This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
Remove redundant field from inversion lists
[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 */
f7c7e32a
DM
129 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
130 implies compiling, so don't emit */
131 regnode emit_dummy; /* placeholder for emit to point to */
830247a4
IZ
132 I32 naughty; /* How bad is this pattern? */
133 I32 sawback; /* Did we see \1, ...? */
134 U32 seen;
135 I32 size; /* Code size. */
c74340f9
YO
136 I32 npar; /* Capture buffer count, (OPEN). */
137 I32 cpar; /* Capture buffer count, (CLOSE). */
e2e6a0f1 138 I32 nestroot; /* root parens we are in - used by accept */
830247a4
IZ
139 I32 extralen;
140 I32 seen_zerolen;
40d049e4
YO
141 regnode **open_parens; /* pointers to open parens */
142 regnode **close_parens; /* pointers to close parens */
143 regnode *opend; /* END node in program */
02daf0ab
YO
144 I32 utf8; /* whether the pattern is utf8 or not */
145 I32 orig_utf8; /* whether the pattern was originally in utf8 */
146 /* XXX use this for future optimisation of case
147 * where pattern must be upgraded to utf8. */
e40e74fe
KW
148 I32 uni_semantics; /* If a d charset modifier should use unicode
149 rules, even if the pattern is not in
150 utf8 */
81714fb9 151 HV *paren_names; /* Paren names */
1f1031fe 152
40d049e4
YO
153 regnode **recurse; /* Recurse regops */
154 I32 recurse_count; /* Number of recurse regops */
b57e4118 155 I32 in_lookbehind;
4624b182 156 I32 contains_locale;
bb3f3ed2 157 I32 override_recoding;
9d53c457 158 I32 in_multi_char_class;
3d2bd50a 159 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 160 within pattern */
b1603ef8
DM
161 int num_code_blocks; /* size of code_blocks[] */
162 int code_index; /* next code_blocks[] slot */
830247a4
IZ
163#if ADD_TO_REGEXEC
164 char *starttry; /* -Dr: where regtry was called. */
165#define RExC_starttry (pRExC_state->starttry)
166#endif
d24ca0c5 167 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 168#ifdef DEBUGGING
be8e71aa 169 const char *lastparse;
3dab1dad 170 I32 lastnum;
1f1031fe 171 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
172#define RExC_lastparse (pRExC_state->lastparse)
173#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 174#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 175#endif
830247a4
IZ
176} RExC_state_t;
177
e2509266 178#define RExC_flags (pRExC_state->flags)
514a91f1 179#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 180#define RExC_precomp (pRExC_state->precomp)
288b8c02 181#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 182#define RExC_rx (pRExC_state->rx)
f8fc2ecf 183#define RExC_rxi (pRExC_state->rxi)
fac92740 184#define RExC_start (pRExC_state->start)
830247a4
IZ
185#define RExC_end (pRExC_state->end)
186#define RExC_parse (pRExC_state->parse)
187#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237
YO
188#ifdef RE_TRACK_PATTERN_OFFSETS
189#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
190#endif
830247a4 191#define RExC_emit (pRExC_state->emit)
f7c7e32a 192#define RExC_emit_dummy (pRExC_state->emit_dummy)
fac92740 193#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 194#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
195#define RExC_naughty (pRExC_state->naughty)
196#define RExC_sawback (pRExC_state->sawback)
197#define RExC_seen (pRExC_state->seen)
198#define RExC_size (pRExC_state->size)
199#define RExC_npar (pRExC_state->npar)
e2e6a0f1 200#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
201#define RExC_extralen (pRExC_state->extralen)
202#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 203#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 204#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 205#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
206#define RExC_open_parens (pRExC_state->open_parens)
207#define RExC_close_parens (pRExC_state->close_parens)
208#define RExC_opend (pRExC_state->opend)
81714fb9 209#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
210#define RExC_recurse (pRExC_state->recurse)
211#define RExC_recurse_count (pRExC_state->recurse_count)
b57e4118 212#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 213#define RExC_contains_locale (pRExC_state->contains_locale)
9d53c457
KW
214#define RExC_override_recoding (pRExC_state->override_recoding)
215#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
830247a4 216
cde0cee5 217
a687059c
LW
218#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
219#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
4d68ffa0 220 ((*s) == '{' && regcurly(s, FALSE)))
a687059c 221
35c8bce7
LW
222#ifdef SPSTART
223#undef SPSTART /* dratted cpp namespace... */
224#endif
a687059c
LW
225/*
226 * Flags to be passed up and down.
227 */
a687059c 228#define WORST 0 /* Worst case. */
a3b492c3 229#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 230
e64f369d 231/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
232 * character. (There needs to be a case: in the switch statement in regexec.c
233 * for any node marked SIMPLE.) Note that this is not the same thing as
234 * REGNODE_SIMPLE */
fda99bee 235#define SIMPLE 0x02
e64f369d 236#define SPSTART 0x04 /* Starts with * or + */
8d9c2815
NC
237#define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
238#define TRYAGAIN 0x10 /* Weeded out a declaration. */
239#define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
a687059c 240
3dab1dad
YO
241#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
242
07be1b83
YO
243/* whether trie related optimizations are enabled */
244#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
245#define TRIE_STUDY_OPT
786e8c11 246#define FULL_TRIE_STUDY
07be1b83
YO
247#define TRIE_STCLASS
248#endif
1de06328
YO
249
250
40d049e4
YO
251
252#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
253#define PBITVAL(paren) (1 << ((paren) & 7))
254#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
255#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
256#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
257
bbd61b5f 258#define REQUIRE_UTF8 STMT_START { \
8d9c2815
NC
259 if (!UTF) { \
260 *flagp = RESTART_UTF8; \
261 return NULL; \
262 } \
bbd61b5f 263 } STMT_END
40d049e4 264
f19b1a63
KW
265/* This converts the named class defined in regcomp.h to its equivalent class
266 * number defined in handy.h. */
267#define namedclass_to_classnum(class) ((int) ((class) / 2))
268#define classnum_to_namedclass(classnum) ((classnum) * 2)
269
1de06328
YO
270/* About scan_data_t.
271
272 During optimisation we recurse through the regexp program performing
273 various inplace (keyhole style) optimisations. In addition study_chunk
274 and scan_commit populate this data structure with information about
275 what strings MUST appear in the pattern. We look for the longest
3b753521 276 string that must appear at a fixed location, and we look for the
1de06328
YO
277 longest string that may appear at a floating location. So for instance
278 in the pattern:
279
280 /FOO[xX]A.*B[xX]BAR/
281
282 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
283 strings (because they follow a .* construct). study_chunk will identify
284 both FOO and BAR as being the longest fixed and floating strings respectively.
285
286 The strings can be composites, for instance
287
288 /(f)(o)(o)/
289
290 will result in a composite fixed substring 'foo'.
291
292 For each string some basic information is maintained:
293
294 - offset or min_offset
295 This is the position the string must appear at, or not before.
296 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
297 characters must match before the string we are searching for.
298 Likewise when combined with minlenp and the length of the string it
1de06328
YO
299 tells us how many characters must appear after the string we have
300 found.
301
302 - max_offset
303 Only used for floating strings. This is the rightmost point that
3b753521 304 the string can appear at. If set to I32 max it indicates that the
1de06328
YO
305 string can occur infinitely far to the right.
306
307 - minlenp
2d608413
KW
308 A pointer to the minimum number of characters of the pattern that the
309 string was found inside. This is important as in the case of positive
1de06328
YO
310 lookahead or positive lookbehind we can have multiple patterns
311 involved. Consider
312
313 /(?=FOO).*F/
314
315 The minimum length of the pattern overall is 3, the minimum length
316 of the lookahead part is 3, but the minimum length of the part that
317 will actually match is 1. So 'FOO's minimum length is 3, but the
318 minimum length for the F is 1. This is important as the minimum length
319 is used to determine offsets in front of and behind the string being
320 looked for. Since strings can be composites this is the length of the
486ec47a 321 pattern at the time it was committed with a scan_commit. Note that
1de06328
YO
322 the length is calculated by study_chunk, so that the minimum lengths
323 are not known until the full pattern has been compiled, thus the
324 pointer to the value.
325
326 - lookbehind
327
328 In the case of lookbehind the string being searched for can be
329 offset past the start point of the final matching string.
330 If this value was just blithely removed from the min_offset it would
331 invalidate some of the calculations for how many chars must match
332 before or after (as they are derived from min_offset and minlen and
333 the length of the string being searched for).
334 When the final pattern is compiled and the data is moved from the
335 scan_data_t structure into the regexp structure the information
336 about lookbehind is factored in, with the information that would
337 have been lost precalculated in the end_shift field for the
338 associated string.
339
340 The fields pos_min and pos_delta are used to store the minimum offset
341 and the delta to the maximum offset at the current point in the pattern.
342
343*/
2c2d71f5
JH
344
345typedef struct scan_data_t {
1de06328
YO
346 /*I32 len_min; unused */
347 /*I32 len_delta; unused */
2c2d71f5
JH
348 I32 pos_min;
349 I32 pos_delta;
350 SV *last_found;
1de06328 351 I32 last_end; /* min value, <0 unless valid. */
2c2d71f5
JH
352 I32 last_start_min;
353 I32 last_start_max;
1de06328
YO
354 SV **longest; /* Either &l_fixed, or &l_float. */
355 SV *longest_fixed; /* longest fixed string found in pattern */
356 I32 offset_fixed; /* offset where it starts */
486ec47a 357 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
358 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
359 SV *longest_float; /* longest floating string found in pattern */
360 I32 offset_float_min; /* earliest point in string it can appear */
361 I32 offset_float_max; /* latest point in string it can appear */
486ec47a 362 I32 *minlen_float; /* pointer to the minlen relevant to the string */
1de06328 363 I32 lookbehind_float; /* is the position of the string modified by LB */
2c2d71f5
JH
364 I32 flags;
365 I32 whilem_c;
cb434fcc 366 I32 *last_closep;
653099ff 367 struct regnode_charclass_class *start_class;
2c2d71f5
JH
368} scan_data_t;
369
a687059c 370/*
e50aee73 371 * Forward declarations for pregcomp()'s friends.
a687059c 372 */
a0d0e21e 373
27da23d5 374static const scan_data_t zero_scan_data =
1de06328 375 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
376
377#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
378#define SF_BEFORE_SEOL 0x0001
379#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
380#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
381#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
382
09b7f37c
CB
383#ifdef NO_UNARY_PLUS
384# define SF_FIX_SHIFT_EOL (0+2)
385# define SF_FL_SHIFT_EOL (0+4)
386#else
387# define SF_FIX_SHIFT_EOL (+2)
388# define SF_FL_SHIFT_EOL (+4)
389#endif
c277df42
IZ
390
391#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
392#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
393
394#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
395#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
396#define SF_IS_INF 0x0040
397#define SF_HAS_PAR 0x0080
398#define SF_IN_PAR 0x0100
399#define SF_HAS_EVAL 0x0200
400#define SCF_DO_SUBSTR 0x0400
653099ff
GS
401#define SCF_DO_STCLASS_AND 0x0800
402#define SCF_DO_STCLASS_OR 0x1000
403#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 404#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 405
786e8c11 406#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
e2e6a0f1 407#define SCF_SEEN_ACCEPT 0x8000
688e0391 408#define SCF_TRIE_DOING_RESTUDY 0x10000
07be1b83 409
43fead97 410#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
411
412/* The enums for all these are ordered so things work out correctly */
a62b1201 413#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
cfaf538b 414#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
00b27cfc 415#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
416#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
417#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1 418#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a725e29c 419#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 420
43fead97 421#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 422
93733859 423#define OOB_NAMEDCLASS -1
b8c5462f 424
8e661ac5
KW
425/* There is no code point that is out-of-bounds, so this is problematic. But
426 * its only current use is to initialize a variable that is always set before
427 * looked at. */
428#define OOB_UNICODE 0xDEADBEEF
429
a0ed51b3
LW
430#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
431#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
432
8615cb43 433
b45f050a
JF
434/* length of regex to show in messages that don't mark a position within */
435#define RegexLengthToShowInErrorMessages 127
436
437/*
438 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
439 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
440 * op/pragma/warn/regcomp.
441 */
7253e4e3
RK
442#define MARKER1 "<-- HERE" /* marker as it appears in the description */
443#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 444
7253e4e3 445#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
446
447/*
448 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
449 * arg. Show regex, up to a maximum length. If it's too long, chop and add
450 * "...".
451 */
58e23c8d 452#define _FAIL(code) STMT_START { \
bfed75c6 453 const char *ellipses = ""; \
ccb2c380
MP
454 IV len = RExC_end - RExC_precomp; \
455 \
456 if (!SIZE_ONLY) \
a5e7bc51 457 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
458 if (len > RegexLengthToShowInErrorMessages) { \
459 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
460 len = RegexLengthToShowInErrorMessages - 10; \
461 ellipses = "..."; \
462 } \
58e23c8d 463 code; \
ccb2c380 464} STMT_END
8615cb43 465
58e23c8d
YO
466#define FAIL(msg) _FAIL( \
467 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
468 msg, (int)len, RExC_precomp, ellipses))
469
470#define FAIL2(msg,arg) _FAIL( \
471 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
472 arg, (int)len, RExC_precomp, ellipses))
473
b45f050a 474/*
b45f050a
JF
475 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
476 */
ccb2c380 477#define Simple_vFAIL(m) STMT_START { \
a28509cc 478 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
479 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
480 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
481} STMT_END
b45f050a
JF
482
483/*
484 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
485 */
ccb2c380
MP
486#define vFAIL(m) STMT_START { \
487 if (!SIZE_ONLY) \
a5e7bc51 488 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
489 Simple_vFAIL(m); \
490} STMT_END
b45f050a
JF
491
492/*
493 * Like Simple_vFAIL(), but accepts two arguments.
494 */
ccb2c380 495#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 496 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
497 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
498 (int)offset, RExC_precomp, RExC_precomp + offset); \
499} STMT_END
b45f050a
JF
500
501/*
502 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
503 */
ccb2c380
MP
504#define vFAIL2(m,a1) STMT_START { \
505 if (!SIZE_ONLY) \
a5e7bc51 506 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
507 Simple_vFAIL2(m, a1); \
508} STMT_END
b45f050a
JF
509
510
511/*
512 * Like Simple_vFAIL(), but accepts three arguments.
513 */
ccb2c380 514#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 515 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
516 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
517 (int)offset, RExC_precomp, RExC_precomp + offset); \
518} STMT_END
b45f050a
JF
519
520/*
521 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
522 */
ccb2c380
MP
523#define vFAIL3(m,a1,a2) STMT_START { \
524 if (!SIZE_ONLY) \
a5e7bc51 525 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
526 Simple_vFAIL3(m, a1, a2); \
527} STMT_END
b45f050a
JF
528
529/*
530 * Like Simple_vFAIL(), but accepts four arguments.
