set_Node_Length(node, RExC_parse - start)
710 /* Get offsets and lengths */
711 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
712 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
714 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
715 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
716 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
720 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
721 #define EXPERIMENTAL_INPLACESCAN
722 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
724 #define DEBUG_STUDYDATA(str,data,depth) \
725 DEBUG_OPTIMISE_MORE_r(if(data){ \
726 PerlIO_printf(Perl_debug_log, \
727 "%*s" str "Pos:%"IVdf"/%"IVdf \
728 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
729 (int)(depth)*2, "", \
730 (IV)((data)->pos_min), \
731 (IV)((data)->pos_delta), \
732 (UV)((data)->flags), \
733 (IV)((data)->whilem_c), \
734 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
735 is_inf ? "INF " : "" \
737 if ((data)->last_found) \
738 PerlIO_printf(Perl_debug_log, \
739 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
740 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
741 SvPVX_const((data)->last_found), \
742 (IV)((data)->last_end), \
743 (IV)((data)->last_start_min), \
744 (IV)((data)->last_start_max), \
745 ((data)->longest && \
746 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
747 SvPVX_const((data)->longest_fixed), \
748 (IV)((data)->offset_fixed), \
749 ((data)->longest && \
750 (data)->longest==&((data)->longest_float)) ? "*" : "", \
751 SvPVX_const((data)->longest_float), \
752 (IV)((data)->offset_float_min), \
753 (IV)((data)->offset_float_max) \
755 PerlIO_printf(Perl_debug_log,"\n"); \
758 /* Mark that we cannot extend a found fixed substring at this point.
759 Update the longest found anchored substring and the longest found
760 floating substrings if needed. */
763 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
764 SSize_t *minlenp, int is_inf)
766 const STRLEN l = CHR_SVLEN(data->last_found);
767 const STRLEN old_l = CHR_SVLEN(*data->longest);
768 GET_RE_DEBUG_FLAGS_DECL;
770 PERL_ARGS_ASSERT_SCAN_COMMIT;
772 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
773 SvSetMagicSV(*data->longest, data->last_found);
774 if (*data->longest == data->longest_fixed) {
775 data->offset_fixed = l ? data->last_start_min : data->pos_min;
776 if (data->flags & SF_BEFORE_EOL)
778 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
780 data->flags &= ~SF_FIX_BEFORE_EOL;
781 data->minlen_fixed=minlenp;
782 data->lookbehind_fixed=0;
784 else { /* *data->longest == data->longest_float */
785 data->offset_float_min = l ? data->last_start_min : data->pos_min;
786 data->offset_float_max = (l
787 ? data->last_start_max
788 : (data->pos_delta == SSize_t_MAX
790 : data->pos_min + data->pos_delta));
792 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
793 data->offset_float_max = SSize_t_MAX;
794 if (data->flags & SF_BEFORE_EOL)
796 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
798 data->flags &= ~SF_FL_BEFORE_EOL;
799 data->minlen_float=minlenp;
800 data->lookbehind_float=0;
803 SvCUR_set(data->last_found, 0);
805 SV * const sv = data->last_found;
806 if (SvUTF8(sv) && SvMAGICAL(sv)) {
807 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
813 data->flags &= ~SF_BEFORE_EOL;
814 DEBUG_STUDYDATA("commit: ",data,0);
817 /* These macros set, clear and test whether the synthetic start class ('ssc',
818 * given by the parameter) matches an empty string (EOS). This uses the
819 * 'next_off' field in the node, to save a bit in the flags field. The ssc
820 * stands alone, so there is never a next_off, so this field is otherwise
821 * unused. The EOS information is used only for compilation, but theoretically
822 * it could be passed on to the execution code. This could be used to store
823 * more than one bit of information, but only this one is currently used. */
824 #define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
825 #define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
826 #define TEST_SSC_EOS(node) cBOOL((node)->next_off)
828 /* Can match anything (initialization) */
830 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
832 PERL_ARGS_ASSERT_CL_ANYTHING;
834 ANYOF_BITMAP_SETALL(cl);
835 cl->flags = ANYOF_UNICODE_ALL;
838 /* If any portion of the regex is to operate under locale rules,
839 * initialization includes it. The reason this isn't done for all regexes
840 * is that the optimizer was written under the assumption that locale was
841 * all-or-nothing. Given the complexity and lack of documentation in the
842 * optimizer, and that there are inadequate test cases for locale, so many
843 * parts of it may not work properly, it is safest to avoid locale unless
845 if (RExC_contains_locale) {
846 ANYOF_CLASS_SETALL(cl); /* /l uses class */
847 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
850 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
854 /* Can match anything (initialization) */
856 S_cl_is_anything(const struct regnode_charclass_class *cl)
860 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
862 for (value = 0; value < ANYOF_MAX; value += 2)
863 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
865 if (!(cl->flags & ANYOF_UNICODE_ALL))
867 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
872 /* Can match anything (initialization) */
874 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
876 PERL_ARGS_ASSERT_CL_INIT;
878 Zero(cl, 1, struct regnode_charclass_class);
880 cl_anything(pRExC_state, cl);
881 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
884 /* These two functions currently do the exact same thing */
885 #define cl_init_zero cl_init
887 /* 'AND' a given class with another one. Can create false positives. 'cl'
888 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
889 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
891 S_cl_and(struct regnode_charclass_class *cl,
892 const struct regnode_charclass_class *and_with)
894 PERL_ARGS_ASSERT_CL_AND;
896 assert(PL_regkind[and_with->type] == ANYOF);
898 /* I (khw) am not sure all these restrictions are necessary XXX */
899 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
900 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
901 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
902 && !(and_with->flags & ANYOF_LOC_FOLD)
903 && !(cl->flags & ANYOF_LOC_FOLD)) {
906 if (and_with->flags & ANYOF_INVERT)
907 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
908 cl->bitmap[i] &= ~and_with->bitmap[i];
910 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
911 cl->bitmap[i] &= and_with->bitmap[i];
912 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
914 if (and_with->flags & ANYOF_INVERT) {
916 /* Here, the and'ed node is inverted. Get the AND of the flags that
917 * aren't affected by the inversion. Those that are affected are
918 * handled individually below */
919 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
920 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
921 cl->flags |= affected_flags;
923 /* We currently don't know how to deal with things that aren't in the
924 * bitmap, but we know that the intersection is no greater than what
925 * is already in cl, so let there be false positives that get sorted
926 * out after the synthetic start class succeeds, and the node is
927 * matched for real. */
929 /* The inversion of these two flags indicate that the resulting
930 * intersection doesn't have them */
931 if (and_with->flags & ANYOF_UNICODE_ALL) {
932 cl->flags &= ~ANYOF_UNICODE_ALL;
934 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
935 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
938 else { /* and'd node is not inverted */
939 U8 outside_bitmap_but_not_utf8; /* Temp variable */
941 if (! ANYOF_NONBITMAP(and_with)) {
943 /* Here 'and_with' doesn't match anything outside the bitmap
944 * (except possibly ANYOF_UNICODE_ALL), which means the
945 * intersection can't either, except for ANYOF_UNICODE_ALL, in
946 * which case we don't know what the intersection is, but it's no
947 * greater than what cl already has, so can just leave it alone,
948 * with possible false positives */
949 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
950 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
951 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
954 else if (! ANYOF_NONBITMAP(cl)) {
956 /* Here, 'and_with' does match something outside the bitmap, and cl
957 * doesn't have a list of things to match outside the bitmap. If
958 * cl can match all code points above 255, the intersection will
959 * be those above-255 code points that 'and_with' matches. If cl
960 * can't match all Unicode code points, it means that it can't
961 * match anything outside the bitmap (since the 'if' that got us
962 * into this block tested for that), so we leave the bitmap empty.
964 if (cl->flags & ANYOF_UNICODE_ALL) {
965 ARG_SET(cl, ARG(and_with));
967 /* and_with's ARG may match things that don't require UTF8.
968 * And now cl's will too, in spite of this being an 'and'. See
969 * the comments below about the kludge */
970 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
974 /* Here, both 'and_with' and cl match something outside the
975 * bitmap. Currently we do not do the intersection, so just match
976 * whatever cl had at the beginning. */
980 /* Take the intersection of the two sets of flags. However, the
981 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
982 * kludge around the fact that this flag is not treated like the others
983 * which are initialized in cl_anything(). The way the optimizer works
984 * is that the synthetic start class (SSC) is initialized to match
985 * anything, and then the first time a real node is encountered, its
986 * values are AND'd with the SSC's with the result being the values of
987 * the real node. However, there are paths through the optimizer where
988 * the AND never gets called, so those initialized bits are set
989 * inappropriately, which is not usually a big deal, as they just cause
990 * false positives in the SSC, which will just mean a probably
991 * imperceptible slow down in execution. However this bit has a
992 * higher false positive consequence in that it can cause utf8.pm,
993 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
994 * bigger slowdown and also causes significant extra memory to be used.
