This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c: Add #undef
[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);
dd58aee1 759 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
ae6bdd37 760 |ANYOF_NON_UTF8_LATIN1_ALL;
3fffb88a
KW
761
762 /* If any portion of the regex is to operate under locale rules,
763 * initialization includes it. The reason this isn't done for all regexes
764 * is that the optimizer was written under the assumption that locale was
765 * all-or-nothing. Given the complexity and lack of documentation in the
766 * optimizer, and that there are inadequate test cases for locale, so many
767 * parts of it may not work properly, it is safest to avoid locale unless
768 * necessary. */
769 if (RExC_contains_locale) {
9d7a1e63 770 ANYOF_CLASS_SETALL(cl); /* /l uses class */
ae6bdd37 771 cl->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3fffb88a 772 }
9d7a1e63
KW
773 else {
774 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
775 }
653099ff
GS
776}
777
778/* Can match anything (initialization) */
779STATIC int
5f66b61c 780S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
781{
782 int value;
783
7918f24d
NC
784 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
785
aaa51d5e 786 for (value = 0; value <= ANYOF_MAX; value += 2)
653099ff
GS
787 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
788 return 1;
1aa99e6b
IH
789 if (!(cl->flags & ANYOF_UNICODE_ALL))
790 return 0;
10edeb5d 791 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 792 return 0;
653099ff
GS
793 return 1;
794}
795
796/* Can match anything (initialization) */
797STATIC void
e755fd73 798S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 799{
7918f24d
NC
800 PERL_ARGS_ASSERT_CL_INIT;
801
8ecf7187 802 Zero(cl, 1, struct regnode_charclass_class);
653099ff 803 cl->type = ANYOF;
3fffb88a 804 cl_anything(pRExC_state, cl);
1411dba4 805 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
806}
807
1051e1c4
KW
808/* These two functions currently do the exact same thing */
809#define cl_init_zero S_cl_init
653099ff 810
dd58aee1
KW
811/* 'AND' a given class with another one. Can create false positives. 'cl'
812 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
813 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 814STATIC void
5f66b61c 815S_cl_and(struct regnode_charclass_class *cl,
a28509cc 816 const struct regnode_charclass_class *and_with)
653099ff 817{
7918f24d 818 PERL_ARGS_ASSERT_CL_AND;
40d049e4
YO
819
820 assert(and_with->type == ANYOF);
1e6ade67 821
c6b76537 822 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
823 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
824 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 825 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
826 && !(and_with->flags & ANYOF_LOC_FOLD)
827 && !(cl->flags & ANYOF_LOC_FOLD)) {
653099ff
GS
828 int i;
829
830 if (and_with->flags & ANYOF_INVERT)
831 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
832 cl->bitmap[i] &= ~and_with->bitmap[i];
833 else
834 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
835 cl->bitmap[i] &= and_with->bitmap[i];
836 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 837
c6b76537 838 if (and_with->flags & ANYOF_INVERT) {
8951c461 839
c6b76537
KW
840 /* Here, the and'ed node is inverted. Get the AND of the flags that
841 * aren't affected by the inversion. Those that are affected are
842 * handled individually below */
843 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
844 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
845 cl->flags |= affected_flags;
846
847 /* We currently don't know how to deal with things that aren't in the
848 * bitmap, but we know that the intersection is no greater than what
849 * is already in cl, so let there be false positives that get sorted
850 * out after the synthetic start class succeeds, and the node is
851 * matched for real. */
852
853 /* The inversion of these two flags indicate that the resulting
854 * intersection doesn't have them */
855 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
856 cl->flags &= ~ANYOF_UNICODE_ALL;
857 }
c6b76537
KW
858 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
859 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 860 }
1aa99e6b 861 }
c6b76537 862 else { /* and'd node is not inverted */
3ad98780
KW
863 U8 outside_bitmap_but_not_utf8; /* Temp variable */
864
137165a6 865 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
866
867 /* Here 'and_with' doesn't match anything outside the bitmap
868 * (except possibly ANYOF_UNICODE_ALL), which means the
869 * intersection can't either, except for ANYOF_UNICODE_ALL, in
870 * which case we don't know what the intersection is, but it's no
871 * greater than what cl already has, so can just leave it alone,
872 * with possible false positives */
873 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
874 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 875 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 876 }
137165a6 877 }
c6b76537
KW
878 else if (! ANYOF_NONBITMAP(cl)) {
879
880 /* Here, 'and_with' does match something outside the bitmap, and cl
881 * doesn't have a list of things to match outside the bitmap. If
882 * cl can match all code points above 255, the intersection will
3ad98780
KW
883 * be those above-255 code points that 'and_with' matches. If cl
884 * can't match all Unicode code points, it means that it can't
885 * match anything outside the bitmap (since the 'if' that got us
886 * into this block tested for that), so we leave the bitmap empty.
887 */
c6b76537
KW
888 if (cl->flags & ANYOF_UNICODE_ALL) {
889 ARG_SET(cl, ARG(and_with));
3ad98780
KW
890
891 /* and_with's ARG may match things that don't require UTF8.
892 * And now cl's will too, in spite of this being an 'and'. See
893 * the comments below about the kludge */
894 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
895 }
896 }
897 else {
898 /* Here, both 'and_with' and cl match something outside the
899 * bitmap. Currently we do not do the intersection, so just match
900 * whatever cl had at the beginning. */
901 }
902
903
3ad98780
KW
904 /* Take the intersection of the two sets of flags. However, the
905 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
906 * kludge around the fact that this flag is not treated like the others
907 * which are initialized in cl_anything(). The way the optimizer works
908 * is that the synthetic start class (SSC) is initialized to match
909 * anything, and then the first time a real node is encountered, its
910 * values are AND'd with the SSC's with the result being the values of
911 * the real node. However, there are paths through the optimizer where
912 * the AND never gets called, so those initialized bits are set
913 * inappropriately, which is not usually a big deal, as they just cause
914 * false positives in the SSC, which will just mean a probably
915 * imperceptible slow down in execution. However this bit has a
916 * higher false positive consequence in that it can cause utf8.pm,
917 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
918 * bigger slowdown and also causes significant extra memory to be used.
919 * In order to prevent this, the code now takes a different tack. The
920 * bit isn't set unless some part of the regular expression needs it,
921 * but once set it won't get cleared. This means that these extra
922 * modules won't get loaded unless there was some path through the
923 * pattern that would have required them anyway, and so any false
924 * positives that occur by not ANDing them out when they could be
925 * aren't as severe as they would be if we treated this bit like all
926 * the others */
927 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
928 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 929 cl->flags &= and_with->flags;
3ad98780 930 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 931 }
653099ff
GS
932}
933
dd58aee1
KW
934/* 'OR' a given class with another one. Can create false positives. 'cl'
935 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
936 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 937STATIC void
3fffb88a 938S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 939{
7918f24d
NC
940 PERL_ARGS_ASSERT_CL_OR;
941
653099ff 942 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
943
944 /* Here, the or'd node is to be inverted. This means we take the
945 * complement of everything not in the bitmap, but currently we don't
946 * know what that is, so give up and match anything */
947 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 948 cl_anything(pRExC_state, cl);
c6b76537 949 }
653099ff
GS
950 /* We do not use
951 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
952 * <= (B1 | !B2) | (CL1 | !CL2)
953 * which is wasteful if CL2 is small, but we ignore CL2:
954 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
955 * XXXX Can we handle case-fold? Unclear:
956 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
957 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
958 */
c6b76537 959 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
960 && !(or_with->flags & ANYOF_LOC_FOLD)
961 && !(cl->flags & ANYOF_LOC_FOLD) ) {
653099ff
GS
962 int i;
963
964 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
965 cl->bitmap[i] |= ~or_with->bitmap[i];
966 } /* XXXX: logic is complicated otherwise */
967 else {
3fffb88a 968 cl_anything(pRExC_state, cl);
653099ff 969 }
c6b76537
KW
970
971 /* And, we can just take the union of the flags that aren't affected
972 * by the inversion */
973 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
974
975 /* For the remaining flags:
976 ANYOF_UNICODE_ALL and inverted means to not match anything above
977 255, which means that the union with cl should just be
978 what cl has in it, so can ignore this flag
979 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
980 is 127-255 to match them, but then invert that, so the
981 union with cl should just be what cl has in it, so can
982 ignore this flag
983 */
984 } else { /* 'or_with' is not inverted */
653099ff
GS
985 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
986 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
987 && (!(or_with->flags & ANYOF_LOC_FOLD)
988 || (cl->flags & ANYOF_LOC_FOLD)) ) {
653099ff
GS
989 int i;
990
991 /* OR char bitmap and class bitmap separately */
992 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
993 cl->bitmap[i] |= or_with->bitmap[i];
9cf78fa0 994 ANYOF_CLASS_OR(or_with, cl);
653099ff
GS
995 }
996 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 997 cl_anything(pRExC_state, cl);
653099ff 998 }
9826f543 999
c6b76537
KW
1000 if (ANYOF_NONBITMAP(or_with)) {
1001
1002 /* Use the added node's outside-the-bit-map match if there isn't a
1003 * conflict. If there is a conflict (both nodes match something
1004 * outside the bitmap, but what they match outside is not the same
1005 * pointer, and hence not easily compared until XXX we extend
1006 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
1007 * match everything outside the bitmap. If that stuff is all above
1008 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
1009 if (! ANYOF_NONBITMAP(cl)) {
1010 ARG_SET(cl, ARG(or_with));
1011 }
1012 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
1013
1014 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1015 cl_anything(pRExC_state, cl);
1016 }
1017 else {
1018 cl->flags |= ANYOF_UNICODE_ALL;
1019 }
c6b76537 1020 }
4c34a693 1021 }
0b9668ee
KW
1022
1023 /* Take the union */
1024 cl->flags |= or_with->flags;
1aa99e6b 1025 }
653099ff
GS
1026}
1027
a3621e74
YO
1028#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1029#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1030#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1031#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1032
3dab1dad
YO
1033
1034#ifdef DEBUGGING
07be1b83 1035/*
2b8b4781
NC
1036 dump_trie(trie,widecharmap,revcharmap)
1037 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1038 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1039
1040 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1041 The _interim_ variants are used for debugging the interim
1042 tables that are used to generate the final compressed
1043 representation which is what dump_trie expects.
1044
486ec47a 1045 Part of the reason for their existence is to provide a form
3dab1dad 1046 of documentation as to how the different representations function.
07be1b83
YO
1047
1048*/
3dab1dad
YO
1049
1050/*
3dab1dad
YO
1051 Dumps the final compressed table form of the trie to Perl_debug_log.
1052 Used for debugging make_trie().
1053*/
b9a59e08 1054
3dab1dad 1055STATIC void
2b8b4781
NC
1056S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1057 AV *revcharmap, U32 depth)
3dab1dad
YO
1058{
1059 U32 state;
ab3bbdeb 1060 SV *sv=sv_newmortal();
55eed653 1061 int colwidth= widecharmap ? 6 : 4;
2e64971a 1062 U16 word;
3dab1dad
YO
1063 GET_RE_DEBUG_FLAGS_DECL;
1064
7918f24d 1065 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1066
3dab1dad
YO
1067 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1068 (int)depth * 2 + 2,"",
1069 "Match","Base","Ofs" );
1070
1071 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1072 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1073 if ( tmp ) {
ab3bbdeb
YO
1074 PerlIO_printf( Perl_debug_log, "%*s",
1075 colwidth,
ddc5bc0f 1076 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1077 PL_colors[0], PL_colors[1],
1078 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1079 PERL_PV_ESCAPE_FIRSTCHAR
1080 )
1081 );
3dab1dad
YO
1082 }
1083 }
1084 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1085 (int)depth * 2 + 2,"");
1086
1087 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1088 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1089 PerlIO_printf( Perl_debug_log, "\n");
1090
1e2e3d02 1091 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1092 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1093
1094 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1095
1096 if ( trie->states[ state ].wordnum ) {
1097 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1098 } else {
1099 PerlIO_printf( Perl_debug_log, "%6s", "" );
1100 }
1101
1102 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1103
1104 if ( base ) {
1105 U32 ofs = 0;
1106
1107 while( ( base + ofs < trie->uniquecharcount ) ||
1108 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1109 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1110 ofs++;
1111
1112 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1113
1114 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1115 if ( ( base + ofs >= trie->uniquecharcount ) &&
1116 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1117 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1118 {
ab3bbdeb
YO
1119 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1120 colwidth,
3dab1dad
YO
1121 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1122 } else {
ab3bbdeb 1123 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1124 }
1125 }
1126
1127 PerlIO_printf( Perl_debug_log, "]");
1128
1129 }
1130 PerlIO_printf( Perl_debug_log, "\n" );
1131 }
2e64971a
DM
1132 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1133 for (word=1; word <= trie->wordcount; word++) {
1134 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1135 (int)word, (int)(trie->wordinfo[word].prev),
1136 (int)(trie->wordinfo[word].len));
1137 }
1138 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1139}
1140/*
3dab1dad
YO
1141 Dumps a fully constructed but uncompressed trie in list form.
1142 List tries normally only are used for construction when the number of
1143 possible chars (trie->uniquecharcount) is very high.
