This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c: Remove unnecessary flag setting
[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"
40b6423c 84extern const struct regexp_engine my_reg_engine;
54df2634
NC
85#else
86# include "regcomp.h"
87#endif
a687059c 88
04e98a4d 89#include "dquote_static.c"
26faadbd 90#include "charclass_invlists.h"
81e983c1 91#include "inline_invlist.c"
1b0f46bf 92#include "unicode_constants.h"
04e98a4d 93
94dc5c2d 94#define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
26faadbd 95#define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
2c61f163 96#define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
94dc5c2d 97
d4cce5f1 98#ifdef op
11343788 99#undef op
d4cce5f1 100#endif /* op */
11343788 101
fe14fcc3 102#ifdef MSDOS
7e4e8c89 103# if defined(BUGGY_MSC6)
fe14fcc3 104 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
7e4e8c89 105# pragma optimize("a",off)
fe14fcc3 106 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
7e4e8c89
NC
107# pragma optimize("w",on )
108# endif /* BUGGY_MSC6 */
fe14fcc3
LW
109#endif /* MSDOS */
110
a687059c
LW
111#ifndef STATIC
112#define STATIC static
113#endif
114
b1603ef8 115
830247a4 116typedef struct RExC_state_t {
514a91f1
DM
117 U32 flags; /* RXf_* are we folding, multilining? */
118 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 119 char *precomp; /* uncompiled string. */
288b8c02 120 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf
YO
121 regexp *rx; /* perl core regexp structure */
122 regexp_internal *rxi; /* internal data for regexp object pprivate field */
fac92740 123 char *start; /* Start of input for compile */
830247a4
IZ
124 char *end; /* End of input for compile */
125 char *parse; /* Input-scan pointer. */
126 I32 whilem_seen; /* number of WHILEM in this expr */
fac92740 127 regnode *emit_start; /* Start of emitted-code area */
3b57cd43 128 regnode *emit_bound; /* First regnode outside of the allocated space */
ffc61ed2 129 regnode *emit; /* Code-emit pointer; &regdummy = don't = compiling */
830247a4
IZ
130 I32 naughty; /* How bad is this pattern? */
131 I32 sawback; /* Did we see \1, ...? */
132 U32 seen;
133 I32 size; /* Code size. */
c74340f9
YO
134 I32 npar; /* Capture buffer count, (OPEN). */
135 I32 cpar; /* Capture buffer count, (CLOSE). */
e2e6a0f1 136 I32 nestroot; /* root parens we are in - used by accept */
830247a4
IZ
137 I32 extralen;
138 I32 seen_zerolen;
40d049e4
YO
139 regnode **open_parens; /* pointers to open parens */
140 regnode **close_parens; /* pointers to close parens */
141 regnode *opend; /* END node in program */
02daf0ab
YO
142 I32 utf8; /* whether the pattern is utf8 or not */
143 I32 orig_utf8; /* whether the pattern was originally in utf8 */
144 /* XXX use this for future optimisation of case
145 * where pattern must be upgraded to utf8. */
e40e74fe
KW
146 I32 uni_semantics; /* If a d charset modifier should use unicode
147 rules, even if the pattern is not in
148 utf8 */
81714fb9 149 HV *paren_names; /* Paren names */
1f1031fe 150
40d049e4
YO
151 regnode **recurse; /* Recurse regops */
152 I32 recurse_count; /* Number of recurse regops */
b57e4118 153 I32 in_lookbehind;
4624b182 154 I32 contains_locale;
bb3f3ed2 155 I32 override_recoding;
9d53c457 156 I32 in_multi_char_class;
3d2bd50a 157 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 158 within pattern */
b1603ef8
DM
159 int num_code_blocks; /* size of code_blocks[] */
160 int code_index; /* next code_blocks[] slot */
830247a4
IZ
161#if ADD_TO_REGEXEC
162 char *starttry; /* -Dr: where regtry was called. */
163#define RExC_starttry (pRExC_state->starttry)
164#endif
d24ca0c5 165 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 166#ifdef DEBUGGING
be8e71aa 167 const char *lastparse;
3dab1dad 168 I32 lastnum;
1f1031fe 169 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
170#define RExC_lastparse (pRExC_state->lastparse)
171#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 172#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 173#endif
830247a4
IZ
174} RExC_state_t;
175
e2509266 176#define RExC_flags (pRExC_state->flags)
514a91f1 177#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 178#define RExC_precomp (pRExC_state->precomp)
288b8c02 179#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 180#define RExC_rx (pRExC_state->rx)
f8fc2ecf 181#define RExC_rxi (pRExC_state->rxi)
fac92740 182#define RExC_start (pRExC_state->start)
830247a4
IZ
183#define RExC_end (pRExC_state->end)
184#define RExC_parse (pRExC_state->parse)
185#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237
YO
186#ifdef RE_TRACK_PATTERN_OFFSETS
187#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
188#endif
830247a4 189#define RExC_emit (pRExC_state->emit)
fac92740 190#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 191#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
192#define RExC_naughty (pRExC_state->naughty)
193#define RExC_sawback (pRExC_state->sawback)
194#define RExC_seen (pRExC_state->seen)
195#define RExC_size (pRExC_state->size)
196#define RExC_npar (pRExC_state->npar)
e2e6a0f1 197#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
198#define RExC_extralen (pRExC_state->extralen)
199#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 200#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 201#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 202#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
203#define RExC_open_parens (pRExC_state->open_parens)
204#define RExC_close_parens (pRExC_state->close_parens)
205#define RExC_opend (pRExC_state->opend)
81714fb9 206#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
207#define RExC_recurse (pRExC_state->recurse)
208#define RExC_recurse_count (pRExC_state->recurse_count)
b57e4118 209#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 210#define RExC_contains_locale (pRExC_state->contains_locale)
9d53c457
KW
211#define RExC_override_recoding (pRExC_state->override_recoding)
212#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
830247a4 213
cde0cee5 214
a687059c
LW
215#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
216#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
217 ((*s) == '{' && regcurly(s)))
a687059c 218
35c8bce7
LW
219#ifdef SPSTART
220#undef SPSTART /* dratted cpp namespace... */
221#endif
a687059c
LW
222/*
223 * Flags to be passed up and down.
224 */
a687059c 225#define WORST 0 /* Worst case. */
a3b492c3 226#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 227
e64f369d 228/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
229 * character. (There needs to be a case: in the switch statement in regexec.c
230 * for any node marked SIMPLE.) Note that this is not the same thing as
231 * REGNODE_SIMPLE */
fda99bee 232#define SIMPLE 0x02
e64f369d 233#define SPSTART 0x04 /* Starts with * or + */
a3b492c3
YO
234#define TRYAGAIN 0x08 /* Weeded out a declaration. */
235#define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
a687059c 236
3dab1dad
YO
237#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
238
07be1b83
YO
239/* whether trie related optimizations are enabled */
240#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
241#define TRIE_STUDY_OPT
786e8c11 242#define FULL_TRIE_STUDY
07be1b83
YO
243#define TRIE_STCLASS
244#endif
1de06328
YO
245
246
40d049e4
YO
247
248#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
249#define PBITVAL(paren) (1 << ((paren) & 7))
250#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
251#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
252#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
253
bbd61b5f
KW
254/* If not already in utf8, do a longjmp back to the beginning */
255#define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
256#define REQUIRE_UTF8 STMT_START { \
257 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
258 } STMT_END
40d049e4 259
f19b1a63
KW
260/* This converts the named class defined in regcomp.h to its equivalent class
261 * number defined in handy.h. */
262#define namedclass_to_classnum(class) ((int) ((class) / 2))
263#define classnum_to_namedclass(classnum) ((classnum) * 2)
264
1de06328
YO
265/* About scan_data_t.
266
267 During optimisation we recurse through the regexp program performing
268 various inplace (keyhole style) optimisations. In addition study_chunk
269 and scan_commit populate this data structure with information about
270 what strings MUST appear in the pattern. We look for the longest
3b753521 271 string that must appear at a fixed location, and we look for the
1de06328
YO
272 longest string that may appear at a floating location. So for instance
273 in the pattern:
274
275 /FOO[xX]A.*B[xX]BAR/
276
277 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
278 strings (because they follow a .* construct). study_chunk will identify
279 both FOO and BAR as being the longest fixed and floating strings respectively.
280
281 The strings can be composites, for instance
282
283 /(f)(o)(o)/
284
285 will result in a composite fixed substring 'foo'.
286
287 For each string some basic information is maintained:
288
289 - offset or min_offset
290 This is the position the string must appear at, or not before.
291 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
292 characters must match before the string we are searching for.
293 Likewise when combined with minlenp and the length of the string it
1de06328
YO
294 tells us how many characters must appear after the string we have
295 found.
296
297 - max_offset
298 Only used for floating strings. This is the rightmost point that
3b753521 299 the string can appear at. If set to I32 max it indicates that the
1de06328
YO
300 string can occur infinitely far to the right.
301
302 - minlenp
2d608413
KW
303 A pointer to the minimum number of characters of the pattern that the
304 string was found inside. This is important as in the case of positive
1de06328
YO
305 lookahead or positive lookbehind we can have multiple patterns
306 involved. Consider
307
308 /(?=FOO).*F/
309
310 The minimum length of the pattern overall is 3, the minimum length
311 of the lookahead part is 3, but the minimum length of the part that
312 will actually match is 1. So 'FOO's minimum length is 3, but the
313 minimum length for the F is 1. This is important as the minimum length
314 is used to determine offsets in front of and behind the string being
315 looked for. Since strings can be composites this is the length of the
486ec47a 316 pattern at the time it was committed with a scan_commit. Note that
1de06328
YO
317 the length is calculated by study_chunk, so that the minimum lengths
318 are not known until the full pattern has been compiled, thus the
319 pointer to the value.
320
321 - lookbehind
322
323 In the case of lookbehind the string being searched for can be
324 offset past the start point of the final matching string.
325 If this value was just blithely removed from the min_offset it would
326 invalidate some of the calculations for how many chars must match
327 before or after (as they are derived from min_offset and minlen and
328 the length of the string being searched for).
329 When the final pattern is compiled and the data is moved from the
330 scan_data_t structure into the regexp structure the information
331 about lookbehind is factored in, with the information that would
332 have been lost precalculated in the end_shift field for the
333 associated string.
334
335 The fields pos_min and pos_delta are used to store the minimum offset
336 and the delta to the maximum offset at the current point in the pattern.
337
338*/
2c2d71f5
JH
339
340typedef struct scan_data_t {
1de06328
YO
341 /*I32 len_min; unused */
342 /*I32 len_delta; unused */
2c2d71f5
JH
343 I32 pos_min;
344 I32 pos_delta;
345 SV *last_found;
1de06328 346 I32 last_end; /* min value, <0 unless valid. */
2c2d71f5
JH
347 I32 last_start_min;
348 I32 last_start_max;
1de06328
YO
349 SV **longest; /* Either &l_fixed, or &l_float. */
350 SV *longest_fixed; /* longest fixed string found in pattern */
351 I32 offset_fixed; /* offset where it starts */
486ec47a 352 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
353 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
354 SV *longest_float; /* longest floating string found in pattern */
355 I32 offset_float_min; /* earliest point in string it can appear */
356 I32 offset_float_max; /* latest point in string it can appear */
486ec47a 357 I32 *minlen_float; /* pointer to the minlen relevant to the string */
1de06328 358 I32 lookbehind_float; /* is the position of the string modified by LB */
2c2d71f5
JH
359 I32 flags;
360 I32 whilem_c;
cb434fcc 361 I32 *last_closep;
653099ff 362 struct regnode_charclass_class *start_class;
2c2d71f5
JH
363} scan_data_t;
364
a687059c 365/*
e50aee73 366 * Forward declarations for pregcomp()'s friends.
a687059c 367 */
a0d0e21e 368
27da23d5 369static const scan_data_t zero_scan_data =
1de06328 370 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
371
372#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
373#define SF_BEFORE_SEOL 0x0001
374#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
375#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
376#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
377
09b7f37c
CB
378#ifdef NO_UNARY_PLUS
379# define SF_FIX_SHIFT_EOL (0+2)
380# define SF_FL_SHIFT_EOL (0+4)
381#else
382# define SF_FIX_SHIFT_EOL (+2)
383# define SF_FL_SHIFT_EOL (+4)
384#endif
c277df42
IZ
385
386#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
387#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
388
389#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
390#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
391#define SF_IS_INF 0x0040
392#define SF_HAS_PAR 0x0080
393#define SF_IN_PAR 0x0100
394#define SF_HAS_EVAL 0x0200
395#define SCF_DO_SUBSTR 0x0400
653099ff
GS
396#define SCF_DO_STCLASS_AND 0x0800
397#define SCF_DO_STCLASS_OR 0x1000
398#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 399#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 400
786e8c11 401#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
e2e6a0f1 402#define SCF_SEEN_ACCEPT 0x8000
07be1b83 403
43fead97 404#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
405
406/* The enums for all these are ordered so things work out correctly */
a62b1201 407#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
cfaf538b 408#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
00b27cfc 409#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
410#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
411#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1 412#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a725e29c 413#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 414
43fead97 415#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 416
93733859 417#define OOB_NAMEDCLASS -1
b8c5462f 418
8e661ac5
KW
419/* There is no code point that is out-of-bounds, so this is problematic. But
420 * its only current use is to initialize a variable that is always set before
421 * looked at. */
422#define OOB_UNICODE 0xDEADBEEF
423
a0ed51b3
LW
424#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
425#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
426
8615cb43 427
b45f050a
JF
428/* length of regex to show in messages that don't mark a position within */
429#define RegexLengthToShowInErrorMessages 127
430
431/*
432 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
433 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
434 * op/pragma/warn/regcomp.
435 */
7253e4e3
RK
436#define MARKER1 "<-- HERE" /* marker as it appears in the description */
437#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 438
7253e4e3 439#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
440
441/*
442 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
443 * arg. Show regex, up to a maximum length. If it's too long, chop and add
444 * "...".
445 */
58e23c8d 446#define _FAIL(code) STMT_START { \
bfed75c6 447 const char *ellipses = ""; \
ccb2c380
MP
448 IV len = RExC_end - RExC_precomp; \
449 \
450 if (!SIZE_ONLY) \
a5e7bc51 451 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
452 if (len > RegexLengthToShowInErrorMessages) { \
453 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
454 len = RegexLengthToShowInErrorMessages - 10; \
455 ellipses = "..."; \
456 } \
58e23c8d 457 code; \
ccb2c380 458} STMT_END
8615cb43 459
58e23c8d
YO
460#define FAIL(msg) _FAIL( \
461 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
462 msg, (int)len, RExC_precomp, ellipses))
463
464#define FAIL2(msg,arg) _FAIL( \
465 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
466 arg, (int)len, RExC_precomp, ellipses))
467
b45f050a 468/*
b45f050a
JF
469 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
470 */
ccb2c380 471#define Simple_vFAIL(m) STMT_START { \
a28509cc 472 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
473 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
474 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
475} STMT_END
b45f050a
JF
476
477/*
478 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
479 */
ccb2c380
MP
480#define vFAIL(m) STMT_START { \
481 if (!SIZE_ONLY) \
a5e7bc51 482 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
483 Simple_vFAIL(m); \
484} STMT_END
b45f050a
JF
485
486/*
487 * Like Simple_vFAIL(), but accepts two arguments.