995 * In order to prevent this, the code now takes a different tack. The
996 * bit isn't set unless some part of the regular expression needs it,
997 * but once set it won't get cleared. This means that these extra
998 * modules won't get loaded unless there was some path through the
999 * pattern that would have required them anyway, and so any false
1000 * positives that occur by not ANDing them out when they could be
1001 * aren't as severe as they would be if we treated this bit like all
1003 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
1004 & ANYOF_NONBITMAP_NON_UTF8;
1005 cl->flags &= and_with->flags;
1006 cl->flags |= outside_bitmap_but_not_utf8;
1010 /* 'OR' a given class with another one. Can create false positives. 'cl'
1011 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
1012 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
1014 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
1016 PERL_ARGS_ASSERT_CL_OR;
1018 if (or_with->flags & ANYOF_INVERT) {
1020 /* Here, the or'd node is to be inverted. This means we take the
1021 * complement of everything not in the bitmap, but currently we don't
1022 * know what that is, so give up and match anything */
1023 if (ANYOF_NONBITMAP(or_with)) {
1024 cl_anything(pRExC_state, cl);
1027 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
1028 * <= (B1 | !B2) | (CL1 | !CL2)
1029 * which is wasteful if CL2 is small, but we ignore CL2:
1030 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
1031 * XXXX Can we handle case-fold? Unclear:
1032 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
1033 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
1035 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1036 && !(or_with->flags & ANYOF_LOC_FOLD)
1037 && !(cl->flags & ANYOF_LOC_FOLD) ) {
1040 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1041 cl->bitmap[i] |= ~or_with->bitmap[i];
1042 } /* XXXX: logic is complicated otherwise */
1044 cl_anything(pRExC_state, cl);
1047 /* And, we can just take the union of the flags that aren't affected
1048 * by the inversion */
1049 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
1051 /* For the remaining flags:
1052 ANYOF_UNICODE_ALL and inverted means to not match anything above
1053 255, which means that the union with cl should just be
1054 what cl has in it, so can ignore this flag
1055 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
1056 is (ASCII) 127-255 to match them, but then invert that, so
1057 the union with cl should just be what cl has in it, so can
1060 } else { /* 'or_with' is not inverted */
1061 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1062 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1063 && (!(or_with->flags & ANYOF_LOC_FOLD)
1064 || (cl->flags & ANYOF_LOC_FOLD)) ) {
1067 /* OR char bitmap and class bitmap separately */
1068 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1069 cl->bitmap[i] |= or_with->bitmap[i];
1070 if (or_with->flags & ANYOF_CLASS) {
1071 ANYOF_CLASS_OR(or_with, cl);
1074 else { /* XXXX: logic is complicated, leave it along for a moment. */
1075 cl_anything(pRExC_state, cl);
1078 if (ANYOF_NONBITMAP(or_with)) {
1080 /* Use the added node's outside-the-bit-map match if there isn't a
1081 * conflict. If there is a conflict (both nodes match something
1082 * outside the bitmap, but what they match outside is not the same
1083 * pointer, and hence not easily compared until XXX we extend
1084 * inversion lists this far), give up and allow the start class to
1085 * match everything outside the bitmap. If that stuff is all above
1086 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1087 if (! ANYOF_NONBITMAP(cl)) {
1088 ARG_SET(cl, ARG(or_with));
1090 else if (ARG(cl) != ARG(or_with)) {
1092 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1093 cl_anything(pRExC_state, cl);
1096 cl->flags |= ANYOF_UNICODE_ALL;
1101 /* Take the union */
1102 cl->flags |= or_with->flags;
1106 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1107 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1108 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1109 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1114 dump_trie(trie,widecharmap,revcharmap)
1115 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1116 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1118 These routines dump out a trie in a somewhat readable format.
1119 The _interim_ variants are used for debugging the interim
1120 tables that are used to generate the final compressed
1121 representation which is what dump_trie expects.
1123 Part of the reason for their existence is to provide a form
1124 of documentation as to how the different representations function.
1129 Dumps the final compressed table form of the trie to Perl_debug_log.
1130 Used for debugging make_trie().
1134 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1135 AV *revcharmap, U32 depth)
1138 SV *sv=sv_newmortal();
1139 int colwidth= widecharmap ? 6 : 4;
1141 GET_RE_DEBUG_FLAGS_DECL;
1143 PERL_ARGS_ASSERT_DUMP_TRIE;
1145 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1146 (int)depth * 2 + 2,"",
1147 "Match","Base","Ofs" );
1149 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1150 SV ** const tmp = av_fetch( revcharmap, state, 0);
1152 PerlIO_printf( Perl_debug_log, "%*s",
1154 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1155 PL_colors[0], PL_colors[1],
1156 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1157 PERL_PV_ESCAPE_FIRSTCHAR
1162 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1163 (int)depth * 2 + 2,"");
1165 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1166 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1167 PerlIO_printf( Perl_debug_log, "\n");
1169 for( state = 1 ; state < trie->statecount ; state++ ) {
1170 const U32 base = trie->states[ state ].trans.base;
1172 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1174 if ( trie->states[ state ].wordnum ) {
1175 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1177 PerlIO_printf( Perl_debug_log, "%6s", "" );
1180 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1185 while( ( base + ofs < trie->uniquecharcount ) ||
1186 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1187 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1190 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1192 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1193 if ( ( base + ofs >= trie->uniquecharcount ) &&
1194 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1195 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1197 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1199 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1201 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1205 PerlIO_printf( Perl_debug_log, "]");
1208 PerlIO_printf( Perl_debug_log, "\n" );
1210 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1211 for (word=1; word <= trie->wordcount; word++) {
1212 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1213 (int)word, (int)(trie->wordinfo[word].prev),
1214 (int)(trie->wordinfo[word].len));
1216 PerlIO_printf(Perl_debug_log, "\n" );
1219 Dumps a fully constructed but uncompressed trie in list form.
1220 List tries normally only are used for construction when the number of
1221 possible chars (trie->uniquecharcount) is very high.
1222 Used for debugging make_trie().
1225 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1226 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1230 SV *sv=sv_newmortal();
1231 int colwidth= widecharmap ? 6 : 4;
1232 GET_RE_DEBUG_FLAGS_DECL;
1234 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1236 /* print out the table precompression. */
1237 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1238 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1239 "------:-----+-----------------\n" );
1241 for( state=1 ; state < next_alloc ; state ++ ) {
1244 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1245 (int)depth * 2 + 2,"", (UV)state );
1246 if ( ! trie->states[ state ].wordnum ) {
1247 PerlIO_printf( Perl_debug_log, "%5s| ","");
1249 PerlIO_printf( Perl_debug_log, "W%4x| ",
1250 trie->states[ state ].wordnum
1253 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1254 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1256 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1258 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1259 PL_colors[0], PL_colors[1],
1260 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1261 PERL_PV_ESCAPE_FIRSTCHAR
1263 TRIE_LIST_ITEM(state,charid).forid,
1264 (UV)TRIE_LIST_ITEM(state,charid).newstate
1267 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1268 (int)((depth * 2) + 14), "");
1271 PerlIO_printf( Perl_debug_log, "\n");
1276 Dumps a fully constructed but uncompressed trie in table form.
1277 This is the normal DFA style state transition table, with a few
1278 twists to facilitate compression later.
1279 Used for debugging make_trie().
1282 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1283 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1288 SV *sv=sv_newmortal();
1289 int colwidth= widecharmap ? 6 : 4;
1290 GET_RE_DEBUG_FLAGS_DECL;
1292 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1295 print out the table precompression so that we can do a visual check
1296 that they are identical.
1299 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1301 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1302 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1304 PerlIO_printf( Perl_debug_log, "%*s",
1306 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1307 PL_colors[0], PL_colors[1],
1308 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1309 PERL_PV_ESCAPE_FIRSTCHAR
1315 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1317 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1318 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1321 PerlIO_printf( Perl_debug_log, "\n" );
1323 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1325 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1326 (int)depth * 2 + 2,"",
1327 (UV)TRIE_NODENUM( state ) );
1329 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1330 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1332 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1334 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1336 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1337 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1339 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1340 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1348 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1349 startbranch: the first branch in the whole branch sequence
1350 first : start branch of sequence of branch-exact nodes.
1351 May be the same as startbranch
1352 last : Thing following the last branch.
1353 May be the same as tail.
1354 tail : item following the branch sequence
1355 count : words in the sequence
1356 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1357 depth : indent depth
1359 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1361 A trie is an N'ary tree where the branches are determined by digital
1362 decomposition of the key. IE, at the root node you look up the 1st character and
1363 follow that branch repeat until you find the end of the branches. Nodes can be
1364 marked as "accepting" meaning they represent a complete word. Eg:
1368 would convert into the following structure. Numbers represent states, letters
1369 following numbers represent valid transitions on the letter from that state, if
1370 the number is in square brackets it represents an accepting state, otherwise it
1371 will be in parenthesis.
1373 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1377 (1) +-i->(6)-+-s->[7]
1379 +-s->(3)-+-h->(4)-+-e->[5]
1381 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1383 This shows that when matching against the string 'hers' we will begin at state 1
1384 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1385 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1386 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1387 single traverse. We store a mapping from accepting to state to which word was
1388 matched, and then when we have multiple possibilities we try to complete the
1389 rest of the regex in the order in which they occured in the alternation.
1391 The only prior NFA like behaviour that would be changed by the TRIE support is
1392 the silent ignoring of duplicate alternations which are of the form:
1394 / (DUPE|DUPE) X? (?{ ... }) Y /x
1396 Thus EVAL blocks following a trie may be called a different number of times with
1397 and without the optimisation. With the optimisations dupes will be silently
1398 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1399 the following demonstrates:
1401 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1403 which prints out 'word' three times, but
1405 'words'=~/(word|word|word)(?{ print $1 })S/
1407 which doesnt print it out at all. This is due to other optimisations kicking in.
