This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
In regcomp.c, Set_Node_Cur_Length() uses parse_start, so explicitly pass it.
[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
07be1b83 408
43fead97 409#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
410
411/* The enums for all these are ordered so things work out correctly */
a62b1201 412#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
cfaf538b 413#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
00b27cfc 414#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
415#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
416#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1 417#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a725e29c 418#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 419
43fead97 420#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 421
93733859 422#define OOB_NAMEDCLASS -1
b8c5462f 423
8e661ac5
KW
424/* There is no code point that is out-of-bounds, so this is problematic. But
425 * its only current use is to initialize a variable that is always set before
426 * looked at. */
427#define OOB_UNICODE 0xDEADBEEF
428
a0ed51b3
LW
429#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
430#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
431
8615cb43 432
b45f050a
JF
433/* length of regex to show in messages that don't mark a position within */
434#define RegexLengthToShowInErrorMessages 127
435
436/*
437 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
438 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
439 * op/pragma/warn/regcomp.
440 */
7253e4e3
RK
441#define MARKER1 "<-- HERE" /* marker as it appears in the description */
442#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 443
7253e4e3 444#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
445
446/*
447 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
448 * arg. Show regex, up to a maximum length. If it's too long, chop and add
449 * "...".
450 */
58e23c8d 451#define _FAIL(code) STMT_START { \
bfed75c6 452 const char *ellipses = ""; \
ccb2c380
MP
453 IV len = RExC_end - RExC_precomp; \
454 \
455 if (!SIZE_ONLY) \
a5e7bc51 456 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
457 if (len > RegexLengthToShowInErrorMessages) { \
458 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
459 len = RegexLengthToShowInErrorMessages - 10; \
460 ellipses = "..."; \
461 } \
58e23c8d 462 code; \
ccb2c380 463} STMT_END
8615cb43 464
58e23c8d
YO
465#define FAIL(msg) _FAIL( \
466 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
467 msg, (int)len, RExC_precomp, ellipses))
468
469#define FAIL2(msg,arg) _FAIL( \
470 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
471 arg, (int)len, RExC_precomp, ellipses))
472
b45f050a 473/*
b45f050a
JF
474 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
475 */
ccb2c380 476#define Simple_vFAIL(m) STMT_START { \
a28509cc 477 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
478 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
479 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
480} STMT_END
b45f050a
JF
481
482/*
483 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
484 */
ccb2c380
MP
485#define vFAIL(m) STMT_START { \
486 if (!SIZE_ONLY) \
a5e7bc51 487 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
488 Simple_vFAIL(m); \
489} STMT_END
b45f050a
JF
490
491/*
492 * Like Simple_vFAIL(), but accepts two arguments.
493 */
ccb2c380 494#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 495 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
496 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
497 (int)offset, RExC_precomp, RExC_precomp + offset); \
498} STMT_END
b45f050a
JF
499
500/*
501 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
502 */
ccb2c380
MP
503#define vFAIL2(m,a1) STMT_START { \
504 if (!SIZE_ONLY) \
a5e7bc51 505 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
506 Simple_vFAIL2(m, a1); \
507} STMT_END
b45f050a
JF
508
509
510/*
511 * Like Simple_vFAIL(), but accepts three arguments.
512 */
ccb2c380 513#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 514 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
515 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
516 (int)offset, RExC_precomp, RExC_precomp + offset); \
517} STMT_END
b45f050a
JF
518
519/*
520 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
521 */
ccb2c380
MP
522#define vFAIL3(m,a1,a2) STMT_START { \
523 if (!SIZE_ONLY) \
a5e7bc51 524 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
525 Simple_vFAIL3(m, a1, a2); \
526} STMT_END
b45f050a
JF
527
528/*
529 * Like Simple_vFAIL(), but accepts four arguments.
530 */
ccb2c380 531#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 532 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
533 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
534 (int)offset, RExC_precomp, RExC_precomp + offset); \
535} STMT_END
b45f050a 536
95db3ffa
KW
537#define vFAIL4(m,a1,a2,a3) STMT_START { \
538 if (!SIZE_ONLY) \
539 SAVEFREESV(RExC_rx_sv); \
540 Simple_vFAIL4(m, a1, a2, a3); \
541} STMT_END
542
5e0a247b
KW
543/* m is not necessarily a "literal string", in this macro */
544#define reg_warn_non_literal_string(loc, m) STMT_START { \
545 const IV offset = loc - RExC_precomp; \
546 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
547 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
548} STMT_END
549
668c081a 550#define ckWARNreg(loc,m) STMT_START { \
a28509cc 551 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
552 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
553 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
554} STMT_END
555
0d6106aa
KW
556#define vWARN_dep(loc, m) STMT_START { \
557 const IV offset = loc - RExC_precomp; \
558 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
559 (int)offset, RExC_precomp, RExC_precomp + offset); \
560} STMT_END
561
147508a2
KW
562#define ckWARNdep(loc,m) STMT_START { \
563 const IV offset = loc - RExC_precomp; \
564 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
565 m REPORT_LOCATION, \
566 (int)offset, RExC_precomp, RExC_precomp + offset); \
567} STMT_END
568
668c081a 569#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 570 const IV offset = loc - RExC_precomp; \
d1d15184 571 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
572 m REPORT_LOCATION, \
573 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
574} STMT_END
575
2335b3d3
KW
576#define ckWARN2regdep(loc,m, a1) STMT_START { \
577 const IV offset = loc - RExC_precomp; \
578 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
579 m REPORT_LOCATION, \
580 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
581} STMT_END
582
668c081a 583#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 584 const IV offset = loc - RExC_precomp; \
668c081a 585 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
586 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
587} STMT_END
588
589#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 590 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
591 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
592 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
593} STMT_END
594
668c081a
NC
595#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
596 const IV offset = loc - RExC_precomp; \
597 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
598 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
599} STMT_END
600
ccb2c380 601#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 602 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
603 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
604 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
605} STMT_END
606
668c081a
NC
607#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
608 const IV offset = loc - RExC_precomp; \
609 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
610 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
611} STMT_END
612
ccb2c380 613#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 614 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
615 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
616 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
617} STMT_END
9d1d55b5 618
8615cb43 619
cd439c50 620/* Allow for side effects in s */
ccb2c380
MP
621#define REGC(c,s) STMT_START { \
622 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
623} STMT_END
cd439c50 624
fac92740
MJD
625/* Macros for recording node offsets. 20001227 mjd@plover.com
626 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
627 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
628 * Element 0 holds the number n.
07be1b83 629 * Position is 1 indexed.
fac92740 630 */
7122b237
YO
631#ifndef RE_TRACK_PATTERN_OFFSETS
632#define Set_Node_Offset_To_R(node,byte)
633#define Set_Node_Offset(node,byte)
634#define Set_Cur_Node_Offset
635#define Set_Node_Length_To_R(node,len)
636#define Set_Node_Length(node,len)
6a86c6ad 637#define Set_Node_Cur_Length(node,start)
7122b237
YO
638#define Node_Offset(n)
639#define Node_Length(n)
640#define Set_Node_Offset_Length(node,offset,len)
641#define ProgLen(ri) ri->u.proglen
642#define SetProgLen(ri,x) ri->u.proglen = x
643#else
644#define ProgLen(ri) ri->u.offsets[0]
645#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
646#define Set_Node_Offset_To_R(node,byte) STMT_START { \
647 if (! SIZE_ONLY) { \
648 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 649 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 650 if((node) < 0) { \
551405c4 651 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
652 } else { \
653 RExC_offsets[2*(node)-1] = (byte); \
654 } \
655 } \
656} STMT_END
657
658#define Set_Node_Offset(node,byte) \
659 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
660#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
661
662#define Set_Node_Length_To_R(node,len) STMT_START { \
663 if (! SIZE_ONLY) { \
664 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 665 __LINE__, (int)(node), (int)(len))); \
ccb2c380 666 if((node) < 0) { \
551405c4 667 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
668 } else { \
669 RExC_offsets[2*(node)] = (len); \
670 } \
671 } \
672} STMT_END
673
674#define Set_Node_Length(node,len) \
675 Set_Node_Length_To_R((node)-RExC_emit_start, len)
6a86c6ad
NC
676#define Set_Node_Cur_Length(node, start) \
677 Set_Node_Length(node, RExC_parse - start)
fac92740
MJD
678
679/* Get offsets and lengths */
680#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
681#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
682
07be1b83
YO
683#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
684 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
685 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
686} STMT_END
7122b237 687#endif
07be1b83
YO
688
689#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
690#define EXPERIMENTAL_INPLACESCAN
f427392e 691#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 692
304ee84b
YO
693#define DEBUG_STUDYDATA(str,data,depth) \
694DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 695 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
696 "%*s" str "Pos:%"IVdf"/%"IVdf \
697 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
698 (int)(depth)*2, "", \
699 (IV)((data)->pos_min), \
700 (IV)((data)->pos_delta), \
304ee84b 701 (UV)((data)->flags), \
1de06328 702 (IV)((data)->whilem_c), \
304ee84b
YO
703 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
704 is_inf ? "INF " : "" \
1de06328
YO
705 ); \
706 if ((data)->last_found) \
707 PerlIO_printf(Perl_debug_log, \
708 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
709 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
710 SvPVX_const((data)->last_found), \
711 (IV)((data)->last_end), \
712 (IV)((data)->last_start_min), \
713 (IV)((data)->last_start_max), \
714 ((data)->longest && \
715 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
716 SvPVX_const((data)->longest_fixed), \
717 (IV)((data)->offset_fixed), \
718 ((data)->longest && \
719 (data)->longest==&((data)->longest_float)) ? "*" : "", \
720 SvPVX_const((data)->longest_float), \
721 (IV)((data)->offset_float_min), \
722 (IV)((data)->offset_float_max) \
723 ); \
724 PerlIO_printf(Perl_debug_log,"\n"); \
725});
726
653099ff 727/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 728 Update the longest found anchored substring and the longest found
653099ff
GS
729 floating substrings if needed. */
730
4327152a 731STATIC void
304ee84b 732S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
c277df42 733{
e1ec3a88
AL
734 const STRLEN l = CHR_SVLEN(data->last_found);
735 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 736 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 737
7918f24d
NC
738 PERL_ARGS_ASSERT_SCAN_COMMIT;
739
c277df42 740 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 741 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
742 if (*data->longest == data->longest_fixed) {
743 data->offset_fixed = l ? data->last_start_min : data->pos_min;
744 if (data->flags & SF_BEFORE_EOL)
b81d288d 745 data->flags
c277df42
IZ
746 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
747 else
748 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 749 data->minlen_fixed=minlenp;
1de06328 750 data->lookbehind_fixed=0;
a0ed51b3 751 }
304ee84b 752 else { /* *data->longest == data->longest_float */
c277df42 753 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
754 data->offset_float_max = (l
755 ? data->last_start_max
9b139d09 756 : (data->pos_delta == I32_MAX ? I32_MAX : data->pos_min + data->pos_delta));
304ee84b 757 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
9051bda5 758 data->offset_float_max = I32_MAX;
c277df42 759 if (data->flags & SF_BEFORE_EOL)
b81d288d 760 data->flags
c277df42
IZ
761 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
762 else
763 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
764 data->minlen_float=minlenp;
765 data->lookbehind_float=0;
c277df42
IZ
766 }
767 }
768 SvCUR_set(data->last_found, 0);
0eda9292 769 {
a28509cc 770 SV * const sv = data->last_found;
097eb12c
AL
771 if (SvUTF8(sv) && SvMAGICAL(sv)) {
772 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
773 if (mg)
774 mg->mg_len = 0;
775 }
0eda9292 776 }
c277df42
IZ
777 data->last_end = -1;
778 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 779 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
780}
781
899d20b9
KW
782/* These macros set, clear and test whether the synthetic start class ('ssc',
783 * given by the parameter) matches an empty string (EOS). This uses the
784 * 'next_off' field in the node, to save a bit in the flags field. The ssc
785 * stands alone, so there is never a next_off, so this field is otherwise
786 * unused. The EOS information is used only for compilation, but theoretically
787 * it could be passed on to the execution code. This could be used to store
788 * more than one bit of information, but only this one is currently used. */
789#define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
790#define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
791#define TEST_SSC_EOS(node) cBOOL((node)->next_off)
792
653099ff
GS
793/* Can match anything (initialization) */
794STATIC void
3fffb88a 795S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 796{
7918f24d
NC
797 PERL_ARGS_ASSERT_CL_ANYTHING;
798
f8bef550 799 ANYOF_BITMAP_SETALL(cl);
899d20b9
KW
800 cl->flags = ANYOF_UNICODE_ALL;
801 SET_SSC_EOS(cl);
3fffb88a
KW
802
803 /* If any portion of the regex is to operate under locale rules,
804 * initialization includes it. The reason this isn't done for all regexes
805 * is that the optimizer was written under the assumption that locale was
806 * all-or-nothing. Given the complexity and lack of documentation in the
807 * optimizer, and that there are inadequate test cases for locale, so many
808 * parts of it may not work properly, it is safest to avoid locale unless
809 * necessary. */
810 if (RExC_contains_locale) {
9d7a1e63 811 ANYOF_CLASS_SETALL(cl); /* /l uses class */
b25a1036 812 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
3fffb88a 813 }
9d7a1e63
KW
814 else {
815 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
816 }
653099ff
GS
817}
818
819/* Can match anything (initialization) */
820STATIC int
5f66b61c 821S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
822{
823 int value;
824
7918f24d
NC
825 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
826
3732a889 827 for (value = 0; value < ANYOF_MAX; value += 2)
653099ff
GS
828 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
829 return 1;
1aa99e6b
IH
830 if (!(cl->flags & ANYOF_UNICODE_ALL))
831 return 0;
10edeb5d 832 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 833 return 0;
653099ff
GS
834 return 1;
835}
836
837/* Can match anything (initialization) */
838STATIC void
e755fd73 839S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 840{
7918f24d
NC
841 PERL_ARGS_ASSERT_CL_INIT;
842
8ecf7187 843 Zero(cl, 1, struct regnode_charclass_class);
653099ff 844 cl->type = ANYOF;
3fffb88a 845 cl_anything(pRExC_state, cl);
1411dba4 846 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
847}
848
1051e1c4
KW
849/* These two functions currently do the exact same thing */
850#define cl_init_zero S_cl_init
653099ff 851
dd58aee1
KW
852/* 'AND' a given class with another one. Can create false positives. 'cl'
853 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
854 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 855STATIC void
5f66b61c 856S_cl_and(struct regnode_charclass_class *cl,
a28509cc 857 const struct regnode_charclass_class *and_with)
653099ff 858{
7918f24d 859 PERL_ARGS_ASSERT_CL_AND;
40d049e4 860
954a2af6 861 assert(PL_regkind[and_with->type] == ANYOF);
1e6ade67 862
c6b76537 863 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
864 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
865 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 866 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
867 && !(and_with->flags & ANYOF_LOC_FOLD)
868 && !(cl->flags & ANYOF_LOC_FOLD)) {
653099ff
GS
869 int i;
870
871 if (and_with->flags & ANYOF_INVERT)
872 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
873 cl->bitmap[i] &= ~and_with->bitmap[i];
874 else
875 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
876 cl->bitmap[i] &= and_with->bitmap[i];
877 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 878
c6b76537 879 if (and_with->flags & ANYOF_INVERT) {
8951c461 880
c6b76537
KW
881 /* Here, the and'ed node is inverted. Get the AND of the flags that
882 * aren't affected by the inversion. Those that are affected are
883 * handled individually below */
884 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
885 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
886 cl->flags |= affected_flags;
887
888 /* We currently don't know how to deal with things that aren't in the
889 * bitmap, but we know that the intersection is no greater than what
890 * is already in cl, so let there be false positives that get sorted
891 * out after the synthetic start class succeeds, and the node is
892 * matched for real. */
893
894 /* The inversion of these two flags indicate that the resulting
895 * intersection doesn't have them */
896 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
897 cl->flags &= ~ANYOF_UNICODE_ALL;
898 }
c6b76537
KW
899 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
900 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 901 }
1aa99e6b 902 }
c6b76537 903 else { /* and'd node is not inverted */
3ad98780
KW
904 U8 outside_bitmap_but_not_utf8; /* Temp variable */
905
137165a6 906 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
907
908 /* Here 'and_with' doesn't match anything outside the bitmap
909 * (except possibly ANYOF_UNICODE_ALL), which means the
910 * intersection can't either, except for ANYOF_UNICODE_ALL, in
911 * which case we don't know what the intersection is, but it's no
912 * greater than what cl already has, so can just leave it alone,
913 * with possible false positives */
914 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
915 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 916 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 917 }
137165a6 918 }
c6b76537
KW
919 else if (! ANYOF_NONBITMAP(cl)) {
920
921 /* Here, 'and_with' does match something outside the bitmap, and cl
922 * doesn't have a list of things to match outside the bitmap. If
923 * cl can match all code points above 255, the intersection will
3ad98780
KW
924 * be those above-255 code points that 'and_with' matches. If cl
925 * can't match all Unicode code points, it means that it can't
926 * match anything outside the bitmap (since the 'if' that got us
927 * into this block tested for that), so we leave the bitmap empty.