488 */
ccb2c380 489#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 490 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
491 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
492 (int)offset, RExC_precomp, RExC_precomp + offset); \
493} STMT_END
b45f050a
JF
494
495/*
496 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
497 */
ccb2c380
MP
498#define vFAIL2(m,a1) STMT_START { \
499 if (!SIZE_ONLY) \
a5e7bc51 500 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
501 Simple_vFAIL2(m, a1); \
502} STMT_END
b45f050a
JF
503
504
505/*
506 * Like Simple_vFAIL(), but accepts three arguments.
507 */
ccb2c380 508#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 509 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
510 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
511 (int)offset, RExC_precomp, RExC_precomp + offset); \
512} STMT_END
b45f050a
JF
513
514/*
515 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
516 */
ccb2c380
MP
517#define vFAIL3(m,a1,a2) STMT_START { \
518 if (!SIZE_ONLY) \
a5e7bc51 519 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
520 Simple_vFAIL3(m, a1, a2); \
521} STMT_END
b45f050a
JF
522
523/*
524 * Like Simple_vFAIL(), but accepts four arguments.
525 */
ccb2c380 526#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 527 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
528 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
529 (int)offset, RExC_precomp, RExC_precomp + offset); \
530} STMT_END
b45f050a 531
668c081a 532#define ckWARNreg(loc,m) STMT_START { \
a28509cc 533 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
534 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
535 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
536} STMT_END
537
668c081a 538#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 539 const IV offset = loc - RExC_precomp; \
d1d15184 540 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
541 m REPORT_LOCATION, \
542 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
543} STMT_END
544
2335b3d3
KW
545#define ckWARN2regdep(loc,m, a1) STMT_START { \
546 const IV offset = loc - RExC_precomp; \
547 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
548 m REPORT_LOCATION, \
549 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
550} STMT_END
551
668c081a 552#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 553 const IV offset = loc - RExC_precomp; \
668c081a 554 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
555 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
556} STMT_END
557
558#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 559 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
560 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
561 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
562} STMT_END
563
668c081a
NC
564#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
565 const IV offset = loc - RExC_precomp; \
566 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
567 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
568} STMT_END
569
ccb2c380 570#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 571 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
572 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
573 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
574} STMT_END
575
668c081a
NC
576#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
577 const IV offset = loc - RExC_precomp; \
578 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
579 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
580} STMT_END
581
ccb2c380 582#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 583 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
584 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
585 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
586} STMT_END
9d1d55b5 587
8615cb43 588
cd439c50 589/* Allow for side effects in s */
ccb2c380
MP
590#define REGC(c,s) STMT_START { \
591 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
592} STMT_END
cd439c50 593
fac92740
MJD
594/* Macros for recording node offsets. 20001227 mjd@plover.com
595 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
596 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
597 * Element 0 holds the number n.
07be1b83 598 * Position is 1 indexed.
fac92740 599 */
7122b237
YO
600#ifndef RE_TRACK_PATTERN_OFFSETS
601#define Set_Node_Offset_To_R(node,byte)
602#define Set_Node_Offset(node,byte)
603#define Set_Cur_Node_Offset
604#define Set_Node_Length_To_R(node,len)
605#define Set_Node_Length(node,len)
606#define Set_Node_Cur_Length(node)
607#define Node_Offset(n)
608#define Node_Length(n)
609#define Set_Node_Offset_Length(node,offset,len)
610#define ProgLen(ri) ri->u.proglen
611#define SetProgLen(ri,x) ri->u.proglen = x
612#else
613#define ProgLen(ri) ri->u.offsets[0]
614#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
615#define Set_Node_Offset_To_R(node,byte) STMT_START { \
616 if (! SIZE_ONLY) { \
617 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 618 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 619 if((node) < 0) { \
551405c4 620 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
621 } else { \
622 RExC_offsets[2*(node)-1] = (byte); \
623 } \
624 } \
625} STMT_END
626
627#define Set_Node_Offset(node,byte) \
628 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
629#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
630
631#define Set_Node_Length_To_R(node,len) STMT_START { \
632 if (! SIZE_ONLY) { \
633 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 634 __LINE__, (int)(node), (int)(len))); \
ccb2c380 635 if((node) < 0) { \
551405c4 636 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
637 } else { \
638 RExC_offsets[2*(node)] = (len); \
639 } \
640 } \
641} STMT_END
642
643#define Set_Node_Length(node,len) \
644 Set_Node_Length_To_R((node)-RExC_emit_start, len)
645#define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
646#define Set_Node_Cur_Length(node) \
647 Set_Node_Length(node, RExC_parse - parse_start)
fac92740
MJD
648
649/* Get offsets and lengths */
650#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
651#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
652
07be1b83
YO
653#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
654 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
655 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
656} STMT_END
7122b237 657#endif
07be1b83
YO
658
659#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
660#define EXPERIMENTAL_INPLACESCAN
f427392e 661#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 662
304ee84b
YO
663#define DEBUG_STUDYDATA(str,data,depth) \
664DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 665 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
666 "%*s" str "Pos:%"IVdf"/%"IVdf \
667 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
668 (int)(depth)*2, "", \
669 (IV)((data)->pos_min), \
670 (IV)((data)->pos_delta), \
304ee84b 671 (UV)((data)->flags), \
1de06328 672 (IV)((data)->whilem_c), \
304ee84b
YO
673 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
674 is_inf ? "INF " : "" \
1de06328
YO
675 ); \
676 if ((data)->last_found) \
677 PerlIO_printf(Perl_debug_log, \
678 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
679 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
680 SvPVX_const((data)->last_found), \
681 (IV)((data)->last_end), \
682 (IV)((data)->last_start_min), \
683 (IV)((data)->last_start_max), \
684 ((data)->longest && \
685 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
686 SvPVX_const((data)->longest_fixed), \
687 (IV)((data)->offset_fixed), \
688 ((data)->longest && \
689 (data)->longest==&((data)->longest_float)) ? "*" : "", \
690 SvPVX_const((data)->longest_float), \
691 (IV)((data)->offset_float_min), \
692 (IV)((data)->offset_float_max) \
693 ); \
694 PerlIO_printf(Perl_debug_log,"\n"); \
695});
696
653099ff 697/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 698 Update the longest found anchored substring and the longest found
653099ff
GS
699 floating substrings if needed. */
700
4327152a 701STATIC void
304ee84b 702S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
c277df42 703{
e1ec3a88
AL
704 const STRLEN l = CHR_SVLEN(data->last_found);
705 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 706 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 707
7918f24d
NC
708 PERL_ARGS_ASSERT_SCAN_COMMIT;
709
c277df42 710 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 711 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
712 if (*data->longest == data->longest_fixed) {
713 data->offset_fixed = l ? data->last_start_min : data->pos_min;
714 if (data->flags & SF_BEFORE_EOL)
b81d288d 715 data->flags
c277df42
IZ
716 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
717 else
718 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 719 data->minlen_fixed=minlenp;
1de06328 720 data->lookbehind_fixed=0;
a0ed51b3 721 }
304ee84b 722 else { /* *data->longest == data->longest_float */
c277df42 723 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
724 data->offset_float_max = (l
725 ? data->last_start_max
c277df42 726 : data->pos_min + data->pos_delta);
304ee84b 727 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
9051bda5 728 data->offset_float_max = I32_MAX;
c277df42 729 if (data->flags & SF_BEFORE_EOL)
b81d288d 730 data->flags
c277df42
IZ
731 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
732 else
733 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
734 data->minlen_float=minlenp;
735 data->lookbehind_float=0;
c277df42
IZ
736 }
737 }
738 SvCUR_set(data->last_found, 0);
0eda9292 739 {
a28509cc 740 SV * const sv = data->last_found;
097eb12c
AL
741 if (SvUTF8(sv) && SvMAGICAL(sv)) {
742 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
743 if (mg)
744 mg->mg_len = 0;
745 }
0eda9292 746 }
c277df42
IZ
747 data->last_end = -1;
748 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 749 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
750}
751
653099ff
GS
752/* Can match anything (initialization) */
753STATIC void
3fffb88a 754S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 755{
7918f24d
NC
756 PERL_ARGS_ASSERT_CL_ANYTHING;
757
f8bef550 758 ANYOF_BITMAP_SETALL(cl);
c81eddd4 759 cl->flags = ANYOF_EOS|ANYOF_UNICODE_ALL;
3fffb88a
KW
760
761 /* If any portion of the regex is to operate under locale rules,
762 * initialization includes it. The reason this isn't done for all regexes
763 * is that the optimizer was written under the assumption that locale was
764 * all-or-nothing. Given the complexity and lack of documentation in the
765 * optimizer, and that there are inadequate test cases for locale, so many
766 * parts of it may not work properly, it is safest to avoid locale unless
767 * necessary. */
768 if (RExC_contains_locale) {
9d7a1e63 769 ANYOF_CLASS_SETALL(cl); /* /l uses class */
b25a1036 770 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
3fffb88a 771 }
9d7a1e63
KW
772 else {
773 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
774 }
653099ff
GS
775}
776
777/* Can match anything (initialization) */
778STATIC int
5f66b61c 779S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
780{
781 int value;
782
7918f24d
NC
783 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
784
aaa51d5e 785 for (value = 0; value <= ANYOF_MAX; value += 2)
653099ff
GS
786 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
787 return 1;
1aa99e6b
IH
788 if (!(cl->flags & ANYOF_UNICODE_ALL))
789 return 0;
10edeb5d 790 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 791 return 0;
653099ff
GS
792 return 1;
793}
794
795/* Can match anything (initialization) */
796STATIC void
e755fd73 797S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 798{
7918f24d
NC
799 PERL_ARGS_ASSERT_CL_INIT;
800
8ecf7187 801 Zero(cl, 1, struct regnode_charclass_class);
653099ff 802 cl->type = ANYOF;
3fffb88a 803 cl_anything(pRExC_state, cl);
1411dba4 804 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
805}
806
1051e1c4
KW
807/* These two functions currently do the exact same thing */
808#define cl_init_zero S_cl_init
653099ff 809
dd58aee1
KW
810/* 'AND' a given class with another one. Can create false positives. 'cl'
811 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
812 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 813STATIC void
5f66b61c 814S_cl_and(struct regnode_charclass_class *cl,
a28509cc 815 const struct regnode_charclass_class *and_with)
653099ff 816{
7918f24d 817 PERL_ARGS_ASSERT_CL_AND;
40d049e4
YO
818
819 assert(and_with->type == ANYOF);
1e6ade67 820
c6b76537 821 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
822 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
823 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 824 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
825 && !(and_with->flags & ANYOF_LOC_FOLD)
826 && !(cl->flags & ANYOF_LOC_FOLD)) {
653099ff
GS
827 int i;
828
829 if (and_with->flags & ANYOF_INVERT)
830 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
831 cl->bitmap[i] &= ~and_with->bitmap[i];
832 else
833 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
834 cl->bitmap[i] &= and_with->bitmap[i];
835 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 836
c6b76537 837 if (and_with->flags & ANYOF_INVERT) {
8951c461 838
c6b76537
KW
839 /* Here, the and'ed node is inverted. Get the AND of the flags that
840 * aren't affected by the inversion. Those that are affected are
841 * handled individually below */
842 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
843 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
844 cl->flags |= affected_flags;
845
846 /* We currently don't know how to deal with things that aren't in the
847 * bitmap, but we know that the intersection is no greater than what
848 * is already in cl, so let there be false positives that get sorted
849 * out after the synthetic start class succeeds, and the node is
850 * matched for real. */
851
852 /* The inversion of these two flags indicate that the resulting
853 * intersection doesn't have them */
854 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
855 cl->flags &= ~ANYOF_UNICODE_ALL;
856 }
c6b76537
KW
857 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
858 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 859 }
1aa99e6b 860 }
c6b76537 861 else { /* and'd node is not inverted */
3ad98780
KW
862 U8 outside_bitmap_but_not_utf8; /* Temp variable */
863
137165a6 864 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
865
866 /* Here 'and_with' doesn't match anything outside the bitmap
867 * (except possibly ANYOF_UNICODE_ALL), which means the
868 * intersection can't either, except for ANYOF_UNICODE_ALL, in
869 * which case we don't know what the intersection is, but it's no
870 * greater than what cl already has, so can just leave it alone,
871 * with possible false positives */
872 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
873 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 874 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 875 }
137165a6 876 }
c6b76537
KW
877 else if (! ANYOF_NONBITMAP(cl)) {
878
879 /* Here, 'and_with' does match something outside the bitmap, and cl
880 * doesn't have a list of things to match outside the bitmap. If
881 * cl can match all code points above 255, the intersection will
3ad98780
KW
882 * be those above-255 code points that 'and_with' matches. If cl
883 * can't match all Unicode code points, it means that it can't
884 * match anything outside the bitmap (since the 'if' that got us
885 * into this block tested for that), so we leave the bitmap empty.
886 */
c6b76537
KW
887 if (cl->flags & ANYOF_UNICODE_ALL) {
888 ARG_SET(cl, ARG(and_with));
3ad98780
KW
889
890 /* and_with's ARG may match things that don't require UTF8.
891 * And now cl's will too, in spite of this being an 'and'. See
892 * the comments below about the kludge */
893 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
894 }
895 }
896 else {
897 /* Here, both 'and_with' and cl match something outside the
898 * bitmap. Currently we do not do the intersection, so just match
899 * whatever cl had at the beginning. */
900 }
901
902
3ad98780
KW
903 /* Take the intersection of the two sets of flags. However, the
904 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
905 * kludge around the fact that this flag is not treated like the others
906 * which are initialized in cl_anything(). The way the optimizer works
907 * is that the synthetic start class (SSC) is initialized to match
908 * anything, and then the first time a real node is encountered, its
909 * values are AND'd with the SSC's with the result being the values of
910 * the real node. However, there are paths through the optimizer where
911 * the AND never gets called, so those initialized bits are set
912 * inappropriately, which is not usually a big deal, as they just cause
913 * false positives in the SSC, which will just mean a probably
914 * imperceptible slow down in execution. However this bit has a
915 * higher false positive consequence in that it can cause utf8.pm,
916 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
917 * bigger slowdown and also causes significant extra memory to be used.