531 */
ccb2c380 532#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 533 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
534 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
535 (int)offset, RExC_precomp, RExC_precomp + offset); \
536} STMT_END
b45f050a 537
95db3ffa
KW
538#define vFAIL4(m,a1,a2,a3) STMT_START { \
539 if (!SIZE_ONLY) \
540 SAVEFREESV(RExC_rx_sv); \
541 Simple_vFAIL4(m, a1, a2, a3); \
542} STMT_END
543
5e0a247b
KW
544/* m is not necessarily a "literal string", in this macro */
545#define reg_warn_non_literal_string(loc, m) STMT_START { \
546 const IV offset = loc - RExC_precomp; \
547 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
548 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
549} STMT_END
550
668c081a 551#define ckWARNreg(loc,m) STMT_START { \
a28509cc 552 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
553 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
554 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
555} STMT_END
556
0d6106aa
KW
557#define vWARN_dep(loc, m) STMT_START { \
558 const IV offset = loc - RExC_precomp; \
559 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
560 (int)offset, RExC_precomp, RExC_precomp + offset); \
561} STMT_END
562
147508a2
KW
563#define ckWARNdep(loc,m) STMT_START { \
564 const IV offset = loc - RExC_precomp; \
565 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
566 m REPORT_LOCATION, \
567 (int)offset, RExC_precomp, RExC_precomp + offset); \
568} STMT_END
569
668c081a 570#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 571 const IV offset = loc - RExC_precomp; \
d1d15184 572 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
573 m REPORT_LOCATION, \
574 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
575} STMT_END
576
b23eb183 577#define ckWARN2reg_d(loc,m, a1) STMT_START { \
2335b3d3 578 const IV offset = loc - RExC_precomp; \
b23eb183 579 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
2335b3d3
KW
580 m REPORT_LOCATION, \
581 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
582} STMT_END
583
668c081a 584#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 585 const IV offset = loc - RExC_precomp; \
668c081a 586 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
587 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
588} STMT_END
589
590#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 591 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
592 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
593 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
594} STMT_END
595
668c081a
NC
596#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
597 const IV offset = loc - RExC_precomp; \
598 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
599 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
600} STMT_END
601
ccb2c380 602#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 603 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
604 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
605 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
606} STMT_END
607
668c081a
NC
608#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
609 const IV offset = loc - RExC_precomp; \
610 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
611 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
612} STMT_END
613
ccb2c380 614#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 615 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
616 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
617 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
618} STMT_END
9d1d55b5 619
8615cb43 620
cd439c50 621/* Allow for side effects in s */
ccb2c380
MP
622#define REGC(c,s) STMT_START { \
623 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
624} STMT_END
cd439c50 625
fac92740
MJD
626/* Macros for recording node offsets. 20001227 mjd@plover.com
627 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
628 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
629 * Element 0 holds the number n.
07be1b83 630 * Position is 1 indexed.
fac92740 631 */
7122b237
YO
632#ifndef RE_TRACK_PATTERN_OFFSETS
633#define Set_Node_Offset_To_R(node,byte)
634#define Set_Node_Offset(node,byte)
635#define Set_Cur_Node_Offset
636#define Set_Node_Length_To_R(node,len)
637#define Set_Node_Length(node,len)
6a86c6ad 638#define Set_Node_Cur_Length(node,start)
7122b237
YO
639#define Node_Offset(n)
640#define Node_Length(n)
641#define Set_Node_Offset_Length(node,offset,len)
642#define ProgLen(ri) ri->u.proglen
643#define SetProgLen(ri,x) ri->u.proglen = x
644#else
645#define ProgLen(ri) ri->u.offsets[0]
646#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
647#define Set_Node_Offset_To_R(node,byte) STMT_START { \
648 if (! SIZE_ONLY) { \
649 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 650 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 651 if((node) < 0) { \
551405c4 652 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
653 } else { \
654 RExC_offsets[2*(node)-1] = (byte); \
655 } \
656 } \
657} STMT_END
658
659#define Set_Node_Offset(node,byte) \
660 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
661#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
662
663#define Set_Node_Length_To_R(node,len) STMT_START { \
664 if (! SIZE_ONLY) { \
665 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 666 __LINE__, (int)(node), (int)(len))); \
ccb2c380 667 if((node) < 0) { \
551405c4 668 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
669 } else { \
670 RExC_offsets[2*(node)] = (len); \
671 } \
672 } \
673} STMT_END
674
675#define Set_Node_Length(node,len) \
676 Set_Node_Length_To_R((node)-RExC_emit_start, len)
6a86c6ad
NC
677#define Set_Node_Cur_Length(node, start) \
678 Set_Node_Length(node, RExC_parse - start)
fac92740
MJD
679
680/* Get offsets and lengths */
681#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
682#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
683
07be1b83
YO
684#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
685 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
686 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
687} STMT_END
7122b237 688#endif
07be1b83
YO
689
690#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
691#define EXPERIMENTAL_INPLACESCAN
f427392e 692#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 693
304ee84b
YO
694#define DEBUG_STUDYDATA(str,data,depth) \
695DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 696 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
697 "%*s" str "Pos:%"IVdf"/%"IVdf \
698 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
699 (int)(depth)*2, "", \
700 (IV)((data)->pos_min), \
701 (IV)((data)->pos_delta), \
304ee84b 702 (UV)((data)->flags), \
1de06328 703 (IV)((data)->whilem_c), \
304ee84b
YO
704 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
705 is_inf ? "INF " : "" \
1de06328
YO
706 ); \
707 if ((data)->last_found) \
708 PerlIO_printf(Perl_debug_log, \
709 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
710 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
711 SvPVX_const((data)->last_found), \
712 (IV)((data)->last_end), \
713 (IV)((data)->last_start_min), \
714 (IV)((data)->last_start_max), \
715 ((data)->longest && \
716 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
717 SvPVX_const((data)->longest_fixed), \
718 (IV)((data)->offset_fixed), \
719 ((data)->longest && \
720 (data)->longest==&((data)->longest_float)) ? "*" : "", \
721 SvPVX_const((data)->longest_float), \
722 (IV)((data)->offset_float_min), \
723 (IV)((data)->offset_float_max) \
724 ); \
725 PerlIO_printf(Perl_debug_log,"\n"); \
726});
727
653099ff 728/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 729 Update the longest found anchored substring and the longest found
653099ff
GS
730 floating substrings if needed. */
731
4327152a 732STATIC void
304ee84b 733S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
c277df42 734{
e1ec3a88
AL
735 const STRLEN l = CHR_SVLEN(data->last_found);
736 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 737 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 738
7918f24d
NC
739 PERL_ARGS_ASSERT_SCAN_COMMIT;
740
c277df42 741 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 742 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
743 if (*data->longest == data->longest_fixed) {
744 data->offset_fixed = l ? data->last_start_min : data->pos_min;
745 if (data->flags & SF_BEFORE_EOL)
b81d288d 746 data->flags
c277df42
IZ
747 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
748 else
749 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 750 data->minlen_fixed=minlenp;
1de06328 751 data->lookbehind_fixed=0;
a0ed51b3 752 }
304ee84b 753 else { /* *data->longest == data->longest_float */
c277df42 754 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
755 data->offset_float_max = (l
756 ? data->last_start_max
9b139d09 757 : (data->pos_delta == I32_MAX ? I32_MAX : data->pos_min + data->pos_delta));
304ee84b 758 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
9051bda5 759 data->offset_float_max = I32_MAX;
c277df42 760 if (data->flags & SF_BEFORE_EOL)
b81d288d 761 data->flags
c277df42
IZ
762 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
763 else
764 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
765 data->minlen_float=minlenp;
766 data->lookbehind_float=0;
c277df42
IZ
767 }
768 }
769 SvCUR_set(data->last_found, 0);
0eda9292 770 {
a28509cc 771 SV * const sv = data->last_found;
097eb12c
AL
772 if (SvUTF8(sv) && SvMAGICAL(sv)) {
773 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
774 if (mg)
775 mg->mg_len = 0;
776 }
0eda9292 777 }
c277df42
IZ
778 data->last_end = -1;
779 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 780 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
781}
782
899d20b9
KW
783/* These macros set, clear and test whether the synthetic start class ('ssc',
784 * given by the parameter) matches an empty string (EOS). This uses the
785 * 'next_off' field in the node, to save a bit in the flags field. The ssc
786 * stands alone, so there is never a next_off, so this field is otherwise
787 * unused. The EOS information is used only for compilation, but theoretically
788 * it could be passed on to the execution code. This could be used to store
789 * more than one bit of information, but only this one is currently used. */
790#define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
791#define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
792#define TEST_SSC_EOS(node) cBOOL((node)->next_off)
793
653099ff
GS
794/* Can match anything (initialization) */
795STATIC void
3fffb88a 796S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 797{
7918f24d
NC
798 PERL_ARGS_ASSERT_CL_ANYTHING;
799
f8bef550 800 ANYOF_BITMAP_SETALL(cl);
899d20b9
KW
801 cl->flags = ANYOF_UNICODE_ALL;
802 SET_SSC_EOS(cl);
3fffb88a
KW
803
804 /* If any portion of the regex is to operate under locale rules,
805 * initialization includes it. The reason this isn't done for all regexes
806 * is that the optimizer was written under the assumption that locale was
807 * all-or-nothing. Given the complexity and lack of documentation in the
808 * optimizer, and that there are inadequate test cases for locale, so many
809 * parts of it may not work properly, it is safest to avoid locale unless
810 * necessary. */
811 if (RExC_contains_locale) {
9d7a1e63 812 ANYOF_CLASS_SETALL(cl); /* /l uses class */
b25a1036 813 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
3fffb88a 814 }
9d7a1e63
KW
815 else {
816 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
817 }
653099ff
GS
818}
819
820/* Can match anything (initialization) */
821STATIC int
5f66b61c 822S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
823{
824 int value;
825
7918f24d
NC
826 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
827
3732a889 828 for (value = 0; value < ANYOF_MAX; value += 2)
653099ff
GS
829 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
830 return 1;
1aa99e6b
IH
831 if (!(cl->flags & ANYOF_UNICODE_ALL))
832 return 0;
10edeb5d 833 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 834 return 0;
653099ff
GS
835 return 1;
836}
837
838/* Can match anything (initialization) */
839STATIC void
e755fd73 840S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 841{
7918f24d
NC
842 PERL_ARGS_ASSERT_CL_INIT;
843
8ecf7187 844 Zero(cl, 1, struct regnode_charclass_class);
653099ff 845 cl->type = ANYOF;
3fffb88a 846 cl_anything(pRExC_state, cl);
1411dba4 847 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
848}
849
1051e1c4
KW
850/* These two functions currently do the exact same thing */
851#define cl_init_zero S_cl_init
653099ff 852
dd58aee1
KW
853/* 'AND' a given class with another one. Can create false positives. 'cl'
854 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
855 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 856STATIC void
5f66b61c 857S_cl_and(struct regnode_charclass_class *cl,
a28509cc 858 const struct regnode_charclass_class *and_with)
653099ff 859{
7918f24d 860 PERL_ARGS_ASSERT_CL_AND;
40d049e4 861
954a2af6 862 assert(PL_regkind[and_with->type] == ANYOF);
1e6ade67 863
c6b76537 864 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
865 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
866 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 867 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
868 && !(and_with->flags & ANYOF_LOC_FOLD)
869 && !(cl->flags & ANYOF_LOC_FOLD)) {
653099ff
GS
870 int i;
871
872 if (and_with->flags & ANYOF_INVERT)
873 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
874 cl->bitmap[i] &= ~and_with->bitmap[i];
875 else
876 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
877 cl->bitmap[i] &= and_with->bitmap[i];
878 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 879
c6b76537 880 if (and_with->flags & ANYOF_INVERT) {
8951c461 881
c6b76537
KW
882 /* Here, the and'ed node is inverted. Get the AND of the flags that
883 * aren't affected by the inversion. Those that are affected are
884 * handled individually below */
885 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
886 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
887 cl->flags |= affected_flags;
888
889 /* We currently don't know how to deal with things that aren't in the
890 * bitmap, but we know that the intersection is no greater than what
891 * is already in cl, so let there be false positives that get sorted
892 * out after the synthetic start class succeeds, and the node is
893 * matched for real. */
894
895 /* The inversion of these two flags indicate that the resulting
896 * intersection doesn't have them */
897 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
898 cl->flags &= ~ANYOF_UNICODE_ALL;
899 }
c6b76537
KW
900 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
901 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 902 }
1aa99e6b 903 }
c6b76537 904 else { /* and'd node is not inverted */
3ad98780
KW
905 U8 outside_bitmap_but_not_utf8; /* Temp variable */
906
137165a6 907 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
908
909 /* Here 'and_with' doesn't match anything outside the bitmap
910 * (except possibly ANYOF_UNICODE_ALL), which means the
911 * intersection can't either, except for ANYOF_UNICODE_ALL, in
912 * which case we don't know what the intersection is, but it's no
913 * greater than what cl already has, so can just leave it alone,
914 * with possible false positives */
915 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
916 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 917 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 918 }
137165a6 919 }
c6b76537
KW
920 else if (! ANYOF_NONBITMAP(cl)) {
921
922 /* Here, 'and_with' does match something outside the bitmap, and cl
923 * doesn't have a list of things to match outside the bitmap. If
924 * cl can match all code points above 255, the intersection will
3ad98780
KW
925 * be those above-255 code points that 'and_with' matches. If cl
926 * can't match all Unicode code points, it means that it can't
927 * match anything outside the bitmap (since the 'if' that got us
928 * into this block tested for that), so we leave the bitmap empty.
929 */
c6b76537
KW
930 if (cl->flags & ANYOF_UNICODE_ALL) {
931 ARG_SET(cl, ARG(and_with));
3ad98780
KW
932
933 /* and_with's ARG may match things that don't require UTF8.
934 * And now cl's will too, in spite of this being an 'and'. See
935 * the comments below about the kludge */
936 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
937 }
938 }
939 else {
940 /* Here, both 'and_with' and cl match something outside the
941 * bitmap. Currently we do not do the intersection, so just match
942 * whatever cl had at the beginning. */
943 }
944
945
3ad98780
KW
946 /* Take the intersection of the two sets of flags. However, the
947 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
948 * kludge around the fact that this flag is not treated like the others
949 * which are initialized in cl_anything(). The way the optimizer works
950 * is that the synthetic start class (SSC) is initialized to match
951 * anything, and then the first time a real node is encountered, its
952 * values are AND'd with the SSC's with the result being the values of
953 * the real node. However, there are paths through the optimizer where
954 * the AND never gets called, so those initialized bits are set
955 * inappropriately, which is not usually a big deal, as they just cause
956 * false positives in the SSC, which will just mean a probably
957 * imperceptible slow down in execution. However this bit has a
958 * higher false positive consequence in that it can cause utf8.pm,
959 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
960 * bigger slowdown and also causes significant extra memory to be used.