928 */
c6b76537
KW
929 if (cl->flags & ANYOF_UNICODE_ALL) {
930 ARG_SET(cl, ARG(and_with));
3ad98780
KW
931
932 /* and_with's ARG may match things that don't require UTF8.
933 * And now cl's will too, in spite of this being an 'and'. See
934 * the comments below about the kludge */
935 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
936 }
937 }
938 else {
939 /* Here, both 'and_with' and cl match something outside the
940 * bitmap. Currently we do not do the intersection, so just match
941 * whatever cl had at the beginning. */
942 }
943
944
3ad98780
KW
945 /* Take the intersection of the two sets of flags. However, the
946 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
947 * kludge around the fact that this flag is not treated like the others
948 * which are initialized in cl_anything(). The way the optimizer works
949 * is that the synthetic start class (SSC) is initialized to match
950 * anything, and then the first time a real node is encountered, its
951 * values are AND'd with the SSC's with the result being the values of
952 * the real node. However, there are paths through the optimizer where
953 * the AND never gets called, so those initialized bits are set
954 * inappropriately, which is not usually a big deal, as they just cause
955 * false positives in the SSC, which will just mean a probably
956 * imperceptible slow down in execution. However this bit has a
957 * higher false positive consequence in that it can cause utf8.pm,
958 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
959 * bigger slowdown and also causes significant extra memory to be used.
960 * In order to prevent this, the code now takes a different tack. The
961 * bit isn't set unless some part of the regular expression needs it,
962 * but once set it won't get cleared. This means that these extra
963 * modules won't get loaded unless there was some path through the
964 * pattern that would have required them anyway, and so any false
965 * positives that occur by not ANDing them out when they could be
966 * aren't as severe as they would be if we treated this bit like all
967 * the others */
968 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
969 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 970 cl->flags &= and_with->flags;
3ad98780 971 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 972 }
653099ff
GS
973}
974
dd58aee1
KW
975/* 'OR' a given class with another one. Can create false positives. 'cl'
976 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
977 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 978STATIC void
3fffb88a 979S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 980{
7918f24d
NC
981 PERL_ARGS_ASSERT_CL_OR;
982
653099ff 983 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
984
985 /* Here, the or'd node is to be inverted. This means we take the
986 * complement of everything not in the bitmap, but currently we don't
987 * know what that is, so give up and match anything */
988 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 989 cl_anything(pRExC_state, cl);
c6b76537 990 }
653099ff
GS
991 /* We do not use
992 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
993 * <= (B1 | !B2) | (CL1 | !CL2)
994 * which is wasteful if CL2 is small, but we ignore CL2:
995 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
996 * XXXX Can we handle case-fold? Unclear:
997 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
998 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
999 */
c6b76537 1000 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
1001 && !(or_with->flags & ANYOF_LOC_FOLD)
1002 && !(cl->flags & ANYOF_LOC_FOLD) ) {
653099ff
GS
1003 int i;
1004
1005 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1006 cl->bitmap[i] |= ~or_with->bitmap[i];
1007 } /* XXXX: logic is complicated otherwise */
1008 else {
3fffb88a 1009 cl_anything(pRExC_state, cl);
653099ff 1010 }
c6b76537
KW
1011
1012 /* And, we can just take the union of the flags that aren't affected
1013 * by the inversion */
1014 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
1015
1016 /* For the remaining flags:
1017 ANYOF_UNICODE_ALL and inverted means to not match anything above
1018 255, which means that the union with cl should just be
1019 what cl has in it, so can ignore this flag
1020 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
1021 is 127-255 to match them, but then invert that, so the
1022 union with cl should just be what cl has in it, so can
1023 ignore this flag
1024 */
1025 } else { /* 'or_with' is not inverted */
653099ff
GS
1026 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1027 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
1028 && (!(or_with->flags & ANYOF_LOC_FOLD)
1029 || (cl->flags & ANYOF_LOC_FOLD)) ) {
653099ff
GS
1030 int i;
1031
1032 /* OR char bitmap and class bitmap separately */
1033 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1034 cl->bitmap[i] |= or_with->bitmap[i];
c96939e4
KW
1035 if (or_with->flags & ANYOF_CLASS) {
1036 ANYOF_CLASS_OR(or_with, cl);
1037 }
653099ff
GS
1038 }
1039 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 1040 cl_anything(pRExC_state, cl);
653099ff 1041 }
9826f543 1042
c6b76537
KW
1043 if (ANYOF_NONBITMAP(or_with)) {
1044
1045 /* Use the added node's outside-the-bit-map match if there isn't a
1046 * conflict. If there is a conflict (both nodes match something
1047 * outside the bitmap, but what they match outside is not the same
1048 * pointer, and hence not easily compared until XXX we extend
1049 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
1050 * match everything outside the bitmap. If that stuff is all above
1051 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
1052 if (! ANYOF_NONBITMAP(cl)) {
1053 ARG_SET(cl, ARG(or_with));
1054 }
1055 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
1056
1057 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1058 cl_anything(pRExC_state, cl);
1059 }
1060 else {
1061 cl->flags |= ANYOF_UNICODE_ALL;
1062 }
c6b76537 1063 }
4c34a693 1064 }
0b9668ee
KW
1065
1066 /* Take the union */
1067 cl->flags |= or_with->flags;
1aa99e6b 1068 }
653099ff
GS
1069}
1070
a3621e74
YO
1071#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1072#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1073#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1074#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1075
3dab1dad
YO
1076
1077#ifdef DEBUGGING
07be1b83 1078/*
2b8b4781
NC
1079 dump_trie(trie,widecharmap,revcharmap)
1080 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1081 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1082
1083 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1084 The _interim_ variants are used for debugging the interim
1085 tables that are used to generate the final compressed
1086 representation which is what dump_trie expects.
1087
486ec47a 1088 Part of the reason for their existence is to provide a form
3dab1dad 1089 of documentation as to how the different representations function.
07be1b83
YO
1090
1091*/
3dab1dad
YO
1092
1093/*
3dab1dad
YO
1094 Dumps the final compressed table form of the trie to Perl_debug_log.
1095 Used for debugging make_trie().
1096*/
b9a59e08 1097
3dab1dad 1098STATIC void
2b8b4781
NC
1099S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1100 AV *revcharmap, U32 depth)
3dab1dad
YO
1101{
1102 U32 state;
ab3bbdeb 1103 SV *sv=sv_newmortal();
55eed653 1104 int colwidth= widecharmap ? 6 : 4;
2e64971a 1105 U16 word;
3dab1dad
YO
1106 GET_RE_DEBUG_FLAGS_DECL;
1107
7918f24d 1108 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1109
3dab1dad
YO
1110 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1111 (int)depth * 2 + 2,"",
1112 "Match","Base","Ofs" );
1113
1114 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1115 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1116 if ( tmp ) {
ab3bbdeb
YO
1117 PerlIO_printf( Perl_debug_log, "%*s",
1118 colwidth,
ddc5bc0f 1119 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1120 PL_colors[0], PL_colors[1],
1121 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1122 PERL_PV_ESCAPE_FIRSTCHAR
1123 )
1124 );
3dab1dad
YO
1125 }
1126 }
1127 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1128 (int)depth * 2 + 2,"");
1129
1130 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1131 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1132 PerlIO_printf( Perl_debug_log, "\n");
1133
1e2e3d02 1134 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1135 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1136
1137 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1138
1139 if ( trie->states[ state ].wordnum ) {
1140 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1141 } else {
1142 PerlIO_printf( Perl_debug_log, "%6s", "" );
1143 }
1144
1145 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1146
1147 if ( base ) {
1148 U32 ofs = 0;
1149
1150 while( ( base + ofs < trie->uniquecharcount ) ||
1151 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1152 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1153 ofs++;
1154
1155 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1156
1157 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1158 if ( ( base + ofs >= trie->uniquecharcount ) &&
1159 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1160 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1161 {
ab3bbdeb
YO
1162 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1163 colwidth,
3dab1dad
YO
1164 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1165 } else {
ab3bbdeb 1166 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1167 }
1168 }
1169
1170 PerlIO_printf( Perl_debug_log, "]");
1171
1172 }
1173 PerlIO_printf( Perl_debug_log, "\n" );
1174 }
2e64971a
DM
1175 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1176 for (word=1; word <= trie->wordcount; word++) {
1177 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1178 (int)word, (int)(trie->wordinfo[word].prev),
1179 (int)(trie->wordinfo[word].len));
1180 }
1181 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1182}
1183/*
3dab1dad
YO
1184 Dumps a fully constructed but uncompressed trie in list form.
1185 List tries normally only are used for construction when the number of
1186 possible chars (trie->uniquecharcount) is very high.
1187 Used for debugging make_trie().
1188*/
1189STATIC void
55eed653 1190S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1191 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1192 U32 depth)
3dab1dad
YO
1193{
1194 U32 state;
ab3bbdeb 1195 SV *sv=sv_newmortal();
55eed653 1196 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1197 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1198
1199 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1200
3dab1dad 1201 /* print out the table precompression. */
ab3bbdeb
YO
1202 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1203 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1204 "------:-----+-----------------\n" );
3dab1dad
YO
1205
1206 for( state=1 ; state < next_alloc ; state ++ ) {
1207 U16 charid;
1208
ab3bbdeb 1209 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1210 (int)depth * 2 + 2,"", (UV)state );
1211 if ( ! trie->states[ state ].wordnum ) {
1212 PerlIO_printf( Perl_debug_log, "%5s| ","");
1213 } else {
1214 PerlIO_printf( Perl_debug_log, "W%4x| ",
1215 trie->states[ state ].wordnum
1216 );
1217 }
1218 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1219 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1220 if ( tmp ) {
1221 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1222 colwidth,
ddc5bc0f 1223 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1224 PL_colors[0], PL_colors[1],
1225 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1226 PERL_PV_ESCAPE_FIRSTCHAR
1227 ) ,
1e2e3d02
YO
1228 TRIE_LIST_ITEM(state,charid).forid,
1229 (UV)TRIE_LIST_ITEM(state,charid).newstate
1230 );
1231 if (!(charid % 10))
664e119d
RGS
1232 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1233 (int)((depth * 2) + 14), "");
1e2e3d02 1234 }
ab3bbdeb
YO
1235 }
1236 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1237 }
1238}
1239
1240/*
3dab1dad
YO
1241 Dumps a fully constructed but uncompressed trie in table form.
1242 This is the normal DFA style state transition table, with a few
1243 twists to facilitate compression later.