918 * In order to prevent this, the code now takes a different tack. The
919 * bit isn't set unless some part of the regular expression needs it,
920 * but once set it won't get cleared. This means that these extra
921 * modules won't get loaded unless there was some path through the
922 * pattern that would have required them anyway, and so any false
923 * positives that occur by not ANDing them out when they could be
924 * aren't as severe as they would be if we treated this bit like all
925 * the others */
926 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
927 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 928 cl->flags &= and_with->flags;
3ad98780 929 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 930 }
653099ff
GS
931}
932
dd58aee1
KW
933/* 'OR' a given class with another one. Can create false positives. 'cl'
934 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
935 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 936STATIC void
3fffb88a 937S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 938{
7918f24d
NC
939 PERL_ARGS_ASSERT_CL_OR;
940
653099ff 941 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
942
943 /* Here, the or'd node is to be inverted. This means we take the
944 * complement of everything not in the bitmap, but currently we don't
945 * know what that is, so give up and match anything */
946 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 947 cl_anything(pRExC_state, cl);
c6b76537 948 }
653099ff
GS
949 /* We do not use
950 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
951 * <= (B1 | !B2) | (CL1 | !CL2)
952 * which is wasteful if CL2 is small, but we ignore CL2:
953 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
954 * XXXX Can we handle case-fold? Unclear:
955 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
956 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
957 */
c6b76537 958 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
959 && !(or_with->flags & ANYOF_LOC_FOLD)
960 && !(cl->flags & ANYOF_LOC_FOLD) ) {
653099ff
GS
961 int i;
962
963 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
964 cl->bitmap[i] |= ~or_with->bitmap[i];
965 } /* XXXX: logic is complicated otherwise */
966 else {
3fffb88a 967 cl_anything(pRExC_state, cl);
653099ff 968 }
c6b76537
KW
969
970 /* And, we can just take the union of the flags that aren't affected
971 * by the inversion */
972 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
973
974 /* For the remaining flags:
975 ANYOF_UNICODE_ALL and inverted means to not match anything above
976 255, which means that the union with cl should just be
977 what cl has in it, so can ignore this flag
978 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
979 is 127-255 to match them, but then invert that, so the
980 union with cl should just be what cl has in it, so can
981 ignore this flag
982 */
983 } else { /* 'or_with' is not inverted */
653099ff
GS
984 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
985 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
986 && (!(or_with->flags & ANYOF_LOC_FOLD)
987 || (cl->flags & ANYOF_LOC_FOLD)) ) {
653099ff
GS
988 int i;
989
990 /* OR char bitmap and class bitmap separately */
991 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
992 cl->bitmap[i] |= or_with->bitmap[i];
9cf78fa0 993 ANYOF_CLASS_OR(or_with, cl);
653099ff
GS
994 }
995 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 996 cl_anything(pRExC_state, cl);
653099ff 997 }
9826f543 998
c6b76537
KW
999 if (ANYOF_NONBITMAP(or_with)) {
1000
1001 /* Use the added node's outside-the-bit-map match if there isn't a
1002 * conflict. If there is a conflict (both nodes match something
1003 * outside the bitmap, but what they match outside is not the same
1004 * pointer, and hence not easily compared until XXX we extend
1005 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
1006 * match everything outside the bitmap. If that stuff is all above
1007 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
1008 if (! ANYOF_NONBITMAP(cl)) {
1009 ARG_SET(cl, ARG(or_with));
1010 }
1011 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
1012
1013 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1014 cl_anything(pRExC_state, cl);
1015 }
1016 else {
1017 cl->flags |= ANYOF_UNICODE_ALL;
1018 }
c6b76537 1019 }
4c34a693 1020 }
0b9668ee
KW
1021
1022 /* Take the union */
1023 cl->flags |= or_with->flags;
1aa99e6b 1024 }
653099ff
GS
1025}
1026
a3621e74
YO
1027#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1028#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1029#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1030#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1031
3dab1dad
YO
1032
1033#ifdef DEBUGGING
07be1b83 1034/*
2b8b4781
NC
1035 dump_trie(trie,widecharmap,revcharmap)
1036 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1037 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1038
1039 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1040 The _interim_ variants are used for debugging the interim
1041 tables that are used to generate the final compressed
1042 representation which is what dump_trie expects.
1043
486ec47a 1044 Part of the reason for their existence is to provide a form
3dab1dad 1045 of documentation as to how the different representations function.
07be1b83
YO
1046
1047*/
3dab1dad
YO
1048
1049/*
3dab1dad
YO
1050 Dumps the final compressed table form of the trie to Perl_debug_log.
1051 Used for debugging make_trie().
1052*/
b9a59e08 1053
3dab1dad 1054STATIC void
2b8b4781
NC
1055S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1056 AV *revcharmap, U32 depth)
3dab1dad
YO
1057{
1058 U32 state;
ab3bbdeb 1059 SV *sv=sv_newmortal();
55eed653 1060 int colwidth= widecharmap ? 6 : 4;
2e64971a 1061 U16 word;
3dab1dad
YO
1062 GET_RE_DEBUG_FLAGS_DECL;
1063
7918f24d 1064 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1065
3dab1dad
YO
1066 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1067 (int)depth * 2 + 2,"",
1068 "Match","Base","Ofs" );
1069
1070 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1071 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1072 if ( tmp ) {
ab3bbdeb
YO
1073 PerlIO_printf( Perl_debug_log, "%*s",
1074 colwidth,
ddc5bc0f 1075 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1076 PL_colors[0], PL_colors[1],
1077 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1078 PERL_PV_ESCAPE_FIRSTCHAR
1079 )
1080 );
3dab1dad
YO
1081 }
1082 }
1083 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1084 (int)depth * 2 + 2,"");
1085
1086 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1087 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1088 PerlIO_printf( Perl_debug_log, "\n");
1089
1e2e3d02 1090 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1091 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1092
1093 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1094
1095 if ( trie->states[ state ].wordnum ) {
1096 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1097 } else {
1098 PerlIO_printf( Perl_debug_log, "%6s", "" );
1099 }
1100
1101 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1102
1103 if ( base ) {
1104 U32 ofs = 0;
1105
1106 while( ( base + ofs < trie->uniquecharcount ) ||
1107 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1108 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1109 ofs++;
1110
1111 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1112
1113 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1114 if ( ( base + ofs >= trie->uniquecharcount ) &&
1115 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1116 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1117 {
ab3bbdeb
YO
1118 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1119 colwidth,
3dab1dad
YO
1120 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1121 } else {
ab3bbdeb 1122 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1123 }
1124 }
1125
1126 PerlIO_printf( Perl_debug_log, "]");
1127
1128 }
1129 PerlIO_printf( Perl_debug_log, "\n" );
1130 }
2e64971a
DM
1131 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1132 for (word=1; word <= trie->wordcount; word++) {
1133 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1134 (int)word, (int)(trie->wordinfo[word].prev),
1135 (int)(trie->wordinfo[word].len));
1136 }
1137 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1138}
1139/*
3dab1dad
YO
1140 Dumps a fully constructed but uncompressed trie in list form.
1141 List tries normally only are used for construction when the number of
1142 possible chars (trie->uniquecharcount) is very high.
1143 Used for debugging make_trie().
1144*/
1145STATIC void
55eed653 1146S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1147 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1148 U32 depth)
3dab1dad
YO
1149{
1150 U32 state;
ab3bbdeb 1151 SV *sv=sv_newmortal();
55eed653 1152 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1153 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1154
1155 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1156
3dab1dad 1157 /* print out the table precompression. */
ab3bbdeb
YO
1158 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1159 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1160 "------:-----+-----------------\n" );
3dab1dad
YO
1161
1162 for( state=1 ; state < next_alloc ; state ++ ) {
1163 U16 charid;
1164
ab3bbdeb 1165 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1166 (int)depth * 2 + 2,"", (UV)state );
1167 if ( ! trie->states[ state ].wordnum ) {
1168 PerlIO_printf( Perl_debug_log, "%5s| ","");
1169 } else {
1170 PerlIO_printf( Perl_debug_log, "W%4x| ",
1171 trie->states[ state ].wordnum
1172 );
1173 }
1174 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1175 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1176 if ( tmp ) {
1177 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1178 colwidth,
ddc5bc0f 1179 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1180 PL_colors[0], PL_colors[1],
1181 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1182 PERL_PV_ESCAPE_FIRSTCHAR
1183 ) ,
1e2e3d02
YO
1184 TRIE_LIST_ITEM(state,charid).forid,
1185 (UV)TRIE_LIST_ITEM(state,charid).newstate
1186 );
1187 if (!(charid % 10))
664e119d
RGS
1188 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1189 (int)((depth * 2) + 14), "");
1e2e3d02 1190 }
ab3bbdeb
YO
1191 }
1192 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1193 }
1194}
1195
1196/*
3dab1dad
YO
1197 Dumps a fully constructed but uncompressed trie in table form.
1198 This is the normal DFA style state transition table, with a few
1199 twists to facilitate compression later.
1200 Used for debugging make_trie().
1201*/
1202STATIC void
55eed653 1203S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1204 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1205 U32 depth)
3dab1dad
YO
1206{
1207 U32 state;
1208 U16 charid;
ab3bbdeb 1209 SV *sv=sv_newmortal();
55eed653 1210 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1211 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1212
1213 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1214
1215 /*
1216 print out the table precompression so that we can do a visual check
1217 that they are identical.
1218 */
1219
1220 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1221
1222 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1223 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1224 if ( tmp ) {
ab3bbdeb
YO
1225 PerlIO_printf( Perl_debug_log, "%*s",
1226 colwidth,
ddc5bc0f 1227 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1228 PL_colors[0], PL_colors[1],
1229 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1230 PERL_PV_ESCAPE_FIRSTCHAR
1231 )
1232 );
3dab1dad
YO
1233 }
1234 }
1235
1236 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1237
1238 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1239 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1240 }
1241
1242 PerlIO_printf( Perl_debug_log, "\n" );
1243
1244 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1245
1246 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1247 (int)depth * 2 + 2,"",
1248 (UV)TRIE_NODENUM( state ) );
1249
1250 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1251 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1252 if (v)
1253 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1254 else
1255 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1256 }
1257 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1258 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1259 } else {
1260 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1261 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1262 }
1263 }
07be1b83 1264}
3dab1dad
YO
1265
1266#endif
1267
2e64971a 1268
786e8c11
YO
1269/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1270 startbranch: the first branch in the whole branch sequence
1271 first : start branch of sequence of branch-exact nodes.
1272 May be the same as startbranch
1273 last : Thing following the last branch.
1274 May be the same as tail.
1275 tail : item following the branch sequence
1276 count : words in the sequence
1277 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1278 depth : indent depth
3dab1dad 1279
786e8c11 1280Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1281
786e8c11
YO
1282A trie is an N'ary tree where the branches are determined by digital
1283decomposition of the key. IE, at the root node you look up the 1st character and
1284follow that branch repeat until you find the end of the branches. Nodes can be
1285marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1286
786e8c11 1287 /he|she|his|hers/
72f13be8 1288
786e8c11
YO
1289would convert into the following structure. Numbers represent states, letters
1290following numbers represent valid transitions on the letter from that state, if
1291the number is in square brackets it represents an accepting state, otherwise it
1292will be in parenthesis.
07be1b83 1293
786e8c11
YO
1294 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1295 | |
1296 | (2)
1297 | |
1298 (1) +-i->(6)-+-s->[7]
1299 |
1300 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1301
786e8c11
YO
1302 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1303
1304This shows that when matching against the string 'hers' we will begin at state 1
1305read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1306then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1307is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1308single traverse. We store a mapping from accepting to state to which word was
1309matched, and then when we have multiple possibilities we try to complete the
1310rest of the regex in the order in which they occured in the alternation.
1311
1312The only prior NFA like behaviour that would be changed by the TRIE support is
1313the silent ignoring of duplicate alternations which are of the form:
1314
1315 / (DUPE|DUPE) X? (?{ ... }) Y /x
1316
4b714af6 1317Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1318and without the optimisation. With the optimisations dupes will be silently
486ec47a 1319ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1320the following demonstrates:
1321
1322 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1323
1324which prints out 'word' three times, but
1325
1326 'words'=~/(word|word|word)(?{ print $1 })S/
1327
1328which doesnt print it out at all. This is due to other optimisations kicking in.
1329
1330Example of what happens on a structural level:
1331
486ec47a 1332The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1333
1334 1: CURLYM[1] {1,32767}(18)
1335 5: BRANCH(8)
1336 6: EXACT <ac>(16)
1337 8: BRANCH(11)
1338 9: EXACT <ad>(16)
1339 11: BRANCH(14)
1340 12: EXACT <ab>(16)
1341 16: SUCCEED(0)
1342 17: NOTHING(18)
1343 18: END(0)
1344
1345This would be optimizable with startbranch=5, first=5, last=16, tail=16
1346and should turn into:
1347
1348 1: CURLYM[1] {1,32767}(18)
1349 5: TRIE(16)
1350 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1351 <ac>
1352 <ad>
1353 <ab>
1354 16: SUCCEED(0)
1355 17: NOTHING(18)
1356 18: END(0)
1357
1358Cases where tail != last would be like /(?foo|bar)baz/:
1359
1360 1: BRANCH(4)
1361 2: EXACT <foo>(8)
1362 4: BRANCH(7)
1363 5: EXACT <bar>(8)
1364 7: TAIL(8)
1365 8: EXACT <baz>(10)
1366 10: END(0)
1367
1368which would be optimizable with startbranch=1, first=1, last=7, tail=8
1369and would end up looking like:
1370
1371 1: TRIE(8)
1372 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1373 <foo>
1374 <bar>
1375 7: TAIL(8)
1376 8: EXACT <baz>(10)
1377 10: END(0)
1378
1379 d = uvuni_to_utf8_flags(d, uv, 0);
1380
1381is the recommended Unicode-aware way of saying
1382
1383 *(d++) = uv;
1384*/
1385
fab2782b 1386#define TRIE_STORE_REVCHAR(val) \
786e8c11 1387 STMT_START { \
73031816 1388 if (UTF) { \
fab2782b 1389 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1390 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
fab2782b 1391 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
73031816
NC
1392 SvCUR_set(zlopp, kapow - flrbbbbb); \
1393 SvPOK_on(zlopp); \
1394 SvUTF8_on(zlopp); \
1395 av_push(revcharmap, zlopp); \
1396 } else { \
fab2782b 1397 char ooooff = (char)val; \
73031816
NC
1398 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1399 } \
1400 } STMT_END
786e8c11 1401
fab2782b
YO
1402#define TRIE_READ_CHAR STMT_START { \
1403 wordlen++; \
1404 if ( UTF ) { \
1405 /* if it is UTF then it is either already folded, or does not need folding */ \
1406 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1407 } \
1408 else if (folder == PL_fold_latin1) { \
1409 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1410 if ( foldlen > 0 ) { \
1411 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1412 foldlen -= len; \
1413 scan += len; \
1414 len = 0; \
1415 } else { \
1416 len = 1; \
1417 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1418 skiplen = UNISKIP(uvc); \
1419 foldlen -= skiplen; \
1420 scan = foldbuf + skiplen; \
1421 } \
1422 } else { \
1423 /* raw data, will be folded later if needed */ \
1424 uvc = (U32)*uc; \
1425 len = 1; \
1426 } \
786e8c11
YO
1427} STMT_END
1428
1429
1430
1431#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1432 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1433 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1434 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1435 } \
1436 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1437 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1438 TRIE_LIST_CUR( state )++; \
1439} STMT_END
07be1b83 1440
786e8c11
YO
1441#define TRIE_LIST_NEW(state) STMT_START { \
1442 Newxz( trie->states[ state ].trans.list, \
1443 4, reg_trie_trans_le ); \
1444 TRIE_LIST_CUR( state ) = 1; \
1445 TRIE_LIST_LEN( state ) = 4; \
1446} STMT_END
07be1b83 1447
786e8c11
YO
1448#define TRIE_HANDLE_WORD(state) STMT_START { \
1449 U16 dupe= trie->states[ state ].wordnum; \
1450 regnode * const noper_next = regnext( noper ); \
1451 \
786e8c11
YO
1452 DEBUG_r({ \
1453 /* store the word for dumping */ \
1454 SV* tmp; \
1455 if (OP(noper) != NOTHING) \
740cce10 1456 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1457 else \
740cce10 1458 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1459 av_push( trie_words, tmp ); \
786e8c11
YO
1460 }); \
1461 \
1462 curword++; \
2e64971a
DM
1463 trie->wordinfo[curword].prev = 0; \
1464 trie->wordinfo[curword].len = wordlen; \
1465 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1466 \
1467 if ( noper_next < tail ) { \
1468 if (!trie->jump) \
c944940b 1469 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1470 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1471 if (!jumper) \
1472 jumper = noper_next; \
1473 if (!nextbranch) \
1474 nextbranch= regnext(cur); \
1475 } \
1476 \
1477 if ( dupe ) { \
2e64971a
DM
1478 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1479 /* chain, so that when the bits of chain are later */\
1480 /* linked together, the dups appear in the chain */\
1481 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1482 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1483 } else { \
1484 /* we haven't inserted this word yet. */ \
1485 trie->states[ state ].wordnum = curword; \
1486 } \
1487} STMT_END
07be1b83 1488
3dab1dad 1489
786e8c11
YO
1490#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1491 ( ( base + charid >= ucharcount \
1492 && base + charid < ubound \
1493 && state == trie->trans[ base - ucharcount + charid ].check \
1494 && trie->trans[ base - ucharcount + charid ].next ) \
1495 ? trie->trans[ base - ucharcount + charid ].next \
1496 : ( state==1 ? special : 0 ) \
1497 )
3dab1dad 1498
786e8c11
YO
1499#define MADE_TRIE 1
1500#define MADE_JUMP_TRIE 2
1501#define MADE_EXACT_TRIE 4
3dab1dad 1502
a3621e74 1503STATIC I32
786e8c11 1504S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1505{
27da23d5 1506 dVAR;
a3621e74
YO
1507 /* first pass, loop through and scan words */
1508 reg_trie_data *trie;
55eed653 1509 HV *widecharmap = NULL;
2b8b4781 1510 AV *revcharmap = newAV();
a3621e74 1511 regnode *cur;
9f7f3913 1512 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1513 STRLEN len = 0;
1514 UV uvc = 0;
1515 U16 curword = 0;
1516 U32 next_alloc = 0;
786e8c11
YO
1517 regnode *jumper = NULL;
1518 regnode *nextbranch = NULL;
7f69552c 1519 regnode *convert = NULL;
2e64971a 1520 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1521 /* we just use folder as a flag in utf8 */
1e696034 1522 const U8 * folder = NULL;
a3621e74 1523
2b8b4781
NC
1524#ifdef DEBUGGING
1525 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1526 AV *trie_words = NULL;
1527 /* along with revcharmap, this only used during construction but both are
1528 * useful during debugging so we store them in the struct when debugging.