1409 Example of what happens on a structural level:
1411 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1413 1: CURLYM[1] {1,32767}(18)
1424 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1425 and should turn into:
1427 1: CURLYM[1] {1,32767}(18)
1429 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1437 Cases where tail != last would be like /(?foo|bar)baz/:
1447 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1448 and would end up looking like:
1451 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1458 d = uvchr_to_utf8_flags(d, uv, 0);
1460 is the recommended Unicode-aware way of saying
1465 #define TRIE_STORE_REVCHAR(val) \
1468 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1469 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1470 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
1471 SvCUR_set(zlopp, kapow - flrbbbbb); \
1474 av_push(revcharmap, zlopp); \
1476 char ooooff = (char)val; \
1477 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1481 /* This gets the next character from the input, folding it if not already
1483 #define TRIE_READ_CHAR STMT_START { \
1486 /* if it is UTF then it is either already folded, or does not need \
1488 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
1490 else if (folder == PL_fold_latin1) { \
1491 /* This folder implies Unicode rules, which in the range expressible \
1492 * by not UTF is the lower case, with the two exceptions, one of \
1493 * which should have been taken care of before calling this */ \
1494 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
1495 uvc = toLOWER_L1(*uc); \
1496 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
1499 /* raw data, will be folded later if needed */ \
1507 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1508 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1509 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1510 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1512 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1513 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1514 TRIE_LIST_CUR( state )++; \
1517 #define TRIE_LIST_NEW(state) STMT_START { \
1518 Newxz( trie->states[ state ].trans.list, \
1519 4, reg_trie_trans_le ); \
1520 TRIE_LIST_CUR( state ) = 1; \
1521 TRIE_LIST_LEN( state ) = 4; \
1524 #define TRIE_HANDLE_WORD(state) STMT_START { \
1525 U16 dupe= trie->states[ state ].wordnum; \
1526 regnode * const noper_next = regnext( noper ); \
1529 /* store the word for dumping */ \
1531 if (OP(noper) != NOTHING) \
1532 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1534 tmp = newSVpvn_utf8( "", 0, UTF ); \
1535 av_push( trie_words, tmp ); \
1539 trie->wordinfo[curword].prev = 0; \
1540 trie->wordinfo[curword].len = wordlen; \
1541 trie->wordinfo[curword].accept = state; \
1543 if ( noper_next < tail ) { \
1545 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1546 trie->jump[curword] = (U16)(noper_next - convert); \
1548 jumper = noper_next; \
1550 nextbranch= regnext(cur); \
1554 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1555 /* chain, so that when the bits of chain are later */\
1556 /* linked together, the dups appear in the chain */\
1557 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1558 trie->wordinfo[dupe].prev = curword; \
1560 /* we haven't inserted this word yet. */ \
1561 trie->states[ state ].wordnum = curword; \
1566 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1567 ( ( base + charid >= ucharcount \
1568 && base + charid < ubound \
1569 && state == trie->trans[ base - ucharcount + charid ].check \
1570 && trie->trans[ base - ucharcount + charid ].next ) \
1571 ? trie->trans[ base - ucharcount + charid ].next \
1572 : ( state==1 ? special : 0 ) \
1576 #define MADE_JUMP_TRIE 2
1577 #define MADE_EXACT_TRIE 4
1580 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1583 /* first pass, loop through and scan words */
1584 reg_trie_data *trie;
1585 HV *widecharmap = NULL;
1586 AV *revcharmap = newAV();
1592 regnode *jumper = NULL;
1593 regnode *nextbranch = NULL;
1594 regnode *convert = NULL;
1595 U32 *prev_states; /* temp array mapping each state to previous one */
1596 /* we just use folder as a flag in utf8 */
1597 const U8 * folder = NULL;
1600 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1601 AV *trie_words = NULL;
1602 /* along with revcharmap, this only used during construction but both are
1603 * useful during debugging so we store them in the struct when debugging.
1606 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1607 STRLEN trie_charcount=0;
1609 SV *re_trie_maxbuff;
1610 GET_RE_DEBUG_FLAGS_DECL;
1612 PERL_ARGS_ASSERT_MAKE_TRIE;
1614 PERL_UNUSED_ARG(depth);
1621 case EXACTFU: folder = PL_fold_latin1; break;
1622 case EXACTF: folder = PL_fold; break;
1623 case EXACTFL: folder = PL_fold_locale; break;
1624 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1627 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1629 trie->startstate = 1;
1630 trie->wordcount = word_count;
1631 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1632 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1634 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1635 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1636 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1639 trie_words = newAV();
1642 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1643 if (!SvIOK(re_trie_maxbuff)) {
1644 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1646 DEBUG_TRIE_COMPILE_r({
1647 PerlIO_printf( Perl_debug_log,
1648 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1649 (int)depth * 2 + 2, "",
1650 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1651 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1655 /* Find the node we are going to overwrite */
1656 if ( first == startbranch && OP( last ) != BRANCH ) {
1657 /* whole branch chain */
1660 /* branch sub-chain */
1661 convert = NEXTOPER( first );
1664 /* -- First loop and Setup --
1666 We first traverse the branches and scan each word to determine if it
1667 contains widechars, and how many unique chars there are, this is
1668 important as we have to build a table with at least as many columns as we
1671 We use an array of integers to represent the character codes 0..255
1672 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1673 native representation of the character value as the key and IV's for the
1676 *TODO* If we keep track of how many times each character is used we can
1677 remap the columns so that the table compression later on is more
1678 efficient in terms of memory by ensuring the most common value is in the
1679 middle and the least common are on the outside. IMO this would be better
1680 than a most to least common mapping as theres a decent chance the most
1681 common letter will share a node with the least common, meaning the node
1682 will not be compressible. With a middle is most common approach the worst
1683 case is when we have the least common nodes twice.
1687 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1688 regnode *noper = NEXTOPER( cur );
1689 const U8 *uc = (U8*)STRING( noper );
1690 const U8 *e = uc + STR_LEN( noper );
1692 U32 wordlen = 0; /* required init */
1693 STRLEN minbytes = 0;
1694 STRLEN maxbytes = 0;
1695 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1697 if (OP(noper) == NOTHING) {
1698 regnode *noper_next= regnext(noper);
1699 if (noper_next != tail && OP(noper_next) == flags) {
1701 uc= (U8*)STRING(noper);
1702 e= uc + STR_LEN(noper);
1703 trie->minlen= STR_LEN(noper);
1710 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1711 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1712 regardless of encoding */
1713 if (OP( noper ) == EXACTFU_SS) {
1714 /* false positives are ok, so just set this */
1715 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
1718 for ( ; uc < e ; uc += len ) {
1719 TRIE_CHARCOUNT(trie)++;
1722 /* Acummulate to the current values, the range in the number of
1723 * bytes that this character could match. The max is presumed to
1724 * be the same as the folded input (which TRIE_READ_CHAR returns),
1725 * except that when this is not in UTF-8, it could be matched
1726 * against a string which is UTF-8, and the variant characters
1727 * could be 2 bytes instead of the 1 here. Likewise, for the
1728 * minimum number of bytes when not folded. When folding, the min
1729 * is assumed to be 1 byte could fold to match the single character
1730 * here, or in the case of a multi-char fold, 1 byte can fold to
1731 * the whole sequence. 'foldlen' is used to denote whether we are
1732 * in such a sequence, skipping the min setting if so. XXX TODO
1733 * Use the exact list of what folds to each character, from
1734 * PL_utf8_foldclosures */
1736 maxbytes += UTF8SKIP(uc);
1738 /* A non-UTF-8 string could be 1 byte to match our 2 */
1739 minbytes += (UTF8_IS_DOWNGRADEABLE_START(*uc))
1745 foldlen -= UTF8SKIP(uc);
1748 foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e);
1754 maxbytes += (UNI_IS_INVARIANT(*uc))
1765 foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e);
1772 U8 folded= folder[ (U8) uvc ];
1773 if ( !trie->charmap[ folded ] ) {
1774 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1775 TRIE_STORE_REVCHAR( folded );
1778 if ( !trie->charmap[ uvc ] ) {
1779 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1780 TRIE_STORE_REVCHAR( uvc );
1783 /* store the codepoint in the bitmap, and its folded
1785 TRIE_BITMAP_SET(trie, uvc);
1787 /* store the folded codepoint */
1788 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1791 /* store first byte of utf8 representation of
1792 variant codepoints */
1793 if (! NATIVE_IS_INVARIANT(uvc)) {
1794 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1797 set_bit = 0; /* We've done our bit :-) */
1802 widecharmap = newHV();
1804 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1807 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1809 if ( !SvTRUE( *svpp ) ) {
1810 sv_setiv( *svpp, ++trie->uniquecharcount );
1811 TRIE_STORE_REVCHAR(uvc);
1815 if( cur == first ) {
1816 trie->minlen = minbytes;
1817 trie->maxlen = maxbytes;
1818 } else if (minbytes < trie->minlen) {
1819 trie->minlen = minbytes;
1820 } else if (maxbytes > trie->maxlen) {
1821 trie->maxlen = maxbytes;
1823 } /* end first pass */
1824 DEBUG_TRIE_COMPILE_r(
1825 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1826 (int)depth * 2 + 2,"",
1827 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1828 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1829 (int)trie->minlen, (int)trie->maxlen )
1833 We now know what we are dealing with in terms of unique chars and
1834 string sizes so we can calculate how much memory a naive
1835 representation using a flat table will take. If it's over a reasonable
1836 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1837 conservative but potentially much slower representation using an array
1840 At the end we convert both representations into the same compressed
1841 form that will be used in regexec.c for matching with. The latter
1842 is a form that cannot be used to construct with but has memory
1843 properties similar to the list form and access properties similar
1844 to the table form making it both suitable for fast searches and
1845 small enough that its feasable to store for the duration of a program.
1847 See the comment in the code where the compressed table is produced
1848 inplace from the flat tabe representation for an explanation of how
1849 the compression works.
1854 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1857 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1859 Second Pass -- Array Of Lists Representation
1861 Each state will be represented by a list of charid:state records
1862 (reg_trie_trans_le) the first such element holds the CUR and LEN
1863 points of the allocated array. (See defines above).
1865 We build the initial structure using the lists, and then convert
1866 it into the compressed table form which allows faster lookups
1867 (but cant be modified once converted).
1870 STRLEN transcount = 1;
1872 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1873 "%*sCompiling trie using list compiler\n",
1874 (int)depth * 2 + 2, ""));
1876 trie->states = (reg_trie_state *)
1877 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1878 sizeof(reg_trie_state) );
1882 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1884 regnode *noper = NEXTOPER( cur );
1885 U8 *uc = (U8*)STRING( noper );
1886 const U8 *e = uc + STR_LEN( noper );
1887 U32 state = 1; /* required init */
1888 U16 charid = 0; /* sanity init */
1889 U32 wordlen = 0; /* required init */
1891 if (OP(noper) == NOTHING) {
1892 regnode *noper_next= regnext(noper);
1893 if (noper_next != tail && OP(noper_next) == flags) {
1895 uc= (U8*)STRING(noper);
1896 e= uc + STR_LEN(noper);
1900 if (OP(noper) != NOTHING) {
1901 for ( ; uc < e ; uc += len ) {
1906 charid = trie->charmap[ uvc ];
1908 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1912 charid=(U16)SvIV( *svpp );
1915 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1922 if ( !trie->states[ state ].trans.list ) {
1923 TRIE_LIST_NEW( state );
1925 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1926 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1927 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1932 newstate = next_alloc++;
1933 prev_states[newstate] = state;
1934 TRIE_LIST_PUSH( state, charid, newstate );
1939 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1943 TRIE_HANDLE_WORD(state);
1945 } /* end second pass */
1947 /* next alloc is the NEXT state to be allocated */
1948 trie->statecount = next_alloc;
1949 trie->states = (reg_trie_state *)
1950 PerlMemShared_realloc( trie->states,
1952 * sizeof(reg_trie_state) );
1954 /* and now dump it out before we compress it */
1955 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1956 revcharmap, next_alloc,
1960 trie->trans = (reg_trie_trans *)
1961 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1968 for( state=1 ; state < next_alloc ; state ++ ) {
1972 DEBUG_TRIE_COMPILE_MORE_r(
1973 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1977 if (trie->states[state].trans.list) {
1978 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1982 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1983 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1984 if ( forid < minid ) {
1986 } else if ( forid > maxid ) {
1990 if ( transcount < tp + maxid - minid + 1) {
1992 trie->trans = (reg_trie_trans *)
1993 PerlMemShared_realloc( trie->trans,
1995 * sizeof(reg_trie_trans) );
1996 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1998 base = trie->uniquecharcount + tp - minid;
1999 if ( maxid == minid ) {
2001 for ( ; zp < tp ; zp++ ) {
2002 if ( ! trie->trans[ zp ].next ) {
2003 base = trie->uniquecharcount + zp - minid;
2004 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
2005 trie->trans[ zp ].check = state;
2011 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
2012 trie->trans[ tp ].check = state;
2017 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2018 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
2019 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
2020 trie->trans[ tid ].check = state;
2022 tp += ( maxid - minid + 1 );
2024 Safefree(trie->states[ state ].trans.list);
2027 DEBUG_TRIE_COMPILE_MORE_r(
2028 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2031 trie->states[ state ].trans.base=base;
2033 trie->lasttrans = tp + 1;
2037 Second Pass -- Flat Table Representation.
2039 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
2040 We know that we will need Charcount+1 trans at most to store the data
2041 (one row per char at worst case) So we preallocate both structures
2042 assuming worst case.
