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 I32 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; ®dummy = don't = compiling */
130 I32 naughty; /* How bad is this pattern? */
131 I32 sawback; /* Did we see \1, ...? */
133 I32 size; /* Code size. */
134 I32 npar; /* Capture buffer count, (OPEN). */
135 I32 cpar; /* Capture buffer count, (CLOSE). */
136 I32 nestroot; /* root parens we are in - used by accept */
139 regnode **open_parens; /* pointers to open parens */
140 regnode **close_parens; /* pointers to close parens */
141 regnode *opend; /* END node in program */
142 I32 utf8; /* whether the pattern is utf8 or not */
143 I32 orig_utf8; /* whether the pattern was originally in utf8 */
144 /* XXX use this for future optimisation of case
145 * where pattern must be upgraded to utf8. */
146 I32 uni_semantics; /* If a d charset modifier should use unicode
147 rules, even if the pattern is not in
149 HV *paren_names; /* Paren names */
151 regnode **recurse; /* Recurse regops */
152 I32 recurse_count; /* Number of recurse regops */
155 I32 override_recoding;
156 I32 in_multi_char_class;
157 struct reg_code_block *code_blocks; /* positions of literal (?{})
159 int num_code_blocks; /* size of code_blocks[] */
160 int code_index; /* next code_blocks[] slot */
162 char *starttry; /* -Dr: where regtry was called. */
163 #define RExC_starttry (pRExC_state->starttry)
165 SV *runtime_code_qr; /* qr with the runtime code blocks */
167 const char *lastparse;
169 AV *paren_name_list; /* idx -> name */
170 #define RExC_lastparse (pRExC_state->lastparse)
171 #define RExC_lastnum (pRExC_state->lastnum)
172 #define RExC_paren_name_list (pRExC_state->paren_name_list)
176 #define RExC_flags (pRExC_state->flags)
177 #define RExC_pm_flags (pRExC_state->pm_flags)
178 #define RExC_precomp (pRExC_state->precomp)
179 #define RExC_rx_sv (pRExC_state->rx_sv)
180 #define RExC_rx (pRExC_state->rx)
181 #define RExC_rxi (pRExC_state->rxi)
182 #define RExC_start (pRExC_state->start)
183 #define RExC_end (pRExC_state->end)
184 #define RExC_parse (pRExC_state->parse)
185 #define RExC_whilem_seen (pRExC_state->whilem_seen)
186 #ifdef RE_TRACK_PATTERN_OFFSETS
187 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
189 #define RExC_emit (pRExC_state->emit)
190 #define RExC_emit_start (pRExC_state->emit_start)
191 #define RExC_emit_bound (pRExC_state->emit_bound)
192 #define RExC_naughty (pRExC_state->naughty)
193 #define RExC_sawback (pRExC_state->sawback)
194 #define RExC_seen (pRExC_state->seen)
195 #define RExC_size (pRExC_state->size)
196 #define RExC_npar (pRExC_state->npar)
197 #define RExC_nestroot (pRExC_state->nestroot)
198 #define RExC_extralen (pRExC_state->extralen)
199 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
200 #define RExC_utf8 (pRExC_state->utf8)
201 #define RExC_uni_semantics (pRExC_state->uni_semantics)
202 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
203 #define RExC_open_parens (pRExC_state->open_parens)
204 #define RExC_close_parens (pRExC_state->close_parens)
205 #define RExC_opend (pRExC_state->opend)
206 #define RExC_paren_names (pRExC_state->paren_names)
207 #define RExC_recurse (pRExC_state->recurse)
208 #define RExC_recurse_count (pRExC_state->recurse_count)
209 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
210 #define RExC_contains_locale (pRExC_state->contains_locale)
211 #define RExC_override_recoding (pRExC_state->override_recoding)
212 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
215 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
216 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
217 ((*s) == '{' && regcurly(s)))
220 #undef SPSTART /* dratted cpp namespace... */
223 * Flags to be passed up and down.
225 #define WORST 0 /* Worst case. */
226 #define HASWIDTH 0x01 /* Known to match non-null strings. */
228 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
229 * character. (There needs to be a case: in the switch statement in regexec.c
230 * for any node marked SIMPLE.) Note that this is not the same thing as
233 #define SPSTART 0x04 /* Starts with * or + */
234 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
235 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
237 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
239 /* whether trie related optimizations are enabled */
240 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
241 #define TRIE_STUDY_OPT
242 #define FULL_TRIE_STUDY
248 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
249 #define PBITVAL(paren) (1 << ((paren) & 7))
250 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
251 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
252 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
254 /* If not already in utf8, do a longjmp back to the beginning */
255 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
256 #define REQUIRE_UTF8 STMT_START { \
257 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
260 /* This converts the named class defined in regcomp.h to its equivalent class
261 * number defined in handy.h. */
262 #define namedclass_to_classnum(class) ((int) ((class) / 2))
263 #define classnum_to_namedclass(classnum) ((classnum) * 2)
265 /* About scan_data_t.
267 During optimisation we recurse through the regexp program performing
268 various inplace (keyhole style) optimisations. In addition study_chunk
269 and scan_commit populate this data structure with information about
270 what strings MUST appear in the pattern. We look for the longest
271 string that must appear at a fixed location, and we look for the
272 longest string that may appear at a floating location. So for instance
277 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
278 strings (because they follow a .* construct). study_chunk will identify
279 both FOO and BAR as being the longest fixed and floating strings respectively.
281 The strings can be composites, for instance
285 will result in a composite fixed substring 'foo'.
287 For each string some basic information is maintained:
289 - offset or min_offset
290 This is the position the string must appear at, or not before.
291 It also implicitly (when combined with minlenp) tells us how many
292 characters must match before the string we are searching for.
293 Likewise when combined with minlenp and the length of the string it
294 tells us how many characters must appear after the string we have
298 Only used for floating strings. This is the rightmost point that
299 the string can appear at. If set to I32 max it indicates that the
300 string can occur infinitely far to the right.
303 A pointer to the minimum number of characters of the pattern that the
304 string was found inside. This is important as in the case of positive
305 lookahead or positive lookbehind we can have multiple patterns
310 The minimum length of the pattern overall is 3, the minimum length
311 of the lookahead part is 3, but the minimum length of the part that
312 will actually match is 1. So 'FOO's minimum length is 3, but the
313 minimum length for the F is 1. This is important as the minimum length
314 is used to determine offsets in front of and behind the string being
315 looked for. Since strings can be composites this is the length of the
316 pattern at the time it was committed with a scan_commit. Note that
317 the length is calculated by study_chunk, so that the minimum lengths
318 are not known until the full pattern has been compiled, thus the
319 pointer to the value.
323 In the case of lookbehind the string being searched for can be
324 offset past the start point of the final matching string.
325 If this value was just blithely removed from the min_offset it would
326 invalidate some of the calculations for how many chars must match
327 before or after (as they are derived from min_offset and minlen and
328 the length of the string being searched for).
329 When the final pattern is compiled and the data is moved from the
330 scan_data_t structure into the regexp structure the information
331 about lookbehind is factored in, with the information that would
332 have been lost precalculated in the end_shift field for the
335 The fields pos_min and pos_delta are used to store the minimum offset
336 and the delta to the maximum offset at the current point in the pattern.
340 typedef struct scan_data_t {
341 /*I32 len_min; unused */
342 /*I32 len_delta; unused */
346 I32 last_end; /* min value, <0 unless valid. */
349 SV **longest; /* Either &l_fixed, or &l_float. */
350 SV *longest_fixed; /* longest fixed string found in pattern */
351 I32 offset_fixed; /* offset where it starts */
352 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
353 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
354 SV *longest_float; /* longest floating string found in pattern */
355 I32 offset_float_min; /* earliest point in string it can appear */
356 I32 offset_float_max; /* latest point in string it can appear */
357 I32 *minlen_float; /* pointer to the minlen relevant to the string */
358 I32 lookbehind_float; /* is the position of the string modified by LB */
362 struct regnode_charclass_class *start_class;
366 * Forward declarations for pregcomp()'s friends.
369 static const scan_data_t zero_scan_data =
370 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
372 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
373 #define SF_BEFORE_SEOL 0x0001
374 #define SF_BEFORE_MEOL 0x0002
375 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
376 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
379 # define SF_FIX_SHIFT_EOL (0+2)
380 # define SF_FL_SHIFT_EOL (0+4)
382 # define SF_FIX_SHIFT_EOL (+2)
383 # define SF_FL_SHIFT_EOL (+4)
386 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
387 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
389 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
390 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
391 #define SF_IS_INF 0x0040
392 #define SF_HAS_PAR 0x0080
393 #define SF_IN_PAR 0x0100
394 #define SF_HAS_EVAL 0x0200
395 #define SCF_DO_SUBSTR 0x0400
396 #define SCF_DO_STCLASS_AND 0x0800
397 #define SCF_DO_STCLASS_OR 0x1000
398 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
399 #define SCF_WHILEM_VISITED_POS 0x2000
401 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
402 #define SCF_SEEN_ACCEPT 0x8000
404 #define UTF cBOOL(RExC_utf8)
406 /* The enums for all these are ordered so things work out correctly */
407 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
408 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
409 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
410 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
411 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
412 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
413 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
415 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
417 #define OOB_NAMEDCLASS -1
419 /* There is no code point that is out-of-bounds, so this is problematic. But
420 * its only current use is to initialize a variable that is always set before
422 #define OOB_UNICODE 0xDEADBEEF
424 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
425 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
428 /* length of regex to show in messages that don't mark a position within */
429 #define RegexLengthToShowInErrorMessages 127
432 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
433 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
434 * op/pragma/warn/regcomp.
436 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
437 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
439 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
442 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
443 * arg. Show regex, up to a maximum length. If it's too long, chop and add
446 #define _FAIL(code) STMT_START { \
447 const char *ellipses = ""; \
448 IV len = RExC_end - RExC_precomp; \
451 SAVEFREESV(RExC_rx_sv); \
452 if (len > RegexLengthToShowInErrorMessages) { \
453 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
454 len = RegexLengthToShowInErrorMessages - 10; \
460 #define FAIL(msg) _FAIL( \
461 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
462 msg, (int)len, RExC_precomp, ellipses))
464 #define FAIL2(msg,arg) _FAIL( \
465 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
466 arg, (int)len, RExC_precomp, ellipses))
469 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
471 #define Simple_vFAIL(m) STMT_START { \
472 const IV offset = RExC_parse - RExC_precomp; \
473 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
474 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
478 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
480 #define vFAIL(m) STMT_START { \
482 SAVEFREESV(RExC_rx_sv); \
487 * Like Simple_vFAIL(), but accepts two arguments.
489 #define Simple_vFAIL2(m,a1) STMT_START { \
490 const IV offset = RExC_parse - RExC_precomp; \
491 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
492 (int)offset, RExC_precomp, RExC_precomp + offset); \
496 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
498 #define vFAIL2(m,a1) STMT_START { \
500 SAVEFREESV(RExC_rx_sv); \
501 Simple_vFAIL2(m, a1); \
506 * Like Simple_vFAIL(), but accepts three arguments.
508 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
509 const IV offset = RExC_parse - RExC_precomp; \
510 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
511 (int)offset, RExC_precomp, RExC_precomp + offset); \
515 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
517 #define vFAIL3(m,a1,a2) STMT_START { \
519 SAVEFREESV(RExC_rx_sv); \
520 Simple_vFAIL3(m, a1, a2); \
524 * Like Simple_vFAIL(), but accepts four arguments.