8e11feef 1529 */
2b8b4781
NC
1530#else
1531 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1532 STRLEN trie_charcount=0;
3dab1dad 1533#endif
2b8b4781 1534 SV *re_trie_maxbuff;
a3621e74 1535 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1536
1537 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1538#ifndef DEBUGGING
1539 PERL_UNUSED_ARG(depth);
1540#endif
a3621e74 1541
1e696034 1542 switch (flags) {
79a81a6e 1543 case EXACT: break;
2f7f8cb1 1544 case EXACTFA:
fab2782b
YO
1545 case EXACTFU_SS:
1546 case EXACTFU_TRICKYFOLD:
1e696034
KW
1547 case EXACTFU: folder = PL_fold_latin1; break;
1548 case EXACTF: folder = PL_fold; break;
1549 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1550 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1551 }
1552
c944940b 1553 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1554 trie->refcount = 1;
3dab1dad 1555 trie->startstate = 1;
786e8c11 1556 trie->wordcount = word_count;
f8fc2ecf 1557 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1558 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1559 if (flags == EXACT)
c944940b 1560 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1561 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1562 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1563
a3621e74 1564 DEBUG_r({
2b8b4781 1565 trie_words = newAV();
a3621e74 1566 });
a3621e74 1567
0111c4fd 1568 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1569 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1570 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1571 }
df826430 1572 DEBUG_TRIE_COMPILE_r({
3dab1dad 1573 PerlIO_printf( Perl_debug_log,
786e8c11 1574 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1575 (int)depth * 2 + 2, "",
1576 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1577 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1578 (int)depth);
3dab1dad 1579 });
7f69552c
YO
1580
1581 /* Find the node we are going to overwrite */
1582 if ( first == startbranch && OP( last ) != BRANCH ) {
1583 /* whole branch chain */
1584 convert = first;
1585 } else {
1586 /* branch sub-chain */
1587 convert = NEXTOPER( first );
1588 }
1589
a3621e74
YO
1590 /* -- First loop and Setup --
1591
1592 We first traverse the branches and scan each word to determine if it
1593 contains widechars, and how many unique chars there are, this is
1594 important as we have to build a table with at least as many columns as we
1595 have unique chars.
1596
1597 We use an array of integers to represent the character codes 0..255
38a44b82 1598 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1599 native representation of the character value as the key and IV's for the
1600 coded index.
1601
1602 *TODO* If we keep track of how many times each character is used we can
1603 remap the columns so that the table compression later on is more
3b753521 1604 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1605 middle and the least common are on the outside. IMO this would be better
1606 than a most to least common mapping as theres a decent chance the most
1607 common letter will share a node with the least common, meaning the node
486ec47a 1608 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1609 case is when we have the least common nodes twice.
1610
1611 */
1612
a3621e74 1613 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1614 regnode *noper = NEXTOPER( cur );
e1ec3a88 1615 const U8 *uc = (U8*)STRING( noper );
df826430 1616 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1617 STRLEN foldlen = 0;
1618 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1619 STRLEN skiplen = 0;
2af232bd 1620 const U8 *scan = (U8*)NULL;
07be1b83 1621 U32 wordlen = 0; /* required init */
02daf0ab
YO
1622 STRLEN chars = 0;
1623 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1624
3dab1dad 1625 if (OP(noper) == NOTHING) {
df826430
YO
1626 regnode *noper_next= regnext(noper);
1627 if (noper_next != tail && OP(noper_next) == flags) {
1628 noper = noper_next;
1629 uc= (U8*)STRING(noper);
1630 e= uc + STR_LEN(noper);
1631 trie->minlen= STR_LEN(noper);
1632 } else {
1633 trie->minlen= 0;
1634 continue;
1635 }
3dab1dad 1636 }
df826430 1637
fab2782b 1638 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1639 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1640 regardless of encoding */
fab2782b
YO
1641 if (OP( noper ) == EXACTFU_SS) {
1642 /* false positives are ok, so just set this */
1643 TRIE_BITMAP_SET(trie,0xDF);
1644 }
1645 }
a3621e74 1646 for ( ; uc < e ; uc += len ) {
3dab1dad 1647 TRIE_CHARCOUNT(trie)++;
a3621e74 1648 TRIE_READ_CHAR;
3dab1dad 1649 chars++;
a3621e74 1650 if ( uvc < 256 ) {
fab2782b
YO
1651 if ( folder ) {
1652 U8 folded= folder[ (U8) uvc ];
1653 if ( !trie->charmap[ folded ] ) {
1654 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1655 TRIE_STORE_REVCHAR( folded );
1656 }
1657 }
a3621e74
YO
1658 if ( !trie->charmap[ uvc ] ) {
1659 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1660 TRIE_STORE_REVCHAR( uvc );
a3621e74 1661 }
02daf0ab 1662 if ( set_bit ) {
62012aee
KW
1663 /* store the codepoint in the bitmap, and its folded
1664 * equivalent. */
fab2782b 1665 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1666
1667 /* store the folded codepoint */
fab2782b 1668 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1669
1670 if ( !UTF ) {
1671 /* store first byte of utf8 representation of
acdf4139
KW
1672 variant codepoints */
1673 if (! UNI_IS_INVARIANT(uvc)) {
1674 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1675 }
1676 }
02daf0ab
YO
1677 set_bit = 0; /* We've done our bit :-) */
1678 }
a3621e74
YO
1679 } else {
1680 SV** svpp;
55eed653
NC
1681 if ( !widecharmap )
1682 widecharmap = newHV();
a3621e74 1683
55eed653 1684 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1685
1686 if ( !svpp )
e4584336 1687 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1688
1689 if ( !SvTRUE( *svpp ) ) {
1690 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1691 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1692 }
1693 }
1694 }
3dab1dad 1695 if( cur == first ) {
fab2782b
YO
1696 trie->minlen = chars;
1697 trie->maxlen = chars;
3dab1dad 1698 } else if (chars < trie->minlen) {
fab2782b 1699 trie->minlen = chars;
3dab1dad 1700 } else if (chars > trie->maxlen) {
fab2782b
YO
1701 trie->maxlen = chars;
1702 }
1703 if (OP( noper ) == EXACTFU_SS) {
1704 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1705 if (trie->minlen > 1)
1706 trie->minlen= 1;
1707 }
1708 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1709 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1710 * - We assume that any such sequence might match a 2 byte string */
1711 if (trie->minlen > 2 )
1712 trie->minlen= 2;
3dab1dad
YO
1713 }
1714
a3621e74
YO
1715 } /* end first pass */
1716 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1717 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1718 (int)depth * 2 + 2,"",
55eed653 1719 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1720 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1721 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1722 );
a3621e74
YO
1723
1724 /*
1725 We now know what we are dealing with in terms of unique chars and
1726 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1727 representation using a flat table will take. If it's over a reasonable
1728 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1729 conservative but potentially much slower representation using an array
1730 of lists.
1731
1732 At the end we convert both representations into the same compressed
1733 form that will be used in regexec.c for matching with. The latter
1734 is a form that cannot be used to construct with but has memory
1735 properties similar to the list form and access properties similar
1736 to the table form making it both suitable for fast searches and
1737 small enough that its feasable to store for the duration of a program.
1738
1739 See the comment in the code where the compressed table is produced
1740 inplace from the flat tabe representation for an explanation of how
1741 the compression works.
1742
1743 */
1744
1745
2e64971a
DM
1746 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1747 prev_states[1] = 0;
1748
3dab1dad 1749 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1750 /*
1751 Second Pass -- Array Of Lists Representation
1752
1753 Each state will be represented by a list of charid:state records
1754 (reg_trie_trans_le) the first such element holds the CUR and LEN
1755 points of the allocated array. (See defines above).
1756
1757 We build the initial structure using the lists, and then convert
1758 it into the compressed table form which allows faster lookups
1759 (but cant be modified once converted).
a3621e74
YO
1760 */
1761
a3621e74
YO
1762 STRLEN transcount = 1;
1763
1e2e3d02
YO
1764 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1765 "%*sCompiling trie using list compiler\n",
1766 (int)depth * 2 + 2, ""));
686b73d4 1767
c944940b
JH
1768 trie->states = (reg_trie_state *)
1769 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1770 sizeof(reg_trie_state) );
a3621e74
YO
1771 TRIE_LIST_NEW(1);
1772 next_alloc = 2;
1773
1774 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1775
df826430 1776 regnode *noper = NEXTOPER( cur );
c445ea15 1777 U8 *uc = (U8*)STRING( noper );
df826430 1778 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1779 U32 state = 1; /* required init */
1780 U16 charid = 0; /* sanity init */
1781 U8 *scan = (U8*)NULL; /* sanity init */
1782 STRLEN foldlen = 0; /* required init */
07be1b83 1783 U32 wordlen = 0; /* required init */
c445ea15 1784 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1785 STRLEN skiplen = 0;
c445ea15 1786
df826430
YO
1787 if (OP(noper) == NOTHING) {
1788 regnode *noper_next= regnext(noper);
1789 if (noper_next != tail && OP(noper_next) == flags) {
1790 noper = noper_next;
1791 uc= (U8*)STRING(noper);
1792 e= uc + STR_LEN(noper);
1793 }
1794 }
1795
3dab1dad 1796 if (OP(noper) != NOTHING) {
786e8c11 1797 for ( ; uc < e ; uc += len ) {
c445ea15 1798
786e8c11 1799 TRIE_READ_CHAR;
c445ea15 1800
786e8c11
YO
1801 if ( uvc < 256 ) {
1802 charid = trie->charmap[ uvc ];
c445ea15 1803 } else {
55eed653 1804 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1805 if ( !svpp ) {
1806 charid = 0;
1807 } else {
1808 charid=(U16)SvIV( *svpp );
1809 }
c445ea15 1810 }
786e8c11
YO
1811 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1812 if ( charid ) {
a3621e74 1813
786e8c11
YO
1814 U16 check;
1815 U32 newstate = 0;
a3621e74 1816
786e8c11
YO
1817 charid--;
1818 if ( !trie->states[ state ].trans.list ) {
1819 TRIE_LIST_NEW( state );
c445ea15 1820 }
786e8c11
YO
1821 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1822 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1823 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1824 break;
1825 }
1826 }
1827 if ( ! newstate ) {
1828 newstate = next_alloc++;
2e64971a 1829 prev_states[newstate] = state;
786e8c11
YO
1830 TRIE_LIST_PUSH( state, charid, newstate );
1831 transcount++;
1832 }
1833 state = newstate;
1834 } else {
1835 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1836 }
a28509cc 1837 }
c445ea15 1838 }
3dab1dad 1839 TRIE_HANDLE_WORD(state);
a3621e74
YO
1840
1841 } /* end second pass */
1842
1e2e3d02
YO
1843 /* next alloc is the NEXT state to be allocated */
1844 trie->statecount = next_alloc;
c944940b
JH
1845 trie->states = (reg_trie_state *)
1846 PerlMemShared_realloc( trie->states,
1847 next_alloc
1848 * sizeof(reg_trie_state) );
a3621e74 1849
3dab1dad 1850 /* and now dump it out before we compress it */
2b8b4781
NC
1851 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1852 revcharmap, next_alloc,
1853 depth+1)
1e2e3d02 1854 );
a3621e74 1855
c944940b
JH
1856 trie->trans = (reg_trie_trans *)
1857 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1858 {
1859 U32 state;
a3621e74
YO
1860 U32 tp = 0;
1861 U32 zp = 0;
1862
1863
1864 for( state=1 ; state < next_alloc ; state ++ ) {
1865 U32 base=0;
1866
1867 /*
1868 DEBUG_TRIE_COMPILE_MORE_r(
1869 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1870 );
1871 */
1872
1873 if (trie->states[state].trans.list) {
1874 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1875 U16 maxid=minid;
a28509cc 1876 U16 idx;
a3621e74
YO
1877
1878 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1879 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1880 if ( forid < minid ) {
1881 minid=forid;
1882 } else if ( forid > maxid ) {
1883 maxid=forid;
1884 }
a3621e74
YO
1885 }
1886 if ( transcount < tp + maxid - minid + 1) {
1887 transcount *= 2;
c944940b
JH
1888 trie->trans = (reg_trie_trans *)
1889 PerlMemShared_realloc( trie->trans,
446bd890
NC
1890 transcount
1891 * sizeof(reg_trie_trans) );
a3621e74
YO
1892 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1893 }
1894 base = trie->uniquecharcount + tp - minid;
1895 if ( maxid == minid ) {
1896 U32 set = 0;
1897 for ( ; zp < tp ; zp++ ) {
1898 if ( ! trie->trans[ zp ].next ) {
1899 base = trie->uniquecharcount + zp - minid;
1900 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1901 trie->trans[ zp ].check = state;
1902 set = 1;
1903 break;
1904 }
1905 }
1906 if ( !set ) {
1907 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1908 trie->trans[ tp ].check = state;
1909 tp++;
1910 zp = tp;
1911 }
1912 } else {
1913 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1914 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1915 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1916 trie->trans[ tid ].check = state;
1917 }
1918 tp += ( maxid - minid + 1 );
1919 }
1920 Safefree(trie->states[ state ].trans.list);
1921 }
1922 /*
1923 DEBUG_TRIE_COMPILE_MORE_r(
1924 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1925 );
1926 */
1927 trie->states[ state ].trans.base=base;
1928 }
cc601c31 1929 trie->lasttrans = tp + 1;
a3621e74
YO
1930 }
1931 } else {
1932 /*
1933 Second Pass -- Flat Table Representation.
1934
1935 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1936 We know that we will need Charcount+1 trans at most to store the data
1937 (one row per char at worst case) So we preallocate both structures
1938 assuming worst case.
