This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
PATCH: [perl #116899]: jump on uninitialised value
[perl5.git] / regcomp.c
CommitLineData
a0d0e21e
LW
1/* regcomp.c
2 */
3
4/*
4ac71550
TC
5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
6 *
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
a0d0e21e
LW
8 */
9
61296642
DM
10/* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
166f8a29 12 * a regular expression.
e4a054ea
DM
13 *
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
166f8a29
DM
18 */
19
a687059c
LW
20/* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
22 */
23
24/* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
27 */
28
e50aee73 29/* The names of the functions have been changed from regcomp and
3b753521 30 * regexec to pregcomp and pregexec in order to avoid conflicts
e50aee73
AD
31 * with the POSIX routines of the same names.
32*/
33
b9d5759e 34#ifdef PERL_EXT_RE_BUILD
54df2634 35#include "re_top.h"
b81d288d 36#endif
56953603 37
a687059c 38/*
e50aee73 39 * pregcomp and pregexec -- regsub and regerror are not used in perl
a687059c
LW
40 *
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
43 *
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
47 *
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
50 * from defects in it.
51 *
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
54 *
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
57 *
58 *
59 **** Alterations to Henry's code are...
60 ****
4bb101f2 61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
1129b882
NC
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
a687059c 64 ****
9ef589d8
LW
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
67
a687059c
LW
68 *
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
72 */
73#include "EXTERN.h"
864dbfa3 74#define PERL_IN_REGCOMP_C
a687059c 75#include "perl.h"
d06ea78c 76
acfe0abc 77#ifndef PERL_IN_XSUB_RE
d06ea78c
GS
78# include "INTERN.h"
79#endif
c277df42
IZ
80
81#define REG_COMP_C
54df2634
NC
82#ifdef PERL_IN_XSUB_RE
83# include "re_comp.h"
40b6423c 84extern const struct regexp_engine my_reg_engine;
54df2634
NC
85#else
86# include "regcomp.h"
87#endif
a687059c 88
04e98a4d 89#include "dquote_static.c"
26faadbd 90#include "charclass_invlists.h"
81e983c1 91#include "inline_invlist.c"
1b0f46bf 92#include "unicode_constants.h"
04e98a4d 93
94dc5c2d 94#define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
26faadbd 95#define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
2c61f163 96#define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
94dc5c2d 97
d4cce5f1 98#ifdef op
11343788 99#undef op
d4cce5f1 100#endif /* op */
11343788 101
fe14fcc3 102#ifdef MSDOS
7e4e8c89 103# if defined(BUGGY_MSC6)
fe14fcc3 104 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
7e4e8c89 105# pragma optimize("a",off)
fe14fcc3 106 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
7e4e8c89
NC
107# pragma optimize("w",on )
108# endif /* BUGGY_MSC6 */
fe14fcc3
LW
109#endif /* MSDOS */
110
a687059c
LW
111#ifndef STATIC
112#define STATIC static
113#endif
114
b1603ef8 115
830247a4 116typedef struct RExC_state_t {
514a91f1
DM
117 U32 flags; /* RXf_* are we folding, multilining? */
118 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 119 char *precomp; /* uncompiled string. */
288b8c02 120 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf
YO
121 regexp *rx; /* perl core regexp structure */
122 regexp_internal *rxi; /* internal data for regexp object pprivate field */
fac92740 123 char *start; /* Start of input for compile */
830247a4
IZ
124 char *end; /* End of input for compile */
125 char *parse; /* Input-scan pointer. */
126 I32 whilem_seen; /* number of WHILEM in this expr */
fac92740 127 regnode *emit_start; /* Start of emitted-code area */
3b57cd43 128 regnode *emit_bound; /* First regnode outside of the allocated space */
ffc61ed2 129 regnode *emit; /* Code-emit pointer; &regdummy = don't = compiling */
830247a4
IZ
130 I32 naughty; /* How bad is this pattern? */
131 I32 sawback; /* Did we see \1, ...? */
132 U32 seen;
133 I32 size; /* Code size. */
c74340f9
YO
134 I32 npar; /* Capture buffer count, (OPEN). */
135 I32 cpar; /* Capture buffer count, (CLOSE). */
e2e6a0f1 136 I32 nestroot; /* root parens we are in - used by accept */
830247a4
IZ
137 I32 extralen;
138 I32 seen_zerolen;
40d049e4
YO
139 regnode **open_parens; /* pointers to open parens */
140 regnode **close_parens; /* pointers to close parens */
141 regnode *opend; /* END node in program */
02daf0ab
YO
142 I32 utf8; /* whether the pattern is utf8 or not */
143 I32 orig_utf8; /* whether the pattern was originally in utf8 */
144 /* XXX use this for future optimisation of case
145 * where pattern must be upgraded to utf8. */
e40e74fe
KW
146 I32 uni_semantics; /* If a d charset modifier should use unicode
147 rules, even if the pattern is not in
148 utf8 */
81714fb9 149 HV *paren_names; /* Paren names */
1f1031fe 150
40d049e4
YO
151 regnode **recurse; /* Recurse regops */
152 I32 recurse_count; /* Number of recurse regops */
b57e4118 153 I32 in_lookbehind;
4624b182 154 I32 contains_locale;
bb3f3ed2 155 I32 override_recoding;
9d53c457 156 I32 in_multi_char_class;
3d2bd50a 157 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 158 within pattern */
b1603ef8
DM
159 int num_code_blocks; /* size of code_blocks[] */
160 int code_index; /* next code_blocks[] slot */
830247a4
IZ
161#if ADD_TO_REGEXEC
162 char *starttry; /* -Dr: where regtry was called. */
163#define RExC_starttry (pRExC_state->starttry)
164#endif
d24ca0c5 165 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 166#ifdef DEBUGGING
be8e71aa 167 const char *lastparse;
3dab1dad 168 I32 lastnum;
1f1031fe 169 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
170#define RExC_lastparse (pRExC_state->lastparse)
171#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 172#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 173#endif
830247a4
IZ
174} RExC_state_t;
175
e2509266 176#define RExC_flags (pRExC_state->flags)
514a91f1 177#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 178#define RExC_precomp (pRExC_state->precomp)
288b8c02 179#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 180#define RExC_rx (pRExC_state->rx)
f8fc2ecf 181#define RExC_rxi (pRExC_state->rxi)
fac92740 182#define RExC_start (pRExC_state->start)
830247a4
IZ
183#define RExC_end (pRExC_state->end)
184#define RExC_parse (pRExC_state->parse)
185#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237
YO
186#ifdef RE_TRACK_PATTERN_OFFSETS
187#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
188#endif
830247a4 189#define RExC_emit (pRExC_state->emit)
fac92740 190#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 191#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
192#define RExC_naughty (pRExC_state->naughty)
193#define RExC_sawback (pRExC_state->sawback)
194#define RExC_seen (pRExC_state->seen)
195#define RExC_size (pRExC_state->size)
196#define RExC_npar (pRExC_state->npar)
e2e6a0f1 197#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
198#define RExC_extralen (pRExC_state->extralen)
199#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 200#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 201#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 202#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
203#define RExC_open_parens (pRExC_state->open_parens)
204#define RExC_close_parens (pRExC_state->close_parens)
205#define RExC_opend (pRExC_state->opend)
81714fb9 206#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
207#define RExC_recurse (pRExC_state->recurse)
208#define RExC_recurse_count (pRExC_state->recurse_count)
b57e4118 209#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 210#define RExC_contains_locale (pRExC_state->contains_locale)
9d53c457
KW
211#define RExC_override_recoding (pRExC_state->override_recoding)
212#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
830247a4 213
cde0cee5 214
a687059c
LW
215#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
216#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
4d68ffa0 217 ((*s) == '{' && regcurly(s, FALSE)))
a687059c 218
35c8bce7
LW
219#ifdef SPSTART
220#undef SPSTART /* dratted cpp namespace... */
221#endif
a687059c
LW
222/*
223 * Flags to be passed up and down.