1144 Used for debugging make_trie().
1145*/
1146STATIC void
55eed653 1147S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1148 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1149 U32 depth)
3dab1dad
YO
1150{
1151 U32 state;
ab3bbdeb 1152 SV *sv=sv_newmortal();
55eed653 1153 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1154 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1155
1156 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1157
3dab1dad 1158 /* print out the table precompression. */
ab3bbdeb
YO
1159 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1160 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1161 "------:-----+-----------------\n" );
3dab1dad
YO
1162
1163 for( state=1 ; state < next_alloc ; state ++ ) {
1164 U16 charid;
1165
ab3bbdeb 1166 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1167 (int)depth * 2 + 2,"", (UV)state );
1168 if ( ! trie->states[ state ].wordnum ) {
1169 PerlIO_printf( Perl_debug_log, "%5s| ","");
1170 } else {
1171 PerlIO_printf( Perl_debug_log, "W%4x| ",
1172 trie->states[ state ].wordnum
1173 );
1174 }
1175 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1176 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1177 if ( tmp ) {
1178 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1179 colwidth,
ddc5bc0f 1180 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1181 PL_colors[0], PL_colors[1],
1182 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1183 PERL_PV_ESCAPE_FIRSTCHAR
1184 ) ,
1e2e3d02
YO
1185 TRIE_LIST_ITEM(state,charid).forid,
1186 (UV)TRIE_LIST_ITEM(state,charid).newstate
1187 );
1188 if (!(charid % 10))
664e119d
RGS
1189 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1190 (int)((depth * 2) + 14), "");
1e2e3d02 1191 }
ab3bbdeb
YO
1192 }
1193 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1194 }
1195}
1196
1197/*
3dab1dad
YO
1198 Dumps a fully constructed but uncompressed trie in table form.
1199 This is the normal DFA style state transition table, with a few
1200 twists to facilitate compression later.
1201 Used for debugging make_trie().
1202*/
1203STATIC void
55eed653 1204S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1205 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1206 U32 depth)
3dab1dad
YO
1207{
1208 U32 state;
1209 U16 charid;
ab3bbdeb 1210 SV *sv=sv_newmortal();
55eed653 1211 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1212 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1213
1214 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1215
1216 /*
1217 print out the table precompression so that we can do a visual check
1218 that they are identical.
1219 */
1220
1221 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1222
1223 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1224 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1225 if ( tmp ) {
ab3bbdeb
YO
1226 PerlIO_printf( Perl_debug_log, "%*s",
1227 colwidth,
ddc5bc0f 1228 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1229 PL_colors[0], PL_colors[1],
1230 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1231 PERL_PV_ESCAPE_FIRSTCHAR
1232 )
1233 );
3dab1dad
YO
1234 }
1235 }
1236
1237 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1238
1239 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1240 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1241 }
1242
1243 PerlIO_printf( Perl_debug_log, "\n" );
1244
1245 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1246
1247 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1248 (int)depth * 2 + 2,"",
1249 (UV)TRIE_NODENUM( state ) );
1250
1251 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1252 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1253 if (v)
1254 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1255 else
1256 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1257 }
1258 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1259 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1260 } else {
1261 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1262 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1263 }
1264 }
07be1b83 1265}
3dab1dad
YO
1266
1267#endif
1268
2e64971a 1269
786e8c11
YO
1270/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1271 startbranch: the first branch in the whole branch sequence
1272 first : start branch of sequence of branch-exact nodes.
1273 May be the same as startbranch
1274 last : Thing following the last branch.
1275 May be the same as tail.
1276 tail : item following the branch sequence
1277 count : words in the sequence
1278 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1279 depth : indent depth
3dab1dad 1280
786e8c11 1281Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1282
786e8c11
YO
1283A trie is an N'ary tree where the branches are determined by digital
1284decomposition of the key. IE, at the root node you look up the 1st character and
1285follow that branch repeat until you find the end of the branches. Nodes can be
1286marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1287
786e8c11 1288 /he|she|his|hers/
72f13be8 1289
786e8c11
YO
1290would convert into the following structure. Numbers represent states, letters
1291following numbers represent valid transitions on the letter from that state, if
1292the number is in square brackets it represents an accepting state, otherwise it
1293will be in parenthesis.
07be1b83 1294
786e8c11
YO
1295 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1296 | |
1297 | (2)
1298 | |
1299 (1) +-i->(6)-+-s->[7]
1300 |
1301 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1302
786e8c11
YO
1303 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1304
1305This shows that when matching against the string 'hers' we will begin at state 1
1306read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1307then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1308is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1309single traverse. We store a mapping from accepting to state to which word was
1310matched, and then when we have multiple possibilities we try to complete the
1311rest of the regex in the order in which they occured in the alternation.
1312
1313The only prior NFA like behaviour that would be changed by the TRIE support is
1314the silent ignoring of duplicate alternations which are of the form:
1315
1316 / (DUPE|DUPE) X? (?{ ... }) Y /x
1317
4b714af6 1318Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1319and without the optimisation. With the optimisations dupes will be silently
486ec47a 1320ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1321the following demonstrates:
1322
1323 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1324
1325which prints out 'word' three times, but
1326
1327 'words'=~/(word|word|word)(?{ print $1 })S/
1328
1329which doesnt print it out at all. This is due to other optimisations kicking in.
1330
1331Example of what happens on a structural level:
1332
486ec47a 1333The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1334
1335 1: CURLYM[1] {1,32767}(18)
1336 5: BRANCH(8)
1337 6: EXACT <ac>(16)
1338 8: BRANCH(11)
1339 9: EXACT <ad>(16)
1340 11: BRANCH(14)
1341 12: EXACT <ab>(16)
1342 16: SUCCEED(0)
1343 17: NOTHING(18)
1344 18: END(0)
1345
1346This would be optimizable with startbranch=5, first=5, last=16, tail=16
1347and should turn into:
1348
1349 1: CURLYM[1] {1,32767}(18)
1350 5: TRIE(16)
1351 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1352 <ac>
1353 <ad>
1354 <ab>
1355 16: SUCCEED(0)
1356 17: NOTHING(18)
1357 18: END(0)
1358
1359Cases where tail != last would be like /(?foo|bar)baz/:
1360
1361 1: BRANCH(4)
1362 2: EXACT <foo>(8)
1363 4: BRANCH(7)
1364 5: EXACT <bar>(8)
1365 7: TAIL(8)
1366 8: EXACT <baz>(10)
1367 10: END(0)
1368
1369which would be optimizable with startbranch=1, first=1, last=7, tail=8
1370and would end up looking like:
1371
1372 1: TRIE(8)
1373 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1374 <foo>
1375 <bar>
1376 7: TAIL(8)
1377 8: EXACT <baz>(10)
1378 10: END(0)
1379
1380 d = uvuni_to_utf8_flags(d, uv, 0);
1381
1382is the recommended Unicode-aware way of saying
1383
1384 *(d++) = uv;
1385*/
1386
fab2782b 1387#define TRIE_STORE_REVCHAR(val) \
786e8c11 1388 STMT_START { \
73031816 1389 if (UTF) { \
fab2782b 1390 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1391 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
fab2782b 1392 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
73031816
NC
1393 SvCUR_set(zlopp, kapow - flrbbbbb); \
1394 SvPOK_on(zlopp); \
1395 SvUTF8_on(zlopp); \
1396 av_push(revcharmap, zlopp); \
1397 } else { \
fab2782b 1398 char ooooff = (char)val; \
73031816
NC
1399 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1400 } \
1401 } STMT_END
786e8c11 1402
fab2782b
YO
1403#define TRIE_READ_CHAR STMT_START { \
1404 wordlen++; \
1405 if ( UTF ) { \
1406 /* if it is UTF then it is either already folded, or does not need folding */ \
1407 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1408 } \
1409 else if (folder == PL_fold_latin1) { \
1410 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1411 if ( foldlen > 0 ) { \
1412 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1413 foldlen -= len; \
1414 scan += len; \
1415 len = 0; \
1416 } else { \
1417 len = 1; \
1418 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1419 skiplen = UNISKIP(uvc); \
1420 foldlen -= skiplen; \
1421 scan = foldbuf + skiplen; \
1422 } \
1423 } else { \
1424 /* raw data, will be folded later if needed */ \
1425 uvc = (U32)*uc; \
1426 len = 1; \
1427 } \
786e8c11
YO
1428} STMT_END
1429
1430
1431
1432#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1433 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1434 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1435 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1436 } \
1437 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1438 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1439 TRIE_LIST_CUR( state )++; \
1440} STMT_END
07be1b83 1441
786e8c11
YO
1442#define TRIE_LIST_NEW(state) STMT_START { \
1443 Newxz( trie->states[ state ].trans.list, \
1444 4, reg_trie_trans_le ); \
1445 TRIE_LIST_CUR( state ) = 1; \
1446 TRIE_LIST_LEN( state ) = 4; \
1447} STMT_END
07be1b83 1448
786e8c11
YO
1449#define TRIE_HANDLE_WORD(state) STMT_START { \
1450 U16 dupe= trie->states[ state ].wordnum; \
1451 regnode * const noper_next = regnext( noper ); \
1452 \
786e8c11
YO
1453 DEBUG_r({ \
1454 /* store the word for dumping */ \
1455 SV* tmp; \
1456 if (OP(noper) != NOTHING) \
740cce10 1457 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1458 else \
740cce10 1459 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1460 av_push( trie_words, tmp ); \
786e8c11
YO
1461 }); \
1462 \
1463 curword++; \
2e64971a
DM
1464 trie->wordinfo[curword].prev = 0; \
1465 trie->wordinfo[curword].len = wordlen; \
1466 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1467 \
1468 if ( noper_next < tail ) { \
1469 if (!trie->jump) \
c944940b 1470 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1471 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1472 if (!jumper) \
1473 jumper = noper_next; \
1474 if (!nextbranch) \
1475 nextbranch= regnext(cur); \
1476 } \
1477 \
1478 if ( dupe ) { \
2e64971a
DM
1479 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1480 /* chain, so that when the bits of chain are later */\
1481 /* linked together, the dups appear in the chain */\
1482 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1483 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1484 } else { \
1485 /* we haven't inserted this word yet. */ \
1486 trie->states[ state ].wordnum = curword; \
1487 } \
1488} STMT_END
07be1b83 1489
3dab1dad 1490
786e8c11
YO
1491#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1492 ( ( base + charid >= ucharcount \
1493 && base + charid < ubound \
1494 && state == trie->trans[ base - ucharcount + charid ].check \
1495 && trie->trans[ base - ucharcount + charid ].next ) \
1496 ? trie->trans[ base - ucharcount + charid ].next \
1497 : ( state==1 ? special : 0 ) \
1498 )
3dab1dad 1499
786e8c11
YO
1500#define MADE_TRIE 1
1501#define MADE_JUMP_TRIE 2
1502#define MADE_EXACT_TRIE 4
3dab1dad 1503
a3621e74 1504STATIC I32
786e8c11 1505S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1506{
27da23d5 1507 dVAR;
a3621e74
YO
1508 /* first pass, loop through and scan words */
1509 reg_trie_data *trie;
55eed653 1510 HV *widecharmap = NULL;
2b8b4781 1511 AV *revcharmap = newAV();
a3621e74 1512 regnode *cur;
9f7f3913 1513 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1514 STRLEN len = 0;
1515 UV uvc = 0;
1516 U16 curword = 0;
1517 U32 next_alloc = 0;
786e8c11
YO
1518 regnode *jumper = NULL;
1519 regnode *nextbranch = NULL;
7f69552c 1520 regnode *convert = NULL;
2e64971a 1521 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1522 /* we just use folder as a flag in utf8 */
1e696034 1523 const U8 * folder = NULL;
a3621e74 1524
2b8b4781
NC
1525#ifdef DEBUGGING
1526 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1527 AV *trie_words = NULL;
1528 /* along with revcharmap, this only used during construction but both are
1529 * useful during debugging so we store them in the struct when debugging.