1939
1940 We then construct the trie using only the .next slots of the entry
1941 structs.
1942
3b753521 1943 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1944 make compression both faster and easier by keeping track of how many non
1945 zero fields are in the node.
1946
1947 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1948 transition.
1949
1950 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1951 number representing the first entry of the node, and state as a
1952 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1953 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1954 are 2 entrys per node. eg:
1955
1956 A B A B
1957 1. 2 4 1. 3 7
1958 2. 0 3 3. 0 5
1959 3. 0 0 5. 0 0
1960 4. 0 0 7. 0 0
1961
1962 The table is internally in the right hand, idx form. However as we also
1963 have to deal with the states array which is indexed by nodenum we have to
1964 use TRIE_NODENUM() to convert.
1965
1966 */
1e2e3d02
YO
1967 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1968 "%*sCompiling trie using table compiler\n",
1969 (int)depth * 2 + 2, ""));
3dab1dad 1970
c944940b
JH
1971 trie->trans = (reg_trie_trans *)
1972 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1973 * trie->uniquecharcount + 1,
1974 sizeof(reg_trie_trans) );
1975 trie->states = (reg_trie_state *)
1976 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1977 sizeof(reg_trie_state) );
a3621e74
YO
1978 next_alloc = trie->uniquecharcount + 1;
1979
3dab1dad 1980
a3621e74
YO
1981 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1982
df826430 1983 regnode *noper = NEXTOPER( cur );
a28509cc 1984 const U8 *uc = (U8*)STRING( noper );
df826430 1985 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1986
1987 U32 state = 1; /* required init */
1988
1989 U16 charid = 0; /* sanity init */
1990 U32 accept_state = 0; /* sanity init */
1991 U8 *scan = (U8*)NULL; /* sanity init */
1992
1993 STRLEN foldlen = 0; /* required init */
07be1b83 1994 U32 wordlen = 0; /* required init */
fab2782b 1995 STRLEN skiplen = 0;
a3621e74
YO
1996 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1997
df826430
YO
1998 if (OP(noper) == NOTHING) {
1999 regnode *noper_next= regnext(noper);
2000 if (noper_next != tail && OP(noper_next) == flags) {
2001 noper = noper_next;
2002 uc= (U8*)STRING(noper);
2003 e= uc + STR_LEN(noper);
2004 }
2005 }
fab2782b 2006
3dab1dad 2007 if ( OP(noper) != NOTHING ) {
786e8c11 2008 for ( ; uc < e ; uc += len ) {
a3621e74 2009
786e8c11 2010 TRIE_READ_CHAR;
a3621e74 2011
786e8c11
YO
2012 if ( uvc < 256 ) {
2013 charid = trie->charmap[ uvc ];
2014 } else {
55eed653 2015 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2016 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2017 }
786e8c11
YO
2018 if ( charid ) {
2019 charid--;
2020 if ( !trie->trans[ state + charid ].next ) {
2021 trie->trans[ state + charid ].next = next_alloc;
2022 trie->trans[ state ].check++;
2e64971a
DM
2023 prev_states[TRIE_NODENUM(next_alloc)]
2024 = TRIE_NODENUM(state);
786e8c11
YO
2025 next_alloc += trie->uniquecharcount;
2026 }
2027 state = trie->trans[ state + charid ].next;
2028 } else {
2029 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2030 }
2031 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2032 }
a3621e74 2033 }
3dab1dad
YO
2034 accept_state = TRIE_NODENUM( state );
2035 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2036
2037 } /* end second pass */
2038
3dab1dad 2039 /* and now dump it out before we compress it */
2b8b4781
NC
2040 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2041 revcharmap,
2042 next_alloc, depth+1));
a3621e74 2043
a3621e74
YO
2044 {
2045 /*
2046 * Inplace compress the table.*
2047
2048 For sparse data sets the table constructed by the trie algorithm will
2049 be mostly 0/FAIL transitions or to put it another way mostly empty.
2050 (Note that leaf nodes will not contain any transitions.)
2051
2052 This algorithm compresses the tables by eliminating most such
2053 transitions, at the cost of a modest bit of extra work during lookup:
2054
2055 - Each states[] entry contains a .base field which indicates the
2056 index in the state[] array wheres its transition data is stored.
2057
3b753521 2058 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2059
2060 - If .base is nonzero then charid is added to it to find an entry in
2061 the trans array.
2062
2063 -If trans[states[state].base+charid].check!=state then the
2064 transition is taken to be a 0/Fail transition. Thus if there are fail
2065 transitions at the front of the node then the .base offset will point
2066 somewhere inside the previous nodes data (or maybe even into a node
2067 even earlier), but the .check field determines if the transition is
2068 valid.
2069
786e8c11 2070 XXX - wrong maybe?
a3621e74 2071 The following process inplace converts the table to the compressed
3b753521 2072 table: We first do not compress the root node 1,and mark all its
a3621e74 2073 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2074 allows us to do a DFA construction from the compressed table later,
2075 and ensures that any .base pointers we calculate later are greater
2076 than 0.
a3621e74
YO
2077
2078 - We set 'pos' to indicate the first entry of the second node.
2079
2080 - We then iterate over the columns of the node, finding the first and
2081 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2082 and set the .check pointers accordingly, and advance pos
2083 appropriately and repreat for the next node. Note that when we copy
2084 the next pointers we have to convert them from the original
2085 NODEIDX form to NODENUM form as the former is not valid post
2086 compression.
2087
2088 - If a node has no transitions used we mark its base as 0 and do not
2089 advance the pos pointer.
2090
2091 - If a node only has one transition we use a second pointer into the
2092 structure to fill in allocated fail transitions from other states.
2093 This pointer is independent of the main pointer and scans forward
2094 looking for null transitions that are allocated to a state. When it
2095 finds one it writes the single transition into the "hole". If the
786e8c11 2096 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2097
2098 - Once compressed we can Renew/realloc the structures to release the
2099 excess space.
2100
2101 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2102 specifically Fig 3.47 and the associated pseudocode.
2103
2104 demq
2105 */
a3b680e6 2106 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2107 U32 state, charid;
a3621e74 2108 U32 pos = 0, zp=0;
1e2e3d02 2109 trie->statecount = laststate;
a3621e74
YO
2110
2111 for ( state = 1 ; state < laststate ; state++ ) {
2112 U8 flag = 0;
a28509cc
AL
2113 const U32 stateidx = TRIE_NODEIDX( state );
2114 const U32 o_used = trie->trans[ stateidx ].check;
2115 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2116 trie->trans[ stateidx ].check = 0;
2117
2118 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2119 if ( flag || trie->trans[ stateidx + charid ].next ) {
2120 if ( trie->trans[ stateidx + charid ].next ) {
2121 if (o_used == 1) {
2122 for ( ; zp < pos ; zp++ ) {
2123 if ( ! trie->trans[ zp ].next ) {
2124 break;
2125 }
2126 }
2127 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2128 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2129 trie->trans[ zp ].check = state;
2130 if ( ++zp > pos ) pos = zp;
2131 break;
2132 }
2133 used--;
2134 }
2135 if ( !flag ) {
2136 flag = 1;
2137 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2138 }
2139 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2140 trie->trans[ pos ].check = state;
2141 pos++;
2142 }
2143 }
2144 }
cc601c31 2145 trie->lasttrans = pos + 1;
c944940b
JH
2146 trie->states = (reg_trie_state *)
2147 PerlMemShared_realloc( trie->states, laststate
2148 * sizeof(reg_trie_state) );
a3621e74 2149 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2150 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2151 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2152 (int)depth * 2 + 2,"",
2153 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2154 (IV)next_alloc,
2155 (IV)pos,
a3621e74
YO
2156 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2157 );
2158
2159 } /* end table compress */
2160 }
1e2e3d02
YO
2161 DEBUG_TRIE_COMPILE_MORE_r(
2162 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2163 (int)depth * 2 + 2, "",
2164 (UV)trie->statecount,
2165 (UV)trie->lasttrans)
2166 );
cc601c31 2167 /* resize the trans array to remove unused space */
c944940b
JH
2168 trie->trans = (reg_trie_trans *)
2169 PerlMemShared_realloc( trie->trans, trie->lasttrans
2170 * sizeof(reg_trie_trans) );
a3621e74 2171
3b753521 2172 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2173 U8 nodetype =(U8)(flags & 0xFF);
2174 char *str=NULL;
786e8c11 2175
07be1b83 2176#ifdef DEBUGGING
e62cc96a 2177 regnode *optimize = NULL;
7122b237
YO
2178#ifdef RE_TRACK_PATTERN_OFFSETS
2179
b57a0404
JH
2180 U32 mjd_offset = 0;
2181 U32 mjd_nodelen = 0;
7122b237
YO
2182#endif /* RE_TRACK_PATTERN_OFFSETS */
2183#endif /* DEBUGGING */
a3621e74 2184 /*
3dab1dad
YO
2185 This means we convert either the first branch or the first Exact,
2186 depending on whether the thing following (in 'last') is a branch
2187 or not and whther first is the startbranch (ie is it a sub part of
2188 the alternation or is it the whole thing.)
3b753521 2189 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2190 the whole branch sequence, including the first.
a3621e74 2191 */
3dab1dad 2192 /* Find the node we are going to overwrite */
7f69552c 2193 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2194 /* branch sub-chain */
3dab1dad 2195 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2196#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2197 DEBUG_r({
2198 mjd_offset= Node_Offset((convert));
2199 mjd_nodelen= Node_Length((convert));
2200 });
7122b237 2201#endif
7f69552c 2202 /* whole branch chain */
7122b237
YO
2203 }
2204#ifdef RE_TRACK_PATTERN_OFFSETS
2205 else {
7f69552c
YO
2206 DEBUG_r({
2207 const regnode *nop = NEXTOPER( convert );
2208 mjd_offset= Node_Offset((nop));
2209 mjd_nodelen= Node_Length((nop));
2210 });
07be1b83
YO
2211 }
2212 DEBUG_OPTIMISE_r(
2213 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2214 (int)depth * 2 + 2, "",
786e8c11 2215 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2216 );
7122b237 2217#endif
3dab1dad
YO
2218 /* But first we check to see if there is a common prefix we can
2219 split out as an EXACT and put in front of the TRIE node. */
2220 trie->startstate= 1;
55eed653 2221 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2222 U32 state;
1e2e3d02 2223 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2224 U32 ofs = 0;
8e11feef
RGS
2225 I32 idx = -1;
2226 U32 count = 0;
2227 const U32 base = trie->states[ state ].trans.base;
a3621e74 2228
3dab1dad 2229 if ( trie->states[state].wordnum )
8e11feef 2230 count = 1;
a3621e74 2231
8e11feef 2232 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2233 if ( ( base + ofs >= trie->uniquecharcount ) &&
2234 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2235 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2236 {
3dab1dad 2237 if ( ++count > 1 ) {
2b8b4781 2238 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2239 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2240 if ( state == 1 ) break;
3dab1dad
YO
2241 if ( count == 2 ) {
2242 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2243 DEBUG_OPTIMISE_r(
8e11feef
RGS
2244 PerlIO_printf(Perl_debug_log,
2245 "%*sNew Start State=%"UVuf" Class: [",
2246 (int)depth * 2 + 2, "",
786e8c11 2247 (UV)state));
be8e71aa 2248 if (idx >= 0) {
2b8b4781 2249 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2250 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2251
3dab1dad 2252 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2253 if ( folder )
2254 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2255 DEBUG_OPTIMISE_r(
f1f66076 2256 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2257 );
8e11feef
RGS
2258 }
2259 }
2260 TRIE_BITMAP_SET(trie,*ch);
2261 if ( folder )
2262 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2263 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2264 }
2265 idx = ofs;
2266 }
3dab1dad
YO
2267 }
2268 if ( count == 1 ) {
2b8b4781 2269 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2270 STRLEN len;
2271 char *ch = SvPV( *tmp, len );
de734bd5
A
2272 DEBUG_OPTIMISE_r({
2273 SV *sv=sv_newmortal();
8e11feef
RGS
2274 PerlIO_printf( Perl_debug_log,
2275 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2276 (int)depth * 2 + 2, "",
de734bd5
A
2277 (UV)state, (UV)idx,
2278 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2279 PL_colors[0], PL_colors[1],
2280 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2281 PERL_PV_ESCAPE_FIRSTCHAR
2282 )
2283 );
2284 });
3dab1dad
YO
2285 if ( state==1 ) {
2286 OP( convert ) = nodetype;
2287 str=STRING(convert);
2288 STR_LEN(convert)=0;
2289 }
c490c714
YO
2290 STR_LEN(convert) += len;
2291 while (len--)
de734bd5 2292 *str++ = *ch++;
8e11feef 2293 } else {
f9049ba1 2294#ifdef DEBUGGING
8e11feef
RGS
2295 if (state>1)
2296 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2297#endif
8e11feef
RGS
2298 break;
2299 }
2300 }
2e64971a 2301 trie->prefixlen = (state-1);
3dab1dad 2302 if (str) {
8e11feef 2303 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2304 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2305 trie->startstate = state;
07be1b83
YO
2306 trie->minlen -= (state - 1);
2307 trie->maxlen -= (state - 1);
33809eae
JH
2308#ifdef DEBUGGING
2309 /* At least the UNICOS C compiler choked on this
2310 * being argument to DEBUG_r(), so let's just have
2311 * it right here. */
2312 if (
2313#ifdef PERL_EXT_RE_BUILD
2314 1
2315#else
2316 DEBUG_r_TEST
2317#endif
2318 ) {
2319 regnode *fix = convert;
2320 U32 word = trie->wordcount;
2321 mjd_nodelen++;
2322 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2323 while( ++fix < n ) {
2324 Set_Node_Offset_Length(fix, 0, 0);
2325 }
2326 while (word--) {
2327 SV ** const tmp = av_fetch( trie_words, word, 0 );
2328 if (tmp) {
2329 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2330 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2331 else
2332 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2333 }
2334 }
2335 }
2336#endif
8e11feef
RGS
2337 if (trie->maxlen) {
2338 convert = n;
2339 } else {
3dab1dad 2340 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2341 DEBUG_r(optimize= n);
3dab1dad
YO
2342 }
2343 }
2344 }
a5ca303d
YO
2345 if (!jumper)
2346 jumper = last;
3dab1dad 2347 if ( trie->maxlen ) {
8e11feef
RGS
2348 NEXT_OFF( convert ) = (U16)(tail - convert);
2349 ARG_SET( convert, data_slot );
786e8c11
YO
2350 /* Store the offset to the first unabsorbed branch in
2351 jump[0], which is otherwise unused by the jump logic.