1244 Used for debugging make_trie().
1245*/
1246STATIC void
55eed653 1247S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1248 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1249 U32 depth)
3dab1dad
YO
1250{
1251 U32 state;
1252 U16 charid;
ab3bbdeb 1253 SV *sv=sv_newmortal();
55eed653 1254 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1255 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1256
1257 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1258
1259 /*
1260 print out the table precompression so that we can do a visual check
1261 that they are identical.
1262 */
1263
1264 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1265
1266 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1267 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1268 if ( tmp ) {
ab3bbdeb
YO
1269 PerlIO_printf( Perl_debug_log, "%*s",
1270 colwidth,
ddc5bc0f 1271 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1272 PL_colors[0], PL_colors[1],
1273 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1274 PERL_PV_ESCAPE_FIRSTCHAR
1275 )
1276 );
3dab1dad
YO
1277 }
1278 }
1279
1280 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1281
1282 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1283 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1284 }
1285
1286 PerlIO_printf( Perl_debug_log, "\n" );
1287
1288 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1289
1290 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1291 (int)depth * 2 + 2,"",
1292 (UV)TRIE_NODENUM( state ) );
1293
1294 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1295 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1296 if (v)
1297 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1298 else
1299 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1300 }
1301 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1302 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1303 } else {
1304 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1305 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1306 }
1307 }
07be1b83 1308}
3dab1dad
YO
1309
1310#endif
1311
2e64971a 1312
786e8c11
YO
1313/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1314 startbranch: the first branch in the whole branch sequence
1315 first : start branch of sequence of branch-exact nodes.
1316 May be the same as startbranch
1317 last : Thing following the last branch.
1318 May be the same as tail.
1319 tail : item following the branch sequence
1320 count : words in the sequence
1321 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1322 depth : indent depth
3dab1dad 1323
786e8c11 1324Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1325
786e8c11
YO
1326A trie is an N'ary tree where the branches are determined by digital
1327decomposition of the key. IE, at the root node you look up the 1st character and
1328follow that branch repeat until you find the end of the branches. Nodes can be
1329marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1330
786e8c11 1331 /he|she|his|hers/
72f13be8 1332
786e8c11
YO
1333would convert into the following structure. Numbers represent states, letters
1334following numbers represent valid transitions on the letter from that state, if
1335the number is in square brackets it represents an accepting state, otherwise it
1336will be in parenthesis.
07be1b83 1337
786e8c11
YO
1338 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1339 | |
1340 | (2)
1341 | |
1342 (1) +-i->(6)-+-s->[7]
1343 |
1344 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1345
786e8c11
YO
1346 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1347
1348This shows that when matching against the string 'hers' we will begin at state 1
1349read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1350then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1351is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1352single traverse. We store a mapping from accepting to state to which word was
1353matched, and then when we have multiple possibilities we try to complete the
1354rest of the regex in the order in which they occured in the alternation.
1355
1356The only prior NFA like behaviour that would be changed by the TRIE support is
1357the silent ignoring of duplicate alternations which are of the form:
1358
1359 / (DUPE|DUPE) X? (?{ ... }) Y /x
1360
4b714af6 1361Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1362and without the optimisation. With the optimisations dupes will be silently
486ec47a 1363ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1364the following demonstrates:
1365
1366 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1367
1368which prints out 'word' three times, but
1369
1370 'words'=~/(word|word|word)(?{ print $1 })S/
1371
1372which doesnt print it out at all. This is due to other optimisations kicking in.
1373
1374Example of what happens on a structural level:
1375
486ec47a 1376The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1377
1378 1: CURLYM[1] {1,32767}(18)
1379 5: BRANCH(8)
1380 6: EXACT <ac>(16)
1381 8: BRANCH(11)
1382 9: EXACT <ad>(16)
1383 11: BRANCH(14)
1384 12: EXACT <ab>(16)
1385 16: SUCCEED(0)
1386 17: NOTHING(18)
1387 18: END(0)
1388
1389This would be optimizable with startbranch=5, first=5, last=16, tail=16
1390and should turn into:
1391
1392 1: CURLYM[1] {1,32767}(18)
1393 5: TRIE(16)
1394 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1395 <ac>
1396 <ad>
1397 <ab>
1398 16: SUCCEED(0)
1399 17: NOTHING(18)
1400 18: END(0)
1401
1402Cases where tail != last would be like /(?foo|bar)baz/:
1403
1404 1: BRANCH(4)
1405 2: EXACT <foo>(8)
1406 4: BRANCH(7)
1407 5: EXACT <bar>(8)
1408 7: TAIL(8)
1409 8: EXACT <baz>(10)
1410 10: END(0)
1411
1412which would be optimizable with startbranch=1, first=1, last=7, tail=8
1413and would end up looking like:
1414
1415 1: TRIE(8)
1416 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1417 <foo>
1418 <bar>
1419 7: TAIL(8)
1420 8: EXACT <baz>(10)
1421 10: END(0)
1422
1423 d = uvuni_to_utf8_flags(d, uv, 0);
1424
1425is the recommended Unicode-aware way of saying
1426
1427 *(d++) = uv;
1428*/
1429
fab2782b 1430#define TRIE_STORE_REVCHAR(val) \
786e8c11 1431 STMT_START { \
73031816 1432 if (UTF) { \
fab2782b 1433 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1434 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
fab2782b 1435 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
73031816
NC
1436 SvCUR_set(zlopp, kapow - flrbbbbb); \
1437 SvPOK_on(zlopp); \
1438 SvUTF8_on(zlopp); \
1439 av_push(revcharmap, zlopp); \
1440 } else { \
fab2782b 1441 char ooooff = (char)val; \
73031816
NC
1442 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1443 } \
1444 } STMT_END
786e8c11 1445
fab2782b
YO
1446#define TRIE_READ_CHAR STMT_START { \
1447 wordlen++; \
1448 if ( UTF ) { \
1449 /* if it is UTF then it is either already folded, or does not need folding */ \
1450 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1451 } \
1452 else if (folder == PL_fold_latin1) { \
1453 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1454 if ( foldlen > 0 ) { \
1455 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1456 foldlen -= len; \
1457 scan += len; \
1458 len = 0; \
1459 } else { \
1460 len = 1; \
51910141 1461 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, FOLD_FLAGS_FULL); \
fab2782b
YO
1462 skiplen = UNISKIP(uvc); \
1463 foldlen -= skiplen; \
1464 scan = foldbuf + skiplen; \
1465 } \
1466 } else { \
1467 /* raw data, will be folded later if needed */ \
1468 uvc = (U32)*uc; \
1469 len = 1; \
1470 } \
786e8c11
YO
1471} STMT_END
1472
1473
1474
1475#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1476 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1477 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1478 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1479 } \
1480 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1481 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1482 TRIE_LIST_CUR( state )++; \
1483} STMT_END
07be1b83 1484
786e8c11
YO
1485#define TRIE_LIST_NEW(state) STMT_START { \
1486 Newxz( trie->states[ state ].trans.list, \
1487 4, reg_trie_trans_le ); \
1488 TRIE_LIST_CUR( state ) = 1; \
1489 TRIE_LIST_LEN( state ) = 4; \
1490} STMT_END
07be1b83 1491
786e8c11
YO
1492#define TRIE_HANDLE_WORD(state) STMT_START { \
1493 U16 dupe= trie->states[ state ].wordnum; \
1494 regnode * const noper_next = regnext( noper ); \
1495 \
786e8c11
YO
1496 DEBUG_r({ \
1497 /* store the word for dumping */ \
1498 SV* tmp; \
1499 if (OP(noper) != NOTHING) \
740cce10 1500 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1501 else \
740cce10 1502 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1503 av_push( trie_words, tmp ); \
786e8c11
YO
1504 }); \
1505 \
1506 curword++; \
2e64971a
DM
1507 trie->wordinfo[curword].prev = 0; \
1508 trie->wordinfo[curword].len = wordlen; \
1509 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1510 \
1511 if ( noper_next < tail ) { \
1512 if (!trie->jump) \
c944940b 1513 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1514 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1515 if (!jumper) \
1516 jumper = noper_next; \
1517 if (!nextbranch) \
1518 nextbranch= regnext(cur); \
1519 } \
1520 \
1521 if ( dupe ) { \
2e64971a
DM
1522 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1523 /* chain, so that when the bits of chain are later */\
1524 /* linked together, the dups appear in the chain */\
1525 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1526 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1527 } else { \
1528 /* we haven't inserted this word yet. */ \
1529 trie->states[ state ].wordnum = curword; \
1530 } \
1531} STMT_END
07be1b83 1532
3dab1dad 1533
786e8c11
YO
1534#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1535 ( ( base + charid >= ucharcount \
1536 && base + charid < ubound \
1537 && state == trie->trans[ base - ucharcount + charid ].check \
1538 && trie->trans[ base - ucharcount + charid ].next ) \
1539 ? trie->trans[ base - ucharcount + charid ].next \
1540 : ( state==1 ? special : 0 ) \
1541 )
3dab1dad 1542
786e8c11
YO
1543#define MADE_TRIE 1
1544#define MADE_JUMP_TRIE 2
1545#define MADE_EXACT_TRIE 4
3dab1dad 1546
a3621e74 1547STATIC I32
786e8c11 1548S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1549{
27da23d5 1550 dVAR;
a3621e74
YO
1551 /* first pass, loop through and scan words */
1552 reg_trie_data *trie;
55eed653 1553 HV *widecharmap = NULL;
2b8b4781 1554 AV *revcharmap = newAV();
a3621e74 1555 regnode *cur;
9f7f3913 1556 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1557 STRLEN len = 0;
1558 UV uvc = 0;
1559 U16 curword = 0;
1560 U32 next_alloc = 0;
786e8c11
YO
1561 regnode *jumper = NULL;
1562 regnode *nextbranch = NULL;
7f69552c 1563 regnode *convert = NULL;
2e64971a 1564 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1565 /* we just use folder as a flag in utf8 */
1e696034 1566 const U8 * folder = NULL;
a3621e74 1567
2b8b4781
NC
1568#ifdef DEBUGGING
1569 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1570 AV *trie_words = NULL;
1571 /* along with revcharmap, this only used during construction but both are
1572 * useful during debugging so we store them in the struct when debugging.
8e11feef 1573 */
2b8b4781
NC
1574#else
1575 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1576 STRLEN trie_charcount=0;
3dab1dad 1577#endif
2b8b4781 1578 SV *re_trie_maxbuff;
a3621e74 1579 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1580
1581 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1582#ifndef DEBUGGING
1583 PERL_UNUSED_ARG(depth);
1584#endif
a3621e74 1585
1e696034 1586 switch (flags) {
79a81a6e 1587 case EXACT: break;
2f7f8cb1 1588 case EXACTFA:
fab2782b
YO
1589 case EXACTFU_SS:
1590 case EXACTFU_TRICKYFOLD:
1e696034
KW
1591 case EXACTFU: folder = PL_fold_latin1; break;
1592 case EXACTF: folder = PL_fold; break;
1593 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1594 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1595 }
1596
c944940b 1597 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1598 trie->refcount = 1;
3dab1dad 1599 trie->startstate = 1;
786e8c11 1600 trie->wordcount = word_count;
f8fc2ecf 1601 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1602 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1603 if (flags == EXACT)
c944940b 1604 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1605 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1606 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1607
a3621e74 1608 DEBUG_r({
2b8b4781 1609 trie_words = newAV();
a3621e74 1610 });
a3621e74 1611
0111c4fd 1612 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1613 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1614 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1615 }
df826430 1616 DEBUG_TRIE_COMPILE_r({
3dab1dad 1617 PerlIO_printf( Perl_debug_log,
786e8c11 1618 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1619 (int)depth * 2 + 2, "",
1620 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1621 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1622 (int)depth);
3dab1dad 1623 });
7f69552c
YO
1624
1625 /* Find the node we are going to overwrite */
1626 if ( first == startbranch && OP( last ) != BRANCH ) {
1627 /* whole branch chain */
1628 convert = first;
1629 } else {
1630 /* branch sub-chain */
1631 convert = NEXTOPER( first );
1632 }
1633
a3621e74
YO
1634 /* -- First loop and Setup --
1635
1636 We first traverse the branches and scan each word to determine if it
1637 contains widechars, and how many unique chars there are, this is
1638 important as we have to build a table with at least as many columns as we
1639 have unique chars.
1640
1641 We use an array of integers to represent the character codes 0..255
38a44b82 1642 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1643 native representation of the character value as the key and IV's for the
1644 coded index.
1645
1646 *TODO* If we keep track of how many times each character is used we can
1647 remap the columns so that the table compression later on is more
3b753521 1648 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1649 middle and the least common are on the outside. IMO this would be better
1650 than a most to least common mapping as theres a decent chance the most
1651 common letter will share a node with the least common, meaning the node
486ec47a 1652 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1653 case is when we have the least common nodes twice.