2044 We then construct the trie using only the .next slots of the entry
2047 We use the .check field of the first entry of the node temporarily to
2048 make compression both faster and easier by keeping track of how many non
2049 zero fields are in the node.
2051 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2054 There are two terms at use here: state as a TRIE_NODEIDX() which is a
2055 number representing the first entry of the node, and state as a
2056 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
2057 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
2058 are 2 entrys per node. eg:
2066 The table is internally in the right hand, idx form. However as we also
2067 have to deal with the states array which is indexed by nodenum we have to
2068 use TRIE_NODENUM() to convert.
2071 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2072 "%*sCompiling trie using table compiler\n",
2073 (int)depth * 2 + 2, ""));
2075 trie->trans = (reg_trie_trans *)
2076 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2077 * trie->uniquecharcount + 1,
2078 sizeof(reg_trie_trans) );
2079 trie->states = (reg_trie_state *)
2080 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2081 sizeof(reg_trie_state) );
2082 next_alloc = trie->uniquecharcount + 1;
2085 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2087 regnode *noper = NEXTOPER( cur );
2088 const U8 *uc = (U8*)STRING( noper );
2089 const U8 *e = uc + STR_LEN( noper );
2091 U32 state = 1; /* required init */
2093 U16 charid = 0; /* sanity init */
2094 U32 accept_state = 0; /* sanity init */
2096 U32 wordlen = 0; /* required init */
2098 if (OP(noper) == NOTHING) {
2099 regnode *noper_next= regnext(noper);
2100 if (noper_next != tail && OP(noper_next) == flags) {
2102 uc= (U8*)STRING(noper);
2103 e= uc + STR_LEN(noper);
2107 if ( OP(noper) != NOTHING ) {
2108 for ( ; uc < e ; uc += len ) {
2113 charid = trie->charmap[ uvc ];
2115 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2116 charid = svpp ? (U16)SvIV(*svpp) : 0;
2120 if ( !trie->trans[ state + charid ].next ) {
2121 trie->trans[ state + charid ].next = next_alloc;
2122 trie->trans[ state ].check++;
2123 prev_states[TRIE_NODENUM(next_alloc)]
2124 = TRIE_NODENUM(state);
2125 next_alloc += trie->uniquecharcount;
2127 state = trie->trans[ state + charid ].next;
2129 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2131 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2134 accept_state = TRIE_NODENUM( state );
2135 TRIE_HANDLE_WORD(accept_state);
2137 } /* end second pass */
2139 /* and now dump it out before we compress it */
2140 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2142 next_alloc, depth+1));
2146 * Inplace compress the table.*
2148 For sparse data sets the table constructed by the trie algorithm will
2149 be mostly 0/FAIL transitions or to put it another way mostly empty.
2150 (Note that leaf nodes will not contain any transitions.)
2152 This algorithm compresses the tables by eliminating most such
2153 transitions, at the cost of a modest bit of extra work during lookup:
2155 - Each states[] entry contains a .base field which indicates the
2156 index in the state[] array wheres its transition data is stored.
2158 - If .base is 0 there are no valid transitions from that node.
2160 - If .base is nonzero then charid is added to it to find an entry in
2163 -If trans[states[state].base+charid].check!=state then the
2164 transition is taken to be a 0/Fail transition. Thus if there are fail
2165 transitions at the front of the node then the .base offset will point
2166 somewhere inside the previous nodes data (or maybe even into a node
2167 even earlier), but the .check field determines if the transition is
2171 The following process inplace converts the table to the compressed
2172 table: We first do not compress the root node 1,and mark all its
2173 .check pointers as 1 and set its .base pointer as 1 as well. This
2174 allows us to do a DFA construction from the compressed table later,
2175 and ensures that any .base pointers we calculate later are greater
2178 - We set 'pos' to indicate the first entry of the second node.
2180 - We then iterate over the columns of the node, finding the first and
2181 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2182 and set the .check pointers accordingly, and advance pos
2183 appropriately and repreat for the next node. Note that when we copy
2184 the next pointers we have to convert them from the original
2185 NODEIDX form to NODENUM form as the former is not valid post
2188 - If a node has no transitions used we mark its base as 0 and do not
2189 advance the pos pointer.
2191 - If a node only has one transition we use a second pointer into the
2192 structure to fill in allocated fail transitions from other states.
2193 This pointer is independent of the main pointer and scans forward
2194 looking for null transitions that are allocated to a state. When it
2195 finds one it writes the single transition into the "hole". If the
2196 pointer doesnt find one the single transition is appended as normal.
2198 - Once compressed we can Renew/realloc the structures to release the
2201 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2202 specifically Fig 3.47 and the associated pseudocode.
2206 const U32 laststate = TRIE_NODENUM( next_alloc );
2209 trie->statecount = laststate;
2211 for ( state = 1 ; state < laststate ; state++ ) {
2213 const U32 stateidx = TRIE_NODEIDX( state );
2214 const U32 o_used = trie->trans[ stateidx ].check;
2215 U32 used = trie->trans[ stateidx ].check;
2216 trie->trans[ stateidx ].check = 0;
2218 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2219 if ( flag || trie->trans[ stateidx + charid ].next ) {
2220 if ( trie->trans[ stateidx + charid ].next ) {
2222 for ( ; zp < pos ; zp++ ) {
2223 if ( ! trie->trans[ zp ].next ) {
2227 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2228 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2229 trie->trans[ zp ].check = state;
2230 if ( ++zp > pos ) pos = zp;
2237 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2239 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2240 trie->trans[ pos ].check = state;
2245 trie->lasttrans = pos + 1;
2246 trie->states = (reg_trie_state *)
2247 PerlMemShared_realloc( trie->states, laststate
2248 * sizeof(reg_trie_state) );
2249 DEBUG_TRIE_COMPILE_MORE_r(
2250 PerlIO_printf( Perl_debug_log,
2251 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2252 (int)depth * 2 + 2,"",
2253 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2256 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2259 } /* end table compress */
2261 DEBUG_TRIE_COMPILE_MORE_r(
2262 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2263 (int)depth * 2 + 2, "",
2264 (UV)trie->statecount,
2265 (UV)trie->lasttrans)
2267 /* resize the trans array to remove unused space */
2268 trie->trans = (reg_trie_trans *)
2269 PerlMemShared_realloc( trie->trans, trie->lasttrans
2270 * sizeof(reg_trie_trans) );
2272 { /* Modify the program and insert the new TRIE node */
2273 U8 nodetype =(U8)(flags & 0xFF);
2277 regnode *optimize = NULL;
2278 #ifdef RE_TRACK_PATTERN_OFFSETS
2281 U32 mjd_nodelen = 0;
2282 #endif /* RE_TRACK_PATTERN_OFFSETS */
2283 #endif /* DEBUGGING */
2285 This means we convert either the first branch or the first Exact,
2286 depending on whether the thing following (in 'last') is a branch
2287 or not and whther first is the startbranch (ie is it a sub part of
2288 the alternation or is it the whole thing.)
2289 Assuming its a sub part we convert the EXACT otherwise we convert
2290 the whole branch sequence, including the first.
2292 /* Find the node we are going to overwrite */
2293 if ( first != startbranch || OP( last ) == BRANCH ) {
2294 /* branch sub-chain */
2295 NEXT_OFF( first ) = (U16)(last - first);
2296 #ifdef RE_TRACK_PATTERN_OFFSETS
2298 mjd_offset= Node_Offset((convert));
2299 mjd_nodelen= Node_Length((convert));
2302 /* whole branch chain */
2304 #ifdef RE_TRACK_PATTERN_OFFSETS
2307 const regnode *nop = NEXTOPER( convert );
2308 mjd_offset= Node_Offset((nop));
2309 mjd_nodelen= Node_Length((nop));
2313 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2314 (int)depth * 2 + 2, "",
2315 (UV)mjd_offset, (UV)mjd_nodelen)
2318 /* But first we check to see if there is a common prefix we can
2319 split out as an EXACT and put in front of the TRIE node. */
2320 trie->startstate= 1;
2321 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2323 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2327 const U32 base = trie->states[ state ].trans.base;
2329 if ( trie->states[state].wordnum )
2332 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2333 if ( ( base + ofs >= trie->uniquecharcount ) &&
2334 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2335 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2337 if ( ++count > 1 ) {
2338 SV **tmp = av_fetch( revcharmap, ofs, 0);
2339 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2340 if ( state == 1 ) break;
2342 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2344 PerlIO_printf(Perl_debug_log,
2345 "%*sNew Start State=%"UVuf" Class: [",
2346 (int)depth * 2 + 2, "",
2349 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2350 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2352 TRIE_BITMAP_SET(trie,*ch);
2354 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2356 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2360 TRIE_BITMAP_SET(trie,*ch);
2362 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2363 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2369 SV **tmp = av_fetch( revcharmap, idx, 0);
2371 char *ch = SvPV( *tmp, len );
2373 SV *sv=sv_newmortal();
2374 PerlIO_printf( Perl_debug_log,
2375 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2376 (int)depth * 2 + 2, "",
2378 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2379 PL_colors[0], PL_colors[1],
2380 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2381 PERL_PV_ESCAPE_FIRSTCHAR
2386 OP( convert ) = nodetype;
2387 str=STRING(convert);
2390 STR_LEN(convert) += len;
2396 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2401 trie->prefixlen = (state-1);
2403 regnode *n = convert+NODE_SZ_STR(convert);
2404 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2405 trie->startstate = state;
2406 trie->minlen -= (state - 1);
2407 trie->maxlen -= (state - 1);
2409 /* At least the UNICOS C compiler choked on this
2410 * being argument to DEBUG_r(), so let's just have
2413 #ifdef PERL_EXT_RE_BUILD
2419 regnode *fix = convert;
2420 U32 word = trie->wordcount;
2422 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2423 while( ++fix < n ) {
2424 Set_Node_Offset_Length(fix, 0, 0);
2427 SV ** const tmp = av_fetch( trie_words, word, 0 );
2429 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2430 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2432 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2440 NEXT_OFF(convert) = (U16)(tail - convert);
2441 DEBUG_r(optimize= n);
2447 if ( trie->maxlen ) {
2448 NEXT_OFF( convert ) = (U16)(tail - convert);
2449 ARG_SET( convert, data_slot );
2450 /* Store the offset to the first unabsorbed branch in
2451 jump[0], which is otherwise unused by the jump logic.