2352 We use this when dumping a trie and during optimisation. */
2353 if (trie->jump)
7f69552c 2354 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2355
6c48061a
YO
2356 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2357 * and there is a bitmap
2358 * and the first "jump target" node we found leaves enough room
2359 * then convert the TRIE node into a TRIEC node, with the bitmap
2360 * embedded inline in the opcode - this is hypothetically faster.
2361 */
2362 if ( !trie->states[trie->startstate].wordnum
2363 && trie->bitmap
2364 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2365 {
2366 OP( convert ) = TRIEC;
2367 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2368 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2369 trie->bitmap= NULL;
2370 } else
2371 OP( convert ) = TRIE;
a3621e74 2372
3dab1dad
YO
2373 /* store the type in the flags */
2374 convert->flags = nodetype;
a5ca303d
YO
2375 DEBUG_r({
2376 optimize = convert
2377 + NODE_STEP_REGNODE
2378 + regarglen[ OP( convert ) ];
2379 });
2380 /* XXX We really should free up the resource in trie now,
2381 as we won't use them - (which resources?) dmq */
3dab1dad 2382 }
a3621e74 2383 /* needed for dumping*/
e62cc96a 2384 DEBUG_r(if (optimize) {
07be1b83 2385 regnode *opt = convert;
bcdf7404 2386
e62cc96a 2387 while ( ++opt < optimize) {
07be1b83
YO
2388 Set_Node_Offset_Length(opt,0,0);
2389 }
786e8c11
YO
2390 /*
2391 Try to clean up some of the debris left after the
2392 optimisation.
a3621e74 2393 */
786e8c11 2394 while( optimize < jumper ) {
07be1b83 2395 mjd_nodelen += Node_Length((optimize));
a3621e74 2396 OP( optimize ) = OPTIMIZED;
07be1b83 2397 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2398 optimize++;
2399 }
07be1b83 2400 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2401 });
2402 } /* end node insert */
2e64971a
DM
2403
2404 /* Finish populating the prev field of the wordinfo array. Walk back
2405 * from each accept state until we find another accept state, and if
2406 * so, point the first word's .prev field at the second word. If the
2407 * second already has a .prev field set, stop now. This will be the
2408 * case either if we've already processed that word's accept state,
3b753521
FN
2409 * or that state had multiple words, and the overspill words were
2410 * already linked up earlier.
2e64971a
DM
2411 */
2412 {
2413 U16 word;
2414 U32 state;
2415 U16 prev;
2416
2417 for (word=1; word <= trie->wordcount; word++) {
2418 prev = 0;
2419 if (trie->wordinfo[word].prev)
2420 continue;
2421 state = trie->wordinfo[word].accept;
2422 while (state) {
2423 state = prev_states[state];
2424 if (!state)
2425 break;
2426 prev = trie->states[state].wordnum;
2427 if (prev)
2428 break;
2429 }
2430 trie->wordinfo[word].prev = prev;
2431 }
2432 Safefree(prev_states);
2433 }
2434
2435
2436 /* and now dump out the compressed format */
2437 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2438
55eed653 2439 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2440#ifdef DEBUGGING
2441 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2442 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2443#else
03e70be4 2444 SvREFCNT_dec_NN(revcharmap);
07be1b83 2445#endif
786e8c11
YO
2446 return trie->jump
2447 ? MADE_JUMP_TRIE
2448 : trie->startstate>1
2449 ? MADE_EXACT_TRIE
2450 : MADE_TRIE;
2451}
2452
2453STATIC void
2454S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2455{
3b753521 2456/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2457
2458 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2459 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2460 ISBN 0-201-10088-6
2461
2462 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2463 suffix of the current state's 'word' that is also a proper prefix of another word in our
2464 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2465 the DFA not to have to restart after its tried and failed a word at a given point, it
2466 simply continues as though it had been matching the other word in the first place.
2467 Consider
2468 'abcdgu'=~/abcdefg|cdgu/
2469 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2470 fail, which would bring us to the state representing 'd' in the second word where we would
2471 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2472 */
2473 /* add a fail transition */
3251b653
NC
2474 const U32 trie_offset = ARG(source);
2475 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2476 U32 *q;
2477 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2478 const U32 numstates = trie->statecount;
786e8c11
YO
2479 const U32 ubound = trie->lasttrans + ucharcount;
2480 U32 q_read = 0;
2481 U32 q_write = 0;
2482 U32 charid;
2483 U32 base = trie->states[ 1 ].trans.base;
2484 U32 *fail;
2485 reg_ac_data *aho;
2486 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2487 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2488
2489 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2490#ifndef DEBUGGING
2491 PERL_UNUSED_ARG(depth);
2492#endif
2493
2494
2495 ARG_SET( stclass, data_slot );
c944940b 2496 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2497 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2498 aho->trie=trie_offset;
446bd890
NC
2499 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2500 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2501 Newxz( q, numstates, U32);
c944940b 2502 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2503 aho->refcount = 1;
2504 fail = aho->fail;
2505 /* initialize fail[0..1] to be 1 so that we always have
2506 a valid final fail state */
2507 fail[ 0 ] = fail[ 1 ] = 1;
2508
2509 for ( charid = 0; charid < ucharcount ; charid++ ) {
2510 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2511 if ( newstate ) {
2512 q[ q_write ] = newstate;
2513 /* set to point at the root */
2514 fail[ q[ q_write++ ] ]=1;
2515 }
2516 }
2517 while ( q_read < q_write) {
2518 const U32 cur = q[ q_read++ % numstates ];
2519 base = trie->states[ cur ].trans.base;
2520
2521 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2522 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2523 if (ch_state) {
2524 U32 fail_state = cur;
2525 U32 fail_base;
2526 do {
2527 fail_state = fail[ fail_state ];
2528 fail_base = aho->states[ fail_state ].trans.base;
2529 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2530
2531 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2532 fail[ ch_state ] = fail_state;
2533 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2534 {
2535 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2536 }
2537 q[ q_write++ % numstates] = ch_state;
2538 }
2539 }
2540 }
2541 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2542 when we fail in state 1, this allows us to use the
2543 charclass scan to find a valid start char. This is based on the principle
2544 that theres a good chance the string being searched contains lots of stuff
2545 that cant be a start char.
2546 */
2547 fail[ 0 ] = fail[ 1 ] = 0;
2548 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2549 PerlIO_printf(Perl_debug_log,
2550 "%*sStclass Failtable (%"UVuf" states): 0",
2551 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2552 );
786e8c11
YO
2553 for( q_read=1; q_read<numstates; q_read++ ) {
2554 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2555 }
2556 PerlIO_printf(Perl_debug_log, "\n");
2557 });
2558 Safefree(q);
2559 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2560}
2561
786e8c11 2562
a3621e74 2563/*
5d1c421c
JH
2564 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2565 * These need to be revisited when a newer toolchain becomes available.
2566 */
2567#if defined(__sparc64__) && defined(__GNUC__)
2568# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2569# undef SPARC64_GCC_WORKAROUND
2570# define SPARC64_GCC_WORKAROUND 1
2571# endif
2572#endif
2573
07be1b83 2574#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2575 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2576 SV * const mysv=sv_newmortal(); \
2577 regnode *Next = regnext(scan); \
2578 regprop(RExC_rx, mysv, scan); \
7f69552c 2579 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2580 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2581 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2582 }});
07be1b83 2583
1de06328 2584
bb914485 2585/* The below joins as many adjacent EXACTish nodes as possible into a single
0a982f06
KW
2586 * one. The regop may be changed if the node(s) contain certain sequences that
2587 * require special handling. The joining is only done if:
bb914485
KW
2588 * 1) there is room in the current conglomerated node to entirely contain the
2589 * next one.
2590 * 2) they are the exact same node type
2591 *
87b8b349 2592 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
2593 * these get optimized out
2594 *
0a982f06
KW
2595 * If a node is to match under /i (folded), the number of characters it matches
2596 * can be different than its character length if it contains a multi-character
2597 * fold. *min_subtract is set to the total delta of the input nodes.
bb914485 2598 *
a0c4c608
KW
2599 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2600 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2601 *
bb914485 2602 * This is as good a place as any to discuss the design of handling these
0a982f06
KW
2603 * multi-character fold sequences. It's been wrong in Perl for a very long
2604 * time. There are three code points in Unicode whose multi-character folds
2605 * were long ago discovered to mess things up. The previous designs for
2606 * dealing with these involved assigning a special node for them. This
2607 * approach doesn't work, as evidenced by this example:
a0c4c608 2608 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
0a982f06
KW
2609 * Both these fold to "sss", but if the pattern is parsed to create a node that
2610 * would match just the \xDF, it won't be able to handle the case where a
bb914485
KW
2611 * successful match would have to cross the node's boundary. The new approach
2612 * that hopefully generally solves the problem generates an EXACTFU_SS node
2613 * that is "sss".
2614 *
0a982f06
KW
2615 * It turns out that there are problems with all multi-character folds, and not
2616 * just these three. Now the code is general, for all such cases, but the
2617 * three still have some special handling. The approach taken is:
2618 * 1) This routine examines each EXACTFish node that could contain multi-
2619 * character fold sequences. It returns in *min_subtract how much to
9d071ca8 2620 * subtract from the the actual length of the string to get a real minimum
0a982f06
KW
2621 * match length; it is 0 if there are no multi-char folds. This delta is
2622 * used by the caller to adjust the min length of the match, and the delta
2623 * between min and max, so that the optimizer doesn't reject these
2624 * possibilities based on size constraints.
2625 * 2) Certain of these sequences require special handling by the trie code,
2626 * so, if found, this code changes the joined node type to special ops:
2627 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2628 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2629 * is used for an EXACTFU node that contains at least one "ss" sequence in
2630 * it. For non-UTF-8 patterns and strings, this is the only case where
2631 * there is a possible fold length change. That means that a regular
2632 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2633 * with length changes, and so can be processed faster. regexec.c takes
2634 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2635 * pre-folded by regcomp.c. This saves effort in regex matching.
87b8b349 2636 * However, the pre-folding isn't done for non-UTF8 patterns because the
0a982f06
KW
2637 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2638 * down by forcing the pattern into UTF8 unless necessary. Also what
2639 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2640 * possibilities for the non-UTF8 patterns are quite simple, except for
2641 * the sharp s. All the ones that don't involve a UTF-8 target string are
2642 * members of a fold-pair, and arrays are set up for all of them so that
2643 * the other member of the pair can be found quickly. Code elsewhere in
2644 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2645 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2646 * described in the next item.
bb914485
KW
2647 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2648 * 'ss' or not is not knowable at compile time. It will match iff the
2649 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2650 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
6012a526 2651 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
bb914485
KW
2652 * described in item 3). An assumption that the optimizer part of
2653 * regexec.c (probably unwittingly) makes is that a character in the
2654 * pattern corresponds to at most a single character in the target string.
2655 * (And I do mean character, and not byte here, unlike other parts of the
2656 * documentation that have never been updated to account for multibyte
2657 * Unicode.) This assumption is wrong only in this case, as all other
2658 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2659 * virtue of having this file pre-fold UTF-8 patterns. I'm
2660 * reluctant to try to change this assumption, so instead the code punts.
9d071ca8
KW
2661 * This routine examines EXACTF nodes for the sharp s, and returns a
2662 * boolean indicating whether or not the node is an EXACTF node that
2663 * contains a sharp s. When it is true, the caller sets a flag that later
2664 * causes the optimizer in this file to not set values for the floating
2665 * and fixed string lengths, and thus avoids the optimizer code in
2666 * regexec.c that makes the invalid assumption. Thus, there is no
2667 * optimization based on string lengths for EXACTF nodes that contain the
2668 * sharp s. This only happens for /id rules (which means the pattern
2669 * isn't in UTF-8).