8e11feef 1530 */
2b8b4781
NC
1531#else
1532 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1533 STRLEN trie_charcount=0;
3dab1dad 1534#endif
2b8b4781 1535 SV *re_trie_maxbuff;
a3621e74 1536 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1537
1538 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1539#ifndef DEBUGGING
1540 PERL_UNUSED_ARG(depth);
1541#endif
a3621e74 1542
1e696034 1543 switch (flags) {
79a81a6e 1544 case EXACT: break;
2f7f8cb1 1545 case EXACTFA:
fab2782b
YO
1546 case EXACTFU_SS:
1547 case EXACTFU_TRICKYFOLD:
1e696034
KW
1548 case EXACTFU: folder = PL_fold_latin1; break;
1549 case EXACTF: folder = PL_fold; break;
1550 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1551 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1552 }
1553
c944940b 1554 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1555 trie->refcount = 1;
3dab1dad 1556 trie->startstate = 1;
786e8c11 1557 trie->wordcount = word_count;
f8fc2ecf 1558 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1559 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1560 if (flags == EXACT)
c944940b 1561 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1562 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1563 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1564
a3621e74 1565 DEBUG_r({
2b8b4781 1566 trie_words = newAV();
a3621e74 1567 });
a3621e74 1568
0111c4fd 1569 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1570 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1571 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1572 }
df826430 1573 DEBUG_TRIE_COMPILE_r({
3dab1dad 1574 PerlIO_printf( Perl_debug_log,
786e8c11 1575 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1576 (int)depth * 2 + 2, "",
1577 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1578 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1579 (int)depth);
3dab1dad 1580 });
7f69552c
YO
1581
1582 /* Find the node we are going to overwrite */
1583 if ( first == startbranch && OP( last ) != BRANCH ) {
1584 /* whole branch chain */
1585 convert = first;
1586 } else {
1587 /* branch sub-chain */
1588 convert = NEXTOPER( first );
1589 }
1590
a3621e74
YO
1591 /* -- First loop and Setup --
1592
1593 We first traverse the branches and scan each word to determine if it
1594 contains widechars, and how many unique chars there are, this is
1595 important as we have to build a table with at least as many columns as we
1596 have unique chars.
1597
1598 We use an array of integers to represent the character codes 0..255
38a44b82 1599 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1600 native representation of the character value as the key and IV's for the
1601 coded index.
1602
1603 *TODO* If we keep track of how many times each character is used we can
1604 remap the columns so that the table compression later on is more
3b753521 1605 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1606 middle and the least common are on the outside. IMO this would be better
1607 than a most to least common mapping as theres a decent chance the most
1608 common letter will share a node with the least common, meaning the node
486ec47a 1609 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1610 case is when we have the least common nodes twice.
1611
1612 */
1613
a3621e74 1614 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1615 regnode *noper = NEXTOPER( cur );
e1ec3a88 1616 const U8 *uc = (U8*)STRING( noper );
df826430 1617 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1618 STRLEN foldlen = 0;
1619 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1620 STRLEN skiplen = 0;
2af232bd 1621 const U8 *scan = (U8*)NULL;
07be1b83 1622 U32 wordlen = 0; /* required init */
02daf0ab
YO
1623 STRLEN chars = 0;
1624 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1625
3dab1dad 1626 if (OP(noper) == NOTHING) {
df826430
YO
1627 regnode *noper_next= regnext(noper);
1628 if (noper_next != tail && OP(noper_next) == flags) {
1629 noper = noper_next;
1630 uc= (U8*)STRING(noper);
1631 e= uc + STR_LEN(noper);
1632 trie->minlen= STR_LEN(noper);
1633 } else {
1634 trie->minlen= 0;
1635 continue;
1636 }
3dab1dad 1637 }
df826430 1638
fab2782b 1639 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1640 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1641 regardless of encoding */
fab2782b
YO
1642 if (OP( noper ) == EXACTFU_SS) {
1643 /* false positives are ok, so just set this */
1644 TRIE_BITMAP_SET(trie,0xDF);
1645 }
1646 }
a3621e74 1647 for ( ; uc < e ; uc += len ) {
3dab1dad 1648 TRIE_CHARCOUNT(trie)++;
a3621e74 1649 TRIE_READ_CHAR;
3dab1dad 1650 chars++;
a3621e74 1651 if ( uvc < 256 ) {
fab2782b
YO
1652 if ( folder ) {
1653 U8 folded= folder[ (U8) uvc ];
1654 if ( !trie->charmap[ folded ] ) {
1655 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1656 TRIE_STORE_REVCHAR( folded );
1657 }
1658 }
a3621e74
YO
1659 if ( !trie->charmap[ uvc ] ) {
1660 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1661 TRIE_STORE_REVCHAR( uvc );
a3621e74 1662 }
02daf0ab 1663 if ( set_bit ) {
62012aee
KW
1664 /* store the codepoint in the bitmap, and its folded
1665 * equivalent. */
fab2782b 1666 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1667
1668 /* store the folded codepoint */
fab2782b 1669 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1670
1671 if ( !UTF ) {
1672 /* store first byte of utf8 representation of
acdf4139
KW
1673 variant codepoints */
1674 if (! UNI_IS_INVARIANT(uvc)) {
1675 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1676 }
1677 }
02daf0ab
YO
1678 set_bit = 0; /* We've done our bit :-) */
1679 }
a3621e74
YO
1680 } else {
1681 SV** svpp;
55eed653
NC
1682 if ( !widecharmap )
1683 widecharmap = newHV();
a3621e74 1684
55eed653 1685 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1686
1687 if ( !svpp )
e4584336 1688 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1689
1690 if ( !SvTRUE( *svpp ) ) {
1691 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1692 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1693 }
1694 }
1695 }
3dab1dad 1696 if( cur == first ) {
fab2782b
YO
1697 trie->minlen = chars;
1698 trie->maxlen = chars;
3dab1dad 1699 } else if (chars < trie->minlen) {
fab2782b 1700 trie->minlen = chars;
3dab1dad 1701 } else if (chars > trie->maxlen) {
fab2782b
YO
1702 trie->maxlen = chars;
1703 }
1704 if (OP( noper ) == EXACTFU_SS) {
1705 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1706 if (trie->minlen > 1)
1707 trie->minlen= 1;
1708 }
1709 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1710 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1711 * - We assume that any such sequence might match a 2 byte string */
1712 if (trie->minlen > 2 )
1713 trie->minlen= 2;
3dab1dad
YO
1714 }
1715
a3621e74
YO
1716 } /* end first pass */
1717 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1718 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1719 (int)depth * 2 + 2,"",
55eed653 1720 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1721 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1722 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1723 );
a3621e74
YO
1724
1725 /*
1726 We now know what we are dealing with in terms of unique chars and
1727 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1728 representation using a flat table will take. If it's over a reasonable
1729 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1730 conservative but potentially much slower representation using an array
1731 of lists.
1732
1733 At the end we convert both representations into the same compressed
1734 form that will be used in regexec.c for matching with. The latter
1735 is a form that cannot be used to construct with but has memory
1736 properties similar to the list form and access properties similar
1737 to the table form making it both suitable for fast searches and
1738 small enough that its feasable to store for the duration of a program.
1739
1740 See the comment in the code where the compressed table is produced
1741 inplace from the flat tabe representation for an explanation of how
1742 the compression works.
1743
1744 */
1745
1746
2e64971a
DM
1747 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1748 prev_states[1] = 0;
1749
3dab1dad 1750 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1751 /*
1752 Second Pass -- Array Of Lists Representation
1753
1754 Each state will be represented by a list of charid:state records
1755 (reg_trie_trans_le) the first such element holds the CUR and LEN
1756 points of the allocated array. (See defines above).
1757
1758 We build the initial structure using the lists, and then convert
1759 it into the compressed table form which allows faster lookups
1760 (but cant be modified once converted).
a3621e74
YO
1761 */
1762
a3621e74
YO
1763 STRLEN transcount = 1;
1764
1e2e3d02
YO
1765 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1766 "%*sCompiling trie using list compiler\n",
1767 (int)depth * 2 + 2, ""));
686b73d4 1768
c944940b
JH
1769 trie->states = (reg_trie_state *)
1770 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1771 sizeof(reg_trie_state) );
a3621e74
YO
1772 TRIE_LIST_NEW(1);
1773 next_alloc = 2;
1774
1775 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1776
df826430 1777 regnode *noper = NEXTOPER( cur );
c445ea15 1778 U8 *uc = (U8*)STRING( noper );
df826430 1779 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1780 U32 state = 1; /* required init */
1781 U16 charid = 0; /* sanity init */
1782 U8 *scan = (U8*)NULL; /* sanity init */
1783 STRLEN foldlen = 0; /* required init */
07be1b83 1784 U32 wordlen = 0; /* required init */
c445ea15 1785 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1786 STRLEN skiplen = 0;
c445ea15 1787
df826430
YO
1788 if (OP(noper) == NOTHING) {
1789 regnode *noper_next= regnext(noper);
1790 if (noper_next != tail && OP(noper_next) == flags) {
1791 noper = noper_next;
1792 uc= (U8*)STRING(noper);
1793 e= uc + STR_LEN(noper);
1794 }
1795 }
1796
3dab1dad 1797 if (OP(noper) != NOTHING) {
786e8c11 1798 for ( ; uc < e ; uc += len ) {
c445ea15 1799
786e8c11 1800 TRIE_READ_CHAR;
c445ea15 1801
786e8c11
YO
1802 if ( uvc < 256 ) {
1803 charid = trie->charmap[ uvc ];
c445ea15 1804 } else {
55eed653 1805 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1806 if ( !svpp ) {
1807 charid = 0;
1808 } else {
1809 charid=(U16)SvIV( *svpp );
1810 }
c445ea15 1811 }
786e8c11
YO
1812 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1813 if ( charid ) {
a3621e74 1814
786e8c11
YO
1815 U16 check;
1816 U32 newstate = 0;
a3621e74 1817
786e8c11
YO
1818 charid--;
1819 if ( !trie->states[ state ].trans.list ) {
1820 TRIE_LIST_NEW( state );
c445ea15 1821 }
786e8c11
YO
1822 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1823 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1824 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1825 break;
1826 }
1827 }
1828 if ( ! newstate ) {
1829 newstate = next_alloc++;
2e64971a 1830 prev_states[newstate] = state;
786e8c11
YO
1831 TRIE_LIST_PUSH( state, charid, newstate );
1832 transcount++;
1833 }
1834 state = newstate;
1835 } else {
1836 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1837 }
a28509cc 1838 }
c445ea15 1839 }
3dab1dad 1840 TRIE_HANDLE_WORD(state);
a3621e74
YO
1841
1842 } /* end second pass */
1843
1e2e3d02
YO
1844 /* next alloc is the NEXT state to be allocated */
1845 trie->statecount = next_alloc;
c944940b
JH
1846 trie->states = (reg_trie_state *)
1847 PerlMemShared_realloc( trie->states,
1848 next_alloc
1849 * sizeof(reg_trie_state) );
a3621e74 1850
3dab1dad 1851 /* and now dump it out before we compress it */
2b8b4781
NC
1852 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1853 revcharmap, next_alloc,
1854 depth+1)
1e2e3d02 1855 );
a3621e74 1856
c944940b
JH
1857 trie->trans = (reg_trie_trans *)
1858 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1859 {
1860 U32 state;
a3621e74
YO
1861 U32 tp = 0;
1862 U32 zp = 0;
1863
1864
1865 for( state=1 ; state < next_alloc ; state ++ ) {
1866 U32 base=0;
1867
1868 /*
1869 DEBUG_TRIE_COMPILE_MORE_r(
1870 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1871 );
1872 */
1873
1874 if (trie->states[state].trans.list) {
1875 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1876 U16 maxid=minid;
a28509cc 1877 U16 idx;
a3621e74
YO
1878
1879 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1880 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1881 if ( forid < minid ) {
1882 minid=forid;
1883 } else if ( forid > maxid ) {
1884 maxid=forid;
1885 }
a3621e74
YO
1886 }
1887 if ( transcount < tp + maxid - minid + 1) {
1888 transcount *= 2;
c944940b
JH
1889 trie->trans = (reg_trie_trans *)
1890 PerlMemShared_realloc( trie->trans,
446bd890
NC
1891 transcount
1892 * sizeof(reg_trie_trans) );
a3621e74
YO
1893 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1894 }
1895 base = trie->uniquecharcount + tp - minid;
1896 if ( maxid == minid ) {
1897 U32 set = 0;
1898 for ( ; zp < tp ; zp++ ) {
1899 if ( ! trie->trans[ zp ].next ) {
1900 base = trie->uniquecharcount + zp - minid;
1901 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1902 trie->trans[ zp ].check = state;
1903 set = 1;
1904 break;
1905 }
1906 }
1907 if ( !set ) {
1908 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1909 trie->trans[ tp ].check = state;
1910 tp++;
1911 zp = tp;
1912 }
1913 } else {
1914 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1915 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1916 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1917 trie->trans[ tid ].check = state;
1918 }
1919 tp += ( maxid - minid + 1 );
1920 }
1921 Safefree(trie->states[ state ].trans.list);
1922 }
1923 /*
1924 DEBUG_TRIE_COMPILE_MORE_r(
1925 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1926 );
1927 */
1928 trie->states[ state ].trans.base=base;
1929 }
cc601c31 1930 trie->lasttrans = tp + 1;
a3621e74
YO
1931 }
1932 } else {
1933 /*
1934 Second Pass -- Flat Table Representation.
1935
1936 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1937 We know that we will need Charcount+1 trans at most to store the data
1938 (one row per char at worst case) So we preallocate both structures
1939 assuming worst case.