2452 We use this when dumping a trie and during optimisation. */
2454 trie->jump[0] = (U16)(nextbranch - convert);
2456 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2457 * and there is a bitmap
2458 * and the first "jump target" node we found leaves enough room
2459 * then convert the TRIE node into a TRIEC node, with the bitmap
2460 * embedded inline in the opcode - this is hypothetically faster.
2462 if ( !trie->states[trie->startstate].wordnum
2464 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2466 OP( convert ) = TRIEC;
2467 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2468 PerlMemShared_free(trie->bitmap);
2471 OP( convert ) = TRIE;
2473 /* store the type in the flags */
2474 convert->flags = nodetype;
2478 + regarglen[ OP( convert ) ];
2480 /* XXX We really should free up the resource in trie now,
2481 as we won't use them - (which resources?) dmq */
2483 /* needed for dumping*/
2484 DEBUG_r(if (optimize) {
2485 regnode *opt = convert;
2487 while ( ++opt < optimize) {
2488 Set_Node_Offset_Length(opt,0,0);
2491 Try to clean up some of the debris left after the
2494 while( optimize < jumper ) {
2495 mjd_nodelen += Node_Length((optimize));
2496 OP( optimize ) = OPTIMIZED;
2497 Set_Node_Offset_Length(optimize,0,0);
2500 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2502 } /* end node insert */
2504 /* Finish populating the prev field of the wordinfo array. Walk back
2505 * from each accept state until we find another accept state, and if
2506 * so, point the first word's .prev field at the second word. If the
2507 * second already has a .prev field set, stop now. This will be the
2508 * case either if we've already processed that word's accept state,
2509 * or that state had multiple words, and the overspill words were
2510 * already linked up earlier.
2517 for (word=1; word <= trie->wordcount; word++) {
2519 if (trie->wordinfo[word].prev)
2521 state = trie->wordinfo[word].accept;
2523 state = prev_states[state];
2526 prev = trie->states[state].wordnum;
2530 trie->wordinfo[word].prev = prev;
2532 Safefree(prev_states);
2536 /* and now dump out the compressed format */
2537 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2539 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2541 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2542 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2544 SvREFCNT_dec_NN(revcharmap);
2548 : trie->startstate>1
2554 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2556 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2558 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2559 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2562 We find the fail state for each state in the trie, this state is the longest proper
2563 suffix of the current state's 'word' that is also a proper prefix of another word in our
2564 trie. State 1 represents the word '' and is thus the default fail state. This allows
2565 the DFA not to have to restart after its tried and failed a word at a given point, it
2566 simply continues as though it had been matching the other word in the first place.
2568 'abcdgu'=~/abcdefg|cdgu/
2569 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2570 fail, which would bring us to the state representing 'd' in the second word where we would
2571 try 'g' and succeed, proceeding to match 'cdgu'.
2573 /* add a fail transition */
2574 const U32 trie_offset = ARG(source);
2575 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2577 const U32 ucharcount = trie->uniquecharcount;
2578 const U32 numstates = trie->statecount;
2579 const U32 ubound = trie->lasttrans + ucharcount;
2583 U32 base = trie->states[ 1 ].trans.base;
2586 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2587 GET_RE_DEBUG_FLAGS_DECL;
2589 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2591 PERL_UNUSED_ARG(depth);
2595 ARG_SET( stclass, data_slot );
2596 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2597 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2598 aho->trie=trie_offset;
2599 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2600 Copy( trie->states, aho->states, numstates, reg_trie_state );
2601 Newxz( q, numstates, U32);
2602 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2605 /* initialize fail[0..1] to be 1 so that we always have
2606 a valid final fail state */
2607 fail[ 0 ] = fail[ 1 ] = 1;
2609 for ( charid = 0; charid < ucharcount ; charid++ ) {
2610 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2612 q[ q_write ] = newstate;
2613 /* set to point at the root */
2614 fail[ q[ q_write++ ] ]=1;
2617 while ( q_read < q_write) {
2618 const U32 cur = q[ q_read++ % numstates ];
2619 base = trie->states[ cur ].trans.base;
2621 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2622 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2624 U32 fail_state = cur;
2627 fail_state = fail[ fail_state ];
2628 fail_base = aho->states[ fail_state ].trans.base;
2629 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2631 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2632 fail[ ch_state ] = fail_state;
2633 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2635 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2637 q[ q_write++ % numstates] = ch_state;
2641 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2642 when we fail in state 1, this allows us to use the
2643 charclass scan to find a valid start char. This is based on the principle
2644 that theres a good chance the string being searched contains lots of stuff
2645 that cant be a start char.
2647 fail[ 0 ] = fail[ 1 ] = 0;
2648 DEBUG_TRIE_COMPILE_r({
2649 PerlIO_printf(Perl_debug_log,
2650 "%*sStclass Failtable (%"UVuf" states): 0",
2651 (int)(depth * 2), "", (UV)numstates
2653 for( q_read=1; q_read<numstates; q_read++ ) {
2654 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2656 PerlIO_printf(Perl_debug_log, "\n");
2659 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2664 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2665 * These need to be revisited when a newer toolchain becomes available.
2667 #if defined(__sparc64__) && defined(__GNUC__)
2668 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2669 # undef SPARC64_GCC_WORKAROUND
2670 # define SPARC64_GCC_WORKAROUND 1
2674 #define DEBUG_PEEP(str,scan,depth) \
2675 DEBUG_OPTIMISE_r({if (scan){ \
2676 SV * const mysv=sv_newmortal(); \
2677 regnode *Next = regnext(scan); \
2678 regprop(RExC_rx, mysv, scan); \
2679 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2680 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2681 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2685 /* The below joins as many adjacent EXACTish nodes as possible into a single
2686 * one. The regop may be changed if the node(s) contain certain sequences that
2687 * require special handling. The joining is only done if:
2688 * 1) there is room in the current conglomerated node to entirely contain the
2690 * 2) they are the exact same node type
2692 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2693 * these get optimized out
2695 * If a node is to match under /i (folded), the number of characters it matches
2696 * can be different than its character length if it contains a multi-character
2697 * fold. *min_subtract is set to the total delta of the input nodes.
2699 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2700 * and contains LATIN SMALL LETTER SHARP S
2702 * This is as good a place as any to discuss the design of handling these
2703 * multi-character fold sequences. It's been wrong in Perl for a very long
2704 * time. There are three code points in Unicode whose multi-character folds
2705 * were long ago discovered to mess things up. The previous designs for
2706 * dealing with these involved assigning a special node for them. This
2707 * approach doesn't work, as evidenced by this example:
2708 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2709 * Both these fold to "sss", but if the pattern is parsed to create a node that
2710 * would match just the \xDF, it won't be able to handle the case where a
2711 * successful match would have to cross the node's boundary. The new approach
2712 * that hopefully generally solves the problem generates an EXACTFU_SS node
2715 * It turns out that there are problems with all multi-character folds, and not
2716 * just these three. Now the code is general, for all such cases. The
2717 * approach taken is:
2718 * 1) This routine examines each EXACTFish node that could contain multi-
2719 * character fold sequences. It returns in *min_subtract how much to
2720 * subtract from the the actual length of the string to get a real minimum
2721 * match length; it is 0 if there are no multi-char folds. This delta is
2722 * used by the caller to adjust the min length of the match, and the delta
2723 * between min and max, so that the optimizer doesn't reject these
2724 * possibilities based on size constraints.
2725 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2726 * is used for an EXACTFU node that contains at least one "ss" sequence in
2727 * it. For non-UTF-8 patterns and strings, this is the only case where
2728 * there is a possible fold length change. That means that a regular
2729 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2730 * with length changes, and so can be processed faster. regexec.c takes
2731 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2732 * pre-folded by regcomp.c. This saves effort in regex matching.
2733 * However, the pre-folding isn't done for non-UTF8 patterns because the
2734 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2735 * down by forcing the pattern into UTF8 unless necessary. Also what
2736 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2737 * possibilities for the non-UTF8 patterns are quite simple, except for
2738 * the sharp s. All the ones that don't involve a UTF-8 target string are
2739 * members of a fold-pair, and arrays are set up for all of them so that
2740 * the other member of the pair can be found quickly. Code elsewhere in
2741 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2742 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2743 * described in the next item.