bb914485 2670 */
1de06328 2671
9d071ca8 2672#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2673 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2674 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2675
be8e71aa 2676STATIC U32
9d071ca8 2677S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, UV *min_subtract, bool *has_exactf_sharp_s, U32 flags,regnode *val, U32 depth) {
07be1b83
YO
2678 /* Merge several consecutive EXACTish nodes into one. */
2679 regnode *n = regnext(scan);
2680 U32 stringok = 1;
2681 regnode *next = scan + NODE_SZ_STR(scan);
2682 U32 merged = 0;
2683 U32 stopnow = 0;
2684#ifdef DEBUGGING
2685 regnode *stop = scan;
72f13be8 2686 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2687#else
d47053eb
RGS
2688 PERL_UNUSED_ARG(depth);
2689#endif
7918f24d
NC
2690
2691 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2692#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2693 PERL_UNUSED_ARG(flags);
2694 PERL_UNUSED_ARG(val);
07be1b83 2695#endif
07be1b83 2696 DEBUG_PEEP("join",scan,depth);
bb914485 2697
3f410cf6
KW
2698 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2699 * EXACT ones that are mergeable to the current one. */
2700 while (n
2701 && (PL_regkind[OP(n)] == NOTHING
2702 || (stringok && OP(n) == OP(scan)))
07be1b83 2703 && NEXT_OFF(n)
3f410cf6
KW
2704 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2705 {
07be1b83
YO
2706
2707 if (OP(n) == TAIL || n > next)
2708 stringok = 0;
2709 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2710 DEBUG_PEEP("skip:",n,depth);
2711 NEXT_OFF(scan) += NEXT_OFF(n);
2712 next = n + NODE_STEP_REGNODE;
2713#ifdef DEBUGGING
2714 if (stringok)
2715 stop = n;
2716#endif
2717 n = regnext(n);
2718 }
2719 else if (stringok) {
786e8c11 2720 const unsigned int oldl = STR_LEN(scan);
07be1b83 2721 regnode * const nnext = regnext(n);
b2230d39 2722
87b8b349
KW
2723 /* XXX I (khw) kind of doubt that this works on platforms where
2724 * U8_MAX is above 255 because of lots of other assumptions */
79a81a6e 2725 /* Don't join if the sum can't fit into a single node */
b2230d39
KW
2726 if (oldl + STR_LEN(n) > U8_MAX)
2727 break;
07be1b83
YO
2728
2729 DEBUG_PEEP("merg",n,depth);
07be1b83 2730 merged++;
b2230d39 2731
07be1b83
YO
2732 NEXT_OFF(scan) += NEXT_OFF(n);
2733 STR_LEN(scan) += STR_LEN(n);
2734 next = n + NODE_SZ_STR(n);
2735 /* Now we can overwrite *n : */
2736 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2737#ifdef DEBUGGING
2738 stop = next - 1;
2739#endif
2740 n = nnext;
2741 if (stopnow) break;
2742 }
2743
d47053eb
RGS
2744#ifdef EXPERIMENTAL_INPLACESCAN
2745 if (flags && !NEXT_OFF(n)) {
2746 DEBUG_PEEP("atch", val, depth);
2747 if (reg_off_by_arg[OP(n)]) {
2748 ARG_SET(n, val - n);
2749 }
2750 else {
2751 NEXT_OFF(n) = val - n;
2752 }
2753 stopnow = 1;
2754 }
07be1b83
YO
2755#endif
2756 }
2c2b7f86 2757
9d071ca8 2758 *min_subtract = 0;
f758bddf 2759 *has_exactf_sharp_s = FALSE;
f646642f 2760
3f410cf6
KW
2761 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2762 * can now analyze for sequences of problematic code points. (Prior to
2763 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2764 * hence missed). The sequences only happen in folding, hence for any
2765 * non-EXACT EXACTish node */
86d6fcad 2766 if (OP(scan) != EXACT) {
0a982f06
KW
2767 const U8 * const s0 = (U8*) STRING(scan);
2768 const U8 * s = s0;
2769 const U8 * const s_end = s0 + STR_LEN(scan);
f758bddf
KW
2770
2771 /* One pass is made over the node's string looking for all the
2772 * possibilities. to avoid some tests in the loop, there are two main
2773 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2774 * non-UTF-8 */
2775 if (UTF) {
86d6fcad 2776
0a982f06
KW
2777 /* Examine the string for a multi-character fold sequence. UTF-8
2778 * patterns have all characters pre-folded by the time this code is
2779 * executed */
2780 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2781 length sequence we are looking for is 2 */
86d6fcad 2782 {
0a982f06
KW
2783 int count = 0;
2784 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2785 if (! len) { /* Not a multi-char fold: get next char */
2786 s += UTF8SKIP(s);
2787 continue;
2788 }
bb914485 2789
0a982f06
KW
2790 /* Nodes with 'ss' require special handling, except for EXACTFL
2791 * and EXACTFA for which there is no multi-char fold to this */
2792 if (len == 2 && *s == 's' && *(s+1) == 's'
2793 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2794 {
2795 count = 2;
2796 OP(scan) = EXACTFU_SS;
2797 s += 2;
2798 }
2799 else if (len == 6 /* len is the same in both ASCII and EBCDIC for these */
2800 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2801 COMBINING_DIAERESIS_UTF8
2802 COMBINING_ACUTE_ACCENT_UTF8,
2803 6)
2804 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2805 COMBINING_DIAERESIS_UTF8
2806 COMBINING_ACUTE_ACCENT_UTF8,
2807 6)))
2808 {
2809 count = 3;
2810
2811 /* These two folds require special handling by trie's, so
2812 * change the node type to indicate this. If EXACTFA and
2813 * EXACTFL were ever to be handled by trie's, this would
2814 * have to be changed. If this node has already been
2815 * changed to EXACTFU_SS in this loop, leave it as is. (I
2816 * (khw) think it doesn't matter in regexec.c for UTF
2817 * patterns, but no need to change it */
2818 if (OP(scan) == EXACTFU) {
2819 OP(scan) = EXACTFU_TRICKYFOLD;
2820 }
2821 s += 6;
2822 }
2823 else { /* Here is a generic multi-char fold. */
2824 const U8* multi_end = s + len;
2825
2826 /* Count how many characters in it. In the case of /l and
2827 * /aa, no folds which contain ASCII code points are
2828 * allowed, so check for those, and skip if found. (In
2829 * EXACTFL, no folds are allowed to any Latin1 code point,
2830 * not just ASCII. But there aren't any of these
2831 * currently, nor ever likely, so don't take the time to
2832 * test for them. The code that generates the
2833 * is_MULTI_foo() macros croaks should one actually get put
2834 * into Unicode .) */
2835 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2836 count = utf8_length(s, multi_end);
2837 s = multi_end;
2838 }
2839 else {
2840 while (s < multi_end) {
2841 if (isASCII(*s)) {
2842 s++;
2843 goto next_iteration;
2844 }
2845 else {
2846 s += UTF8SKIP(s);
2847 }
2848 count++;
2849 }
2850 }
2851 }
f758bddf 2852
0a982f06
KW
2853 /* The delta is how long the sequence is minus 1 (1 is how long
2854 * the character that folds to the sequence is) */
2855 *min_subtract += count - 1;
2856 next_iteration: ;
bb914485
KW
2857 }
2858 }
f758bddf 2859 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
bb914485 2860
0a982f06
KW
2861 /* Here, the pattern is not UTF-8. Look for the multi-char folds
2862 * that are all ASCII. As in the above case, EXACTFL and EXACTFA
2863 * nodes can't have multi-char folds to this range (and there are
2db97d41 2864 * no existing ones in the upper latin1 range). In the EXACTF
0a982f06
KW
2865 * case we look also for the sharp s, which can be in the final
2866 * position. Otherwise we can stop looking 1 byte earlier because
2867 * have to find at least two characters for a multi-fold */
f758bddf
KW
2868 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2869
0a982f06
KW
2870 /* The below is perhaps overboard, but this allows us to save a
2871 * test each time through the loop at the expense of a mask. This
2872 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2873 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2874 * are 64. This uses an exclusive 'or' to find that bit and then
2875 * inverts it to form a mask, with just a single 0, in the bit
2876 * position where 'S' and 's' differ. */
2877 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2878 const U8 s_masked = 's' & S_or_s_mask;
2879
2880 while (s < upper) {
40b1ba4f 2881 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
0a982f06
KW
2882 if (! len) { /* Not a multi-char fold. */
2883 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2884 {
2885 *has_exactf_sharp_s = TRUE;
2886 }
2887 s++;
2888 continue;
2889 }
2890
2891 if (len == 2
2892 && ((*s & S_or_s_mask) == s_masked)
2893 && ((*(s+1) & S_or_s_mask) == s_masked))
2894 {
2895
2896 /* EXACTF nodes need to know that the minimum length
2897 * changed so that a sharp s in the string can match this
2898 * ss in the pattern, but they remain EXACTF nodes, as they
2899 * won't match this unless the target string is is UTF-8,
2900 * which we don't know until runtime */
2901 if (OP(scan) != EXACTF) {
2902 OP(scan) = EXACTFU_SS;
2903 }
86d6fcad 2904 }
0a982f06
KW
2905
2906 *min_subtract += len - 1;
2907 s += len;
86d6fcad
KW
2908 }
2909 }
07be1b83 2910 }
3f410cf6 2911
07be1b83 2912#ifdef DEBUGGING
bb789b09
DM
2913 /* Allow dumping but overwriting the collection of skipped
2914 * ops and/or strings with fake optimized ops */
07be1b83
YO
2915 n = scan + NODE_SZ_STR(scan);
2916 while (n <= stop) {
bb789b09
DM
2917 OP(n) = OPTIMIZED;
2918 FLAGS(n) = 0;
2919 NEXT_OFF(n) = 0;
07be1b83
YO
2920 n++;
2921 }
2922#endif
2923 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2924 return stopnow;
2925}
2926
486ec47a 2927/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2928 Finds fixed substrings. */
2929
a0288114 2930/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
2931 to the position after last scanned or to NULL. */
2932
40d049e4
YO
2933#define INIT_AND_WITHP \
2934 assert(!and_withp); \
2935 Newx(and_withp,1,struct regnode_charclass_class); \
2936 SAVEFREEPV(and_withp)
07be1b83 2937
b515a41d 2938/* this is a chain of data about sub patterns we are processing that
486ec47a 2939 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
2940 we can simulate recursion without losing state. */
2941struct scan_frame;
2942typedef struct scan_frame {
2943 regnode *last; /* last node to process in this frame */
2944 regnode *next; /* next node to process when last is reached */
2945 struct scan_frame *prev; /*previous frame*/
2946 I32 stop; /* what stopparen do we use */
2947} scan_frame;
2948
304ee84b
YO
2949
2950#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2951
76e3520e 2952STATIC I32
40d049e4 2953S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 2954 I32 *minlenp, I32 *deltap,
40d049e4
YO
2955 regnode *last,
2956 scan_data_t *data,
2957 I32 stopparen,
2958 U8* recursed,
2959 struct regnode_charclass_class *and_withp,
2960 U32 flags, U32 depth)
c277df42
IZ
2961 /* scanp: Start here (read-write). */
2962 /* deltap: Write maxlen-minlen here. */
2963 /* last: Stop before this one. */
40d049e4
YO
2964 /* data: string data about the pattern */
2965 /* stopparen: treat close N as END */
2966 /* recursed: which subroutines have we recursed into */
2967 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 2968{
97aff369 2969 dVAR;
2d608413
KW
2970 I32 min = 0; /* There must be at least this number of characters to match */
2971 I32 pars = 0, code;
c277df42
IZ
2972 regnode *scan = *scanp, *next;
2973 I32 delta = 0;
2974 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 2975 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
2976 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2977 scan_data_t data_fake;
a3621e74 2978 SV *re_trie_maxbuff = NULL;
786e8c11 2979 regnode *first_non_open = scan;
e2e6a0f1 2980 I32 stopmin = I32_MAX;
8aa23a47 2981 scan_frame *frame = NULL;
a3621e74 2982 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 2983
7918f24d
NC
2984 PERL_ARGS_ASSERT_STUDY_CHUNK;
2985
13a24bad 2986#ifdef DEBUGGING
40d049e4 2987 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 2988#endif
40d049e4 2989
786e8c11 2990 if ( depth == 0 ) {
40d049e4 2991 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
2992 first_non_open=regnext(first_non_open);
2993 }
2994
b81d288d 2995
8aa23a47
YO
2996 fake_study_recurse:
2997 while ( scan && OP(scan) != END && scan < last ){
2d608413
KW
2998 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
2999 node length to get a real minimum (because
3000 the folded version may be shorter) */
f758bddf 3001 bool has_exactf_sharp_s = FALSE;
8aa23a47 3002 /* Peephole optimizer: */
304ee84b 3003 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 3004 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
3005
3006 /* Its not clear to khw or hv why this is done here, and not in the
3007 * clauses that deal with EXACT nodes. khw's guess is that it's
3008 * because of a previous design */
9d071ca8 3009 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
3010
3011 /* Follow the next-chain of the current node and optimize
3012 away all the NOTHINGs from it. */
3013 if (OP(scan) != CURLYX) {
3014 const int max = (reg_off_by_arg[OP(scan)]
3015 ? I32_MAX
3016 /* I32 may be smaller than U16 on CRAYs! */
3017 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3018 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3019 int noff;
3020 regnode *n = scan;
686b73d4 3021
8aa23a47
YO
3022 /* Skip NOTHING and LONGJMP. */
3023 while ((n = regnext(n))
3024 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3025 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3026 && off + noff < max)
3027 off += noff;
3028 if (reg_off_by_arg[OP(scan)])
3029 ARG(scan) = off;
3030 else
3031 NEXT_OFF(scan) = off;
3032 }
a3621e74 3033
c277df42 3034
8aa23a47
YO
3035
3036 /* The principal pseudo-switch. Cannot be a switch, since we
3037 look into several different things. */
3038 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3039 || OP(scan) == IFTHEN) {
3040 next = regnext(scan);
3041 code = OP(scan);
3042 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3043
8aa23a47
YO
3044 if (OP(next) == code || code == IFTHEN) {
3045 /* NOTE - There is similar code to this block below for handling
3046 TRIE nodes on a re-study. If you change stuff here check there
3047 too. */
3048 I32 max1 = 0, min1 = I32_MAX, num = 0;
3049 struct regnode_charclass_class accum;
3050 regnode * const startbranch=scan;
686b73d4 3051
8aa23a47 3052 if (flags & SCF_DO_SUBSTR)
304ee84b 3053 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3054 if (flags & SCF_DO_STCLASS)
e755fd73 3055 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3056
3057 while (OP(scan) == code) {
3058 I32 deltanext, minnext, f = 0, fake;
3059 struct regnode_charclass_class this_class;
3060
3061 num++;
3062 data_fake.flags = 0;
3063 if (data) {
3064 data_fake.whilem_c = data->whilem_c;
3065 data_fake.last_closep = data->last_closep;
3066 }
3067 else
3068 data_fake.last_closep = &fake;
58e23c8d
YO
3069
3070 data_fake.pos_delta = delta;
8aa23a47
YO
3071 next = regnext(scan);
3072 scan = NEXTOPER(scan);
3073 if (code != BRANCH)
c277df42 3074 scan = NEXTOPER(scan);
8aa23a47 3075 if (flags & SCF_DO_STCLASS) {
e755fd73 3076 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3077 data_fake.start_class = &this_class;
3078 f = SCF_DO_STCLASS_AND;
58e23c8d 3079 }
8aa23a47
YO
3080 if (flags & SCF_WHILEM_VISITED_POS)
3081 f |= SCF_WHILEM_VISITED_POS;
3082
3083 /* we suppose the run is continuous, last=next...*/
3084 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3085 next, &data_fake,
3086 stopparen, recursed, NULL, f,depth+1);
3087 if (min1 > minnext)
3088 min1 = minnext;
3089 if (max1 < minnext + deltanext)
3090 max1 = minnext + deltanext;
3091 if (deltanext == I32_MAX)
3092 is_inf = is_inf_internal = 1;
3093 scan = next;
3094 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3095 pars++;
3096 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3097 if ( stopmin > minnext)
3098 stopmin = min + min1;
3099 flags &= ~SCF_DO_SUBSTR;
3100 if (data)
3101 data->flags |= SCF_SEEN_ACCEPT;
3102 }
3103 if (data) {
3104 if (data_fake.flags & SF_HAS_EVAL)
3105 data->flags |= SF_HAS_EVAL;
3106 data->whilem_c = data_fake.whilem_c;
3dab1dad 3107 }
8aa23a47 3108 if (flags & SCF_DO_STCLASS)
3fffb88a 3109 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3110 }
3111 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3112 min1 = 0;
3113 if (flags & SCF_DO_SUBSTR) {
3114 data->pos_min += min1;
3115 data->pos_delta += max1 - min1;
3116 if (max1 != min1 || is_inf)
3117 data->longest = &(data->longest_float);
3118 }
3119 min += min1;
3120 delta += max1 - min1;
3121 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3122 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3123 if (min1) {
3124 cl_and(data->start_class, and_withp);
3125 flags &= ~SCF_DO_STCLASS;
653099ff 3126 }
8aa23a47
YO
3127 }
3128 else if (flags & SCF_DO_STCLASS_AND) {
3129 if (min1) {
3130 cl_and(data->start_class, &accum);
3131 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3132 }
8aa23a47
YO
3133 else {
3134 /* Switch to OR mode: cache the old value of
3135 * data->start_class */
3136 INIT_AND_WITHP;
3137 StructCopy(data->start_class, and_withp,
3138 struct regnode_charclass_class);
3139 flags &= ~SCF_DO_STCLASS_AND;
3140 StructCopy(&accum, data->start_class,
3141 struct regnode_charclass_class);
3142 flags |= SCF_DO_STCLASS_OR;
3143 data->start_class->flags |= ANYOF_EOS;
de0c8cb8 3144 }
8aa23a47 3145 }
a3621e74 3146
8aa23a47
YO
3147 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3148 /* demq.
a3621e74 3149
8aa23a47
YO
3150 Assuming this was/is a branch we are dealing with: 'scan' now
3151 points at the item that follows the branch sequence, whatever
3152 it is. We now start at the beginning of the sequence and look
3153 for subsequences of
a3621e74 3154
8aa23a47
YO
3155 BRANCH->EXACT=>x1
3156 BRANCH->EXACT=>x2
3157 tail
a3621e74 3158
8aa23a47 3159 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3160
486ec47a 3161 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3162 element into a trie and then add the subsequent branch exact
3163 strings to the trie.
a3621e74 3164
8aa23a47 3165 We have two cases
a3621e74 3166
3b753521 3167 1. patterns where the whole set of branches can be converted.
a3621e74 3168
8aa23a47 3169 2. patterns where only a subset can be converted.
a3621e74 3170
8aa23a47
YO
3171 In case 1 we can replace the whole set with a single regop
3172 for the trie. In case 2 we need to keep the start and end
3b753521 3173 branches so
a3621e74 3174
8aa23a47
YO
3175 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3176 becomes BRANCH TRIE; BRANCH X;
786e8c11 3177
8aa23a47
YO
3178 There is an additional case, that being where there is a
3179 common prefix, which gets split out into an EXACT like node
3180 preceding the TRIE node.
a3621e74 3181
8aa23a47
YO
3182 If x(1..n)==tail then we can do a simple trie, if not we make
3183 a "jump" trie, such that when we match the appropriate word
486ec47a 3184 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3185 a nested if into a case structure of sorts.
b515a41d 3186
8aa23a47 3187 */
686b73d4 3188
8aa23a47
YO
3189 int made=0;
3190 if (!re_trie_maxbuff) {
3191 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3192 if (!SvIOK(re_trie_maxbuff))
3193 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3194 }
3195 if ( SvIV(re_trie_maxbuff)>=0 ) {
3196 regnode *cur;
3197 regnode *first = (regnode *)NULL;
3198 regnode *last = (regnode *)NULL;
3199 regnode *tail = scan;
fab2782b 3200 U8 trietype = 0;
8aa23a47 3201 U32 count=0;
a3621e74
YO
3202
3203#ifdef DEBUGGING
8aa23a47 3204 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3205#endif
8aa23a47
YO
3206 /* var tail is used because there may be a TAIL
3207 regop in the way. Ie, the exacts will point to the
3208 thing following the TAIL, but the last branch will
3209 point at the TAIL. So we advance tail. If we
3210 have nested (?:) we may have to move through several
3211 tails.
