This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c: Modify some comments
[perl5.git] / regcomp.c
CommitLineData
a0d0e21e
LW
1/* regcomp.c
2 */
3
4/*
4ac71550
TC
5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
6 *
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
a0d0e21e
LW
8 */
9
61296642
DM
10/* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
166f8a29 12 * a regular expression.
e4a054ea
DM
13 *
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.
166f8a29
DM
18 */
19
a687059c
LW
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!
22 */
23
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.
27 */
28
e50aee73 29/* The names of the functions have been changed from regcomp and
3b753521 30 * regexec to pregcomp and pregexec in order to avoid conflicts
e50aee73
AD
31 * with the POSIX routines of the same names.
32*/
33
b9d5759e 34#ifdef PERL_EXT_RE_BUILD
54df2634 35#include "re_top.h"
b81d288d 36#endif
56953603 37
a687059c 38/*
e50aee73 39 * pregcomp and pregexec -- regsub and regerror are not used in perl
a687059c
LW
40 *
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
43 *
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:
47 *
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
50 * from defects in it.
51 *
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
54 *
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
57 *
58 *
59 **** Alterations to Henry's code are...
60 ****
4bb101f2 61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
1129b882
NC
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
a687059c 64 ****
9ef589d8
LW
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.
67
a687059c
LW
68 *
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.
72 */
73#include "EXTERN.h"
864dbfa3 74#define PERL_IN_REGCOMP_C
a687059c 75#include "perl.h"
d06ea78c 76
acfe0abc 77#ifndef PERL_IN_XSUB_RE
d06ea78c
GS
78# include "INTERN.h"
79#endif
c277df42
IZ
80
81#define REG_COMP_C
54df2634
NC
82#ifdef PERL_IN_XSUB_RE
83# include "re_comp.h"
84#else
85# include "regcomp.h"
86#endif
a687059c 87
04e98a4d
AD
88#include "dquote_static.c"
89
d4cce5f1 90#ifdef op
11343788 91#undef op
d4cce5f1 92#endif /* op */
11343788 93
fe14fcc3 94#ifdef MSDOS
7e4e8c89 95# if defined(BUGGY_MSC6)
fe14fcc3 96 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
7e4e8c89 97# pragma optimize("a",off)
fe14fcc3 98 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
7e4e8c89
NC
99# pragma optimize("w",on )
100# endif /* BUGGY_MSC6 */
fe14fcc3
LW
101#endif /* MSDOS */
102
a687059c
LW
103#ifndef STATIC
104#define STATIC static
105#endif
106
830247a4 107typedef struct RExC_state_t {
e2509266 108 U32 flags; /* are we folding, multilining? */
830247a4 109 char *precomp; /* uncompiled string. */
288b8c02 110 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf
YO
111 regexp *rx; /* perl core regexp structure */
112 regexp_internal *rxi; /* internal data for regexp object pprivate field */
fac92740 113 char *start; /* Start of input for compile */
830247a4
IZ
114 char *end; /* End of input for compile */
115 char *parse; /* Input-scan pointer. */
116 I32 whilem_seen; /* number of WHILEM in this expr */
fac92740 117 regnode *emit_start; /* Start of emitted-code area */
3b57cd43 118 regnode *emit_bound; /* First regnode outside of the allocated space */
ffc61ed2 119 regnode *emit; /* Code-emit pointer; &regdummy = don't = compiling */
830247a4
IZ
120 I32 naughty; /* How bad is this pattern? */
121 I32 sawback; /* Did we see \1, ...? */
122 U32 seen;
123 I32 size; /* Code size. */
c74340f9
YO
124 I32 npar; /* Capture buffer count, (OPEN). */
125 I32 cpar; /* Capture buffer count, (CLOSE). */
e2e6a0f1 126 I32 nestroot; /* root parens we are in - used by accept */
830247a4
IZ
127 I32 extralen;
128 I32 seen_zerolen;
129 I32 seen_evals;
40d049e4
YO
130 regnode **open_parens; /* pointers to open parens */
131 regnode **close_parens; /* pointers to close parens */
132 regnode *opend; /* END node in program */
02daf0ab
YO
133 I32 utf8; /* whether the pattern is utf8 or not */
134 I32 orig_utf8; /* whether the pattern was originally in utf8 */
135 /* XXX use this for future optimisation of case
136 * where pattern must be upgraded to utf8. */
e40e74fe
KW
137 I32 uni_semantics; /* If a d charset modifier should use unicode
138 rules, even if the pattern is not in
139 utf8 */
81714fb9 140 HV *paren_names; /* Paren names */
1f1031fe 141
40d049e4
YO
142 regnode **recurse; /* Recurse regops */
143 I32 recurse_count; /* Number of recurse regops */
b57e4118 144 I32 in_lookbehind;
4624b182 145 I32 contains_locale;
bb3f3ed2 146 I32 override_recoding;
830247a4
IZ
147#if ADD_TO_REGEXEC
148 char *starttry; /* -Dr: where regtry was called. */
149#define RExC_starttry (pRExC_state->starttry)
150#endif
3dab1dad 151#ifdef DEBUGGING
be8e71aa 152 const char *lastparse;
3dab1dad 153 I32 lastnum;
1f1031fe 154 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
155#define RExC_lastparse (pRExC_state->lastparse)
156#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 157#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 158#endif
830247a4
IZ
159} RExC_state_t;
160
e2509266 161#define RExC_flags (pRExC_state->flags)
830247a4 162#define RExC_precomp (pRExC_state->precomp)
288b8c02 163#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 164#define RExC_rx (pRExC_state->rx)
f8fc2ecf 165#define RExC_rxi (pRExC_state->rxi)
fac92740 166#define RExC_start (pRExC_state->start)
830247a4
IZ
167#define RExC_end (pRExC_state->end)
168#define RExC_parse (pRExC_state->parse)
169#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237
YO
170#ifdef RE_TRACK_PATTERN_OFFSETS
171#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
172#endif
830247a4 173#define RExC_emit (pRExC_state->emit)
fac92740 174#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 175#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
176#define RExC_naughty (pRExC_state->naughty)
177#define RExC_sawback (pRExC_state->sawback)
178#define RExC_seen (pRExC_state->seen)
179#define RExC_size (pRExC_state->size)
180#define RExC_npar (pRExC_state->npar)
e2e6a0f1 181#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
182#define RExC_extralen (pRExC_state->extralen)
183#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
184#define RExC_seen_evals (pRExC_state->seen_evals)
1aa99e6b 185#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 186#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 187#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
188#define RExC_open_parens (pRExC_state->open_parens)
189#define RExC_close_parens (pRExC_state->close_parens)
190#define RExC_opend (pRExC_state->opend)
81714fb9 191#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
192#define RExC_recurse (pRExC_state->recurse)
193#define RExC_recurse_count (pRExC_state->recurse_count)
b57e4118 194#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 195#define RExC_contains_locale (pRExC_state->contains_locale)
bb3f3ed2 196#define RExC_override_recoding (pRExC_state->override_recoding)
830247a4 197
cde0cee5 198
a687059c
LW
199#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
200#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
201 ((*s) == '{' && regcurly(s)))
a687059c 202
35c8bce7
LW
203#ifdef SPSTART
204#undef SPSTART /* dratted cpp namespace... */
205#endif
a687059c
LW
206/*
207 * Flags to be passed up and down.
208 */
a687059c 209#define WORST 0 /* Worst case. */
a3b492c3 210#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee
KW
211
212/* Simple enough to be STAR/PLUS operand, in an EXACT node must be a single
d7b56a3c 213 * character, and if utf8, must be invariant. Note that this is not the same thing as REGNODE_SIMPLE */
fda99bee 214#define SIMPLE 0x02
a3b492c3
YO
215#define SPSTART 0x04 /* Starts with * or +. */
216#define TRYAGAIN 0x08 /* Weeded out a declaration. */
217#define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
a687059c 218
3dab1dad
YO
219#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
220
07be1b83
YO
221/* whether trie related optimizations are enabled */
222#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
223#define TRIE_STUDY_OPT
786e8c11 224#define FULL_TRIE_STUDY
07be1b83
YO
225#define TRIE_STCLASS
226#endif
1de06328
YO
227
228
40d049e4
YO
229
230#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
231#define PBITVAL(paren) (1 << ((paren) & 7))
232#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
233#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
234#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
235
bbd61b5f
KW
236/* If not already in utf8, do a longjmp back to the beginning */
237#define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
238#define REQUIRE_UTF8 STMT_START { \
239 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
240 } STMT_END
40d049e4 241
1de06328
YO
242/* About scan_data_t.
243
244 During optimisation we recurse through the regexp program performing
245 various inplace (keyhole style) optimisations. In addition study_chunk
246 and scan_commit populate this data structure with information about
247 what strings MUST appear in the pattern. We look for the longest
3b753521 248 string that must appear at a fixed location, and we look for the
1de06328
YO
249 longest string that may appear at a floating location. So for instance
250 in the pattern:
251
252 /FOO[xX]A.*B[xX]BAR/
253
254 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
255 strings (because they follow a .* construct). study_chunk will identify
256 both FOO and BAR as being the longest fixed and floating strings respectively.
257
258 The strings can be composites, for instance
259
260 /(f)(o)(o)/
261
262 will result in a composite fixed substring 'foo'.
263
264 For each string some basic information is maintained:
265
266 - offset or min_offset
267 This is the position the string must appear at, or not before.
268 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
269 characters must match before the string we are searching for.
270 Likewise when combined with minlenp and the length of the string it
1de06328
YO
271 tells us how many characters must appear after the string we have
272 found.
273
274 - max_offset
275 Only used for floating strings. This is the rightmost point that
3b753521 276 the string can appear at. If set to I32 max it indicates that the
1de06328
YO
277 string can occur infinitely far to the right.
278
279 - minlenp
280 A pointer to the minimum length of the pattern that the string
281 was found inside. This is important as in the case of positive
282 lookahead or positive lookbehind we can have multiple patterns
283 involved. Consider
284
285 /(?=FOO).*F/
286
287 The minimum length of the pattern overall is 3, the minimum length
288 of the lookahead part is 3, but the minimum length of the part that
289 will actually match is 1. So 'FOO's minimum length is 3, but the
290 minimum length for the F is 1. This is important as the minimum length
291 is used to determine offsets in front of and behind the string being
292 looked for. Since strings can be composites this is the length of the
486ec47a 293 pattern at the time it was committed with a scan_commit. Note that
1de06328
YO
294 the length is calculated by study_chunk, so that the minimum lengths
295 are not known until the full pattern has been compiled, thus the
296 pointer to the value.
297
298 - lookbehind
299
300 In the case of lookbehind the string being searched for can be
301 offset past the start point of the final matching string.
302 If this value was just blithely removed from the min_offset it would
303 invalidate some of the calculations for how many chars must match
304 before or after (as they are derived from min_offset and minlen and
305 the length of the string being searched for).
306 When the final pattern is compiled and the data is moved from the
307 scan_data_t structure into the regexp structure the information
308 about lookbehind is factored in, with the information that would
309 have been lost precalculated in the end_shift field for the
310 associated string.
311
312 The fields pos_min and pos_delta are used to store the minimum offset
313 and the delta to the maximum offset at the current point in the pattern.
314
315*/
2c2d71f5
JH
316
317typedef struct scan_data_t {
1de06328
YO
318 /*I32 len_min; unused */
319 /*I32 len_delta; unused */
2c2d71f5
JH
320 I32 pos_min;
321 I32 pos_delta;
322 SV *last_found;
1de06328 323 I32 last_end; /* min value, <0 unless valid. */
2c2d71f5
JH
324 I32 last_start_min;
325 I32 last_start_max;
1de06328
YO
326 SV **longest; /* Either &l_fixed, or &l_float. */
327 SV *longest_fixed; /* longest fixed string found in pattern */
328 I32 offset_fixed; /* offset where it starts */
486ec47a 329 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
330 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
331 SV *longest_float; /* longest floating string found in pattern */
332 I32 offset_float_min; /* earliest point in string it can appear */
333 I32 offset_float_max; /* latest point in string it can appear */
486ec47a 334 I32 *minlen_float; /* pointer to the minlen relevant to the string */
1de06328 335 I32 lookbehind_float; /* is the position of the string modified by LB */
2c2d71f5
JH
336 I32 flags;
337 I32 whilem_c;
cb434fcc 338 I32 *last_closep;
653099ff 339 struct regnode_charclass_class *start_class;
2c2d71f5
JH
340} scan_data_t;
341
a687059c 342/*
e50aee73 343 * Forward declarations for pregcomp()'s friends.
a687059c 344 */
a0d0e21e 345
27da23d5 346static const scan_data_t zero_scan_data =
1de06328 347 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
348
349#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
350#define SF_BEFORE_SEOL 0x0001
351#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
352#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
353#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
354
09b7f37c
CB
355#ifdef NO_UNARY_PLUS
356# define SF_FIX_SHIFT_EOL (0+2)
357# define SF_FL_SHIFT_EOL (0+4)
358#else
359# define SF_FIX_SHIFT_EOL (+2)
360# define SF_FL_SHIFT_EOL (+4)
361#endif
c277df42
IZ
362
363#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
364#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
365
366#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
367#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
368#define SF_IS_INF 0x0040
369#define SF_HAS_PAR 0x0080
370#define SF_IN_PAR 0x0100
371#define SF_HAS_EVAL 0x0200
372#define SCF_DO_SUBSTR 0x0400
653099ff
GS
373#define SCF_DO_STCLASS_AND 0x0800
374#define SCF_DO_STCLASS_OR 0x1000
375#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 376#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 377
786e8c11 378#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
e2e6a0f1 379#define SCF_SEEN_ACCEPT 0x8000
07be1b83 380
43fead97 381#define UTF cBOOL(RExC_utf8)
a62b1201
KW
382#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
383#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
384#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
385#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
386#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1
KW
387#define MORE_ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
388#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a62b1201 389
43fead97 390#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 391
ffc61ed2 392#define OOB_UNICODE 12345678
93733859 393#define OOB_NAMEDCLASS -1
b8c5462f 394
a0ed51b3
LW
395#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
396#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
397
8615cb43 398
b45f050a
JF
399/* length of regex to show in messages that don't mark a position within */
400#define RegexLengthToShowInErrorMessages 127
401
402/*
403 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
404 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
405 * op/pragma/warn/regcomp.
406 */
7253e4e3
RK
407#define MARKER1 "<-- HERE" /* marker as it appears in the description */
408#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 409
7253e4e3 410#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
411
412/*
413 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
414 * arg. Show regex, up to a maximum length. If it's too long, chop and add
415 * "...".
416 */
58e23c8d 417#define _FAIL(code) STMT_START { \
bfed75c6 418 const char *ellipses = ""; \
ccb2c380
MP
419 IV len = RExC_end - RExC_precomp; \
420 \
421 if (!SIZE_ONLY) \
288b8c02 422 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
423 if (len > RegexLengthToShowInErrorMessages) { \
424 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
425 len = RegexLengthToShowInErrorMessages - 10; \
426 ellipses = "..."; \
427 } \
58e23c8d 428 code; \
ccb2c380 429} STMT_END
8615cb43 430
58e23c8d
YO
431#define FAIL(msg) _FAIL( \
432 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
433 msg, (int)len, RExC_precomp, ellipses))
434
435#define FAIL2(msg,arg) _FAIL( \
436 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
437 arg, (int)len, RExC_precomp, ellipses))
438
b45f050a 439/*
b45f050a
JF
440 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
441 */
ccb2c380 442#define Simple_vFAIL(m) STMT_START { \
a28509cc 443 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
444 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
445 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
446} STMT_END
b45f050a
JF
447
448/*
449 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
450 */
ccb2c380
MP
451#define vFAIL(m) STMT_START { \
452 if (!SIZE_ONLY) \
288b8c02 453 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
454 Simple_vFAIL(m); \
455} STMT_END
b45f050a
JF
456
457/*
458 * Like Simple_vFAIL(), but accepts two arguments.