526 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
527 const IV offset = RExC_parse - RExC_precomp; \
528 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
529 (int)offset, RExC_precomp, RExC_precomp + offset); \
532 #define vFAIL4(m,a1,a2,a3) STMT_START { \
534 SAVEFREESV(RExC_rx_sv); \
535 Simple_vFAIL4(m, a1, a2, a3); \
538 #define ckWARNreg(loc,m) STMT_START { \
539 const IV offset = loc - RExC_precomp; \
540 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
541 (int)offset, RExC_precomp, RExC_precomp + offset); \
544 #define ckWARNregdep(loc,m) STMT_START { \
545 const IV offset = loc - RExC_precomp; \
546 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
548 (int)offset, RExC_precomp, RExC_precomp + offset); \
551 #define ckWARN2regdep(loc,m, a1) STMT_START { \
552 const IV offset = loc - RExC_precomp; \
553 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
555 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
558 #define ckWARN2reg(loc, m, a1) STMT_START { \
559 const IV offset = loc - RExC_precomp; \
560 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
561 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
564 #define vWARN3(loc, m, a1, a2) STMT_START { \
565 const IV offset = loc - RExC_precomp; \
566 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
567 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
570 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
571 const IV offset = loc - RExC_precomp; \
572 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
573 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
576 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
577 const IV offset = loc - RExC_precomp; \
578 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
579 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
582 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
583 const IV offset = loc - RExC_precomp; \
584 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
585 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
588 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
589 const IV offset = loc - RExC_precomp; \
590 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
591 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
595 /* Allow for side effects in s */
596 #define REGC(c,s) STMT_START { \
597 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
600 /* Macros for recording node offsets. 20001227 mjd@plover.com
601 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
602 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
603 * Element 0 holds the number n.
604 * Position is 1 indexed.
606 #ifndef RE_TRACK_PATTERN_OFFSETS
607 #define Set_Node_Offset_To_R(node,byte)
608 #define Set_Node_Offset(node,byte)
609 #define Set_Cur_Node_Offset
610 #define Set_Node_Length_To_R(node,len)
611 #define Set_Node_Length(node,len)
612 #define Set_Node_Cur_Length(node)
613 #define Node_Offset(n)
614 #define Node_Length(n)
615 #define Set_Node_Offset_Length(node,offset,len)
616 #define ProgLen(ri) ri->u.proglen
617 #define SetProgLen(ri,x) ri->u.proglen = x
619 #define ProgLen(ri) ri->u.offsets[0]
620 #define SetProgLen(ri,x) ri->u.offsets[0] = x
621 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
623 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
624 __LINE__, (int)(node), (int)(byte))); \
626 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
628 RExC_offsets[2*(node)-1] = (byte); \
633 #define Set_Node_Offset(node,byte) \
634 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
635 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
637 #define Set_Node_Length_To_R(node,len) STMT_START { \
639 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
640 __LINE__, (int)(node), (int)(len))); \
642 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
644 RExC_offsets[2*(node)] = (len); \
649 #define Set_Node_Length(node,len) \
650 Set_Node_Length_To_R((node)-RExC_emit_start, len)
651 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
652 #define Set_Node_Cur_Length(node) \
653 Set_Node_Length(node, RExC_parse - parse_start)
655 /* Get offsets and lengths */
656 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
657 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
659 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
660 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
661 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
665 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
666 #define EXPERIMENTAL_INPLACESCAN
667 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
669 #define DEBUG_STUDYDATA(str,data,depth) \
670 DEBUG_OPTIMISE_MORE_r(if(data){ \
671 PerlIO_printf(Perl_debug_log, \
672 "%*s" str "Pos:%"IVdf"/%"IVdf \
673 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
674 (int)(depth)*2, "", \
675 (IV)((data)->pos_min), \
676 (IV)((data)->pos_delta), \
677 (UV)((data)->flags), \
678 (IV)((data)->whilem_c), \
679 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
680 is_inf ? "INF " : "" \
682 if ((data)->last_found) \
683 PerlIO_printf(Perl_debug_log, \
684 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
685 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
686 SvPVX_const((data)->last_found), \
687 (IV)((data)->last_end), \
688 (IV)((data)->last_start_min), \
689 (IV)((data)->last_start_max), \
690 ((data)->longest && \
691 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
692 SvPVX_const((data)->longest_fixed), \
693 (IV)((data)->offset_fixed), \
694 ((data)->longest && \
695 (data)->longest==&((data)->longest_float)) ? "*" : "", \
696 SvPVX_const((data)->longest_float), \
697 (IV)((data)->offset_float_min), \
698 (IV)((data)->offset_float_max) \
700 PerlIO_printf(Perl_debug_log,"\n"); \
703 /* Mark that we cannot extend a found fixed substring at this point.
704 Update the longest found anchored substring and the longest found
705 floating substrings if needed. */
708 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
710 const STRLEN l = CHR_SVLEN(data->last_found);
711 const STRLEN old_l = CHR_SVLEN(*data->longest);
712 GET_RE_DEBUG_FLAGS_DECL;
714 PERL_ARGS_ASSERT_SCAN_COMMIT;
716 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
717 SvSetMagicSV(*data->longest, data->last_found);
718 if (*data->longest == data->longest_fixed) {
719 data->offset_fixed = l ? data->last_start_min : data->pos_min;
720 if (data->flags & SF_BEFORE_EOL)
722 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
724 data->flags &= ~SF_FIX_BEFORE_EOL;
725 data->minlen_fixed=minlenp;
726 data->lookbehind_fixed=0;
728 else { /* *data->longest == data->longest_float */
729 data->offset_float_min = l ? data->last_start_min : data->pos_min;
730 data->offset_float_max = (l
731 ? data->last_start_max
732 : data->pos_min + data->pos_delta);
733 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
734 data->offset_float_max = I32_MAX;
735 if (data->flags & SF_BEFORE_EOL)
737 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
739 data->flags &= ~SF_FL_BEFORE_EOL;
740 data->minlen_float=minlenp;
741 data->lookbehind_float=0;
744 SvCUR_set(data->last_found, 0);
746 SV * const sv = data->last_found;
747 if (SvUTF8(sv) && SvMAGICAL(sv)) {
748 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
754 data->flags &= ~SF_BEFORE_EOL;
755 DEBUG_STUDYDATA("commit: ",data,0);
758 /* These macros set, clear and test whether the synthetic start class ('ssc',
759 * given by the parameter) matches an empty string (EOS). This uses the
760 * 'next_off' field in the node, to save a bit in the flags field. The ssc
761 * stands alone, so there is never a next_off, so this field is otherwise
762 * unused. The EOS information is used only for compilation, but theoretically
763 * it could be passed on to the execution code. This could be used to store
764 * more than one bit of information, but only this one is currently used. */
765 #define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
766 #define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
767 #define TEST_SSC_EOS(node) cBOOL((node)->next_off)
769 /* Can match anything (initialization) */
771 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
773 PERL_ARGS_ASSERT_CL_ANYTHING;
775 ANYOF_BITMAP_SETALL(cl);
776 cl->flags = ANYOF_UNICODE_ALL;
779 /* If any portion of the regex is to operate under locale rules,
780 * initialization includes it. The reason this isn't done for all regexes
781 * is that the optimizer was written under the assumption that locale was
782 * all-or-nothing. Given the complexity and lack of documentation in the
783 * optimizer, and that there are inadequate test cases for locale, so many
784 * parts of it may not work properly, it is safest to avoid locale unless
786 if (RExC_contains_locale) {
787 ANYOF_CLASS_SETALL(cl); /* /l uses class */
788 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
791 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
795 /* Can match anything (initialization) */
797 S_cl_is_anything(const struct regnode_charclass_class *cl)
801 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
803 for (value = 0; value < ANYOF_MAX; value += 2)
804 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
806 if (!(cl->flags & ANYOF_UNICODE_ALL))
808 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
813 /* Can match anything (initialization) */
815 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
817 PERL_ARGS_ASSERT_CL_INIT;
819 Zero(cl, 1, struct regnode_charclass_class);
821 cl_anything(pRExC_state, cl);
822 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
825 /* These two functions currently do the exact same thing */
826 #define cl_init_zero S_cl_init
828 /* 'AND' a given class with another one. Can create false positives. 'cl'
829 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
830 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
832 S_cl_and(struct regnode_charclass_class *cl,
833 const struct regnode_charclass_class *and_with)
835 PERL_ARGS_ASSERT_CL_AND;
837 assert(PL_regkind[and_with->type] == ANYOF);
839 /* I (khw) am not sure all these restrictions are necessary XXX */
840 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
841 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
842 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
843 && !(and_with->flags & ANYOF_LOC_FOLD)
844 && !(cl->flags & ANYOF_LOC_FOLD)) {
847 if (and_with->flags & ANYOF_INVERT)
848 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
849 cl->bitmap[i] &= ~and_with->bitmap[i];
851 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
852 cl->bitmap[i] &= and_with->bitmap[i];
853 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
855 if (and_with->flags & ANYOF_INVERT) {
857 /* Here, the and'ed node is inverted. Get the AND of the flags that
858 * aren't affected by the inversion. Those that are affected are
859 * handled individually below */
860 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
861 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
862 cl->flags |= affected_flags;
864 /* We currently don't know how to deal with things that aren't in the
865 * bitmap, but we know that the intersection is no greater than what
866 * is already in cl, so let there be false positives that get sorted
867 * out after the synthetic start class succeeds, and the node is
868 * matched for real. */
870 /* The inversion of these two flags indicate that the resulting
871 * intersection doesn't have them */
872 if (and_with->flags & ANYOF_UNICODE_ALL) {
873 cl->flags &= ~ANYOF_UNICODE_ALL;
875 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
876 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
879 else { /* and'd node is not inverted */
880 U8 outside_bitmap_but_not_utf8; /* Temp variable */
882 if (! ANYOF_NONBITMAP(and_with)) {
884 /* Here 'and_with' doesn't match anything outside the bitmap
885 * (except possibly ANYOF_UNICODE_ALL), which means the
886 * intersection can't either, except for ANYOF_UNICODE_ALL, in
887 * which case we don't know what the intersection is, but it's no
888 * greater than what cl already has, so can just leave it alone,
889 * with possible false positives */
890 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
891 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
892 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
895 else if (! ANYOF_NONBITMAP(cl)) {
897 /* Here, 'and_with' does match something outside the bitmap, and cl
898 * doesn't have a list of things to match outside the bitmap. If
899 * cl can match all code points above 255, the intersection will
900 * be those above-255 code points that 'and_with' matches. If cl
901 * can't match all Unicode code points, it means that it can't
902 * match anything outside the bitmap (since the 'if' that got us
903 * into this block tested for that), so we leave the bitmap empty.
905 if (cl->flags & ANYOF_UNICODE_ALL) {
906 ARG_SET(cl, ARG(and_with));
908 /* and_with's ARG may match things that don't require UTF8.
909 * And now cl's will too, in spite of this being an 'and'. See
910 * the comments below about the kludge */
911 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
915 /* Here, both 'and_with' and cl match something outside the
916 * bitmap. Currently we do not do the intersection, so just match
917 * whatever cl had at the beginning. */
921 /* Take the intersection of the two sets of flags. However, the
922 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
923 * kludge around the fact that this flag is not treated like the others
924 * which are initialized in cl_anything(). The way the optimizer works
925 * is that the synthetic start class (SSC) is initialized to match
926 * anything, and then the first time a real node is encountered, its
927 * values are AND'd with the SSC's with the result being the values of
928 * the real node. However, there are paths through the optimizer where
929 * the AND never gets called, so those initialized bits are set
930 * inappropriately, which is not usually a big deal, as they just cause
931 * false positives in the SSC, which will just mean a probably
932 * imperceptible slow down in execution. However this bit has a
933 * higher false positive consequence in that it can cause utf8.pm,
934 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
935 * bigger slowdown and also causes significant extra memory to be used.