3212 */
3213
3214 while ( OP( tail ) == TAIL ) {
3215 /* this is the TAIL generated by (?:) */
3216 tail = regnext( tail );
3217 }
a3621e74 3218
8aa23a47 3219
df826430 3220 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3221 regprop(RExC_rx, mysv, tail );
3222 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3223 (int)depth * 2 + 2, "",
3224 "Looking for TRIE'able sequences. Tail node is: ",
3225 SvPV_nolen_const( mysv )
3226 );
3227 });
3228
3229 /*
3230
fab2782b
YO
3231 Step through the branches
3232 cur represents each branch,
3233 noper is the first thing to be matched as part of that branch
3234 noper_next is the regnext() of that node.
3235
3236 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3237 via a "jump trie" but we also support building with NOJUMPTRIE,
3238 which restricts the trie logic to structures like /FOO|BAR/.
3239
3240 If noper is a trieable nodetype then the branch is a possible optimization
3241 target. If we are building under NOJUMPTRIE then we require that noper_next
3242 is the same as scan (our current position in the regex program).
3243
3244 Once we have two or more consecutive such branches we can create a
3245 trie of the EXACT's contents and stitch it in place into the program.
3246
3247 If the sequence represents all of the branches in the alternation we
3248 replace the entire thing with a single TRIE node.
3249
3250 Otherwise when it is a subsequence we need to stitch it in place and
3251 replace only the relevant branches. This means the first branch has
3252 to remain as it is used by the alternation logic, and its next pointer,
3253 and needs to be repointed at the item on the branch chain following
3254 the last branch we have optimized away.
3255
3256 This could be either a BRANCH, in which case the subsequence is internal,
3257 or it could be the item following the branch sequence in which case the
3258 subsequence is at the end (which does not necessarily mean the first node
3259 is the start of the alternation).
3260
3261 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3262
3263 optype | trietype
3264 ----------------+-----------
3265 NOTHING | NOTHING
3266 EXACT | EXACT
3267 EXACTFU | EXACTFU
3268 EXACTFU_SS | EXACTFU
3269 EXACTFU_TRICKYFOLD | EXACTFU
3270 EXACTFA | 0
3271
8aa23a47
YO
3272
3273 */
fab2782b
YO
3274#define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3275 ( EXACT == (X) ) ? EXACT : \
3276 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3277 0 )
8aa23a47
YO
3278
3279 /* dont use tail as the end marker for this traverse */
3280 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3281 regnode * const noper = NEXTOPER( cur );
fab2782b
YO
3282 U8 noper_type = OP( noper );
3283 U8 noper_trietype = TRIE_TYPE( noper_type );
b515a41d 3284#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3285 regnode * const noper_next = regnext( noper );
df826430
YO
3286 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3287 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
b515a41d
YO
3288#endif
3289
df826430 3290 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3291 regprop(RExC_rx, mysv, cur);
3292 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3293 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3294
3295 regprop(RExC_rx, mysv, noper);
3296 PerlIO_printf( Perl_debug_log, " -> %s",
3297 SvPV_nolen_const(mysv));
3298
3299 if ( noper_next ) {
3300 regprop(RExC_rx, mysv, noper_next );
3301 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3302 SvPV_nolen_const(mysv));
3303 }
df826430
YO
3304 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3305 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3306 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3307 );
8aa23a47 3308 });
fab2782b
YO
3309
3310 /* Is noper a trieable nodetype that can be merged with the
3311 * current trie (if there is one)? */
3312 if ( noper_trietype
3313 &&
3314 (
df826430
YO
3315 ( noper_trietype == NOTHING)
3316 || ( trietype == NOTHING )
a40630bf 3317 || ( trietype == noper_trietype )
fab2782b 3318 )
786e8c11 3319#ifdef NOJUMPTRIE
8aa23a47 3320 && noper_next == tail
786e8c11 3321#endif
8aa23a47
YO
3322 && count < U16_MAX)
3323 {
fab2782b
YO
3324 /* Handle mergable triable node
3325 * Either we are the first node in a new trieable sequence,
3326 * in which case we do some bookkeeping, otherwise we update
3327 * the end pointer. */
fab2782b 3328 if ( !first ) {
3b6759a6 3329 first = cur;
df826430
YO
3330 if ( noper_trietype == NOTHING ) {
3331#if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3332 regnode * const noper_next = regnext( noper );
3b6759a6 3333 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
df826430
YO
3334 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3335#endif
3336
190c1910 3337 if ( noper_next_trietype ) {
df826430 3338 trietype = noper_next_trietype;
190c1910
YO
3339 } else if (noper_next_type) {
3340 /* a NOTHING regop is 1 regop wide. We need at least two
3341 * for a trie so we can't merge this in */
3342 first = NULL;
3343 }
3344 } else {
3345 trietype = noper_trietype;
3b6759a6 3346 }
8aa23a47 3347 } else {
fab2782b
YO
3348 if ( trietype == NOTHING )
3349 trietype = noper_trietype;
8aa23a47
YO
3350 last = cur;
3351 }
df826430
YO
3352 if (first)
3353 count++;
fab2782b
YO
3354 } /* end handle mergable triable node */
3355 else {
3356 /* handle unmergable node -
3357 * noper may either be a triable node which can not be tried
3358 * together with the current trie, or a non triable node */
729aaeb5
YO
3359 if ( last ) {
3360 /* If last is set and trietype is not NOTHING then we have found
3361 * at least two triable branch sequences in a row of a similar
3362 * trietype so we can turn them into a trie. If/when we
3363 * allow NOTHING to start a trie sequence this condition will be
3364 * required, and it isn't expensive so we leave it in for now. */
e6351b37 3365 if ( trietype && trietype != NOTHING )
729aaeb5
YO
3366 make_trie( pRExC_state,
3367 startbranch, first, cur, tail, count,
3368 trietype, depth+1 );
fab2782b 3369 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
8aa23a47 3370 }
fab2782b 3371 if ( noper_trietype
786e8c11 3372#ifdef NOJUMPTRIE
8aa23a47 3373 && noper_next == tail
786e8c11 3374#endif
8aa23a47 3375 ){
fab2782b 3376 /* noper is triable, so we can start a new trie sequence */
8aa23a47
YO
3377 count = 1;
3378 first = cur;
fab2782b
YO
3379 trietype = noper_trietype;
3380 } else if (first) {
3381 /* if we already saw a first but the current node is not triable then we have
3382 * to reset the first information. */
8aa23a47
YO
3383 count = 0;
3384 first = NULL;
fab2782b 3385 trietype = 0;
8aa23a47 3386 }
fab2782b
YO
3387 } /* end handle unmergable node */
3388 } /* loop over branches */
df826430 3389 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3390 regprop(RExC_rx, mysv, cur);
3391 PerlIO_printf( Perl_debug_log,
3392 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3393 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3394
3395 });
e6351b37 3396 if ( last && trietype ) {
3b6759a6
YO
3397 if ( trietype != NOTHING ) {
3398 /* the last branch of the sequence was part of a trie,
3399 * so we have to construct it here outside of the loop
3400 */
3401 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
686b73d4 3402#ifdef TRIE_STUDY_OPT
3b6759a6
YO
3403 if ( ((made == MADE_EXACT_TRIE &&
3404 startbranch == first)
3405 || ( first_non_open == first )) &&
3406 depth==0 ) {
3407 flags |= SCF_TRIE_RESTUDY;
3408 if ( startbranch == first
3409 && scan == tail )
3410 {
3411 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3412 }
8aa23a47 3413 }
3dab1dad 3414#endif
3b6759a6
YO
3415 } else {
3416 /* at this point we know whatever we have is a NOTHING sequence/branch
3417 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3418 */
3419 if ( startbranch == first ) {
3420 regnode *opt;
3421 /* the entire thing is a NOTHING sequence, something like this:
3422 * (?:|) So we can turn it into a plain NOTHING op. */
3423 DEBUG_TRIE_COMPILE_r({
3424 regprop(RExC_rx, mysv, cur);
3425 PerlIO_printf( Perl_debug_log,
3426 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3427 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3428
3429 });
3430 OP(startbranch)= NOTHING;
3431 NEXT_OFF(startbranch)= tail - startbranch;
3432 for ( opt= startbranch + 1; opt < tail ; opt++ )
3433 OP(opt)= OPTIMIZED;
3434 }
3435 }
fab2782b
YO
3436 } /* end if ( last) */
3437 } /* TRIE_MAXBUF is non zero */
8aa23a47
YO
3438
3439 } /* do trie */
3440
653099ff 3441 }
8aa23a47
YO
3442 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3443 scan = NEXTOPER(NEXTOPER(scan));
3444 } else /* single branch is optimized. */
3445 scan = NEXTOPER(scan);
3446 continue;
3447 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3448 scan_frame *newframe = NULL;
3449 I32 paren;
3450 regnode *start;
3451 regnode *end;
3452
3453 if (OP(scan) != SUSPEND) {
3454 /* set the pointer */
3455 if (OP(scan) == GOSUB) {
3456 paren = ARG(scan);
3457 RExC_recurse[ARG2L(scan)] = scan;
3458 start = RExC_open_parens[paren-1];
3459 end = RExC_close_parens[paren-1];
3460 } else {
3461 paren = 0;
f8fc2ecf 3462 start = RExC_rxi->program + 1;
8aa23a47
YO
3463 end = RExC_opend;
3464 }
3465 if (!recursed) {
3466 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3467 SAVEFREEPV(recursed);
3468 }
3469 if (!PAREN_TEST(recursed,paren+1)) {
3470 PAREN_SET(recursed,paren+1);
3471 Newx(newframe,1,scan_frame);
3472 } else {
3473 if (flags & SCF_DO_SUBSTR) {
304ee84b 3474 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3475 data->longest = &(data->longest_float);
3476 }
3477 is_inf = is_inf_internal = 1;
3478 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3479 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3480 flags &= ~SCF_DO_STCLASS;
3481 }
3482 } else {
3483 Newx(newframe,1,scan_frame);
3484 paren = stopparen;
3485 start = scan+2;
3486 end = regnext(scan);
3487 }
3488 if (newframe) {
3489 assert(start);
3490 assert(end);
3491 SAVEFREEPV(newframe);
3492 newframe->next = regnext(scan);
3493 newframe->last = last;
3494 newframe->stop = stopparen;
3495 newframe->prev = frame;
3496
3497 frame = newframe;
3498 scan = start;
3499 stopparen = paren;
3500 last = end;
3501
3502 continue;
3503 }
3504 }
3505 else if (OP(scan) == EXACT) {
3506 I32 l = STR_LEN(scan);
3507 UV uc;
3508 if (UTF) {
3509 const U8 * const s = (U8*)STRING(scan);
4b88fb76 3510 uc = utf8_to_uvchr_buf(s, s + l, NULL);
8aa23a47 3511 l = utf8_length(s, s + l);
8aa23a47
YO
3512 } else {
3513 uc = *((U8*)STRING(scan));
3514 }
3515 min += l;
3516 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3517 /* The code below prefers earlier match for fixed
3518 offset, later match for variable offset. */
3519 if (data->last_end == -1) { /* Update the start info. */
3520 data->last_start_min = data->pos_min;
3521 data->last_start_max = is_inf
3522 ? I32_MAX : data->pos_min + data->pos_delta;
b515a41d 3523 }
8aa23a47
YO
3524 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3525 if (UTF)
3526 SvUTF8_on(data->last_found);
3527 {
3528 SV * const sv = data->last_found;
3529 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3530 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3531 if (mg && mg->mg_len >= 0)
3532 mg->mg_len += utf8_length((U8*)STRING(scan),
3533 (U8*)STRING(scan)+STR_LEN(scan));
b515a41d 3534 }
8aa23a47
YO
3535 data->last_end = data->pos_min + l;
3536 data->pos_min += l; /* As in the first entry. */
3537 data->flags &= ~SF_BEFORE_EOL;
3538 }
3539 if (flags & SCF_DO_STCLASS_AND) {
3540 /* Check whether it is compatible with what we know already! */
3541 int compat = 1;
3542
54251c2e 3543
486ec47a 3544 /* If compatible, we or it in below. It is compatible if is
54251c2e
KW
3545 * in the bitmp and either 1) its bit or its fold is set, or 2)
3546 * it's for a locale. Even if there isn't unicode semantics
3547 * here, at runtime there may be because of matching against a
3548 * utf8 string, so accept a possible false positive for
3549 * latin1-range folds */
8aa23a47 3550 if (uc >= 0x100 ||
b25a1036 3551 (!(data->start_class->flags & ANYOF_LOCALE)
8aa23a47 3552 && !ANYOF_BITMAP_TEST(data->start_class, uc)
538b546e 3553 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
54251c2e 3554 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
8aa23a47 3555 )
d18bf9dc 3556 {
8aa23a47 3557 compat = 0;
d18bf9dc 3558 }
8aa23a47
YO
3559 ANYOF_CLASS_ZERO(data->start_class);
3560 ANYOF_BITMAP_ZERO(data->start_class);
3561 if (compat)
3562 ANYOF_BITMAP_SET(data->start_class, uc);
d18bf9dc
KW
3563 else if (uc >= 0x100) {
3564 int i;
3565
3566 /* Some Unicode code points fold to the Latin1 range; as
3567 * XXX temporary code, instead of figuring out if this is
3568 * one, just assume it is and set all the start class bits
3569 * that could be some such above 255 code point's fold
3570 * which will generate fals positives. As the code
3571 * elsewhere that does compute the fold settles down, it
3572 * can be extracted out and re-used here */
3573 for (i = 0; i < 256; i++){
94dc5c2d 3574 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
d18bf9dc
KW
3575 ANYOF_BITMAP_SET(data->start_class, i);
3576 }