459 */
ccb2c380 460#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 461 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
462 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
463 (int)offset, RExC_precomp, RExC_precomp + offset); \
464} STMT_END
b45f050a
JF
465
466/*
467 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
468 */
ccb2c380
MP
469#define vFAIL2(m,a1) STMT_START { \
470 if (!SIZE_ONLY) \
288b8c02 471 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
472 Simple_vFAIL2(m, a1); \
473} STMT_END
b45f050a
JF
474
475
476/*
477 * Like Simple_vFAIL(), but accepts three arguments.
478 */
ccb2c380 479#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 480 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
481 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
482 (int)offset, RExC_precomp, RExC_precomp + offset); \
483} STMT_END
b45f050a
JF
484
485/*
486 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
487 */
ccb2c380
MP
488#define vFAIL3(m,a1,a2) STMT_START { \
489 if (!SIZE_ONLY) \
288b8c02 490 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
ccb2c380
MP
491 Simple_vFAIL3(m, a1, a2); \
492} STMT_END
b45f050a
JF
493
494/*
495 * Like Simple_vFAIL(), but accepts four arguments.
496 */
ccb2c380 497#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 498 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
499 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
500 (int)offset, RExC_precomp, RExC_precomp + offset); \
501} STMT_END
b45f050a 502
668c081a 503#define ckWARNreg(loc,m) STMT_START { \
a28509cc 504 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
505 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
506 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
507} STMT_END
508
668c081a 509#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 510 const IV offset = loc - RExC_precomp; \
d1d15184 511 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
512 m REPORT_LOCATION, \
513 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
514} STMT_END
515
2335b3d3
KW
516#define ckWARN2regdep(loc,m, a1) STMT_START { \
517 const IV offset = loc - RExC_precomp; \
518 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
519 m REPORT_LOCATION, \
520 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
521} STMT_END
522
668c081a 523#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 524 const IV offset = loc - RExC_precomp; \
668c081a 525 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
526 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
527} STMT_END
528
529#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 530 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
531 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
532 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
533} STMT_END
534
668c081a
NC
535#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
536 const IV offset = loc - RExC_precomp; \
537 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
538 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
539} STMT_END
540
ccb2c380 541#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 542 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
543 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
544 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
545} STMT_END
546
668c081a
NC
547#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
548 const IV offset = loc - RExC_precomp; \
549 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
550 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
551} STMT_END
552
ccb2c380 553#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 554 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
555 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
556 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
557} STMT_END
9d1d55b5 558
8615cb43 559
cd439c50 560/* Allow for side effects in s */
ccb2c380
MP
561#define REGC(c,s) STMT_START { \
562 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
563} STMT_END
cd439c50 564
fac92740
MJD
565/* Macros for recording node offsets. 20001227 mjd@plover.com
566 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
567 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
568 * Element 0 holds the number n.
07be1b83 569 * Position is 1 indexed.
fac92740 570 */
7122b237
YO
571#ifndef RE_TRACK_PATTERN_OFFSETS
572#define Set_Node_Offset_To_R(node,byte)
573#define Set_Node_Offset(node,byte)
574#define Set_Cur_Node_Offset
575#define Set_Node_Length_To_R(node,len)
576#define Set_Node_Length(node,len)
577#define Set_Node_Cur_Length(node)
578#define Node_Offset(n)
579#define Node_Length(n)
580#define Set_Node_Offset_Length(node,offset,len)
581#define ProgLen(ri) ri->u.proglen
582#define SetProgLen(ri,x) ri->u.proglen = x
583#else
584#define ProgLen(ri) ri->u.offsets[0]
585#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
586#define Set_Node_Offset_To_R(node,byte) STMT_START { \
587 if (! SIZE_ONLY) { \
588 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 589 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 590 if((node) < 0) { \
551405c4 591 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
592 } else { \
593 RExC_offsets[2*(node)-1] = (byte); \
594 } \
595 } \
596} STMT_END
597
598#define Set_Node_Offset(node,byte) \
599 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
600#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
601
602#define Set_Node_Length_To_R(node,len) STMT_START { \
603 if (! SIZE_ONLY) { \
604 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 605 __LINE__, (int)(node), (int)(len))); \
ccb2c380 606 if((node) < 0) { \
551405c4 607 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
608 } else { \
609 RExC_offsets[2*(node)] = (len); \
610 } \
611 } \
612} STMT_END
613
614#define Set_Node_Length(node,len) \
615 Set_Node_Length_To_R((node)-RExC_emit_start, len)
616#define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
617#define Set_Node_Cur_Length(node) \
618 Set_Node_Length(node, RExC_parse - parse_start)
fac92740
MJD
619
620/* Get offsets and lengths */
621#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
622#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
623
07be1b83
YO
624#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
625 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
626 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
627} STMT_END
7122b237 628#endif
07be1b83
YO
629
630#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
631#define EXPERIMENTAL_INPLACESCAN
f427392e 632#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 633
304ee84b
YO
634#define DEBUG_STUDYDATA(str,data,depth) \
635DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 636 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
637 "%*s" str "Pos:%"IVdf"/%"IVdf \
638 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
639 (int)(depth)*2, "", \
640 (IV)((data)->pos_min), \
641 (IV)((data)->pos_delta), \
304ee84b 642 (UV)((data)->flags), \
1de06328 643 (IV)((data)->whilem_c), \
304ee84b
YO
644 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
645 is_inf ? "INF " : "" \
1de06328
YO
646 ); \
647 if ((data)->last_found) \
648 PerlIO_printf(Perl_debug_log, \
649 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
650 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
651 SvPVX_const((data)->last_found), \
652 (IV)((data)->last_end), \
653 (IV)((data)->last_start_min), \
654 (IV)((data)->last_start_max), \
655 ((data)->longest && \
656 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
657 SvPVX_const((data)->longest_fixed), \
658 (IV)((data)->offset_fixed), \
659 ((data)->longest && \
660 (data)->longest==&((data)->longest_float)) ? "*" : "", \
661 SvPVX_const((data)->longest_float), \
662 (IV)((data)->offset_float_min), \
663 (IV)((data)->offset_float_max) \
664 ); \
665 PerlIO_printf(Perl_debug_log,"\n"); \
666});
667
acfe0abc 668static void clear_re(pTHX_ void *r);
4327152a 669
653099ff 670/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 671 Update the longest found anchored substring and the longest found
653099ff
GS
672 floating substrings if needed. */
673
4327152a 674STATIC void
304ee84b 675S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
c277df42 676{
e1ec3a88
AL
677 const STRLEN l = CHR_SVLEN(data->last_found);
678 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 679 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 680
7918f24d
NC
681 PERL_ARGS_ASSERT_SCAN_COMMIT;
682
c277df42 683 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 684 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
685 if (*data->longest == data->longest_fixed) {
686 data->offset_fixed = l ? data->last_start_min : data->pos_min;
687 if (data->flags & SF_BEFORE_EOL)
b81d288d 688 data->flags
c277df42
IZ
689 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
690 else
691 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 692 data->minlen_fixed=minlenp;
1de06328 693 data->lookbehind_fixed=0;
a0ed51b3 694 }
304ee84b 695 else { /* *data->longest == data->longest_float */
c277df42 696 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
697 data->offset_float_max = (l
698 ? data->last_start_max
c277df42 699 : data->pos_min + data->pos_delta);
304ee84b 700 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
9051bda5 701 data->offset_float_max = I32_MAX;
c277df42 702 if (data->flags & SF_BEFORE_EOL)
b81d288d 703 data->flags
c277df42
IZ
704 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
705 else
706 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
707 data->minlen_float=minlenp;
708 data->lookbehind_float=0;
c277df42
IZ
709 }
710 }
711 SvCUR_set(data->last_found, 0);
0eda9292 712 {
a28509cc 713 SV * const sv = data->last_found;
097eb12c
AL
714 if (SvUTF8(sv) && SvMAGICAL(sv)) {
715 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
716 if (mg)
717 mg->mg_len = 0;
718 }
0eda9292 719 }
c277df42
IZ
720 data->last_end = -1;
721 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 722 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
723}
724
653099ff
GS
725/* Can match anything (initialization) */
726STATIC void
3fffb88a 727S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 728{
7918f24d
NC
729 PERL_ARGS_ASSERT_CL_ANYTHING;
730
f8bef550 731 ANYOF_BITMAP_SETALL(cl);
dd58aee1 732 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
3ad98780 733 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
3fffb88a
KW
734
735 /* If any portion of the regex is to operate under locale rules,
736 * initialization includes it. The reason this isn't done for all regexes
737 * is that the optimizer was written under the assumption that locale was
738 * all-or-nothing. Given the complexity and lack of documentation in the
739 * optimizer, and that there are inadequate test cases for locale, so many
740 * parts of it may not work properly, it is safest to avoid locale unless
741 * necessary. */
742 if (RExC_contains_locale) {
9d7a1e63 743 ANYOF_CLASS_SETALL(cl); /* /l uses class */
3fffb88a
KW
744 cl->flags |= ANYOF_LOCALE;
745 }
9d7a1e63
KW
746 else {
747 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
748 }
653099ff
GS
749}
750
751/* Can match anything (initialization) */
752STATIC int
5f66b61c 753S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
754{
755 int value;
756
7918f24d
NC
757 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
758
aaa51d5e 759 for (value = 0; value <= ANYOF_MAX; value += 2)
653099ff
GS
760 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
761 return 1;
1aa99e6b
IH
762 if (!(cl->flags & ANYOF_UNICODE_ALL))
763 return 0;
10edeb5d 764 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 765 return 0;
653099ff
GS
766 return 1;
767}
768
769/* Can match anything (initialization) */
770STATIC void
e755fd73 771S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 772{
7918f24d
NC
773 PERL_ARGS_ASSERT_CL_INIT;
774
8ecf7187 775 Zero(cl, 1, struct regnode_charclass_class);
653099ff 776 cl->type = ANYOF;
3fffb88a 777 cl_anything(pRExC_state, cl);
1411dba4 778 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
779}
780
1051e1c4
KW
781/* These two functions currently do the exact same thing */
782#define cl_init_zero S_cl_init
653099ff 783
dd58aee1
KW
784/* 'AND' a given class with another one. Can create false positives. 'cl'
785 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
786 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 787STATIC void
5f66b61c 788S_cl_and(struct regnode_charclass_class *cl,
a28509cc 789 const struct regnode_charclass_class *and_with)
653099ff 790{
7918f24d 791 PERL_ARGS_ASSERT_CL_AND;
40d049e4
YO
792
793 assert(and_with->type == ANYOF);
1e6ade67 794
c6b76537 795 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
796 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
797 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 798 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
799 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
800 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
653099ff
GS
801 int i;
802
803 if (and_with->flags & ANYOF_INVERT)
804 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
805 cl->bitmap[i] &= ~and_with->bitmap[i];
806 else
807 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
808 cl->bitmap[i] &= and_with->bitmap[i];
809 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 810
c6b76537 811 if (and_with->flags & ANYOF_INVERT) {
8951c461 812
c6b76537
KW
813 /* Here, the and'ed node is inverted. Get the AND of the flags that
814 * aren't affected by the inversion. Those that are affected are
815 * handled individually below */
816 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
817 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
818 cl->flags |= affected_flags;
819
820 /* We currently don't know how to deal with things that aren't in the
821 * bitmap, but we know that the intersection is no greater than what
822 * is already in cl, so let there be false positives that get sorted
823 * out after the synthetic start class succeeds, and the node is
824 * matched for real. */
825
826 /* The inversion of these two flags indicate that the resulting
827 * intersection doesn't have them */
828 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
829 cl->flags &= ~ANYOF_UNICODE_ALL;
830 }
c6b76537
KW
831 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
832 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 833 }
1aa99e6b 834 }
c6b76537 835 else { /* and'd node is not inverted */
3ad98780
KW
836 U8 outside_bitmap_but_not_utf8; /* Temp variable */
837
137165a6 838 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
839
840 /* Here 'and_with' doesn't match anything outside the bitmap
841 * (except possibly ANYOF_UNICODE_ALL), which means the
842 * intersection can't either, except for ANYOF_UNICODE_ALL, in
843 * which case we don't know what the intersection is, but it's no
844 * greater than what cl already has, so can just leave it alone,
845 * with possible false positives */
846 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
847 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 848 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 849 }
137165a6 850 }
c6b76537
KW
851 else if (! ANYOF_NONBITMAP(cl)) {
852
853 /* Here, 'and_with' does match something outside the bitmap, and cl
854 * doesn't have a list of things to match outside the bitmap. If
855 * cl can match all code points above 255, the intersection will
3ad98780
KW
856 * be those above-255 code points that 'and_with' matches. If cl
857 * can't match all Unicode code points, it means that it can't
858 * match anything outside the bitmap (since the 'if' that got us
859 * into this block tested for that), so we leave the bitmap empty.
860 */
c6b76537
KW
861 if (cl->flags & ANYOF_UNICODE_ALL) {
862 ARG_SET(cl, ARG(and_with));
3ad98780
KW
863
864 /* and_with's ARG may match things that don't require UTF8.
865 * And now cl's will too, in spite of this being an 'and'. See
866 * the comments below about the kludge */
867 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
868 }
869 }
870 else {
871 /* Here, both 'and_with' and cl match something outside the
872 * bitmap. Currently we do not do the intersection, so just match
873 * whatever cl had at the beginning. */
874 }
875
876
3ad98780
KW
877 /* Take the intersection of the two sets of flags. However, the
878 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
879 * kludge around the fact that this flag is not treated like the others
880 * which are initialized in cl_anything(). The way the optimizer works
881 * is that the synthetic start class (SSC) is initialized to match
882 * anything, and then the first time a real node is encountered, its
883 * values are AND'd with the SSC's with the result being the values of
884 * the real node. However, there are paths through the optimizer where
885 * the AND never gets called, so those initialized bits are set
886 * inappropriately, which is not usually a big deal, as they just cause
887 * false positives in the SSC, which will just mean a probably
888 * imperceptible slow down in execution. However this bit has a
889 * higher false positive consequence in that it can cause utf8.pm,
890 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
891 * bigger slowdown and also causes significant extra memory to be used.
892 * In order to prevent this, the code now takes a different tack. The
893 * bit isn't set unless some part of the regular expression needs it,
894 * but once set it won't get cleared. This means that these extra
895 * modules won't get loaded unless there was some path through the
896 * pattern that would have required them anyway, and so any false
897 * positives that occur by not ANDing them out when they could be
898 * aren't as severe as they would be if we treated this bit like all
899 * the others */
900 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
901 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 902 cl->flags &= and_with->flags;
3ad98780 903 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 904 }
653099ff
GS
905}
906
dd58aee1
KW
907/* 'OR' a given class with another one. Can create false positives. 'cl'
908 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
909 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 910STATIC void
3fffb88a 911S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 912{
7918f24d
NC
913 PERL_ARGS_ASSERT_CL_OR;
914
653099ff 915 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
916
917 /* Here, the or'd node is to be inverted. This means we take the
918 * complement of everything not in the bitmap, but currently we don't
919 * know what that is, so give up and match anything */
920 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 921 cl_anything(pRExC_state, cl);
c6b76537 922 }
653099ff
GS
923 /* We do not use
924 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
925 * <= (B1 | !B2) | (CL1 | !CL2)
926 * which is wasteful if CL2 is small, but we ignore CL2:
927 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
928 * XXXX Can we handle case-fold? Unclear:
929 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
930 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
931 */
c6b76537 932 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
933 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
934 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
653099ff
GS
935 int i;
936
937 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
938 cl->bitmap[i] |= ~or_with->bitmap[i];
939 } /* XXXX: logic is complicated otherwise */
940 else {
3fffb88a 941 cl_anything(pRExC_state, cl);
653099ff 942 }
c6b76537
KW
943
944 /* And, we can just take the union of the flags that aren't affected
945 * by the inversion */
946 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
947
948 /* For the remaining flags:
949 ANYOF_UNICODE_ALL and inverted means to not match anything above
950 255, which means that the union with cl should just be
951 what cl has in it, so can ignore this flag
952 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
953 is 127-255 to match them, but then invert that, so the
954 union with cl should just be what cl has in it, so can
955 ignore this flag
956 */
957 } else { /* 'or_with' is not inverted */
653099ff
GS
958 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
959 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
39065660
KW
960 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
961 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
653099ff
GS
962 int i;
963
964 /* OR char bitmap and class bitmap separately */
965 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
966 cl->bitmap[i] |= or_with->bitmap[i];
1e6ade67 967 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
653099ff
GS
968 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
969 cl->classflags[i] |= or_with->classflags[i];
970 cl->flags |= ANYOF_CLASS;
971 }
972 }
973 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 974 cl_anything(pRExC_state, cl);
653099ff 975 }
9826f543 976
c6b76537
KW
977 if (ANYOF_NONBITMAP(or_with)) {
978
979 /* Use the added node's outside-the-bit-map match if there isn't a
980 * conflict. If there is a conflict (both nodes match something
981 * outside the bitmap, but what they match outside is not the same
982 * pointer, and hence not easily compared until XXX we extend
983 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
984 * match everything outside the bitmap. If that stuff is all above
985 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
986 if (! ANYOF_NONBITMAP(cl)) {
987 ARG_SET(cl, ARG(or_with));
988 }
989 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
990
991 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
992 cl_anything(pRExC_state, cl);
993 }
994 else {
995 cl->flags |= ANYOF_UNICODE_ALL;
996 }
c6b76537 997 }
4c34a693 998 }
0b9668ee
KW
999
1000 /* Take the union */
1001 cl->flags |= or_with->flags;
1aa99e6b 1002 }
653099ff
GS
1003}
1004
a3621e74
YO
1005#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1006#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1007#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1008#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1009
3dab1dad
YO
1010
1011#ifdef DEBUGGING
07be1b83 1012/*
2b8b4781
NC
1013 dump_trie(trie,widecharmap,revcharmap)
1014 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1015 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1016
1017 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1018 The _interim_ variants are used for debugging the interim
1019 tables that are used to generate the final compressed
1020 representation which is what dump_trie expects.