936 * In order to prevent this, the code now takes a different tack. The
937 * bit isn't set unless some part of the regular expression needs it,
938 * but once set it won't get cleared. This means that these extra
939 * modules won't get loaded unless there was some path through the
940 * pattern that would have required them anyway, and so any false
941 * positives that occur by not ANDing them out when they could be
942 * aren't as severe as they would be if we treated this bit like all
944 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
945 & ANYOF_NONBITMAP_NON_UTF8;
946 cl->flags &= and_with->flags;
947 cl->flags |= outside_bitmap_but_not_utf8;
951 /* 'OR' a given class with another one. Can create false positives. 'cl'
952 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
953 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
955 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
957 PERL_ARGS_ASSERT_CL_OR;
959 if (or_with->flags & ANYOF_INVERT) {
961 /* Here, the or'd node is to be inverted. This means we take the
962 * complement of everything not in the bitmap, but currently we don't
963 * know what that is, so give up and match anything */
964 if (ANYOF_NONBITMAP(or_with)) {
965 cl_anything(pRExC_state, cl);
968 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
969 * <= (B1 | !B2) | (CL1 | !CL2)
970 * which is wasteful if CL2 is small, but we ignore CL2:
971 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
972 * XXXX Can we handle case-fold? Unclear:
973 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
974 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
976 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
977 && !(or_with->flags & ANYOF_LOC_FOLD)
978 && !(cl->flags & ANYOF_LOC_FOLD) ) {
981 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
982 cl->bitmap[i] |= ~or_with->bitmap[i];
983 } /* XXXX: logic is complicated otherwise */
985 cl_anything(pRExC_state, cl);
988 /* And, we can just take the union of the flags that aren't affected
989 * by the inversion */
990 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
992 /* For the remaining flags:
993 ANYOF_UNICODE_ALL and inverted means to not match anything above
994 255, which means that the union with cl should just be
995 what cl has in it, so can ignore this flag
996 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
997 is 127-255 to match them, but then invert that, so the
998 union with cl should just be what cl has in it, so can
1001 } else { /* 'or_with' is not inverted */
1002 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1003 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1004 && (!(or_with->flags & ANYOF_LOC_FOLD)
1005 || (cl->flags & ANYOF_LOC_FOLD)) ) {
1008 /* OR char bitmap and class bitmap separately */
1009 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1010 cl->bitmap[i] |= or_with->bitmap[i];
1011 ANYOF_CLASS_OR(or_with, cl);
1013 else { /* XXXX: logic is complicated, leave it along for a moment. */
1014 cl_anything(pRExC_state, cl);
1017 if (ANYOF_NONBITMAP(or_with)) {
1019 /* Use the added node's outside-the-bit-map match if there isn't a
1020 * conflict. If there is a conflict (both nodes match something
1021 * outside the bitmap, but what they match outside is not the same
1022 * pointer, and hence not easily compared until XXX we extend
1023 * inversion lists this far), give up and allow the start class to
1024 * match everything outside the bitmap. If that stuff is all above
1025 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1026 if (! ANYOF_NONBITMAP(cl)) {
1027 ARG_SET(cl, ARG(or_with));
1029 else if (ARG(cl) != ARG(or_with)) {
1031 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1032 cl_anything(pRExC_state, cl);
1035 cl->flags |= ANYOF_UNICODE_ALL;
1040 /* Take the union */
1041 cl->flags |= or_with->flags;
1045 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1046 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1047 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1048 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1053 dump_trie(trie,widecharmap,revcharmap)
1054 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1055 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1057 These routines dump out a trie in a somewhat readable format.
1058 The _interim_ variants are used for debugging the interim
1059 tables that are used to generate the final compressed
1060 representation which is what dump_trie expects.
1062 Part of the reason for their existence is to provide a form
1063 of documentation as to how the different representations function.
1068 Dumps the final compressed table form of the trie to Perl_debug_log.
1069 Used for debugging make_trie().
1073 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1074 AV *revcharmap, U32 depth)
1077 SV *sv=sv_newmortal();
1078 int colwidth= widecharmap ? 6 : 4;
1080 GET_RE_DEBUG_FLAGS_DECL;
1082 PERL_ARGS_ASSERT_DUMP_TRIE;
1084 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1085 (int)depth * 2 + 2,"",
1086 "Match","Base","Ofs" );
1088 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1089 SV ** const tmp = av_fetch( revcharmap, state, 0);
1091 PerlIO_printf( Perl_debug_log, "%*s",
1093 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1094 PL_colors[0], PL_colors[1],
1095 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1096 PERL_PV_ESCAPE_FIRSTCHAR
1101 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1102 (int)depth * 2 + 2,"");
1104 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1105 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1106 PerlIO_printf( Perl_debug_log, "\n");
1108 for( state = 1 ; state < trie->statecount ; state++ ) {
1109 const U32 base = trie->states[ state ].trans.base;
1111 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1113 if ( trie->states[ state ].wordnum ) {
1114 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1116 PerlIO_printf( Perl_debug_log, "%6s", "" );
1119 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1124 while( ( base + ofs < trie->uniquecharcount ) ||
1125 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1126 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1129 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1131 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1132 if ( ( base + ofs >= trie->uniquecharcount ) &&
1133 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1134 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1136 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1138 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1140 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1144 PerlIO_printf( Perl_debug_log, "]");
1147 PerlIO_printf( Perl_debug_log, "\n" );
1149 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1150 for (word=1; word <= trie->wordcount; word++) {
1151 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1152 (int)word, (int)(trie->wordinfo[word].prev),
1153 (int)(trie->wordinfo[word].len));
1155 PerlIO_printf(Perl_debug_log, "\n" );
1158 Dumps a fully constructed but uncompressed trie in list form.
1159 List tries normally only are used for construction when the number of
1160 possible chars (trie->uniquecharcount) is very high.
1161 Used for debugging make_trie().
1164 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1165 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1169 SV *sv=sv_newmortal();
1170 int colwidth= widecharmap ? 6 : 4;
1171 GET_RE_DEBUG_FLAGS_DECL;
1173 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1175 /* print out the table precompression. */
1176 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1177 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1178 "------:-----+-----------------\n" );
1180 for( state=1 ; state < next_alloc ; state ++ ) {
1183 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1184 (int)depth * 2 + 2,"", (UV)state );
1185 if ( ! trie->states[ state ].wordnum ) {
1186 PerlIO_printf( Perl_debug_log, "%5s| ","");
1188 PerlIO_printf( Perl_debug_log, "W%4x| ",
1189 trie->states[ state ].wordnum
1192 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1193 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1195 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1197 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1198 PL_colors[0], PL_colors[1],
1199 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1200 PERL_PV_ESCAPE_FIRSTCHAR
1202 TRIE_LIST_ITEM(state,charid).forid,
1203 (UV)TRIE_LIST_ITEM(state,charid).newstate
1206 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1207 (int)((depth * 2) + 14), "");
1210 PerlIO_printf( Perl_debug_log, "\n");
1215 Dumps a fully constructed but uncompressed trie in table form.
1216 This is the normal DFA style state transition table, with a few
1217 twists to facilitate compression later.
1218 Used for debugging make_trie().
1221 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1222 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1227 SV *sv=sv_newmortal();
1228 int colwidth= widecharmap ? 6 : 4;
1229 GET_RE_DEBUG_FLAGS_DECL;
1231 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1234 print out the table precompression so that we can do a visual check
1235 that they are identical.
1238 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1240 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1241 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1243 PerlIO_printf( Perl_debug_log, "%*s",
1245 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1246 PL_colors[0], PL_colors[1],
1247 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1248 PERL_PV_ESCAPE_FIRSTCHAR
1254 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1256 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1257 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1260 PerlIO_printf( Perl_debug_log, "\n" );
1262 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1264 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1265 (int)depth * 2 + 2,"",
1266 (UV)TRIE_NODENUM( state ) );
1268 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1269 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1271 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1273 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1275 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1276 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1278 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1279 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1287 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1288 startbranch: the first branch in the whole branch sequence
1289 first : start branch of sequence of branch-exact nodes.
1290 May be the same as startbranch
1291 last : Thing following the last branch.
1292 May be the same as tail.
1293 tail : item following the branch sequence
1294 count : words in the sequence
1295 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1296 depth : indent depth
1298 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1300 A trie is an N'ary tree where the branches are determined by digital
1301 decomposition of the key. IE, at the root node you look up the 1st character and
1302 follow that branch repeat until you find the end of the branches. Nodes can be
1303 marked as "accepting" meaning they represent a complete word. Eg:
1307 would convert into the following structure. Numbers represent states, letters
1308 following numbers represent valid transitions on the letter from that state, if
1309 the number is in square brackets it represents an accepting state, otherwise it
1310 will be in parenthesis.
1312 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1316 (1) +-i->(6)-+-s->[7]
1318 +-s->(3)-+-h->(4)-+-e->[5]
1320 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1322 This shows that when matching against the string 'hers' we will begin at state 1
1323 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1324 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1325 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1326 single traverse. We store a mapping from accepting to state to which word was
1327 matched, and then when we have multiple possibilities we try to complete the
1328 rest of the regex in the order in which they occured in the alternation.
1330 The only prior NFA like behaviour that would be changed by the TRIE support is
1331 the silent ignoring of duplicate alternations which are of the form:
1333 / (DUPE|DUPE) X? (?{ ... }) Y /x
1335 Thus EVAL blocks following a trie may be called a different number of times with
1336 and without the optimisation. With the optimisations dupes will be silently
1337 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1338 the following demonstrates:
1340 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1342 which prints out 'word' three times, but
1344 'words'=~/(word|word|word)(?{ print $1 })S/
1346 which doesnt print it out at all. This is due to other optimisations kicking in.
1348 Example of what happens on a structural level:
1350 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1352 1: CURLYM[1] {1,32767}(18)
1363 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1364 and should turn into:
1366 1: CURLYM[1] {1,32767}(18)
1368 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1376 Cases where tail != last would be like /(?foo|bar)baz/:
1386 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1387 and would end up looking like:
1390 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1397 d = uvuni_to_utf8_flags(d, uv, 0);
1399 is the recommended Unicode-aware way of saying
1404 #define TRIE_STORE_REVCHAR(val) \
1407 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1408 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1409 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1410 SvCUR_set(zlopp, kapow - flrbbbbb); \
1413 av_push(revcharmap, zlopp); \
1415 char ooooff = (char)val; \
1416 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1420 #define TRIE_READ_CHAR STMT_START { \
1423 /* if it is UTF then it is either already folded, or does not need folding */ \
1424 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1426 else if (folder == PL_fold_latin1) { \
1427 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1428 if ( foldlen > 0 ) { \
1429 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1435 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1436 skiplen = UNISKIP(uvc); \
1437 foldlen -= skiplen; \
1438 scan = foldbuf + skiplen; \
1441 /* raw data, will be folded later if needed */ \
1449 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1450 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1451 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1452 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1454 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1455 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1456 TRIE_LIST_CUR( state )++; \
1459 #define TRIE_LIST_NEW(state) STMT_START { \
1460 Newxz( trie->states[ state ].trans.list, \
1461 4, reg_trie_trans_le ); \
1462 TRIE_LIST_CUR( state ) = 1; \
1463 TRIE_LIST_LEN( state ) = 4; \
1466 #define TRIE_HANDLE_WORD(state) STMT_START { \
1467 U16 dupe= trie->states[ state ].wordnum; \
1468 regnode * const noper_next = regnext( noper ); \
1471 /* store the word for dumping */ \
1473 if (OP(noper) != NOTHING) \
1474 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1476 tmp = newSVpvn_utf8( "", 0, UTF ); \
1477 av_push( trie_words, tmp ); \
1481 trie->wordinfo[curword].prev = 0; \
1482 trie->wordinfo[curword].len = wordlen; \
1483 trie->wordinfo[curword].accept = state; \
1485 if ( noper_next < tail ) { \
1487 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1488 trie->jump[curword] = (U16)(noper_next - convert); \
1490 jumper = noper_next; \
1492 nextbranch= regnext(cur); \
1496 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1497 /* chain, so that when the bits of chain are later */\
1498 /* linked together, the dups appear in the chain */\
1499 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1500 trie->wordinfo[dupe].prev = curword; \
1502 /* we haven't inserted this word yet. */ \
1503 trie->states[ state ].wordnum = curword; \
1508 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1509 ( ( base + charid >= ucharcount \
1510 && base + charid < ubound \
1511 && state == trie->trans[ base - ucharcount + charid ].check \
1512 && trie->trans[ base - ucharcount + charid ].next ) \
1513 ? trie->trans[ base - ucharcount + charid ].next \
1514 : ( state==1 ? special : 0 ) \
1518 #define MADE_JUMP_TRIE 2
1519 #define MADE_EXACT_TRIE 4
1522 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1525 /* first pass, loop through and scan words */
1526 reg_trie_data *trie;
1527 HV *widecharmap = NULL;
1528 AV *revcharmap = newAV();
1530 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1535 regnode *jumper = NULL;
1536 regnode *nextbranch = NULL;
1537 regnode *convert = NULL;
1538 U32 *prev_states; /* temp array mapping each state to previous one */
1539 /* we just use folder as a flag in utf8 */
1540 const U8 * folder = NULL;
1543 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1544 AV *trie_words = NULL;
1545 /* along with revcharmap, this only used during construction but both are
1546 * useful during debugging so we store them in the struct when debugging.