224 */
a687059c 225#define WORST 0 /* Worst case. */
a3b492c3 226#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 227
e64f369d 228/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
229 * character. (There needs to be a case: in the switch statement in regexec.c
230 * for any node marked SIMPLE.) Note that this is not the same thing as
231 * REGNODE_SIMPLE */
fda99bee 232#define SIMPLE 0x02
e64f369d 233#define SPSTART 0x04 /* Starts with * or + */
a3b492c3
YO
234#define TRYAGAIN 0x08 /* Weeded out a declaration. */
235#define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
a687059c 236
3dab1dad
YO
237#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
238
07be1b83
YO
239/* whether trie related optimizations are enabled */
240#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
241#define TRIE_STUDY_OPT
786e8c11 242#define FULL_TRIE_STUDY
07be1b83
YO
243#define TRIE_STCLASS
244#endif
1de06328
YO
245
246
40d049e4
YO
247
248#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
249#define PBITVAL(paren) (1 << ((paren) & 7))
250#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
251#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
252#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
253
bbd61b5f
KW
254/* If not already in utf8, do a longjmp back to the beginning */
255#define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
256#define REQUIRE_UTF8 STMT_START { \
257 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
258 } STMT_END
40d049e4 259
f19b1a63
KW
260/* This converts the named class defined in regcomp.h to its equivalent class
261 * number defined in handy.h. */
262#define namedclass_to_classnum(class) ((int) ((class) / 2))
263#define classnum_to_namedclass(classnum) ((classnum) * 2)
264
1de06328
YO
265/* About scan_data_t.
266
267 During optimisation we recurse through the regexp program performing
268 various inplace (keyhole style) optimisations. In addition study_chunk
269 and scan_commit populate this data structure with information about
270 what strings MUST appear in the pattern. We look for the longest
3b753521 271 string that must appear at a fixed location, and we look for the
1de06328
YO
272 longest string that may appear at a floating location. So for instance
273 in the pattern:
274
275 /FOO[xX]A.*B[xX]BAR/
276
277 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
278 strings (because they follow a .* construct). study_chunk will identify
279 both FOO and BAR as being the longest fixed and floating strings respectively.
280
281 The strings can be composites, for instance
282
283 /(f)(o)(o)/
284
285 will result in a composite fixed substring 'foo'.
286
287 For each string some basic information is maintained:
288
289 - offset or min_offset
290 This is the position the string must appear at, or not before.
291 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
292 characters must match before the string we are searching for.
293 Likewise when combined with minlenp and the length of the string it
1de06328
YO
294 tells us how many characters must appear after the string we have
295 found.
296
297 - max_offset
298 Only used for floating strings. This is the rightmost point that
3b753521 299 the string can appear at. If set to I32 max it indicates that the
1de06328
YO
300 string can occur infinitely far to the right.
301
302 - minlenp
2d608413
KW
303 A pointer to the minimum number of characters of the pattern that the
304 string was found inside. This is important as in the case of positive
1de06328
YO
305 lookahead or positive lookbehind we can have multiple patterns
306 involved. Consider
307
308 /(?=FOO).*F/
309
310 The minimum length of the pattern overall is 3, the minimum length
311 of the lookahead part is 3, but the minimum length of the part that
312 will actually match is 1. So 'FOO's minimum length is 3, but the
313 minimum length for the F is 1. This is important as the minimum length
314 is used to determine offsets in front of and behind the string being
315 looked for. Since strings can be composites this is the length of the
486ec47a 316 pattern at the time it was committed with a scan_commit. Note that
1de06328
YO
317 the length is calculated by study_chunk, so that the minimum lengths
318 are not known until the full pattern has been compiled, thus the
319 pointer to the value.
320
321 - lookbehind
322
323 In the case of lookbehind the string being searched for can be
324 offset past the start point of the final matching string.
325 If this value was just blithely removed from the min_offset it would
326 invalidate some of the calculations for how many chars must match
327 before or after (as they are derived from min_offset and minlen and
328 the length of the string being searched for).
329 When the final pattern is compiled and the data is moved from the
330 scan_data_t structure into the regexp structure the information
331 about lookbehind is factored in, with the information that would
332 have been lost precalculated in the end_shift field for the
333 associated string.
334
335 The fields pos_min and pos_delta are used to store the minimum offset
336 and the delta to the maximum offset at the current point in the pattern.
337
338*/
2c2d71f5
JH
339
340typedef struct scan_data_t {
1de06328
YO
341 /*I32 len_min; unused */
342 /*I32 len_delta; unused */
2c2d71f5
JH
343 I32 pos_min;
344 I32 pos_delta;
345 SV *last_found;
1de06328 346 I32 last_end; /* min value, <0 unless valid. */
2c2d71f5
JH
347 I32 last_start_min;
348 I32 last_start_max;
1de06328
YO
349 SV **longest; /* Either &l_fixed, or &l_float. */
350 SV *longest_fixed; /* longest fixed string found in pattern */
351 I32 offset_fixed; /* offset where it starts */
486ec47a 352 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
353 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
354 SV *longest_float; /* longest floating string found in pattern */
355 I32 offset_float_min; /* earliest point in string it can appear */
356 I32 offset_float_max; /* latest point in string it can appear */
486ec47a 357 I32 *minlen_float; /* pointer to the minlen relevant to the string */
1de06328 358 I32 lookbehind_float; /* is the position of the string modified by LB */
2c2d71f5
JH
359 I32 flags;
360 I32 whilem_c;
cb434fcc 361 I32 *last_closep;
653099ff 362 struct regnode_charclass_class *start_class;
2c2d71f5
JH
363} scan_data_t;
364
a687059c 365/*
e50aee73 366 * Forward declarations for pregcomp()'s friends.
a687059c 367 */
a0d0e21e 368
27da23d5 369static const scan_data_t zero_scan_data =
1de06328 370 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
371
372#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
373#define SF_BEFORE_SEOL 0x0001
374#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
375#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
376#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
377
09b7f37c
CB
378#ifdef NO_UNARY_PLUS
379# define SF_FIX_SHIFT_EOL (0+2)
380# define SF_FL_SHIFT_EOL (0+4)
381#else
382# define SF_FIX_SHIFT_EOL (+2)
383# define SF_FL_SHIFT_EOL (+4)
384#endif
c277df42
IZ
385
386#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
387#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
388
389#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
390#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
391#define SF_IS_INF 0x0040
392#define SF_HAS_PAR 0x0080
393#define SF_IN_PAR 0x0100
394#define SF_HAS_EVAL 0x0200
395#define SCF_DO_SUBSTR 0x0400
653099ff
GS
396#define SCF_DO_STCLASS_AND 0x0800
397#define SCF_DO_STCLASS_OR 0x1000
398#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 399#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 400
786e8c11 401#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
e2e6a0f1 402#define SCF_SEEN_ACCEPT 0x8000
07be1b83 403
43fead97 404#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
405
406/* The enums for all these are ordered so things work out correctly */
a62b1201 407#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
cfaf538b 408#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
00b27cfc 409#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
410#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
411#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1 412#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a725e29c 413#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 414
43fead97 415#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 416
93733859 417#define OOB_NAMEDCLASS -1
b8c5462f 418
8e661ac5
KW
419/* There is no code point that is out-of-bounds, so this is problematic. But
420 * its only current use is to initialize a variable that is always set before
421 * looked at. */
422#define OOB_UNICODE 0xDEADBEEF
423
a0ed51b3
LW
424#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
425#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
426
8615cb43 427
b45f050a
JF
428/* length of regex to show in messages that don't mark a position within */
429#define RegexLengthToShowInErrorMessages 127
430
431/*
432 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
433 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
434 * op/pragma/warn/regcomp.
435 */
7253e4e3
RK
436#define MARKER1 "<-- HERE" /* marker as it appears in the description */
437#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 438
7253e4e3 439#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
440
441/*
442 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
443 * arg. Show regex, up to a maximum length. If it's too long, chop and add
444 * "...".