1654
1655 */
1656
a3621e74 1657 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1658 regnode *noper = NEXTOPER( cur );
e1ec3a88 1659 const U8 *uc = (U8*)STRING( noper );
df826430 1660 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1661 STRLEN foldlen = 0;
1662 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1663 STRLEN skiplen = 0;
2af232bd 1664 const U8 *scan = (U8*)NULL;
07be1b83 1665 U32 wordlen = 0; /* required init */
02daf0ab
YO
1666 STRLEN chars = 0;
1667 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1668
3dab1dad 1669 if (OP(noper) == NOTHING) {
df826430
YO
1670 regnode *noper_next= regnext(noper);
1671 if (noper_next != tail && OP(noper_next) == flags) {
1672 noper = noper_next;
1673 uc= (U8*)STRING(noper);
1674 e= uc + STR_LEN(noper);
1675 trie->minlen= STR_LEN(noper);
1676 } else {
1677 trie->minlen= 0;
1678 continue;
1679 }
3dab1dad 1680 }
df826430 1681
fab2782b 1682 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1683 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1684 regardless of encoding */
fab2782b
YO
1685 if (OP( noper ) == EXACTFU_SS) {
1686 /* false positives are ok, so just set this */
1687 TRIE_BITMAP_SET(trie,0xDF);
1688 }
1689 }
a3621e74 1690 for ( ; uc < e ; uc += len ) {
3dab1dad 1691 TRIE_CHARCOUNT(trie)++;
a3621e74 1692 TRIE_READ_CHAR;
3dab1dad 1693 chars++;
a3621e74 1694 if ( uvc < 256 ) {
fab2782b
YO
1695 if ( folder ) {
1696 U8 folded= folder[ (U8) uvc ];
1697 if ( !trie->charmap[ folded ] ) {
1698 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1699 TRIE_STORE_REVCHAR( folded );
1700 }
1701 }
a3621e74
YO
1702 if ( !trie->charmap[ uvc ] ) {
1703 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1704 TRIE_STORE_REVCHAR( uvc );
a3621e74 1705 }
02daf0ab 1706 if ( set_bit ) {
62012aee
KW
1707 /* store the codepoint in the bitmap, and its folded
1708 * equivalent. */
fab2782b 1709 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1710
1711 /* store the folded codepoint */
fab2782b 1712 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1713
1714 if ( !UTF ) {
1715 /* store first byte of utf8 representation of
acdf4139
KW
1716 variant codepoints */
1717 if (! UNI_IS_INVARIANT(uvc)) {
1718 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1719 }
1720 }
02daf0ab
YO
1721 set_bit = 0; /* We've done our bit :-) */
1722 }
a3621e74
YO
1723 } else {
1724 SV** svpp;
55eed653
NC
1725 if ( !widecharmap )
1726 widecharmap = newHV();
a3621e74 1727
55eed653 1728 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1729
1730 if ( !svpp )
e4584336 1731 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1732
1733 if ( !SvTRUE( *svpp ) ) {
1734 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1735 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1736 }
1737 }
1738 }
3dab1dad 1739 if( cur == first ) {
fab2782b
YO
1740 trie->minlen = chars;
1741 trie->maxlen = chars;
3dab1dad 1742 } else if (chars < trie->minlen) {
fab2782b 1743 trie->minlen = chars;
3dab1dad 1744 } else if (chars > trie->maxlen) {
fab2782b
YO
1745 trie->maxlen = chars;
1746 }
1747 if (OP( noper ) == EXACTFU_SS) {
1748 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1749 if (trie->minlen > 1)
1750 trie->minlen= 1;
1751 }
1752 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1753 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1754 * - We assume that any such sequence might match a 2 byte string */
1755 if (trie->minlen > 2 )
1756 trie->minlen= 2;
3dab1dad
YO
1757 }
1758
a3621e74
YO
1759 } /* end first pass */
1760 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1761 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1762 (int)depth * 2 + 2,"",
55eed653 1763 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1764 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1765 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1766 );
a3621e74
YO
1767
1768 /*
1769 We now know what we are dealing with in terms of unique chars and
1770 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1771 representation using a flat table will take. If it's over a reasonable
1772 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1773 conservative but potentially much slower representation using an array
1774 of lists.
1775
1776 At the end we convert both representations into the same compressed
1777 form that will be used in regexec.c for matching with. The latter
1778 is a form that cannot be used to construct with but has memory
1779 properties similar to the list form and access properties similar
1780 to the table form making it both suitable for fast searches and
1781 small enough that its feasable to store for the duration of a program.
1782
1783 See the comment in the code where the compressed table is produced
1784 inplace from the flat tabe representation for an explanation of how
1785 the compression works.
1786
1787 */
1788
1789
2e64971a
DM
1790 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1791 prev_states[1] = 0;
1792
3dab1dad 1793 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1794 /*
1795 Second Pass -- Array Of Lists Representation
1796
1797 Each state will be represented by a list of charid:state records
1798 (reg_trie_trans_le) the first such element holds the CUR and LEN
1799 points of the allocated array. (See defines above).
1800
1801 We build the initial structure using the lists, and then convert
1802 it into the compressed table form which allows faster lookups
1803 (but cant be modified once converted).
a3621e74
YO
1804 */
1805
a3621e74
YO
1806 STRLEN transcount = 1;
1807
1e2e3d02
YO
1808 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1809 "%*sCompiling trie using list compiler\n",
1810 (int)depth * 2 + 2, ""));
686b73d4 1811
c944940b
JH
1812 trie->states = (reg_trie_state *)
1813 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1814 sizeof(reg_trie_state) );
a3621e74
YO
1815 TRIE_LIST_NEW(1);
1816 next_alloc = 2;
1817
1818 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1819
df826430 1820 regnode *noper = NEXTOPER( cur );
c445ea15 1821 U8 *uc = (U8*)STRING( noper );
df826430 1822 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1823 U32 state = 1; /* required init */
1824 U16 charid = 0; /* sanity init */
1825 U8 *scan = (U8*)NULL; /* sanity init */
1826 STRLEN foldlen = 0; /* required init */
07be1b83 1827 U32 wordlen = 0; /* required init */
c445ea15 1828 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1829 STRLEN skiplen = 0;
c445ea15 1830
df826430
YO
1831 if (OP(noper) == NOTHING) {
1832 regnode *noper_next= regnext(noper);
1833 if (noper_next != tail && OP(noper_next) == flags) {
1834 noper = noper_next;
1835 uc= (U8*)STRING(noper);
1836 e= uc + STR_LEN(noper);
1837 }
1838 }
1839
3dab1dad 1840 if (OP(noper) != NOTHING) {
786e8c11 1841 for ( ; uc < e ; uc += len ) {
c445ea15 1842
786e8c11 1843 TRIE_READ_CHAR;
c445ea15 1844
786e8c11
YO
1845 if ( uvc < 256 ) {
1846 charid = trie->charmap[ uvc ];
c445ea15 1847 } else {
55eed653 1848 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1849 if ( !svpp ) {
1850 charid = 0;
1851 } else {
1852 charid=(U16)SvIV( *svpp );
1853 }
c445ea15 1854 }
786e8c11
YO
1855 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1856 if ( charid ) {
a3621e74 1857
786e8c11
YO
1858 U16 check;
1859 U32 newstate = 0;
a3621e74 1860
786e8c11
YO
1861 charid--;
1862 if ( !trie->states[ state ].trans.list ) {
1863 TRIE_LIST_NEW( state );
c445ea15 1864 }
786e8c11
YO
1865 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1866 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1867 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1868 break;
1869 }
1870 }
1871 if ( ! newstate ) {
1872 newstate = next_alloc++;
2e64971a 1873 prev_states[newstate] = state;
786e8c11
YO
1874 TRIE_LIST_PUSH( state, charid, newstate );
1875 transcount++;
1876 }
1877 state = newstate;
1878 } else {
1879 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1880 }
a28509cc 1881 }
c445ea15 1882 }
3dab1dad 1883 TRIE_HANDLE_WORD(state);
a3621e74
YO
1884
1885 } /* end second pass */
1886
1e2e3d02
YO
1887 /* next alloc is the NEXT state to be allocated */
1888 trie->statecount = next_alloc;
c944940b
JH
1889 trie->states = (reg_trie_state *)
1890 PerlMemShared_realloc( trie->states,
1891 next_alloc
1892 * sizeof(reg_trie_state) );
a3621e74 1893
3dab1dad 1894 /* and now dump it out before we compress it */
2b8b4781
NC
1895 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1896 revcharmap, next_alloc,
1897 depth+1)
1e2e3d02 1898 );
a3621e74 1899
c944940b
JH
1900 trie->trans = (reg_trie_trans *)
1901 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1902 {
1903 U32 state;
a3621e74
YO
1904 U32 tp = 0;
1905 U32 zp = 0;
1906
1907
1908 for( state=1 ; state < next_alloc ; state ++ ) {
1909 U32 base=0;
1910
1911 /*
1912 DEBUG_TRIE_COMPILE_MORE_r(
1913 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1914 );
1915 */
1916
1917 if (trie->states[state].trans.list) {
1918 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1919 U16 maxid=minid;
a28509cc 1920 U16 idx;
a3621e74
YO
1921
1922 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1923 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1924 if ( forid < minid ) {
1925 minid=forid;
1926 } else if ( forid > maxid ) {
1927 maxid=forid;
1928 }
a3621e74
YO
1929 }
1930 if ( transcount < tp + maxid - minid + 1) {
1931 transcount *= 2;
c944940b
JH
1932 trie->trans = (reg_trie_trans *)
1933 PerlMemShared_realloc( trie->trans,
446bd890
NC
1934 transcount
1935 * sizeof(reg_trie_trans) );
a3621e74
YO
1936 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1937 }
1938 base = trie->uniquecharcount + tp - minid;
1939 if ( maxid == minid ) {
1940 U32 set = 0;
1941 for ( ; zp < tp ; zp++ ) {
1942 if ( ! trie->trans[ zp ].next ) {
1943 base = trie->uniquecharcount + zp - minid;
1944 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1945 trie->trans[ zp ].check = state;
1946 set = 1;
1947 break;
1948 }
1949 }
1950 if ( !set ) {
1951 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1952 trie->trans[ tp ].check = state;
1953 tp++;
1954 zp = tp;
1955 }
1956 } else {
1957 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1958 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1959 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1960 trie->trans[ tid ].check = state;
1961 }
1962 tp += ( maxid - minid + 1 );
1963 }
1964 Safefree(trie->states[ state ].trans.list);
1965 }
1966 /*
1967 DEBUG_TRIE_COMPILE_MORE_r(
1968 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1969 );
1970 */
1971 trie->states[ state ].trans.base=base;
1972 }
cc601c31 1973 trie->lasttrans = tp + 1;
a3621e74
YO
1974 }
1975 } else {
1976 /*
1977 Second Pass -- Flat Table Representation.
1978
1979 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1980 We know that we will need Charcount+1 trans at most to store the data
1981 (one row per char at worst case) So we preallocate both structures
1982 assuming worst case.
1983
1984 We then construct the trie using only the .next slots of the entry
1985 structs.
1986
3b753521 1987 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1988 make compression both faster and easier by keeping track of how many non
1989 zero fields are in the node.
1990
1991 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1992 transition.
1993
1994 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1995 number representing the first entry of the node, and state as a
1996 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1997 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1998 are 2 entrys per node. eg:
1999
2000 A B A B
2001 1. 2 4 1. 3 7
2002 2. 0 3 3. 0 5
2003 3. 0 0 5. 0 0
2004 4. 0 0 7. 0 0
2005
2006 The table is internally in the right hand, idx form. However as we also
2007 have to deal with the states array which is indexed by nodenum we have to
2008 use TRIE_NODENUM() to convert.
2009
2010 */
1e2e3d02
YO
2011 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2012 "%*sCompiling trie using table compiler\n",
2013 (int)depth * 2 + 2, ""));
3dab1dad 2014
c944940b
JH
2015 trie->trans = (reg_trie_trans *)
2016 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2017 * trie->uniquecharcount + 1,
2018 sizeof(reg_trie_trans) );
2019 trie->states = (reg_trie_state *)
2020 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2021 sizeof(reg_trie_state) );
a3621e74
YO
2022 next_alloc = trie->uniquecharcount + 1;
2023
3dab1dad 2024
a3621e74
YO
2025 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2026
df826430 2027 regnode *noper = NEXTOPER( cur );
a28509cc 2028 const U8 *uc = (U8*)STRING( noper );
df826430 2029 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
2030
2031 U32 state = 1; /* required init */
2032
2033 U16 charid = 0; /* sanity init */
2034 U32 accept_state = 0; /* sanity init */
2035 U8 *scan = (U8*)NULL; /* sanity init */
2036
2037 STRLEN foldlen = 0; /* required init */
07be1b83 2038 U32 wordlen = 0; /* required init */
fab2782b 2039 STRLEN skiplen = 0;
a3621e74
YO
2040 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2041
df826430
YO
2042 if (OP(noper) == NOTHING) {
2043 regnode *noper_next= regnext(noper);
2044 if (noper_next != tail && OP(noper_next) == flags) {
2045 noper = noper_next;
2046 uc= (U8*)STRING(noper);
2047 e= uc + STR_LEN(noper);
2048 }
2049 }
fab2782b 2050
3dab1dad 2051 if ( OP(noper) != NOTHING ) {
786e8c11 2052 for ( ; uc < e ; uc += len ) {
a3621e74 2053
786e8c11 2054 TRIE_READ_CHAR;
a3621e74 2055
786e8c11
YO
2056 if ( uvc < 256 ) {
2057 charid = trie->charmap[ uvc ];
2058 } else {
55eed653 2059 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2060 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2061 }
786e8c11
YO
2062 if ( charid ) {
2063 charid--;
2064 if ( !trie->trans[ state + charid ].next ) {
2065 trie->trans[ state + charid ].next = next_alloc;
2066 trie->trans[ state ].check++;
2e64971a
DM
2067 prev_states[TRIE_NODENUM(next_alloc)]
2068 = TRIE_NODENUM(state);
786e8c11
YO
2069 next_alloc += trie->uniquecharcount;
2070 }
2071 state = trie->trans[ state + charid ].next;
2072 } else {
2073 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2074 }
2075 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2076 }
a3621e74 2077 }
3dab1dad
YO
2078 accept_state = TRIE_NODENUM( state );
2079 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2080
2081 } /* end second pass */
2082
3dab1dad 2083 /* and now dump it out before we compress it */
2b8b4781
NC
2084 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2085 revcharmap,
2086 next_alloc, depth+1));
a3621e74 2087
a3621e74
YO
2088 {
2089 /*
2090 * Inplace compress the table.*
2091
2092 For sparse data sets the table constructed by the trie algorithm will
2093 be mostly 0/FAIL transitions or to put it another way mostly empty.