961 * In order to prevent this, the code now takes a different tack. The
962 * bit isn't set unless some part of the regular expression needs it,
963 * but once set it won't get cleared. This means that these extra
964 * modules won't get loaded unless there was some path through the
965 * pattern that would have required them anyway, and so any false
966 * positives that occur by not ANDing them out when they could be
967 * aren't as severe as they would be if we treated this bit like all
968 * the others */
969 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
970 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 971 cl->flags &= and_with->flags;
3ad98780 972 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 973 }
653099ff
GS
974}
975
dd58aee1
KW
976/* 'OR' a given class with another one. Can create false positives. 'cl'
977 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
978 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 979STATIC void
3fffb88a 980S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 981{
7918f24d
NC
982 PERL_ARGS_ASSERT_CL_OR;
983
653099ff 984 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
985
986 /* Here, the or'd node is to be inverted. This means we take the
987 * complement of everything not in the bitmap, but currently we don't
988 * know what that is, so give up and match anything */
989 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 990 cl_anything(pRExC_state, cl);
c6b76537 991 }
653099ff
GS
992 /* We do not use
993 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
994 * <= (B1 | !B2) | (CL1 | !CL2)
995 * which is wasteful if CL2 is small, but we ignore CL2:
996 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
997 * XXXX Can we handle case-fold? Unclear:
998 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
999 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
1000 */
c6b76537 1001 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
1002 && !(or_with->flags & ANYOF_LOC_FOLD)
1003 && !(cl->flags & ANYOF_LOC_FOLD) ) {
653099ff
GS
1004 int i;
1005
1006 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1007 cl->bitmap[i] |= ~or_with->bitmap[i];
1008 } /* XXXX: logic is complicated otherwise */
1009 else {
3fffb88a 1010 cl_anything(pRExC_state, cl);
653099ff 1011 }
c6b76537
KW
1012
1013 /* And, we can just take the union of the flags that aren't affected
1014 * by the inversion */
1015 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
1016
1017 /* For the remaining flags:
1018 ANYOF_UNICODE_ALL and inverted means to not match anything above
1019 255, which means that the union with cl should just be
1020 what cl has in it, so can ignore this flag
1021 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
1022 is 127-255 to match them, but then invert that, so the
1023 union with cl should just be what cl has in it, so can
1024 ignore this flag
1025 */
1026 } else { /* 'or_with' is not inverted */
653099ff
GS
1027 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1028 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
1029 && (!(or_with->flags & ANYOF_LOC_FOLD)
1030 || (cl->flags & ANYOF_LOC_FOLD)) ) {
653099ff
GS
1031 int i;
1032
1033 /* OR char bitmap and class bitmap separately */
1034 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1035 cl->bitmap[i] |= or_with->bitmap[i];
c96939e4
KW
1036 if (or_with->flags & ANYOF_CLASS) {
1037 ANYOF_CLASS_OR(or_with, cl);
1038 }
653099ff
GS
1039 }
1040 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 1041 cl_anything(pRExC_state, cl);
653099ff 1042 }
9826f543 1043
c6b76537
KW
1044 if (ANYOF_NONBITMAP(or_with)) {
1045
1046 /* Use the added node's outside-the-bit-map match if there isn't a
1047 * conflict. If there is a conflict (both nodes match something
1048 * outside the bitmap, but what they match outside is not the same
1049 * pointer, and hence not easily compared until XXX we extend
1050 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
1051 * match everything outside the bitmap. If that stuff is all above
1052 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
1053 if (! ANYOF_NONBITMAP(cl)) {
1054 ARG_SET(cl, ARG(or_with));
1055 }
1056 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
1057
1058 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1059 cl_anything(pRExC_state, cl);
1060 }
1061 else {
1062 cl->flags |= ANYOF_UNICODE_ALL;
1063 }
c6b76537 1064 }
4c34a693 1065 }
0b9668ee
KW
1066
1067 /* Take the union */
1068 cl->flags |= or_with->flags;
1aa99e6b 1069 }
653099ff
GS
1070}
1071
a3621e74
YO
1072#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1073#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1074#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1075#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1076
3dab1dad
YO
1077
1078#ifdef DEBUGGING
07be1b83 1079/*
2b8b4781
NC
1080 dump_trie(trie,widecharmap,revcharmap)
1081 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1082 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1083
1084 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1085 The _interim_ variants are used for debugging the interim
1086 tables that are used to generate the final compressed
1087 representation which is what dump_trie expects.
1088
486ec47a 1089 Part of the reason for their existence is to provide a form
3dab1dad 1090 of documentation as to how the different representations function.
07be1b83
YO
1091
1092*/
3dab1dad
YO
1093
1094/*
3dab1dad
YO
1095 Dumps the final compressed table form of the trie to Perl_debug_log.
1096 Used for debugging make_trie().
1097*/
b9a59e08 1098
3dab1dad 1099STATIC void
2b8b4781
NC
1100S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1101 AV *revcharmap, U32 depth)
3dab1dad
YO
1102{
1103 U32 state;
ab3bbdeb 1104 SV *sv=sv_newmortal();
55eed653 1105 int colwidth= widecharmap ? 6 : 4;
2e64971a 1106 U16 word;
3dab1dad
YO
1107 GET_RE_DEBUG_FLAGS_DECL;
1108
7918f24d 1109 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1110
3dab1dad
YO
1111 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1112 (int)depth * 2 + 2,"",
1113 "Match","Base","Ofs" );
1114
1115 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1116 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1117 if ( tmp ) {
ab3bbdeb
YO
1118 PerlIO_printf( Perl_debug_log, "%*s",
1119 colwidth,
ddc5bc0f 1120 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1121 PL_colors[0], PL_colors[1],
1122 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1123 PERL_PV_ESCAPE_FIRSTCHAR
1124 )
1125 );
3dab1dad
YO
1126 }
1127 }
1128 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1129 (int)depth * 2 + 2,"");
1130
1131 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1132 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1133 PerlIO_printf( Perl_debug_log, "\n");
1134
1e2e3d02 1135 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1136 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1137
1138 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1139
1140 if ( trie->states[ state ].wordnum ) {
1141 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1142 } else {
1143 PerlIO_printf( Perl_debug_log, "%6s", "" );
1144 }
1145
1146 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1147
1148 if ( base ) {
1149 U32 ofs = 0;
1150
1151 while( ( base + ofs < trie->uniquecharcount ) ||
1152 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1153 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1154 ofs++;
1155
1156 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1157
1158 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1159 if ( ( base + ofs >= trie->uniquecharcount ) &&
1160 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1161 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1162 {
ab3bbdeb
YO
1163 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1164 colwidth,
3dab1dad
YO
1165 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1166 } else {
ab3bbdeb 1167 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1168 }
1169 }
1170
1171 PerlIO_printf( Perl_debug_log, "]");
1172
1173 }
1174 PerlIO_printf( Perl_debug_log, "\n" );
1175 }
2e64971a
DM
1176 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1177 for (word=1; word <= trie->wordcount; word++) {
1178 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1179 (int)word, (int)(trie->wordinfo[word].prev),
1180 (int)(trie->wordinfo[word].len));
1181 }
1182 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1183}
1184/*
3dab1dad
YO
1185 Dumps a fully constructed but uncompressed trie in list form.
1186 List tries normally only are used for construction when the number of
1187 possible chars (trie->uniquecharcount) is very high.
1188 Used for debugging make_trie().
1189*/
1190STATIC void
55eed653 1191S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1192 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1193 U32 depth)
3dab1dad
YO
1194{
1195 U32 state;
ab3bbdeb 1196 SV *sv=sv_newmortal();
55eed653 1197 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1198 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1199
1200 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1201
3dab1dad 1202 /* print out the table precompression. */
ab3bbdeb
YO
1203 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1204 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1205 "------:-----+-----------------\n" );
3dab1dad
YO
1206
1207 for( state=1 ; state < next_alloc ; state ++ ) {
1208 U16 charid;
1209
ab3bbdeb 1210 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1211 (int)depth * 2 + 2,"", (UV)state );
1212 if ( ! trie->states[ state ].wordnum ) {
1213 PerlIO_printf( Perl_debug_log, "%5s| ","");
1214 } else {
1215 PerlIO_printf( Perl_debug_log, "W%4x| ",
1216 trie->states[ state ].wordnum
1217 );
1218 }
1219 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1220 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1221 if ( tmp ) {
1222 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1223 colwidth,
ddc5bc0f 1224 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1225 PL_colors[0], PL_colors[1],
1226 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1227 PERL_PV_ESCAPE_FIRSTCHAR
1228 ) ,
1e2e3d02
YO
1229 TRIE_LIST_ITEM(state,charid).forid,
1230 (UV)TRIE_LIST_ITEM(state,charid).newstate
1231 );
1232 if (!(charid % 10))
664e119d
RGS
1233 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1234 (int)((depth * 2) + 14), "");
1e2e3d02 1235 }
ab3bbdeb
YO
1236 }
1237 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1238 }
1239}
1240
1241/*
3dab1dad
YO
1242 Dumps a fully constructed but uncompressed trie in table form.
1243 This is the normal DFA style state transition table, with a few
1244 twists to facilitate compression later.
1245 Used for debugging make_trie().
1246*/
1247STATIC void
55eed653 1248S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1249 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1250 U32 depth)
3dab1dad
YO
1251{
1252 U32 state;
1253 U16 charid;
ab3bbdeb 1254 SV *sv=sv_newmortal();
55eed653 1255 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1256 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1257
1258 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1259
1260 /*
1261 print out the table precompression so that we can do a visual check
1262 that they are identical.
1263 */
1264
1265 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1266
1267 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1268 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1269 if ( tmp ) {
ab3bbdeb
YO
1270 PerlIO_printf( Perl_debug_log, "%*s",
1271 colwidth,
ddc5bc0f 1272 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1273 PL_colors[0], PL_colors[1],
1274 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1275 PERL_PV_ESCAPE_FIRSTCHAR
1276 )
1277 );
3dab1dad
YO
1278 }
1279 }
1280
1281 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1282
1283 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1284 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1285 }
1286
1287 PerlIO_printf( Perl_debug_log, "\n" );
1288
1289 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1290
1291 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1292 (int)depth * 2 + 2,"",
1293 (UV)TRIE_NODENUM( state ) );
1294
1295 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1296 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1297 if (v)
1298 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1299 else
1300 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1301 }
1302 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1303 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1304 } else {
1305 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1306 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1307 }
1308 }
07be1b83 1309}
3dab1dad
YO
1310
1311#endif
1312
2e64971a 1313
786e8c11
YO
1314/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1315 startbranch: the first branch in the whole branch sequence
1316 first : start branch of sequence of branch-exact nodes.
1317 May be the same as startbranch
1318 last : Thing following the last branch.
1319 May be the same as tail.
1320 tail : item following the branch sequence
1321 count : words in the sequence
1322 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1323 depth : indent depth
3dab1dad 1324
786e8c11 1325Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1326
786e8c11
YO
1327A trie is an N'ary tree where the branches are determined by digital
1328decomposition of the key. IE, at the root node you look up the 1st character and
1329follow that branch repeat until you find the end of the branches. Nodes can be
1330marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1331
786e8c11 1332 /he|she|his|hers/
72f13be8 1333
786e8c11
YO
1334would convert into the following structure. Numbers represent states, letters
1335following numbers represent valid transitions on the letter from that state, if
1336the number is in square brackets it represents an accepting state, otherwise it
1337will be in parenthesis.
07be1b83 1338
786e8c11
YO
1339 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1340 | |
1341 | (2)
1342 | |
1343 (1) +-i->(6)-+-s->[7]
1344 |
1345 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1346
786e8c11
YO
1347 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1348
1349This shows that when matching against the string 'hers' we will begin at state 1
1350read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1351then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1352is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1353single traverse. We store a mapping from accepting to state to which word was
1354matched, and then when we have multiple possibilities we try to complete the
1355rest of the regex in the order in which they occured in the alternation.
1356
1357The only prior NFA like behaviour that would be changed by the TRIE support is
1358the silent ignoring of duplicate alternations which are of the form:
1359
1360 / (DUPE|DUPE) X? (?{ ... }) Y /x
1361
4b714af6 1362Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1363and without the optimisation. With the optimisations dupes will be silently
486ec47a 1364ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1365the following demonstrates:
1366
1367 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1368
1369which prints out 'word' three times, but
1370
1371 'words'=~/(word|word|word)(?{ print $1 })S/
1372
1373which doesnt print it out at all. This is due to other optimisations kicking in.