1021
486ec47a 1022 Part of the reason for their existence is to provide a form
3dab1dad 1023 of documentation as to how the different representations function.
07be1b83
YO
1024
1025*/
3dab1dad
YO
1026
1027/*
3dab1dad
YO
1028 Dumps the final compressed table form of the trie to Perl_debug_log.
1029 Used for debugging make_trie().
1030*/
b9a59e08 1031
3dab1dad 1032STATIC void
2b8b4781
NC
1033S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1034 AV *revcharmap, U32 depth)
3dab1dad
YO
1035{
1036 U32 state;
ab3bbdeb 1037 SV *sv=sv_newmortal();
55eed653 1038 int colwidth= widecharmap ? 6 : 4;
2e64971a 1039 U16 word;
3dab1dad
YO
1040 GET_RE_DEBUG_FLAGS_DECL;
1041
7918f24d 1042 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1043
3dab1dad
YO
1044 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1045 (int)depth * 2 + 2,"",
1046 "Match","Base","Ofs" );
1047
1048 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1049 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1050 if ( tmp ) {
ab3bbdeb
YO
1051 PerlIO_printf( Perl_debug_log, "%*s",
1052 colwidth,
ddc5bc0f 1053 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1054 PL_colors[0], PL_colors[1],
1055 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1056 PERL_PV_ESCAPE_FIRSTCHAR
1057 )
1058 );
3dab1dad
YO
1059 }
1060 }
1061 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1062 (int)depth * 2 + 2,"");
1063
1064 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1065 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1066 PerlIO_printf( Perl_debug_log, "\n");
1067
1e2e3d02 1068 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1069 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1070
1071 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1072
1073 if ( trie->states[ state ].wordnum ) {
1074 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1075 } else {
1076 PerlIO_printf( Perl_debug_log, "%6s", "" );
1077 }
1078
1079 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1080
1081 if ( base ) {
1082 U32 ofs = 0;
1083
1084 while( ( base + ofs < trie->uniquecharcount ) ||
1085 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1086 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1087 ofs++;
1088
1089 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1090
1091 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1092 if ( ( base + ofs >= trie->uniquecharcount ) &&
1093 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1094 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1095 {
ab3bbdeb
YO
1096 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1097 colwidth,
3dab1dad
YO
1098 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1099 } else {
ab3bbdeb 1100 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1101 }
1102 }
1103
1104 PerlIO_printf( Perl_debug_log, "]");
1105
1106 }
1107 PerlIO_printf( Perl_debug_log, "\n" );
1108 }
2e64971a
DM
1109 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1110 for (word=1; word <= trie->wordcount; word++) {
1111 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1112 (int)word, (int)(trie->wordinfo[word].prev),
1113 (int)(trie->wordinfo[word].len));
1114 }
1115 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1116}
1117/*
3dab1dad
YO
1118 Dumps a fully constructed but uncompressed trie in list form.
1119 List tries normally only are used for construction when the number of
1120 possible chars (trie->uniquecharcount) is very high.
1121 Used for debugging make_trie().
1122*/
1123STATIC void
55eed653 1124S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1125 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1126 U32 depth)
3dab1dad
YO
1127{
1128 U32 state;
ab3bbdeb 1129 SV *sv=sv_newmortal();
55eed653 1130 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1131 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1132
1133 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1134
3dab1dad 1135 /* print out the table precompression. */
ab3bbdeb
YO
1136 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1137 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1138 "------:-----+-----------------\n" );
3dab1dad
YO
1139
1140 for( state=1 ; state < next_alloc ; state ++ ) {
1141 U16 charid;
1142
ab3bbdeb 1143 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1144 (int)depth * 2 + 2,"", (UV)state );
1145 if ( ! trie->states[ state ].wordnum ) {
1146 PerlIO_printf( Perl_debug_log, "%5s| ","");
1147 } else {
1148 PerlIO_printf( Perl_debug_log, "W%4x| ",
1149 trie->states[ state ].wordnum
1150 );
1151 }
1152 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1153 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1154 if ( tmp ) {
1155 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1156 colwidth,
ddc5bc0f 1157 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1158 PL_colors[0], PL_colors[1],
1159 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1160 PERL_PV_ESCAPE_FIRSTCHAR
1161 ) ,
1e2e3d02
YO
1162 TRIE_LIST_ITEM(state,charid).forid,
1163 (UV)TRIE_LIST_ITEM(state,charid).newstate
1164 );
1165 if (!(charid % 10))
664e119d
RGS
1166 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1167 (int)((depth * 2) + 14), "");
1e2e3d02 1168 }
ab3bbdeb
YO
1169 }
1170 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1171 }
1172}
1173
1174/*
3dab1dad
YO
1175 Dumps a fully constructed but uncompressed trie in table form.
1176 This is the normal DFA style state transition table, with a few
1177 twists to facilitate compression later.
1178 Used for debugging make_trie().
1179*/
1180STATIC void
55eed653 1181S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1182 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1183 U32 depth)
3dab1dad
YO
1184{
1185 U32 state;
1186 U16 charid;
ab3bbdeb 1187 SV *sv=sv_newmortal();
55eed653 1188 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1189 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1190
1191 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1192
1193 /*
1194 print out the table precompression so that we can do a visual check
1195 that they are identical.
1196 */
1197
1198 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1199
1200 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1201 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1202 if ( tmp ) {
ab3bbdeb
YO
1203 PerlIO_printf( Perl_debug_log, "%*s",
1204 colwidth,
ddc5bc0f 1205 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1206 PL_colors[0], PL_colors[1],
1207 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1208 PERL_PV_ESCAPE_FIRSTCHAR
1209 )
1210 );
3dab1dad
YO
1211 }
1212 }
1213
1214 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1215
1216 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1217 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1218 }
1219
1220 PerlIO_printf( Perl_debug_log, "\n" );
1221
1222 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1223
1224 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1225 (int)depth * 2 + 2,"",
1226 (UV)TRIE_NODENUM( state ) );
1227
1228 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1229 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1230 if (v)
1231 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1232 else
1233 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1234 }
1235 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1236 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1237 } else {
1238 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1239 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1240 }
1241 }
07be1b83 1242}
3dab1dad
YO
1243
1244#endif
1245
2e64971a 1246
786e8c11
YO
1247/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1248 startbranch: the first branch in the whole branch sequence
1249 first : start branch of sequence of branch-exact nodes.
1250 May be the same as startbranch
1251 last : Thing following the last branch.
1252 May be the same as tail.
1253 tail : item following the branch sequence
1254 count : words in the sequence
1255 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1256 depth : indent depth
3dab1dad 1257
786e8c11 1258Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1259
786e8c11
YO
1260A trie is an N'ary tree where the branches are determined by digital
1261decomposition of the key. IE, at the root node you look up the 1st character and
1262follow that branch repeat until you find the end of the branches. Nodes can be
1263marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1264
786e8c11 1265 /he|she|his|hers/
72f13be8 1266
786e8c11
YO
1267would convert into the following structure. Numbers represent states, letters
1268following numbers represent valid transitions on the letter from that state, if
1269the number is in square brackets it represents an accepting state, otherwise it
1270will be in parenthesis.
07be1b83 1271
786e8c11
YO
1272 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1273 | |
1274 | (2)
1275 | |
1276 (1) +-i->(6)-+-s->[7]
1277 |
1278 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1279
786e8c11
YO
1280 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1281
1282This shows that when matching against the string 'hers' we will begin at state 1
1283read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1284then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1285is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1286single traverse. We store a mapping from accepting to state to which word was
1287matched, and then when we have multiple possibilities we try to complete the
1288rest of the regex in the order in which they occured in the alternation.
1289
1290The only prior NFA like behaviour that would be changed by the TRIE support is
1291the silent ignoring of duplicate alternations which are of the form:
1292
1293 / (DUPE|DUPE) X? (?{ ... }) Y /x
1294
4b714af6 1295Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1296and without the optimisation. With the optimisations dupes will be silently
486ec47a 1297ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1298the following demonstrates:
1299
1300 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1301
1302which prints out 'word' three times, but
1303
1304 'words'=~/(word|word|word)(?{ print $1 })S/
1305
1306which doesnt print it out at all. This is due to other optimisations kicking in.
1307
1308Example of what happens on a structural level:
1309
486ec47a 1310The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1311
1312 1: CURLYM[1] {1,32767}(18)
1313 5: BRANCH(8)
1314 6: EXACT <ac>(16)
1315 8: BRANCH(11)
1316 9: EXACT <ad>(16)
1317 11: BRANCH(14)
1318 12: EXACT <ab>(16)
1319 16: SUCCEED(0)
1320 17: NOTHING(18)
1321 18: END(0)
1322
1323This would be optimizable with startbranch=5, first=5, last=16, tail=16
1324and should turn into:
1325
1326 1: CURLYM[1] {1,32767}(18)
1327 5: TRIE(16)
1328 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1329 <ac>
1330 <ad>
1331 <ab>
1332 16: SUCCEED(0)
1333 17: NOTHING(18)
1334 18: END(0)
1335
1336Cases where tail != last would be like /(?foo|bar)baz/:
1337
1338 1: BRANCH(4)
1339 2: EXACT <foo>(8)
1340 4: BRANCH(7)
1341 5: EXACT <bar>(8)
1342 7: TAIL(8)
1343 8: EXACT <baz>(10)
1344 10: END(0)
1345
1346which would be optimizable with startbranch=1, first=1, last=7, tail=8
1347and would end up looking like:
1348
1349 1: TRIE(8)
1350 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1351 <foo>
1352 <bar>
1353 7: TAIL(8)
1354 8: EXACT <baz>(10)
1355 10: END(0)
1356
1357 d = uvuni_to_utf8_flags(d, uv, 0);
1358
1359is the recommended Unicode-aware way of saying
1360
1361 *(d++) = uv;
1362*/
1363
1e2e3d02 1364#define TRIE_STORE_REVCHAR \
786e8c11 1365 STMT_START { \
73031816
NC
1366 if (UTF) { \
1367 SV *zlopp = newSV(2); \
88c9ea1e
CB
1368 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1369 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, uvc & 0xFF); \
73031816
NC
1370 SvCUR_set(zlopp, kapow - flrbbbbb); \
1371 SvPOK_on(zlopp); \
1372 SvUTF8_on(zlopp); \
1373 av_push(revcharmap, zlopp); \
1374 } else { \
6bdeddd2 1375 char ooooff = (char)uvc; \
73031816
NC
1376 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1377 } \
1378 } STMT_END
786e8c11
YO
1379
1380#define TRIE_READ_CHAR STMT_START { \
1381 wordlen++; \
1382 if ( UTF ) { \
1383 if ( folder ) { \
1384 if ( foldlen > 0 ) { \
1385 uvc = utf8n_to_uvuni( scan, UTF8_MAXLEN, &len, uniflags ); \
1386 foldlen -= len; \
1387 scan += len; \
1388 len = 0; \
1389 } else { \
c81f2f9e
KW
1390 len = UTF8SKIP(uc);\
1391 uvc = to_utf8_fold( uc, foldbuf, &foldlen); \
786e8c11
YO
1392 foldlen -= UNISKIP( uvc ); \
1393 scan = foldbuf + UNISKIP( uvc ); \
1394 } \
1395 } else { \
1396 uvc = utf8n_to_uvuni( (const U8*)uc, UTF8_MAXLEN, &len, uniflags);\
1397 } \
1398 } else { \
1399 uvc = (U32)*uc; \
1400 len = 1; \
1401 } \
1402} STMT_END
1403
1404
1405
1406#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1407 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1408 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1409 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1410 } \
1411 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1412 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1413 TRIE_LIST_CUR( state )++; \
1414} STMT_END
07be1b83 1415
786e8c11
YO
1416#define TRIE_LIST_NEW(state) STMT_START { \
1417 Newxz( trie->states[ state ].trans.list, \
1418 4, reg_trie_trans_le ); \
1419 TRIE_LIST_CUR( state ) = 1; \
1420 TRIE_LIST_LEN( state ) = 4; \
1421} STMT_END
07be1b83 1422
786e8c11
YO
1423#define TRIE_HANDLE_WORD(state) STMT_START { \
1424 U16 dupe= trie->states[ state ].wordnum; \
1425 regnode * const noper_next = regnext( noper ); \
1426 \
786e8c11
YO
1427 DEBUG_r({ \
1428 /* store the word for dumping */ \
1429 SV* tmp; \
1430 if (OP(noper) != NOTHING) \
740cce10 1431 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1432 else \
740cce10 1433 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1434 av_push( trie_words, tmp ); \
786e8c11
YO
1435 }); \
1436 \
1437 curword++; \
2e64971a
DM
1438 trie->wordinfo[curword].prev = 0; \
1439 trie->wordinfo[curword].len = wordlen; \
1440 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1441 \
1442 if ( noper_next < tail ) { \
1443 if (!trie->jump) \
c944940b 1444 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1445 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1446 if (!jumper) \
1447 jumper = noper_next; \
1448 if (!nextbranch) \
1449 nextbranch= regnext(cur); \
1450 } \
1451 \
1452 if ( dupe ) { \
2e64971a
DM
1453 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1454 /* chain, so that when the bits of chain are later */\
1455 /* linked together, the dups appear in the chain */\
1456 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1457 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1458 } else { \
1459 /* we haven't inserted this word yet. */ \
1460 trie->states[ state ].wordnum = curword; \
1461 } \
1462} STMT_END
07be1b83 1463
3dab1dad 1464
786e8c11
YO
1465#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1466 ( ( base + charid >= ucharcount \
1467 && base + charid < ubound \
1468 && state == trie->trans[ base - ucharcount + charid ].check \
1469 && trie->trans[ base - ucharcount + charid ].next ) \
1470 ? trie->trans[ base - ucharcount + charid ].next \
1471 : ( state==1 ? special : 0 ) \
1472 )
3dab1dad 1473
786e8c11
YO
1474#define MADE_TRIE 1
1475#define MADE_JUMP_TRIE 2
1476#define MADE_EXACT_TRIE 4
3dab1dad 1477
a3621e74 1478STATIC I32
786e8c11 1479S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1480{
27da23d5 1481 dVAR;
a3621e74
YO
1482 /* first pass, loop through and scan words */
1483 reg_trie_data *trie;
55eed653 1484 HV *widecharmap = NULL;
2b8b4781 1485 AV *revcharmap = newAV();
a3621e74 1486 regnode *cur;
9f7f3913 1487 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1488 STRLEN len = 0;
1489 UV uvc = 0;
1490 U16 curword = 0;
1491 U32 next_alloc = 0;
786e8c11
YO
1492 regnode *jumper = NULL;
1493 regnode *nextbranch = NULL;
7f69552c 1494 regnode *convert = NULL;
2e64971a 1495 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1496 /* we just use folder as a flag in utf8 */
1e696034 1497 const U8 * folder = NULL;
a3621e74 1498
2b8b4781
NC
1499#ifdef DEBUGGING
1500 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1501 AV *trie_words = NULL;
1502 /* along with revcharmap, this only used during construction but both are
1503 * useful during debugging so we store them in the struct when debugging.