2744 * 3) A problem remains for the sharp s in EXACTF and EXACTFA nodes when the
2745 * pattern isn't in UTF-8. (BTW, there cannot be an EXACTF node with a
2746 * UTF-8 pattern.) An assumption that the optimizer part of regexec.c
2747 * (probably unwittingly, in Perl_regexec_flags()) makes is that a
2748 * character in the pattern corresponds to at most a single character in
2749 * the target string. (And I do mean character, and not byte here, unlike
2750 * other parts of the documentation that have never been updated to
2751 * account for multibyte Unicode.) sharp s in EXACTF nodes can match the
2752 * two character string 'ss'; in EXACTFA nodes it can match
2753 * "\x{17F}\x{17F}". These violate the assumption, and they are the only
2754 * instances where it is violated. I'm reluctant to try to change the
2755 * assumption, as the code involved is impenetrable to me (khw), so
2756 * instead the code here punts. This routine examines (when the pattern
2757 * isn't UTF-8) EXACTF and EXACTFA nodes for the sharp s, and returns a
2758 * boolean indicating whether or not the node contains a sharp s. When it
2759 * is true, the caller sets a flag that later causes the optimizer in this
2760 * file to not set values for the floating and fixed string lengths, and
2761 * thus avoids the optimizer code in regexec.c that makes the invalid
2762 * assumption. Thus, there is no optimization based on string lengths for
2763 * non-UTF8-pattern EXACTF and EXACTFA nodes that contain the sharp s.
2764 * (The reason the assumption is wrong only in these two cases is that all
2765 * other non-UTF-8 folds are 1-1; and, for UTF-8 patterns, we pre-fold all
2766 * other folds to their expanded versions. We can't prefold sharp s to
2767 * 'ss' in EXACTF nodes because we don't know at compile time if it
2768 * actually matches 'ss' or not. It will match iff the target string is
2769 * in UTF-8, unlike the EXACTFU nodes, where it always matches; and
2770 * EXACTFA and EXACTFL where it never does. In an EXACTFA node in a UTF-8
2771 * pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the problem;
2772 * but in a non-UTF8 pattern, folding it to that above-Latin1 string would
2773 * require the pattern to be forced into UTF-8, the overhead of which we
2776 * Similarly, the code that generates tries doesn't currently handle
2777 * not-already-folded multi-char folds, and it looks like a pain to change
2778 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
2779 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
2780 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
2781 * using /iaa matching will be doing so almost entirely with ASCII
2782 * strings, so this should rarely be encountered in practice */
2784 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2785 if (PL_regkind[OP(scan)] == EXACT) \
2786 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2789 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, UV *min_subtract, bool *has_exactf_sharp_s, U32 flags,regnode *val, U32 depth) {
2790 /* Merge several consecutive EXACTish nodes into one. */
2791 regnode *n = regnext(scan);
2793 regnode *next = scan + NODE_SZ_STR(scan);
2797 regnode *stop = scan;
2798 GET_RE_DEBUG_FLAGS_DECL;
2800 PERL_UNUSED_ARG(depth);
2803 PERL_ARGS_ASSERT_JOIN_EXACT;
2804 #ifndef EXPERIMENTAL_INPLACESCAN
2805 PERL_UNUSED_ARG(flags);
2806 PERL_UNUSED_ARG(val);
2808 DEBUG_PEEP("join",scan,depth);
2810 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2811 * EXACT ones that are mergeable to the current one. */
2813 && (PL_regkind[OP(n)] == NOTHING
2814 || (stringok && OP(n) == OP(scan)))
2816 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2819 if (OP(n) == TAIL || n > next)
2821 if (PL_regkind[OP(n)] == NOTHING) {
2822 DEBUG_PEEP("skip:",n,depth);
2823 NEXT_OFF(scan) += NEXT_OFF(n);
2824 next = n + NODE_STEP_REGNODE;
2831 else if (stringok) {
2832 const unsigned int oldl = STR_LEN(scan);
2833 regnode * const nnext = regnext(n);
2835 /* XXX I (khw) kind of doubt that this works on platforms where
2836 * U8_MAX is above 255 because of lots of other assumptions */
2837 /* Don't join if the sum can't fit into a single node */
2838 if (oldl + STR_LEN(n) > U8_MAX)
2841 DEBUG_PEEP("merg",n,depth);
2844 NEXT_OFF(scan) += NEXT_OFF(n);
2845 STR_LEN(scan) += STR_LEN(n);
2846 next = n + NODE_SZ_STR(n);
2847 /* Now we can overwrite *n : */
2848 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2856 #ifdef EXPERIMENTAL_INPLACESCAN
2857 if (flags && !NEXT_OFF(n)) {
2858 DEBUG_PEEP("atch", val, depth);
2859 if (reg_off_by_arg[OP(n)]) {
2860 ARG_SET(n, val - n);
2863 NEXT_OFF(n) = val - n;
2871 *has_exactf_sharp_s = FALSE;
2873 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2874 * can now analyze for sequences of problematic code points. (Prior to
2875 * this final joining, sequences could have been split over boundaries, and
2876 * hence missed). The sequences only happen in folding, hence for any
2877 * non-EXACT EXACTish node */
2878 if (OP(scan) != EXACT) {
2879 const U8 * const s0 = (U8*) STRING(scan);
2881 const U8 * const s_end = s0 + STR_LEN(scan);
2883 /* One pass is made over the node's string looking for all the
2884 * possibilities. to avoid some tests in the loop, there are two main
2885 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2889 /* Examine the string for a multi-character fold sequence. UTF-8
2890 * patterns have all characters pre-folded by the time this code is
2892 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2893 length sequence we are looking for is 2 */
2896 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2897 if (! len) { /* Not a multi-char fold: get next char */
2902 /* Nodes with 'ss' require special handling, except for EXACTFL
2903 * and EXACTFA-ish for which there is no multi-char fold to
2905 if (len == 2 && *s == 's' && *(s+1) == 's'
2906 && OP(scan) != EXACTFL
2907 && OP(scan) != EXACTFA
2908 && OP(scan) != EXACTFA_NO_TRIE)
2911 OP(scan) = EXACTFU_SS;
2914 else { /* Here is a generic multi-char fold. */
2915 const U8* multi_end = s + len;
2917 /* Count how many characters in it. In the case of /l and
2918 * /aa, no folds which contain ASCII code points are
2919 * allowed, so check for those, and skip if found. (In
2920 * EXACTFL, no folds are allowed to any Latin1 code point,
2921 * not just ASCII. But there aren't any of these
2922 * currently, nor ever likely, so don't take the time to
2923 * test for them. The code that generates the
2924 * is_MULTI_foo() macros croaks should one actually get put
2925 * into Unicode .) */
2926 if (OP(scan) != EXACTFL
2927 && OP(scan) != EXACTFA
2928 && OP(scan) != EXACTFA_NO_TRIE)
2930 count = utf8_length(s, multi_end);
2934 while (s < multi_end) {
2937 goto next_iteration;
2947 /* The delta is how long the sequence is minus 1 (1 is how long
2948 * the character that folds to the sequence is) */
2949 *min_subtract += count - 1;
2953 else if (OP(scan) == EXACTFA) {
2955 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
2956 * fold to the ASCII range (and there are no existing ones in the
2957 * upper latin1 range). But, as outlined in the comments preceding
2958 * this function, we need to flag any occurrences of the sharp s.
2959 * This character forbids trie formation (because of added
2962 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
2963 OP(scan) = EXACTFA_NO_TRIE;
2964 *has_exactf_sharp_s = TRUE;
2971 else if (OP(scan) != EXACTFL) {
2973 /* Non-UTF-8 pattern, not EXACTFA nor EXACTFL node. Look for the
2974 * multi-char folds that are all Latin1. (This code knows that
2975 * there are no current multi-char folds possible with EXACTFL,
2976 * relying on fold_grind.t to catch any errors if the very unlikely
2977 * event happens that some get added in future Unicode versions.)
2978 * As explained in the comments preceding this function, we look
2979 * also for the sharp s in EXACTF nodes; it can be in the final
2980 * position. Otherwise we can stop looking 1 byte earlier because
2981 * have to find at least two characters for a multi-fold */
2982 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2985 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2986 if (! len) { /* Not a multi-char fold. */
2987 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2989 *has_exactf_sharp_s = TRUE;
2996 && isARG2_lower_or_UPPER_ARG1('s', *s)
2997 && isARG2_lower_or_UPPER_ARG1('s', *(s+1)))
3000 /* EXACTF nodes need to know that the minimum length
3001 * changed so that a sharp s in the string can match this
3002 * ss in the pattern, but they remain EXACTF nodes, as they
3003 * won't match this unless the target string is is UTF-8,
3004 * which we don't know until runtime */
3005 if (OP(scan) != EXACTF) {
3006 OP(scan) = EXACTFU_SS;
3010 *min_subtract += len - 1;
3017 /* Allow dumping but overwriting the collection of skipped
3018 * ops and/or strings with fake optimized ops */
3019 n = scan + NODE_SZ_STR(scan);
3027 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3031 /* REx optimizer. Converts nodes into quicker variants "in place".