1374
1375Example of what happens on a structural level:
1376
486ec47a 1377The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1378
1379 1: CURLYM[1] {1,32767}(18)
1380 5: BRANCH(8)
1381 6: EXACT <ac>(16)
1382 8: BRANCH(11)
1383 9: EXACT <ad>(16)
1384 11: BRANCH(14)
1385 12: EXACT <ab>(16)
1386 16: SUCCEED(0)
1387 17: NOTHING(18)
1388 18: END(0)
1389
1390This would be optimizable with startbranch=5, first=5, last=16, tail=16
1391and should turn into:
1392
1393 1: CURLYM[1] {1,32767}(18)
1394 5: TRIE(16)
1395 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1396 <ac>
1397 <ad>
1398 <ab>
1399 16: SUCCEED(0)
1400 17: NOTHING(18)
1401 18: END(0)
1402
1403Cases where tail != last would be like /(?foo|bar)baz/:
1404
1405 1: BRANCH(4)
1406 2: EXACT <foo>(8)
1407 4: BRANCH(7)
1408 5: EXACT <bar>(8)
1409 7: TAIL(8)
1410 8: EXACT <baz>(10)
1411 10: END(0)
1412
1413which would be optimizable with startbranch=1, first=1, last=7, tail=8
1414and would end up looking like:
1415
1416 1: TRIE(8)
1417 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1418 <foo>
1419 <bar>
1420 7: TAIL(8)
1421 8: EXACT <baz>(10)
1422 10: END(0)
1423
1424 d = uvuni_to_utf8_flags(d, uv, 0);
1425
1426is the recommended Unicode-aware way of saying
1427
1428 *(d++) = uv;
1429*/
1430
fab2782b 1431#define TRIE_STORE_REVCHAR(val) \
786e8c11 1432 STMT_START { \
73031816 1433 if (UTF) { \
fab2782b 1434 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1435 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
fab2782b 1436 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
73031816
NC
1437 SvCUR_set(zlopp, kapow - flrbbbbb); \
1438 SvPOK_on(zlopp); \
1439 SvUTF8_on(zlopp); \
1440 av_push(revcharmap, zlopp); \
1441 } else { \
fab2782b 1442 char ooooff = (char)val; \
73031816
NC
1443 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1444 } \
1445 } STMT_END
786e8c11 1446
fab2782b
YO
1447#define TRIE_READ_CHAR STMT_START { \
1448 wordlen++; \
1449 if ( UTF ) { \
1450 /* if it is UTF then it is either already folded, or does not need folding */ \
1451 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1452 } \
1453 else if (folder == PL_fold_latin1) { \
1454 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1455 if ( foldlen > 0 ) { \
1456 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1457 foldlen -= len; \
1458 scan += len; \
1459 len = 0; \
1460 } else { \
1461 len = 1; \
51910141 1462 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, FOLD_FLAGS_FULL); \
fab2782b
YO
1463 skiplen = UNISKIP(uvc); \
1464 foldlen -= skiplen; \
1465 scan = foldbuf + skiplen; \
1466 } \
1467 } else { \
1468 /* raw data, will be folded later if needed */ \
1469 uvc = (U32)*uc; \
1470 len = 1; \
1471 } \
786e8c11
YO
1472} STMT_END
1473
1474
1475
1476#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1477 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1478 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1479 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1480 } \
1481 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1482 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1483 TRIE_LIST_CUR( state )++; \
1484} STMT_END
07be1b83 1485
786e8c11
YO
1486#define TRIE_LIST_NEW(state) STMT_START { \
1487 Newxz( trie->states[ state ].trans.list, \
1488 4, reg_trie_trans_le ); \
1489 TRIE_LIST_CUR( state ) = 1; \
1490 TRIE_LIST_LEN( state ) = 4; \
1491} STMT_END
07be1b83 1492
786e8c11
YO
1493#define TRIE_HANDLE_WORD(state) STMT_START { \
1494 U16 dupe= trie->states[ state ].wordnum; \
1495 regnode * const noper_next = regnext( noper ); \
1496 \
786e8c11
YO
1497 DEBUG_r({ \
1498 /* store the word for dumping */ \
1499 SV* tmp; \
1500 if (OP(noper) != NOTHING) \
740cce10 1501 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1502 else \
740cce10 1503 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1504 av_push( trie_words, tmp ); \
786e8c11
YO
1505 }); \
1506 \
1507 curword++; \
2e64971a
DM
1508 trie->wordinfo[curword].prev = 0; \
1509 trie->wordinfo[curword].len = wordlen; \
1510 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1511 \
1512 if ( noper_next < tail ) { \
1513 if (!trie->jump) \
c944940b 1514 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1515 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1516 if (!jumper) \
1517 jumper = noper_next; \
1518 if (!nextbranch) \
1519 nextbranch= regnext(cur); \
1520 } \
1521 \
1522 if ( dupe ) { \
2e64971a
DM
1523 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1524 /* chain, so that when the bits of chain are later */\
1525 /* linked together, the dups appear in the chain */\
1526 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1527 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1528 } else { \
1529 /* we haven't inserted this word yet. */ \
1530 trie->states[ state ].wordnum = curword; \
1531 } \
1532} STMT_END
07be1b83 1533
3dab1dad 1534
786e8c11
YO
1535#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1536 ( ( base + charid >= ucharcount \
1537 && base + charid < ubound \
1538 && state == trie->trans[ base - ucharcount + charid ].check \
1539 && trie->trans[ base - ucharcount + charid ].next ) \
1540 ? trie->trans[ base - ucharcount + charid ].next \
1541 : ( state==1 ? special : 0 ) \
1542 )
3dab1dad 1543
786e8c11
YO
1544#define MADE_TRIE 1
1545#define MADE_JUMP_TRIE 2
1546#define MADE_EXACT_TRIE 4
3dab1dad 1547
a3621e74 1548STATIC I32
786e8c11 1549S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1550{
27da23d5 1551 dVAR;
a3621e74
YO
1552 /* first pass, loop through and scan words */
1553 reg_trie_data *trie;
55eed653 1554 HV *widecharmap = NULL;
2b8b4781 1555 AV *revcharmap = newAV();
a3621e74 1556 regnode *cur;
9f7f3913 1557 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1558 STRLEN len = 0;
1559 UV uvc = 0;
1560 U16 curword = 0;
1561 U32 next_alloc = 0;
786e8c11
YO
1562 regnode *jumper = NULL;
1563 regnode *nextbranch = NULL;
7f69552c 1564 regnode *convert = NULL;
2e64971a 1565 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1566 /* we just use folder as a flag in utf8 */
1e696034 1567 const U8 * folder = NULL;
a3621e74 1568
2b8b4781
NC
1569#ifdef DEBUGGING
1570 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1571 AV *trie_words = NULL;
1572 /* along with revcharmap, this only used during construction but both are
1573 * useful during debugging so we store them in the struct when debugging.
8e11feef 1574 */
2b8b4781
NC
1575#else
1576 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1577 STRLEN trie_charcount=0;
3dab1dad 1578#endif
2b8b4781 1579 SV *re_trie_maxbuff;
a3621e74 1580 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1581
1582 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1583#ifndef DEBUGGING
1584 PERL_UNUSED_ARG(depth);
1585#endif
a3621e74 1586
1e696034 1587 switch (flags) {
79a81a6e 1588 case EXACT: break;
2f7f8cb1 1589 case EXACTFA:
fab2782b
YO
1590 case EXACTFU_SS:
1591 case EXACTFU_TRICKYFOLD:
1e696034
KW
1592 case EXACTFU: folder = PL_fold_latin1; break;
1593 case EXACTF: folder = PL_fold; break;
1594 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1595 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1596 }
1597
c944940b 1598 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1599 trie->refcount = 1;
3dab1dad 1600 trie->startstate = 1;
786e8c11 1601 trie->wordcount = word_count;
f8fc2ecf 1602 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1603 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1604 if (flags == EXACT)
c944940b 1605 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1606 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1607 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1608
a3621e74 1609 DEBUG_r({
2b8b4781 1610 trie_words = newAV();
a3621e74 1611 });
a3621e74 1612
0111c4fd 1613 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1614 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1615 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1616 }
df826430 1617 DEBUG_TRIE_COMPILE_r({
3dab1dad 1618 PerlIO_printf( Perl_debug_log,
786e8c11 1619 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1620 (int)depth * 2 + 2, "",
1621 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1622 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1623 (int)depth);
3dab1dad 1624 });
7f69552c
YO
1625
1626 /* Find the node we are going to overwrite */
1627 if ( first == startbranch && OP( last ) != BRANCH ) {
1628 /* whole branch chain */
1629 convert = first;
1630 } else {
1631 /* branch sub-chain */
1632 convert = NEXTOPER( first );
1633 }
1634
a3621e74
YO
1635 /* -- First loop and Setup --
1636
1637 We first traverse the branches and scan each word to determine if it
1638 contains widechars, and how many unique chars there are, this is
1639 important as we have to build a table with at least as many columns as we
1640 have unique chars.
1641
1642 We use an array of integers to represent the character codes 0..255
38a44b82 1643 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1644 native representation of the character value as the key and IV's for the
1645 coded index.
1646
1647 *TODO* If we keep track of how many times each character is used we can
1648 remap the columns so that the table compression later on is more
3b753521 1649 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1650 middle and the least common are on the outside. IMO this would be better
1651 than a most to least common mapping as theres a decent chance the most
1652 common letter will share a node with the least common, meaning the node
486ec47a 1653 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1654 case is when we have the least common nodes twice.
1655
1656 */
1657
a3621e74 1658 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1659 regnode *noper = NEXTOPER( cur );
e1ec3a88 1660 const U8 *uc = (U8*)STRING( noper );
df826430 1661 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1662 STRLEN foldlen = 0;
1663 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1664 STRLEN skiplen = 0;
2af232bd 1665 const U8 *scan = (U8*)NULL;
07be1b83 1666 U32 wordlen = 0; /* required init */
02daf0ab
YO
1667 STRLEN chars = 0;
1668 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1669
3dab1dad 1670 if (OP(noper) == NOTHING) {
df826430
YO
1671 regnode *noper_next= regnext(noper);
1672 if (noper_next != tail && OP(noper_next) == flags) {
1673 noper = noper_next;
1674 uc= (U8*)STRING(noper);
1675 e= uc + STR_LEN(noper);
1676 trie->minlen= STR_LEN(noper);
1677 } else {
1678 trie->minlen= 0;
1679 continue;
1680 }
3dab1dad 1681 }
df826430 1682
fab2782b 1683 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1684 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1685 regardless of encoding */
fab2782b
YO
1686 if (OP( noper ) == EXACTFU_SS) {
1687 /* false positives are ok, so just set this */
1688 TRIE_BITMAP_SET(trie,0xDF);
1689 }
1690 }
a3621e74 1691 for ( ; uc < e ; uc += len ) {
3dab1dad 1692 TRIE_CHARCOUNT(trie)++;
a3621e74 1693 TRIE_READ_CHAR;
3dab1dad 1694 chars++;
a3621e74 1695 if ( uvc < 256 ) {
fab2782b
YO
1696 if ( folder ) {
1697 U8 folded= folder[ (U8) uvc ];
1698 if ( !trie->charmap[ folded ] ) {
1699 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1700 TRIE_STORE_REVCHAR( folded );
1701 }
1702 }
a3621e74
YO
1703 if ( !trie->charmap[ uvc ] ) {
1704 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1705 TRIE_STORE_REVCHAR( uvc );
a3621e74 1706 }
02daf0ab 1707 if ( set_bit ) {
62012aee
KW
1708 /* store the codepoint in the bitmap, and its folded
1709 * equivalent. */
fab2782b 1710 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1711
1712 /* store the folded codepoint */
fab2782b 1713 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1714
1715 if ( !UTF ) {
1716 /* store first byte of utf8 representation of
acdf4139
KW
1717 variant codepoints */
1718 if (! UNI_IS_INVARIANT(uvc)) {
1719 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1720 }
1721 }
02daf0ab
YO
1722 set_bit = 0; /* We've done our bit :-) */
1723 }
a3621e74
YO
1724 } else {
1725 SV** svpp;
55eed653
NC
1726 if ( !widecharmap )
1727 widecharmap = newHV();
a3621e74 1728
55eed653 1729 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1730
1731 if ( !svpp )
e4584336 1732 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1733
1734 if ( !SvTRUE( *svpp ) ) {
1735 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1736 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1737 }
1738 }
1739 }
3dab1dad 1740 if( cur == first ) {
fab2782b
YO
1741 trie->minlen = chars;
1742 trie->maxlen = chars;
3dab1dad 1743 } else if (chars < trie->minlen) {
fab2782b 1744 trie->minlen = chars;
3dab1dad 1745 } else if (chars > trie->maxlen) {
fab2782b
YO
1746 trie->maxlen = chars;
1747 }
1748 if (OP( noper ) == EXACTFU_SS) {
1749 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1750 if (trie->minlen > 1)
1751 trie->minlen= 1;
1752 }
1753 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1754 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1755 * - We assume that any such sequence might match a 2 byte string */
1756 if (trie->minlen > 2 )
1757 trie->minlen= 2;
3dab1dad
YO
1758 }
1759
a3621e74
YO
1760 } /* end first pass */
1761 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1762 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1763 (int)depth * 2 + 2,"",
55eed653 1764 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1765 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1766 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1767 );
a3621e74
YO
1768
1769 /*
1770 We now know what we are dealing with in terms of unique chars and
1771 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1772 representation using a flat table will take. If it's over a reasonable
1773 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1774 conservative but potentially much slower representation using an array
1775 of lists.
1776
1777 At the end we convert both representations into the same compressed
1778 form that will be used in regexec.c for matching with. The latter
1779 is a form that cannot be used to construct with but has memory
1780 properties similar to the list form and access properties similar
1781 to the table form making it both suitable for fast searches and
1782 small enough that its feasable to store for the duration of a program.
1783
1784 See the comment in the code where the compressed table is produced
1785 inplace from the flat tabe representation for an explanation of how
1786 the compression works.
1787
1788 */
1789
1790
2e64971a
DM
1791 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1792 prev_states[1] = 0;
1793
3dab1dad 1794 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1795 /*
1796 Second Pass -- Array Of Lists Representation
1797
1798 Each state will be represented by a list of charid:state records
1799 (reg_trie_trans_le) the first such element holds the CUR and LEN
1800 points of the allocated array. (See defines above).
1801
1802 We build the initial structure using the lists, and then convert
1803 it into the compressed table form which allows faster lookups
1804 (but cant be modified once converted).
a3621e74
YO
1805 */
1806
a3621e74
YO
1807 STRLEN transcount = 1;
1808
1e2e3d02
YO
1809 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1810 "%*sCompiling trie using list compiler\n",
1811 (int)depth * 2 + 2, ""));
686b73d4 1812
c944940b
JH
1813 trie->states = (reg_trie_state *)
1814 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1815 sizeof(reg_trie_state) );
a3621e74
YO
1816 TRIE_LIST_NEW(1);
1817 next_alloc = 2;
1818
1819 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1820
df826430 1821 regnode *noper = NEXTOPER( cur );
c445ea15 1822 U8 *uc = (U8*)STRING( noper );
df826430 1823 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1824 U32 state = 1; /* required init */
1825 U16 charid = 0; /* sanity init */
1826 U8 *scan = (U8*)NULL; /* sanity init */
1827 STRLEN foldlen = 0; /* required init */
07be1b83 1828 U32 wordlen = 0; /* required init */
c445ea15 1829 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1830 STRLEN skiplen = 0;
c445ea15 1831
df826430
YO
1832 if (OP(noper) == NOTHING) {
1833 regnode *noper_next= regnext(noper);
1834 if (noper_next != tail && OP(noper_next) == flags) {
1835 noper = noper_next;
1836 uc= (U8*)STRING(noper);
1837 e= uc + STR_LEN(noper);
1838 }
1839 }
1840
3dab1dad 1841 if (OP(noper) != NOTHING) {
786e8c11 1842 for ( ; uc < e ; uc += len ) {
c445ea15 1843
786e8c11 1844 TRIE_READ_CHAR;
c445ea15 1845
786e8c11
YO
1846 if ( uvc < 256 ) {
1847 charid = trie->charmap[ uvc ];
c445ea15 1848 } else {
55eed653 1849 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1850 if ( !svpp ) {
1851 charid = 0;
1852 } else {
1853 charid=(U16)SvIV( *svpp );
1854 }
c445ea15 1855 }
786e8c11
YO
1856 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1857 if ( charid ) {
a3621e74 1858
786e8c11
YO
1859 U16 check;
1860 U32 newstate = 0;
a3621e74 1861
786e8c11
YO
1862 charid--;
1863 if ( !trie->states[ state ].trans.list ) {
1864 TRIE_LIST_NEW( state );
c445ea15 1865 }
786e8c11
YO
1866 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1867 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1868 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1869 break;
1870 }
1871 }
1872 if ( ! newstate ) {
1873 newstate = next_alloc++;
2e64971a 1874 prev_states[newstate] = state;
786e8c11
YO
1875 TRIE_LIST_PUSH( state, charid, newstate );
1876 transcount++;
1877 }
1878 state = newstate;
1879 } else {
1880 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1881 }
a28509cc 1882 }
c445ea15 1883 }
3dab1dad 1884 TRIE_HANDLE_WORD(state);
a3621e74
YO
1885
1886 } /* end second pass */
1887
1e2e3d02
YO
1888 /* next alloc is the NEXT state to be allocated */
1889 trie->statecount = next_alloc;
c944940b
JH
1890 trie->states = (reg_trie_state *)
1891 PerlMemShared_realloc( trie->states,
1892 next_alloc
1893 * sizeof(reg_trie_state) );
a3621e74 1894
3dab1dad 1895 /* and now dump it out before we compress it */
2b8b4781
NC
1896 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1897 revcharmap, next_alloc,
1898 depth+1)
1e2e3d02 1899 );
a3621e74 1900
c944940b
JH
1901 trie->trans = (reg_trie_trans *)
1902 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1903 {
1904 U32 state;
a3621e74
YO
1905 U32 tp = 0;
1906 U32 zp = 0;
1907
1908
1909 for( state=1 ; state < next_alloc ; state ++ ) {
1910 U32 base=0;
1911
1912 /*
1913 DEBUG_TRIE_COMPILE_MORE_r(
1914 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1915 );
1916 */
1917
1918 if (trie->states[state].trans.list) {
1919 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1920 U16 maxid=minid;
a28509cc 1921 U16 idx;
a3621e74
YO
1922
1923 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1924 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1925 if ( forid < minid ) {
1926 minid=forid;
1927 } else if ( forid > maxid ) {
1928 maxid=forid;
1929 }
a3621e74
YO
1930 }
1931 if ( transcount < tp + maxid - minid + 1) {
1932 transcount *= 2;
c944940b
JH
1933 trie->trans = (reg_trie_trans *)
1934 PerlMemShared_realloc( trie->trans,
446bd890
NC
1935 transcount
1936 * sizeof(reg_trie_trans) );
a3621e74
YO
1937 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1938 }
1939 base = trie->uniquecharcount + tp - minid;
1940 if ( maxid == minid ) {
1941 U32 set = 0;
1942 for ( ; zp < tp ; zp++ ) {
1943 if ( ! trie->trans[ zp ].next ) {
1944 base = trie->uniquecharcount + zp - minid;
1945 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1946 trie->trans[ zp ].check = state;
1947 set = 1;
1948 break;
1949 }
1950 }
1951 if ( !set ) {
1952 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1953 trie->trans[ tp ].check = state;
1954 tp++;
1955 zp = tp;
1956 }
1957 } else {
1958 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1959 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1960 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1961 trie->trans[ tid ].check = state;
1962 }
1963 tp += ( maxid - minid + 1 );
1964 }
1965 Safefree(trie->states[ state ].trans.list);
1966 }
1967 /*
1968 DEBUG_TRIE_COMPILE_MORE_r(
1969 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1970 );
1971 */
1972 trie->states[ state ].trans.base=base;
1973 }
cc601c31 1974 trie->lasttrans = tp + 1;
a3621e74
YO
1975 }
1976 } else {
1977 /*
1978 Second Pass -- Flat Table Representation.
1979
1980 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1981 We know that we will need Charcount+1 trans at most to store the data
1982 (one row per char at worst case) So we preallocate both structures
1983 assuming worst case.