8e11feef 1504 */
2b8b4781
NC
1505#else
1506 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1507 STRLEN trie_charcount=0;
3dab1dad 1508#endif
2b8b4781 1509 SV *re_trie_maxbuff;
a3621e74 1510 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1511
1512 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1513#ifndef DEBUGGING
1514 PERL_UNUSED_ARG(depth);
1515#endif
a3621e74 1516
1e696034 1517 switch (flags) {
c46d03cf 1518 case EXACT: break;
2f7f8cb1 1519 case EXACTFA:
1e696034
KW
1520 case EXACTFU: folder = PL_fold_latin1; break;
1521 case EXACTF: folder = PL_fold; break;
1522 case EXACTFL: folder = PL_fold_locale; break;
c46d03cf 1523 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u", (unsigned) flags );
1e696034
KW
1524 }
1525
c944940b 1526 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1527 trie->refcount = 1;
3dab1dad 1528 trie->startstate = 1;
786e8c11 1529 trie->wordcount = word_count;
f8fc2ecf 1530 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1531 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
3dab1dad 1532 if (!(UTF && folder))
c944940b 1533 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1534 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1535 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1536
a3621e74 1537 DEBUG_r({
2b8b4781 1538 trie_words = newAV();
a3621e74 1539 });
a3621e74 1540
0111c4fd 1541 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1542 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1543 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1544 }
3dab1dad
YO
1545 DEBUG_OPTIMISE_r({
1546 PerlIO_printf( Perl_debug_log,
786e8c11 1547 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1548 (int)depth * 2 + 2, "",
1549 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1550 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1551 (int)depth);
3dab1dad 1552 });
7f69552c
YO
1553
1554 /* Find the node we are going to overwrite */
1555 if ( first == startbranch && OP( last ) != BRANCH ) {
1556 /* whole branch chain */
1557 convert = first;
1558 } else {
1559 /* branch sub-chain */
1560 convert = NEXTOPER( first );
1561 }
1562
a3621e74
YO
1563 /* -- First loop and Setup --
1564
1565 We first traverse the branches and scan each word to determine if it
1566 contains widechars, and how many unique chars there are, this is
1567 important as we have to build a table with at least as many columns as we
1568 have unique chars.
1569
1570 We use an array of integers to represent the character codes 0..255
38a44b82 1571 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1572 native representation of the character value as the key and IV's for the
1573 coded index.
1574
1575 *TODO* If we keep track of how many times each character is used we can
1576 remap the columns so that the table compression later on is more
3b753521 1577 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1578 middle and the least common are on the outside. IMO this would be better
1579 than a most to least common mapping as theres a decent chance the most
1580 common letter will share a node with the least common, meaning the node
486ec47a 1581 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1582 case is when we have the least common nodes twice.
1583
1584 */
1585
a3621e74 1586 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
c445ea15 1587 regnode * const noper = NEXTOPER( cur );
e1ec3a88 1588 const U8 *uc = (U8*)STRING( noper );
a28509cc 1589 const U8 * const e = uc + STR_LEN( noper );
a3621e74
YO
1590 STRLEN foldlen = 0;
1591 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2af232bd 1592 const U8 *scan = (U8*)NULL;
07be1b83 1593 U32 wordlen = 0; /* required init */
02daf0ab
YO
1594 STRLEN chars = 0;
1595 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1596
3dab1dad
YO
1597 if (OP(noper) == NOTHING) {
1598 trie->minlen= 0;
1599 continue;
1600 }
02daf0ab
YO
1601 if ( set_bit ) /* bitmap only alloced when !(UTF&&Folding) */
1602 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1603 regardless of encoding */
1604
a3621e74 1605 for ( ; uc < e ; uc += len ) {
3dab1dad 1606 TRIE_CHARCOUNT(trie)++;
a3621e74 1607 TRIE_READ_CHAR;
3dab1dad 1608 chars++;
a3621e74
YO
1609 if ( uvc < 256 ) {
1610 if ( !trie->charmap[ uvc ] ) {
1611 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1612 if ( folder )
1613 trie->charmap[ folder[ uvc ] ] = trie->charmap[ uvc ];
3dab1dad 1614 TRIE_STORE_REVCHAR;
a3621e74 1615 }
02daf0ab 1616 if ( set_bit ) {
62012aee
KW
1617 /* store the codepoint in the bitmap, and its folded
1618 * equivalent. */
02daf0ab 1619 TRIE_BITMAP_SET(trie,uvc);
0921ee73
T
1620
1621 /* store the folded codepoint */
1622 if ( folder ) TRIE_BITMAP_SET(trie,folder[ uvc ]);
1623
1624 if ( !UTF ) {
1625 /* store first byte of utf8 representation of
acdf4139
KW
1626 variant codepoints */
1627 if (! UNI_IS_INVARIANT(uvc)) {
1628 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1629 }
1630 }
02daf0ab
YO
1631 set_bit = 0; /* We've done our bit :-) */
1632 }
a3621e74
YO
1633 } else {
1634 SV** svpp;
55eed653
NC
1635 if ( !widecharmap )
1636 widecharmap = newHV();
a3621e74 1637
55eed653 1638 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1639
1640 if ( !svpp )
e4584336 1641 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1642
1643 if ( !SvTRUE( *svpp ) ) {
1644 sv_setiv( *svpp, ++trie->uniquecharcount );
3dab1dad 1645 TRIE_STORE_REVCHAR;
a3621e74
YO
1646 }
1647 }
1648 }
3dab1dad
YO
1649 if( cur == first ) {
1650 trie->minlen=chars;
1651 trie->maxlen=chars;
1652 } else if (chars < trie->minlen) {
1653 trie->minlen=chars;
1654 } else if (chars > trie->maxlen) {
1655 trie->maxlen=chars;
1656 }
1657
a3621e74
YO
1658 } /* end first pass */
1659 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1660 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1661 (int)depth * 2 + 2,"",
55eed653 1662 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1663 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1664 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1665 );
a3621e74
YO
1666
1667 /*
1668 We now know what we are dealing with in terms of unique chars and
1669 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1670 representation using a flat table will take. If it's over a reasonable
1671 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1672 conservative but potentially much slower representation using an array
1673 of lists.
1674
1675 At the end we convert both representations into the same compressed
1676 form that will be used in regexec.c for matching with. The latter
1677 is a form that cannot be used to construct with but has memory
1678 properties similar to the list form and access properties similar
1679 to the table form making it both suitable for fast searches and
1680 small enough that its feasable to store for the duration of a program.
1681
1682 See the comment in the code where the compressed table is produced
1683 inplace from the flat tabe representation for an explanation of how
1684 the compression works.
1685
1686 */
1687
1688
2e64971a
DM
1689 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1690 prev_states[1] = 0;
1691
3dab1dad 1692 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1693 /*
1694 Second Pass -- Array Of Lists Representation
1695
1696 Each state will be represented by a list of charid:state records
1697 (reg_trie_trans_le) the first such element holds the CUR and LEN
1698 points of the allocated array. (See defines above).
1699
1700 We build the initial structure using the lists, and then convert
1701 it into the compressed table form which allows faster lookups
1702 (but cant be modified once converted).
a3621e74
YO
1703 */
1704
a3621e74
YO
1705 STRLEN transcount = 1;
1706
1e2e3d02
YO
1707 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1708 "%*sCompiling trie using list compiler\n",
1709 (int)depth * 2 + 2, ""));
686b73d4 1710
c944940b
JH
1711 trie->states = (reg_trie_state *)
1712 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1713 sizeof(reg_trie_state) );
a3621e74
YO
1714 TRIE_LIST_NEW(1);
1715 next_alloc = 2;
1716
1717 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1718
c445ea15
AL
1719 regnode * const noper = NEXTOPER( cur );
1720 U8 *uc = (U8*)STRING( noper );
1721 const U8 * const e = uc + STR_LEN( noper );
1722 U32 state = 1; /* required init */
1723 U16 charid = 0; /* sanity init */
1724 U8 *scan = (U8*)NULL; /* sanity init */
1725 STRLEN foldlen = 0; /* required init */
07be1b83 1726 U32 wordlen = 0; /* required init */
c445ea15
AL
1727 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1728
3dab1dad 1729 if (OP(noper) != NOTHING) {
786e8c11 1730 for ( ; uc < e ; uc += len ) {
c445ea15 1731
786e8c11 1732 TRIE_READ_CHAR;
c445ea15 1733
786e8c11
YO
1734 if ( uvc < 256 ) {
1735 charid = trie->charmap[ uvc ];
c445ea15 1736 } else {
55eed653 1737 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1738 if ( !svpp ) {
1739 charid = 0;
1740 } else {
1741 charid=(U16)SvIV( *svpp );
1742 }
c445ea15 1743 }
786e8c11
YO
1744 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1745 if ( charid ) {
a3621e74 1746
786e8c11
YO
1747 U16 check;
1748 U32 newstate = 0;
a3621e74 1749
786e8c11
YO
1750 charid--;
1751 if ( !trie->states[ state ].trans.list ) {
1752 TRIE_LIST_NEW( state );
c445ea15 1753 }
786e8c11
YO
1754 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1755 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1756 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1757 break;
1758 }
1759 }
1760 if ( ! newstate ) {
1761 newstate = next_alloc++;
2e64971a 1762 prev_states[newstate] = state;
786e8c11
YO
1763 TRIE_LIST_PUSH( state, charid, newstate );
1764 transcount++;
1765 }
1766 state = newstate;
1767 } else {
1768 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1769 }
a28509cc 1770 }
c445ea15 1771 }
3dab1dad 1772 TRIE_HANDLE_WORD(state);
a3621e74
YO
1773
1774 } /* end second pass */
1775
1e2e3d02
YO
1776 /* next alloc is the NEXT state to be allocated */
1777 trie->statecount = next_alloc;
c944940b
JH
1778 trie->states = (reg_trie_state *)
1779 PerlMemShared_realloc( trie->states,
1780 next_alloc
1781 * sizeof(reg_trie_state) );
a3621e74 1782
3dab1dad 1783 /* and now dump it out before we compress it */
2b8b4781
NC
1784 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1785 revcharmap, next_alloc,
1786 depth+1)
1e2e3d02 1787 );
a3621e74 1788
c944940b
JH
1789 trie->trans = (reg_trie_trans *)
1790 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1791 {
1792 U32 state;
a3621e74
YO
1793 U32 tp = 0;
1794 U32 zp = 0;
1795
1796
1797 for( state=1 ; state < next_alloc ; state ++ ) {
1798 U32 base=0;
1799
1800 /*
1801 DEBUG_TRIE_COMPILE_MORE_r(
1802 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1803 );
1804 */
1805
1806 if (trie->states[state].trans.list) {
1807 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1808 U16 maxid=minid;
a28509cc 1809 U16 idx;
a3621e74
YO
1810
1811 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1812 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1813 if ( forid < minid ) {
1814 minid=forid;
1815 } else if ( forid > maxid ) {
1816 maxid=forid;
1817 }
a3621e74
YO
1818 }
1819 if ( transcount < tp + maxid - minid + 1) {
1820 transcount *= 2;
c944940b
JH
1821 trie->trans = (reg_trie_trans *)
1822 PerlMemShared_realloc( trie->trans,
446bd890
NC
1823 transcount
1824 * sizeof(reg_trie_trans) );
a3621e74
YO
1825 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1826 }
1827 base = trie->uniquecharcount + tp - minid;
1828 if ( maxid == minid ) {
1829 U32 set = 0;
1830 for ( ; zp < tp ; zp++ ) {
1831 if ( ! trie->trans[ zp ].next ) {
1832 base = trie->uniquecharcount + zp - minid;
1833 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1834 trie->trans[ zp ].check = state;
1835 set = 1;
1836 break;
1837 }
1838 }
1839 if ( !set ) {
1840 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1841 trie->trans[ tp ].check = state;
1842 tp++;
1843 zp = tp;
1844 }
1845 } else {
1846 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1847 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1848 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1849 trie->trans[ tid ].check = state;
1850 }
1851 tp += ( maxid - minid + 1 );
1852 }
1853 Safefree(trie->states[ state ].trans.list);
1854 }
1855 /*
1856 DEBUG_TRIE_COMPILE_MORE_r(
1857 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1858 );
1859 */
1860 trie->states[ state ].trans.base=base;
1861 }
cc601c31 1862 trie->lasttrans = tp + 1;
a3621e74
YO
1863 }
1864 } else {
1865 /*
1866 Second Pass -- Flat Table Representation.
1867
1868 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1869 We know that we will need Charcount+1 trans at most to store the data
1870 (one row per char at worst case) So we preallocate both structures
1871 assuming worst case.
1872
1873 We then construct the trie using only the .next slots of the entry
1874 structs.
1875
3b753521 1876 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1877 make compression both faster and easier by keeping track of how many non
1878 zero fields are in the node.
1879
1880 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1881 transition.
1882
1883 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1884 number representing the first entry of the node, and state as a
1885 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1886 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1887 are 2 entrys per node. eg:
1888
1889 A B A B
1890 1. 2 4 1. 3 7
1891 2. 0 3 3. 0 5
1892 3. 0 0 5. 0 0
1893 4. 0 0 7. 0 0
1894
1895 The table is internally in the right hand, idx form. However as we also
1896 have to deal with the states array which is indexed by nodenum we have to
1897 use TRIE_NODENUM() to convert.
1898
1899 */
1e2e3d02
YO
1900 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1901 "%*sCompiling trie using table compiler\n",
1902 (int)depth * 2 + 2, ""));
3dab1dad 1903
c944940b
JH
1904 trie->trans = (reg_trie_trans *)
1905 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1906 * trie->uniquecharcount + 1,
1907 sizeof(reg_trie_trans) );
1908 trie->states = (reg_trie_state *)
1909 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1910 sizeof(reg_trie_state) );
a3621e74
YO
1911 next_alloc = trie->uniquecharcount + 1;
1912
3dab1dad 1913
a3621e74
YO
1914 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1915
c445ea15 1916 regnode * const noper = NEXTOPER( cur );
a28509cc
AL
1917 const U8 *uc = (U8*)STRING( noper );
1918 const U8 * const e = uc + STR_LEN( noper );
a3621e74
YO
1919
1920 U32 state = 1; /* required init */
1921
1922 U16 charid = 0; /* sanity init */
1923 U32 accept_state = 0; /* sanity init */
1924 U8 *scan = (U8*)NULL; /* sanity init */
1925
1926 STRLEN foldlen = 0; /* required init */
07be1b83 1927 U32 wordlen = 0; /* required init */
a3621e74
YO
1928 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1929
3dab1dad 1930 if ( OP(noper) != NOTHING ) {
786e8c11 1931 for ( ; uc < e ; uc += len ) {
a3621e74 1932
786e8c11 1933 TRIE_READ_CHAR;
a3621e74 1934
786e8c11
YO
1935 if ( uvc < 256 ) {
1936 charid = trie->charmap[ uvc ];
1937 } else {
55eed653 1938 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 1939 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 1940 }
786e8c11
YO
1941 if ( charid ) {
1942 charid--;
1943 if ( !trie->trans[ state + charid ].next ) {
1944 trie->trans[ state + charid ].next = next_alloc;
1945 trie->trans[ state ].check++;
2e64971a
DM
1946 prev_states[TRIE_NODENUM(next_alloc)]
1947 = TRIE_NODENUM(state);
786e8c11
YO
1948 next_alloc += trie->uniquecharcount;
1949 }
1950 state = trie->trans[ state + charid ].next;
1951 } else {
1952 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1953 }
1954 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 1955 }
a3621e74 1956 }
3dab1dad
YO
1957 accept_state = TRIE_NODENUM( state );
1958 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
1959
1960 } /* end second pass */
1961
3dab1dad 1962 /* and now dump it out before we compress it */
2b8b4781
NC
1963 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
1964 revcharmap,
1965 next_alloc, depth+1));
a3621e74 1966
a3621e74
YO
1967 {
1968 /*
1969 * Inplace compress the table.*
1970
1971 For sparse data sets the table constructed by the trie algorithm will
1972 be mostly 0/FAIL transitions or to put it another way mostly empty.