3032 Finds fixed substrings. */
3034 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3035 to the position after last scanned or to NULL. */
3037 #define INIT_AND_WITHP \
3038 assert(!and_withp); \
3039 Newx(and_withp,1,struct regnode_charclass_class); \
3040 SAVEFREEPV(and_withp)
3042 /* this is a chain of data about sub patterns we are processing that
3043 need to be handled separately/specially in study_chunk. Its so
3044 we can simulate recursion without losing state. */
3046 typedef struct scan_frame {
3047 regnode *last; /* last node to process in this frame */
3048 regnode *next; /* next node to process when last is reached */
3049 struct scan_frame *prev; /*previous frame*/
3050 I32 stop; /* what stopparen do we use */
3054 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
3057 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3058 SSize_t *minlenp, SSize_t *deltap,
3063 struct regnode_charclass_class *and_withp,
3064 U32 flags, U32 depth)
3065 /* scanp: Start here (read-write). */
3066 /* deltap: Write maxlen-minlen here. */
3067 /* last: Stop before this one. */
3068 /* data: string data about the pattern */
3069 /* stopparen: treat close N as END */
3070 /* recursed: which subroutines have we recursed into */
3071 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3074 /* There must be at least this number of characters to match */
3077 regnode *scan = *scanp, *next;
3079 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3080 int is_inf_internal = 0; /* The studied chunk is infinite */
3081 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3082 scan_data_t data_fake;
3083 SV *re_trie_maxbuff = NULL;
3084 regnode *first_non_open = scan;
3085 SSize_t stopmin = SSize_t_MAX;
3086 scan_frame *frame = NULL;
3087 GET_RE_DEBUG_FLAGS_DECL;
3089 PERL_ARGS_ASSERT_STUDY_CHUNK;
3092 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3096 while (first_non_open && OP(first_non_open) == OPEN)
3097 first_non_open=regnext(first_non_open);
3102 while ( scan && OP(scan) != END && scan < last ){
3103 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3104 node length to get a real minimum (because
3105 the folded version may be shorter) */
3106 bool has_exactf_sharp_s = FALSE;
3107 /* Peephole optimizer: */
3108 DEBUG_STUDYDATA("Peep:", data,depth);
3109 DEBUG_PEEP("Peep",scan,depth);
3111 /* Its not clear to khw or hv why this is done here, and not in the
3112 * clauses that deal with EXACT nodes. khw's guess is that it's
3113 * because of a previous design */
3114 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3116 /* Follow the next-chain of the current node and optimize
3117 away all the NOTHINGs from it. */
3118 if (OP(scan) != CURLYX) {
3119 const int max = (reg_off_by_arg[OP(scan)]
3121 /* I32 may be smaller than U16 on CRAYs! */
3122 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3123 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3127 /* Skip NOTHING and LONGJMP. */
3128 while ((n = regnext(n))
3129 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3130 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3131 && off + noff < max)
3133 if (reg_off_by_arg[OP(scan)])
3136 NEXT_OFF(scan) = off;
3141 /* The principal pseudo-switch. Cannot be a switch, since we
3142 look into several different things. */
3143 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3144 || OP(scan) == IFTHEN) {
3145 next = regnext(scan);
3147 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3149 if (OP(next) == code || code == IFTHEN) {
3150 /* NOTE - There is similar code to this block below for handling
3151 TRIE nodes on a re-study. If you change stuff here check there
3153 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
3154 struct regnode_charclass_class accum;
3155 regnode * const startbranch=scan;
3157 if (flags & SCF_DO_SUBSTR)
3158 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3159 if (flags & SCF_DO_STCLASS)
3160 cl_init_zero(pRExC_state, &accum);
3162 while (OP(scan) == code) {
3163 SSize_t deltanext, minnext, fake;
3165 struct regnode_charclass_class this_class;
3168 data_fake.flags = 0;
3170 data_fake.whilem_c = data->whilem_c;
3171 data_fake.last_closep = data->last_closep;
3174 data_fake.last_closep = &fake;
3176 data_fake.pos_delta = delta;
3177 next = regnext(scan);
3178 scan = NEXTOPER(scan);
3180 scan = NEXTOPER(scan);
3181 if (flags & SCF_DO_STCLASS) {
3182 cl_init(pRExC_state, &this_class);
3183 data_fake.start_class = &this_class;
3184 f = SCF_DO_STCLASS_AND;
3186 if (flags & SCF_WHILEM_VISITED_POS)
3187 f |= SCF_WHILEM_VISITED_POS;
3189 /* we suppose the run is continuous, last=next...*/
3190 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3192 stopparen, recursed, NULL, f,depth+1);
3195 if (deltanext == SSize_t_MAX) {
3196 is_inf = is_inf_internal = 1;
3198 } else if (max1 < minnext + deltanext)
3199 max1 = minnext + deltanext;
3201 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3203 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3204 if ( stopmin > minnext)
3205 stopmin = min + min1;
3206 flags &= ~SCF_DO_SUBSTR;
3208 data->flags |= SCF_SEEN_ACCEPT;
3211 if (data_fake.flags & SF_HAS_EVAL)
3212 data->flags |= SF_HAS_EVAL;
3213 data->whilem_c = data_fake.whilem_c;
3215 if (flags & SCF_DO_STCLASS)
3216 cl_or(pRExC_state, &accum, &this_class);
3218 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3220 if (flags & SCF_DO_SUBSTR) {
3221 data->pos_min += min1;
3222 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3223 data->pos_delta = SSize_t_MAX;
3225 data->pos_delta += max1 - min1;
3226 if (max1 != min1 || is_inf)
3227 data->longest = &(data->longest_float);
3230 if (delta == SSize_t_MAX
3231 || SSize_t_MAX - delta - (max1 - min1) < 0)
3232 delta = SSize_t_MAX;
3234 delta += max1 - min1;
3235 if (flags & SCF_DO_STCLASS_OR) {
3236 cl_or(pRExC_state, data->start_class, &accum);
3238 cl_and(data->start_class, and_withp);
3239 flags &= ~SCF_DO_STCLASS;
3242 else if (flags & SCF_DO_STCLASS_AND) {
3244 cl_and(data->start_class, &accum);
3245 flags &= ~SCF_DO_STCLASS;
3248 /* Switch to OR mode: cache the old value of
3249 * data->start_class */
3251 StructCopy(data->start_class, and_withp,
3252 struct regnode_charclass_class);
3253 flags &= ~SCF_DO_STCLASS_AND;
3254 StructCopy(&accum, data->start_class,
3255 struct regnode_charclass_class);
3256 flags |= SCF_DO_STCLASS_OR;
3257 SET_SSC_EOS(data->start_class);
3261 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3264 Assuming this was/is a branch we are dealing with: 'scan' now
3265 points at the item that follows the branch sequence, whatever
3266 it is. We now start at the beginning of the sequence and look
3273 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3275 If we can find such a subsequence we need to turn the first
3276 element into a trie and then add the subsequent branch exact
3277 strings to the trie.
3281 1. patterns where the whole set of branches can be converted.
3283 2. patterns where only a subset can be converted.
3285 In case 1 we can replace the whole set with a single regop
3286 for the trie. In case 2 we need to keep the start and end
3289 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3290 becomes BRANCH TRIE; BRANCH X;
3292 There is an additional case, that being where there is a
3293 common prefix, which gets split out into an EXACT like node
3294 preceding the TRIE node.
3296 If x(1..n)==tail then we can do a simple trie, if not we make
3297 a "jump" trie, such that when we match the appropriate word
3298 we "jump" to the appropriate tail node. Essentially we turn
3299 a nested if into a case structure of sorts.
3304 if (!re_trie_maxbuff) {
3305 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3306 if (!SvIOK(re_trie_maxbuff))
3307 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3309 if ( SvIV(re_trie_maxbuff)>=0 ) {
3311 regnode *first = (regnode *)NULL;
3312 regnode *last = (regnode *)NULL;
3313 regnode *tail = scan;
3318 SV * const mysv = sv_newmortal(); /* for dumping */
3320 /* var tail is used because there may be a TAIL
3321 regop in the way. Ie, the exacts will point to the
3322 thing following the TAIL, but the last branch will
3323 point at the TAIL. So we advance tail. If we
3324 have nested (?:) we may have to move through several
3328 while ( OP( tail ) == TAIL ) {
3329 /* this is the TAIL generated by (?:) */
3330 tail = regnext( tail );
3334 DEBUG_TRIE_COMPILE_r({
3335 regprop(RExC_rx, mysv, tail );
3336 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3337 (int)depth * 2 + 2, "",
3338 "Looking for TRIE'able sequences. Tail node is: ",
3339 SvPV_nolen_const( mysv )
3345 Step through the branches
3346 cur represents each branch,
3347 noper is the first thing to be matched as part of that branch
3348 noper_next is the regnext() of that node.
3350 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3351 via a "jump trie" but we also support building with NOJUMPTRIE,
3352 which restricts the trie logic to structures like /FOO|BAR/.
3354 If noper is a trieable nodetype then the branch is a possible optimization
3355 target. If we are building under NOJUMPTRIE then we require that noper_next
3356 is the same as scan (our current position in the regex program).
3358 Once we have two or more consecutive such branches we can create a
3359 trie of the EXACT's contents and stitch it in place into the program.
3361 If the sequence represents all of the branches in the alternation we
3362 replace the entire thing with a single TRIE node.
3364 Otherwise when it is a subsequence we need to stitch it in place and
3365 replace only the relevant branches. This means the first branch has
3366 to remain as it is used by the alternation logic, and its next pointer,
3367 and needs to be repointed at the item on the branch chain following
3368 the last branch we have optimized away.
3370 This could be either a BRANCH, in which case the subsequence is internal,
3371 or it could be the item following the branch sequence in which case the
3372 subsequence is at the end (which does not necessarily mean the first node
3373 is the start of the alternation).