2094 (Note that leaf nodes will not contain any transitions.)
2095
2096 This algorithm compresses the tables by eliminating most such
2097 transitions, at the cost of a modest bit of extra work during lookup:
2098
2099 - Each states[] entry contains a .base field which indicates the
2100 index in the state[] array wheres its transition data is stored.
2101
3b753521 2102 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2103
2104 - If .base is nonzero then charid is added to it to find an entry in
2105 the trans array.
2106
2107 -If trans[states[state].base+charid].check!=state then the
2108 transition is taken to be a 0/Fail transition. Thus if there are fail
2109 transitions at the front of the node then the .base offset will point
2110 somewhere inside the previous nodes data (or maybe even into a node
2111 even earlier), but the .check field determines if the transition is
2112 valid.
2113
786e8c11 2114 XXX - wrong maybe?
a3621e74 2115 The following process inplace converts the table to the compressed
3b753521 2116 table: We first do not compress the root node 1,and mark all its
a3621e74 2117 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2118 allows us to do a DFA construction from the compressed table later,
2119 and ensures that any .base pointers we calculate later are greater
2120 than 0.
a3621e74
YO
2121
2122 - We set 'pos' to indicate the first entry of the second node.
2123
2124 - We then iterate over the columns of the node, finding the first and
2125 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2126 and set the .check pointers accordingly, and advance pos
2127 appropriately and repreat for the next node. Note that when we copy
2128 the next pointers we have to convert them from the original
2129 NODEIDX form to NODENUM form as the former is not valid post
2130 compression.
2131
2132 - If a node has no transitions used we mark its base as 0 and do not
2133 advance the pos pointer.
2134
2135 - If a node only has one transition we use a second pointer into the
2136 structure to fill in allocated fail transitions from other states.
2137 This pointer is independent of the main pointer and scans forward
2138 looking for null transitions that are allocated to a state. When it
2139 finds one it writes the single transition into the "hole". If the
786e8c11 2140 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2141
2142 - Once compressed we can Renew/realloc the structures to release the
2143 excess space.
2144
2145 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2146 specifically Fig 3.47 and the associated pseudocode.
2147
2148 demq
2149 */
a3b680e6 2150 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2151 U32 state, charid;
a3621e74 2152 U32 pos = 0, zp=0;
1e2e3d02 2153 trie->statecount = laststate;
a3621e74
YO
2154
2155 for ( state = 1 ; state < laststate ; state++ ) {
2156 U8 flag = 0;
a28509cc
AL
2157 const U32 stateidx = TRIE_NODEIDX( state );
2158 const U32 o_used = trie->trans[ stateidx ].check;
2159 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2160 trie->trans[ stateidx ].check = 0;
2161
2162 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2163 if ( flag || trie->trans[ stateidx + charid ].next ) {
2164 if ( trie->trans[ stateidx + charid ].next ) {
2165 if (o_used == 1) {
2166 for ( ; zp < pos ; zp++ ) {
2167 if ( ! trie->trans[ zp ].next ) {
2168 break;
2169 }
2170 }
2171 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2172 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2173 trie->trans[ zp ].check = state;
2174 if ( ++zp > pos ) pos = zp;
2175 break;
2176 }
2177 used--;
2178 }
2179 if ( !flag ) {
2180 flag = 1;
2181 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2182 }
2183 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2184 trie->trans[ pos ].check = state;
2185 pos++;
2186 }
2187 }
2188 }
cc601c31 2189 trie->lasttrans = pos + 1;
c944940b
JH
2190 trie->states = (reg_trie_state *)
2191 PerlMemShared_realloc( trie->states, laststate
2192 * sizeof(reg_trie_state) );
a3621e74 2193 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2194 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2195 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2196 (int)depth * 2 + 2,"",
2197 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2198 (IV)next_alloc,
2199 (IV)pos,
a3621e74
YO
2200 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2201 );
2202
2203 } /* end table compress */
2204 }
1e2e3d02
YO
2205 DEBUG_TRIE_COMPILE_MORE_r(
2206 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2207 (int)depth * 2 + 2, "",
2208 (UV)trie->statecount,
2209 (UV)trie->lasttrans)
2210 );
cc601c31 2211 /* resize the trans array to remove unused space */
c944940b
JH
2212 trie->trans = (reg_trie_trans *)
2213 PerlMemShared_realloc( trie->trans, trie->lasttrans
2214 * sizeof(reg_trie_trans) );
a3621e74 2215
3b753521 2216 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2217 U8 nodetype =(U8)(flags & 0xFF);
2218 char *str=NULL;
786e8c11 2219
07be1b83 2220#ifdef DEBUGGING
e62cc96a 2221 regnode *optimize = NULL;
7122b237
YO
2222#ifdef RE_TRACK_PATTERN_OFFSETS
2223
b57a0404
JH
2224 U32 mjd_offset = 0;
2225 U32 mjd_nodelen = 0;
7122b237
YO
2226#endif /* RE_TRACK_PATTERN_OFFSETS */
2227#endif /* DEBUGGING */
a3621e74 2228 /*
3dab1dad
YO
2229 This means we convert either the first branch or the first Exact,
2230 depending on whether the thing following (in 'last') is a branch
2231 or not and whther first is the startbranch (ie is it a sub part of
2232 the alternation or is it the whole thing.)
3b753521 2233 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2234 the whole branch sequence, including the first.
a3621e74 2235 */
3dab1dad 2236 /* Find the node we are going to overwrite */
7f69552c 2237 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2238 /* branch sub-chain */
3dab1dad 2239 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2240#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2241 DEBUG_r({
2242 mjd_offset= Node_Offset((convert));
2243 mjd_nodelen= Node_Length((convert));
2244 });
7122b237 2245#endif
7f69552c 2246 /* whole branch chain */
7122b237
YO
2247 }
2248#ifdef RE_TRACK_PATTERN_OFFSETS
2249 else {
7f69552c
YO
2250 DEBUG_r({
2251 const regnode *nop = NEXTOPER( convert );
2252 mjd_offset= Node_Offset((nop));
2253 mjd_nodelen= Node_Length((nop));
2254 });
07be1b83
YO
2255 }
2256 DEBUG_OPTIMISE_r(
2257 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2258 (int)depth * 2 + 2, "",
786e8c11 2259 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2260 );
7122b237 2261#endif
3dab1dad
YO
2262 /* But first we check to see if there is a common prefix we can
2263 split out as an EXACT and put in front of the TRIE node. */
2264 trie->startstate= 1;
55eed653 2265 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2266 U32 state;
1e2e3d02 2267 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2268 U32 ofs = 0;
8e11feef
RGS
2269 I32 idx = -1;
2270 U32 count = 0;
2271 const U32 base = trie->states[ state ].trans.base;
a3621e74 2272
3dab1dad 2273 if ( trie->states[state].wordnum )
8e11feef 2274 count = 1;
a3621e74 2275
8e11feef 2276 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2277 if ( ( base + ofs >= trie->uniquecharcount ) &&
2278 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2279 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2280 {
3dab1dad 2281 if ( ++count > 1 ) {
2b8b4781 2282 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2283 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2284 if ( state == 1 ) break;
3dab1dad
YO
2285 if ( count == 2 ) {
2286 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2287 DEBUG_OPTIMISE_r(
8e11feef
RGS
2288 PerlIO_printf(Perl_debug_log,
2289 "%*sNew Start State=%"UVuf" Class: [",
2290 (int)depth * 2 + 2, "",
786e8c11 2291 (UV)state));
be8e71aa 2292 if (idx >= 0) {
2b8b4781 2293 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2294 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2295
3dab1dad 2296 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2297 if ( folder )
2298 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2299 DEBUG_OPTIMISE_r(
f1f66076 2300 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2301 );
8e11feef
RGS
2302 }
2303 }
2304 TRIE_BITMAP_SET(trie,*ch);
2305 if ( folder )
2306 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2307 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2308 }
2309 idx = ofs;
2310 }
3dab1dad
YO
2311 }
2312 if ( count == 1 ) {
2b8b4781 2313 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2314 STRLEN len;
2315 char *ch = SvPV( *tmp, len );
de734bd5
A
2316 DEBUG_OPTIMISE_r({
2317 SV *sv=sv_newmortal();
8e11feef
RGS
2318 PerlIO_printf( Perl_debug_log,
2319 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2320 (int)depth * 2 + 2, "",
de734bd5
A
2321 (UV)state, (UV)idx,
2322 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2323 PL_colors[0], PL_colors[1],
2324 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2325 PERL_PV_ESCAPE_FIRSTCHAR
2326 )
2327 );
2328 });
3dab1dad
YO
2329 if ( state==1 ) {
2330 OP( convert ) = nodetype;
2331 str=STRING(convert);
2332 STR_LEN(convert)=0;
2333 }
c490c714
YO
2334 STR_LEN(convert) += len;
2335 while (len--)
de734bd5 2336 *str++ = *ch++;
8e11feef 2337 } else {
f9049ba1 2338#ifdef DEBUGGING
8e11feef
RGS
2339 if (state>1)
2340 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2341#endif
8e11feef
RGS
2342 break;
2343 }
2344 }
2e64971a 2345 trie->prefixlen = (state-1);
3dab1dad 2346 if (str) {
8e11feef 2347 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2348 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2349 trie->startstate = state;
07be1b83
YO
2350 trie->minlen -= (state - 1);
2351 trie->maxlen -= (state - 1);
33809eae
JH
2352#ifdef DEBUGGING
2353 /* At least the UNICOS C compiler choked on this
2354 * being argument to DEBUG_r(), so let's just have
2355 * it right here. */
2356 if (
2357#ifdef PERL_EXT_RE_BUILD
2358 1
2359#else
2360 DEBUG_r_TEST
2361#endif
2362 ) {
2363 regnode *fix = convert;
2364 U32 word = trie->wordcount;
2365 mjd_nodelen++;
2366 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2367 while( ++fix < n ) {
2368 Set_Node_Offset_Length(fix, 0, 0);
2369 }
2370 while (word--) {
2371 SV ** const tmp = av_fetch( trie_words, word, 0 );
2372 if (tmp) {
2373 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2374 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2375 else
2376 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2377 }
2378 }
2379 }
2380#endif
8e11feef
RGS
2381 if (trie->maxlen) {
2382 convert = n;
2383 } else {
3dab1dad 2384 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2385 DEBUG_r(optimize= n);
3dab1dad
YO
2386 }
2387 }
2388 }
a5ca303d
YO
2389 if (!jumper)
2390 jumper = last;
3dab1dad 2391 if ( trie->maxlen ) {
8e11feef
RGS
2392 NEXT_OFF( convert ) = (U16)(tail - convert);
2393 ARG_SET( convert, data_slot );
786e8c11
YO
2394 /* Store the offset to the first unabsorbed branch in
2395 jump[0], which is otherwise unused by the jump logic.
2396 We use this when dumping a trie and during optimisation. */
2397 if (trie->jump)
7f69552c 2398 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2399
6c48061a
YO
2400 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2401 * and there is a bitmap
2402 * and the first "jump target" node we found leaves enough room
2403 * then convert the TRIE node into a TRIEC node, with the bitmap
2404 * embedded inline in the opcode - this is hypothetically faster.
2405 */
2406 if ( !trie->states[trie->startstate].wordnum
2407 && trie->bitmap
2408 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2409 {
2410 OP( convert ) = TRIEC;
2411 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2412 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2413 trie->bitmap= NULL;
2414 } else
2415 OP( convert ) = TRIE;
a3621e74 2416
3dab1dad
YO
2417 /* store the type in the flags */
2418 convert->flags = nodetype;
a5ca303d
YO
2419 DEBUG_r({
2420 optimize = convert
2421 + NODE_STEP_REGNODE
2422 + regarglen[ OP( convert ) ];
2423 });
2424 /* XXX We really should free up the resource in trie now,
2425 as we won't use them - (which resources?) dmq */
3dab1dad 2426 }
a3621e74 2427 /* needed for dumping*/
e62cc96a 2428 DEBUG_r(if (optimize) {
07be1b83 2429 regnode *opt = convert;
bcdf7404 2430
e62cc96a 2431 while ( ++opt < optimize) {
07be1b83
YO
2432 Set_Node_Offset_Length(opt,0,0);
2433 }
786e8c11
YO
2434 /*
2435 Try to clean up some of the debris left after the
2436 optimisation.
a3621e74 2437 */
786e8c11 2438 while( optimize < jumper ) {
07be1b83 2439 mjd_nodelen += Node_Length((optimize));
a3621e74 2440 OP( optimize ) = OPTIMIZED;
07be1b83 2441 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2442 optimize++;
2443 }
07be1b83 2444 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2445 });
2446 } /* end node insert */
2e64971a
DM
2447
2448 /* Finish populating the prev field of the wordinfo array. Walk back
2449 * from each accept state until we find another accept state, and if
2450 * so, point the first word's .prev field at the second word. If the
2451 * second already has a .prev field set, stop now. This will be the
2452 * case either if we've already processed that word's accept state,
3b753521
FN
2453 * or that state had multiple words, and the overspill words were
2454 * already linked up earlier.
2e64971a
DM
2455 */
2456 {
2457 U16 word;
2458 U32 state;
2459 U16 prev;
2460
2461 for (word=1; word <= trie->wordcount; word++) {
2462 prev = 0;
2463 if (trie->wordinfo[word].prev)
2464 continue;
2465 state = trie->wordinfo[word].accept;
2466 while (state) {
2467 state = prev_states[state];
2468 if (!state)
2469 break;
2470 prev = trie->states[state].wordnum;
2471 if (prev)
2472 break;
2473 }
2474 trie->wordinfo[word].prev = prev;
2475 }
2476 Safefree(prev_states);
2477 }
2478
2479
2480 /* and now dump out the compressed format */
2481 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2482
55eed653 2483 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2484#ifdef DEBUGGING
2485 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2486 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2487#else
03e70be4 2488 SvREFCNT_dec_NN(revcharmap);
07be1b83 2489#endif
786e8c11
YO
2490 return trie->jump
2491 ? MADE_JUMP_TRIE
2492 : trie->startstate>1
2493 ? MADE_EXACT_TRIE
2494 : MADE_TRIE;
2495}
2496
2497STATIC void
2498S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2499{
3b753521 2500/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2501
2502 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2503 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2504 ISBN 0-201-10088-6
2505
2506 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2507 suffix of the current state's 'word' that is also a proper prefix of another word in our
2508 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2509 the DFA not to have to restart after its tried and failed a word at a given point, it
2510 simply continues as though it had been matching the other word in the first place.