445 */
58e23c8d 446#define _FAIL(code) STMT_START { \
bfed75c6 447 const char *ellipses = ""; \
ccb2c380
MP
448 IV len = RExC_end - RExC_precomp; \
449 \
450 if (!SIZE_ONLY) \
a5e7bc51 451 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
452 if (len > RegexLengthToShowInErrorMessages) { \
453 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
454 len = RegexLengthToShowInErrorMessages - 10; \
455 ellipses = "..."; \
456 } \
58e23c8d 457 code; \
ccb2c380 458} STMT_END
8615cb43 459
58e23c8d
YO
460#define FAIL(msg) _FAIL( \
461 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
462 msg, (int)len, RExC_precomp, ellipses))
463
464#define FAIL2(msg,arg) _FAIL( \
465 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
466 arg, (int)len, RExC_precomp, ellipses))
467
b45f050a 468/*
b45f050a
JF
469 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
470 */
ccb2c380 471#define Simple_vFAIL(m) STMT_START { \
a28509cc 472 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
473 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
474 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
475} STMT_END
b45f050a
JF
476
477/*
478 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
479 */
ccb2c380
MP
480#define vFAIL(m) STMT_START { \
481 if (!SIZE_ONLY) \
a5e7bc51 482 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
483 Simple_vFAIL(m); \
484} STMT_END
b45f050a
JF
485
486/*
487 * Like Simple_vFAIL(), but accepts two arguments.
488 */
ccb2c380 489#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 490 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
491 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
492 (int)offset, RExC_precomp, RExC_precomp + offset); \
493} STMT_END
b45f050a
JF
494
495/*
496 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
497 */
ccb2c380
MP
498#define vFAIL2(m,a1) STMT_START { \
499 if (!SIZE_ONLY) \
a5e7bc51 500 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
501 Simple_vFAIL2(m, a1); \
502} STMT_END
b45f050a
JF
503
504
505/*
506 * Like Simple_vFAIL(), but accepts three arguments.
507 */
ccb2c380 508#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 509 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
510 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
511 (int)offset, RExC_precomp, RExC_precomp + offset); \
512} STMT_END
b45f050a
JF
513
514/*
515 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
516 */
ccb2c380
MP
517#define vFAIL3(m,a1,a2) STMT_START { \
518 if (!SIZE_ONLY) \
a5e7bc51 519 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
520 Simple_vFAIL3(m, a1, a2); \
521} STMT_END
b45f050a
JF
522
523/*
524 * Like Simple_vFAIL(), but accepts four arguments.
525 */
ccb2c380 526#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 527 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
528 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
529 (int)offset, RExC_precomp, RExC_precomp + offset); \
530} STMT_END
b45f050a 531
95db3ffa
KW
532#define vFAIL4(m,a1,a2,a3) STMT_START { \
533 if (!SIZE_ONLY) \
534 SAVEFREESV(RExC_rx_sv); \
535 Simple_vFAIL4(m, a1, a2, a3); \
536} STMT_END
537
5e0a247b
KW
538/* m is not necessarily a "literal string", in this macro */
539#define reg_warn_non_literal_string(loc, m) STMT_START { \
540 const IV offset = loc - RExC_precomp; \
541 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
542 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
543} STMT_END
544
668c081a 545#define ckWARNreg(loc,m) STMT_START { \
a28509cc 546 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
547 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
548 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
549} STMT_END
550
0d6106aa
KW
551#define vWARN_dep(loc, m) STMT_START { \
552 const IV offset = loc - RExC_precomp; \
553 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
554 (int)offset, RExC_precomp, RExC_precomp + offset); \
555} STMT_END
556
147508a2
KW
557#define ckWARNdep(loc,m) STMT_START { \
558 const IV offset = loc - RExC_precomp; \
559 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
560 m REPORT_LOCATION, \
561 (int)offset, RExC_precomp, RExC_precomp + offset); \
562} STMT_END
563
668c081a 564#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 565 const IV offset = loc - RExC_precomp; \
d1d15184 566 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
567 m REPORT_LOCATION, \
568 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
569} STMT_END
570
2335b3d3
KW
571#define ckWARN2regdep(loc,m, a1) STMT_START { \
572 const IV offset = loc - RExC_precomp; \
573 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
574 m REPORT_LOCATION, \
575 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
576} STMT_END
577
668c081a 578#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 579 const IV offset = loc - RExC_precomp; \
668c081a 580 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
581 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
582} STMT_END
583
584#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 585 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
586 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
587 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
588} STMT_END
589
668c081a
NC
590#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
591 const IV offset = loc - RExC_precomp; \
592 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
593 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
594} STMT_END
595
ccb2c380 596#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 597 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
598 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
599 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
600} STMT_END
601
668c081a
NC
602#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
603 const IV offset = loc - RExC_precomp; \
604 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
605 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
606} STMT_END
607
ccb2c380 608#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 609 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
610 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
611 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
612} STMT_END
9d1d55b5 613
8615cb43 614
cd439c50 615/* Allow for side effects in s */
ccb2c380
MP
616#define REGC(c,s) STMT_START { \
617 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
618} STMT_END
cd439c50 619
fac92740
MJD
620/* Macros for recording node offsets. 20001227 mjd@plover.com
621 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
622 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
623 * Element 0 holds the number n.
07be1b83 624 * Position is 1 indexed.
fac92740 625 */
7122b237
YO
626#ifndef RE_TRACK_PATTERN_OFFSETS
627#define Set_Node_Offset_To_R(node,byte)
628#define Set_Node_Offset(node,byte)
629#define Set_Cur_Node_Offset
630#define Set_Node_Length_To_R(node,len)
631#define Set_Node_Length(node,len)
632#define Set_Node_Cur_Length(node)
633#define Node_Offset(n)
634#define Node_Length(n)
635#define Set_Node_Offset_Length(node,offset,len)
636#define ProgLen(ri) ri->u.proglen
637#define SetProgLen(ri,x) ri->u.proglen = x
638#else
639#define ProgLen(ri) ri->u.offsets[0]
640#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
641#define Set_Node_Offset_To_R(node,byte) STMT_START { \
642 if (! SIZE_ONLY) { \
643 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 644 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 645 if((node) < 0) { \
551405c4 646 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
647 } else { \
648 RExC_offsets[2*(node)-1] = (byte); \
649 } \
650 } \
651} STMT_END
652
653#define Set_Node_Offset(node,byte) \
654 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
655#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
656
657#define Set_Node_Length_To_R(node,len) STMT_START { \
658 if (! SIZE_ONLY) { \
659 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 660 __LINE__, (int)(node), (int)(len))); \
ccb2c380 661 if((node) < 0) { \
551405c4 662 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
663 } else { \
664 RExC_offsets[2*(node)] = (len); \
665 } \
666 } \
667} STMT_END
668
669#define Set_Node_Length(node,len) \
670 Set_Node_Length_To_R((node)-RExC_emit_start, len)
671#define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
672#define Set_Node_Cur_Length(node) \
673 Set_Node_Length(node, RExC_parse - parse_start)
fac92740
MJD
674
675/* Get offsets and lengths */
676#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
677#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
678
07be1b83
YO
679#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
680 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
681 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
682} STMT_END
7122b237 683#endif
07be1b83
YO
684
685#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
686#define EXPERIMENTAL_INPLACESCAN
f427392e 687#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 688
304ee84b
YO
689#define DEBUG_STUDYDATA(str,data,depth) \
690DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 691 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
692 "%*s" str "Pos:%"IVdf"/%"IVdf \
693 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
694 (int)(depth)*2, "", \
695 (IV)((data)->pos_min), \
696 (IV)((data)->pos_delta), \
304ee84b 697 (UV)((data)->flags), \