1973 (Note that leaf nodes will not contain any transitions.)
1974
1975 This algorithm compresses the tables by eliminating most such
1976 transitions, at the cost of a modest bit of extra work during lookup:
1977
1978 - Each states[] entry contains a .base field which indicates the
1979 index in the state[] array wheres its transition data is stored.
1980
3b753521 1981 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
1982
1983 - If .base is nonzero then charid is added to it to find an entry in
1984 the trans array.
1985
1986 -If trans[states[state].base+charid].check!=state then the
1987 transition is taken to be a 0/Fail transition. Thus if there are fail
1988 transitions at the front of the node then the .base offset will point
1989 somewhere inside the previous nodes data (or maybe even into a node
1990 even earlier), but the .check field determines if the transition is
1991 valid.
1992
786e8c11 1993 XXX - wrong maybe?
a3621e74 1994 The following process inplace converts the table to the compressed
3b753521 1995 table: We first do not compress the root node 1,and mark all its
a3621e74 1996 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
1997 allows us to do a DFA construction from the compressed table later,
1998 and ensures that any .base pointers we calculate later are greater
1999 than 0.
a3621e74
YO
2000
2001 - We set 'pos' to indicate the first entry of the second node.
2002
2003 - We then iterate over the columns of the node, finding the first and
2004 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2005 and set the .check pointers accordingly, and advance pos
2006 appropriately and repreat for the next node. Note that when we copy
2007 the next pointers we have to convert them from the original
2008 NODEIDX form to NODENUM form as the former is not valid post
2009 compression.
2010
2011 - If a node has no transitions used we mark its base as 0 and do not
2012 advance the pos pointer.
2013
2014 - If a node only has one transition we use a second pointer into the
2015 structure to fill in allocated fail transitions from other states.
2016 This pointer is independent of the main pointer and scans forward
2017 looking for null transitions that are allocated to a state. When it
2018 finds one it writes the single transition into the "hole". If the
786e8c11 2019 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2020
2021 - Once compressed we can Renew/realloc the structures to release the
2022 excess space.
2023
2024 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2025 specifically Fig 3.47 and the associated pseudocode.
2026
2027 demq
2028 */
a3b680e6 2029 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2030 U32 state, charid;
a3621e74 2031 U32 pos = 0, zp=0;
1e2e3d02 2032 trie->statecount = laststate;
a3621e74
YO
2033
2034 for ( state = 1 ; state < laststate ; state++ ) {
2035 U8 flag = 0;
a28509cc
AL
2036 const U32 stateidx = TRIE_NODEIDX( state );
2037 const U32 o_used = trie->trans[ stateidx ].check;
2038 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2039 trie->trans[ stateidx ].check = 0;
2040
2041 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2042 if ( flag || trie->trans[ stateidx + charid ].next ) {
2043 if ( trie->trans[ stateidx + charid ].next ) {
2044 if (o_used == 1) {
2045 for ( ; zp < pos ; zp++ ) {
2046 if ( ! trie->trans[ zp ].next ) {
2047 break;
2048 }
2049 }
2050 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2051 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2052 trie->trans[ zp ].check = state;
2053 if ( ++zp > pos ) pos = zp;
2054 break;
2055 }
2056 used--;
2057 }
2058 if ( !flag ) {
2059 flag = 1;
2060 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2061 }
2062 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2063 trie->trans[ pos ].check = state;
2064 pos++;
2065 }
2066 }
2067 }
cc601c31 2068 trie->lasttrans = pos + 1;
c944940b
JH
2069 trie->states = (reg_trie_state *)
2070 PerlMemShared_realloc( trie->states, laststate
2071 * sizeof(reg_trie_state) );
a3621e74 2072 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2073 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2074 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2075 (int)depth * 2 + 2,"",
2076 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2077 (IV)next_alloc,
2078 (IV)pos,
a3621e74
YO
2079 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2080 );
2081
2082 } /* end table compress */
2083 }
1e2e3d02
YO
2084 DEBUG_TRIE_COMPILE_MORE_r(
2085 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2086 (int)depth * 2 + 2, "",
2087 (UV)trie->statecount,
2088 (UV)trie->lasttrans)
2089 );
cc601c31 2090 /* resize the trans array to remove unused space */
c944940b
JH
2091 trie->trans = (reg_trie_trans *)
2092 PerlMemShared_realloc( trie->trans, trie->lasttrans
2093 * sizeof(reg_trie_trans) );
a3621e74 2094
3b753521 2095 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2096 U8 nodetype =(U8)(flags & 0xFF);
2097 char *str=NULL;
786e8c11 2098
07be1b83 2099#ifdef DEBUGGING
e62cc96a 2100 regnode *optimize = NULL;
7122b237
YO
2101#ifdef RE_TRACK_PATTERN_OFFSETS
2102
b57a0404
JH
2103 U32 mjd_offset = 0;
2104 U32 mjd_nodelen = 0;
7122b237
YO
2105#endif /* RE_TRACK_PATTERN_OFFSETS */
2106#endif /* DEBUGGING */
a3621e74 2107 /*
3dab1dad
YO
2108 This means we convert either the first branch or the first Exact,
2109 depending on whether the thing following (in 'last') is a branch
2110 or not and whther first is the startbranch (ie is it a sub part of
2111 the alternation or is it the whole thing.)
3b753521 2112 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2113 the whole branch sequence, including the first.
a3621e74 2114 */
3dab1dad 2115 /* Find the node we are going to overwrite */
7f69552c 2116 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2117 /* branch sub-chain */
3dab1dad 2118 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2119#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2120 DEBUG_r({
2121 mjd_offset= Node_Offset((convert));
2122 mjd_nodelen= Node_Length((convert));
2123 });
7122b237 2124#endif
7f69552c 2125 /* whole branch chain */
7122b237
YO
2126 }
2127#ifdef RE_TRACK_PATTERN_OFFSETS
2128 else {
7f69552c
YO
2129 DEBUG_r({
2130 const regnode *nop = NEXTOPER( convert );
2131 mjd_offset= Node_Offset((nop));
2132 mjd_nodelen= Node_Length((nop));
2133 });
07be1b83
YO
2134 }
2135 DEBUG_OPTIMISE_r(
2136 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2137 (int)depth * 2 + 2, "",
786e8c11 2138 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2139 );
7122b237 2140#endif
3dab1dad
YO
2141 /* But first we check to see if there is a common prefix we can
2142 split out as an EXACT and put in front of the TRIE node. */
2143 trie->startstate= 1;
55eed653 2144 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2145 U32 state;
1e2e3d02 2146 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2147 U32 ofs = 0;
8e11feef
RGS
2148 I32 idx = -1;
2149 U32 count = 0;
2150 const U32 base = trie->states[ state ].trans.base;
a3621e74 2151
3dab1dad 2152 if ( trie->states[state].wordnum )
8e11feef 2153 count = 1;
a3621e74 2154
8e11feef 2155 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2156 if ( ( base + ofs >= trie->uniquecharcount ) &&
2157 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2158 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2159 {
3dab1dad 2160 if ( ++count > 1 ) {
2b8b4781 2161 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2162 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2163 if ( state == 1 ) break;
3dab1dad
YO
2164 if ( count == 2 ) {
2165 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2166 DEBUG_OPTIMISE_r(
8e11feef
RGS
2167 PerlIO_printf(Perl_debug_log,
2168 "%*sNew Start State=%"UVuf" Class: [",
2169 (int)depth * 2 + 2, "",
786e8c11 2170 (UV)state));
be8e71aa 2171 if (idx >= 0) {
2b8b4781 2172 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2173 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2174
3dab1dad 2175 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2176 if ( folder )
2177 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2178 DEBUG_OPTIMISE_r(
f1f66076 2179 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2180 );
8e11feef
RGS
2181 }
2182 }
2183 TRIE_BITMAP_SET(trie,*ch);
2184 if ( folder )
2185 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2186 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2187 }
2188 idx = ofs;
2189 }
3dab1dad
YO
2190 }
2191 if ( count == 1 ) {
2b8b4781 2192 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2193 STRLEN len;
2194 char *ch = SvPV( *tmp, len );
de734bd5
A
2195 DEBUG_OPTIMISE_r({
2196 SV *sv=sv_newmortal();
8e11feef
RGS
2197 PerlIO_printf( Perl_debug_log,
2198 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2199 (int)depth * 2 + 2, "",
de734bd5
A
2200 (UV)state, (UV)idx,
2201 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2202 PL_colors[0], PL_colors[1],
2203 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2204 PERL_PV_ESCAPE_FIRSTCHAR
2205 )
2206 );
2207 });
3dab1dad
YO
2208 if ( state==1 ) {
2209 OP( convert ) = nodetype;
2210 str=STRING(convert);
2211 STR_LEN(convert)=0;
2212 }
c490c714
YO
2213 STR_LEN(convert) += len;
2214 while (len--)
de734bd5 2215 *str++ = *ch++;
8e11feef 2216 } else {
f9049ba1 2217#ifdef DEBUGGING
8e11feef
RGS
2218 if (state>1)
2219 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2220#endif
8e11feef
RGS
2221 break;
2222 }
2223 }
2e64971a 2224 trie->prefixlen = (state-1);
3dab1dad 2225 if (str) {
8e11feef 2226 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2227 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2228 trie->startstate = state;
07be1b83
YO
2229 trie->minlen -= (state - 1);
2230 trie->maxlen -= (state - 1);
33809eae
JH
2231#ifdef DEBUGGING
2232 /* At least the UNICOS C compiler choked on this
2233 * being argument to DEBUG_r(), so let's just have
2234 * it right here. */
2235 if (
2236#ifdef PERL_EXT_RE_BUILD
2237 1
2238#else
2239 DEBUG_r_TEST
2240#endif
2241 ) {
2242 regnode *fix = convert;
2243 U32 word = trie->wordcount;
2244 mjd_nodelen++;
2245 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2246 while( ++fix < n ) {
2247 Set_Node_Offset_Length(fix, 0, 0);
2248 }
2249 while (word--) {
2250 SV ** const tmp = av_fetch( trie_words, word, 0 );
2251 if (tmp) {
2252 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2253 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2254 else
2255 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2256 }
2257 }
2258 }
2259#endif
8e11feef
RGS
2260 if (trie->maxlen) {
2261 convert = n;
2262 } else {
3dab1dad 2263 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2264 DEBUG_r(optimize= n);
3dab1dad
YO
2265 }
2266 }
2267 }
a5ca303d
YO
2268 if (!jumper)
2269 jumper = last;
3dab1dad 2270 if ( trie->maxlen ) {
8e11feef
RGS
2271 NEXT_OFF( convert ) = (U16)(tail - convert);
2272 ARG_SET( convert, data_slot );
786e8c11
YO
2273 /* Store the offset to the first unabsorbed branch in
2274 jump[0], which is otherwise unused by the jump logic.
2275 We use this when dumping a trie and during optimisation. */
2276 if (trie->jump)
7f69552c 2277 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2278
6c48061a
YO
2279 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2280 * and there is a bitmap
2281 * and the first "jump target" node we found leaves enough room
2282 * then convert the TRIE node into a TRIEC node, with the bitmap
2283 * embedded inline in the opcode - this is hypothetically faster.
2284 */
2285 if ( !trie->states[trie->startstate].wordnum
2286 && trie->bitmap
2287 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2288 {
2289 OP( convert ) = TRIEC;
2290 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2291 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2292 trie->bitmap= NULL;
2293 } else
2294 OP( convert ) = TRIE;
a3621e74 2295
3dab1dad
YO
2296 /* store the type in the flags */
2297 convert->flags = nodetype;
a5ca303d
YO
2298 DEBUG_r({
2299 optimize = convert
2300 + NODE_STEP_REGNODE
2301 + regarglen[ OP( convert ) ];
2302 });
2303 /* XXX We really should free up the resource in trie now,
2304 as we won't use them - (which resources?) dmq */
3dab1dad 2305 }
a3621e74 2306 /* needed for dumping*/
e62cc96a 2307 DEBUG_r(if (optimize) {
07be1b83 2308 regnode *opt = convert;
bcdf7404 2309
e62cc96a 2310 while ( ++opt < optimize) {
07be1b83
YO
2311 Set_Node_Offset_Length(opt,0,0);
2312 }
786e8c11
YO
2313 /*
2314 Try to clean up some of the debris left after the
2315 optimisation.
a3621e74 2316 */
786e8c11 2317 while( optimize < jumper ) {
07be1b83 2318 mjd_nodelen += Node_Length((optimize));
a3621e74 2319 OP( optimize ) = OPTIMIZED;
07be1b83 2320 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2321 optimize++;
2322 }
07be1b83 2323 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2324 });
2325 } /* end node insert */
2e64971a
DM
2326
2327 /* Finish populating the prev field of the wordinfo array. Walk back
2328 * from each accept state until we find another accept state, and if
2329 * so, point the first word's .prev field at the second word. If the
2330 * second already has a .prev field set, stop now. This will be the
2331 * case either if we've already processed that word's accept state,
3b753521
FN
2332 * or that state had multiple words, and the overspill words were
2333 * already linked up earlier.
2e64971a
DM
2334 */
2335 {
2336 U16 word;
2337 U32 state;
2338 U16 prev;
2339
2340 for (word=1; word <= trie->wordcount; word++) {
2341 prev = 0;
2342 if (trie->wordinfo[word].prev)
2343 continue;
2344 state = trie->wordinfo[word].accept;
2345 while (state) {
2346 state = prev_states[state];
2347 if (!state)
2348 break;
2349 prev = trie->states[state].wordnum;
2350 if (prev)
2351 break;
2352 }
2353 trie->wordinfo[word].prev = prev;
2354 }
2355 Safefree(prev_states);
2356 }
2357
2358
2359 /* and now dump out the compressed format */
2360 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2361
55eed653 2362 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2363#ifdef DEBUGGING
2364 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2365 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2366#else
2367 SvREFCNT_dec(revcharmap);
07be1b83 2368#endif
786e8c11
YO
2369 return trie->jump
2370 ? MADE_JUMP_TRIE
2371 : trie->startstate>1
2372 ? MADE_EXACT_TRIE
2373 : MADE_TRIE;
2374}
2375
2376STATIC void
2377S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2378{
3b753521 2379/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2380
2381 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2382 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2383 ISBN 0-201-10088-6
2384
2385 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2386 suffix of the current state's 'word' that is also a proper prefix of another word in our
2387 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2388 the DFA not to have to restart after its tried and failed a word at a given point, it
2389 simply continues as though it had been matching the other word in the first place.
2390 Consider
2391 'abcdgu'=~/abcdefg|cdgu/
2392 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2393 fail, which would bring us to the state representing 'd' in the second word where we would
2394 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2395 */
2396 /* add a fail transition */
3251b653
NC
2397 const U32 trie_offset = ARG(source);
2398 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2399 U32 *q;
2400 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2401 const U32 numstates = trie->statecount;
786e8c11
YO
2402 const U32 ubound = trie->lasttrans + ucharcount;
2403 U32 q_read = 0;
2404 U32 q_write = 0;
2405 U32 charid;
2406 U32 base = trie->states[ 1 ].trans.base;
2407 U32 *fail;
2408 reg_ac_data *aho;
2409 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2410 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2411
2412 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2413#ifndef DEBUGGING
2414 PERL_UNUSED_ARG(depth);
2415#endif
2416
2417
2418 ARG_SET( stclass, data_slot );
c944940b 2419 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2420 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2421 aho->trie=trie_offset;
446bd890
NC
2422 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2423 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2424 Newxz( q, numstates, U32);
c944940b 2425 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2426 aho->refcount = 1;
2427 fail = aho->fail;
2428 /* initialize fail[0..1] to be 1 so that we always have
2429 a valid final fail state */
2430 fail[ 0 ] = fail[ 1 ] = 1;
2431
2432 for ( charid = 0; charid < ucharcount ; charid++ ) {
2433 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2434 if ( newstate ) {
2435 q[ q_write ] = newstate;
2436 /* set to point at the root */
2437 fail[ q[ q_write++ ] ]=1;
2438 }
2439 }
2440 while ( q_read < q_write) {
2441 const U32 cur = q[ q_read++ % numstates ];
2442 base = trie->states[ cur ].trans.base;
2443
2444 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2445 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2446 if (ch_state) {
2447 U32 fail_state = cur;
2448 U32 fail_base;
2449 do {
2450 fail_state = fail[ fail_state ];
2451 fail_base = aho->states[ fail_state ].trans.base;
2452 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2453
2454 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2455 fail[ ch_state ] = fail_state;
2456 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2457 {
2458 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2459 }
2460 q[ q_write++ % numstates] = ch_state;
2461 }
2462 }
2463 }
2464 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2465 when we fail in state 1, this allows us to use the
2466 charclass scan to find a valid start char. This is based on the principle
2467 that theres a good chance the string being searched contains lots of stuff
2468 that cant be a start char.