1940
1941 We then construct the trie using only the .next slots of the entry
1942 structs.
1943
3b753521 1944 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1945 make compression both faster and easier by keeping track of how many non
1946 zero fields are in the node.
1947
1948 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1949 transition.
1950
1951 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1952 number representing the first entry of the node, and state as a
1953 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1954 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1955 are 2 entrys per node. eg:
1956
1957 A B A B
1958 1. 2 4 1. 3 7
1959 2. 0 3 3. 0 5
1960 3. 0 0 5. 0 0
1961 4. 0 0 7. 0 0
1962
1963 The table is internally in the right hand, idx form. However as we also
1964 have to deal with the states array which is indexed by nodenum we have to
1965 use TRIE_NODENUM() to convert.
1966
1967 */
1e2e3d02
YO
1968 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1969 "%*sCompiling trie using table compiler\n",
1970 (int)depth * 2 + 2, ""));
3dab1dad 1971
c944940b
JH
1972 trie->trans = (reg_trie_trans *)
1973 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1974 * trie->uniquecharcount + 1,
1975 sizeof(reg_trie_trans) );
1976 trie->states = (reg_trie_state *)
1977 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1978 sizeof(reg_trie_state) );
a3621e74
YO
1979 next_alloc = trie->uniquecharcount + 1;
1980
3dab1dad 1981
a3621e74
YO
1982 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1983
df826430 1984 regnode *noper = NEXTOPER( cur );
a28509cc 1985 const U8 *uc = (U8*)STRING( noper );
df826430 1986 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1987
1988 U32 state = 1; /* required init */
1989
1990 U16 charid = 0; /* sanity init */
1991 U32 accept_state = 0; /* sanity init */
1992 U8 *scan = (U8*)NULL; /* sanity init */
1993
1994 STRLEN foldlen = 0; /* required init */
07be1b83 1995 U32 wordlen = 0; /* required init */
fab2782b 1996 STRLEN skiplen = 0;
a3621e74
YO
1997 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1998
df826430
YO
1999 if (OP(noper) == NOTHING) {
2000 regnode *noper_next= regnext(noper);
2001 if (noper_next != tail && OP(noper_next) == flags) {
2002 noper = noper_next;
2003 uc= (U8*)STRING(noper);
2004 e= uc + STR_LEN(noper);
2005 }
2006 }
fab2782b 2007
3dab1dad 2008 if ( OP(noper) != NOTHING ) {
786e8c11 2009 for ( ; uc < e ; uc += len ) {
a3621e74 2010
786e8c11 2011 TRIE_READ_CHAR;
a3621e74 2012
786e8c11
YO
2013 if ( uvc < 256 ) {
2014 charid = trie->charmap[ uvc ];
2015 } else {
55eed653 2016 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2017 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2018 }
786e8c11
YO
2019 if ( charid ) {
2020 charid--;
2021 if ( !trie->trans[ state + charid ].next ) {
2022 trie->trans[ state + charid ].next = next_alloc;
2023 trie->trans[ state ].check++;
2e64971a
DM
2024 prev_states[TRIE_NODENUM(next_alloc)]
2025 = TRIE_NODENUM(state);
786e8c11
YO
2026 next_alloc += trie->uniquecharcount;
2027 }
2028 state = trie->trans[ state + charid ].next;
2029 } else {
2030 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2031 }
2032 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2033 }
a3621e74 2034 }
3dab1dad
YO
2035 accept_state = TRIE_NODENUM( state );
2036 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2037
2038 } /* end second pass */
2039
3dab1dad 2040 /* and now dump it out before we compress it */
2b8b4781
NC
2041 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2042 revcharmap,
2043 next_alloc, depth+1));
a3621e74 2044
a3621e74
YO
2045 {
2046 /*
2047 * Inplace compress the table.*
2048
2049 For sparse data sets the table constructed by the trie algorithm will
2050 be mostly 0/FAIL transitions or to put it another way mostly empty.
2051 (Note that leaf nodes will not contain any transitions.)
2052
2053 This algorithm compresses the tables by eliminating most such
2054 transitions, at the cost of a modest bit of extra work during lookup:
2055
2056 - Each states[] entry contains a .base field which indicates the
2057 index in the state[] array wheres its transition data is stored.
2058
3b753521 2059 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2060
2061 - If .base is nonzero then charid is added to it to find an entry in
2062 the trans array.
2063
2064 -If trans[states[state].base+charid].check!=state then the
2065 transition is taken to be a 0/Fail transition. Thus if there are fail
2066 transitions at the front of the node then the .base offset will point
2067 somewhere inside the previous nodes data (or maybe even into a node
2068 even earlier), but the .check field determines if the transition is
2069 valid.
2070
786e8c11 2071 XXX - wrong maybe?
a3621e74 2072 The following process inplace converts the table to the compressed
3b753521 2073 table: We first do not compress the root node 1,and mark all its
a3621e74 2074 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2075 allows us to do a DFA construction from the compressed table later,
2076 and ensures that any .base pointers we calculate later are greater
2077 than 0.
a3621e74
YO
2078
2079 - We set 'pos' to indicate the first entry of the second node.
2080
2081 - We then iterate over the columns of the node, finding the first and
2082 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2083 and set the .check pointers accordingly, and advance pos
2084 appropriately and repreat for the next node. Note that when we copy
2085 the next pointers we have to convert them from the original
2086 NODEIDX form to NODENUM form as the former is not valid post
2087 compression.
2088
2089 - If a node has no transitions used we mark its base as 0 and do not
2090 advance the pos pointer.
2091
2092 - If a node only has one transition we use a second pointer into the
2093 structure to fill in allocated fail transitions from other states.
2094 This pointer is independent of the main pointer and scans forward
2095 looking for null transitions that are allocated to a state. When it
2096 finds one it writes the single transition into the "hole". If the
786e8c11 2097 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2098
2099 - Once compressed we can Renew/realloc the structures to release the
2100 excess space.
2101
2102 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2103 specifically Fig 3.47 and the associated pseudocode.
2104
2105 demq
2106 */
a3b680e6 2107 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2108 U32 state, charid;
a3621e74 2109 U32 pos = 0, zp=0;
1e2e3d02 2110 trie->statecount = laststate;
a3621e74
YO
2111
2112 for ( state = 1 ; state < laststate ; state++ ) {
2113 U8 flag = 0;
a28509cc
AL
2114 const U32 stateidx = TRIE_NODEIDX( state );
2115 const U32 o_used = trie->trans[ stateidx ].check;
2116 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2117 trie->trans[ stateidx ].check = 0;
2118
2119 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2120 if ( flag || trie->trans[ stateidx + charid ].next ) {
2121 if ( trie->trans[ stateidx + charid ].next ) {
2122 if (o_used == 1) {
2123 for ( ; zp < pos ; zp++ ) {
2124 if ( ! trie->trans[ zp ].next ) {
2125 break;
2126 }
2127 }
2128 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2129 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2130 trie->trans[ zp ].check = state;
2131 if ( ++zp > pos ) pos = zp;
2132 break;
2133 }
2134 used--;
2135 }
2136 if ( !flag ) {
2137 flag = 1;
2138 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2139 }
2140 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2141 trie->trans[ pos ].check = state;
2142 pos++;
2143 }
2144 }
2145 }
cc601c31 2146 trie->lasttrans = pos + 1;
c944940b
JH
2147 trie->states = (reg_trie_state *)
2148 PerlMemShared_realloc( trie->states, laststate
2149 * sizeof(reg_trie_state) );
a3621e74 2150 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2151 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2152 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2153 (int)depth * 2 + 2,"",
2154 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2155 (IV)next_alloc,
2156 (IV)pos,
a3621e74
YO
2157 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2158 );
2159
2160 } /* end table compress */
2161 }
1e2e3d02
YO
2162 DEBUG_TRIE_COMPILE_MORE_r(
2163 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2164 (int)depth * 2 + 2, "",
2165 (UV)trie->statecount,
2166 (UV)trie->lasttrans)
2167 );
cc601c31 2168 /* resize the trans array to remove unused space */
c944940b
JH
2169 trie->trans = (reg_trie_trans *)
2170 PerlMemShared_realloc( trie->trans, trie->lasttrans
2171 * sizeof(reg_trie_trans) );
a3621e74 2172
3b753521 2173 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2174 U8 nodetype =(U8)(flags & 0xFF);
2175 char *str=NULL;
786e8c11 2176
07be1b83 2177#ifdef DEBUGGING
e62cc96a 2178 regnode *optimize = NULL;
7122b237
YO
2179#ifdef RE_TRACK_PATTERN_OFFSETS
2180
b57a0404
JH
2181 U32 mjd_offset = 0;
2182 U32 mjd_nodelen = 0;
7122b237
YO
2183#endif /* RE_TRACK_PATTERN_OFFSETS */
2184#endif /* DEBUGGING */
a3621e74 2185 /*
3dab1dad
YO
2186 This means we convert either the first branch or the first Exact,
2187 depending on whether the thing following (in 'last') is a branch
2188 or not and whther first is the startbranch (ie is it a sub part of
2189 the alternation or is it the whole thing.)