1549 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1550 STRLEN trie_charcount=0;
1552 SV *re_trie_maxbuff;
1553 GET_RE_DEBUG_FLAGS_DECL;
1555 PERL_ARGS_ASSERT_MAKE_TRIE;
1557 PERL_UNUSED_ARG(depth);
1564 case EXACTFU_TRICKYFOLD:
1565 case EXACTFU: folder = PL_fold_latin1; break;
1566 case EXACTF: folder = PL_fold; break;
1567 case EXACTFL: folder = PL_fold_locale; break;
1568 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1571 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1573 trie->startstate = 1;
1574 trie->wordcount = word_count;
1575 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1576 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1578 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1579 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1580 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1583 trie_words = newAV();
1586 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1587 if (!SvIOK(re_trie_maxbuff)) {
1588 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1590 DEBUG_TRIE_COMPILE_r({
1591 PerlIO_printf( Perl_debug_log,
1592 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1593 (int)depth * 2 + 2, "",
1594 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1595 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1599 /* Find the node we are going to overwrite */
1600 if ( first == startbranch && OP( last ) != BRANCH ) {
1601 /* whole branch chain */
1604 /* branch sub-chain */
1605 convert = NEXTOPER( first );
1608 /* -- First loop and Setup --
1610 We first traverse the branches and scan each word to determine if it
1611 contains widechars, and how many unique chars there are, this is
1612 important as we have to build a table with at least as many columns as we
1615 We use an array of integers to represent the character codes 0..255
1616 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1617 native representation of the character value as the key and IV's for the
1620 *TODO* If we keep track of how many times each character is used we can
1621 remap the columns so that the table compression later on is more
1622 efficient in terms of memory by ensuring the most common value is in the
1623 middle and the least common are on the outside. IMO this would be better
1624 than a most to least common mapping as theres a decent chance the most
1625 common letter will share a node with the least common, meaning the node
1626 will not be compressible. With a middle is most common approach the worst
1627 case is when we have the least common nodes twice.
1631 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1632 regnode *noper = NEXTOPER( cur );
1633 const U8 *uc = (U8*)STRING( noper );
1634 const U8 *e = uc + STR_LEN( noper );
1636 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1638 const U8 *scan = (U8*)NULL;
1639 U32 wordlen = 0; /* required init */
1641 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1643 if (OP(noper) == NOTHING) {
1644 regnode *noper_next= regnext(noper);
1645 if (noper_next != tail && OP(noper_next) == flags) {
1647 uc= (U8*)STRING(noper);
1648 e= uc + STR_LEN(noper);
1649 trie->minlen= STR_LEN(noper);
1656 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1657 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1658 regardless of encoding */
1659 if (OP( noper ) == EXACTFU_SS) {
1660 /* false positives are ok, so just set this */
1661 TRIE_BITMAP_SET(trie,0xDF);
1664 for ( ; uc < e ; uc += len ) {
1665 TRIE_CHARCOUNT(trie)++;
1670 U8 folded= folder[ (U8) uvc ];
1671 if ( !trie->charmap[ folded ] ) {
1672 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1673 TRIE_STORE_REVCHAR( folded );
1676 if ( !trie->charmap[ uvc ] ) {
1677 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1678 TRIE_STORE_REVCHAR( uvc );
1681 /* store the codepoint in the bitmap, and its folded
1683 TRIE_BITMAP_SET(trie, uvc);
1685 /* store the folded codepoint */
1686 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1689 /* store first byte of utf8 representation of
1690 variant codepoints */
1691 if (! UNI_IS_INVARIANT(uvc)) {
1692 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1695 set_bit = 0; /* We've done our bit :-) */
1700 widecharmap = newHV();
1702 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1705 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1707 if ( !SvTRUE( *svpp ) ) {
1708 sv_setiv( *svpp, ++trie->uniquecharcount );
1709 TRIE_STORE_REVCHAR(uvc);
1713 if( cur == first ) {
1714 trie->minlen = chars;
1715 trie->maxlen = chars;
1716 } else if (chars < trie->minlen) {
1717 trie->minlen = chars;
1718 } else if (chars > trie->maxlen) {
1719 trie->maxlen = chars;
1721 if (OP( noper ) == EXACTFU_SS) {
1722 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1723 if (trie->minlen > 1)
1726 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1727 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1728 * - We assume that any such sequence might match a 2 byte string */
1729 if (trie->minlen > 2 )
1733 } /* end first pass */
1734 DEBUG_TRIE_COMPILE_r(
1735 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1736 (int)depth * 2 + 2,"",
1737 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1738 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1739 (int)trie->minlen, (int)trie->maxlen )
1743 We now know what we are dealing with in terms of unique chars and
1744 string sizes so we can calculate how much memory a naive
1745 representation using a flat table will take. If it's over a reasonable
1746 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1747 conservative but potentially much slower representation using an array
1750 At the end we convert both representations into the same compressed
1751 form that will be used in regexec.c for matching with. The latter
1752 is a form that cannot be used to construct with but has memory
1753 properties similar to the list form and access properties similar
1754 to the table form making it both suitable for fast searches and
1755 small enough that its feasable to store for the duration of a program.
1757 See the comment in the code where the compressed table is produced
1758 inplace from the flat tabe representation for an explanation of how
1759 the compression works.
1764 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1767 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1769 Second Pass -- Array Of Lists Representation
1771 Each state will be represented by a list of charid:state records
1772 (reg_trie_trans_le) the first such element holds the CUR and LEN
1773 points of the allocated array. (See defines above).
1775 We build the initial structure using the lists, and then convert
1776 it into the compressed table form which allows faster lookups
1777 (but cant be modified once converted).
1780 STRLEN transcount = 1;
1782 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1783 "%*sCompiling trie using list compiler\n",
1784 (int)depth * 2 + 2, ""));
1786 trie->states = (reg_trie_state *)
1787 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1788 sizeof(reg_trie_state) );
1792 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1794 regnode *noper = NEXTOPER( cur );
1795 U8 *uc = (U8*)STRING( noper );
1796 const U8 *e = uc + STR_LEN( noper );
1797 U32 state = 1; /* required init */
1798 U16 charid = 0; /* sanity init */
1799 U8 *scan = (U8*)NULL; /* sanity init */
1800 STRLEN foldlen = 0; /* required init */
1801 U32 wordlen = 0; /* required init */
1802 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1805 if (OP(noper) == NOTHING) {
1806 regnode *noper_next= regnext(noper);
1807 if (noper_next != tail && OP(noper_next) == flags) {
1809 uc= (U8*)STRING(noper);
1810 e= uc + STR_LEN(noper);
1814 if (OP(noper) != NOTHING) {
1815 for ( ; uc < e ; uc += len ) {
1820 charid = trie->charmap[ uvc ];
1822 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1826 charid=(U16)SvIV( *svpp );
1829 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1836 if ( !trie->states[ state ].trans.list ) {
1837 TRIE_LIST_NEW( state );
1839 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1840 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1841 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1846 newstate = next_alloc++;
1847 prev_states[newstate] = state;
1848 TRIE_LIST_PUSH( state, charid, newstate );
1853 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1857 TRIE_HANDLE_WORD(state);
1859 } /* end second pass */
1861 /* next alloc is the NEXT state to be allocated */
1862 trie->statecount = next_alloc;
1863 trie->states = (reg_trie_state *)
1864 PerlMemShared_realloc( trie->states,
1866 * sizeof(reg_trie_state) );
1868 /* and now dump it out before we compress it */
1869 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1870 revcharmap, next_alloc,
1874 trie->trans = (reg_trie_trans *)
1875 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1882 for( state=1 ; state < next_alloc ; state ++ ) {
1886 DEBUG_TRIE_COMPILE_MORE_r(
1887 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1891 if (trie->states[state].trans.list) {
1892 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1896 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1897 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1898 if ( forid < minid ) {
1900 } else if ( forid > maxid ) {
1904 if ( transcount < tp + maxid - minid + 1) {
1906 trie->trans = (reg_trie_trans *)
1907 PerlMemShared_realloc( trie->trans,
1909 * sizeof(reg_trie_trans) );
1910 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1912 base = trie->uniquecharcount + tp - minid;
1913 if ( maxid == minid ) {
1915 for ( ; zp < tp ; zp++ ) {
1916 if ( ! trie->trans[ zp ].next ) {
1917 base = trie->uniquecharcount + zp - minid;
1918 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1919 trie->trans[ zp ].check = state;
1925 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1926 trie->trans[ tp ].check = state;
1931 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1932 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1933 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1934 trie->trans[ tid ].check = state;
1936 tp += ( maxid - minid + 1 );
1938 Safefree(trie->states[ state ].trans.list);
1941 DEBUG_TRIE_COMPILE_MORE_r(
1942 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1945 trie->states[ state ].trans.base=base;
1947 trie->lasttrans = tp + 1;
1951 Second Pass -- Flat Table Representation.
1953 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1954 We know that we will need Charcount+1 trans at most to store the data
1955 (one row per char at worst case) So we preallocate both structures
1956 assuming worst case.
1958 We then construct the trie using only the .next slots of the entry
1961 We use the .check field of the first entry of the node temporarily to
1962 make compression both faster and easier by keeping track of how many non
1963 zero fields are in the node.
1965 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1968 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1969 number representing the first entry of the node, and state as a
1970 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1971 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1972 are 2 entrys per node. eg:
1980 The table is internally in the right hand, idx form. However as we also
1981 have to deal with the states array which is indexed by nodenum we have to
1982 use TRIE_NODENUM() to convert.
1985 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1986 "%*sCompiling trie using table compiler\n",
1987 (int)depth * 2 + 2, ""));
1989 trie->trans = (reg_trie_trans *)
1990 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1991 * trie->uniquecharcount + 1,
1992 sizeof(reg_trie_trans) );
1993 trie->states = (reg_trie_state *)
1994 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1995 sizeof(reg_trie_state) );
1996 next_alloc = trie->uniquecharcount + 1;
1999 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2001 regnode *noper = NEXTOPER( cur );
2002 const U8 *uc = (U8*)STRING( noper );
2003 const U8 *e = uc + STR_LEN( noper );
2005 U32 state = 1; /* required init */
2007 U16 charid = 0; /* sanity init */
2008 U32 accept_state = 0; /* sanity init */
2009 U8 *scan = (U8*)NULL; /* sanity init */
2011 STRLEN foldlen = 0; /* required init */
2012 U32 wordlen = 0; /* required init */
2014 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2016 if (OP(noper) == NOTHING) {
2017 regnode *noper_next= regnext(noper);
2018 if (noper_next != tail && OP(noper_next) == flags) {
2020 uc= (U8*)STRING(noper);
2021 e= uc + STR_LEN(noper);
2025 if ( OP(noper) != NOTHING ) {
2026 for ( ; uc < e ; uc += len ) {
2031 charid = trie->charmap[ uvc ];
2033 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2034 charid = svpp ? (U16)SvIV(*svpp) : 0;
2038 if ( !trie->trans[ state + charid ].next ) {
2039 trie->trans[ state + charid ].next = next_alloc;
2040 trie->trans[ state ].check++;
2041 prev_states[TRIE_NODENUM(next_alloc)]
2042 = TRIE_NODENUM(state);
2043 next_alloc += trie->uniquecharcount;
2045 state = trie->trans[ state + charid ].next;
2047 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2049 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2052 accept_state = TRIE_NODENUM( state );
2053 TRIE_HANDLE_WORD(accept_state);
2055 } /* end second pass */
2057 /* and now dump it out before we compress it */
2058 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2060 next_alloc, depth+1));
2064 * Inplace compress the table.*
2066 For sparse data sets the table constructed by the trie algorithm will
2067 be mostly 0/FAIL transitions or to put it another way mostly empty.
2068 (Note that leaf nodes will not contain any transitions.)