1984
1985 We then construct the trie using only the .next slots of the entry
1986 structs.
1987
3b753521 1988 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1989 make compression both faster and easier by keeping track of how many non
1990 zero fields are in the node.
1991
1992 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1993 transition.
1994
1995 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1996 number representing the first entry of the node, and state as a
1997 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1998 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1999 are 2 entrys per node. eg:
2000
2001 A B A B
2002 1. 2 4 1. 3 7
2003 2. 0 3 3. 0 5
2004 3. 0 0 5. 0 0
2005 4. 0 0 7. 0 0
2006
2007 The table is internally in the right hand, idx form. However as we also
2008 have to deal with the states array which is indexed by nodenum we have to
2009 use TRIE_NODENUM() to convert.
2010
2011 */
1e2e3d02
YO
2012 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2013 "%*sCompiling trie using table compiler\n",
2014 (int)depth * 2 + 2, ""));
3dab1dad 2015
c944940b
JH
2016 trie->trans = (reg_trie_trans *)
2017 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2018 * trie->uniquecharcount + 1,
2019 sizeof(reg_trie_trans) );
2020 trie->states = (reg_trie_state *)
2021 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2022 sizeof(reg_trie_state) );
a3621e74
YO
2023 next_alloc = trie->uniquecharcount + 1;
2024
3dab1dad 2025
a3621e74
YO
2026 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2027
df826430 2028 regnode *noper = NEXTOPER( cur );
a28509cc 2029 const U8 *uc = (U8*)STRING( noper );
df826430 2030 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
2031
2032 U32 state = 1; /* required init */
2033
2034 U16 charid = 0; /* sanity init */
2035 U32 accept_state = 0; /* sanity init */
2036 U8 *scan = (U8*)NULL; /* sanity init */
2037
2038 STRLEN foldlen = 0; /* required init */
07be1b83 2039 U32 wordlen = 0; /* required init */
fab2782b 2040 STRLEN skiplen = 0;
a3621e74
YO
2041 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2042
df826430
YO
2043 if (OP(noper) == NOTHING) {
2044 regnode *noper_next= regnext(noper);
2045 if (noper_next != tail && OP(noper_next) == flags) {
2046 noper = noper_next;
2047 uc= (U8*)STRING(noper);
2048 e= uc + STR_LEN(noper);
2049 }
2050 }
fab2782b 2051
3dab1dad 2052 if ( OP(noper) != NOTHING ) {
786e8c11 2053 for ( ; uc < e ; uc += len ) {
a3621e74 2054
786e8c11 2055 TRIE_READ_CHAR;
a3621e74 2056
786e8c11
YO
2057 if ( uvc < 256 ) {
2058 charid = trie->charmap[ uvc ];
2059 } else {
55eed653 2060 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2061 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2062 }
786e8c11
YO
2063 if ( charid ) {
2064 charid--;
2065 if ( !trie->trans[ state + charid ].next ) {
2066 trie->trans[ state + charid ].next = next_alloc;
2067 trie->trans[ state ].check++;
2e64971a
DM
2068 prev_states[TRIE_NODENUM(next_alloc)]
2069 = TRIE_NODENUM(state);
786e8c11
YO
2070 next_alloc += trie->uniquecharcount;
2071 }
2072 state = trie->trans[ state + charid ].next;
2073 } else {
2074 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2075 }
2076 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2077 }
a3621e74 2078 }
3dab1dad
YO
2079 accept_state = TRIE_NODENUM( state );
2080 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2081
2082 } /* end second pass */
2083
3dab1dad 2084 /* and now dump it out before we compress it */
2b8b4781
NC
2085 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2086 revcharmap,
2087 next_alloc, depth+1));
a3621e74 2088
a3621e74
YO
2089 {
2090 /*
2091 * Inplace compress the table.*
2092
2093 For sparse data sets the table constructed by the trie algorithm will
2094 be mostly 0/FAIL transitions or to put it another way mostly empty.
2095 (Note that leaf nodes will not contain any transitions.)
2096
2097 This algorithm compresses the tables by eliminating most such
2098 transitions, at the cost of a modest bit of extra work during lookup:
2099
2100 - Each states[] entry contains a .base field which indicates the
2101 index in the state[] array wheres its transition data is stored.
2102
3b753521 2103 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2104
2105 - If .base is nonzero then charid is added to it to find an entry in
2106 the trans array.
2107
2108 -If trans[states[state].base+charid].check!=state then the
2109 transition is taken to be a 0/Fail transition. Thus if there are fail
2110 transitions at the front of the node then the .base offset will point
2111 somewhere inside the previous nodes data (or maybe even into a node
2112 even earlier), but the .check field determines if the transition is
2113 valid.
2114
786e8c11 2115 XXX - wrong maybe?
a3621e74 2116 The following process inplace converts the table to the compressed
3b753521 2117 table: We first do not compress the root node 1,and mark all its
a3621e74 2118 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2119 allows us to do a DFA construction from the compressed table later,
2120 and ensures that any .base pointers we calculate later are greater
2121 than 0.
a3621e74
YO
2122
2123 - We set 'pos' to indicate the first entry of the second node.
2124
2125 - We then iterate over the columns of the node, finding the first and
2126 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2127 and set the .check pointers accordingly, and advance pos
2128 appropriately and repreat for the next node. Note that when we copy
2129 the next pointers we have to convert them from the original
2130 NODEIDX form to NODENUM form as the former is not valid post
2131 compression.
2132
2133 - If a node has no transitions used we mark its base as 0 and do not
2134 advance the pos pointer.
2135
2136 - If a node only has one transition we use a second pointer into the
2137 structure to fill in allocated fail transitions from other states.
2138 This pointer is independent of the main pointer and scans forward
2139 looking for null transitions that are allocated to a state. When it
2140 finds one it writes the single transition into the "hole". If the
786e8c11 2141 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2142
2143 - Once compressed we can Renew/realloc the structures to release the
2144 excess space.
2145
2146 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2147 specifically Fig 3.47 and the associated pseudocode.
2148
2149 demq
2150 */
a3b680e6 2151 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2152 U32 state, charid;
a3621e74 2153 U32 pos = 0, zp=0;
1e2e3d02 2154 trie->statecount = laststate;
a3621e74
YO
2155
2156 for ( state = 1 ; state < laststate ; state++ ) {
2157 U8 flag = 0;
a28509cc
AL
2158 const U32 stateidx = TRIE_NODEIDX( state );
2159 const U32 o_used = trie->trans[ stateidx ].check;
2160 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2161 trie->trans[ stateidx ].check = 0;
2162
2163 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2164 if ( flag || trie->trans[ stateidx + charid ].next ) {
2165 if ( trie->trans[ stateidx + charid ].next ) {
2166 if (o_used == 1) {
2167 for ( ; zp < pos ; zp++ ) {
2168 if ( ! trie->trans[ zp ].next ) {
2169 break;
2170 }
2171 }
2172 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2173 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2174 trie->trans[ zp ].check = state;
2175 if ( ++zp > pos ) pos = zp;
2176 break;
2177 }
2178 used--;
2179 }
2180 if ( !flag ) {
2181 flag = 1;
2182 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2183 }
2184 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2185 trie->trans[ pos ].check = state;
2186 pos++;
2187 }
2188 }
2189 }
cc601c31 2190 trie->lasttrans = pos + 1;
c944940b
JH
2191 trie->states = (reg_trie_state *)
2192 PerlMemShared_realloc( trie->states, laststate
2193 * sizeof(reg_trie_state) );
a3621e74 2194 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2195 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2196 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2197 (int)depth * 2 + 2,"",
2198 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2199 (IV)next_alloc,
2200 (IV)pos,
a3621e74
YO
2201 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2202 );
2203
2204 } /* end table compress */
2205 }
1e2e3d02
YO
2206 DEBUG_TRIE_COMPILE_MORE_r(
2207 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2208 (int)depth * 2 + 2, "",
2209 (UV)trie->statecount,
2210 (UV)trie->lasttrans)
2211 );
cc601c31 2212 /* resize the trans array to remove unused space */
c944940b
JH
2213 trie->trans = (reg_trie_trans *)
2214 PerlMemShared_realloc( trie->trans, trie->lasttrans
2215 * sizeof(reg_trie_trans) );
a3621e74 2216
3b753521 2217 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2218 U8 nodetype =(U8)(flags & 0xFF);
2219 char *str=NULL;
786e8c11 2220
07be1b83 2221#ifdef DEBUGGING
e62cc96a 2222 regnode *optimize = NULL;
7122b237
YO
2223#ifdef RE_TRACK_PATTERN_OFFSETS
2224
b57a0404
JH
2225 U32 mjd_offset = 0;
2226 U32 mjd_nodelen = 0;
7122b237
YO
2227#endif /* RE_TRACK_PATTERN_OFFSETS */
2228#endif /* DEBUGGING */
a3621e74 2229 /*
3dab1dad
YO
2230 This means we convert either the first branch or the first Exact,
2231 depending on whether the thing following (in 'last') is a branch
2232 or not and whther first is the startbranch (ie is it a sub part of
2233 the alternation or is it the whole thing.)
3b753521 2234 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2235 the whole branch sequence, including the first.
a3621e74 2236 */
3dab1dad 2237 /* Find the node we are going to overwrite */
7f69552c 2238 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2239 /* branch sub-chain */
3dab1dad 2240 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2241#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2242 DEBUG_r({
2243 mjd_offset= Node_Offset((convert));
2244 mjd_nodelen= Node_Length((convert));
2245 });
7122b237 2246#endif
7f69552c 2247 /* whole branch chain */
7122b237
YO
2248 }
2249#ifdef RE_TRACK_PATTERN_OFFSETS
2250 else {
7f69552c
YO
2251 DEBUG_r({
2252 const regnode *nop = NEXTOPER( convert );
2253 mjd_offset= Node_Offset((nop));
2254 mjd_nodelen= Node_Length((nop));
2255 });
07be1b83
YO
2256 }
2257 DEBUG_OPTIMISE_r(
2258 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2259 (int)depth * 2 + 2, "",
786e8c11 2260 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2261 );
7122b237 2262#endif
3dab1dad
YO
2263 /* But first we check to see if there is a common prefix we can
2264 split out as an EXACT and put in front of the TRIE node. */
2265 trie->startstate= 1;
55eed653 2266 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2267 U32 state;
1e2e3d02 2268 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2269 U32 ofs = 0;
8e11feef
RGS
2270 I32 idx = -1;
2271 U32 count = 0;
2272 const U32 base = trie->states[ state ].trans.base;
a3621e74 2273
3dab1dad 2274 if ( trie->states[state].wordnum )
8e11feef 2275 count = 1;
a3621e74 2276
8e11feef 2277 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2278 if ( ( base + ofs >= trie->uniquecharcount ) &&
2279 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2280 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2281 {
3dab1dad 2282 if ( ++count > 1 ) {
2b8b4781 2283 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2284 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2285 if ( state == 1 ) break;
3dab1dad
YO
2286 if ( count == 2 ) {
2287 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2288 DEBUG_OPTIMISE_r(
8e11feef
RGS
2289 PerlIO_printf(Perl_debug_log,
2290 "%*sNew Start State=%"UVuf" Class: [",
2291 (int)depth * 2 + 2, "",
786e8c11 2292 (UV)state));
be8e71aa 2293 if (idx >= 0) {
2b8b4781 2294 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2295 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2296
3dab1dad 2297 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2298 if ( folder )
2299 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2300 DEBUG_OPTIMISE_r(
f1f66076 2301 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2302 );
8e11feef
RGS
2303 }
2304 }
2305 TRIE_BITMAP_SET(trie,*ch);
2306 if ( folder )
2307 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2308 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2309 }
2310 idx = ofs;
2311 }
3dab1dad
YO
2312 }
2313 if ( count == 1 ) {
2b8b4781 2314 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2315 STRLEN len;
2316 char *ch = SvPV( *tmp, len );
de734bd5
A
2317 DEBUG_OPTIMISE_r({
2318 SV *sv=sv_newmortal();
8e11feef
RGS
2319 PerlIO_printf( Perl_debug_log,
2320 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2321 (int)depth * 2 + 2, "",
de734bd5
A
2322 (UV)state, (UV)idx,
2323 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2324 PL_colors[0], PL_colors[1],
2325 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2326 PERL_PV_ESCAPE_FIRSTCHAR
2327 )
2328 );
2329 });
3dab1dad
YO
2330 if ( state==1 ) {
2331 OP( convert ) = nodetype;
2332 str=STRING(convert);
2333 STR_LEN(convert)=0;
2334 }
c490c714
YO
2335 STR_LEN(convert) += len;
2336 while (len--)
de734bd5 2337 *str++ = *ch++;
8e11feef 2338 } else {
f9049ba1 2339#ifdef DEBUGGING
8e11feef
RGS
2340 if (state>1)
2341 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2342#endif
8e11feef
RGS
2343 break;
2344 }
2345 }
2e64971a 2346 trie->prefixlen = (state-1);
3dab1dad 2347 if (str) {
8e11feef 2348 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2349 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2350 trie->startstate = state;
07be1b83
YO
2351 trie->minlen -= (state - 1);
2352 trie->maxlen -= (state - 1);
33809eae
JH
2353#ifdef DEBUGGING
2354 /* At least the UNICOS C compiler choked on this
2355 * being argument to DEBUG_r(), so let's just have
2356 * it right here. */
2357 if (
2358#ifdef PERL_EXT_RE_BUILD
2359 1
2360#else
2361 DEBUG_r_TEST
2362#endif
2363 ) {
2364 regnode *fix = convert;
2365 U32 word = trie->wordcount;
2366 mjd_nodelen++;
2367 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2368 while( ++fix < n ) {
2369 Set_Node_Offset_Length(fix, 0, 0);
2370 }
2371 while (word--) {
2372 SV ** const tmp = av_fetch( trie_words, word, 0 );
2373 if (tmp) {
2374 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2375 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2376 else
2377 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2378 }
2379 }
2380 }
2381#endif
8e11feef
RGS
2382 if (trie->maxlen) {
2383 convert = n;
2384 } else {
3dab1dad 2385 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2386 DEBUG_r(optimize= n);
3dab1dad
YO
2387 }
2388 }
2389 }
a5ca303d
YO
2390 if (!jumper)
2391 jumper = last;
3dab1dad 2392 if ( trie->maxlen ) {
8e11feef
RGS
2393 NEXT_OFF( convert ) = (U16)(tail - convert);
2394 ARG_SET( convert, data_slot );
786e8c11
YO
2395 /* Store the offset to the first unabsorbed branch in
2396 jump[0], which is otherwise unused by the jump logic.