2469 */
2470 fail[ 0 ] = fail[ 1 ] = 0;
2471 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2472 PerlIO_printf(Perl_debug_log,
2473 "%*sStclass Failtable (%"UVuf" states): 0",
2474 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2475 );
786e8c11
YO
2476 for( q_read=1; q_read<numstates; q_read++ ) {
2477 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2478 }
2479 PerlIO_printf(Perl_debug_log, "\n");
2480 });
2481 Safefree(q);
2482 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2483}
2484
786e8c11 2485
a3621e74 2486/*
5d1c421c
JH
2487 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2488 * These need to be revisited when a newer toolchain becomes available.
2489 */
2490#if defined(__sparc64__) && defined(__GNUC__)
2491# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2492# undef SPARC64_GCC_WORKAROUND
2493# define SPARC64_GCC_WORKAROUND 1
2494# endif
2495#endif
2496
07be1b83 2497#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2498 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2499 SV * const mysv=sv_newmortal(); \
2500 regnode *Next = regnext(scan); \
2501 regprop(RExC_rx, mysv, scan); \
7f69552c 2502 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2503 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2504 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2505 }});
07be1b83 2506
1de06328 2507
bb914485
KW
2508/* The below joins as many adjacent EXACTish nodes as possible into a single
2509 * one, and looks for problematic sequences of characters whose folds vs.
2510 * non-folds have sufficiently different lengths, that the optimizer would be
2511 * fooled into rejecting legitimate matches of them, and the trie construction
2512 * code can't cope with them. The joining is only done if:
2513 * 1) there is room in the current conglomerated node to entirely contain the
2514 * next one.
2515 * 2) they are the exact same node type
2516 *
2517 * The adjacent nodes actually may be separated by NOTHING kind nodes, and
2518 * these get optimized out
2519 *
2520 * If there are problematic code sequences, *min_change is set to the delta
2521 * that the minimum size of the node can be off from its actual size. And, the
2522 * node type of the result is changed to reflect that it contains these
2523 * sequences.
2524 *
2525 * This is as good a place as any to discuss the design of handling these
2526 * problematic sequences. It's been wrong in Perl for a very long time. There
2527 * are three code points in Unicode whose folded lengths differ so much from
2528 * the un-folded lengths that it causes problems for the optimizer and trie
2529 * construction. Why only these are problematic, and not others where lengths
2530 * also differ is something I (khw) do not understand. New versions of Unicode
2531 * might add more such code points. Hopefully the logic in fold_grind.t that
287722f3 2532 * figures out what to test (in part by verifying that each size-combination
bb914485 2533 * gets tested) will catch any that do come along, so they can be added to the
287722f3
KW
2534 * special handling below. The chances of new ones are actually rather small,
2535 * as most, if not all, of the world's scripts that have casefolding have
2536 * already been encoded by Unicode. Also, a number of Unicode's decisions were
2537 * made to allow compatibility with pre-existing standards, and almost all of
2538 * those have already been dealt with. These would otherwise be the most
2539 * likely candidates for generating further tricky sequences. In other words,
2540 * Unicode by itself is unlikely to add new ones unless it is for compatibility
2541 * with pre-existing standards.
bb914485
KW
2542 *
2543 * The previous designs for dealing with these involved assigning a special
2544 * node for them. This approach doesn't work, as evidenced by this example:
2545 * "\xDFs" =~ /s\xDF/ui
2546 * Both these fold to "sss", but if the pattern is parsed to create a node of
2547 * that would match just the \xDF, it won't be able to handle the case where a
2548 * successful match would have to cross the node's boundary. The new approach
2549 * that hopefully generally solves the problem generates an EXACTFU_SS node
2550 * that is "sss".
2551 *
2552 * There are a number of components to the approach (a lot of work for just
2553 * three code points!):
2554 * 1) This routine examines each EXACTFish node that could contain the
2555 * problematic sequences, and if found, returns in *min_change the total
2556 * delta between the actual length of the string and one that could match
2557 * it. This delta is used by the caller to adjust the min length of the
2558 * match, and the delta between min and max, so that the optimizer doesn't
2559 * reject these possibilities based on size constraints
2560 * 2) These sequences are not currently correctly handled by the trie code
2561 * either, so it changes the joined node type to ops that are not handled
2562 * by trie's, those new ops being EXACTFU_SS and EXACTFU_NO_TRIE.
2563 * 3) This is sufficient for the two Greek sequences (described below), but
2564 * the one involving the Sharp s (\xDF) needs more. The node type
2565 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2566 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2567 * case where there is a possible fold length change. That means that a
2568 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2569 * itself with length changes, and so can be processed faster. regexec.c
2570 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2571 * is pre-folded by regcomp.c. This saves effort in regex matching.
2572 * However, probably mostly for historical reasons, the pre-folding isn't
2573 * done for non-UTF8 patterns. The fold possibilities for these are quite
2574 * simple, except for the sharp s. All the ones that don't involve a
2575 * UTF-8 target string are members of a fold-pair, and arrays are set up
2576 * for all of them that quickly find the other member of the pair. It
2577 * might actually be faster to pre-fold these, but it isn't currently
2578 * done, except for the sharp s. Code elsewhere in this file makes sure
2579 * that it gets folded to 'ss', even if the pattern isn't UTF-8. This
2580 * avoids the issues describe in the next item.
2581 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2582 * 'ss' or not is not knowable at compile time. It will match iff the
2583 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2584 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2585 * it can't be folded to "ss" at compile time, unlike EXACTFU does as
2586 * described in item 3). An assumption that the optimizer part of
2587 * regexec.c (probably unwittingly) makes is that a character in the
2588 * pattern corresponds to at most a single character in the target string.
2589 * (And I do mean character, and not byte here, unlike other parts of the
2590 * documentation that have never been updated to account for multibyte
2591 * Unicode.) This assumption is wrong only in this case, as all other
2592 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2593 * virtue of having this file pre-fold UTF-8 patterns. I'm
2594 * reluctant to try to change this assumption, so instead the code punts.
2595 * Elsewhere in this file, each EXACTF node is examined for the sharp s.
2596 * If found, a flag is set that later causes the optimizer in this file to
2597 * not set values for the floating and fixed string lengths, and thus
2598 * avoid the optimizer code in regexec.c that makes this invalid
2599 * assumption. Thus, there is no optimization based on string lengths for
2600 * EXACTF nodes that contain the sharp s. This only happens for /id rules
2601 * (which means the pattern isn't in UTF-8).
2602 */
1de06328 2603
f646642f 2604#define JOIN_EXACT(scan,min_change,flags) \
07be1b83 2605 if (PL_regkind[OP(scan)] == EXACT) \
f646642f 2606 join_exact(pRExC_state,(scan),(min_change),(flags),NULL,depth+1)
07be1b83 2607
be8e71aa 2608STATIC U32
f646642f 2609S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, IV *min_change, U32 flags,regnode *val, U32 depth) {
07be1b83
YO
2610 /* Merge several consecutive EXACTish nodes into one. */
2611 regnode *n = regnext(scan);
2612 U32 stringok = 1;
2613 regnode *next = scan + NODE_SZ_STR(scan);
2614 U32 merged = 0;
2615 U32 stopnow = 0;
2616#ifdef DEBUGGING
2617 regnode *stop = scan;
72f13be8 2618 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2619#else
d47053eb
RGS
2620 PERL_UNUSED_ARG(depth);
2621#endif
7918f24d
NC
2622
2623 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2624#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2625 PERL_UNUSED_ARG(flags);
2626 PERL_UNUSED_ARG(val);
07be1b83 2627#endif
07be1b83 2628 DEBUG_PEEP("join",scan,depth);
bb914485 2629
3f410cf6
KW
2630 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2631 * EXACT ones that are mergeable to the current one. */
2632 while (n
2633 && (PL_regkind[OP(n)] == NOTHING
2634 || (stringok && OP(n) == OP(scan)))
07be1b83 2635 && NEXT_OFF(n)
3f410cf6
KW
2636 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2637 {
07be1b83
YO
2638
2639 if (OP(n) == TAIL || n > next)
2640 stringok = 0;
2641 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2642 DEBUG_PEEP("skip:",n,depth);
2643 NEXT_OFF(scan) += NEXT_OFF(n);
2644 next = n + NODE_STEP_REGNODE;
2645#ifdef DEBUGGING
2646 if (stringok)
2647 stop = n;
2648#endif
2649 n = regnext(n);
2650 }
2651 else if (stringok) {
786e8c11 2652 const unsigned int oldl = STR_LEN(scan);
07be1b83 2653 regnode * const nnext = regnext(n);
b2230d39
KW
2654
2655 if (oldl + STR_LEN(n) > U8_MAX)
2656 break;
07be1b83
YO
2657
2658 DEBUG_PEEP("merg",n,depth);
07be1b83 2659 merged++;
b2230d39 2660
07be1b83
YO
2661 NEXT_OFF(scan) += NEXT_OFF(n);
2662 STR_LEN(scan) += STR_LEN(n);
2663 next = n + NODE_SZ_STR(n);
2664 /* Now we can overwrite *n : */
2665 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2666#ifdef DEBUGGING
2667 stop = next - 1;
2668#endif
2669 n = nnext;
2670 if (stopnow) break;
2671 }
2672
d47053eb
RGS
2673#ifdef EXPERIMENTAL_INPLACESCAN
2674 if (flags && !NEXT_OFF(n)) {
2675 DEBUG_PEEP("atch", val, depth);
2676 if (reg_off_by_arg[OP(n)]) {
2677 ARG_SET(n, val - n);
2678 }
2679 else {
2680 NEXT_OFF(n) = val - n;
2681 }
2682 stopnow = 1;
2683 }
07be1b83
YO
2684#endif
2685 }
ced7f090
KW
2686#define GREEK_SMALL_LETTER_IOTA_WITH_DIALYTIKA_AND_TONOS 0x0390
2687#define IOTA_D_T GREEK_SMALL_LETTER_IOTA_WITH_DIALYTIKA_AND_TONOS
2688#define GREEK_SMALL_LETTER_UPSILON_WITH_DIALYTIKA_AND_TONOS 0x03B0
2689#define UPSILON_D_T GREEK_SMALL_LETTER_UPSILON_WITH_DIALYTIKA_AND_TONOS
2c2b7f86 2690
f646642f
KW
2691 *min_change = 0;
2692
3f410cf6
KW
2693 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2694 * can now analyze for sequences of problematic code points. (Prior to
2695 * this final joining, sequences could have been split over boundaries, and
2696 * hence missed). The sequences only happen in folding */
86d6fcad
KW
2697 if (OP(scan) != EXACT) {
2698 char *s, *t;
2699 char * s0 = STRING(scan);
2700 char * const s_end = s0 + STR_LEN(scan);
2701
2702 /* First we look at the sequences that can occur only in UTF-8 strings.
2703 * The sequences are of length 6 */
2704 if (UTF && STR_LEN(scan) >= 6) {
2705
2706 /* Two problematic code points in Unicode casefolding of EXACT
2707 * nodes:
2708 *
2709 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2710 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2711 *
2712 * which casefold to
2713 *
2714 * Unicode UTF-8
2715 *
2716 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2717 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2718 *
2719 * This means that in case-insensitive matching (or "loose
2720 * matching", as Unicode calls it), an EXACTF of length six (the
2721 * UTF-8 encoded byte length of the above casefolded versions) can
2722 * match a target string of length two (the byte length of UTF-8
2723 * encoded U+0390 or U+03B0). This would rather mess up the
2724 * minimum length computation. (there are other code points that
2725 * also fold to these two sequences, but the delta is smaller)
2726 *
2727 * What we'll do is to look for the tail four bytes, and then peek
2728 * at the preceding two bytes to see whether we need to decrease
2729 * the minimum length by four (six minus two).
2730 *
2731 * Thanks to the design of UTF-8, there cannot be false matches:
2732 * A sequence of valid UTF-8 bytes cannot be a subsequence of
2733 * another valid sequence of UTF-8 bytes. */
3f410cf6 2734
e294cc5d 2735#ifdef EBCDIC /* RD tunifold greek 0390 and 03B0 */
86d6fcad
KW
2736 const char U390_first_byte = '\xb4';
2737 const char U390_2nd_byte = '\x68';
2738 const char U3B0_first_byte = '\xb5';
2739 const char U3B0_2nd_byte = '\x46';
2740 const char tail[] = "\xaf\x49\xaf\x42";
e294cc5d 2741#else
86d6fcad
KW
2742 const char U390_first_byte = '\xce';
2743 const char U390_2nd_byte = '\xb9';
2744 const char U3B0_first_byte = '\xcf';
2745 const char U3B0_2nd_byte = '\x85';
2746 const char tail[] = "\xcc\x88\xcc\x81";
e294cc5d 2747#endif
86d6fcad
KW
2748 const STRLEN tail_len = sizeof(tail) - 1;
2749 for (s = s0 + 2; /* +2 is to skip the non-tail */
2750 s <= s_end - tail_len
2751 && (t = ninstr(s, s_end, tail, tail + tail_len));
2752 s = t + tail_len)
2753 {
2754 if ((t[-1] == U390_2nd_byte && t[-2] == U390_first_byte)
2755 || (t[-1] == U3B0_2nd_byte && t[-2] == U3B0_first_byte))
2756 {
f646642f 2757 *min_change -= 4;
bb914485
KW
2758
2759 /* This can't currently be handled by tries, so change the
2760 * node type to indicate this. */
2761 if (OP(scan) == EXACTFU) {
2762 OP(scan) = EXACTFU_NO_TRIE;
2763 }
2764 }
2765 }
2766 }
2767
2768 /* The third problematic sequence is 'ss', which can match just the
2769 * single byte LATIN SMALL LETTER SHARP S, and it can do it in both
2770 * non- and UTF-8. Code elsewhere in this file makes sure, however,
2771 * that the sharp s gets folded to 'ss' under Unicode rules even if not
2772 * UTF-8. */
2773 if (STR_LEN(scan) >= 2
2774 && (OP(scan) == EXACTFU
2775 || OP(scan) == EXACTFU_NO_TRIE /* The code above could have
2776 set to this node type */
2777 || OP(scan) == EXACTF))
2778 {
2779 /* The string will be folded to 'ss' if it's in UTF-8, but it could
2780 * include capital 'S' instead of lower case when not UTF-8. We
2781 * could have different code to handle the two cases, but this is
2782 * not necessary since both S and s are invariants under UTF-8; and
2783 * not worth it, especially because we can use just one test for
2784 * either 'S' or 's' each * time through the loop (plus a mask).