2070 This algorithm compresses the tables by eliminating most such
2071 transitions, at the cost of a modest bit of extra work during lookup:
2073 - Each states[] entry contains a .base field which indicates the
2074 index in the state[] array wheres its transition data is stored.
2076 - If .base is 0 there are no valid transitions from that node.
2078 - If .base is nonzero then charid is added to it to find an entry in
2081 -If trans[states[state].base+charid].check!=state then the
2082 transition is taken to be a 0/Fail transition. Thus if there are fail
2083 transitions at the front of the node then the .base offset will point
2084 somewhere inside the previous nodes data (or maybe even into a node
2085 even earlier), but the .check field determines if the transition is
2089 The following process inplace converts the table to the compressed
2090 table: We first do not compress the root node 1,and mark all its
2091 .check pointers as 1 and set its .base pointer as 1 as well. This
2092 allows us to do a DFA construction from the compressed table later,
2093 and ensures that any .base pointers we calculate later are greater
2096 - We set 'pos' to indicate the first entry of the second node.
2098 - We then iterate over the columns of the node, finding the first and
2099 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2100 and set the .check pointers accordingly, and advance pos
2101 appropriately and repreat for the next node. Note that when we copy
2102 the next pointers we have to convert them from the original
2103 NODEIDX form to NODENUM form as the former is not valid post
2106 - If a node has no transitions used we mark its base as 0 and do not
2107 advance the pos pointer.
2109 - If a node only has one transition we use a second pointer into the
2110 structure to fill in allocated fail transitions from other states.
2111 This pointer is independent of the main pointer and scans forward
2112 looking for null transitions that are allocated to a state. When it
2113 finds one it writes the single transition into the "hole". If the
2114 pointer doesnt find one the single transition is appended as normal.
2116 - Once compressed we can Renew/realloc the structures to release the
2119 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2120 specifically Fig 3.47 and the associated pseudocode.
2124 const U32 laststate = TRIE_NODENUM( next_alloc );
2127 trie->statecount = laststate;
2129 for ( state = 1 ; state < laststate ; state++ ) {
2131 const U32 stateidx = TRIE_NODEIDX( state );
2132 const U32 o_used = trie->trans[ stateidx ].check;
2133 U32 used = trie->trans[ stateidx ].check;
2134 trie->trans[ stateidx ].check = 0;
2136 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2137 if ( flag || trie->trans[ stateidx + charid ].next ) {
2138 if ( trie->trans[ stateidx + charid ].next ) {
2140 for ( ; zp < pos ; zp++ ) {
2141 if ( ! trie->trans[ zp ].next ) {
2145 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2146 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2147 trie->trans[ zp ].check = state;
2148 if ( ++zp > pos ) pos = zp;
2155 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2157 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2158 trie->trans[ pos ].check = state;
2163 trie->lasttrans = pos + 1;
2164 trie->states = (reg_trie_state *)
2165 PerlMemShared_realloc( trie->states, laststate
2166 * sizeof(reg_trie_state) );
2167 DEBUG_TRIE_COMPILE_MORE_r(
2168 PerlIO_printf( Perl_debug_log,
2169 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2170 (int)depth * 2 + 2,"",
2171 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2174 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2177 } /* end table compress */
2179 DEBUG_TRIE_COMPILE_MORE_r(
2180 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2181 (int)depth * 2 + 2, "",
2182 (UV)trie->statecount,
2183 (UV)trie->lasttrans)
2185 /* resize the trans array to remove unused space */
2186 trie->trans = (reg_trie_trans *)
2187 PerlMemShared_realloc( trie->trans, trie->lasttrans
2188 * sizeof(reg_trie_trans) );
2190 { /* Modify the program and insert the new TRIE node */
2191 U8 nodetype =(U8)(flags & 0xFF);
2195 regnode *optimize = NULL;
2196 #ifdef RE_TRACK_PATTERN_OFFSETS
2199 U32 mjd_nodelen = 0;
2200 #endif /* RE_TRACK_PATTERN_OFFSETS */
2201 #endif /* DEBUGGING */
2203 This means we convert either the first branch or the first Exact,
2204 depending on whether the thing following (in 'last') is a branch
2205 or not and whther first is the startbranch (ie is it a sub part of
2206 the alternation or is it the whole thing.)
2207 Assuming its a sub part we convert the EXACT otherwise we convert
2208 the whole branch sequence, including the first.
2210 /* Find the node we are going to overwrite */
2211 if ( first != startbranch || OP( last ) == BRANCH ) {
2212 /* branch sub-chain */
2213 NEXT_OFF( first ) = (U16)(last - first);
2214 #ifdef RE_TRACK_PATTERN_OFFSETS
2216 mjd_offset= Node_Offset((convert));
2217 mjd_nodelen= Node_Length((convert));
2220 /* whole branch chain */
2222 #ifdef RE_TRACK_PATTERN_OFFSETS
2225 const regnode *nop = NEXTOPER( convert );
2226 mjd_offset= Node_Offset((nop));
2227 mjd_nodelen= Node_Length((nop));
2231 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2232 (int)depth * 2 + 2, "",
2233 (UV)mjd_offset, (UV)mjd_nodelen)
2236 /* But first we check to see if there is a common prefix we can
2237 split out as an EXACT and put in front of the TRIE node. */
2238 trie->startstate= 1;
2239 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2241 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2245 const U32 base = trie->states[ state ].trans.base;
2247 if ( trie->states[state].wordnum )
2250 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2251 if ( ( base + ofs >= trie->uniquecharcount ) &&
2252 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2253 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2255 if ( ++count > 1 ) {
2256 SV **tmp = av_fetch( revcharmap, ofs, 0);
2257 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2258 if ( state == 1 ) break;
2260 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2262 PerlIO_printf(Perl_debug_log,
2263 "%*sNew Start State=%"UVuf" Class: [",
2264 (int)depth * 2 + 2, "",
2267 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2268 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2270 TRIE_BITMAP_SET(trie,*ch);
2272 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2274 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2278 TRIE_BITMAP_SET(trie,*ch);
2280 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2281 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2287 SV **tmp = av_fetch( revcharmap, idx, 0);
2289 char *ch = SvPV( *tmp, len );
2291 SV *sv=sv_newmortal();
2292 PerlIO_printf( Perl_debug_log,
2293 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2294 (int)depth * 2 + 2, "",
2296 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2297 PL_colors[0], PL_colors[1],
2298 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2299 PERL_PV_ESCAPE_FIRSTCHAR
2304 OP( convert ) = nodetype;
2305 str=STRING(convert);
2308 STR_LEN(convert) += len;
2314 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2319 trie->prefixlen = (state-1);
2321 regnode *n = convert+NODE_SZ_STR(convert);
2322 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2323 trie->startstate = state;
2324 trie->minlen -= (state - 1);
2325 trie->maxlen -= (state - 1);
2327 /* At least the UNICOS C compiler choked on this
2328 * being argument to DEBUG_r(), so let's just have
2331 #ifdef PERL_EXT_RE_BUILD
2337 regnode *fix = convert;
2338 U32 word = trie->wordcount;
2340 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2341 while( ++fix < n ) {
2342 Set_Node_Offset_Length(fix, 0, 0);
2345 SV ** const tmp = av_fetch( trie_words, word, 0 );
2347 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2348 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2350 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2358 NEXT_OFF(convert) = (U16)(tail - convert);
2359 DEBUG_r(optimize= n);
2365 if ( trie->maxlen ) {
2366 NEXT_OFF( convert ) = (U16)(tail - convert);
2367 ARG_SET( convert, data_slot );
2368 /* Store the offset to the first unabsorbed branch in
2369 jump[0], which is otherwise unused by the jump logic.
2370 We use this when dumping a trie and during optimisation. */
2372 trie->jump[0] = (U16)(nextbranch - convert);
2374 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2375 * and there is a bitmap
2376 * and the first "jump target" node we found leaves enough room
2377 * then convert the TRIE node into a TRIEC node, with the bitmap
2378 * embedded inline in the opcode - this is hypothetically faster.
2380 if ( !trie->states[trie->startstate].wordnum
2382 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2384 OP( convert ) = TRIEC;
2385 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2386 PerlMemShared_free(trie->bitmap);
2389 OP( convert ) = TRIE;
2391 /* store the type in the flags */
2392 convert->flags = nodetype;
2396 + regarglen[ OP( convert ) ];
2398 /* XXX We really should free up the resource in trie now,
2399 as we won't use them - (which resources?) dmq */
2401 /* needed for dumping*/
2402 DEBUG_r(if (optimize) {
2403 regnode *opt = convert;
2405 while ( ++opt < optimize) {
2406 Set_Node_Offset_Length(opt,0,0);
2409 Try to clean up some of the debris left after the
2412 while( optimize < jumper ) {
2413 mjd_nodelen += Node_Length((optimize));
2414 OP( optimize ) = OPTIMIZED;
2415 Set_Node_Offset_Length(optimize,0,0);
2418 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2420 } /* end node insert */
2422 /* Finish populating the prev field of the wordinfo array. Walk back
2423 * from each accept state until we find another accept state, and if
2424 * so, point the first word's .prev field at the second word. If the
2425 * second already has a .prev field set, stop now. This will be the
2426 * case either if we've already processed that word's accept state,
2427 * or that state had multiple words, and the overspill words were
2428 * already linked up earlier.
2435 for (word=1; word <= trie->wordcount; word++) {
2437 if (trie->wordinfo[word].prev)
2439 state = trie->wordinfo[word].accept;
2441 state = prev_states[state];
2444 prev = trie->states[state].wordnum;
2448 trie->wordinfo[word].prev = prev;
2450 Safefree(prev_states);
2454 /* and now dump out the compressed format */
2455 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2457 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2459 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2460 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2462 SvREFCNT_dec_NN(revcharmap);
2466 : trie->startstate>1
2472 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2474 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2476 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2477 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2480 We find the fail state for each state in the trie, this state is the longest proper
2481 suffix of the current state's 'word' that is also a proper prefix of another word in our
2482 trie. State 1 represents the word '' and is thus the default fail state. This allows
2483 the DFA not to have to restart after its tried and failed a word at a given point, it
2484 simply continues as though it had been matching the other word in the first place.
2486 'abcdgu'=~/abcdefg|cdgu/
2487 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2488 fail, which would bring us to the state representing 'd' in the second word where we would
2489 try 'g' and succeed, proceeding to match 'cdgu'.
2491 /* add a fail transition */
2492 const U32 trie_offset = ARG(source);
2493 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2495 const U32 ucharcount = trie->uniquecharcount;
2496 const U32 numstates = trie->statecount;
2497 const U32 ubound = trie->lasttrans + ucharcount;
2501 U32 base = trie->states[ 1 ].trans.base;
2504 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2505 GET_RE_DEBUG_FLAGS_DECL;
2507 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2509 PERL_UNUSED_ARG(depth);
2513 ARG_SET( stclass, data_slot );
2514 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2515 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2516 aho->trie=trie_offset;
2517 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2518 Copy( trie->states, aho->states, numstates, reg_trie_state );
2519 Newxz( q, numstates, U32);
2520 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2523 /* initialize fail[0..1] to be 1 so that we always have
2524 a valid final fail state */
2525 fail[ 0 ] = fail[ 1 ] = 1;
2527 for ( charid = 0; charid < ucharcount ; charid++ ) {
2528 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2530 q[ q_write ] = newstate;
2531 /* set to point at the root */
2532 fail[ q[ q_write++ ] ]=1;
2535 while ( q_read < q_write) {
2536 const U32 cur = q[ q_read++ % numstates ];
2537 base = trie->states[ cur ].trans.base;
2539 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2540 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2542 U32 fail_state = cur;
2545 fail_state = fail[ fail_state ];
2546 fail_base = aho->states[ fail_state ].trans.base;
2547 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2549 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2550 fail[ ch_state ] = fail_state;
2551 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2553 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2555 q[ q_write++ % numstates] = ch_state;
2559 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2560 when we fail in state 1, this allows us to use the
2561 charclass scan to find a valid start char. This is based on the principle
2562 that theres a good chance the string being searched contains lots of stuff
2563 that cant be a start char.