2397 We use this when dumping a trie and during optimisation. */
2398 if (trie->jump)
7f69552c 2399 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2400
6c48061a
YO
2401 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2402 * and there is a bitmap
2403 * and the first "jump target" node we found leaves enough room
2404 * then convert the TRIE node into a TRIEC node, with the bitmap
2405 * embedded inline in the opcode - this is hypothetically faster.
2406 */
2407 if ( !trie->states[trie->startstate].wordnum
2408 && trie->bitmap
2409 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2410 {
2411 OP( convert ) = TRIEC;
2412 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2413 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2414 trie->bitmap= NULL;
2415 } else
2416 OP( convert ) = TRIE;
a3621e74 2417
3dab1dad
YO
2418 /* store the type in the flags */
2419 convert->flags = nodetype;
a5ca303d
YO
2420 DEBUG_r({
2421 optimize = convert
2422 + NODE_STEP_REGNODE
2423 + regarglen[ OP( convert ) ];
2424 });
2425 /* XXX We really should free up the resource in trie now,
2426 as we won't use them - (which resources?) dmq */
3dab1dad 2427 }
a3621e74 2428 /* needed for dumping*/
e62cc96a 2429 DEBUG_r(if (optimize) {
07be1b83 2430 regnode *opt = convert;
bcdf7404 2431
e62cc96a 2432 while ( ++opt < optimize) {
07be1b83
YO
2433 Set_Node_Offset_Length(opt,0,0);
2434 }
786e8c11
YO
2435 /*
2436 Try to clean up some of the debris left after the
2437 optimisation.
a3621e74 2438 */
786e8c11 2439 while( optimize < jumper ) {
07be1b83 2440 mjd_nodelen += Node_Length((optimize));
a3621e74 2441 OP( optimize ) = OPTIMIZED;
07be1b83 2442 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2443 optimize++;
2444 }
07be1b83 2445 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2446 });
2447 } /* end node insert */
2e64971a
DM
2448
2449 /* Finish populating the prev field of the wordinfo array. Walk back
2450 * from each accept state until we find another accept state, and if
2451 * so, point the first word's .prev field at the second word. If the
2452 * second already has a .prev field set, stop now. This will be the
2453 * case either if we've already processed that word's accept state,
3b753521
FN
2454 * or that state had multiple words, and the overspill words were
2455 * already linked up earlier.
2e64971a
DM
2456 */
2457 {
2458 U16 word;
2459 U32 state;
2460 U16 prev;
2461
2462 for (word=1; word <= trie->wordcount; word++) {
2463 prev = 0;
2464 if (trie->wordinfo[word].prev)
2465 continue;
2466 state = trie->wordinfo[word].accept;
2467 while (state) {
2468 state = prev_states[state];
2469 if (!state)
2470 break;
2471 prev = trie->states[state].wordnum;
2472 if (prev)
2473 break;
2474 }
2475 trie->wordinfo[word].prev = prev;
2476 }
2477 Safefree(prev_states);
2478 }
2479
2480
2481 /* and now dump out the compressed format */
2482 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2483
55eed653 2484 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2485#ifdef DEBUGGING
2486 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2487 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2488#else
03e70be4 2489 SvREFCNT_dec_NN(revcharmap);
07be1b83 2490#endif
786e8c11
YO
2491 return trie->jump
2492 ? MADE_JUMP_TRIE
2493 : trie->startstate>1
2494 ? MADE_EXACT_TRIE
2495 : MADE_TRIE;
2496}
2497
2498STATIC void
2499S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2500{
3b753521 2501/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2502
2503 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2504 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2505 ISBN 0-201-10088-6
2506
2507 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2508 suffix of the current state's 'word' that is also a proper prefix of another word in our
2509 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2510 the DFA not to have to restart after its tried and failed a word at a given point, it
2511 simply continues as though it had been matching the other word in the first place.
2512 Consider
2513 'abcdgu'=~/abcdefg|cdgu/
2514 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2515 fail, which would bring us to the state representing 'd' in the second word where we would
2516 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2517 */
2518 /* add a fail transition */
3251b653
NC
2519 const U32 trie_offset = ARG(source);
2520 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2521 U32 *q;
2522 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2523 const U32 numstates = trie->statecount;
786e8c11
YO
2524 const U32 ubound = trie->lasttrans + ucharcount;
2525 U32 q_read = 0;
2526 U32 q_write = 0;
2527 U32 charid;
2528 U32 base = trie->states[ 1 ].trans.base;
2529 U32 *fail;
2530 reg_ac_data *aho;
2531 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2532 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2533
2534 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2535#ifndef DEBUGGING
2536 PERL_UNUSED_ARG(depth);
2537#endif
2538
2539
2540 ARG_SET( stclass, data_slot );
c944940b 2541 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2542 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2543 aho->trie=trie_offset;
446bd890
NC
2544 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2545 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2546 Newxz( q, numstates, U32);
c944940b 2547 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2548 aho->refcount = 1;
2549 fail = aho->fail;
2550 /* initialize fail[0..1] to be 1 so that we always have
2551 a valid final fail state */
2552 fail[ 0 ] = fail[ 1 ] = 1;
2553
2554 for ( charid = 0; charid < ucharcount ; charid++ ) {
2555 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2556 if ( newstate ) {
2557 q[ q_write ] = newstate;
2558 /* set to point at the root */
2559 fail[ q[ q_write++ ] ]=1;
2560 }
2561 }
2562 while ( q_read < q_write) {
2563 const U32 cur = q[ q_read++ % numstates ];
2564 base = trie->states[ cur ].trans.base;
2565
2566 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2567 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2568 if (ch_state) {
2569 U32 fail_state = cur;
2570 U32 fail_base;
2571 do {
2572 fail_state = fail[ fail_state ];
2573 fail_base = aho->states[ fail_state ].trans.base;
2574 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2575
2576 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2577 fail[ ch_state ] = fail_state;
2578 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2579 {
2580 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2581 }
2582 q[ q_write++ % numstates] = ch_state;
2583 }
2584 }
2585 }
2586 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2587 when we fail in state 1, this allows us to use the
2588 charclass scan to find a valid start char. This is based on the principle
2589 that theres a good chance the string being searched contains lots of stuff
2590 that cant be a start char.
2591 */
2592 fail[ 0 ] = fail[ 1 ] = 0;
2593 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2594 PerlIO_printf(Perl_debug_log,
2595 "%*sStclass Failtable (%"UVuf" states): 0",
2596 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2597 );
786e8c11
YO
2598 for( q_read=1; q_read<numstates; q_read++ ) {
2599 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2600 }
2601 PerlIO_printf(Perl_debug_log, "\n");
2602 });
2603 Safefree(q);
2604 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2605}
2606
786e8c11 2607
a3621e74 2608/*
5d1c421c
JH
2609 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2610 * These need to be revisited when a newer toolchain becomes available.
2611 */
2612#if defined(__sparc64__) && defined(__GNUC__)
2613# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2614# undef SPARC64_GCC_WORKAROUND
2615# define SPARC64_GCC_WORKAROUND 1
2616# endif
2617#endif
2618
07be1b83 2619#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2620 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2621 SV * const mysv=sv_newmortal(); \
2622 regnode *Next = regnext(scan); \
2623 regprop(RExC_rx, mysv, scan); \
7f69552c 2624 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2625 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2626 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2627 }});
07be1b83 2628
1de06328 2629
bb914485 2630/* The below joins as many adjacent EXACTish nodes as possible into a single
0a982f06
KW
2631 * one. The regop may be changed if the node(s) contain certain sequences that
2632 * require special handling. The joining is only done if:
bb914485
KW
2633 * 1) there is room in the current conglomerated node to entirely contain the
2634 * next one.
2635 * 2) they are the exact same node type
2636 *
87b8b349 2637 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
2638 * these get optimized out
2639 *
0a982f06
KW
2640 * If a node is to match under /i (folded), the number of characters it matches
2641 * can be different than its character length if it contains a multi-character
2642 * fold. *min_subtract is set to the total delta of the input nodes.
bb914485 2643 *
a0c4c608
KW
2644 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2645 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2646 *
bb914485 2647 * This is as good a place as any to discuss the design of handling these
0a982f06
KW
2648 * multi-character fold sequences. It's been wrong in Perl for a very long
2649 * time. There are three code points in Unicode whose multi-character folds
2650 * were long ago discovered to mess things up. The previous designs for
2651 * dealing with these involved assigning a special node for them. This
2652 * approach doesn't work, as evidenced by this example:
a0c4c608 2653 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
0a982f06
KW
2654 * Both these fold to "sss", but if the pattern is parsed to create a node that
2655 * would match just the \xDF, it won't be able to handle the case where a
bb914485
KW
2656 * successful match would have to cross the node's boundary. The new approach
2657 * that hopefully generally solves the problem generates an EXACTFU_SS node
2658 * that is "sss".
2659 *
0a982f06
KW
2660 * It turns out that there are problems with all multi-character folds, and not
2661 * just these three. Now the code is general, for all such cases, but the
2662 * three still have some special handling. The approach taken is:
2663 * 1) This routine examines each EXACTFish node that could contain multi-
2664 * character fold sequences. It returns in *min_subtract how much to
9d071ca8 2665 * subtract from the the actual length of the string to get a real minimum
0a982f06
KW
2666 * match length; it is 0 if there are no multi-char folds. This delta is
2667 * used by the caller to adjust the min length of the match, and the delta
2668 * between min and max, so that the optimizer doesn't reject these
2669 * possibilities based on size constraints.
2670 * 2) Certain of these sequences require special handling by the trie code,
2671 * so, if found, this code changes the joined node type to special ops:
2672 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2673 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2674 * is used for an EXACTFU node that contains at least one "ss" sequence in
2675 * it. For non-UTF-8 patterns and strings, this is the only case where
2676 * there is a possible fold length change. That means that a regular
2677 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2678 * with length changes, and so can be processed faster. regexec.c takes
2679 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2680 * pre-folded by regcomp.c. This saves effort in regex matching.
87b8b349 2681 * However, the pre-folding isn't done for non-UTF8 patterns because the
0a982f06
KW
2682 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2683 * down by forcing the pattern into UTF8 unless necessary. Also what
2684 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2685 * possibilities for the non-UTF8 patterns are quite simple, except for
2686 * the sharp s. All the ones that don't involve a UTF-8 target string are
2687 * members of a fold-pair, and arrays are set up for all of them so that
2688 * the other member of the pair can be found quickly. Code elsewhere in
2689 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2690 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2691 * described in the next item.
1ca267a5
KW
2692 * 4) A problem remains for the sharp s in EXACTF and EXACTFA nodes when the
2693 * pattern isn't in UTF-8. (BTW, there cannot be an EXACTF node with a
2694 * UTF-8 pattern.) An assumption that the optimizer part of regexec.c
2695 * (probably unwittingly, in Perl_regexec_flags()) makes is that a
2696 * character in the pattern corresponds to at most a single character in
2697 * the target string. (And I do mean character, and not byte here, unlike
2698 * other parts of the documentation that have never been updated to
2699 * account for multibyte Unicode.) sharp s in EXACTF nodes can match the
2700 * two character string 'ss'; in EXACTFA nodes it can match
2701 * "\x{17F}\x{17F}". These violate the assumption, and they are the only
2702 * instances where it is violated. I'm reluctant to try to change the
2703 * assumption, as the code involved is impenetrable to me (khw), so
2704 * instead the code here punts. This routine examines (when the pattern
2705 * isn't UTF-8) EXACTF and EXACTFA nodes for the sharp s, and returns a
2706 * boolean indicating whether or not the node contains a sharp s. When it
2707 * is true, the caller sets a flag that later causes the optimizer in this
2708 * file to not set values for the floating and fixed string lengths, and
2709 * thus avoids the optimizer code in regexec.c that makes the invalid
2710 * assumption. Thus, there is no optimization based on string lengths for
2711 * non-UTF8-pattern EXACTF and EXACTFA nodes that contain the sharp s.