1de06328 698 (IV)((data)->whilem_c), \
304ee84b
YO
699 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
700 is_inf ? "INF " : "" \
1de06328
YO
701 ); \
702 if ((data)->last_found) \
703 PerlIO_printf(Perl_debug_log, \
704 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
705 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
706 SvPVX_const((data)->last_found), \
707 (IV)((data)->last_end), \
708 (IV)((data)->last_start_min), \
709 (IV)((data)->last_start_max), \
710 ((data)->longest && \
711 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
712 SvPVX_const((data)->longest_fixed), \
713 (IV)((data)->offset_fixed), \
714 ((data)->longest && \
715 (data)->longest==&((data)->longest_float)) ? "*" : "", \
716 SvPVX_const((data)->longest_float), \
717 (IV)((data)->offset_float_min), \
718 (IV)((data)->offset_float_max) \
719 ); \
720 PerlIO_printf(Perl_debug_log,"\n"); \
721});
722
653099ff 723/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 724 Update the longest found anchored substring and the longest found
653099ff
GS
725 floating substrings if needed. */
726
4327152a 727STATIC void
304ee84b 728S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
c277df42 729{
e1ec3a88
AL
730 const STRLEN l = CHR_SVLEN(data->last_found);
731 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 732 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 733
7918f24d
NC
734 PERL_ARGS_ASSERT_SCAN_COMMIT;
735
c277df42 736 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 737 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
738 if (*data->longest == data->longest_fixed) {
739 data->offset_fixed = l ? data->last_start_min : data->pos_min;
740 if (data->flags & SF_BEFORE_EOL)
b81d288d 741 data->flags
c277df42
IZ
742 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
743 else
744 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 745 data->minlen_fixed=minlenp;
1de06328 746 data->lookbehind_fixed=0;
a0ed51b3 747 }
304ee84b 748 else { /* *data->longest == data->longest_float */
c277df42 749 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
750 data->offset_float_max = (l
751 ? data->last_start_max
c277df42 752 : data->pos_min + data->pos_delta);
304ee84b 753 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
9051bda5 754 data->offset_float_max = I32_MAX;
c277df42 755 if (data->flags & SF_BEFORE_EOL)
b81d288d 756 data->flags
c277df42
IZ
757 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
758 else
759 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
760 data->minlen_float=minlenp;
761 data->lookbehind_float=0;
c277df42
IZ
762 }
763 }
764 SvCUR_set(data->last_found, 0);
0eda9292 765 {
a28509cc 766 SV * const sv = data->last_found;
097eb12c
AL
767 if (SvUTF8(sv) && SvMAGICAL(sv)) {
768 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
769 if (mg)
770 mg->mg_len = 0;
771 }
0eda9292 772 }
c277df42
IZ
773 data->last_end = -1;
774 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 775 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
776}
777
899d20b9
KW
778/* These macros set, clear and test whether the synthetic start class ('ssc',
779 * given by the parameter) matches an empty string (EOS). This uses the
780 * 'next_off' field in the node, to save a bit in the flags field. The ssc
781 * stands alone, so there is never a next_off, so this field is otherwise
782 * unused. The EOS information is used only for compilation, but theoretically
783 * it could be passed on to the execution code. This could be used to store
784 * more than one bit of information, but only this one is currently used. */
785#define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
786#define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
787#define TEST_SSC_EOS(node) cBOOL((node)->next_off)
788
653099ff
GS
789/* Can match anything (initialization) */
790STATIC void
3fffb88a 791S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 792{
7918f24d
NC
793 PERL_ARGS_ASSERT_CL_ANYTHING;
794
f8bef550 795 ANYOF_BITMAP_SETALL(cl);
899d20b9
KW
796 cl->flags = ANYOF_UNICODE_ALL;
797 SET_SSC_EOS(cl);
3fffb88a
KW
798
799 /* If any portion of the regex is to operate under locale rules,
800 * initialization includes it. The reason this isn't done for all regexes
801 * is that the optimizer was written under the assumption that locale was
802 * all-or-nothing. Given the complexity and lack of documentation in the
803 * optimizer, and that there are inadequate test cases for locale, so many
804 * parts of it may not work properly, it is safest to avoid locale unless
805 * necessary. */
806 if (RExC_contains_locale) {
9d7a1e63 807 ANYOF_CLASS_SETALL(cl); /* /l uses class */
b25a1036 808 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
3fffb88a 809 }
9d7a1e63
KW
810 else {
811 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
812 }
653099ff
GS
813}
814
815/* Can match anything (initialization) */
816STATIC int
5f66b61c 817S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
818{
819 int value;
820
7918f24d
NC
821 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
822
3732a889 823 for (value = 0; value < ANYOF_MAX; value += 2)
653099ff
GS
824 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
825 return 1;
1aa99e6b
IH
826 if (!(cl->flags & ANYOF_UNICODE_ALL))
827 return 0;
10edeb5d 828 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 829 return 0;
653099ff
GS
830 return 1;
831}
832
833/* Can match anything (initialization) */
834STATIC void
e755fd73 835S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 836{
7918f24d
NC
837 PERL_ARGS_ASSERT_CL_INIT;
838
8ecf7187 839 Zero(cl, 1, struct regnode_charclass_class);
653099ff 840 cl->type = ANYOF;
3fffb88a 841 cl_anything(pRExC_state, cl);
1411dba4 842 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
843}
844
1051e1c4
KW
845/* These two functions currently do the exact same thing */
846#define cl_init_zero S_cl_init
653099ff 847
dd58aee1
KW
848/* 'AND' a given class with another one. Can create false positives. 'cl'
849 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
850 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 851STATIC void
5f66b61c 852S_cl_and(struct regnode_charclass_class *cl,
a28509cc 853 const struct regnode_charclass_class *and_with)
653099ff 854{
7918f24d 855 PERL_ARGS_ASSERT_CL_AND;
40d049e4 856
954a2af6 857 assert(PL_regkind[and_with->type] == ANYOF);
1e6ade67 858
c6b76537 859 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
860 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
861 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 862 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
863 && !(and_with->flags & ANYOF_LOC_FOLD)
864 && !(cl->flags & ANYOF_LOC_FOLD)) {
653099ff
GS
865 int i;
866
867 if (and_with->flags & ANYOF_INVERT)
868 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
869 cl->bitmap[i] &= ~and_with->bitmap[i];
870 else
871 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
872 cl->bitmap[i] &= and_with->bitmap[i];
873 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 874
c6b76537 875 if (and_with->flags & ANYOF_INVERT) {
8951c461 876
c6b76537
KW
877 /* Here, the and'ed node is inverted. Get the AND of the flags that
878 * aren't affected by the inversion. Those that are affected are
879 * handled individually below */
880 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
881 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
882 cl->flags |= affected_flags;
883
884 /* We currently don't know how to deal with things that aren't in the
885 * bitmap, but we know that the intersection is no greater than what
886 * is already in cl, so let there be false positives that get sorted
887 * out after the synthetic start class succeeds, and the node is
888 * matched for real. */
889
890 /* The inversion of these two flags indicate that the resulting
891 * intersection doesn't have them */
892 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
893 cl->flags &= ~ANYOF_UNICODE_ALL;
894 }
c6b76537
KW
895 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
896 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 897 }
1aa99e6b 898 }
c6b76537 899 else { /* and'd node is not inverted */
3ad98780
KW
900 U8 outside_bitmap_but_not_utf8; /* Temp variable */
901
137165a6 902 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
903
904 /* Here 'and_with' doesn't match anything outside the bitmap
905 * (except possibly ANYOF_UNICODE_ALL), which means the
906 * intersection can't either, except for ANYOF_UNICODE_ALL, in
907 * which case we don't know what the intersection is, but it's no
908 * greater than what cl already has, so can just leave it alone,
909 * with possible false positives */
910 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
911 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 912 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 913 }
137165a6 914 }
c6b76537
KW
915 else if (! ANYOF_NONBITMAP(cl)) {
916
917 /* Here, 'and_with' does match something outside the bitmap, and cl
918 * doesn't have a list of things to match outside the bitmap. If
919 * cl can match all code points above 255, the intersection will
3ad98780
KW
920 * be those above-255 code points that 'and_with' matches. If cl
921 * can't match all Unicode code points, it means that it can't
922 * match anything outside the bitmap (since the 'if' that got us
923 * into this block tested for that), so we leave the bitmap empty.