2565 fail[ 0 ] = fail[ 1 ] = 0;
2566 DEBUG_TRIE_COMPILE_r({
2567 PerlIO_printf(Perl_debug_log,
2568 "%*sStclass Failtable (%"UVuf" states): 0",
2569 (int)(depth * 2), "", (UV)numstates
2571 for( q_read=1; q_read<numstates; q_read++ ) {
2572 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2574 PerlIO_printf(Perl_debug_log, "\n");
2577 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2582 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2583 * These need to be revisited when a newer toolchain becomes available.
2585 #if defined(__sparc64__) && defined(__GNUC__)
2586 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2587 # undef SPARC64_GCC_WORKAROUND
2588 # define SPARC64_GCC_WORKAROUND 1
2592 #define DEBUG_PEEP(str,scan,depth) \
2593 DEBUG_OPTIMISE_r({if (scan){ \
2594 SV * const mysv=sv_newmortal(); \
2595 regnode *Next = regnext(scan); \
2596 regprop(RExC_rx, mysv, scan); \
2597 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2598 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2599 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2603 /* The below joins as many adjacent EXACTish nodes as possible into a single
2604 * one. The regop may be changed if the node(s) contain certain sequences that
2605 * require special handling. The joining is only done if:
2606 * 1) there is room in the current conglomerated node to entirely contain the
2608 * 2) they are the exact same node type
2610 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2611 * these get optimized out
2613 * If a node is to match under /i (folded), the number of characters it matches
2614 * can be different than its character length if it contains a multi-character
2615 * fold. *min_subtract is set to the total delta of the input nodes.
2617 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2618 * and contains LATIN SMALL LETTER SHARP S
2620 * This is as good a place as any to discuss the design of handling these
2621 * multi-character fold sequences. It's been wrong in Perl for a very long
2622 * time. There are three code points in Unicode whose multi-character folds
2623 * were long ago discovered to mess things up. The previous designs for
2624 * dealing with these involved assigning a special node for them. This
2625 * approach doesn't work, as evidenced by this example:
2626 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2627 * Both these fold to "sss", but if the pattern is parsed to create a node that
2628 * would match just the \xDF, it won't be able to handle the case where a
2629 * successful match would have to cross the node's boundary. The new approach
2630 * that hopefully generally solves the problem generates an EXACTFU_SS node
2633 * It turns out that there are problems with all multi-character folds, and not
2634 * just these three. Now the code is general, for all such cases, but the
2635 * three still have some special handling. The approach taken is:
2636 * 1) This routine examines each EXACTFish node that could contain multi-
2637 * character fold sequences. It returns in *min_subtract how much to
2638 * subtract from the the actual length of the string to get a real minimum
2639 * match length; it is 0 if there are no multi-char folds. This delta is
2640 * used by the caller to adjust the min length of the match, and the delta
2641 * between min and max, so that the optimizer doesn't reject these
2642 * possibilities based on size constraints.
2643 * 2) Certain of these sequences require special handling by the trie code,
2644 * so, if found, this code changes the joined node type to special ops:
2645 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2646 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2647 * is used for an EXACTFU node that contains at least one "ss" sequence in
2648 * it. For non-UTF-8 patterns and strings, this is the only case where
2649 * there is a possible fold length change. That means that a regular
2650 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2651 * with length changes, and so can be processed faster. regexec.c takes
2652 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2653 * pre-folded by regcomp.c. This saves effort in regex matching.
2654 * However, the pre-folding isn't done for non-UTF8 patterns because the
2655 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2656 * down by forcing the pattern into UTF8 unless necessary. Also what
2657 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2658 * possibilities for the non-UTF8 patterns are quite simple, except for
2659 * the sharp s. All the ones that don't involve a UTF-8 target string are
2660 * members of a fold-pair, and arrays are set up for all of them so that
2661 * the other member of the pair can be found quickly. Code elsewhere in
2662 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2663 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2664 * described in the next item.
2665 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2666 * 'ss' or not is not knowable at compile time. It will match iff the
2667 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2668 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2669 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
2670 * described in item 3). An assumption that the optimizer part of
2671 * regexec.c (probably unwittingly) makes is that a character in the
2672 * pattern corresponds to at most a single character in the target string.
2673 * (And I do mean character, and not byte here, unlike other parts of the
2674 * documentation that have never been updated to account for multibyte
2675 * Unicode.) This assumption is wrong only in this case, as all other
2676 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2677 * virtue of having this file pre-fold UTF-8 patterns. I'm
2678 * reluctant to try to change this assumption, so instead the code punts.
2679 * This routine examines EXACTF nodes for the sharp s, and returns a
2680 * boolean indicating whether or not the node is an EXACTF node that
2681 * contains a sharp s. When it is true, the caller sets a flag that later
2682 * causes the optimizer in this file to not set values for the floating
2683 * and fixed string lengths, and thus avoids the optimizer code in
2684 * regexec.c that makes the invalid assumption. Thus, there is no
2685 * optimization based on string lengths for EXACTF nodes that contain the
2686 * sharp s. This only happens for /id rules (which means the pattern
2690 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2691 if (PL_regkind[OP(scan)] == EXACT) \
2692 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2695 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) {
2696 /* Merge several consecutive EXACTish nodes into one. */
2697 regnode *n = regnext(scan);
2699 regnode *next = scan + NODE_SZ_STR(scan);
2703 regnode *stop = scan;
2704 GET_RE_DEBUG_FLAGS_DECL;
2706 PERL_UNUSED_ARG(depth);
2709 PERL_ARGS_ASSERT_JOIN_EXACT;
2710 #ifndef EXPERIMENTAL_INPLACESCAN
2711 PERL_UNUSED_ARG(flags);
2712 PERL_UNUSED_ARG(val);
2714 DEBUG_PEEP("join",scan,depth);
2716 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2717 * EXACT ones that are mergeable to the current one. */
2719 && (PL_regkind[OP(n)] == NOTHING
2720 || (stringok && OP(n) == OP(scan)))
2722 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2725 if (OP(n) == TAIL || n > next)
2727 if (PL_regkind[OP(n)] == NOTHING) {
2728 DEBUG_PEEP("skip:",n,depth);
2729 NEXT_OFF(scan) += NEXT_OFF(n);
2730 next = n + NODE_STEP_REGNODE;
2737 else if (stringok) {
2738 const unsigned int oldl = STR_LEN(scan);
2739 regnode * const nnext = regnext(n);
2741 /* XXX I (khw) kind of doubt that this works on platforms where
2742 * U8_MAX is above 255 because of lots of other assumptions */
2743 /* Don't join if the sum can't fit into a single node */
2744 if (oldl + STR_LEN(n) > U8_MAX)
2747 DEBUG_PEEP("merg",n,depth);
2750 NEXT_OFF(scan) += NEXT_OFF(n);
2751 STR_LEN(scan) += STR_LEN(n);
2752 next = n + NODE_SZ_STR(n);
2753 /* Now we can overwrite *n : */
2754 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2762 #ifdef EXPERIMENTAL_INPLACESCAN
2763 if (flags && !NEXT_OFF(n)) {
2764 DEBUG_PEEP("atch", val, depth);
2765 if (reg_off_by_arg[OP(n)]) {
2766 ARG_SET(n, val - n);
2769 NEXT_OFF(n) = val - n;
2777 *has_exactf_sharp_s = FALSE;
2779 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2780 * can now analyze for sequences of problematic code points. (Prior to
2781 * this final joining, sequences could have been split over boundaries, and
2782 * hence missed). The sequences only happen in folding, hence for any
2783 * non-EXACT EXACTish node */
2784 if (OP(scan) != EXACT) {
2785 const U8 * const s0 = (U8*) STRING(scan);
2787 const U8 * const s_end = s0 + STR_LEN(scan);
2789 /* One pass is made over the node's string looking for all the
2790 * possibilities. to avoid some tests in the loop, there are two main
2791 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2795 /* Examine the string for a multi-character fold sequence. UTF-8
2796 * patterns have all characters pre-folded by the time this code is
2798 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2799 length sequence we are looking for is 2 */
2802 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2803 if (! len) { /* Not a multi-char fold: get next char */
2808 /* Nodes with 'ss' require special handling, except for EXACTFL
2809 * and EXACTFA for which there is no multi-char fold to this */
2810 if (len == 2 && *s == 's' && *(s+1) == 's'
2811 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2814 OP(scan) = EXACTFU_SS;
2817 else if (len == 6 /* len is the same in both ASCII and EBCDIC for these */
2818 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2819 COMBINING_DIAERESIS_UTF8
2820 COMBINING_ACUTE_ACCENT_UTF8,
2822 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2823 COMBINING_DIAERESIS_UTF8
2824 COMBINING_ACUTE_ACCENT_UTF8,
2829 /* These two folds require special handling by trie's, so
2830 * change the node type to indicate this. If EXACTFA and
2831 * EXACTFL were ever to be handled by trie's, this would
2832 * have to be changed. If this node has already been
2833 * changed to EXACTFU_SS in this loop, leave it as is. (I
2834 * (khw) think it doesn't matter in regexec.c for UTF
2835 * patterns, but no need to change it */
2836 if (OP(scan) == EXACTFU) {
2837 OP(scan) = EXACTFU_TRICKYFOLD;
2841 else { /* Here is a generic multi-char fold. */
2842 const U8* multi_end = s + len;
2844 /* Count how many characters in it. In the case of /l and
2845 * /aa, no folds which contain ASCII code points are
2846 * allowed, so check for those, and skip if found. (In
2847 * EXACTFL, no folds are allowed to any Latin1 code point,
2848 * not just ASCII. But there aren't any of these
2849 * currently, nor ever likely, so don't take the time to
2850 * test for them. The code that generates the
2851 * is_MULTI_foo() macros croaks should one actually get put
2852 * into Unicode .) */
2853 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2854 count = utf8_length(s, multi_end);
2858 while (s < multi_end) {
2861 goto next_iteration;
2871 /* The delta is how long the sequence is minus 1 (1 is how long
2872 * the character that folds to the sequence is) */
2873 *min_subtract += count - 1;
2877 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2879 /* Here, the pattern is not UTF-8. Look for the multi-char folds
2880 * that are all ASCII. As in the above case, EXACTFL and EXACTFA
2881 * nodes can't have multi-char folds to this range (and there are
2882 * no existing ones in the upper latin1 range). In the EXACTF
2883 * case we look also for the sharp s, which can be in the final
2884 * position. Otherwise we can stop looking 1 byte earlier because
2885 * have to find at least two characters for a multi-fold */
2886 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2888 /* The below is perhaps overboard, but this allows us to save a
2889 * test each time through the loop at the expense of a mask. This
2890 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2891 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2892 * are 64. This uses an exclusive 'or' to find that bit and then
2893 * inverts it to form a mask, with just a single 0, in the bit
2894 * position where 'S' and 's' differ. */
2895 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2896 const U8 s_masked = 's' & S_or_s_mask;
2899 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2900 if (! len) { /* Not a multi-char fold. */
2901 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2903 *has_exactf_sharp_s = TRUE;
2910 && ((*s & S_or_s_mask) == s_masked)
2911 && ((*(s+1) & S_or_s_mask) == s_masked))
2914 /* EXACTF nodes need to know that the minimum length
2915 * changed so that a sharp s in the string can match this
2916 * ss in the pattern, but they remain EXACTF nodes, as they
2917 * won't match this unless the target string is is UTF-8,
2918 * which we don't know until runtime */
2919 if (OP(scan) != EXACTF) {
2920 OP(scan) = EXACTFU_SS;
2924 *min_subtract += len - 1;
2931 /* Allow dumping but overwriting the collection of skipped
2932 * ops and/or strings with fake optimized ops */
2933 n = scan + NODE_SZ_STR(scan);
2941 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2945 /* REx optimizer. Converts nodes into quicker variants "in place".