2712 * (The reason the assumption is wrong only in these two cases is that all
2713 * other non-UTF-8 folds are 1-1; and, for UTF-8 patterns, we pre-fold all
2714 * other folds to their expanded versions. We can't prefold sharp s to
2715 * 'ss' in EXACTF nodes because we don't know at compile time if it
2716 * actually matches 'ss' or not. It will match iff the target string is
2717 * in UTF-8, unlike the EXACTFU nodes, where it always matches; and
2718 * EXACTFA and EXACTFL where it never does. In an EXACTFA node in a UTF-8
2719 * pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the problem;
2720 * but in a non-UTF8 pattern, folding it to that above-Latin1 string would
2721 * require the pattern to be forced into UTF-8, the overhead of which we
2722 * want to avoid.)
bb914485 2723 */
1de06328 2724
9d071ca8 2725#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2726 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2727 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2728
be8e71aa 2729STATIC U32
9d071ca8 2730S_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
2731 /* Merge several consecutive EXACTish nodes into one. */
2732 regnode *n = regnext(scan);
2733 U32 stringok = 1;
2734 regnode *next = scan + NODE_SZ_STR(scan);
2735 U32 merged = 0;
2736 U32 stopnow = 0;
2737#ifdef DEBUGGING
2738 regnode *stop = scan;
72f13be8 2739 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2740#else
d47053eb
RGS
2741 PERL_UNUSED_ARG(depth);
2742#endif
7918f24d
NC
2743
2744 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2745#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2746 PERL_UNUSED_ARG(flags);
2747 PERL_UNUSED_ARG(val);
07be1b83 2748#endif
07be1b83 2749 DEBUG_PEEP("join",scan,depth);
bb914485 2750
3f410cf6
KW
2751 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2752 * EXACT ones that are mergeable to the current one. */
2753 while (n
2754 && (PL_regkind[OP(n)] == NOTHING
2755 || (stringok && OP(n) == OP(scan)))
07be1b83 2756 && NEXT_OFF(n)
3f410cf6
KW
2757 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2758 {
07be1b83
YO
2759
2760 if (OP(n) == TAIL || n > next)
2761 stringok = 0;
2762 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2763 DEBUG_PEEP("skip:",n,depth);
2764 NEXT_OFF(scan) += NEXT_OFF(n);
2765 next = n + NODE_STEP_REGNODE;
2766#ifdef DEBUGGING
2767 if (stringok)
2768 stop = n;
2769#endif
2770 n = regnext(n);
2771 }
2772 else if (stringok) {
786e8c11 2773 const unsigned int oldl = STR_LEN(scan);
07be1b83 2774 regnode * const nnext = regnext(n);
b2230d39 2775
87b8b349
KW
2776 /* XXX I (khw) kind of doubt that this works on platforms where
2777 * U8_MAX is above 255 because of lots of other assumptions */
79a81a6e 2778 /* Don't join if the sum can't fit into a single node */
b2230d39
KW
2779 if (oldl + STR_LEN(n) > U8_MAX)
2780 break;
07be1b83
YO
2781
2782 DEBUG_PEEP("merg",n,depth);
07be1b83 2783 merged++;
b2230d39 2784
07be1b83
YO
2785 NEXT_OFF(scan) += NEXT_OFF(n);
2786 STR_LEN(scan) += STR_LEN(n);
2787 next = n + NODE_SZ_STR(n);
2788 /* Now we can overwrite *n : */
2789 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2790#ifdef DEBUGGING
2791 stop = next - 1;
2792#endif
2793 n = nnext;
2794 if (stopnow) break;
2795 }
2796
d47053eb
RGS
2797#ifdef EXPERIMENTAL_INPLACESCAN
2798 if (flags && !NEXT_OFF(n)) {
2799 DEBUG_PEEP("atch", val, depth);
2800 if (reg_off_by_arg[OP(n)]) {
2801 ARG_SET(n, val - n);
2802 }
2803 else {
2804 NEXT_OFF(n) = val - n;
2805 }
2806 stopnow = 1;
2807 }
07be1b83
YO
2808#endif
2809 }
2c2b7f86 2810
9d071ca8 2811 *min_subtract = 0;
f758bddf 2812 *has_exactf_sharp_s = FALSE;
f646642f 2813
3f410cf6
KW
2814 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2815 * can now analyze for sequences of problematic code points. (Prior to
2816 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2817 * hence missed). The sequences only happen in folding, hence for any
2818 * non-EXACT EXACTish node */
86d6fcad 2819 if (OP(scan) != EXACT) {
0a982f06
KW
2820 const U8 * const s0 = (U8*) STRING(scan);
2821 const U8 * s = s0;
2822 const U8 * const s_end = s0 + STR_LEN(scan);
f758bddf
KW
2823
2824 /* One pass is made over the node's string looking for all the
2825 * possibilities. to avoid some tests in the loop, there are two main
2826 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2827 * non-UTF-8 */
2828 if (UTF) {
86d6fcad 2829
0a982f06
KW
2830 /* Examine the string for a multi-character fold sequence. UTF-8
2831 * patterns have all characters pre-folded by the time this code is
2832 * executed */
2833 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2834 length sequence we are looking for is 2 */
86d6fcad 2835 {
0a982f06
KW
2836 int count = 0;
2837 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2838 if (! len) { /* Not a multi-char fold: get next char */
2839 s += UTF8SKIP(s);
2840 continue;
2841 }
bb914485 2842
0a982f06
KW
2843 /* Nodes with 'ss' require special handling, except for EXACTFL
2844 * and EXACTFA for which there is no multi-char fold to this */
2845 if (len == 2 && *s == 's' && *(s+1) == 's'
2846 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2847 {
2848 count = 2;
2849 OP(scan) = EXACTFU_SS;
2850 s += 2;
2851 }
accd961f
KW
2852 else if (len == 6 /* len is the same in both ASCII and EBCDIC
2853 for these */
0a982f06
KW
2854 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2855 COMBINING_DIAERESIS_UTF8
2856 COMBINING_ACUTE_ACCENT_UTF8,
2857 6)
2858 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2859 COMBINING_DIAERESIS_UTF8
2860 COMBINING_ACUTE_ACCENT_UTF8,
2861 6)))
2862 {
2863 count = 3;
2864
2865 /* These two folds require special handling by trie's, so
2866 * change the node type to indicate this. If EXACTFA and
2867 * EXACTFL were ever to be handled by trie's, this would
2868 * have to be changed. If this node has already been
2869 * changed to EXACTFU_SS in this loop, leave it as is. (I
2870 * (khw) think it doesn't matter in regexec.c for UTF
2871 * patterns, but no need to change it */
2872 if (OP(scan) == EXACTFU) {
2873 OP(scan) = EXACTFU_TRICKYFOLD;
2874 }
2875 s += 6;
2876 }
2877 else { /* Here is a generic multi-char fold. */
2878 const U8* multi_end = s + len;
2879
2880 /* Count how many characters in it. In the case of /l and
2881 * /aa, no folds which contain ASCII code points are
2882 * allowed, so check for those, and skip if found. (In
2883 * EXACTFL, no folds are allowed to any Latin1 code point,
2884 * not just ASCII. But there aren't any of these
2885 * currently, nor ever likely, so don't take the time to
2886 * test for them. The code that generates the
2887 * is_MULTI_foo() macros croaks should one actually get put
2888 * into Unicode .) */
2889 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2890 count = utf8_length(s, multi_end);
2891 s = multi_end;
2892 }
2893 else {
2894 while (s < multi_end) {
2895 if (isASCII(*s)) {
2896 s++;
2897 goto next_iteration;
2898 }
2899 else {
2900 s += UTF8SKIP(s);
2901 }
2902 count++;
2903 }
2904 }
2905 }
f758bddf 2906
0a982f06
KW
2907 /* The delta is how long the sequence is minus 1 (1 is how long
2908 * the character that folds to the sequence is) */
2909 *min_subtract += count - 1;
2910 next_iteration: ;
bb914485
KW
2911 }
2912 }
1ca267a5
KW
2913 else if (OP(scan) == EXACTFA) {
2914
2915 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
2916 * fold to the ASCII range (and there are no existing ones in the
2917 * upper latin1 range). But, as outlined in the comments preceding
2918 * this function, we need to flag any occurrences of the sharp s */
2919 while (s < s_end) {
2920 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
2921 *has_exactf_sharp_s = TRUE;
2922 break;
2923 }
2924 s++;
2925 continue;
2926 }
2927 }
2928 else if (OP(scan) != EXACTFL) {
2929
2930 /* Non-UTF-8 pattern, not EXACTFA nor EXACTFL node. Look for the
2931 * multi-char folds that are all Latin1. (This code knows that
2932 * there are no current multi-char folds possible with EXACTFL,
2933 * relying on fold_grind.t to catch any errors if the very unlikely
2934 * event happens that some get added in future Unicode versions.)
2935 * As explained in the comments preceding this function, we look
2936 * also for the sharp s in EXACTF nodes; it can be in the final
0a982f06
KW
2937 * position. Otherwise we can stop looking 1 byte earlier because
2938 * have to find at least two characters for a multi-fold */
f758bddf
KW
2939 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2940
0a982f06
KW
2941 /* The below is perhaps overboard, but this allows us to save a
2942 * test each time through the loop at the expense of a mask. This
2943 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2944 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2945 * are 64. This uses an exclusive 'or' to find that bit and then
2946 * inverts it to form a mask, with just a single 0, in the bit
2947 * position where 'S' and 's' differ. */
2948 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2949 const U8 s_masked = 's' & S_or_s_mask;
2950
2951 while (s < upper) {
40b1ba4f 2952 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
0a982f06
KW
2953 if (! len) { /* Not a multi-char fold. */
2954 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2955 {
2956 *has_exactf_sharp_s = TRUE;
2957 }
2958 s++;
2959 continue;
2960 }
2961
2962 if (len == 2
2963 && ((*s & S_or_s_mask) == s_masked)
2964 && ((*(s+1) & S_or_s_mask) == s_masked))
2965 {
2966
2967 /* EXACTF nodes need to know that the minimum length
2968 * changed so that a sharp s in the string can match this
2969 * ss in the pattern, but they remain EXACTF nodes, as they
2970 * won't match this unless the target string is is UTF-8,
2971 * which we don't know until runtime */
2972 if (OP(scan) != EXACTF) {
2973 OP(scan) = EXACTFU_SS;
2974 }
86d6fcad 2975 }
0a982f06
KW
2976
2977 *min_subtract += len - 1;
2978 s += len;
86d6fcad
KW
2979 }
2980 }
07be1b83 2981 }
3f410cf6 2982
07be1b83 2983#ifdef DEBUGGING
bb789b09
DM
2984 /* Allow dumping but overwriting the collection of skipped
2985 * ops and/or strings with fake optimized ops */
07be1b83
YO
2986 n = scan + NODE_SZ_STR(scan);
2987 while (n <= stop) {
bb789b09
DM
2988 OP(n) = OPTIMIZED;
2989 FLAGS(n) = 0;
2990 NEXT_OFF(n) = 0;
07be1b83
YO
2991 n++;
2992 }
2993#endif
2994 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2995 return stopnow;
2996}
2997
486ec47a 2998/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2999 Finds fixed substrings. */
3000
a0288114 3001/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
3002 to the position after last scanned or to NULL. */
3003
40d049e4
YO
3004#define INIT_AND_WITHP \
3005 assert(!and_withp); \
3006 Newx(and_withp,1,struct regnode_charclass_class); \
3007 SAVEFREEPV(and_withp)
07be1b83 3008
b515a41d 3009/* this is a chain of data about sub patterns we are processing that
486ec47a 3010 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
3011 we can simulate recursion without losing state. */
3012struct scan_frame;
3013typedef struct scan_frame {
3014 regnode *last; /* last node to process in this frame */
3015 regnode *next; /* next node to process when last is reached */
3016 struct scan_frame *prev; /*previous frame*/
3017 I32 stop; /* what stopparen do we use */
3018} scan_frame;
3019
304ee84b
YO
3020
3021#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
3022
76e3520e 3023STATIC I32
40d049e4 3024S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 3025 I32 *minlenp, I32 *deltap,
40d049e4
YO
3026 regnode *last,
3027 scan_data_t *data,
3028 I32 stopparen,
3029 U8* recursed,
3030 struct regnode_charclass_class *and_withp,
3031 U32 flags, U32 depth)
c277df42
IZ
3032 /* scanp: Start here (read-write). */
3033 /* deltap: Write maxlen-minlen here. */
3034 /* last: Stop before this one. */
40d049e4
YO
3035 /* data: string data about the pattern */
3036 /* stopparen: treat close N as END */
3037 /* recursed: which subroutines have we recursed into */
3038 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 3039{
97aff369 3040 dVAR;
2d608413
KW
3041 I32 min = 0; /* There must be at least this number of characters to match */
3042 I32 pars = 0, code;
c277df42
IZ
3043 regnode *scan = *scanp, *next;
3044 I32 delta = 0;
3045 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 3046 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
3047 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3048 scan_data_t data_fake;
a3621e74 3049 SV *re_trie_maxbuff = NULL;
786e8c11 3050 regnode *first_non_open = scan;
e2e6a0f1 3051 I32 stopmin = I32_MAX;
8aa23a47 3052 scan_frame *frame = NULL;
a3621e74 3053 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 3054
7918f24d
NC
3055 PERL_ARGS_ASSERT_STUDY_CHUNK;
3056
13a24bad 3057#ifdef DEBUGGING
40d049e4 3058 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 3059#endif
40d049e4 3060
786e8c11 3061 if ( depth == 0 ) {
40d049e4 3062 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
3063 first_non_open=regnext(first_non_open);
3064 }
3065
b81d288d 3066
8aa23a47
YO
3067 fake_study_recurse:
3068 while ( scan && OP(scan) != END && scan < last ){
2d608413
KW
3069 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3070 node length to get a real minimum (because
3071 the folded version may be shorter) */
f758bddf 3072 bool has_exactf_sharp_s = FALSE;
8aa23a47 3073 /* Peephole optimizer: */
304ee84b 3074 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 3075 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
3076
3077 /* Its not clear to khw or hv why this is done here, and not in the
3078 * clauses that deal with EXACT nodes. khw's guess is that it's
3079 * because of a previous design */
9d071ca8 3080 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
3081
3082 /* Follow the next-chain of the current node and optimize
3083 away all the NOTHINGs from it. */
3084 if (OP(scan) != CURLYX) {
3085 const int max = (reg_off_by_arg[OP(scan)]
3086 ? I32_MAX
3087 /* I32 may be smaller than U16 on CRAYs! */
3088 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3089 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3090 int noff;
3091 regnode *n = scan;
686b73d4 3092
8aa23a47
YO
3093 /* Skip NOTHING and LONGJMP. */
3094 while ((n = regnext(n))
3095 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3096 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3097 && off + noff < max)
3098 off += noff;
3099 if (reg_off_by_arg[OP(scan)])
3100 ARG(scan) = off;
3101 else
3102 NEXT_OFF(scan) = off;
3103 }
a3621e74 3104
c277df42 3105
8aa23a47
YO
3106
3107 /* The principal pseudo-switch. Cannot be a switch, since we
3108 look into several different things. */
3109 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3110 || OP(scan) == IFTHEN) {
3111 next = regnext(scan);
3112 code = OP(scan);
3113 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3114
8aa23a47
YO
3115 if (OP(next) == code || code == IFTHEN) {
3116 /* NOTE - There is similar code to this block below for handling
3117 TRIE nodes on a re-study. If you change stuff here check there
3118 too. */
3119 I32 max1 = 0, min1 = I32_MAX, num = 0;
3120 struct regnode_charclass_class accum;
3121 regnode * const startbranch=scan;
686b73d4 3122
8aa23a47 3123 if (flags & SCF_DO_SUBSTR)
304ee84b 3124 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3125 if (flags & SCF_DO_STCLASS)
e755fd73 3126 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3127
3128 while (OP(scan) == code) {
3129 I32 deltanext, minnext, f = 0, fake;
3130 struct regnode_charclass_class this_class;
3131
3132 num++;
3133 data_fake.flags = 0;
3134 if (data) {
3135 data_fake.whilem_c = data->whilem_c;
3136 data_fake.last_closep = data->last_closep;
3137 }
3138 else
3139 data_fake.last_closep = &fake;
58e23c8d
YO
3140
3141 data_fake.pos_delta = delta;
8aa23a47
YO
3142 next = regnext(scan);
3143 scan = NEXTOPER(scan);
3144 if (code != BRANCH)
c277df42 3145 scan = NEXTOPER(scan);
8aa23a47 3146 if (flags & SCF_DO_STCLASS) {
e755fd73 3147 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3148 data_fake.start_class = &this_class;
3149 f = SCF_DO_STCLASS_AND;
58e23c8d 3150 }
8aa23a47
YO
3151 if (flags & SCF_WHILEM_VISITED_POS)
3152 f |= SCF_WHILEM_VISITED_POS;
3153
3154 /* we suppose the run is continuous, last=next...*/
3155 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3156 next, &data_fake,
3157 stopparen, recursed, NULL, f,depth+1);
3158 if (min1 > minnext)
3159 min1 = minnext;
9b139d09 3160 if (deltanext == I32_MAX) {
8aa23a47 3161 is_inf = is_inf_internal = 1;
9b139d09
GG
3162 max1 = I32_MAX;
3163 } else if (max1 < minnext + deltanext)
3164 max1 = minnext + deltanext;
8aa23a47
YO
3165 scan = next;
3166 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3167 pars++;
3168 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3169 if ( stopmin > minnext)
3170 stopmin = min + min1;
3171 flags &= ~SCF_DO_SUBSTR;
3172 if (data)
3173 data->flags |= SCF_SEEN_ACCEPT;
3174 }
3175 if (data) {
3176 if (data_fake.flags & SF_HAS_EVAL)
3177 data->flags |= SF_HAS_EVAL;
3178 data->whilem_c = data_fake.whilem_c;
3dab1dad 3179 }
8aa23a47 3180 if (flags & SCF_DO_STCLASS)
3fffb88a 3181 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3182 }
3183 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3184 min1 = 0;
3185 if (flags & SCF_DO_SUBSTR) {
3186 data->pos_min += min1;
9b139d09
GG
3187 if (data->pos_delta >= I32_MAX - (max1 - min1))
3188 data->pos_delta = I32_MAX;
3189 else
3190 data->pos_delta += max1 - min1;
8aa23a47
YO
3191 if (max1 != min1 || is_inf)
3192 data->longest = &(data->longest_float);
3193 }
3194 min += min1;
9b139d09
GG
3195 if (delta == I32_MAX || I32_MAX - delta - (max1 - min1) < 0)
3196 delta = I32_MAX;
3197 else
3198 delta += max1 - min1;
8aa23a47 3199 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3200 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3201 if (min1) {
3202 cl_and(data->start_class, and_withp);
3203 flags &= ~SCF_DO_STCLASS;
653099ff 3204 }
8aa23a47
YO
3205 }
3206 else if (flags & SCF_DO_STCLASS_AND) {
3207 if (min1) {
3208 cl_and(data->start_class, &accum);
3209 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3210 }
8aa23a47
YO
3211 else {
3212 /* Switch to OR mode: cache the old value of
3213 * data->start_class */
3214 INIT_AND_WITHP;
3215 StructCopy(data->start_class, and_withp,
3216 struct regnode_charclass_class);
3217 flags &= ~SCF_DO_STCLASS_AND;
3218 StructCopy(&accum, data->start_class,
3219 struct regnode_charclass_class);
3220 flags |= SCF_DO_STCLASS_OR;
899d20b9 3221 SET_SSC_EOS(data->start_class);
de0c8cb8 3222 }
8aa23a47 3223 }
a3621e74 3224
8aa23a47
YO
3225 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3226 /* demq.