2511 Consider
2512 'abcdgu'=~/abcdefg|cdgu/
2513 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2514 fail, which would bring us to the state representing 'd' in the second word where we would
2515 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2516 */
2517 /* add a fail transition */
3251b653
NC
2518 const U32 trie_offset = ARG(source);
2519 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2520 U32 *q;
2521 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2522 const U32 numstates = trie->statecount;
786e8c11
YO
2523 const U32 ubound = trie->lasttrans + ucharcount;
2524 U32 q_read = 0;
2525 U32 q_write = 0;
2526 U32 charid;
2527 U32 base = trie->states[ 1 ].trans.base;
2528 U32 *fail;
2529 reg_ac_data *aho;
2530 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2531 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2532
2533 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2534#ifndef DEBUGGING
2535 PERL_UNUSED_ARG(depth);
2536#endif
2537
2538
2539 ARG_SET( stclass, data_slot );
c944940b 2540 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2541 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2542 aho->trie=trie_offset;
446bd890
NC
2543 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2544 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2545 Newxz( q, numstates, U32);
c944940b 2546 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2547 aho->refcount = 1;
2548 fail = aho->fail;
2549 /* initialize fail[0..1] to be 1 so that we always have
2550 a valid final fail state */
2551 fail[ 0 ] = fail[ 1 ] = 1;
2552
2553 for ( charid = 0; charid < ucharcount ; charid++ ) {
2554 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2555 if ( newstate ) {
2556 q[ q_write ] = newstate;
2557 /* set to point at the root */
2558 fail[ q[ q_write++ ] ]=1;
2559 }
2560 }
2561 while ( q_read < q_write) {
2562 const U32 cur = q[ q_read++ % numstates ];
2563 base = trie->states[ cur ].trans.base;
2564
2565 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2566 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2567 if (ch_state) {
2568 U32 fail_state = cur;
2569 U32 fail_base;
2570 do {
2571 fail_state = fail[ fail_state ];
2572 fail_base = aho->states[ fail_state ].trans.base;
2573 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2574
2575 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2576 fail[ ch_state ] = fail_state;
2577 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2578 {
2579 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2580 }
2581 q[ q_write++ % numstates] = ch_state;
2582 }
2583 }
2584 }
2585 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2586 when we fail in state 1, this allows us to use the
2587 charclass scan to find a valid start char. This is based on the principle
2588 that theres a good chance the string being searched contains lots of stuff
2589 that cant be a start char.
2590 */
2591 fail[ 0 ] = fail[ 1 ] = 0;
2592 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2593 PerlIO_printf(Perl_debug_log,
2594 "%*sStclass Failtable (%"UVuf" states): 0",
2595 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2596 );
786e8c11
YO
2597 for( q_read=1; q_read<numstates; q_read++ ) {
2598 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2599 }
2600 PerlIO_printf(Perl_debug_log, "\n");
2601 });
2602 Safefree(q);
2603 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2604}
2605
786e8c11 2606
a3621e74 2607/*
5d1c421c
JH
2608 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2609 * These need to be revisited when a newer toolchain becomes available.
2610 */
2611#if defined(__sparc64__) && defined(__GNUC__)
2612# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2613# undef SPARC64_GCC_WORKAROUND
2614# define SPARC64_GCC_WORKAROUND 1
2615# endif
2616#endif
2617
07be1b83 2618#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2619 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2620 SV * const mysv=sv_newmortal(); \
2621 regnode *Next = regnext(scan); \
2622 regprop(RExC_rx, mysv, scan); \
7f69552c 2623 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2624 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2625 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2626 }});
07be1b83 2627
1de06328 2628
bb914485 2629/* The below joins as many adjacent EXACTish nodes as possible into a single
0a982f06
KW
2630 * one. The regop may be changed if the node(s) contain certain sequences that
2631 * require special handling. The joining is only done if:
bb914485
KW
2632 * 1) there is room in the current conglomerated node to entirely contain the
2633 * next one.
2634 * 2) they are the exact same node type
2635 *
87b8b349 2636 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
2637 * these get optimized out
2638 *
0a982f06
KW
2639 * If a node is to match under /i (folded), the number of characters it matches
2640 * can be different than its character length if it contains a multi-character
2641 * fold. *min_subtract is set to the total delta of the input nodes.
bb914485 2642 *
a0c4c608
KW
2643 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2644 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2645 *
bb914485 2646 * This is as good a place as any to discuss the design of handling these
0a982f06
KW
2647 * multi-character fold sequences. It's been wrong in Perl for a very long
2648 * time. There are three code points in Unicode whose multi-character folds
2649 * were long ago discovered to mess things up. The previous designs for
2650 * dealing with these involved assigning a special node for them. This
2651 * approach doesn't work, as evidenced by this example:
a0c4c608 2652 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
0a982f06
KW
2653 * Both these fold to "sss", but if the pattern is parsed to create a node that
2654 * would match just the \xDF, it won't be able to handle the case where a
bb914485
KW
2655 * successful match would have to cross the node's boundary. The new approach
2656 * that hopefully generally solves the problem generates an EXACTFU_SS node
2657 * that is "sss".
2658 *
0a982f06
KW
2659 * It turns out that there are problems with all multi-character folds, and not
2660 * just these three. Now the code is general, for all such cases, but the
2661 * three still have some special handling. The approach taken is:
2662 * 1) This routine examines each EXACTFish node that could contain multi-
2663 * character fold sequences. It returns in *min_subtract how much to
9d071ca8 2664 * subtract from the the actual length of the string to get a real minimum
0a982f06
KW
2665 * match length; it is 0 if there are no multi-char folds. This delta is
2666 * used by the caller to adjust the min length of the match, and the delta
2667 * between min and max, so that the optimizer doesn't reject these
2668 * possibilities based on size constraints.
2669 * 2) Certain of these sequences require special handling by the trie code,
2670 * so, if found, this code changes the joined node type to special ops:
2671 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2672 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2673 * is used for an EXACTFU node that contains at least one "ss" sequence in
2674 * it. For non-UTF-8 patterns and strings, this is the only case where
2675 * there is a possible fold length change. That means that a regular
2676 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2677 * with length changes, and so can be processed faster. regexec.c takes
2678 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2679 * pre-folded by regcomp.c. This saves effort in regex matching.
87b8b349 2680 * However, the pre-folding isn't done for non-UTF8 patterns because the
0a982f06
KW
2681 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2682 * down by forcing the pattern into UTF8 unless necessary. Also what
2683 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2684 * possibilities for the non-UTF8 patterns are quite simple, except for
2685 * the sharp s. All the ones that don't involve a UTF-8 target string are
2686 * members of a fold-pair, and arrays are set up for all of them so that
2687 * the other member of the pair can be found quickly. Code elsewhere in
2688 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2689 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2690 * described in the next item.
1ca267a5
KW
2691 * 4) A problem remains for the sharp s in EXACTF and EXACTFA nodes when the
2692 * pattern isn't in UTF-8. (BTW, there cannot be an EXACTF node with a
2693 * UTF-8 pattern.) An assumption that the optimizer part of regexec.c
2694 * (probably unwittingly, in Perl_regexec_flags()) makes is that a
2695 * character in the pattern corresponds to at most a single character in
2696 * the target string. (And I do mean character, and not byte here, unlike
2697 * other parts of the documentation that have never been updated to
2698 * account for multibyte Unicode.) sharp s in EXACTF nodes can match the
2699 * two character string 'ss'; in EXACTFA nodes it can match
2700 * "\x{17F}\x{17F}". These violate the assumption, and they are the only
2701 * instances where it is violated. I'm reluctant to try to change the
2702 * assumption, as the code involved is impenetrable to me (khw), so
2703 * instead the code here punts. This routine examines (when the pattern
2704 * isn't UTF-8) EXACTF and EXACTFA nodes for the sharp s, and returns a
2705 * boolean indicating whether or not the node contains a sharp s. When it
2706 * is true, the caller sets a flag that later causes the optimizer in this
2707 * file to not set values for the floating and fixed string lengths, and
2708 * thus avoids the optimizer code in regexec.c that makes the invalid
2709 * assumption. Thus, there is no optimization based on string lengths for
2710 * non-UTF8-pattern EXACTF and EXACTFA nodes that contain the sharp s.
2711 * (The reason the assumption is wrong only in these two cases is that all
2712 * other non-UTF-8 folds are 1-1; and, for UTF-8 patterns, we pre-fold all
2713 * other folds to their expanded versions. We can't prefold sharp s to
2714 * 'ss' in EXACTF nodes because we don't know at compile time if it
2715 * actually matches 'ss' or not. It will match iff the target string is
2716 * in UTF-8, unlike the EXACTFU nodes, where it always matches; and
2717 * EXACTFA and EXACTFL where it never does. In an EXACTFA node in a UTF-8
2718 * pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the problem;
2719 * but in a non-UTF8 pattern, folding it to that above-Latin1 string would
2720 * require the pattern to be forced into UTF-8, the overhead of which we
2721 * want to avoid.)
bb914485 2722 */
1de06328 2723
9d071ca8 2724#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2725 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2726 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2727
be8e71aa 2728STATIC U32
9d071ca8 2729S_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
2730 /* Merge several consecutive EXACTish nodes into one. */
2731 regnode *n = regnext(scan);
2732 U32 stringok = 1;
2733 regnode *next = scan + NODE_SZ_STR(scan);
2734 U32 merged = 0;
2735 U32 stopnow = 0;
2736#ifdef DEBUGGING
2737 regnode *stop = scan;
72f13be8 2738 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2739#else
d47053eb
RGS
2740 PERL_UNUSED_ARG(depth);
2741#endif
7918f24d
NC
2742
2743 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2744#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2745 PERL_UNUSED_ARG(flags);
2746 PERL_UNUSED_ARG(val);
07be1b83 2747#endif
07be1b83 2748 DEBUG_PEEP("join",scan,depth);
bb914485 2749
3f410cf6
KW
2750 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2751 * EXACT ones that are mergeable to the current one. */
2752 while (n
2753 && (PL_regkind[OP(n)] == NOTHING
2754 || (stringok && OP(n) == OP(scan)))
07be1b83 2755 && NEXT_OFF(n)
3f410cf6
KW
2756 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2757 {
07be1b83
YO
2758
2759 if (OP(n) == TAIL || n > next)
2760 stringok = 0;
2761 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2762 DEBUG_PEEP("skip:",n,depth);
2763 NEXT_OFF(scan) += NEXT_OFF(n);
2764 next = n + NODE_STEP_REGNODE;
2765#ifdef DEBUGGING
2766 if (stringok)
2767 stop = n;
2768#endif
2769 n = regnext(n);
2770 }
2771 else if (stringok) {
786e8c11 2772 const unsigned int oldl = STR_LEN(scan);
07be1b83 2773 regnode * const nnext = regnext(n);
b2230d39 2774
87b8b349
KW
2775 /* XXX I (khw) kind of doubt that this works on platforms where
2776 * U8_MAX is above 255 because of lots of other assumptions */
79a81a6e 2777 /* Don't join if the sum can't fit into a single node */
b2230d39
KW
2778 if (oldl + STR_LEN(n) > U8_MAX)
2779 break;
07be1b83
YO
2780
2781 DEBUG_PEEP("merg",n,depth);
07be1b83 2782 merged++;
b2230d39 2783
07be1b83
YO
2784 NEXT_OFF(scan) += NEXT_OFF(n);
2785 STR_LEN(scan) += STR_LEN(n);
2786 next = n + NODE_SZ_STR(n);
2787 /* Now we can overwrite *n : */
2788 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2789#ifdef DEBUGGING
2790 stop = next - 1;
2791#endif
2792 n = nnext;
2793 if (stopnow) break;
2794 }
2795
d47053eb
RGS
2796#ifdef EXPERIMENTAL_INPLACESCAN
2797 if (flags && !NEXT_OFF(n)) {
2798 DEBUG_PEEP("atch", val, depth);
2799 if (reg_off_by_arg[OP(n)]) {
2800 ARG_SET(n, val - n);
2801 }
2802 else {
2803 NEXT_OFF(n) = val - n;
2804 }
2805 stopnow = 1;
2806 }
07be1b83
YO
2807#endif
2808 }
2c2b7f86 2809
9d071ca8 2810 *min_subtract = 0;
f758bddf 2811 *has_exactf_sharp_s = FALSE;
f646642f 2812
3f410cf6
KW
2813 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2814 * can now analyze for sequences of problematic code points. (Prior to
2815 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2816 * hence missed). The sequences only happen in folding, hence for any
2817 * non-EXACT EXACTish node */
86d6fcad 2818 if (OP(scan) != EXACT) {
0a982f06
KW
2819 const U8 * const s0 = (U8*) STRING(scan);
2820 const U8 * s = s0;
2821 const U8 * const s_end = s0 + STR_LEN(scan);
f758bddf
KW
2822
2823 /* One pass is made over the node's string looking for all the
2824 * possibilities. to avoid some tests in the loop, there are two main
2825 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2826 * non-UTF-8 */
2827 if (UTF) {
86d6fcad 2828
0a982f06
KW
2829 /* Examine the string for a multi-character fold sequence. UTF-8
2830 * patterns have all characters pre-folded by the time this code is
2831 * executed */
2832 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2833 length sequence we are looking for is 2 */
86d6fcad 2834 {
0a982f06
KW
2835 int count = 0;
2836 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2837 if (! len) { /* Not a multi-char fold: get next char */
2838 s += UTF8SKIP(s);
2839 continue;
2840 }
bb914485 2841
0a982f06
KW
2842 /* Nodes with 'ss' require special handling, except for EXACTFL
2843 * and EXACTFA for which there is no multi-char fold to this */
2844 if (len == 2 && *s == 's' && *(s+1) == 's'
2845 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2846 {
2847 count = 2;
2848 OP(scan) = EXACTFU_SS;
2849 s += 2;
2850 }
accd961f
KW
2851 else if (len == 6 /* len is the same in both ASCII and EBCDIC
2852 for these */
0a982f06
KW
2853 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2854 COMBINING_DIAERESIS_UTF8
2855 COMBINING_ACUTE_ACCENT_UTF8,
2856 6)
2857 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2858 COMBINING_DIAERESIS_UTF8
2859 COMBINING_ACUTE_ACCENT_UTF8,
2860 6)))
2861 {
2862 count = 3;
2863
2864 /* These two folds require special handling by trie's, so
2865 * change the node type to indicate this. If EXACTFA and
2866 * EXACTFL were ever to be handled by trie's, this would
2867 * have to be changed. If this node has already been
2868 * changed to EXACTFU_SS in this loop, leave it as is. (I
2869 * (khw) think it doesn't matter in regexec.c for UTF
2870 * patterns, but no need to change it */
2871 if (OP(scan) == EXACTFU) {
2872 OP(scan) = EXACTFU_TRICKYFOLD;
2873 }
2874 s += 6;
2875 }
2876 else { /* Here is a generic multi-char fold. */
2877 const U8* multi_end = s + len;
2878
2879 /* Count how many characters in it. In the case of /l and
2880 * /aa, no folds which contain ASCII code points are
2881 * allowed, so check for those, and skip if found. (In
2882 * EXACTFL, no folds are allowed to any Latin1 code point,
2883 * not just ASCII. But there aren't any of these
2884 * currently, nor ever likely, so don't take the time to
2885 * test for them. The code that generates the
2886 * is_MULTI_foo() macros croaks should one actually get put
2887 * into Unicode .) */
2888 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2889 count = utf8_length(s, multi_end);
2890 s = multi_end;
2891 }
2892 else {
2893 while (s < multi_end) {
2894 if (isASCII(*s)) {
2895 s++;
2896 goto next_iteration;
2897 }
2898 else {
2899 s += UTF8SKIP(s);
2900 }
2901 count++;
2902 }
2903 }
2904 }
f758bddf 2905
0a982f06
KW
2906 /* The delta is how long the sequence is minus 1 (1 is how long
2907 * the character that folds to the sequence is) */
2908 *min_subtract += count - 1;
2909 next_iteration: ;
bb914485
KW
2910 }
2911 }
1ca267a5
KW
2912 else if (OP(scan) == EXACTFA) {
2913
2914 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
2915 * fold to the ASCII range (and there are no existing ones in the
2916 * upper latin1 range). But, as outlined in the comments preceding
2917 * this function, we need to flag any occurrences of the sharp s */
2918 while (s < s_end) {
2919 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
2920 *has_exactf_sharp_s = TRUE;
2921 break;
2922 }
2923 s++;
2924 continue;
2925 }
2926 }
2927 else if (OP(scan) != EXACTFL) {
2928
2929 /* Non-UTF-8 pattern, not EXACTFA nor EXACTFL node. Look for the
2930 * multi-char folds that are all Latin1. (This code knows that
2931 * there are no current multi-char folds possible with EXACTFL,
2932 * relying on fold_grind.t to catch any errors if the very unlikely
2933 * event happens that some get added in future Unicode versions.)