2785 * Ths is because on both EBCDIC and ASCII machines, 'S' and 's'
2786 * differ by a single bit. On ASCII they are 32 apart; on EBCDIC,
2787 * they are 64. This uses an exclusive 'or' to find that bit and
2788 * then inverts it to form a mask, with just a single 0, in the bit
2789 * position where 'S' and 's' differ. */
2790 const char S_or_s_mask = ~ ('S' ^ 's');
2791 const char s_masked = 's' & S_or_s_mask;
2792
2793 for (s = s0; s < s_end - 1; s++) {
2794 if (((*s & S_or_s_mask) == s_masked)
2795 && ((*(s+1) & S_or_s_mask) == s_masked))
2796 {
2797 s++;
2798 *min_change -= 1;
2799
2800 /* EXACTFU_SS also isn't trie'able, so don't have to
2801 * preserve EXACTFU_NO_TRIE. EXACTF is also not trie'able,
2802 * and because we essentially punt the optimizations in its
2803 * case, we don't need to indicate that it has an ss */
2804 if (OP(scan) == EXACTFU || OP(scan) == EXACTFU_NO_TRIE) {
2805 OP(scan) = EXACTFU_SS;
2806 }
86d6fcad
KW
2807 }
2808 }
2809 }
07be1b83 2810 }
3f410cf6 2811
07be1b83 2812#ifdef DEBUGGING
bb789b09
DM
2813 /* Allow dumping but overwriting the collection of skipped
2814 * ops and/or strings with fake optimized ops */
07be1b83
YO
2815 n = scan + NODE_SZ_STR(scan);
2816 while (n <= stop) {
bb789b09
DM
2817 OP(n) = OPTIMIZED;
2818 FLAGS(n) = 0;
2819 NEXT_OFF(n) = 0;
07be1b83
YO
2820 n++;
2821 }
2822#endif
2823 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2824 return stopnow;
2825}
2826
486ec47a 2827/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2828 Finds fixed substrings. */
2829
a0288114 2830/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
2831 to the position after last scanned or to NULL. */
2832
40d049e4
YO
2833#define INIT_AND_WITHP \
2834 assert(!and_withp); \
2835 Newx(and_withp,1,struct regnode_charclass_class); \
2836 SAVEFREEPV(and_withp)
07be1b83 2837
b515a41d 2838/* this is a chain of data about sub patterns we are processing that
486ec47a 2839 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
2840 we can simulate recursion without losing state. */
2841struct scan_frame;
2842typedef struct scan_frame {
2843 regnode *last; /* last node to process in this frame */
2844 regnode *next; /* next node to process when last is reached */
2845 struct scan_frame *prev; /*previous frame*/
2846 I32 stop; /* what stopparen do we use */
2847} scan_frame;
2848
304ee84b
YO
2849
2850#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2851
e1d1eefb
YO
2852#define CASE_SYNST_FNC(nAmE) \
2853case nAmE: \
2854 if (flags & SCF_DO_STCLASS_AND) { \
2855 for (value = 0; value < 256; value++) \
2856 if (!is_ ## nAmE ## _cp(value)) \
2857 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2858 } \
2859 else { \
2860 for (value = 0; value < 256; value++) \
2861 if (is_ ## nAmE ## _cp(value)) \
2862 ANYOF_BITMAP_SET(data->start_class, value); \
2863 } \
2864 break; \
2865case N ## nAmE: \
2866 if (flags & SCF_DO_STCLASS_AND) { \
2867 for (value = 0; value < 256; value++) \
2868 if (is_ ## nAmE ## _cp(value)) \
2869 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2870 } \
2871 else { \
2872 for (value = 0; value < 256; value++) \
2873 if (!is_ ## nAmE ## _cp(value)) \
2874 ANYOF_BITMAP_SET(data->start_class, value); \
2875 } \
2876 break
2877
2878
2879
76e3520e 2880STATIC I32
40d049e4 2881S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 2882 I32 *minlenp, I32 *deltap,
40d049e4
YO
2883 regnode *last,
2884 scan_data_t *data,
2885 I32 stopparen,
2886 U8* recursed,
2887 struct regnode_charclass_class *and_withp,
2888 U32 flags, U32 depth)
c277df42
IZ
2889 /* scanp: Start here (read-write). */
2890 /* deltap: Write maxlen-minlen here. */
2891 /* last: Stop before this one. */
40d049e4
YO
2892 /* data: string data about the pattern */
2893 /* stopparen: treat close N as END */
2894 /* recursed: which subroutines have we recursed into */
2895 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 2896{
97aff369 2897 dVAR;
c277df42
IZ
2898 I32 min = 0, pars = 0, code;
2899 regnode *scan = *scanp, *next;
2900 I32 delta = 0;
2901 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 2902 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
2903 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2904 scan_data_t data_fake;
a3621e74 2905 SV *re_trie_maxbuff = NULL;
786e8c11 2906 regnode *first_non_open = scan;
e2e6a0f1 2907 I32 stopmin = I32_MAX;
8aa23a47 2908 scan_frame *frame = NULL;
a3621e74 2909 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 2910
7918f24d
NC
2911 PERL_ARGS_ASSERT_STUDY_CHUNK;
2912
13a24bad 2913#ifdef DEBUGGING
40d049e4 2914 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 2915#endif
40d049e4 2916
786e8c11 2917 if ( depth == 0 ) {
40d049e4 2918 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
2919 first_non_open=regnext(first_non_open);
2920 }
2921
b81d288d 2922
8aa23a47
YO
2923 fake_study_recurse:
2924 while ( scan && OP(scan) != END && scan < last ){
f646642f 2925 IV min_change = 0;
8aa23a47 2926 /* Peephole optimizer: */
304ee84b 2927 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 2928 DEBUG_PEEP("Peep",scan,depth);
f646642f 2929 JOIN_EXACT(scan,&min_change,0);
8aa23a47
YO
2930
2931 /* Follow the next-chain of the current node and optimize
2932 away all the NOTHINGs from it. */
2933 if (OP(scan) != CURLYX) {
2934 const int max = (reg_off_by_arg[OP(scan)]
2935 ? I32_MAX
2936 /* I32 may be smaller than U16 on CRAYs! */
2937 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
2938 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
2939 int noff;
2940 regnode *n = scan;
686b73d4 2941
8aa23a47
YO
2942 /* Skip NOTHING and LONGJMP. */
2943 while ((n = regnext(n))
2944 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
2945 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
2946 && off + noff < max)
2947 off += noff;
2948 if (reg_off_by_arg[OP(scan)])
2949 ARG(scan) = off;
2950 else
2951 NEXT_OFF(scan) = off;
2952 }
a3621e74 2953
c277df42 2954
8aa23a47
YO
2955
2956 /* The principal pseudo-switch. Cannot be a switch, since we
2957 look into several different things. */
2958 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
2959 || OP(scan) == IFTHEN) {
2960 next = regnext(scan);
2961 code = OP(scan);
2962 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 2963
8aa23a47
YO
2964 if (OP(next) == code || code == IFTHEN) {
2965 /* NOTE - There is similar code to this block below for handling
2966 TRIE nodes on a re-study. If you change stuff here check there
2967 too. */
2968 I32 max1 = 0, min1 = I32_MAX, num = 0;
2969 struct regnode_charclass_class accum;
2970 regnode * const startbranch=scan;
686b73d4 2971
8aa23a47 2972 if (flags & SCF_DO_SUBSTR)
304ee84b 2973 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 2974 if (flags & SCF_DO_STCLASS)
e755fd73 2975 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
2976
2977 while (OP(scan) == code) {
2978 I32 deltanext, minnext, f = 0, fake;
2979 struct regnode_charclass_class this_class;
2980
2981 num++;
2982 data_fake.flags = 0;
2983 if (data) {
2984 data_fake.whilem_c = data->whilem_c;
2985 data_fake.last_closep = data->last_closep;
2986 }
2987 else
2988 data_fake.last_closep = &fake;
58e23c8d
YO
2989
2990 data_fake.pos_delta = delta;
8aa23a47
YO
2991 next = regnext(scan);
2992 scan = NEXTOPER(scan);
2993 if (code != BRANCH)
c277df42 2994 scan = NEXTOPER(scan);
8aa23a47 2995 if (flags & SCF_DO_STCLASS) {
e755fd73 2996 cl_init(pRExC_state, &this_class);
8aa23a47
YO
2997 data_fake.start_class = &this_class;
2998 f = SCF_DO_STCLASS_AND;
58e23c8d 2999 }
8aa23a47
YO
3000 if (flags & SCF_WHILEM_VISITED_POS)
3001 f |= SCF_WHILEM_VISITED_POS;
3002
3003 /* we suppose the run is continuous, last=next...*/
3004 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3005 next, &data_fake,
3006 stopparen, recursed, NULL, f,depth+1);
3007 if (min1 > minnext)
3008 min1 = minnext;
3009 if (max1 < minnext + deltanext)
3010 max1 = minnext + deltanext;
3011 if (deltanext == I32_MAX)
3012 is_inf = is_inf_internal = 1;
3013 scan = next;
3014 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3015 pars++;
3016 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3017 if ( stopmin > minnext)
3018 stopmin = min + min1;
3019 flags &= ~SCF_DO_SUBSTR;
3020 if (data)
3021 data->flags |= SCF_SEEN_ACCEPT;
3022 }
3023 if (data) {
3024 if (data_fake.flags & SF_HAS_EVAL)
3025 data->flags |= SF_HAS_EVAL;
3026 data->whilem_c = data_fake.whilem_c;
3dab1dad 3027 }
8aa23a47 3028 if (flags & SCF_DO_STCLASS)
3fffb88a 3029 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3030 }
3031 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3032 min1 = 0;
3033 if (flags & SCF_DO_SUBSTR) {
3034 data->pos_min += min1;
3035 data->pos_delta += max1 - min1;
3036 if (max1 != min1 || is_inf)
3037 data->longest = &(data->longest_float);
3038 }
3039 min += min1;
3040 delta += max1 - min1;
3041 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3042 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3043 if (min1) {
3044 cl_and(data->start_class, and_withp);
3045 flags &= ~SCF_DO_STCLASS;
653099ff 3046 }
8aa23a47
YO
3047 }
3048 else if (flags & SCF_DO_STCLASS_AND) {
3049 if (min1) {
3050 cl_and(data->start_class, &accum);
3051 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3052 }
8aa23a47
YO
3053 else {
3054 /* Switch to OR mode: cache the old value of
3055 * data->start_class */
3056 INIT_AND_WITHP;
3057 StructCopy(data->start_class, and_withp,
3058 struct regnode_charclass_class);
3059 flags &= ~SCF_DO_STCLASS_AND;
3060 StructCopy(&accum, data->start_class,
3061 struct regnode_charclass_class);
3062 flags |= SCF_DO_STCLASS_OR;
3063 data->start_class->flags |= ANYOF_EOS;
de0c8cb8 3064 }
8aa23a47 3065 }
a3621e74 3066
8aa23a47
YO
3067 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3068 /* demq.
a3621e74 3069
8aa23a47
YO
3070 Assuming this was/is a branch we are dealing with: 'scan' now
3071 points at the item that follows the branch sequence, whatever
3072 it is. We now start at the beginning of the sequence and look
3073 for subsequences of
a3621e74 3074
8aa23a47
YO
3075 BRANCH->EXACT=>x1
3076 BRANCH->EXACT=>x2
3077 tail
a3621e74 3078
8aa23a47 3079 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3080
486ec47a 3081 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3082 element into a trie and then add the subsequent branch exact
3083 strings to the trie.
a3621e74 3084
8aa23a47 3085 We have two cases
a3621e74 3086
3b753521 3087 1. patterns where the whole set of branches can be converted.
a3621e74 3088
8aa23a47 3089 2. patterns where only a subset can be converted.
a3621e74 3090
8aa23a47
YO
3091 In case 1 we can replace the whole set with a single regop
3092 for the trie. In case 2 we need to keep the start and end
3b753521 3093 branches so
a3621e74 3094
8aa23a47
YO
3095 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3096 becomes BRANCH TRIE; BRANCH X;
786e8c11 3097
8aa23a47
YO
3098 There is an additional case, that being where there is a
3099 common prefix, which gets split out into an EXACT like node
3100 preceding the TRIE node.
a3621e74 3101
8aa23a47
YO
3102 If x(1..n)==tail then we can do a simple trie, if not we make
3103 a "jump" trie, such that when we match the appropriate word
486ec47a 3104 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3105 a nested if into a case structure of sorts.
b515a41d 3106
8aa23a47 3107 */
686b73d4 3108
8aa23a47
YO
3109 int made=0;
3110 if (!re_trie_maxbuff) {
3111 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3112 if (!SvIOK(re_trie_maxbuff))
3113 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3114 }
3115 if ( SvIV(re_trie_maxbuff)>=0 ) {
3116 regnode *cur;
3117 regnode *first = (regnode *)NULL;
3118 regnode *last = (regnode *)NULL;
3119 regnode *tail = scan;
3120 U8 optype = 0;
3121 U32 count=0;
a3621e74
YO
3122
3123#ifdef DEBUGGING
8aa23a47 3124 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3125#endif
8aa23a47
YO
3126 /* var tail is used because there may be a TAIL
3127 regop in the way. Ie, the exacts will point to the
3128 thing following the TAIL, but the last branch will
3129 point at the TAIL. So we advance tail. If we
3130 have nested (?:) we may have to move through several
3131 tails.
3132 */
3133
3134 while ( OP( tail ) == TAIL ) {
3135 /* this is the TAIL generated by (?:) */
3136 tail = regnext( tail );
3137 }
a3621e74 3138
8aa23a47
YO
3139
3140 DEBUG_OPTIMISE_r({
3141 regprop(RExC_rx, mysv, tail );
3142 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3143 (int)depth * 2 + 2, "",
3144 "Looking for TRIE'able sequences. Tail node is: ",
3145 SvPV_nolen_const( mysv )
3146 );
3147 });
3148
3149 /*
3150
3151 step through the branches, cur represents each
3152 branch, noper is the first thing to be matched
3153 as part of that branch and noper_next is the
3154 regnext() of that node. if noper is an EXACT
3155 and noper_next is the same as scan (our current
3156 position in the regex) then the EXACT branch is
3157 a possible optimization target. Once we have
486ec47a 3158 two or more consecutive such branches we can
8aa23a47
YO
3159 create a trie of the EXACT's contents and stich
3160 it in place. If the sequence represents all of
3161 the branches we eliminate the whole thing and
3162 replace it with a single TRIE. If it is a
3163 subsequence then we need to stitch it in. This
3164 means the first branch has to remain, and needs
3165 to be repointed at the item on the branch chain
3166 following the last branch optimized. This could
3167 be either a BRANCH, in which case the
3168 subsequence is internal, or it could be the
3169 item following the branch sequence in which
3170 case the subsequence is at the end.
3171
3172 */
3173
3174 /* dont use tail as the end marker for this traverse */
3175 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3176 regnode * const noper = NEXTOPER( cur );
b515a41d 3177#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3178 regnode * const noper_next = regnext( noper );
b515a41d
YO
3179#endif
3180
8aa23a47
YO
3181 DEBUG_OPTIMISE_r({
3182 regprop(RExC_rx, mysv, cur);
3183 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3184 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3185
3186 regprop(RExC_rx, mysv, noper);
3187 PerlIO_printf( Perl_debug_log, " -> %s",
3188 SvPV_nolen_const(mysv));
3189
3190 if ( noper_next ) {
3191 regprop(RExC_rx, mysv, noper_next );
3192 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3193 SvPV_nolen_const(mysv));
3194 }
3195 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d)\n",
3196 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur) );
3197 });
3198 if ( (((first && optype!=NOTHING) ? OP( noper ) == optype
3199 : PL_regkind[ OP( noper ) ] == EXACT )
3200 || OP(noper) == NOTHING )
786e8c11 3201#ifdef NOJUMPTRIE
8aa23a47 3202 && noper_next == tail
786e8c11 3203#endif
8aa23a47
YO
3204 && count < U16_MAX)
3205 {
3206 count++;
3207 if ( !first || optype == NOTHING ) {
3208 if (!first) first = cur;
3209 optype = OP( noper );
3210 } else {
3211 last = cur;
3212 }
3213 } else {
a0a388a1 3214/*
fbebf34e
KW
3215 Currently the trie logic handles case insensitive matching properly only
3216 when the pattern is UTF-8 and the node is EXACTFU (thus forcing unicode
3217 semantics).
0abd0d78
YO
3218
3219 If/when this is fixed the following define can be swapped
3220 in below to fully enable trie logic.