3b753521 2190 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2191 the whole branch sequence, including the first.
a3621e74 2192 */
3dab1dad 2193 /* Find the node we are going to overwrite */
7f69552c 2194 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2195 /* branch sub-chain */
3dab1dad 2196 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2197#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2198 DEBUG_r({
2199 mjd_offset= Node_Offset((convert));
2200 mjd_nodelen= Node_Length((convert));
2201 });
7122b237 2202#endif
7f69552c 2203 /* whole branch chain */
7122b237
YO
2204 }
2205#ifdef RE_TRACK_PATTERN_OFFSETS
2206 else {
7f69552c
YO
2207 DEBUG_r({
2208 const regnode *nop = NEXTOPER( convert );
2209 mjd_offset= Node_Offset((nop));
2210 mjd_nodelen= Node_Length((nop));
2211 });
07be1b83
YO
2212 }
2213 DEBUG_OPTIMISE_r(
2214 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2215 (int)depth * 2 + 2, "",
786e8c11 2216 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2217 );
7122b237 2218#endif
3dab1dad
YO
2219 /* But first we check to see if there is a common prefix we can
2220 split out as an EXACT and put in front of the TRIE node. */
2221 trie->startstate= 1;
55eed653 2222 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2223 U32 state;
1e2e3d02 2224 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2225 U32 ofs = 0;
8e11feef
RGS
2226 I32 idx = -1;
2227 U32 count = 0;
2228 const U32 base = trie->states[ state ].trans.base;
a3621e74 2229
3dab1dad 2230 if ( trie->states[state].wordnum )
8e11feef 2231 count = 1;
a3621e74 2232
8e11feef 2233 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2234 if ( ( base + ofs >= trie->uniquecharcount ) &&
2235 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2236 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2237 {
3dab1dad 2238 if ( ++count > 1 ) {
2b8b4781 2239 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2240 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2241 if ( state == 1 ) break;
3dab1dad
YO
2242 if ( count == 2 ) {
2243 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2244 DEBUG_OPTIMISE_r(
8e11feef
RGS
2245 PerlIO_printf(Perl_debug_log,
2246 "%*sNew Start State=%"UVuf" Class: [",
2247 (int)depth * 2 + 2, "",
786e8c11 2248 (UV)state));
be8e71aa 2249 if (idx >= 0) {
2b8b4781 2250 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2251 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2252
3dab1dad 2253 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2254 if ( folder )
2255 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2256 DEBUG_OPTIMISE_r(
f1f66076 2257 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2258 );
8e11feef
RGS
2259 }
2260 }
2261 TRIE_BITMAP_SET(trie,*ch);
2262 if ( folder )
2263 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2264 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2265 }
2266 idx = ofs;
2267 }
3dab1dad
YO
2268 }
2269 if ( count == 1 ) {
2b8b4781 2270 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2271 STRLEN len;
2272 char *ch = SvPV( *tmp, len );
de734bd5
A
2273 DEBUG_OPTIMISE_r({
2274 SV *sv=sv_newmortal();
8e11feef
RGS
2275 PerlIO_printf( Perl_debug_log,
2276 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2277 (int)depth * 2 + 2, "",
de734bd5
A
2278 (UV)state, (UV)idx,
2279 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2280 PL_colors[0], PL_colors[1],
2281 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2282 PERL_PV_ESCAPE_FIRSTCHAR
2283 )
2284 );
2285 });
3dab1dad
YO
2286 if ( state==1 ) {
2287 OP( convert ) = nodetype;
2288 str=STRING(convert);
2289 STR_LEN(convert)=0;
2290 }
c490c714
YO
2291 STR_LEN(convert) += len;
2292 while (len--)
de734bd5 2293 *str++ = *ch++;
8e11feef 2294 } else {
f9049ba1 2295#ifdef DEBUGGING
8e11feef
RGS
2296 if (state>1)
2297 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2298#endif
8e11feef
RGS
2299 break;
2300 }
2301 }
2e64971a 2302 trie->prefixlen = (state-1);
3dab1dad 2303 if (str) {
8e11feef 2304 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2305 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2306 trie->startstate = state;
07be1b83
YO
2307 trie->minlen -= (state - 1);
2308 trie->maxlen -= (state - 1);
33809eae
JH
2309#ifdef DEBUGGING
2310 /* At least the UNICOS C compiler choked on this
2311 * being argument to DEBUG_r(), so let's just have
2312 * it right here. */
2313 if (
2314#ifdef PERL_EXT_RE_BUILD
2315 1
2316#else
2317 DEBUG_r_TEST
2318#endif
2319 ) {
2320 regnode *fix = convert;
2321 U32 word = trie->wordcount;
2322 mjd_nodelen++;
2323 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2324 while( ++fix < n ) {
2325 Set_Node_Offset_Length(fix, 0, 0);
2326 }
2327 while (word--) {
2328 SV ** const tmp = av_fetch( trie_words, word, 0 );
2329 if (tmp) {
2330 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2331 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2332 else
2333 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2334 }
2335 }
2336 }
2337#endif
8e11feef
RGS
2338 if (trie->maxlen) {
2339 convert = n;
2340 } else {
3dab1dad 2341 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2342 DEBUG_r(optimize= n);
3dab1dad
YO
2343 }
2344 }
2345 }
a5ca303d
YO
2346 if (!jumper)
2347 jumper = last;
3dab1dad 2348 if ( trie->maxlen ) {
8e11feef
RGS
2349 NEXT_OFF( convert ) = (U16)(tail - convert);
2350 ARG_SET( convert, data_slot );
786e8c11
YO
2351 /* Store the offset to the first unabsorbed branch in
2352 jump[0], which is otherwise unused by the jump logic.
2353 We use this when dumping a trie and during optimisation. */
2354 if (trie->jump)
7f69552c 2355 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2356
6c48061a
YO
2357 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2358 * and there is a bitmap
2359 * and the first "jump target" node we found leaves enough room
2360 * then convert the TRIE node into a TRIEC node, with the bitmap
2361 * embedded inline in the opcode - this is hypothetically faster.
2362 */
2363 if ( !trie->states[trie->startstate].wordnum
2364 && trie->bitmap
2365 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2366 {
2367 OP( convert ) = TRIEC;
2368 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2369 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2370 trie->bitmap= NULL;
2371 } else
2372 OP( convert ) = TRIE;
a3621e74 2373
3dab1dad
YO
2374 /* store the type in the flags */
2375 convert->flags = nodetype;
a5ca303d
YO
2376 DEBUG_r({
2377 optimize = convert
2378 + NODE_STEP_REGNODE
2379 + regarglen[ OP( convert ) ];
2380 });
2381 /* XXX We really should free up the resource in trie now,
2382 as we won't use them - (which resources?) dmq */
3dab1dad 2383 }
a3621e74 2384 /* needed for dumping*/
e62cc96a 2385 DEBUG_r(if (optimize) {
07be1b83 2386 regnode *opt = convert;
bcdf7404 2387
e62cc96a 2388 while ( ++opt < optimize) {
07be1b83
YO
2389 Set_Node_Offset_Length(opt,0,0);
2390 }
786e8c11
YO
2391 /*
2392 Try to clean up some of the debris left after the
2393 optimisation.
a3621e74 2394 */
786e8c11 2395 while( optimize < jumper ) {
07be1b83 2396 mjd_nodelen += Node_Length((optimize));
a3621e74 2397 OP( optimize ) = OPTIMIZED;
07be1b83 2398 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2399 optimize++;
2400 }
07be1b83 2401 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2402 });
2403 } /* end node insert */
2e64971a
DM
2404
2405 /* Finish populating the prev field of the wordinfo array. Walk back
2406 * from each accept state until we find another accept state, and if
2407 * so, point the first word's .prev field at the second word. If the
2408 * second already has a .prev field set, stop now. This will be the
2409 * case either if we've already processed that word's accept state,
3b753521
FN
2410 * or that state had multiple words, and the overspill words were
2411 * already linked up earlier.
2e64971a
DM
2412 */
2413 {
2414 U16 word;
2415 U32 state;
2416 U16 prev;
2417
2418 for (word=1; word <= trie->wordcount; word++) {
2419 prev = 0;
2420 if (trie->wordinfo[word].prev)
2421 continue;
2422 state = trie->wordinfo[word].accept;
2423 while (state) {
2424 state = prev_states[state];
2425 if (!state)
2426 break;
2427 prev = trie->states[state].wordnum;
2428 if (prev)
2429 break;
2430 }
2431 trie->wordinfo[word].prev = prev;
2432 }
2433 Safefree(prev_states);
2434 }
2435
2436
2437 /* and now dump out the compressed format */
2438 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2439
55eed653 2440 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2441#ifdef DEBUGGING
2442 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2443 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2444#else
03e70be4 2445 SvREFCNT_dec_NN(revcharmap);
07be1b83 2446#endif
786e8c11
YO
2447 return trie->jump
2448 ? MADE_JUMP_TRIE
2449 : trie->startstate>1
2450 ? MADE_EXACT_TRIE
2451 : MADE_TRIE;
2452}
2453
2454STATIC void
2455S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2456{
3b753521 2457/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2458
2459 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2460 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2461 ISBN 0-201-10088-6
2462
2463 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2464 suffix of the current state's 'word' that is also a proper prefix of another word in our
2465 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2466 the DFA not to have to restart after its tried and failed a word at a given point, it
2467 simply continues as though it had been matching the other word in the first place.
2468 Consider
2469 'abcdgu'=~/abcdefg|cdgu/
2470 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2471 fail, which would bring us to the state representing 'd' in the second word where we would
2472 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2473 */
2474 /* add a fail transition */
3251b653
NC
2475 const U32 trie_offset = ARG(source);
2476 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2477 U32 *q;
2478 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2479 const U32 numstates = trie->statecount;
786e8c11
YO
2480 const U32 ubound = trie->lasttrans + ucharcount;
2481 U32 q_read = 0;
2482 U32 q_write = 0;
2483 U32 charid;
2484 U32 base = trie->states[ 1 ].trans.base;
2485 U32 *fail;
2486 reg_ac_data *aho;
2487 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2488 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2489
2490 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2491#ifndef DEBUGGING
2492 PERL_UNUSED_ARG(depth);
2493#endif
2494
2495
2496 ARG_SET( stclass, data_slot );
c944940b 2497 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2498 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2499 aho->trie=trie_offset;
446bd890
NC
2500 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2501 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2502 Newxz( q, numstates, U32);
c944940b 2503 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2504 aho->refcount = 1;
2505 fail = aho->fail;
2506 /* initialize fail[0..1] to be 1 so that we always have
2507 a valid final fail state */
2508 fail[ 0 ] = fail[ 1 ] = 1;
2509
2510 for ( charid = 0; charid < ucharcount ; charid++ ) {
2511 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2512 if ( newstate ) {
2513 q[ q_write ] = newstate;
2514 /* set to point at the root */
2515 fail[ q[ q_write++ ] ]=1;
2516 }
2517 }
2518 while ( q_read < q_write) {
2519 const U32 cur = q[ q_read++ % numstates ];
2520 base = trie->states[ cur ].trans.base;
2521
2522 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2523 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2524 if (ch_state) {
2525 U32 fail_state = cur;
2526 U32 fail_base;
2527 do {
2528 fail_state = fail[ fail_state ];
2529 fail_base = aho->states[ fail_state ].trans.base;
2530 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2531
2532 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2533 fail[ ch_state ] = fail_state;
2534 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2535 {
2536 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2537 }
2538 q[ q_write++ % numstates] = ch_state;
2539 }
2540 }
2541 }
2542 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2543 when we fail in state 1, this allows us to use the
2544 charclass scan to find a valid start char. This is based on the principle
2545 that theres a good chance the string being searched contains lots of stuff
2546 that cant be a start char.
2547 */
2548 fail[ 0 ] = fail[ 1 ] = 0;
2549 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2550 PerlIO_printf(Perl_debug_log,
2551 "%*sStclass Failtable (%"UVuf" states): 0",
2552 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2553 );
786e8c11
YO
2554 for( q_read=1; q_read<numstates; q_read++ ) {
2555 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2556 }
2557 PerlIO_printf(Perl_debug_log, "\n");
2558 });
2559 Safefree(q);
2560 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2561}
2562
786e8c11 2563
a3621e74 2564/*
5d1c421c
JH
2565 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2566 * These need to be revisited when a newer toolchain becomes available.
2567 */
2568#if defined(__sparc64__) && defined(__GNUC__)
2569# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2570# undef SPARC64_GCC_WORKAROUND
2571# define SPARC64_GCC_WORKAROUND 1
2572# endif
2573#endif
2574
07be1b83 2575#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2576 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2577 SV * const mysv=sv_newmortal(); \
2578 regnode *Next = regnext(scan); \
2579 regprop(RExC_rx, mysv, scan); \
7f69552c 2580 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2581 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2582 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2583 }});
07be1b83 2584
1de06328 2585
bb914485 2586/* The below joins as many adjacent EXACTish nodes as possible into a single
0a982f06
KW
2587 * one. The regop may be changed if the node(s) contain certain sequences that
2588 * require special handling. The joining is only done if:
bb914485
KW
2589 * 1) there is room in the current conglomerated node to entirely contain the
2590 * next one.
2591 * 2) they are the exact same node type
2592 *
87b8b349 2593 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
2594 * these get optimized out
2595 *
0a982f06
KW
2596 * If a node is to match under /i (folded), the number of characters it matches
2597 * can be different than its character length if it contains a multi-character
2598 * fold. *min_subtract is set to the total delta of the input nodes.
bb914485 2599 *
a0c4c608
KW
2600 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2601 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2602 *
bb914485 2603 * This is as good a place as any to discuss the design of handling these
0a982f06
KW
2604 * multi-character fold sequences. It's been wrong in Perl for a very long
2605 * time. There are three code points in Unicode whose multi-character folds
2606 * were long ago discovered to mess things up. The previous designs for
2607 * dealing with these involved assigning a special node for them. This
2608 * approach doesn't work, as evidenced by this example:
a0c4c608 2609 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
0a982f06
KW
2610 * Both these fold to "sss", but if the pattern is parsed to create a node that
2611 * would match just the \xDF, it won't be able to handle the case where a
bb914485
KW
2612 * successful match would have to cross the node's boundary. The new approach
2613 * that hopefully generally solves the problem generates an EXACTFU_SS node
2614 * that is "sss".