2934 * As explained in the comments preceding this function, we look
2935 * also for the sharp s in EXACTF nodes; it can be in the final
0a982f06
KW
2936 * position. Otherwise we can stop looking 1 byte earlier because
2937 * have to find at least two characters for a multi-fold */
f758bddf
KW
2938 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2939
0a982f06
KW
2940 /* The below is perhaps overboard, but this allows us to save a
2941 * test each time through the loop at the expense of a mask. This
2942 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2943 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2944 * are 64. This uses an exclusive 'or' to find that bit and then
2945 * inverts it to form a mask, with just a single 0, in the bit
2946 * position where 'S' and 's' differ. */
2947 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2948 const U8 s_masked = 's' & S_or_s_mask;
2949
2950 while (s < upper) {
40b1ba4f 2951 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
0a982f06
KW
2952 if (! len) { /* Not a multi-char fold. */
2953 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2954 {
2955 *has_exactf_sharp_s = TRUE;
2956 }
2957 s++;
2958 continue;
2959 }
2960
2961 if (len == 2
2962 && ((*s & S_or_s_mask) == s_masked)
2963 && ((*(s+1) & S_or_s_mask) == s_masked))
2964 {
2965
2966 /* EXACTF nodes need to know that the minimum length
2967 * changed so that a sharp s in the string can match this
2968 * ss in the pattern, but they remain EXACTF nodes, as they
2969 * won't match this unless the target string is is UTF-8,
2970 * which we don't know until runtime */
2971 if (OP(scan) != EXACTF) {
2972 OP(scan) = EXACTFU_SS;
2973 }
86d6fcad 2974 }
0a982f06
KW
2975
2976 *min_subtract += len - 1;
2977 s += len;
86d6fcad
KW
2978 }
2979 }
07be1b83 2980 }
3f410cf6 2981
07be1b83 2982#ifdef DEBUGGING
bb789b09
DM
2983 /* Allow dumping but overwriting the collection of skipped
2984 * ops and/or strings with fake optimized ops */
07be1b83
YO
2985 n = scan + NODE_SZ_STR(scan);
2986 while (n <= stop) {
bb789b09
DM
2987 OP(n) = OPTIMIZED;
2988 FLAGS(n) = 0;
2989 NEXT_OFF(n) = 0;
07be1b83
YO
2990 n++;
2991 }
2992#endif
2993 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2994 return stopnow;
2995}
2996
486ec47a 2997/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2998 Finds fixed substrings. */
2999
a0288114 3000/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
3001 to the position after last scanned or to NULL. */
3002
40d049e4
YO
3003#define INIT_AND_WITHP \
3004 assert(!and_withp); \
3005 Newx(and_withp,1,struct regnode_charclass_class); \
3006 SAVEFREEPV(and_withp)
07be1b83 3007
b515a41d 3008/* this is a chain of data about sub patterns we are processing that
486ec47a 3009 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
3010 we can simulate recursion without losing state. */
3011struct scan_frame;
3012typedef struct scan_frame {
3013 regnode *last; /* last node to process in this frame */
3014 regnode *next; /* next node to process when last is reached */
3015 struct scan_frame *prev; /*previous frame*/
3016 I32 stop; /* what stopparen do we use */
3017} scan_frame;
3018
304ee84b
YO
3019
3020#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
3021
76e3520e 3022STATIC I32
40d049e4 3023S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 3024 I32 *minlenp, I32 *deltap,
40d049e4
YO
3025 regnode *last,
3026 scan_data_t *data,
3027 I32 stopparen,
3028 U8* recursed,
3029 struct regnode_charclass_class *and_withp,
3030 U32 flags, U32 depth)
c277df42
IZ
3031 /* scanp: Start here (read-write). */
3032 /* deltap: Write maxlen-minlen here. */
3033 /* last: Stop before this one. */
40d049e4
YO
3034 /* data: string data about the pattern */
3035 /* stopparen: treat close N as END */
3036 /* recursed: which subroutines have we recursed into */
3037 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 3038{
97aff369 3039 dVAR;
2d608413
KW
3040 I32 min = 0; /* There must be at least this number of characters to match */
3041 I32 pars = 0, code;
c277df42
IZ
3042 regnode *scan = *scanp, *next;
3043 I32 delta = 0;
3044 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 3045 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
3046 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3047 scan_data_t data_fake;
a3621e74 3048 SV *re_trie_maxbuff = NULL;
786e8c11 3049 regnode *first_non_open = scan;
e2e6a0f1 3050 I32 stopmin = I32_MAX;
8aa23a47 3051 scan_frame *frame = NULL;
a3621e74 3052 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 3053
7918f24d
NC
3054 PERL_ARGS_ASSERT_STUDY_CHUNK;
3055
13a24bad 3056#ifdef DEBUGGING
40d049e4 3057 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 3058#endif
40d049e4 3059
786e8c11 3060 if ( depth == 0 ) {
40d049e4 3061 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
3062 first_non_open=regnext(first_non_open);
3063 }
3064
b81d288d 3065
8aa23a47
YO
3066 fake_study_recurse:
3067 while ( scan && OP(scan) != END && scan < last ){
2d608413
KW
3068 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3069 node length to get a real minimum (because
3070 the folded version may be shorter) */
f758bddf 3071 bool has_exactf_sharp_s = FALSE;
8aa23a47 3072 /* Peephole optimizer: */
304ee84b 3073 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 3074 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
3075
3076 /* Its not clear to khw or hv why this is done here, and not in the
3077 * clauses that deal with EXACT nodes. khw's guess is that it's
3078 * because of a previous design */
9d071ca8 3079 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
3080
3081 /* Follow the next-chain of the current node and optimize
3082 away all the NOTHINGs from it. */
3083 if (OP(scan) != CURLYX) {
3084 const int max = (reg_off_by_arg[OP(scan)]
3085 ? I32_MAX
3086 /* I32 may be smaller than U16 on CRAYs! */
3087 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3088 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3089 int noff;
3090 regnode *n = scan;
686b73d4 3091
8aa23a47
YO
3092 /* Skip NOTHING and LONGJMP. */
3093 while ((n = regnext(n))
3094 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3095 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3096 && off + noff < max)
3097 off += noff;
3098 if (reg_off_by_arg[OP(scan)])
3099 ARG(scan) = off;
3100 else
3101 NEXT_OFF(scan) = off;
3102 }
a3621e74 3103
c277df42 3104
8aa23a47
YO
3105
3106 /* The principal pseudo-switch. Cannot be a switch, since we
3107 look into several different things. */
3108 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3109 || OP(scan) == IFTHEN) {
3110 next = regnext(scan);
3111 code = OP(scan);
3112 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3113
8aa23a47
YO
3114 if (OP(next) == code || code == IFTHEN) {
3115 /* NOTE - There is similar code to this block below for handling
3116 TRIE nodes on a re-study. If you change stuff here check there
3117 too. */
3118 I32 max1 = 0, min1 = I32_MAX, num = 0;
3119 struct regnode_charclass_class accum;
3120 regnode * const startbranch=scan;
686b73d4 3121
8aa23a47 3122 if (flags & SCF_DO_SUBSTR)
304ee84b 3123 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3124 if (flags & SCF_DO_STCLASS)
e755fd73 3125 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3126
3127 while (OP(scan) == code) {
3128 I32 deltanext, minnext, f = 0, fake;
3129 struct regnode_charclass_class this_class;
3130
3131 num++;
3132 data_fake.flags = 0;
3133 if (data) {
3134 data_fake.whilem_c = data->whilem_c;
3135 data_fake.last_closep = data->last_closep;
3136 }
3137 else
3138 data_fake.last_closep = &fake;
58e23c8d
YO
3139
3140 data_fake.pos_delta = delta;
8aa23a47
YO
3141 next = regnext(scan);
3142 scan = NEXTOPER(scan);
3143 if (code != BRANCH)
c277df42 3144 scan = NEXTOPER(scan);
8aa23a47 3145 if (flags & SCF_DO_STCLASS) {
e755fd73 3146 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3147 data_fake.start_class = &this_class;
3148 f = SCF_DO_STCLASS_AND;
58e23c8d 3149 }
8aa23a47
YO
3150 if (flags & SCF_WHILEM_VISITED_POS)
3151 f |= SCF_WHILEM_VISITED_POS;
3152
3153 /* we suppose the run is continuous, last=next...*/
3154 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3155 next, &data_fake,
3156 stopparen, recursed, NULL, f,depth+1);
3157 if (min1 > minnext)
3158 min1 = minnext;
9b139d09 3159 if (deltanext == I32_MAX) {
8aa23a47 3160 is_inf = is_inf_internal = 1;
9b139d09
GG
3161 max1 = I32_MAX;
3162 } else if (max1 < minnext + deltanext)
3163 max1 = minnext + deltanext;
8aa23a47
YO
3164 scan = next;
3165 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3166 pars++;
3167 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3168 if ( stopmin > minnext)
3169 stopmin = min + min1;
3170 flags &= ~SCF_DO_SUBSTR;
3171 if (data)
3172 data->flags |= SCF_SEEN_ACCEPT;
3173 }
3174 if (data) {
3175 if (data_fake.flags & SF_HAS_EVAL)
3176 data->flags |= SF_HAS_EVAL;
3177 data->whilem_c = data_fake.whilem_c;
3dab1dad 3178 }
8aa23a47 3179 if (flags & SCF_DO_STCLASS)
3fffb88a 3180 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3181 }
3182 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3183 min1 = 0;
3184 if (flags & SCF_DO_SUBSTR) {
3185 data->pos_min += min1;
9b139d09
GG
3186 if (data->pos_delta >= I32_MAX - (max1 - min1))
3187 data->pos_delta = I32_MAX;
3188 else
3189 data->pos_delta += max1 - min1;
8aa23a47
YO
3190 if (max1 != min1 || is_inf)
3191 data->longest = &(data->longest_float);
3192 }
3193 min += min1;
9b139d09
GG
3194 if (delta == I32_MAX || I32_MAX - delta - (max1 - min1) < 0)
3195 delta = I32_MAX;
3196 else
3197 delta += max1 - min1;
8aa23a47 3198 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3199 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3200 if (min1) {
3201 cl_and(data->start_class, and_withp);
3202 flags &= ~SCF_DO_STCLASS;
653099ff 3203 }
8aa23a47
YO
3204 }
3205 else if (flags & SCF_DO_STCLASS_AND) {
3206 if (min1) {
3207 cl_and(data->start_class, &accum);
3208 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3209 }
8aa23a47
YO
3210 else {
3211 /* Switch to OR mode: cache the old value of
3212 * data->start_class */
3213 INIT_AND_WITHP;
3214 StructCopy(data->start_class, and_withp,
3215 struct regnode_charclass_class);
3216 flags &= ~SCF_DO_STCLASS_AND;
3217 StructCopy(&accum, data->start_class,
3218 struct regnode_charclass_class);
3219 flags |= SCF_DO_STCLASS_OR;
899d20b9 3220 SET_SSC_EOS(data->start_class);
de0c8cb8 3221 }
8aa23a47 3222 }
a3621e74 3223
8aa23a47
YO
3224 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3225 /* demq.