2946 Finds fixed substrings. */
2948 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2949 to the position after last scanned or to NULL. */
2951 #define INIT_AND_WITHP \
2952 assert(!and_withp); \
2953 Newx(and_withp,1,struct regnode_charclass_class); \
2954 SAVEFREEPV(and_withp)
2956 /* this is a chain of data about sub patterns we are processing that
2957 need to be handled separately/specially in study_chunk. Its so
2958 we can simulate recursion without losing state. */
2960 typedef struct scan_frame {
2961 regnode *last; /* last node to process in this frame */
2962 regnode *next; /* next node to process when last is reached */
2963 struct scan_frame *prev; /*previous frame*/
2964 I32 stop; /* what stopparen do we use */
2968 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2971 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
2972 I32 *minlenp, I32 *deltap,
2977 struct regnode_charclass_class *and_withp,
2978 U32 flags, U32 depth)
2979 /* scanp: Start here (read-write). */
2980 /* deltap: Write maxlen-minlen here. */
2981 /* last: Stop before this one. */
2982 /* data: string data about the pattern */
2983 /* stopparen: treat close N as END */
2984 /* recursed: which subroutines have we recursed into */
2985 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
2988 I32 min = 0; /* There must be at least this number of characters to match */
2990 regnode *scan = *scanp, *next;
2992 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
2993 int is_inf_internal = 0; /* The studied chunk is infinite */
2994 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2995 scan_data_t data_fake;
2996 SV *re_trie_maxbuff = NULL;
2997 regnode *first_non_open = scan;
2998 I32 stopmin = I32_MAX;
2999 scan_frame *frame = NULL;
3000 GET_RE_DEBUG_FLAGS_DECL;
3002 PERL_ARGS_ASSERT_STUDY_CHUNK;
3005 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3009 while (first_non_open && OP(first_non_open) == OPEN)
3010 first_non_open=regnext(first_non_open);
3015 while ( scan && OP(scan) != END && scan < last ){
3016 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3017 node length to get a real minimum (because
3018 the folded version may be shorter) */
3019 bool has_exactf_sharp_s = FALSE;
3020 /* Peephole optimizer: */
3021 DEBUG_STUDYDATA("Peep:", data,depth);
3022 DEBUG_PEEP("Peep",scan,depth);
3024 /* Its not clear to khw or hv why this is done here, and not in the
3025 * clauses that deal with EXACT nodes. khw's guess is that it's
3026 * because of a previous design */
3027 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3029 /* Follow the next-chain of the current node and optimize
3030 away all the NOTHINGs from it. */
3031 if (OP(scan) != CURLYX) {
3032 const int max = (reg_off_by_arg[OP(scan)]
3034 /* I32 may be smaller than U16 on CRAYs! */
3035 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3036 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3040 /* Skip NOTHING and LONGJMP. */
3041 while ((n = regnext(n))
3042 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3043 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3044 && off + noff < max)
3046 if (reg_off_by_arg[OP(scan)])
3049 NEXT_OFF(scan) = off;
3054 /* The principal pseudo-switch. Cannot be a switch, since we
3055 look into several different things. */
3056 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3057 || OP(scan) == IFTHEN) {
3058 next = regnext(scan);
3060 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3062 if (OP(next) == code || code == IFTHEN) {
3063 /* NOTE - There is similar code to this block below for handling
3064 TRIE nodes on a re-study. If you change stuff here check there
3066 I32 max1 = 0, min1 = I32_MAX, num = 0;
3067 struct regnode_charclass_class accum;
3068 regnode * const startbranch=scan;
3070 if (flags & SCF_DO_SUBSTR)
3071 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3072 if (flags & SCF_DO_STCLASS)
3073 cl_init_zero(pRExC_state, &accum);
3075 while (OP(scan) == code) {
3076 I32 deltanext, minnext, f = 0, fake;
3077 struct regnode_charclass_class this_class;
3080 data_fake.flags = 0;
3082 data_fake.whilem_c = data->whilem_c;
3083 data_fake.last_closep = data->last_closep;
3086 data_fake.last_closep = &fake;
3088 data_fake.pos_delta = delta;
3089 next = regnext(scan);
3090 scan = NEXTOPER(scan);
3092 scan = NEXTOPER(scan);
3093 if (flags & SCF_DO_STCLASS) {
3094 cl_init(pRExC_state, &this_class);
3095 data_fake.start_class = &this_class;
3096 f = SCF_DO_STCLASS_AND;
3098 if (flags & SCF_WHILEM_VISITED_POS)
3099 f |= SCF_WHILEM_VISITED_POS;
3101 /* we suppose the run is continuous, last=next...*/
3102 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3104 stopparen, recursed, NULL, f,depth+1);
3107 if (max1 < minnext + deltanext)
3108 max1 = minnext + deltanext;
3109 if (deltanext == I32_MAX)
3110 is_inf = is_inf_internal = 1;
3112 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3114 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3115 if ( stopmin > minnext)
3116 stopmin = min + min1;
3117 flags &= ~SCF_DO_SUBSTR;
3119 data->flags |= SCF_SEEN_ACCEPT;
3122 if (data_fake.flags & SF_HAS_EVAL)
3123 data->flags |= SF_HAS_EVAL;
3124 data->whilem_c = data_fake.whilem_c;
3126 if (flags & SCF_DO_STCLASS)
3127 cl_or(pRExC_state, &accum, &this_class);
3129 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3131 if (flags & SCF_DO_SUBSTR) {
3132 data->pos_min += min1;
3133 data->pos_delta += max1 - min1;
3134 if (max1 != min1 || is_inf)
3135 data->longest = &(data->longest_float);
3138 delta += max1 - min1;
3139 if (flags & SCF_DO_STCLASS_OR) {
3140 cl_or(pRExC_state, data->start_class, &accum);
3142 cl_and(data->start_class, and_withp);
3143 flags &= ~SCF_DO_STCLASS;
3146 else if (flags & SCF_DO_STCLASS_AND) {
3148 cl_and(data->start_class, &accum);
3149 flags &= ~SCF_DO_STCLASS;
3152 /* Switch to OR mode: cache the old value of
3153 * data->start_class */
3155 StructCopy(data->start_class, and_withp,
3156 struct regnode_charclass_class);
3157 flags &= ~SCF_DO_STCLASS_AND;
3158 StructCopy(&accum, data->start_class,
3159 struct regnode_charclass_class);
3160 flags |= SCF_DO_STCLASS_OR;
3161 SET_SSC_EOS(data->start_class);
3165 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3168 Assuming this was/is a branch we are dealing with: 'scan' now
3169 points at the item that follows the branch sequence, whatever
3170 it is. We now start at the beginning of the sequence and look
3177 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3179 If we can find such a subsequence we need to turn the first
3180 element into a trie and then add the subsequent branch exact
3181 strings to the trie.
3185 1. patterns where the whole set of branches can be converted.
3187 2. patterns where only a subset can be converted.
3189 In case 1 we can replace the whole set with a single regop
3190 for the trie. In case 2 we need to keep the start and end
3193 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3194 becomes BRANCH TRIE; BRANCH X;
3196 There is an additional case, that being where there is a
3197 common prefix, which gets split out into an EXACT like node
3198 preceding the TRIE node.
3200 If x(1..n)==tail then we can do a simple trie, if not we make
3201 a "jump" trie, such that when we match the appropriate word
3202 we "jump" to the appropriate tail node. Essentially we turn
3203 a nested if into a case structure of sorts.
3208 if (!re_trie_maxbuff) {
3209 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3210 if (!SvIOK(re_trie_maxbuff))
3211 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3213 if ( SvIV(re_trie_maxbuff)>=0 ) {
3215 regnode *first = (regnode *)NULL;
3216 regnode *last = (regnode *)NULL;
3217 regnode *tail = scan;
3222 SV * const mysv = sv_newmortal(); /* for dumping */
3224 /* var tail is used because there may be a TAIL
3225 regop in the way. Ie, the exacts will point to the
3226 thing following the TAIL, but the last branch will
3227 point at the TAIL. So we advance tail. If we
3228 have nested (?:) we may have to move through several
3232 while ( OP( tail ) == TAIL ) {
3233 /* this is the TAIL generated by (?:) */
3234 tail = regnext( tail );
3238 DEBUG_TRIE_COMPILE_r({
3239 regprop(RExC_rx, mysv, tail );
3240 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3241 (int)depth * 2 + 2, "",
3242 "Looking for TRIE'able sequences. Tail node is: ",
3243 SvPV_nolen_const( mysv )
3249 Step through the branches
3250 cur represents each branch,
3251 noper is the first thing to be matched as part of that branch
3252 noper_next is the regnext() of that node.
3254 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3255 via a "jump trie" but we also support building with NOJUMPTRIE,
3256 which restricts the trie logic to structures like /FOO|BAR/.
3258 If noper is a trieable nodetype then the branch is a possible optimization
3259 target. If we are building under NOJUMPTRIE then we require that noper_next
3260 is the same as scan (our current position in the regex program).
3262 Once we have two or more consecutive such branches we can create a
3263 trie of the EXACT's contents and stitch it in place into the program.
3265 If the sequence represents all of the branches in the alternation we
3266 replace the entire thing with a single TRIE node.
3268 Otherwise when it is a subsequence we need to stitch it in place and
3269 replace only the relevant branches. This means the first branch has
3270 to remain as it is used by the alternation logic, and its next pointer,
3271 and needs to be repointed at the item on the branch chain following
3272 the last branch we have optimized away.
3274 This could be either a BRANCH, in which case the subsequence is internal,
3275 or it could be the item following the branch sequence in which case the
3276 subsequence is at the end (which does not necessarily mean the first node
3277 is the start of the alternation).