2615 *
0a982f06
KW
2616 * It turns out that there are problems with all multi-character folds, and not
2617 * just these three. Now the code is general, for all such cases, but the
2618 * three still have some special handling. The approach taken is:
2619 * 1) This routine examines each EXACTFish node that could contain multi-
2620 * character fold sequences. It returns in *min_subtract how much to
9d071ca8 2621 * subtract from the the actual length of the string to get a real minimum
0a982f06
KW
2622 * match length; it is 0 if there are no multi-char folds. This delta is
2623 * used by the caller to adjust the min length of the match, and the delta
2624 * between min and max, so that the optimizer doesn't reject these
2625 * possibilities based on size constraints.
2626 * 2) Certain of these sequences require special handling by the trie code,
2627 * so, if found, this code changes the joined node type to special ops:
2628 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2629 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2630 * is used for an EXACTFU node that contains at least one "ss" sequence in
2631 * it. For non-UTF-8 patterns and strings, this is the only case where
2632 * there is a possible fold length change. That means that a regular
2633 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2634 * with length changes, and so can be processed faster. regexec.c takes
2635 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2636 * pre-folded by regcomp.c. This saves effort in regex matching.
87b8b349 2637 * However, the pre-folding isn't done for non-UTF8 patterns because the
0a982f06
KW
2638 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2639 * down by forcing the pattern into UTF8 unless necessary. Also what
2640 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2641 * possibilities for the non-UTF8 patterns are quite simple, except for
2642 * the sharp s. All the ones that don't involve a UTF-8 target string are
2643 * members of a fold-pair, and arrays are set up for all of them so that
2644 * the other member of the pair can be found quickly. Code elsewhere in
2645 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2646 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2647 * described in the next item.
bb914485
KW
2648 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2649 * 'ss' or not is not knowable at compile time. It will match iff the
2650 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2651 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
6012a526 2652 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
bb914485
KW
2653 * described in item 3). An assumption that the optimizer part of
2654 * regexec.c (probably unwittingly) makes is that a character in the
2655 * pattern corresponds to at most a single character in the target string.
2656 * (And I do mean character, and not byte here, unlike other parts of the
2657 * documentation that have never been updated to account for multibyte
2658 * Unicode.) This assumption is wrong only in this case, as all other
2659 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2660 * virtue of having this file pre-fold UTF-8 patterns. I'm
2661 * reluctant to try to change this assumption, so instead the code punts.
9d071ca8
KW
2662 * This routine examines EXACTF nodes for the sharp s, and returns a
2663 * boolean indicating whether or not the node is an EXACTF node that
2664 * contains a sharp s. When it is true, the caller sets a flag that later
2665 * causes the optimizer in this file to not set values for the floating
2666 * and fixed string lengths, and thus avoids the optimizer code in
2667 * regexec.c that makes the invalid assumption. Thus, there is no
2668 * optimization based on string lengths for EXACTF nodes that contain the
2669 * sharp s. This only happens for /id rules (which means the pattern
2670 * isn't in UTF-8).
bb914485 2671 */
1de06328 2672
9d071ca8 2673#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2674 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2675 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2676
be8e71aa 2677STATIC U32
9d071ca8 2678S_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
2679 /* Merge several consecutive EXACTish nodes into one. */
2680 regnode *n = regnext(scan);
2681 U32 stringok = 1;
2682 regnode *next = scan + NODE_SZ_STR(scan);
2683 U32 merged = 0;
2684 U32 stopnow = 0;
2685#ifdef DEBUGGING
2686 regnode *stop = scan;
72f13be8 2687 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2688#else
d47053eb
RGS
2689 PERL_UNUSED_ARG(depth);
2690#endif
7918f24d
NC
2691
2692 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2693#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2694 PERL_UNUSED_ARG(flags);
2695 PERL_UNUSED_ARG(val);
07be1b83 2696#endif
07be1b83 2697 DEBUG_PEEP("join",scan,depth);
bb914485 2698
3f410cf6
KW
2699 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2700 * EXACT ones that are mergeable to the current one. */
2701 while (n
2702 && (PL_regkind[OP(n)] == NOTHING
2703 || (stringok && OP(n) == OP(scan)))
07be1b83 2704 && NEXT_OFF(n)
3f410cf6
KW
2705 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2706 {
07be1b83
YO
2707
2708 if (OP(n) == TAIL || n > next)
2709 stringok = 0;
2710 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2711 DEBUG_PEEP("skip:",n,depth);
2712 NEXT_OFF(scan) += NEXT_OFF(n);
2713 next = n + NODE_STEP_REGNODE;
2714#ifdef DEBUGGING
2715 if (stringok)
2716 stop = n;
2717#endif
2718 n = regnext(n);
2719 }
2720 else if (stringok) {
786e8c11 2721 const unsigned int oldl = STR_LEN(scan);
07be1b83 2722 regnode * const nnext = regnext(n);
b2230d39 2723
87b8b349
KW
2724 /* XXX I (khw) kind of doubt that this works on platforms where
2725 * U8_MAX is above 255 because of lots of other assumptions */
79a81a6e 2726 /* Don't join if the sum can't fit into a single node */
b2230d39
KW
2727 if (oldl + STR_LEN(n) > U8_MAX)
2728 break;
07be1b83
YO
2729
2730 DEBUG_PEEP("merg",n,depth);
07be1b83 2731 merged++;
b2230d39 2732
07be1b83
YO
2733 NEXT_OFF(scan) += NEXT_OFF(n);
2734 STR_LEN(scan) += STR_LEN(n);
2735 next = n + NODE_SZ_STR(n);
2736 /* Now we can overwrite *n : */
2737 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2738#ifdef DEBUGGING
2739 stop = next - 1;
2740#endif
2741 n = nnext;
2742 if (stopnow) break;
2743 }
2744
d47053eb
RGS
2745#ifdef EXPERIMENTAL_INPLACESCAN
2746 if (flags && !NEXT_OFF(n)) {
2747 DEBUG_PEEP("atch", val, depth);
2748 if (reg_off_by_arg[OP(n)]) {
2749 ARG_SET(n, val - n);
2750 }
2751 else {
2752 NEXT_OFF(n) = val - n;
2753 }
2754 stopnow = 1;
2755 }
07be1b83
YO
2756#endif
2757 }
2c2b7f86 2758
9d071ca8 2759 *min_subtract = 0;
f758bddf 2760 *has_exactf_sharp_s = FALSE;
f646642f 2761
3f410cf6
KW
2762 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2763 * can now analyze for sequences of problematic code points. (Prior to
2764 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2765 * hence missed). The sequences only happen in folding, hence for any
2766 * non-EXACT EXACTish node */
86d6fcad 2767 if (OP(scan) != EXACT) {
0a982f06
KW
2768 const U8 * const s0 = (U8*) STRING(scan);
2769 const U8 * s = s0;
2770 const U8 * const s_end = s0 + STR_LEN(scan);
f758bddf
KW
2771
2772 /* One pass is made over the node's string looking for all the
2773 * possibilities. to avoid some tests in the loop, there are two main
2774 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2775 * non-UTF-8 */
2776 if (UTF) {
86d6fcad 2777
0a982f06
KW
2778 /* Examine the string for a multi-character fold sequence. UTF-8
2779 * patterns have all characters pre-folded by the time this code is
2780 * executed */
2781 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2782 length sequence we are looking for is 2 */
86d6fcad 2783 {
0a982f06
KW
2784 int count = 0;
2785 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2786 if (! len) { /* Not a multi-char fold: get next char */
2787 s += UTF8SKIP(s);
2788 continue;
2789 }
bb914485 2790
0a982f06
KW
2791 /* Nodes with 'ss' require special handling, except for EXACTFL
2792 * and EXACTFA for which there is no multi-char fold to this */
2793 if (len == 2 && *s == 's' && *(s+1) == 's'
2794 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2795 {
2796 count = 2;
2797 OP(scan) = EXACTFU_SS;
2798 s += 2;
2799 }
2800 else if (len == 6 /* len is the same in both ASCII and EBCDIC for these */
2801 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2802 COMBINING_DIAERESIS_UTF8
2803 COMBINING_ACUTE_ACCENT_UTF8,
2804 6)
2805 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2806 COMBINING_DIAERESIS_UTF8
2807 COMBINING_ACUTE_ACCENT_UTF8,
2808 6)))
2809 {
2810 count = 3;
2811
2812 /* These two folds require special handling by trie's, so
2813 * change the node type to indicate this. If EXACTFA and
2814 * EXACTFL were ever to be handled by trie's, this would
2815 * have to be changed. If this node has already been
2816 * changed to EXACTFU_SS in this loop, leave it as is. (I
2817 * (khw) think it doesn't matter in regexec.c for UTF
2818 * patterns, but no need to change it */
2819 if (OP(scan) == EXACTFU) {
2820 OP(scan) = EXACTFU_TRICKYFOLD;
2821 }
2822 s += 6;
2823 }
2824 else { /* Here is a generic multi-char fold. */
2825 const U8* multi_end = s + len;
2826
2827 /* Count how many characters in it. In the case of /l and
2828 * /aa, no folds which contain ASCII code points are
2829 * allowed, so check for those, and skip if found. (In
2830 * EXACTFL, no folds are allowed to any Latin1 code point,
2831 * not just ASCII. But there aren't any of these
2832 * currently, nor ever likely, so don't take the time to
2833 * test for them. The code that generates the
2834 * is_MULTI_foo() macros croaks should one actually get put
2835 * into Unicode .) */
2836 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2837 count = utf8_length(s, multi_end);
2838 s = multi_end;
2839 }
2840 else {
2841 while (s < multi_end) {
2842 if (isASCII(*s)) {
2843 s++;
2844 goto next_iteration;
2845 }
2846 else {
2847 s += UTF8SKIP(s);
2848 }
2849 count++;
2850 }
2851 }
2852 }
f758bddf 2853
0a982f06
KW
2854 /* The delta is how long the sequence is minus 1 (1 is how long
2855 * the character that folds to the sequence is) */
2856 *min_subtract += count - 1;
2857 next_iteration: ;
bb914485
KW
2858 }
2859 }
f758bddf 2860 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
bb914485 2861
0a982f06
KW
2862 /* Here, the pattern is not UTF-8. Look for the multi-char folds
2863 * that are all ASCII. As in the above case, EXACTFL and EXACTFA
2864 * nodes can't have multi-char folds to this range (and there are
2db97d41 2865 * no existing ones in the upper latin1 range). In the EXACTF
0a982f06
KW
2866 * case we look also for the sharp s, which can be in the final
2867 * position. Otherwise we can stop looking 1 byte earlier because
2868 * have to find at least two characters for a multi-fold */
f758bddf
KW
2869 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2870
0a982f06
KW
2871 /* The below is perhaps overboard, but this allows us to save a
2872 * test each time through the loop at the expense of a mask. This
2873 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2874 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2875 * are 64. This uses an exclusive 'or' to find that bit and then
2876 * inverts it to form a mask, with just a single 0, in the bit
2877 * position where 'S' and 's' differ. */
2878 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2879 const U8 s_masked = 's' & S_or_s_mask;
2880
2881 while (s < upper) {
40b1ba4f 2882 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
0a982f06
KW
2883 if (! len) { /* Not a multi-char fold. */
2884 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2885 {
2886 *has_exactf_sharp_s = TRUE;
2887 }
2888 s++;
2889 continue;
2890 }
2891
2892 if (len == 2
2893 && ((*s & S_or_s_mask) == s_masked)
2894 && ((*(s+1) & S_or_s_mask) == s_masked))
2895 {
2896
2897 /* EXACTF nodes need to know that the minimum length
2898 * changed so that a sharp s in the string can match this
2899 * ss in the pattern, but they remain EXACTF nodes, as they
2900 * won't match this unless the target string is is UTF-8,
2901 * which we don't know until runtime */
2902 if (OP(scan) != EXACTF) {