924 */
c6b76537
KW
925 if (cl->flags & ANYOF_UNICODE_ALL) {
926 ARG_SET(cl, ARG(and_with));
3ad98780
KW
927
928 /* and_with's ARG may match things that don't require UTF8.
929 * And now cl's will too, in spite of this being an 'and'. See
930 * the comments below about the kludge */
931 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
932 }
933 }
934 else {
935 /* Here, both 'and_with' and cl match something outside the
936 * bitmap. Currently we do not do the intersection, so just match
937 * whatever cl had at the beginning. */
938 }
939
940
3ad98780
KW
941 /* Take the intersection of the two sets of flags. However, the
942 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
943 * kludge around the fact that this flag is not treated like the others
944 * which are initialized in cl_anything(). The way the optimizer works
945 * is that the synthetic start class (SSC) is initialized to match
946 * anything, and then the first time a real node is encountered, its
947 * values are AND'd with the SSC's with the result being the values of
948 * the real node. However, there are paths through the optimizer where
949 * the AND never gets called, so those initialized bits are set
950 * inappropriately, which is not usually a big deal, as they just cause
951 * false positives in the SSC, which will just mean a probably
952 * imperceptible slow down in execution. However this bit has a
953 * higher false positive consequence in that it can cause utf8.pm,
954 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
955 * bigger slowdown and also causes significant extra memory to be used.
956 * In order to prevent this, the code now takes a different tack. The
957 * bit isn't set unless some part of the regular expression needs it,
958 * but once set it won't get cleared. This means that these extra
959 * modules won't get loaded unless there was some path through the
960 * pattern that would have required them anyway, and so any false
961 * positives that occur by not ANDing them out when they could be
962 * aren't as severe as they would be if we treated this bit like all
963 * the others */
964 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
965 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 966 cl->flags &= and_with->flags;
3ad98780 967 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 968 }
653099ff
GS
969}
970
dd58aee1
KW
971/* 'OR' a given class with another one. Can create false positives. 'cl'
972 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
973 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 974STATIC void
3fffb88a 975S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 976{
7918f24d
NC
977 PERL_ARGS_ASSERT_CL_OR;
978
653099ff 979 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
980
981 /* Here, the or'd node is to be inverted. This means we take the
982 * complement of everything not in the bitmap, but currently we don't
983 * know what that is, so give up and match anything */
984 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 985 cl_anything(pRExC_state, cl);
c6b76537 986 }
653099ff
GS
987 /* We do not use
988 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
989 * <= (B1 | !B2) | (CL1 | !CL2)
990 * which is wasteful if CL2 is small, but we ignore CL2:
991 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
992 * XXXX Can we handle case-fold? Unclear:
993 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
994 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
995 */
c6b76537 996 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
997 && !(or_with->flags & ANYOF_LOC_FOLD)
998 && !(cl->flags & ANYOF_LOC_FOLD) ) {
653099ff
GS
999 int i;
1000
1001 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1002 cl->bitmap[i] |= ~or_with->bitmap[i];
1003 } /* XXXX: logic is complicated otherwise */
1004 else {
3fffb88a 1005 cl_anything(pRExC_state, cl);
653099ff 1006 }
c6b76537
KW
1007
1008 /* And, we can just take the union of the flags that aren't affected
1009 * by the inversion */
1010 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
1011
1012 /* For the remaining flags:
1013 ANYOF_UNICODE_ALL and inverted means to not match anything above
1014 255, which means that the union with cl should just be
1015 what cl has in it, so can ignore this flag
1016 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
1017 is 127-255 to match them, but then invert that, so the
1018 union with cl should just be what cl has in it, so can
1019 ignore this flag
1020 */
1021 } else { /* 'or_with' is not inverted */
653099ff
GS
1022 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1023 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
1024 && (!(or_with->flags & ANYOF_LOC_FOLD)
1025 || (cl->flags & ANYOF_LOC_FOLD)) ) {
653099ff
GS
1026 int i;
1027
1028 /* OR char bitmap and class bitmap separately */
1029 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1030 cl->bitmap[i] |= or_with->bitmap[i];
c96939e4
KW
1031 if (or_with->flags & ANYOF_CLASS) {
1032 ANYOF_CLASS_OR(or_with, cl);
1033 }
653099ff
GS
1034 }
1035 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 1036 cl_anything(pRExC_state, cl);
653099ff 1037 }
9826f543 1038
c6b76537
KW
1039 if (ANYOF_NONBITMAP(or_with)) {
1040
1041 /* Use the added node's outside-the-bit-map match if there isn't a
1042 * conflict. If there is a conflict (both nodes match something
1043 * outside the bitmap, but what they match outside is not the same
1044 * pointer, and hence not easily compared until XXX we extend
1045 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
1046 * match everything outside the bitmap. If that stuff is all above
1047 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
1048 if (! ANYOF_NONBITMAP(cl)) {
1049 ARG_SET(cl, ARG(or_with));
1050 }
1051 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
1052
1053 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1054 cl_anything(pRExC_state, cl);
1055 }
1056 else {
1057 cl->flags |= ANYOF_UNICODE_ALL;
1058 }
c6b76537 1059 }
4c34a693 1060 }
0b9668ee
KW
1061
1062 /* Take the union */
1063 cl->flags |= or_with->flags;
1aa99e6b 1064 }
653099ff
GS
1065}
1066
a3621e74
YO
1067#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1068#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1069#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1070#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1071
3dab1dad
YO
1072
1073#ifdef DEBUGGING
07be1b83 1074/*
2b8b4781
NC
1075 dump_trie(trie,widecharmap,revcharmap)
1076 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1077 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1078
1079 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1080 The _interim_ variants are used for debugging the interim
1081 tables that are used to generate the final compressed
1082 representation which is what dump_trie expects.
1083
486ec47a 1084 Part of the reason for their existence is to provide a form
3dab1dad 1085 of documentation as to how the different representations function.
07be1b83
YO
1086
1087*/
3dab1dad
YO
1088
1089/*
3dab1dad
YO
1090 Dumps the final compressed table form of the trie to Perl_debug_log.
1091 Used for debugging make_trie().
1092*/
b9a59e08 1093
3dab1dad 1094STATIC void
2b8b4781
NC
1095S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1096 AV *revcharmap, U32 depth)
3dab1dad
YO
1097{
1098 U32 state;
ab3bbdeb 1099 SV *sv=sv_newmortal();
55eed653 1100 int colwidth= widecharmap ? 6 : 4;
2e64971a 1101 U16 word;
3dab1dad
YO
1102 GET_RE_DEBUG_FLAGS_DECL;
1103
7918f24d 1104 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1105
3dab1dad
YO
1106 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1107 (int)depth * 2 + 2,"",
1108 "Match","Base","Ofs" );
1109
1110 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1111 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1112 if ( tmp ) {
ab3bbdeb
YO
1113 PerlIO_printf( Perl_debug_log, "%*s",
1114 colwidth,
ddc5bc0f 1115 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1116 PL_colors[0], PL_colors[1],
1117 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1118 PERL_PV_ESCAPE_FIRSTCHAR
1119 )
1120 );
3dab1dad
YO
1121 }
1122 }
1123 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1124 (int)depth * 2 + 2,"");
1125
1126 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1127 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1128 PerlIO_printf( Perl_debug_log, "\n");
1129
1e2e3d02 1130 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1131 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1132
1133 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1134
1135 if ( trie->states[ state ].wordnum ) {
1136 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1137 } else {
1138 PerlIO_printf( Perl_debug_log, "%6s", "" );
1139 }
1140
1141 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1142
1143 if ( base ) {
1144 U32 ofs = 0;
1145
1146 while( ( base + ofs < trie->uniquecharcount ) ||
1147 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1148 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1149 ofs++;
1150
1151 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1152
1153 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1154 if ( ( base + ofs >= trie->uniquecharcount ) &&
1155 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1156 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1157 {
ab3bbdeb
YO
1158 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1159 colwidth,
3dab1dad
YO
1160 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1161 } else {
ab3bbdeb 1162 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1163 }
1164 }
1165
1166 PerlIO_printf( Perl_debug_log, "]");
1167
1168 }
1169 PerlIO_printf( Perl_debug_log, "\n" );
1170 }
2e64971a
DM
1171 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1172 for (word=1; word <= trie->wordcount; word++) {
1173 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1174 (int)word, (int)(trie->wordinfo[word].prev),
1175 (int)(trie->wordinfo[word].len));
1176 }
1177 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1178}
1179/*
3dab1dad
YO
1180 Dumps a fully constructed but uncompressed trie in list form.
1181 List tries normally only are used for construction when the number of
1182 possible chars (trie->uniquecharcount) is very high.