3375 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3378 ----------------+-----------
3382 EXACTFU_SS | EXACTFU
3387 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3388 ( EXACT == (X) ) ? EXACT : \
3389 ( EXACTFU == (X) || EXACTFU_SS == (X) ) ? EXACTFU : \
3390 ( EXACTFA == (X) ) ? EXACTFA : \
3393 /* dont use tail as the end marker for this traverse */
3394 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3395 regnode * const noper = NEXTOPER( cur );
3396 U8 noper_type = OP( noper );
3397 U8 noper_trietype = TRIE_TYPE( noper_type );
3398 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3399 regnode * const noper_next = regnext( noper );
3400 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3401 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3404 DEBUG_TRIE_COMPILE_r({
3405 regprop(RExC_rx, mysv, cur);
3406 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3407 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3409 regprop(RExC_rx, mysv, noper);
3410 PerlIO_printf( Perl_debug_log, " -> %s",
3411 SvPV_nolen_const(mysv));
3414 regprop(RExC_rx, mysv, noper_next );
3415 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3416 SvPV_nolen_const(mysv));
3418 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3419 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3420 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3424 /* Is noper a trieable nodetype that can be merged with the
3425 * current trie (if there is one)? */
3429 ( noper_trietype == NOTHING)
3430 || ( trietype == NOTHING )
3431 || ( trietype == noper_trietype )
3434 && noper_next == tail
3438 /* Handle mergable triable node
3439 * Either we are the first node in a new trieable sequence,
3440 * in which case we do some bookkeeping, otherwise we update
3441 * the end pointer. */
3444 if ( noper_trietype == NOTHING ) {
3445 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3446 regnode * const noper_next = regnext( noper );
3447 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3448 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3451 if ( noper_next_trietype ) {
3452 trietype = noper_next_trietype;
3453 } else if (noper_next_type) {
3454 /* a NOTHING regop is 1 regop wide. We need at least two
3455 * for a trie so we can't merge this in */
3459 trietype = noper_trietype;
3462 if ( trietype == NOTHING )
3463 trietype = noper_trietype;
3468 } /* end handle mergable triable node */
3470 /* handle unmergable node -
3471 * noper may either be a triable node which can not be tried
3472 * together with the current trie, or a non triable node */
3474 /* If last is set and trietype is not NOTHING then we have found
3475 * at least two triable branch sequences in a row of a similar
3476 * trietype so we can turn them into a trie. If/when we
3477 * allow NOTHING to start a trie sequence this condition will be
3478 * required, and it isn't expensive so we leave it in for now. */
3479 if ( trietype && trietype != NOTHING )
3480 make_trie( pRExC_state,
3481 startbranch, first, cur, tail, count,
3482 trietype, depth+1 );
3483 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3487 && noper_next == tail
3490 /* noper is triable, so we can start a new trie sequence */
3493 trietype = noper_trietype;
3495 /* if we already saw a first but the current node is not triable then we have
3496 * to reset the first information. */
3501 } /* end handle unmergable node */
3502 } /* loop over branches */
3503 DEBUG_TRIE_COMPILE_r({
3504 regprop(RExC_rx, mysv, cur);
3505 PerlIO_printf( Perl_debug_log,
3506 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3507 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3510 if ( last && trietype ) {
3511 if ( trietype != NOTHING ) {
3512 /* the last branch of the sequence was part of a trie,
3513 * so we have to construct it here outside of the loop
3515 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3516 #ifdef TRIE_STUDY_OPT
3517 if ( ((made == MADE_EXACT_TRIE &&
3518 startbranch == first)
3519 || ( first_non_open == first )) &&
3521 flags |= SCF_TRIE_RESTUDY;
3522 if ( startbranch == first
3525 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3530 /* at this point we know whatever we have is a NOTHING sequence/branch
3531 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3533 if ( startbranch == first ) {
3535 /* the entire thing is a NOTHING sequence, something like this:
3536 * (?:|) So we can turn it into a plain NOTHING op. */
3537 DEBUG_TRIE_COMPILE_r({
3538 regprop(RExC_rx, mysv, cur);
3539 PerlIO_printf( Perl_debug_log,
3540 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3541 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3544 OP(startbranch)= NOTHING;
3545 NEXT_OFF(startbranch)= tail - startbranch;
3546 for ( opt= startbranch + 1; opt < tail ; opt++ )
3550 } /* end if ( last) */
3551 } /* TRIE_MAXBUF is non zero */
3556 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3557 scan = NEXTOPER(NEXTOPER(scan));
3558 } else /* single branch is optimized. */
3559 scan = NEXTOPER(scan);
3561 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3562 scan_frame *newframe = NULL;
3567 if (OP(scan) != SUSPEND) {
3568 /* set the pointer */
3569 if (OP(scan) == GOSUB) {
3571 RExC_recurse[ARG2L(scan)] = scan;
3572 start = RExC_open_parens[paren-1];
3573 end = RExC_close_parens[paren-1];
3576 start = RExC_rxi->program + 1;
3580 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3581 SAVEFREEPV(recursed);
3583 if (!PAREN_TEST(recursed,paren+1)) {
3584 PAREN_SET(recursed,paren+1);
3585 Newx(newframe,1,scan_frame);
3587 if (flags & SCF_DO_SUBSTR) {
3588 SCAN_COMMIT(pRExC_state,data,minlenp);
3589 data->longest = &(data->longest_float);
3591 is_inf = is_inf_internal = 1;
3592 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3593 cl_anything(pRExC_state, data->start_class);
3594 flags &= ~SCF_DO_STCLASS;
3597 Newx(newframe,1,scan_frame);
3600 end = regnext(scan);
3605 SAVEFREEPV(newframe);
3606 newframe->next = regnext(scan);
3607 newframe->last = last;
3608 newframe->stop = stopparen;
3609 newframe->prev = frame;
3619 else if (OP(scan) == EXACT) {
3620 SSize_t l = STR_LEN(scan);
3623 const U8 * const s = (U8*)STRING(scan);
3624 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3625 l = utf8_length(s, s + l);
3627 uc = *((U8*)STRING(scan));
3630 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3631 /* The code below prefers earlier match for fixed
3632 offset, later match for variable offset. */
3633 if (data->last_end == -1) { /* Update the start info. */
3634 data->last_start_min = data->pos_min;
3635 data->last_start_max = is_inf
3636 ? SSize_t_MAX : data->pos_min + data->pos_delta;
3638 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3640 SvUTF8_on(data->last_found);
3642 SV * const sv = data->last_found;
3643 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3644 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3645 if (mg && mg->mg_len >= 0)
3646 mg->mg_len += utf8_length((U8*)STRING(scan),
3647 (U8*)STRING(scan)+STR_LEN(scan));
3649 data->last_end = data->pos_min + l;
3650 data->pos_min += l; /* As in the first entry. */
3651 data->flags &= ~SF_BEFORE_EOL;
3653 if (flags & SCF_DO_STCLASS_AND) {
3654 /* Check whether it is compatible with what we know already! */
3658 /* If compatible, we or it in below. It is compatible if is
3659 * in the bitmp and either 1) its bit or its fold is set, or 2)
3660 * it's for a locale. Even if there isn't unicode semantics
3661 * here, at runtime there may be because of matching against a
3662 * utf8 string, so accept a possible false positive for
3663 * latin1-range folds */
3665 (!(data->start_class->flags & ANYOF_LOCALE)
3666 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3667 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3668 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3673 ANYOF_CLASS_ZERO(data->start_class);
3674 ANYOF_BITMAP_ZERO(data->start_class);
3676 ANYOF_BITMAP_SET(data->start_class, uc);
3677 else if (uc >= 0x100) {
3680 /* Some Unicode code points fold to the Latin1 range; as
3681 * XXX temporary code, instead of figuring out if this is
3682 * one, just assume it is and set all the start class bits
3683 * that could be some such above 255 code point's fold
3684 * which will generate fals positives. As the code
3685 * elsewhere that does compute the fold settles down, it
3686 * can be extracted out and re-used here */
3687 for (i = 0; i < 256; i++){
3688 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3689 ANYOF_BITMAP_SET(data->start_class, i);
3693 CLEAR_SSC_EOS(data->start_class);
3695 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3697 else if (flags & SCF_DO_STCLASS_OR) {
3698 /* false positive possible if the class is case-folded */
3700 ANYOF_BITMAP_SET(data->start_class, uc);
3702 data->start_class->flags |= ANYOF_UNICODE_ALL;
3703 CLEAR_SSC_EOS(data->start_class);
3704 cl_and(data->start_class, and_withp);
3706 flags &= ~SCF_DO_STCLASS;
3708 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3709 SSize_t l = STR_LEN(scan);
3710 UV uc = *((U8*)STRING(scan));
3712 /* Search for fixed substrings supports EXACT only. */
3713 if (flags & SCF_DO_SUBSTR) {
3715 SCAN_COMMIT(pRExC_state, data, minlenp);
3718 const U8 * const s = (U8 *)STRING(scan);
3719 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3720 l = utf8_length(s, s + l);
3722 if (has_exactf_sharp_s) {
3723 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3725 min += l - min_subtract;
3727 delta += min_subtract;
3728 if (flags & SCF_DO_SUBSTR) {
3729 data->pos_min += l - min_subtract;
3730 if (data->pos_min < 0) {
3733 data->pos_delta += min_subtract;
3735 data->longest = &(data->longest_float);
3738 if (flags & SCF_DO_STCLASS_AND) {
3739 /* Check whether it is compatible with what we know already! */
3742 (!(data->start_class->flags & ANYOF_LOCALE)
3743 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3744 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3748 ANYOF_CLASS_ZERO(data->start_class);
3749 ANYOF_BITMAP_ZERO(data->start_class);
3751 ANYOF_BITMAP_SET(data->start_class, uc);
3752 CLEAR_SSC_EOS(data->start_class);
3753 if (OP(scan) == EXACTFL) {
3754 /* XXX This set is probably no longer necessary, and
3755 * probably wrong as LOCALE now is on in the initial
3757 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3761 /* Also set the other member of the fold pair. In case
3762 * that unicode semantics is called for at runtime, use
3763 * the full latin1 fold. (Can't do this for locale,
3764 * because not known until runtime) */
3765 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3767 /* All other (EXACTFL handled above) folds except under
3768 * /iaa that include s, S, and sharp_s also may include
3770 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE)
3772 if (uc == 's' || uc == 'S') {
3773 ANYOF_BITMAP_SET(data->start_class,
3774 LATIN_SMALL_LETTER_SHARP_S);
3776 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3777 ANYOF_BITMAP_SET(data->start_class, 's');
3778 ANYOF_BITMAP_SET(data->start_class, 'S');
3783 else if (uc >= 0x100) {
3785 for (i = 0; i < 256; i++){
3786 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3787 ANYOF_BITMAP_SET(data->start_class, i);
3792 else if (flags & SCF_DO_STCLASS_OR) {
3793 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3794 /* false positive possible if the class is case-folded.
3795 Assume that the locale settings are the same... */
3797 ANYOF_BITMAP_SET(data->start_class, uc);
3798 if (OP(scan) != EXACTFL) {
3800 /* And set the other member of the fold pair, but
3801 * can't do that in locale because not known until
3803 ANYOF_BITMAP_SET(data->start_class,
3804 PL_fold_latin1[uc]);
3806 /* All folds except under /iaa that include s, S,
3807 * and sharp_s also may include the others */
3808 if (OP(scan) != EXACTFA
3809 && OP(scan) != EXACTFA_NO_TRIE)
3811 if (uc == 's' || uc == 'S') {
3812 ANYOF_BITMAP_SET(data->start_class,
3813 LATIN_SMALL_LETTER_SHARP_S);
3815 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3816 ANYOF_BITMAP_SET(data->start_class, 's');
3817 ANYOF_BITMAP_SET(data->start_class, 'S');
3822 CLEAR_SSC_EOS(data->start_class);
3824 cl_and(data->start_class, and_withp);
3826 flags &= ~SCF_DO_STCLASS;
3828 else if (REGNODE_VARIES(OP(scan))) {
3829 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
3830 I32 fl = 0, f = flags;
3831 regnode * const oscan = scan;
3832 struct regnode_charclass_class this_class;
3833 struct regnode_charclass_class *oclass = NULL;
3834 I32 next_is_eval = 0;
3836 switch (PL_regkind[OP(scan)]) {
3837 case WHILEM: /* End of (?:...)* . */
3838 scan = NEXTOPER(scan);
3841 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3842 next = NEXTOPER(scan);
3843 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3845 maxcount = REG_INFTY;
3846 next = regnext(scan);
3847 scan = NEXTOPER(scan);
3851 if (flags & SCF_DO_SUBSTR)
https://perl5.git.perl.org / perl5.git / blob