3221
a0a388a1 3222#define TRIE_TYPE_IS_SAFE 1
0abd0d78
YO
3223
3224*/
fbebf34e 3225#define TRIE_TYPE_IS_SAFE ((UTF && optype == EXACTFU) || optype==EXACT)
0abd0d78 3226
a0a388a1 3227 if ( last && TRIE_TYPE_IS_SAFE ) {
8aa23a47
YO
3228 make_trie( pRExC_state,
3229 startbranch, first, cur, tail, count,
3230 optype, depth+1 );
3231 }
3232 if ( PL_regkind[ OP( noper ) ] == EXACT
786e8c11 3233#ifdef NOJUMPTRIE
8aa23a47 3234 && noper_next == tail
786e8c11 3235#endif
8aa23a47
YO
3236 ){
3237 count = 1;
3238 first = cur;
3239 optype = OP( noper );
3240 } else {
3241 count = 0;
3242 first = NULL;
3243 optype = 0;
3244 }
3245 last = NULL;
3246 }
3247 }
3248 DEBUG_OPTIMISE_r({
3249 regprop(RExC_rx, mysv, cur);
3250 PerlIO_printf( Perl_debug_log,
3251 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3252 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3253
3254 });
a0a388a1
YO
3255
3256 if ( last && TRIE_TYPE_IS_SAFE ) {
8aa23a47 3257 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, optype, depth+1 );
686b73d4 3258#ifdef TRIE_STUDY_OPT
8aa23a47
YO
3259 if ( ((made == MADE_EXACT_TRIE &&
3260 startbranch == first)
3261 || ( first_non_open == first )) &&
3262 depth==0 ) {
3263 flags |= SCF_TRIE_RESTUDY;
3264 if ( startbranch == first
3265 && scan == tail )
3266 {
3267 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3268 }
3269 }
3dab1dad 3270#endif
8aa23a47
YO
3271 }
3272 }
3273
3274 } /* do trie */
3275
653099ff 3276 }
8aa23a47
YO
3277 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3278 scan = NEXTOPER(NEXTOPER(scan));
3279 } else /* single branch is optimized. */
3280 scan = NEXTOPER(scan);
3281 continue;
3282 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3283 scan_frame *newframe = NULL;
3284 I32 paren;
3285 regnode *start;
3286 regnode *end;
3287
3288 if (OP(scan) != SUSPEND) {
3289 /* set the pointer */
3290 if (OP(scan) == GOSUB) {
3291 paren = ARG(scan);
3292 RExC_recurse[ARG2L(scan)] = scan;
3293 start = RExC_open_parens[paren-1];
3294 end = RExC_close_parens[paren-1];
3295 } else {
3296 paren = 0;
f8fc2ecf 3297 start = RExC_rxi->program + 1;
8aa23a47
YO
3298 end = RExC_opend;
3299 }
3300 if (!recursed) {
3301 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3302 SAVEFREEPV(recursed);
3303 }
3304 if (!PAREN_TEST(recursed,paren+1)) {
3305 PAREN_SET(recursed,paren+1);
3306 Newx(newframe,1,scan_frame);
3307 } else {
3308 if (flags & SCF_DO_SUBSTR) {
304ee84b 3309 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3310 data->longest = &(data->longest_float);
3311 }
3312 is_inf = is_inf_internal = 1;
3313 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3314 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3315 flags &= ~SCF_DO_STCLASS;
3316 }
3317 } else {
3318 Newx(newframe,1,scan_frame);
3319 paren = stopparen;
3320 start = scan+2;
3321 end = regnext(scan);
3322 }
3323 if (newframe) {
3324 assert(start);
3325 assert(end);
3326 SAVEFREEPV(newframe);
3327 newframe->next = regnext(scan);
3328 newframe->last = last;
3329 newframe->stop = stopparen;
3330 newframe->prev = frame;
3331
3332 frame = newframe;
3333 scan = start;
3334 stopparen = paren;
3335 last = end;
3336
3337 continue;
3338 }
3339 }
3340 else if (OP(scan) == EXACT) {
3341 I32 l = STR_LEN(scan);
3342 UV uc;
3343 if (UTF) {
3344 const U8 * const s = (U8*)STRING(scan);
3345 l = utf8_length(s, s + l);
3346 uc = utf8_to_uvchr(s, NULL);
3347 } else {
3348 uc = *((U8*)STRING(scan));
3349 }
3350 min += l;
3351 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3352 /* The code below prefers earlier match for fixed
3353 offset, later match for variable offset. */
3354 if (data->last_end == -1) { /* Update the start info. */
3355 data->last_start_min = data->pos_min;
3356 data->last_start_max = is_inf
3357 ? I32_MAX : data->pos_min + data->pos_delta;
b515a41d 3358 }
8aa23a47
YO
3359 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3360 if (UTF)
3361 SvUTF8_on(data->last_found);
3362 {
3363 SV * const sv = data->last_found;
3364 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3365 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3366 if (mg && mg->mg_len >= 0)
3367 mg->mg_len += utf8_length((U8*)STRING(scan),
3368 (U8*)STRING(scan)+STR_LEN(scan));
b515a41d 3369 }
8aa23a47
YO
3370 data->last_end = data->pos_min + l;
3371 data->pos_min += l; /* As in the first entry. */
3372 data->flags &= ~SF_BEFORE_EOL;
3373 }
3374 if (flags & SCF_DO_STCLASS_AND) {
3375 /* Check whether it is compatible with what we know already! */
3376 int compat = 1;
3377
54251c2e 3378
486ec47a 3379 /* If compatible, we or it in below. It is compatible if is
54251c2e
KW
3380 * in the bitmp and either 1) its bit or its fold is set, or 2)
3381 * it's for a locale. Even if there isn't unicode semantics
3382 * here, at runtime there may be because of matching against a
3383 * utf8 string, so accept a possible false positive for
3384 * latin1-range folds */
8aa23a47
YO
3385 if (uc >= 0x100 ||
3386 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3387 && !ANYOF_BITMAP_TEST(data->start_class, uc)
39065660 3388 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
54251c2e 3389 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
8aa23a47 3390 )
d18bf9dc 3391 {
8aa23a47 3392 compat = 0;
d18bf9dc 3393 }
8aa23a47
YO
3394 ANYOF_CLASS_ZERO(data->start_class);
3395 ANYOF_BITMAP_ZERO(data->start_class);
3396 if (compat)
3397 ANYOF_BITMAP_SET(data->start_class, uc);
d18bf9dc
KW
3398 else if (uc >= 0x100) {
3399 int i;
3400
3401 /* Some Unicode code points fold to the Latin1 range; as
3402 * XXX temporary code, instead of figuring out if this is
3403 * one, just assume it is and set all the start class bits
3404 * that could be some such above 255 code point's fold
3405 * which will generate fals positives. As the code
3406 * elsewhere that does compute the fold settles down, it
3407 * can be extracted out and re-used here */
3408 for (i = 0; i < 256; i++){
3409 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3410 ANYOF_BITMAP_SET(data->start_class, i);
3411 }
3412 }
3413 }
8aa23a47
YO
3414 data->start_class->flags &= ~ANYOF_EOS;
3415 if (uc < 0x100)
3416 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3417 }
3418 else if (flags & SCF_DO_STCLASS_OR) {
3419 /* false positive possible if the class is case-folded */
3420 if (uc < 0x100)
3421 ANYOF_BITMAP_SET(data->start_class, uc);
3422 else
3423 data->start_class->flags |= ANYOF_UNICODE_ALL;
3424 data->start_class->flags &= ~ANYOF_EOS;
3425 cl_and(data->start_class, and_withp);
3426 }
3427 flags &= ~SCF_DO_STCLASS;
3428 }
3429 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3430 I32 l = STR_LEN(scan);
3431 UV uc = *((U8*)STRING(scan));
3432
3433 /* Search for fixed substrings supports EXACT only. */
3434 if (flags & SCF_DO_SUBSTR) {
3435 assert(data);
304ee84b 3436 SCAN_COMMIT(pRExC_state, data, minlenp);
8aa23a47
YO
3437 }
3438 if (UTF) {
3439 const U8 * const s = (U8 *)STRING(scan);
3440 l = utf8_length(s, s + l);
3441 uc = utf8_to_uvchr(s, NULL);
3442 }
bb914485
KW
3443 else if (OP(scan) == EXACTF) {
3444 if (memchr(STRING(scan), LATIN_SMALL_LETTER_SHARP_S, l)) {
3445 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3446 }
3447 }
f646642f
KW
3448 min += l + min_change;
3449 if (min < 0) {
3450 min = 0;
3451 }
d2197104 3452 delta += abs(min_change);
f646642f
KW
3453 if (flags & SCF_DO_SUBSTR) {
3454 data->pos_min += l + min_change;
3455 if (data->pos_min < 0) {
3456 data->pos_min = 0;
3457 }
d2197104
KW
3458 data->pos_delta += abs(min_change);
3459 if (min_change) {
3460 data->longest = &(data->longest_float);
3461 }
f646642f 3462 }
8aa23a47
YO
3463 if (flags & SCF_DO_STCLASS_AND) {
3464 /* Check whether it is compatible with what we know already! */
3465 int compat = 1;
8aa23a47 3466 if (uc >= 0x100 ||
54251c2e
KW
3467 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3468 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3469 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3470 {
8aa23a47 3471 compat = 0;
54251c2e 3472 }
8aa23a47
YO
3473 ANYOF_CLASS_ZERO(data->start_class);
3474 ANYOF_BITMAP_ZERO(data->start_class);
3475 if (compat) {
3476 ANYOF_BITMAP_SET(data->start_class, uc);
653099ff 3477 data->start_class->flags &= ~ANYOF_EOS;
39065660 3478 data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
970c8436 3479 if (OP(scan) == EXACTFL) {
af302e7f
KW
3480 /* XXX This set is probably no longer necessary, and
3481 * probably wrong as LOCALE now is on in the initial
3482 * state */
8aa23a47 3483 data->start_class->flags |= ANYOF_LOCALE;
970c8436
KW
3484 }
3485 else {
3486
54251c2e
KW
3487 /* Also set the other member of the fold pair. In case
3488 * that unicode semantics is called for at runtime, use
3489 * the full latin1 fold. (Can't do this for locale,
3490 * because not known until runtime */
3491 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
e22b340a
KW
3492
3493 /* All folds except under /iaa that include s, S, and
3494 * sharp_s also may include the others */
3495 if (OP(scan) != EXACTFA) {
3496 if (uc == 's' || uc == 'S') {
3497 ANYOF_BITMAP_SET(data->start_class,
3498 LATIN_SMALL_LETTER_SHARP_S);
3499 }
3500 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3501 ANYOF_BITMAP_SET(data->start_class, 's');
3502 ANYOF_BITMAP_SET(data->start_class, 'S');
3503 }
3504 }
970c8436 3505 }
653099ff 3506 }
d18bf9dc
KW
3507 else if (uc >= 0x100) {
3508 int i;
3509 for (i = 0; i < 256; i++){
3510 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3511 ANYOF_BITMAP_SET(data->start_class, i);
3512 }
3513 }
3514 }
8aa23a47
YO
3515 }
3516 else if (flags & SCF_DO_STCLASS_OR) {
39065660 3517 if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
8aa23a47
YO
3518 /* false positive possible if the class is case-folded.
3519 Assume that the locale settings are the same... */
970c8436 3520 if (uc < 0x100) {
1aa99e6b 3521 ANYOF_BITMAP_SET(data->start_class, uc);
970c8436
KW
3522 if (OP(scan) != EXACTFL) {
3523
3524 /* And set the other member of the fold pair, but
3525 * can't do that in locale because not known until
3526 * run-time */
3527 ANYOF_BITMAP_SET(data->start_class,
54251c2e 3528 PL_fold_latin1[uc]);
e22b340a
KW
3529
3530 /* All folds except under /iaa that include s, S,
3531 * and sharp_s also may include the others */
3532 if (OP(scan) != EXACTFA) {
3533 if (uc == 's' || uc == 'S') {
3534 ANYOF_BITMAP_SET(data->start_class,
3535 LATIN_SMALL_LETTER_SHARP_S);
3536 }
3537 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3538 ANYOF_BITMAP_SET(data->start_class, 's');
3539 ANYOF_BITMAP_SET(data->start_class, 'S');
3540 }
3541 }
970c8436
KW
3542 }
3543 }
653099ff
GS
3544 data->start_class->flags &= ~ANYOF_EOS;
3545 }
8aa23a47 3546 cl_and(data->start_class, and_withp);
653099ff 3547 }
8aa23a47
YO
3548 flags &= ~SCF_DO_STCLASS;
3549 }
e52fc539 3550 else if (REGNODE_VARIES(OP(scan))) {
8aa23a47
YO
3551 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3552 I32 f = flags, pos_before = 0;
3553 regnode * const oscan = scan;
3554 struct regnode_charclass_class this_class;
3555 struct regnode_charclass_class *oclass = NULL;
3556 I32 next_is_eval = 0;
3557
3558 switch (PL_regkind[OP(scan)]) {
3559 case WHILEM: /* End of (?:...)* . */
3560 scan = NEXTOPER(scan);
3561 goto finish;
3562 case PLUS:
3563 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3564 next = NEXTOPER(scan);
3565 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3566 mincount = 1;
3567 maxcount = REG_INFTY;
3568 next = regnext(scan);
3569 scan = NEXTOPER(scan);
3570 goto do_curly;
3571 }
3572 }
3573 if (flags & SCF_DO_SUBSTR)
3574 data->pos_min++;
3575 min++;
3576 /* Fall through. */
3577 case STAR:
3578 if (flags & SCF_DO_STCLASS) {
3579 mincount = 0;
3580 maxcount = REG_INFTY;
3581 next = regnext(scan);
3582 scan = NEXTOPER(scan);
3583 goto do_curly;
3584 }
3585 is_inf = is_inf_internal = 1;
3586 scan = regnext(scan);
c277df42 3587 if (flags & SCF_DO_SUBSTR) {
304ee84b 3588 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
8aa23a47 3589 data->longest = &(data->longest_float);
c277df42 3590 }
8aa23a47
YO
3591 goto optimize_curly_tail;
3592 case CURLY:
3593 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3594 && (scan->flags == stopparen))
3595 {
3596 mincount = 1;
3597 maxcount = 1;
3598 } else {
3599 mincount = ARG1(scan);
3600 maxcount = ARG2(scan);
653099ff 3601 }
8aa23a47
YO
3602 next = regnext(scan);
3603 if (OP(scan) == CURLYX) {
3604 I32 lp = (data ? *(data->last_closep) : 0);
3605 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
653099ff 3606 }
8aa23a47
YO
3607 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3608 next_is_eval = (OP(scan) == EVAL);
3609 do_curly:
3610 if (flags & SCF_DO_SUBSTR) {
304ee84b 3611 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
8aa23a47 3612 pos_before = data->pos_min;
b45f050a 3613 }
8aa23a47
YO
3614 if (data) {
3615 fl = data->flags;
3616 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3617 if (is_inf)
3618 data->flags |= SF_IS_INF;
3619 }
3620 if (flags & SCF_DO_STCLASS) {
e755fd73 3621 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3622 oclass = data->start_class;
3623 data->start_class = &this_class;
3624 f |= SCF_DO_STCLASS_AND;
3625 f &= ~SCF_DO_STCLASS_OR;
3626 }
779bcb7d
NC
3627 /* Exclude from super-linear cache processing any {n,m}
3628 regops for which the combination of input pos and regex
3629 pos is not enough information to determine if a match
3630 will be possible.
3631
3632 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3633 regex pos at the \s*, the prospects for a match depend not
3634 only on the input position but also on how many (bar\s*)
3635 repeats into the {4,8} we are. */
3636 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
8aa23a47 3637 f &= ~SCF_WHILEM_VISITED_POS;
b45f050a 3638
8aa23a47
YO
3639 /* This will finish on WHILEM, setting scan, or on NULL: */
3640 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3641 last, data, stopparen, recursed, NULL,
3642 (mincount == 0
3643 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
b515a41d 3644
8aa23a47
YO
3645 if (flags & SCF_DO_STCLASS)
3646 data->start_class = oclass;
3647 if (mincount == 0 || minnext == 0) {
3648 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3649 cl_or(pRExC_state, data->start_class, &this_class);
8aa23a47
YO
3650 }
3651 else if (flags & SCF_DO_STCLASS_AND) {
3652 /* Switch to OR mode: cache the old value of
3653 * data->start_class */
3654 INIT_AND_WITHP;
3655 StructCopy(data->start_class, and_withp,
3656 struct regnode_charclass_class);