2903 OP(scan) = EXACTFU_SS;
2904 }
86d6fcad 2905 }
0a982f06
KW
2906
2907 *min_subtract += len - 1;
2908 s += len;
86d6fcad
KW
2909 }
2910 }
07be1b83 2911 }
3f410cf6 2912
07be1b83 2913#ifdef DEBUGGING
bb789b09
DM
2914 /* Allow dumping but overwriting the collection of skipped
2915 * ops and/or strings with fake optimized ops */
07be1b83
YO
2916 n = scan + NODE_SZ_STR(scan);
2917 while (n <= stop) {
bb789b09
DM
2918 OP(n) = OPTIMIZED;
2919 FLAGS(n) = 0;
2920 NEXT_OFF(n) = 0;
07be1b83
YO
2921 n++;
2922 }
2923#endif
2924 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2925 return stopnow;
2926}
2927
486ec47a 2928/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2929 Finds fixed substrings. */
2930
a0288114 2931/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
2932 to the position after last scanned or to NULL. */
2933
40d049e4
YO
2934#define INIT_AND_WITHP \
2935 assert(!and_withp); \
2936 Newx(and_withp,1,struct regnode_charclass_class); \
2937 SAVEFREEPV(and_withp)
07be1b83 2938
b515a41d 2939/* this is a chain of data about sub patterns we are processing that
486ec47a 2940 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
2941 we can simulate recursion without losing state. */
2942struct scan_frame;
2943typedef struct scan_frame {
2944 regnode *last; /* last node to process in this frame */
2945 regnode *next; /* next node to process when last is reached */
2946 struct scan_frame *prev; /*previous frame*/
2947 I32 stop; /* what stopparen do we use */
2948} scan_frame;
2949
304ee84b
YO
2950
2951#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2952
76e3520e 2953STATIC I32
40d049e4 2954S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 2955 I32 *minlenp, I32 *deltap,
40d049e4
YO
2956 regnode *last,
2957 scan_data_t *data,
2958 I32 stopparen,
2959 U8* recursed,
2960 struct regnode_charclass_class *and_withp,
2961 U32 flags, U32 depth)
c277df42
IZ
2962 /* scanp: Start here (read-write). */
2963 /* deltap: Write maxlen-minlen here. */
2964 /* last: Stop before this one. */
40d049e4
YO
2965 /* data: string data about the pattern */
2966 /* stopparen: treat close N as END */
2967 /* recursed: which subroutines have we recursed into */
2968 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 2969{
97aff369 2970 dVAR;
2d608413
KW
2971 I32 min = 0; /* There must be at least this number of characters to match */
2972 I32 pars = 0, code;
c277df42
IZ
2973 regnode *scan = *scanp, *next;
2974 I32 delta = 0;
2975 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 2976 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
2977 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
2978 scan_data_t data_fake;
a3621e74 2979 SV *re_trie_maxbuff = NULL;
786e8c11 2980 regnode *first_non_open = scan;
e2e6a0f1 2981 I32 stopmin = I32_MAX;
8aa23a47 2982 scan_frame *frame = NULL;
a3621e74 2983 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 2984
7918f24d
NC
2985 PERL_ARGS_ASSERT_STUDY_CHUNK;
2986
13a24bad 2987#ifdef DEBUGGING
40d049e4 2988 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 2989#endif
40d049e4 2990
786e8c11 2991 if ( depth == 0 ) {
40d049e4 2992 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
2993 first_non_open=regnext(first_non_open);
2994 }
2995
b81d288d 2996
8aa23a47
YO
2997 fake_study_recurse:
2998 while ( scan && OP(scan) != END && scan < last ){
2d608413
KW
2999 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3000 node length to get a real minimum (because
3001 the folded version may be shorter) */
f758bddf 3002 bool has_exactf_sharp_s = FALSE;
8aa23a47 3003 /* Peephole optimizer: */
304ee84b 3004 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 3005 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
3006
3007 /* Its not clear to khw or hv why this is done here, and not in the
3008 * clauses that deal with EXACT nodes. khw's guess is that it's
3009 * because of a previous design */
9d071ca8 3010 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
3011
3012 /* Follow the next-chain of the current node and optimize
3013 away all the NOTHINGs from it. */
3014 if (OP(scan) != CURLYX) {
3015 const int max = (reg_off_by_arg[OP(scan)]
3016 ? I32_MAX
3017 /* I32 may be smaller than U16 on CRAYs! */
3018 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3019 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3020 int noff;
3021 regnode *n = scan;
686b73d4 3022
8aa23a47
YO
3023 /* Skip NOTHING and LONGJMP. */
3024 while ((n = regnext(n))
3025 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3026 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3027 && off + noff < max)
3028 off += noff;
3029 if (reg_off_by_arg[OP(scan)])
3030 ARG(scan) = off;
3031 else
3032 NEXT_OFF(scan) = off;
3033 }
a3621e74 3034
c277df42 3035
8aa23a47
YO
3036
3037 /* The principal pseudo-switch. Cannot be a switch, since we
3038 look into several different things. */
3039 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3040 || OP(scan) == IFTHEN) {
3041 next = regnext(scan);
3042 code = OP(scan);
3043 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3044
8aa23a47
YO
3045 if (OP(next) == code || code == IFTHEN) {
3046 /* NOTE - There is similar code to this block below for handling
3047 TRIE nodes on a re-study. If you change stuff here check there
3048 too. */
3049 I32 max1 = 0, min1 = I32_MAX, num = 0;
3050 struct regnode_charclass_class accum;
3051 regnode * const startbranch=scan;
686b73d4 3052
8aa23a47 3053 if (flags & SCF_DO_SUBSTR)
304ee84b 3054 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3055 if (flags & SCF_DO_STCLASS)
e755fd73 3056 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3057
3058 while (OP(scan) == code) {
3059 I32 deltanext, minnext, f = 0, fake;
3060 struct regnode_charclass_class this_class;
3061
3062 num++;
3063 data_fake.flags = 0;
3064 if (data) {
3065 data_fake.whilem_c = data->whilem_c;
3066 data_fake.last_closep = data->last_closep;
3067 }
3068 else
3069 data_fake.last_closep = &fake;
58e23c8d
YO
3070
3071 data_fake.pos_delta = delta;
8aa23a47
YO
3072 next = regnext(scan);
3073 scan = NEXTOPER(scan);
3074 if (code != BRANCH)
c277df42 3075 scan = NEXTOPER(scan);
8aa23a47 3076 if (flags & SCF_DO_STCLASS) {
e755fd73 3077 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3078 data_fake.start_class = &this_class;
3079 f = SCF_DO_STCLASS_AND;
58e23c8d 3080 }
8aa23a47
YO
3081 if (flags & SCF_WHILEM_VISITED_POS)
3082 f |= SCF_WHILEM_VISITED_POS;
3083
3084 /* we suppose the run is continuous, last=next...*/
3085 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3086 next, &data_fake,
3087 stopparen, recursed, NULL, f,depth+1);
3088 if (min1 > minnext)
3089 min1 = minnext;
3090 if (max1 < minnext + deltanext)
3091 max1 = minnext + deltanext;
3092 if (deltanext == I32_MAX)
3093 is_inf = is_inf_internal = 1;
3094 scan = next;
3095 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3096 pars++;
3097 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3098 if ( stopmin > minnext)
3099 stopmin = min + min1;
3100 flags &= ~SCF_DO_SUBSTR;
3101 if (data)
3102 data->flags |= SCF_SEEN_ACCEPT;
3103 }
3104 if (data) {
3105 if (data_fake.flags & SF_HAS_EVAL)
3106 data->flags |= SF_HAS_EVAL;
3107 data->whilem_c = data_fake.whilem_c;
3dab1dad 3108 }
8aa23a47 3109 if (flags & SCF_DO_STCLASS)
3fffb88a 3110 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3111 }
3112 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3113 min1 = 0;
3114 if (flags & SCF_DO_SUBSTR) {
3115 data->pos_min += min1;
3116 data->pos_delta += max1 - min1;
3117 if (max1 != min1 || is_inf)
3118 data->longest = &(data->longest_float);
3119 }
3120 min += min1;
3121 delta += max1 - min1;
3122 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3123 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3124 if (min1) {
3125 cl_and(data->start_class, and_withp);
3126 flags &= ~SCF_DO_STCLASS;
653099ff 3127 }
8aa23a47
YO
3128 }
3129 else if (flags & SCF_DO_STCLASS_AND) {
3130 if (min1) {
3131 cl_and(data->start_class, &accum);
3132 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3133 }
8aa23a47
YO
3134 else {
3135 /* Switch to OR mode: cache the old value of
3136 * data->start_class */
3137 INIT_AND_WITHP;
3138 StructCopy(data->start_class, and_withp,
3139 struct regnode_charclass_class);
3140 flags &= ~SCF_DO_STCLASS_AND;
3141 StructCopy(&accum, data->start_class,
3142 struct regnode_charclass_class);
3143 flags |= SCF_DO_STCLASS_OR;
3144 data->start_class->flags |= ANYOF_EOS;
de0c8cb8 3145 }
8aa23a47 3146 }
a3621e74 3147
8aa23a47
YO
3148 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3149 /* demq.
a3621e74 3150
8aa23a47
YO
3151 Assuming this was/is a branch we are dealing with: 'scan' now
3152 points at the item that follows the branch sequence, whatever
3153 it is. We now start at the beginning of the sequence and look
3154 for subsequences of
a3621e74 3155
8aa23a47
YO
3156 BRANCH->EXACT=>x1
3157 BRANCH->EXACT=>x2
3158 tail
a3621e74 3159
8aa23a47 3160 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3161
486ec47a 3162 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3163 element into a trie and then add the subsequent branch exact
3164 strings to the trie.
a3621e74 3165
8aa23a47 3166 We have two cases
a3621e74 3167
3b753521 3168 1. patterns where the whole set of branches can be converted.
a3621e74 3169
8aa23a47 3170 2. patterns where only a subset can be converted.
a3621e74 3171
8aa23a47
YO
3172 In case 1 we can replace the whole set with a single regop
3173 for the trie. In case 2 we need to keep the start and end
3b753521 3174 branches so
a3621e74 3175
8aa23a47
YO
3176 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3177 becomes BRANCH TRIE; BRANCH X;
786e8c11 3178
8aa23a47
YO
3179 There is an additional case, that being where there is a
3180 common prefix, which gets split out into an EXACT like node
3181 preceding the TRIE node.
a3621e74 3182
8aa23a47
YO
3183 If x(1..n)==tail then we can do a simple trie, if not we make
3184 a "jump" trie, such that when we match the appropriate word
486ec47a 3185 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3186 a nested if into a case structure of sorts.
b515a41d 3187
8aa23a47 3188 */
686b73d4 3189
8aa23a47
YO
3190 int made=0;
3191 if (!re_trie_maxbuff) {
3192 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3193 if (!SvIOK(re_trie_maxbuff))
3194 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3195 }
3196 if ( SvIV(re_trie_maxbuff)>=0 ) {
3197 regnode *cur;
3198 regnode *first = (regnode *)NULL;
3199 regnode *last = (regnode *)NULL;
3200 regnode *tail = scan;
fab2782b 3201 U8 trietype = 0;
8aa23a47 3202 U32 count=0;
a3621e74
YO
3203
3204#ifdef DEBUGGING
8aa23a47 3205 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3206#endif
8aa23a47
YO
3207 /* var tail is used because there may be a TAIL
3208 regop in the way. Ie, the exacts will point to the
3209 thing following the TAIL, but the last branch will
3210 point at the TAIL. So we advance tail. If we
3211 have nested (?:) we may have to move through several
3212 tails.
3213 */
3214
3215 while ( OP( tail ) == TAIL ) {
3216 /* this is the TAIL generated by (?:) */
3217 tail = regnext( tail );
3218 }
a3621e74 3219
8aa23a47 3220
df826430 3221 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3222 regprop(RExC_rx, mysv, tail );
3223 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3224 (int)depth * 2 + 2, "",
3225 "Looking for TRIE'able sequences. Tail node is: ",
3226 SvPV_nolen_const( mysv )
3227 );
3228 });
3229
3230 /*
3231
fab2782b
YO
3232 Step through the branches
3233 cur represents each branch,
3234 noper is the first thing to be matched as part of that branch
3235 noper_next is the regnext() of that node.
3236
3237 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3238 via a "jump trie" but we also support building with NOJUMPTRIE,
3239 which restricts the trie logic to structures like /FOO|BAR/.
3240
3241 If noper is a trieable nodetype then the branch is a possible optimization
3242 target. If we are building under NOJUMPTRIE then we require that noper_next
3243 is the same as scan (our current position in the regex program).
3244
3245 Once we have two or more consecutive such branches we can create a
3246 trie of the EXACT's contents and stitch it in place into the program.