1183 Used for debugging make_trie().
1184*/
1185STATIC void
55eed653 1186S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1187 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1188 U32 depth)
3dab1dad
YO
1189{
1190 U32 state;
ab3bbdeb 1191 SV *sv=sv_newmortal();
55eed653 1192 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1193 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1194
1195 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1196
3dab1dad 1197 /* print out the table precompression. */
ab3bbdeb
YO
1198 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1199 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1200 "------:-----+-----------------\n" );
3dab1dad
YO
1201
1202 for( state=1 ; state < next_alloc ; state ++ ) {
1203 U16 charid;
1204
ab3bbdeb 1205 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1206 (int)depth * 2 + 2,"", (UV)state );
1207 if ( ! trie->states[ state ].wordnum ) {
1208 PerlIO_printf( Perl_debug_log, "%5s| ","");
1209 } else {
1210 PerlIO_printf( Perl_debug_log, "W%4x| ",
1211 trie->states[ state ].wordnum
1212 );
1213 }
1214 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1215 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1216 if ( tmp ) {
1217 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1218 colwidth,
ddc5bc0f 1219 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1220 PL_colors[0], PL_colors[1],
1221 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1222 PERL_PV_ESCAPE_FIRSTCHAR
1223 ) ,
1e2e3d02
YO
1224 TRIE_LIST_ITEM(state,charid).forid,
1225 (UV)TRIE_LIST_ITEM(state,charid).newstate
1226 );
1227 if (!(charid % 10))
664e119d
RGS
1228 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1229 (int)((depth * 2) + 14), "");
1e2e3d02 1230 }
ab3bbdeb
YO
1231 }
1232 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1233 }
1234}
1235
1236/*
3dab1dad
YO
1237 Dumps a fully constructed but uncompressed trie in table form.
1238 This is the normal DFA style state transition table, with a few
1239 twists to facilitate compression later.
1240 Used for debugging make_trie().
1241*/
1242STATIC void
55eed653 1243S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1244 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1245 U32 depth)
3dab1dad
YO
1246{
1247 U32 state;
1248 U16 charid;
ab3bbdeb 1249 SV *sv=sv_newmortal();
55eed653 1250 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1251 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1252
1253 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1254
1255 /*
1256 print out the table precompression so that we can do a visual check
1257 that they are identical.
1258 */
1259
1260 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1261
1262 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1263 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1264 if ( tmp ) {
ab3bbdeb
YO
1265 PerlIO_printf( Perl_debug_log, "%*s",
1266 colwidth,
ddc5bc0f 1267 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1268 PL_colors[0], PL_colors[1],
1269 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1270 PERL_PV_ESCAPE_FIRSTCHAR
1271 )
1272 );
3dab1dad
YO
1273 }
1274 }
1275
1276 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1277
1278 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1279 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1280 }
1281
1282 PerlIO_printf( Perl_debug_log, "\n" );
1283
1284 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1285
1286 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1287 (int)depth * 2 + 2,"",
1288 (UV)TRIE_NODENUM( state ) );
1289
1290 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1291 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1292 if (v)
1293 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1294 else
1295 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1296 }
1297 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1298 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1299 } else {
1300 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1301 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1302 }
1303 }
07be1b83 1304}
3dab1dad
YO
1305
1306#endif
1307
2e64971a 1308
786e8c11
YO
1309/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1310 startbranch: the first branch in the whole branch sequence
1311 first : start branch of sequence of branch-exact nodes.
1312 May be the same as startbranch
1313 last : Thing following the last branch.
1314 May be the same as tail.
1315 tail : item following the branch sequence
1316 count : words in the sequence
1317 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1318 depth : indent depth
3dab1dad 1319
786e8c11 1320Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1321
786e8c11
YO
1322A trie is an N'ary tree where the branches are determined by digital
1323decomposition of the key. IE, at the root node you look up the 1st character and
1324follow that branch repeat until you find the end of the branches. Nodes can be
1325marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1326
786e8c11 1327 /he|she|his|hers/
72f13be8 1328
786e8c11
YO
1329would convert into the following structure. Numbers represent states, letters
1330following numbers represent valid transitions on the letter from that state, if
1331the number is in square brackets it represents an accepting state, otherwise it
1332will be in parenthesis.
07be1b83 1333
786e8c11
YO
1334 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1335 | |
1336 | (2)
1337 | |
1338 (1) +-i->(6)-+-s->[7]
1339 |
1340 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1341
786e8c11
YO
1342 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1343
1344This shows that when matching against the string 'hers' we will begin at state 1
1345read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1346then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1347is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1348single traverse. We store a mapping from accepting to state to which word was
1349matched, and then when we have multiple possibilities we try to complete the
1350rest of the regex in the order in which they occured in the alternation.
1351
1352The only prior NFA like behaviour that would be changed by the TRIE support is
1353the silent ignoring of duplicate alternations which are of the form:
1354
1355 / (DUPE|DUPE) X? (?{ ... }) Y /x
1356
4b714af6 1357Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1358and without the optimisation. With the optimisations dupes will be silently
486ec47a 1359ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1360the following demonstrates:
1361
1362 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1363
1364which prints out 'word' three times, but
1365
1366 'words'=~/(word|word|word)(?{ print $1 })S/
1367
1368which doesnt print it out at all. This is due to other optimisations kicking in.