a3621e74 3227
8aa23a47
YO
3228 Assuming this was/is a branch we are dealing with: 'scan' now
3229 points at the item that follows the branch sequence, whatever
3230 it is. We now start at the beginning of the sequence and look
3231 for subsequences of
a3621e74 3232
8aa23a47
YO
3233 BRANCH->EXACT=>x1
3234 BRANCH->EXACT=>x2
3235 tail
a3621e74 3236
8aa23a47 3237 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3238
486ec47a 3239 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3240 element into a trie and then add the subsequent branch exact
3241 strings to the trie.
a3621e74 3242
8aa23a47 3243 We have two cases
a3621e74 3244
3b753521 3245 1. patterns where the whole set of branches can be converted.
a3621e74 3246
8aa23a47 3247 2. patterns where only a subset can be converted.
a3621e74 3248
8aa23a47
YO
3249 In case 1 we can replace the whole set with a single regop
3250 for the trie. In case 2 we need to keep the start and end
3b753521 3251 branches so
a3621e74 3252
8aa23a47
YO
3253 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3254 becomes BRANCH TRIE; BRANCH X;
786e8c11 3255
8aa23a47
YO
3256 There is an additional case, that being where there is a
3257 common prefix, which gets split out into an EXACT like node
3258 preceding the TRIE node.
a3621e74 3259
8aa23a47
YO
3260 If x(1..n)==tail then we can do a simple trie, if not we make
3261 a "jump" trie, such that when we match the appropriate word
486ec47a 3262 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3263 a nested if into a case structure of sorts.
b515a41d 3264
8aa23a47 3265 */
686b73d4 3266
8aa23a47
YO
3267 int made=0;
3268 if (!re_trie_maxbuff) {
3269 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3270 if (!SvIOK(re_trie_maxbuff))
3271 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3272 }
3273 if ( SvIV(re_trie_maxbuff)>=0 ) {
3274 regnode *cur;
3275 regnode *first = (regnode *)NULL;
3276 regnode *last = (regnode *)NULL;
3277 regnode *tail = scan;
fab2782b 3278 U8 trietype = 0;
8aa23a47 3279 U32 count=0;
a3621e74
YO
3280
3281#ifdef DEBUGGING
8aa23a47 3282 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3283#endif
8aa23a47
YO
3284 /* var tail is used because there may be a TAIL
3285 regop in the way. Ie, the exacts will point to the
3286 thing following the TAIL, but the last branch will
3287 point at the TAIL. So we advance tail. If we
3288 have nested (?:) we may have to move through several
3289 tails.
3290 */
3291
3292 while ( OP( tail ) == TAIL ) {
3293 /* this is the TAIL generated by (?:) */
3294 tail = regnext( tail );
3295 }
a3621e74 3296
8aa23a47 3297
df826430 3298 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3299 regprop(RExC_rx, mysv, tail );
3300 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3301 (int)depth * 2 + 2, "",
3302 "Looking for TRIE'able sequences. Tail node is: ",
3303 SvPV_nolen_const( mysv )
3304 );
3305 });
3306
3307 /*
3308
fab2782b
YO
3309 Step through the branches
3310 cur represents each branch,
3311 noper is the first thing to be matched as part of that branch
3312 noper_next is the regnext() of that node.
3313
3314 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3315 via a "jump trie" but we also support building with NOJUMPTRIE,
3316 which restricts the trie logic to structures like /FOO|BAR/.
3317
3318 If noper is a trieable nodetype then the branch is a possible optimization
3319 target. If we are building under NOJUMPTRIE then we require that noper_next
3320 is the same as scan (our current position in the regex program).
3321
3322 Once we have two or more consecutive such branches we can create a
3323 trie of the EXACT's contents and stitch it in place into the program.
3324
3325 If the sequence represents all of the branches in the alternation we
3326 replace the entire thing with a single TRIE node.
3327
3328 Otherwise when it is a subsequence we need to stitch it in place and
3329 replace only the relevant branches. This means the first branch has
3330 to remain as it is used by the alternation logic, and its next pointer,
3331 and needs to be repointed at the item on the branch chain following
3332 the last branch we have optimized away.
3333
3334 This could be either a BRANCH, in which case the subsequence is internal,
3335 or it could be the item following the branch sequence in which case the
3336 subsequence is at the end (which does not necessarily mean the first node
3337 is the start of the alternation).
3338
3339 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3340
3341 optype | trietype
3342 ----------------+-----------
3343 NOTHING | NOTHING
3344 EXACT | EXACT
3345 EXACTFU | EXACTFU
3346 EXACTFU_SS | EXACTFU
3347 EXACTFU_TRICKYFOLD | EXACTFU
3348 EXACTFA | 0
3349
8aa23a47
YO
3350
3351 */
fab2782b
YO
3352#define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3353 ( EXACT == (X) ) ? EXACT : \
3354 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3355 0 )
8aa23a47
YO
3356
3357 /* dont use tail as the end marker for this traverse */
3358 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3359 regnode * const noper = NEXTOPER( cur );
fab2782b
YO
3360 U8 noper_type = OP( noper );
3361 U8 noper_trietype = TRIE_TYPE( noper_type );
b515a41d 3362#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3363 regnode * const noper_next = regnext( noper );
df826430
YO
3364 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3365 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
b515a41d
YO
3366#endif
3367
df826430 3368 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3369 regprop(RExC_rx, mysv, cur);
3370 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3371 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3372
3373 regprop(RExC_rx, mysv, noper);
3374 PerlIO_printf( Perl_debug_log, " -> %s",
3375 SvPV_nolen_const(mysv));
3376
3377 if ( noper_next ) {
3378 regprop(RExC_rx, mysv, noper_next );
3379 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3380 SvPV_nolen_const(mysv));
3381 }
df826430
YO
3382 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3383 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3384 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3385 );
8aa23a47 3386 });
fab2782b
YO
3387
3388 /* Is noper a trieable nodetype that can be merged with the
3389 * current trie (if there is one)? */
3390 if ( noper_trietype
3391 &&
3392 (
df826430
YO
3393 ( noper_trietype == NOTHING)
3394 || ( trietype == NOTHING )
a40630bf 3395 || ( trietype == noper_trietype )
fab2782b 3396 )
786e8c11 3397#ifdef NOJUMPTRIE
8aa23a47 3398 && noper_next == tail
786e8c11 3399#endif
8aa23a47
YO
3400 && count < U16_MAX)
3401 {
fab2782b
YO
3402 /* Handle mergable triable node
3403 * Either we are the first node in a new trieable sequence,
3404 * in which case we do some bookkeeping, otherwise we update
3405 * the end pointer. */
fab2782b 3406 if ( !first ) {
3b6759a6 3407 first = cur;
df826430
YO
3408 if ( noper_trietype == NOTHING ) {
3409#if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3410 regnode * const noper_next = regnext( noper );
3b6759a6 3411 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
df826430
YO
3412 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3413#endif
3414
190c1910 3415 if ( noper_next_trietype ) {
df826430 3416 trietype = noper_next_trietype;
190c1910
YO
3417 } else if (noper_next_type) {
3418 /* a NOTHING regop is 1 regop wide. We need at least two
3419 * for a trie so we can't merge this in */
3420 first = NULL;
3421 }
3422 } else {
3423 trietype = noper_trietype;
3b6759a6 3424 }
8aa23a47 3425 } else {
fab2782b
YO
3426 if ( trietype == NOTHING )
3427 trietype = noper_trietype;
8aa23a47
YO
3428 last = cur;
3429 }
df826430
YO
3430 if (first)
3431 count++;
fab2782b
YO
3432 } /* end handle mergable triable node */
3433 else {
3434 /* handle unmergable node -
3435 * noper may either be a triable node which can not be tried
3436 * together with the current trie, or a non triable node */
729aaeb5
YO
3437 if ( last ) {
3438 /* If last is set and trietype is not NOTHING then we have found
3439 * at least two triable branch sequences in a row of a similar
3440 * trietype so we can turn them into a trie. If/when we
3441 * allow NOTHING to start a trie sequence this condition will be
3442 * required, and it isn't expensive so we leave it in for now. */
e6351b37 3443 if ( trietype && trietype != NOTHING )
729aaeb5
YO
3444 make_trie( pRExC_state,
3445 startbranch, first, cur, tail, count,
3446 trietype, depth+1 );
fab2782b 3447 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
8aa23a47 3448 }
fab2782b 3449 if ( noper_trietype
786e8c11 3450#ifdef NOJUMPTRIE
8aa23a47 3451 && noper_next == tail
786e8c11 3452#endif
8aa23a47 3453 ){
fab2782b 3454 /* noper is triable, so we can start a new trie sequence */
8aa23a47
YO
3455 count = 1;
3456 first = cur;
fab2782b
YO
3457 trietype = noper_trietype;
3458 } else if (first) {
3459 /* if we already saw a first but the current node is not triable then we have
3460 * to reset the first information. */
8aa23a47
YO
3461 count = 0;
3462 first = NULL;
fab2782b 3463 trietype = 0;
8aa23a47 3464 }
fab2782b
YO
3465 } /* end handle unmergable node */
3466 } /* loop over branches */
df826430 3467 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3468 regprop(RExC_rx, mysv, cur);
3469 PerlIO_printf( Perl_debug_log,
3470 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3471 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3472
3473 });
e6351b37 3474 if ( last && trietype ) {
3b6759a6
YO
3475 if ( trietype != NOTHING ) {
3476 /* the last branch of the sequence was part of a trie,
3477 * so we have to construct it here outside of the loop
3478 */
3479 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
686b73d4 3480#ifdef TRIE_STUDY_OPT
3b6759a6
YO
3481 if ( ((made == MADE_EXACT_TRIE &&
3482 startbranch == first)
3483 || ( first_non_open == first )) &&
3484 depth==0 ) {
3485 flags |= SCF_TRIE_RESTUDY;
3486 if ( startbranch == first
3487 && scan == tail )
3488 {
3489 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3490 }
8aa23a47 3491 }
3dab1dad 3492#endif
3b6759a6
YO
3493 } else {
3494 /* at this point we know whatever we have is a NOTHING sequence/branch
3495 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3496 */
3497 if ( startbranch == first ) {
3498 regnode *opt;
3499 /* the entire thing is a NOTHING sequence, something like this:
3500 * (?:|) So we can turn it into a plain NOTHING op. */
3501 DEBUG_TRIE_COMPILE_r({
3502 regprop(RExC_rx, mysv, cur);
3503 PerlIO_printf( Perl_debug_log,
3504 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3505 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3506
3507 });
3508 OP(startbranch)= NOTHING;
3509 NEXT_OFF(startbranch)= tail - startbranch;
3510 for ( opt= startbranch + 1; opt < tail ; opt++ )
3511 OP(opt)= OPTIMIZED;
3512 }
3513 }
fab2782b
YO
3514 } /* end if ( last) */
3515 } /* TRIE_MAXBUF is non zero */
8aa23a47
YO
3516
3517 } /* do trie */
3518
653099ff 3519 }
8aa23a47
YO
3520 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3521 scan = NEXTOPER(NEXTOPER(scan));
3522 } else /* single branch is optimized. */
3523 scan = NEXTOPER(scan);
3524 continue;
3525 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3526 scan_frame *newframe = NULL;
3527 I32 paren;
3528 regnode *start;
3529 regnode *end;
3530
3531 if (OP(scan) != SUSPEND) {
3532 /* set the pointer */
3533 if (OP(scan) == GOSUB) {
3534 paren = ARG(scan);
3535 RExC_recurse[ARG2L(scan)] = scan;
3536 start = RExC_open_parens[paren-1];
3537 end = RExC_close_parens[paren-1];
3538 } else {
3539 paren = 0;
f8fc2ecf 3540 start = RExC_rxi->program + 1;
8aa23a47
YO
3541 end = RExC_opend;
3542 }
3543 if (!recursed) {
3544 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3545 SAVEFREEPV(recursed);
3546 }
3547 if (!PAREN_TEST(recursed,paren+1)) {
3548 PAREN_SET(recursed,paren+1);
3549 Newx(newframe,1,scan_frame);
3550 } else {
3551 if (flags & SCF_DO_SUBSTR) {
304ee84b 3552 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3553 data->longest = &(data->longest_float);
3554 }
3555 is_inf = is_inf_internal = 1;
3556 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3557 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3558 flags &= ~SCF_DO_STCLASS;
3559 }
3560 } else {
3561 Newx(newframe,1,scan_frame);
3562 paren = stopparen;
3563 start = scan+2;
3564 end = regnext(scan);
3565 }
3566 if (newframe) {
3567 assert(start);
3568 assert(end);
3569 SAVEFREEPV(newframe