3247
3248 If the sequence represents all of the branches in the alternation we
3249 replace the entire thing with a single TRIE node.
3250
3251 Otherwise when it is a subsequence we need to stitch it in place and
3252 replace only the relevant branches. This means the first branch has
3253 to remain as it is used by the alternation logic, and its next pointer,
3254 and needs to be repointed at the item on the branch chain following
3255 the last branch we have optimized away.
3256
3257 This could be either a BRANCH, in which case the subsequence is internal,
3258 or it could be the item following the branch sequence in which case the
3259 subsequence is at the end (which does not necessarily mean the first node
3260 is the start of the alternation).
3261
3262 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3263
3264 optype | trietype
3265 ----------------+-----------
3266 NOTHING | NOTHING
3267 EXACT | EXACT
3268 EXACTFU | EXACTFU
3269 EXACTFU_SS | EXACTFU
3270 EXACTFU_TRICKYFOLD | EXACTFU
3271 EXACTFA | 0
3272
8aa23a47
YO
3273
3274 */
fab2782b
YO
3275#define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3276 ( EXACT == (X) ) ? EXACT : \
3277 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3278 0 )
8aa23a47
YO
3279
3280 /* dont use tail as the end marker for this traverse */
3281 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3282 regnode * const noper = NEXTOPER( cur );
fab2782b
YO
3283 U8 noper_type = OP( noper );
3284 U8 noper_trietype = TRIE_TYPE( noper_type );
b515a41d 3285#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3286 regnode * const noper_next = regnext( noper );
df826430
YO
3287 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3288 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
b515a41d
YO
3289#endif
3290
df826430 3291 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3292 regprop(RExC_rx, mysv, cur);
3293 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3294 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3295
3296 regprop(RExC_rx, mysv, noper);
3297 PerlIO_printf( Perl_debug_log, " -> %s",
3298 SvPV_nolen_const(mysv));
3299
3300 if ( noper_next ) {
3301 regprop(RExC_rx, mysv, noper_next );
3302 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3303 SvPV_nolen_const(mysv));
3304 }
df826430
YO
3305 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3306 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3307 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3308 );
8aa23a47 3309 });
fab2782b
YO
3310
3311 /* Is noper a trieable nodetype that can be merged with the
3312 * current trie (if there is one)? */
3313 if ( noper_trietype
3314 &&
3315 (
df826430
YO
3316 ( noper_trietype == NOTHING)
3317 || ( trietype == NOTHING )
a40630bf 3318 || ( trietype == noper_trietype )
fab2782b 3319 )
786e8c11 3320#ifdef NOJUMPTRIE
8aa23a47 3321 && noper_next == tail
786e8c11 3322#endif
8aa23a47
YO
3323 && count < U16_MAX)
3324 {
fab2782b
YO
3325 /* Handle mergable triable node
3326 * Either we are the first node in a new trieable sequence,
3327 * in which case we do some bookkeeping, otherwise we update
3328 * the end pointer. */
fab2782b 3329 if ( !first ) {
3b6759a6 3330 first = cur;
df826430
YO
3331 if ( noper_trietype == NOTHING ) {
3332#if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3333 regnode * const noper_next = regnext( noper );
3b6759a6 3334 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
df826430
YO
3335 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3336#endif
3337
190c1910 3338 if ( noper_next_trietype ) {
df826430 3339 trietype = noper_next_trietype;
190c1910
YO
3340 } else if (noper_next_type) {
3341 /* a NOTHING regop is 1 regop wide. We need at least two
3342 * for a trie so we can't merge this in */
3343 first = NULL;
3344 }
3345 } else {
3346 trietype = noper_trietype;
3b6759a6 3347 }
8aa23a47 3348 } else {
fab2782b
YO
3349 if ( trietype == NOTHING )
3350 trietype = noper_trietype;
8aa23a47
YO
3351 last = cur;
3352 }
df826430
YO
3353 if (first)
3354 count++;
fab2782b
YO
3355 } /* end handle mergable triable node */
3356 else {
3357 /* handle unmergable node -
3358 * noper may either be a triable node which can not be tried
3359 * together with the current trie, or a non triable node */
729aaeb5
YO
3360 if ( last ) {
3361 /* If last is set and trietype is not NOTHING then we have found
3362 * at least two triable branch sequences in a row of a similar
3363 * trietype so we can turn them into a trie. If/when we
3364 * allow NOTHING to start a trie sequence this condition will be
3365 * required, and it isn't expensive so we leave it in for now. */
e6351b37 3366 if ( trietype && trietype != NOTHING )
729aaeb5
YO
3367 make_trie( pRExC_state,
3368 startbranch, first, cur, tail, count,
3369 trietype, depth+1 );
fab2782b 3370 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
8aa23a47 3371 }
fab2782b 3372 if ( noper_trietype
786e8c11 3373#ifdef NOJUMPTRIE
8aa23a47 3374 && noper_next == tail
786e8c11 3375#endif
8aa23a47 3376 ){
fab2782b 3377 /* noper is triable, so we can start a new trie sequence */
8aa23a47
YO
3378 count = 1;
3379 first = cur;
fab2782b
YO
3380 trietype = noper_trietype;
3381 } else if (first) {
3382 /* if we already saw a first but the current node is not triable then we have
3383 * to reset the first information. */
8aa23a47
YO
3384 count = 0;
3385 first = NULL;
fab2782b 3386 trietype = 0;
8aa23a47 3387 }
fab2782b
YO
3388 } /* end handle unmergable node */
3389 } /* loop over branches */
df826430 3390 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3391 regprop(RExC_rx, mysv, cur);
3392 PerlIO_printf( Perl_debug_log,
3393 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3394 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3395
3396 });
e6351b37 3397 if ( last && trietype ) {
3b6759a6
YO
3398 if ( trietype != NOTHING ) {
3399 /* the last branch of the sequence was part of a trie,
3400 * so we have to construct it here outside of the loop
3401 */
3402 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
686b73d4 3403#ifdef TRIE_STUDY_OPT
3b6759a6
YO
3404 if ( ((made == MADE_EXACT_TRIE &&
3405 startbranch == first)
3406 || ( first_non_open == first )) &&
3407 depth==0 ) {
3408 flags |= SCF_TRIE_RESTUDY;
3409 if ( startbranch == first
3410 && scan == tail )
3411 {
3412 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3413 }
8aa23a47 3414 }
3dab1dad 3415#endif
3b6759a6
YO
3416 } else {
3417 /* at this point we know whatever we have is a NOTHING sequence/branch
3418 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3419 */
3420 if ( startbranch == first ) {
3421 regnode *opt;
3422 /* the entire thing is a NOTHING sequence, something like this:
3423 * (?:|) So we can turn it into a plain NOTHING op. */
3424 DEBUG_TRIE_COMPILE_r({
3425 regprop(RExC_rx, mysv, cur);
3426 PerlIO_printf( Perl_debug_log,
3427 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3428 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3429
3430 });
3431 OP(startbranch)= NOTHING;
3432 NEXT_OFF(startbranch)= tail - startbranch;
3433 for ( opt= startbranch + 1; opt < tail ; opt++ )
3434 OP(opt)= OPTIMIZED;
3435 }
3436 }
fab2782b
YO
3437 } /* end if ( last) */
3438 } /* TRIE_MAXBUF is non zero */
8aa23a47
YO
3439
3440 } /* do trie */
3441
653099ff 3442 }
8aa23a47
YO
3443 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3444 scan = NEXTOPER(NEXTOPER(scan));
3445 } else /* single branch is optimized. */
3446 scan = NEXTOPER(scan);
3447 continue;
3448 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3449 scan_frame *newframe = NULL;
3450 I32 paren;
3451 regnode *start;
3452 regnode *end;
3453
3454 if (OP(scan) != SUSPEND) {
3455 /* set the pointer */
3456 if (OP(scan) == GOSUB) {
3457 paren = ARG(scan);
3458 RExC_recurse[ARG2L(scan)] = scan;
3459 start = RExC_open_parens[paren-1];
3460 end = RExC_close_parens[paren-1];
3461 } else {
3462 paren = 0;
f8fc2ecf 3463 start = RExC_rxi->program + 1;
8aa23a47
YO
3464 end = RExC_opend;
3465 }
3466 if (!recursed) {
3467 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3468 SAVEFREEPV(recursed);
3469 }
3470 if (!PAREN_TEST(recursed,paren+1)) {
3471 PAREN_SET(recursed,paren+1);
3472 Newx(newframe,1,scan_frame);
3473 } else {
3474 if (flags & SCF_DO_SUBSTR) {
304ee84b 3475 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3476 data->longest = &(data->longest_float);
3477 }
3478 is_inf = is_inf_internal = 1;
3479 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3480 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3481 flags &= ~SCF_DO_STCLASS;
3482 }
3483 } else {
3484 Newx(newframe,1,scan_frame);
3485 paren = stopparen;
3486 start = scan+2;
3487 end = regnext(scan);
3488 }
3489 if (newframe) {
3490 assert(start);
3491 assert(end);
3492 SAVEFREEPV(newframe);
3493 newframe->next = regnext(scan);
3494 newframe->last = last;
3495 newframe->stop = stopparen;
3496 newframe->prev = frame;
3497
3498 frame = newframe;
3499 scan = start;
3500 stopparen = paren;
3501 last = end;
3502
3503 continue;
3504 }
3505 }
3506 else if (OP(scan) == EXACT) {
3507 I32 l = STR_LEN(scan);
3508 UV uc;
3509 if (UTF) {
3510 const U8 * const s = (U8*)STRING(scan);
4b88fb76 3511 uc = utf8_to_uvchr_buf(s, s + l, NULL);
8aa23a47 3512 l = utf8_length(s, s + l);
8aa23a47
YO
3513 } else {
3514 uc = *((U8*)STRING(scan));
3515 }
3516 min += l;
3517 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3518 /* The code below prefers earlier match for fixed
3519 offset, later match for variable offset. */
3520 if (data->last_end == -1) { /* Update the start info. */
3521 data->last_start_min = data->pos_min;
3522 data->last_start_max = is_inf
3523 ? I32_MAX : data->pos_min + data->pos_delta;
b515a41d 3524 }
8aa23a47
YO
3525 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3526 if (UTF)
3527 SvUTF8_on(data->last_found);
3528 {
3529 SV * const sv = data->last_found;
3530 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3531 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3532 if (mg && mg->mg_len >= 0)
3533 mg->mg_len += utf8_length((U8*)STRING(scan),
3534 (U8*)STRING(scan)+STR_LEN(scan));
b515a41d 3535 }
8aa23a47
YO
3536 data->last_end = data->pos_min + l;
3537 data->pos_min += l; /* As in the first entry. */
3538 data->flags &= ~SF_BEFORE_EOL;
3539 }
3540 if (flags & SCF_DO_STCLASS_AND) {
3541 /* Check whether it is compatible with what we know already! */
3542 int compat = 1;
3543
54251c2e 3544
486ec47a 3545 /* If compatible, we or it in below. It is compatible if is
54251c2e
KW
3546 * in the bitmp and either 1) its bit or its fold is set, or 2)
3547 * it's for a locale. Even if there isn't unicode semantics
3548 * here, at runtime there may be because of matching against a
3549 * utf8 string, so accept a possible false positive for
3550 * latin1-range folds */
8aa23a47
YO
3551 if (uc >= 0x100 ||
3552 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3553 && !ANYOF_BITMAP_TEST(data->start_class, uc)
538b546e 3554 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
54251c2e 3555 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
8aa23a47 3556 )
d18bf9dc 3557 {
8aa23a47 3558 compat = 0;
d18bf9dc 3559 }
8aa23a47
YO
3560 ANYOF_CLASS_ZERO(data->start_class);
3561 ANYOF_BITMAP_ZERO(data->start_class);
3562 if (compat)
3563 ANYOF_BITMAP_SET(data->start_class, uc);
d18bf9dc
KW
3564 else if (uc >= 0x100) {
3565 int i;
3566
3567 /* Some Unicode code points fold to the Latin1 range; as
3568 * XXX temporary code, instead of figuring out if this is
3569 * one, just assume it is and set all the start class bits
3570 * that could be some such above 255 code point's fold
3571 * which will generate fals positives. As the code
3572 * elsewhere that does compute the fold settles down, it
3573 * can be extracted out and re-used here */
3574 for (i = 0; i < 256; i++){
94dc5c2d 3575 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
d18bf9dc
KW
3576 ANYOF_BITMAP_SET(data->start_class, i);