a3621e74 3226
8aa23a47
YO
3227 Assuming this was/is a branch we are dealing with: 'scan' now
3228 points at the item that follows the branch sequence, whatever
3229 it is. We now start at the beginning of the sequence and look
3230 for subsequences of
a3621e74 3231
8aa23a47
YO
3232 BRANCH->EXACT=>x1
3233 BRANCH->EXACT=>x2
3234 tail
a3621e74 3235
8aa23a47 3236 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3237
486ec47a 3238 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3239 element into a trie and then add the subsequent branch exact
3240 strings to the trie.
a3621e74 3241
8aa23a47 3242 We have two cases
a3621e74 3243
3b753521 3244 1. patterns where the whole set of branches can be converted.
a3621e74 3245
8aa23a47 3246 2. patterns where only a subset can be converted.
a3621e74 3247
8aa23a47
YO
3248 In case 1 we can replace the whole set with a single regop
3249 for the trie. In case 2 we need to keep the start and end
3b753521 3250 branches so
a3621e74 3251
8aa23a47
YO
3252 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3253 becomes BRANCH TRIE; BRANCH X;
786e8c11 3254
8aa23a47
YO
3255 There is an additional case, that being where there is a
3256 common prefix, which gets split out into an EXACT like node
3257 preceding the TRIE node.
a3621e74 3258
8aa23a47
YO
3259 If x(1..n)==tail then we can do a simple trie, if not we make
3260 a "jump" trie, such that when we match the appropriate word
486ec47a 3261 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3262 a nested if into a case structure of sorts.
b515a41d 3263
8aa23a47 3264 */
686b73d4 3265
8aa23a47
YO
3266 int made=0;
3267 if (!re_trie_maxbuff) {
3268 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3269 if (!SvIOK(re_trie_maxbuff))
3270 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3271 }
3272 if ( SvIV(re_trie_maxbuff)>=0 ) {
3273 regnode *cur;
3274 regnode *first = (regnode *)NULL;
3275 regnode *last = (regnode *)NULL;
3276 regnode *tail = scan;
fab2782b 3277 U8 trietype = 0;
8aa23a47 3278 U32 count=0;
a3621e74
YO
3279
3280#ifdef DEBUGGING
8aa23a47 3281 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3282#endif
8aa23a47
YO
3283 /* var tail is used because there may be a TAIL
3284 regop in the way. Ie, the exacts will point to the
3285 thing following the TAIL, but the last branch will
3286 point at the TAIL. So we advance tail. If we
3287 have nested (?:) we may have to move through several
3288 tails.
3289 */
3290
3291 while ( OP( tail ) == TAIL ) {
3292 /* this is the TAIL generated by (?:) */
3293 tail = regnext( tail );
3294 }
a3621e74 3295
8aa23a47 3296
df826430 3297 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3298 regprop(RExC_rx, mysv, tail );
3299 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3300 (int)depth * 2 + 2, "",
3301 "Looking for TRIE'able sequences. Tail node is: ",
3302 SvPV_nolen_const( mysv )
3303 );
3304 });
3305
3306 /*
3307
fab2782b
YO
3308 Step through the branches
3309 cur represents each branch,
3310 noper is the first thing to be matched as part of that branch
3311 noper_next is the regnext() of that node.
3312
3313 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3314 via a "jump trie" but we also support building with NOJUMPTRIE,
3315 which restricts the trie logic to structures like /FOO|BAR/.
3316
3317 If noper is a trieable nodetype then the branch is a possible optimization
3318 target. If we are building under NOJUMPTRIE then we require that noper_next
3319 is the same as scan (our current position in the regex program).
3320
3321 Once we have two or more consecutive such branches we can create a
3322 trie of the EXACT's contents and stitch it in place into the program.
3323
3324 If the sequence represents all of the branches in the alternation we
3325 replace the entire thing with a single TRIE node.
3326
3327 Otherwise when it is a subsequence we need to stitch it in place and
3328 replace only the relevant branches. This means the first branch has
3329 to remain as it is used by the alternation logic, and its next pointer,
3330 and needs to be repointed at the item on the branch chain following
3331 the last branch we have optimized away.
3332
3333 This could be either a BRANCH, in which case the subsequence is internal,
3334 or it could be the item following the branch sequence in which case the
3335 subsequence is at the end (which does not necessarily mean the first node
3336 is the start of the alternation).
3337
3338 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3339
3340 optype | trietype
3341 ----------------+-----------
3342 NOTHING | NOTHING
3343 EXACT | EXACT
3344 EXACTFU | EXACTFU
3345 EXACTFU_SS | EXACTFU
3346 EXACTFU_TRICKYFOLD | EXACTFU
3347 EXACTFA | 0
3348
8aa23a47
YO
3349
3350 */
fab2782b
YO
3351#define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3352 ( EXACT == (X) ) ? EXACT : \
3353 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3354 0 )
8aa23a47
YO
3355
3356 /* dont use tail as the end marker for this traverse */
3357 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3358 regnode * const noper = NEXTOPER( cur );
fab2782b
YO
3359 U8 noper_type = OP( noper );
3360 U8 noper_trietype = TRIE_TYPE( noper_type );
b515a41d 3361#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3362 regnode * const noper_next = regnext( noper );
df826430
YO
3363 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3364 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
b515a41d
YO
3365#endif
3366
df826430 3367 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3368 regprop(RExC_rx, mysv, cur);
3369 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3370 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3371
3372 regprop(RExC_rx, mysv, noper);
3373 PerlIO_printf( Perl_debug_log, " -> %s",
3374 SvPV_nolen_const(mysv));
3375
3376 if ( noper_next ) {
3377 regprop(RExC_rx, mysv, noper_next );
3378 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3379 SvPV_nolen_const(mysv));
3380 }
df826430
YO
3381 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3382 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3383 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3384 );
8aa23a47 3385 });
fab2782b
YO
3386
3387 /* Is noper a trieable nodetype that can be merged with the
3388 * current trie (if there is one)? */
3389 if ( noper_trietype
3390 &&
3391 (
df826430
YO
3392 ( noper_trietype == NOTHING)
3393 || ( trietype == NOTHING )
a40630bf 3394 || ( trietype == noper_trietype )
fab2782b 3395 )
786e8c11 3396#ifdef NOJUMPTRIE
8aa23a47 3397 && noper_next == tail
786e8c11 3398#endif
8aa23a47
YO
3399 && count < U16_MAX)
3400 {
fab2782b
YO
3401 /* Handle mergable triable node
3402 * Either we are the first node in a new trieable sequence,
3403 * in which case we do some bookkeeping, otherwise we update
3404 * the end pointer. */
fab2782b 3405 if ( !first ) {
3b6759a6 3406 first = cur;
df826430
YO
3407 if ( noper_trietype == NOTHING ) {
3408#if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3409 regnode * const noper_next = regnext( noper );
3b6759a6 3410 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
df826430
YO
3411 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3412#endif
3413
190c1910 3414 if ( noper_next_trietype ) {
df826430 3415 trietype = noper_next_trietype;
190c1910
YO
3416 } else if (noper_next_type) {
3417 /* a NOTHING regop is 1 regop wide. We need at least two
3418 * for a trie so we can't merge this in */
3419 first = NULL;
3420 }
3421 } else {
3422 trietype = noper_trietype;
3b6759a6 3423 }
8aa23a47 3424 } else {
fab2782b
YO
3425 if ( trietype == NOTHING )
3426 trietype = noper_trietype;
8aa23a47
YO
3427 last = cur;
3428 }
df826430
YO
3429 if (first)
3430 count++;
fab2782b
YO
3431 } /* end handle mergable triable node */
3432 else {
3433 /* handle unmergable node -
3434 * noper may either be a triable node which can not be tried
3435 * together with the current trie, or a non triable node */
729aaeb5
YO
3436 if ( last ) {
3437 /* If last is set and trietype is not NOTHING then we have found
3438 * at least two triable branch sequences in a row of a similar
3439 * trietype so we can turn them into a trie. If/when we
3440 * allow NOTHING to start a trie sequence this condition will be
3441 * required, and it isn't expensive so we leave it in for now. */
e6351b37 3442 if ( trietype && trietype != NOTHING )
729aaeb5
YO
3443 make_trie( pRExC_state,
3444 startbranch, first, cur, tail, count,
3445 trietype, depth+1 );
fab2782b 3446 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
8aa23a47 3447 }
fab2782b 3448 if ( noper_trietype
786e8c11 3449#ifdef NOJUMPTRIE
8aa23a47 3450 && noper_next == tail
786e8c11 3451#endif
8aa23a47 3452 ){
fab2782b 3453 /* noper is triable, so we can start a new trie sequence */
8aa23a47
YO
3454 count = 1;
3455 first = cur;
fab2782b
YO
3456 trietype = noper_trietype;
3457 } else if (first) {
3458 /* if we already saw a first but the current node is not triable then we have
3459 * to reset the first information. */
8aa23a47
YO
3460 count = 0;
3461 first = NULL;
fab2782b 3462 trietype = 0;
8aa23a47 3463 }
fab2782b
YO
3464 } /* end handle unmergable node */
3465 } /* loop over branches */
df826430 3466 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3467 regprop(RExC_rx, mysv, cur);
3468 PerlIO_printf( Perl_debug_log,
3469 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3470 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3471
3472 });
e6351b37 3473 if ( last && trietype ) {
3b6759a6
YO
3474 if ( trietype != NOTHING ) {
3475 /* the last branch of the sequence was part of a trie,
3476 * so we have to construct it here outside of the loop
3477 */
3478 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
686b73d4 3479#ifdef TRIE_STUDY_OPT
3b6759a6
YO
3480 if ( ((made == MADE_EXACT_TRIE &&
3481 startbranch == first)
3482 || ( first_non_open == first )) &&
3483 depth==0 ) {
3484 flags |= SCF_TRIE_RESTUDY;
3485 if ( startbranch == first
3486 && scan == tail )
3487 {
3488 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3489 }
8aa23a47 3490 }
3dab1dad 3491#endif
3b6759a6
YO
3492 } else {
3493 /* at this point we know whatever we have is a NOTHING sequence/branch
3494 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3495 */
3496 if ( startbranch == first ) {
3497 regnode *opt;
3498 /* the entire thing is a NOTHING sequence, something like this:
3499 * (?:|) So we can turn it into a plain NOTHING op. */
3500 DEBUG_TRIE_COMPILE_r({
3501 regprop(RExC_rx, mysv, cur);
3502 PerlIO_printf( Perl_debug_log,
3503 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3504 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3505
3506 });
3507 OP(startbranch)= NOTHING;
3508 NEXT_OFF(startbranch)= tail - startbranch;
3509 for ( opt= startbranch + 1; opt < tail ; opt++ )
3510 OP(opt)= OPTIMIZED;
3511 }
3512 }
fab2782b
YO
3513 } /* end if ( last) */
3514 } /* TRIE_MAXBUF is non zero */
8aa23a47
YO
3515
3516 } /* do trie */
3517
653099ff 3518 }
8aa23a47
YO
3519 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3520 scan = NEXTOPER(NEXTOPER(scan));
3521 } else /* single branch is optimized. */
3522 scan = NEXTOPER(scan);
3523 continue;
3524 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3525 scan_frame *newframe = NULL;
3526 I32 paren;
3527 regnode *start;
3528 regnode *end;
3529
3530 if (OP(scan) != SUSPEND) {
3531 /* set the pointer */
3532 if (OP(scan) == GOSUB) {
3533 paren = ARG(scan);
3534 RExC_recurse[ARG2L(scan)] = scan;
3535 start = RExC_open_parens[paren-1];
3536 end = RExC_close_parens[paren-1];
3537 } else {
3538 paren = 0;
f8fc2ecf 3539 start = RExC_rxi->program + 1;
8aa23a47
YO
3540 end = RExC_opend;
3541 }
3542 if (!recursed) {
3543 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3544 SAVEFREEPV(recursed);
3545 }
3546 if (!PAREN_TEST(recursed,paren+1)) {
3547 PAREN_SET(recursed,paren+1);
3548 Newx(newframe,1,scan_frame);
3549 } else {
3550 if (flags & SCF_DO_SUBSTR) {
304ee84b 3551 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3552 data->longest = &(data->longest_float);
3553 }
3554 is_inf = is_inf_internal = 1;
3555 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3556 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3557 flags &= ~SCF_DO_STCLASS;
3558 }
3559 } else {
3560 Newx(newframe,1,scan_frame);
3561 paren = stopparen;
3562 start = scan+2;
3563 end = regnext(scan);
3564 }
3565 if (newframe) {
3566 assert(start);
3567 assert(end);
3568 SAVEFREEPV(newframe);
3569 newframe->next = regnext(scan);
3570 newframe->last = last;
3571 newframe->stop = stopparen;
3572