1369
1370Example of what happens on a structural level:
1371
486ec47a 1372The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1373
1374 1: CURLYM[1] {1,32767}(18)
1375 5: BRANCH(8)
1376 6: EXACT <ac>(16)
1377 8: BRANCH(11)
1378 9: EXACT <ad>(16)
1379 11: BRANCH(14)
1380 12: EXACT <ab>(16)
1381 16: SUCCEED(0)
1382 17: NOTHING(18)
1383 18: END(0)
1384
1385This would be optimizable with startbranch=5, first=5, last=16, tail=16
1386and should turn into:
1387
1388 1: CURLYM[1] {1,32767}(18)
1389 5: TRIE(16)
1390 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1391 <ac>
1392 <ad>
1393 <ab>
1394 16: SUCCEED(0)
1395 17: NOTHING(18)
1396 18: END(0)
1397
1398Cases where tail != last would be like /(?foo|bar)baz/:
1399
1400 1: BRANCH(4)
1401 2: EXACT <foo>(8)
1402 4: BRANCH(7)
1403 5: EXACT <bar>(8)
1404 7: TAIL(8)
1405 8: EXACT <baz>(10)
1406 10: END(0)
1407
1408which would be optimizable with startbranch=1, first=1, last=7, tail=8
1409and would end up looking like:
1410
1411 1: TRIE(8)
1412 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1413 <foo>
1414 <bar>
1415 7: TAIL(8)
1416 8: EXACT <baz>(10)
1417 10: END(0)
1418
1419 d = uvuni_to_utf8_flags(d, uv, 0);
1420
1421is the recommended Unicode-aware way of saying
1422
1423 *(d++) = uv;
1424*/
1425
fab2782b 1426#define TRIE_STORE_REVCHAR(val) \
786e8c11 1427 STMT_START { \
73031816 1428 if (UTF) { \
fab2782b 1429 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1430 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
fab2782b 1431 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
73031816
NC
1432 SvCUR_set(zlopp, kapow - flrbbbbb); \
1433 SvPOK_on(zlopp); \
1434 SvUTF8_on(zlopp); \
1435 av_push(revcharmap, zlopp); \
1436 } else { \
fab2782b 1437 char ooooff = (char)val; \
73031816
NC
1438 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1439 } \
1440 } STMT_END
786e8c11 1441
fab2782b
YO
1442#define TRIE_READ_CHAR STMT_START { \
1443 wordlen++; \
1444 if ( UTF ) { \
1445 /* if it is UTF then it is either already folded, or does not need folding */ \
1446 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1447 } \
1448 else if (folder == PL_fold_latin1) { \
1449 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1450 if ( foldlen > 0 ) { \
1451 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1452 foldlen -= len; \
1453 scan += len; \
1454 len = 0; \
1455 } else { \
1456 len = 1; \
1457 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1458 skiplen = UNISKIP(uvc); \
1459 foldlen -= skiplen; \
1460 scan = foldbuf + skiplen; \
1461 } \
1462 } else { \
1463 /* raw data, will be folded later if needed */ \
1464 uvc = (U32)*uc; \
1465 len = 1; \
1466 } \
786e8c11
YO
1467} STMT_END
1468
1469
1470
1471#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1472 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1473 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1474 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1475 } \
1476 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1477 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1478 TRIE_LIST_CUR( state )++; \
1479} STMT_END
07be1b83 1480
786e8c11
YO
1481#define TRIE_LIST_NEW(state) STMT_START { \
1482 Newxz( trie->states[ state ].trans.list, \
1483 4, reg_trie_trans_le ); \
1484 TRIE_LIST_CUR( state ) = 1; \
1485 TRIE_LIST_LEN( state ) = 4; \
1486} STMT_END
07be1b83 1487
786e8c11
YO
1488#define TRIE_HANDLE_WORD(state) STMT_START { \
1489 U16 dupe= trie->states[ state ].wordnum; \
1490 regnode * const noper_next = regnext( noper ); \
1491 \
786e8c11
YO
1492 DEBUG_r({ \
1493 /* store the word for dumping */ \
1494 SV* tmp; \
1495 if (OP(noper) != NOTHING) \
740cce10 1496 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1497 else \
740cce10 1498 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1499 av_push( trie_words, tmp ); \
786e8c11
YO
1500 }); \
1501 \
1502 curword++; \
2e64971a
DM
1503 trie->wordinfo[curword].prev = 0; \
1504 trie->wordinfo[curword].len = wordlen; \
1505 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1506 \
1507 if ( noper_next < tail ) { \
1508 if (!trie->jump) \
c944940b 1509 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1510 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1511 if (!jumper) \
1512 jumper = noper_next; \
1513 if (!nextbranch) \
1514 nextbranch= regnext(cur); \
1515 } \
1516 \
1517 if ( dupe ) { \
2e64971a
DM
1518 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1519 /* chain, so that when the bits of chain are later */\
1520 /* linked together, the dups appear in the chain */\
1521 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1522 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1523 } else { \
1524 /* we haven't inserted this word yet. */ \
1525 trie->states[ state ].wordnum = curword; \
1526 } \
1527} STMT_END
07be1b83 1528
3dab1dad 1529
786e8c11
YO
1530#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1531 ( ( base + charid >= ucharcount \
1532 && base + charid < ubound \
1533 && state == trie->trans[ base - ucharcount + charid ].check \
1534 && trie->trans[ base - ucharcount + charid ].next ) \
1535 ? trie->trans[ base - ucharcount + charid ].next \
1536 : ( state==1 ? special : 0 ) \
1537 )
3dab1dad 1538
786e8c11
YO
1539#define MADE_TRIE 1
1540#define MADE_JUMP_TRIE 2
1541#define MADE_EXACT_TRIE 4
3dab1dad 1542
a3621e74 1543STATIC I32
786e8c11 1544S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1545{
27da23d5 1546 dVAR;
a3621e74
YO
1547 /* first pass, loop through and scan words */
1548 reg_trie_data *trie;
55eed653 1549 HV *widecharmap = NULL;
2b8b4781 1550 AV *revcharmap = newAV();
a3621e74 1551 regnode *cur;
9f7f3913 1552 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1553 STRLEN len = 0;
1554 UV uvc = 0;
1555 U16 curword = 0;
1556 U32 next_alloc = 0;
786e8c11
YO
1557 regnode *jumper = NULL;
1558 regnode *nextbranch = NULL;
7f69552c 1559 regnode *convert = NULL;
2e64971a 1560 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1561 /* we just use folder as a flag in utf8 */
1e696034 1562 const U8 * folder = NULL;
a3621e74 1563
2b8b4781
NC
1564#ifdef DEBUGGING
1565 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1566 AV *trie_words = NULL;
1567 /* along with revcharmap, this only used during construction but both are
1568 * useful during debugging so we store them in the struct when debugging.
8e11feef 1569 */
2b8b4781
NC
1570#else
1571 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1572 STRLEN trie_charcount=0;
3dab1dad 1573#endif
2b8b4781 1574 SV *re_trie_maxbuff;
a3621e74 1575 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1576
1577 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1578#ifndef DEBUGGING
1579 PERL_UNUSED_ARG(depth);
1580#endif
a3621e74 1581
1e696034 1582 switch (flags) {
79a81a6e 1583 case EXACT: break;
2f7f8cb1 1584 case EXACTFA:
fab2782b
YO
1585 case EXACTFU_SS:
1586 case EXACTFU_TRICKYFOLD:
1e696034
KW
1587 case EXACTFU: folder = PL_fold_latin1; break;
1588 case EXACTF: folder = PL_fold; break;
1589 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1590 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1591 }
1592
c944940b 1593 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1594 trie->refcount = 1;
3dab1dad 1595 trie->startstate = 1;
786e8c11 1596 trie->wordcount = word_count;
f8fc2ecf 1597 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1598 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1599 if (flags == EXACT)
c944940b 1600 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1601 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1602 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1603
a3621e74 1604 DEBUG_r({
2b8b4781 1605 trie_words = newAV();
a3621e74 1606 });
a3621e74 1607
0111c4fd 1608 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1609 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1610 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1611 }
df826430 1612 DEBUG_TRIE_COMPILE_r({
3dab1dad 1613 PerlIO_printf( Perl_debug_log,
786e8c11 1614 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1615 (int)depth * 2 + 2, "",
1616 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1617 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1618 (int)depth);
3dab1dad 1619 });
7f69552c
YO
1620
1621 /* Find the node we are going to overwrite */
1622 if ( first == startbranch && OP( last ) != BRANCH ) {
1623 /* whole branch chain */
1624 convert = first;
1625 } else {
1626 /* branch sub-chain */
1627 convert = NEXTOPER( first );
1628 }
1629
a3621e74
YO
1630 /* -- First loop and Setup --
1631
1632 We first traverse the branches and scan each word to determine if it
1633 contains widechars, and how many unique chars there are, this is
1634 important as we have to build a table with at least as many columns as we
1635 have unique chars.
1636
1637 We use an array of integers to represent the character codes 0..255
38a44b82 1638 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1639 native representation of the character value as the key and IV's for the
1640 coded index.
1641
1642 *TODO* If we keep track of how many times each character is used we can
1643 remap the columns so that the table compression later on is more
3b753521 1644 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1645 middle and the least common are on the outside. IMO this would be better
1646 than a most to least common mapping as theres a decent chance the most
1647 common letter will share a node with the least common, meaning the node
486ec47a 1648 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1649 case is when we have the least common nodes twice.