3279 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3282 ----------------+-----------
3286 EXACTFU_SS | EXACTFU
3287 EXACTFU_TRICKYFOLD | EXACTFU
3292 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3293 ( EXACT == (X) ) ? EXACT : \
3294 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3297 /* dont use tail as the end marker for this traverse */
3298 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3299 regnode * const noper = NEXTOPER( cur );
3300 U8 noper_type = OP( noper );
3301 U8 noper_trietype = TRIE_TYPE( noper_type );
3302 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3303 regnode * const noper_next = regnext( noper );
3304 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3305 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3308 DEBUG_TRIE_COMPILE_r({
3309 regprop(RExC_rx, mysv, cur);
3310 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3311 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3313 regprop(RExC_rx, mysv, noper);
3314 PerlIO_printf( Perl_debug_log, " -> %s",
3315 SvPV_nolen_const(mysv));
3318 regprop(RExC_rx, mysv, noper_next );
3319 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3320 SvPV_nolen_const(mysv));
3322 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3323 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3324 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3328 /* Is noper a trieable nodetype that can be merged with the
3329 * current trie (if there is one)? */
3333 ( noper_trietype == NOTHING)
3334 || ( trietype == NOTHING )
3335 || ( trietype == noper_trietype )
3338 && noper_next == tail
3342 /* Handle mergable triable node
3343 * Either we are the first node in a new trieable sequence,
3344 * in which case we do some bookkeeping, otherwise we update
3345 * the end pointer. */
3348 if ( noper_trietype == NOTHING ) {
3349 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3350 regnode * const noper_next = regnext( noper );
3351 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3352 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3355 if ( noper_next_trietype ) {
3356 trietype = noper_next_trietype;
3357 } else if (noper_next_type) {
3358 /* a NOTHING regop is 1 regop wide. We need at least two
3359 * for a trie so we can't merge this in */
3363 trietype = noper_trietype;
3366 if ( trietype == NOTHING )
3367 trietype = noper_trietype;
3372 } /* end handle mergable triable node */
3374 /* handle unmergable node -
3375 * noper may either be a triable node which can not be tried
3376 * together with the current trie, or a non triable node */
3378 /* If last is set and trietype is not NOTHING then we have found
3379 * at least two triable branch sequences in a row of a similar
3380 * trietype so we can turn them into a trie. If/when we
3381 * allow NOTHING to start a trie sequence this condition will be
3382 * required, and it isn't expensive so we leave it in for now. */
3383 if ( trietype && trietype != NOTHING )
3384 make_trie( pRExC_state,
3385 startbranch, first, cur, tail, count,
3386 trietype, depth+1 );
3387 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3391 && noper_next == tail
3394 /* noper is triable, so we can start a new trie sequence */
3397 trietype = noper_trietype;
3399 /* if we already saw a first but the current node is not triable then we have
3400 * to reset the first information. */
3405 } /* end handle unmergable node */
3406 } /* loop over branches */
3407 DEBUG_TRIE_COMPILE_r({
3408 regprop(RExC_rx, mysv, cur);
3409 PerlIO_printf( Perl_debug_log,
3410 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3411 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3414 if ( last && trietype ) {
3415 if ( trietype != NOTHING ) {
3416 /* the last branch of the sequence was part of a trie,
3417 * so we have to construct it here outside of the loop
3419 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3420 #ifdef TRIE_STUDY_OPT
3421 if ( ((made == MADE_EXACT_TRIE &&
3422 startbranch == first)
3423 || ( first_non_open == first )) &&
3425 flags |= SCF_TRIE_RESTUDY;
3426 if ( startbranch == first
3429 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3434 /* at this point we know whatever we have is a NOTHING sequence/branch
3435 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3437 if ( startbranch == first ) {
3439 /* the entire thing is a NOTHING sequence, something like this:
3440 * (?:|) So we can turn it into a plain NOTHING op. */
3441 DEBUG_TRIE_COMPILE_r({
3442 regprop(RExC_rx, mysv, cur);
3443 PerlIO_printf( Perl_debug_log,
3444 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3445 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3448 OP(startbranch)= NOTHING;
3449 NEXT_OFF(startbranch)= tail - startbranch;
3450 for ( opt= startbranch + 1; opt < tail ; opt++ )
3454 } /* end if ( last) */
3455 } /* TRIE_MAXBUF is non zero */
3460 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3461 scan = NEXTOPER(NEXTOPER(scan));
3462 } else /* single branch is optimized. */
3463 scan = NEXTOPER(scan);
3465 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3466 scan_frame *newframe = NULL;
3471 if (OP(scan) != SUSPEND) {
3472 /* set the pointer */
3473 if (OP(scan) == GOSUB) {
3475 RExC_recurse[ARG2L(scan)] = scan;
3476 start = RExC_open_parens[paren-1];
3477 end = RExC_close_parens[paren-1];
3480 start = RExC_rxi->program + 1;
3484 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3485 SAVEFREEPV(recursed);
3487 if (!PAREN_TEST(recursed,paren+1)) {
3488 PAREN_SET(recursed,paren+1);
3489 Newx(newframe,1,scan_frame);
3491 if (flags & SCF_DO_SUBSTR) {
3492 SCAN_COMMIT(pRExC_state,data,minlenp);
3493 data->longest = &(data->longest_float);
3495 is_inf = is_inf_internal = 1;
3496 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3497 cl_anything(pRExC_state, data->start_class);
3498 flags &= ~SCF_DO_STCLASS;
3501 Newx(newframe,1,scan_frame);
3504 end = regnext(scan);
3509 SAVEFREEPV(newframe);
3510 newframe->next = regnext(scan);
3511 newframe->last = last;
3512 newframe->stop = stopparen;
3513 newframe->prev = frame;
3523 else if (OP(scan) == EXACT) {
3524 I32 l = STR_LEN(scan);
3527 const U8 * const s = (U8*)STRING(scan);
3528 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3529 l = utf8_length(s, s + l);
3531 uc = *((U8*)STRING(scan));
3534 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3535 /* The code below prefers earlier match for fixed
3536 offset, later match for variable offset. */
3537 if (data->last_end == -1) { /* Update the start info. */
3538 data->last_start_min = data->pos_min;
3539 data->last_start_max = is_inf
3540 ? I32_MAX : data->pos_min + data->pos_delta;
3542 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3544 SvUTF8_on(data->last_found);
3546 SV * const sv = data->last_found;
3547 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3548 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3549 if (mg && mg->mg_len >= 0)
3550 mg->mg_len += utf8_length((U8*)STRING(scan),
3551 (U8*)STRING(scan)+STR_LEN(scan));
3553 data->last_end = data->pos_min + l;
3554 data->pos_min += l; /* As in the first entry. */
3555 data->flags &= ~SF_BEFORE_EOL;
3557 if (flags & SCF_DO_STCLASS_AND) {
3558 /* Check whether it is compatible with what we know already! */
3562 /* If compatible, we or it in below. It is compatible if is
3563 * in the bitmp and either 1) its bit or its fold is set, or 2)
3564 * it's for a locale. Even if there isn't unicode semantics
3565 * here, at runtime there may be because of matching against a
3566 * utf8 string, so accept a possible false positive for
3567 * latin1-range folds */
3569 (!(data->start_class->flags & ANYOF_LOCALE)
3570 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3571 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3572 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3577 ANYOF_CLASS_ZERO(data->start_class);
3578 ANYOF_BITMAP_ZERO(data->start_class);
3580 ANYOF_BITMAP_SET(data->start_class, uc);
3581 else if (uc >= 0x100) {
3584 /* Some Unicode code points fold to the Latin1 range; as
3585 * XXX temporary code, instead of figuring out if this is
3586 * one, just assume it is and set all the start class bits
3587 * that could be some such above 255 code point's fold
3588 * which will generate fals positives. As the code
3589 * elsewhere that does compute the fold settles down, it
3590 * can be extracted out and re-used here */
3591 for (i = 0; i < 256; i++){
3592 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3593 ANYOF_BITMAP_SET(data->start_class, i);
3597 CLEAR_SSC_EOS(data->start_class);
3599 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3601 else if (flags & SCF_DO_STCLASS_OR) {
3602 /* false positive possible if the class is case-folded */
3604 ANYOF_BITMAP_SET(data->start_class, uc);
3606 data->start_class->flags |= ANYOF_UNICODE_ALL;
3607 CLEAR_SSC_EOS(data->start_class);
3608 cl_and(data->start_class, and_withp);
3610 flags &= ~SCF_DO_STCLASS;
3612 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3613 I32 l = STR_LEN(scan);
3614 UV uc = *((U8*)STRING(scan));
3616 /* Search for fixed substrings supports EXACT only. */
3617 if (flags & SCF_DO_SUBSTR) {
3619 SCAN_COMMIT(pRExC_state, data, minlenp);
3622 const U8 * const s = (U8 *)STRING(scan);
3623 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3624 l = utf8_length(s, s + l);
3626 if (has_exactf_sharp_s) {
3627 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3629 min += l - min_subtract;
3631 delta += min_subtract;
3632 if (flags & SCF_DO_SUBSTR) {
3633 data->pos_min += l - min_subtract;
3634 if (data->pos_min < 0) {
3637 data->pos_delta += min_subtract;
3639 data->longest = &(data->longest_float);
3642 if (flags & SCF_DO_STCLASS_AND) {
3643 /* Check whether it is compatible with what we know already! */
3646 (!(data->start_class->flags & ANYOF_LOCALE)
3647 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3648 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3652 ANYOF_CLASS_ZERO(data->start_class);
3653 ANYOF_BITMAP_ZERO(data->start_class);
3655 ANYOF_BITMAP_SET(data->start_class, uc);
3656 CLEAR_SSC_EOS(data->start_class);
3657 if (OP(scan) == EXACTFL) {
3658 /* XXX This set is probably no longer necessary, and
3659 * probably wrong as LOCALE now is on in the initial
3661 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3665 /* Also set the other member of the fold pair. In case
3666 * that unicode semantics is called for at runtime, use
3667 * the full latin1 fold. (Can't do this for locale,
3668 * because not known until runtime) */
3669 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3671 /* All other (EXACTFL handled above) folds except under
3672 * /iaa that include s, S, and sharp_s also may include
3674 if (OP(scan) != EXACTFA) {
3675 if (uc == 's' || uc == 'S') {
3676 ANYOF_BITMAP_SET(data->start_class,
3677 LATIN_SMALL_LETTER_SHARP_S);
3679 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3680 ANYOF_BITMAP_SET(data->start_class, 's');
3681 ANYOF_BITMAP_SET(data->start_class, 'S');
3686 else if (uc >= 0x100) {
3688 for (i = 0; i < 256; i++){
3689 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3690 ANYOF_BITMAP_SET(data->start_class, i);
3695 else if (flags & SCF_DO_STCLASS_OR) {
3696 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3697 /* false positive possible if the class is case-folded.
3698 Assume that the locale settings are the same... */
3700 ANYOF_BITMAP_SET(data->start_class, uc);
3701 if (OP(scan) != EXACTFL) {
3703 /* And set the other member of the fold pair, but
3704 * can't do that in locale because not known until
3706 ANYOF_BITMAP_SET(data->start_class,
3707 PL_fold_latin1[uc]);
3709 /* All folds except under /iaa that include s, S,
3710 * and sharp_s also may include the others */
3711 if (OP(scan) != EXACTFA) {
3712 if (uc == 's' || uc == 'S') {
3713 ANYOF_BITMAP_SET(data->start_class,
3714 LATIN_SMALL_LETTER_SHARP_S);
3716 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3717 ANYOF_BITMAP_SET(data->start_class, 's');
3718 ANYOF_BITMAP_SET(data->start_class, 'S');
3723 CLEAR_SSC_EOS(data->start_class);
3725 cl_and(data->start_class, and_withp);
3727 flags &= ~SCF_DO_STCLASS;
3729 else if (REGNODE_VARIES(OP(scan))) {
3730 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3731 I32 f = flags, pos_before = 0;
3732 regnode * const oscan = scan;
3733 struct regnode_charclass_class this_class;
3734 struct regnode_charclass_class *oclass = NULL;
3735 I32 next_is_eval = 0;
3737 switch (PL_regkind[OP(scan)]) {
3738 case WHILEM: /* End of (?:...)* . */
3739 scan = NEXTOPER(scan);
3742 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3743 next = NEXTOPER(scan);
3744 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3746 maxcount = REG_INFTY;
3747 next = regnext(scan);
3748 scan = NEXTOPER(scan);
3752 if (flags & SCF_DO_SUBSTR)
3757 if (flags & SCF_DO_STCLASS) {
3759 maxcount = REG_INFTY;
3760 next = regnext(scan);
3761 scan = NEXTOPER(scan);
3764 is_inf = is_inf_internal = 1;
3765 scan = regnext(scan);
3766 if (flags & SCF_DO_SUBSTR) {
3767 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3768 data->longest = &(data->longest_float);
3770 goto optimize_curly_tail;
3772 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3773 && (scan->flags == stopparen))
3778 mincount = ARG1(scan);
3779 maxcount = ARG2(scan);
3781 next = regnext(scan);
3782 if (OP(scan) == CURLYX) {
3783 I32 lp = (data ? *(data->last_closep) : 0);
3784 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3786 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3787 next_is_eval = (OP(scan) == EVAL);
3789 if (flags & SCF_DO_SUBSTR) {
3790 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3791 pos_before = data->pos_min;
3795 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3797 data->flags |= SF_IS_INF;
3799 if (flags & SCF_DO_STCLASS) {
3800 cl_init(pRExC_state, &this_class);
3801 oclass = data->start_class;
3802 data->start_class = &this_class;
3803 f |= SCF_DO_STCLASS_AND;
3804 f &= ~SCF_DO_STCLASS_OR;
3806 /* Exclude from super-linear cache processing any {n,m}
3807 regops for which the combination of input pos and regex
3808 pos is not enough information to determine if a match
3811 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3812 regex pos at the \s*, the prospects for a match depend not
3813 only on the input position but also on how many (bar\s*)
3814 repeats into the {4,8} we are. */
3815 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3816 f &= ~SCF_WHILEM_VISITED_POS;
3818 /* This will finish on WHILEM, setting scan, or on NULL: */
3819 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3820 last, data, stopparen, recursed, NULL,
3822 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3824 if (flags & SCF_DO_STCLASS)
3825 data->start_class = oclass;
3826 if (mincount == 0 || minnext == 0) {
3827 if (flags & SCF_DO_STCLASS_OR) {
3828 cl_or(pRExC_state, data->start_class, &this_class);
3830 else if (flags & SCF_DO_STCLASS_AND) {
3831 /* Switch to OR mode: cache the old value of
3832 * data->start_class */
3834 StructCopy(data->start_class, and_withp,
3835 struct regnode_charclass_class);
3836 flags &= ~SCF_DO_STCLASS_AND;
3837 StructCopy(&this_class, data->start_class,
https://perl5.git.perl.org / perl5.git / blob