1650
1651 */
1652
a3621e74 1653 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1654 regnode *noper = NEXTOPER( cur );
e1ec3a88 1655 const U8 *uc = (U8*)STRING( noper );
df826430 1656 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1657 STRLEN foldlen = 0;
1658 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1659 STRLEN skiplen = 0;
2af232bd 1660 const U8 *scan = (U8*)NULL;
07be1b83 1661 U32 wordlen = 0; /* required init */
02daf0ab
YO
1662 STRLEN chars = 0;
1663 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1664
3dab1dad 1665 if (OP(noper) == NOTHING) {
df826430
YO
1666 regnode *noper_next= regnext(noper);
1667 if (noper_next != tail && OP(noper_next) == flags) {
1668 noper = noper_next;
1669 uc= (U8*)STRING(noper);
1670 e= uc + STR_LEN(noper);
1671 trie->minlen= STR_LEN(noper);
1672 } else {
1673 trie->minlen= 0;
1674 continue;
1675 }
3dab1dad 1676 }
df826430 1677
fab2782b 1678 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1679 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1680 regardless of encoding */
fab2782b
YO
1681 if (OP( noper ) == EXACTFU_SS) {
1682 /* false positives are ok, so just set this */
1683 TRIE_BITMAP_SET(trie,0xDF);
1684 }
1685 }
a3621e74 1686 for ( ; uc < e ; uc += len ) {
3dab1dad 1687 TRIE_CHARCOUNT(trie)++;
a3621e74 1688 TRIE_READ_CHAR;
3dab1dad 1689 chars++;
a3621e74 1690 if ( uvc < 256 ) {
fab2782b
YO
1691 if ( folder ) {
1692 U8 folded= folder[ (U8) uvc ];
1693 if ( !trie->charmap[ folded ] ) {
1694 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1695 TRIE_STORE_REVCHAR( folded );
1696 }
1697 }
a3621e74
YO
1698 if ( !trie->charmap[ uvc ] ) {
1699 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1700 TRIE_STORE_REVCHAR( uvc );
a3621e74 1701 }
02daf0ab 1702 if ( set_bit ) {
62012aee
KW
1703 /* store the codepoint in the bitmap, and its folded
1704 * equivalent. */
fab2782b 1705 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1706
1707 /* store the folded codepoint */
fab2782b 1708 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1709
1710 if ( !UTF ) {
1711 /* store first byte of utf8 representation of
acdf4139
KW
1712 variant codepoints */
1713 if (! UNI_IS_INVARIANT(uvc)) {
1714 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1715 }
1716 }
02daf0ab
YO
1717 set_bit = 0; /* We've done our bit :-) */
1718 }
a3621e74
YO
1719 } else {
1720 SV** svpp;
55eed653
NC
1721 if ( !widecharmap )
1722 widecharmap = newHV();
a3621e74 1723
55eed653 1724 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1725
1726 if ( !svpp )
e4584336 1727 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1728
1729 if ( !SvTRUE( *svpp ) ) {
1730 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1731 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1732 }
1733 }
1734 }
3dab1dad 1735 if( cur == first ) {
fab2782b
YO
1736 trie->minlen = chars;
1737 trie->maxlen = chars;
3dab1dad 1738 } else if (chars < trie->minlen) {
fab2782b 1739 trie->minlen = chars;
3dab1dad 1740 } else if (chars > trie->maxlen) {
fab2782b
YO
1741 trie->maxlen = chars;
1742 }
1743 if (OP( noper ) == EXACTFU_SS) {
1744 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1745 if (trie->minlen > 1)
1746 trie->minlen= 1;
1747 }
1748 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1749 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1750 * - We assume that any such sequence might match a 2 byte string */
1751 if (trie->minlen > 2 )
1752 trie->minlen= 2;
3dab1dad
YO
1753 }
1754
a3621e74
YO
1755 } /* end first pass */
1756 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1757 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1758 (int)depth * 2 + 2,"",
55eed653 1759 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1760 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1761 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1762 );
a3621e74
YO
1763
1764 /*
1765 We now know what we are dealing with in terms of unique chars and
1766 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1767 representation using a flat table will take. If it's over a reasonable
1768 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1769 conservative but potentially much slower representation using an array
1770 of lists.
1771
1772 At the end we convert both representations into the same compressed
1773 form that will be used in regexec.c for matching with. The latter
1774 is a form that cannot be used to construct with but has memory
1775 properties similar to the list form and access properties similar
1776 to the table form making it both suitable for fast searches and
1777 small enough that its feasable to store for the duration of a program.
1778
1779 See the comment in the code where the compressed table is produced
1780 inplace from the flat tabe representation for an explanation of how
1781 the compression works.
1782
1783 */
1784
1785
2e64971a
DM
1786 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1787 prev_states[1] = 0;
1788
3dab1dad 1789 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1790 /*
1791 Second Pass -- Array Of Lists Representation
1792
1793 Each state will be represented by a list of charid:state records
1794 (reg_trie_trans_le) the first such element holds the CUR and LEN
1795 points of the allocated array. (See defines above).
1796
1797 We build the initial structure using the lists, and then convert
1798 it into the compressed table form which allows faster lookups
1799 (but cant be modified once converted).
a3621e74
YO
1800 */
1801
a3621e74
YO
1802 STRLEN transcount = 1;
1803
1e2e3d02
YO
1804 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1805 "%*sCompiling trie using list compiler\n",
1806 (int)depth * 2 + 2, ""));
686b73d4 1807
c944940b
JH
1808 trie->states = (reg_trie_state *)
1809 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1810 sizeof(reg_trie_state) );
a3621e74
YO
1811 TRIE_LIST_NEW(1);
1812 next_alloc = 2;
1813
1814 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1815
df826430 1816 regnode *noper = NEXTOPER( cur );
c445ea15 1817 U8 *uc = (U8*)STRING( noper );
df826430 1818 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1819 U32 state = 1; /* required init */
1820 U16 charid = 0; /* sanity init */
1821 U8 *scan = (U8*)NULL; /* sanity init */
1822 STRLEN foldlen = 0; /* required init */
07be1b83 1823 U32 wordlen = 0; /* required init */
c445ea15 1824 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1825 STRLEN skiplen = 0;
c445ea15 1826
df826430
YO
1827 if (OP(noper) == NOTHING) {
1828 regnode *noper_next= regnext(noper);
1829 if (noper_next != tail && OP(noper_next) == flags) {
1830 noper = noper_next;
1831 uc= (U8*)STRING(noper);
1832 e= uc + STR_LEN(noper);
1833 }
1834 }
1835
3dab1dad 1836 if (OP(noper) != NOTHING) {
786e8c11 1837 for ( ; uc < e ; uc += len ) {
c445ea15 1838
786e8c11 1839 TRIE_READ_CHAR;
c445ea15 1840
786e8c11
YO
1841 if ( uvc < 256 ) {
1842 charid = trie->charmap[ uvc ];
c445ea15 1843 } else {
55eed653 1844 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1845 if ( !svpp ) {
1846 charid = 0;
1847 } else {
1848 charid=(U16)SvIV( *svpp );
1849 }
c445ea15 1850 }
786e8c11
YO
1851 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1852 if ( charid ) {
a3621e74 1853
786e8c11
YO
1854 U16 check;
1855 U32 newstate = 0;
a3621e74 1856
786e8c11
YO
1857 charid--;
1858 if ( !trie->states[ state ].trans.list ) {
1859 TRIE_LIST_NEW( state );
c445ea15 1860 }
786e8c11
YO
1861 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1862 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1863 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1864 break;
1865 }
1866 }
1867 if ( ! newstate ) {
1868 newstate = next_alloc++;
2e64971a 1869 prev_states[newstate] = state;
786e8c11
YO
1870 TRIE_LIST_PUSH( state, charid, newstate );
1871 transcount++;
1872 }
1873 state = newstate;
1874 } else {
1875 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1876 }
a28509cc 1877 }
c445ea15 1878 }
3dab1dad 1879 TRIE_HANDLE_WORD(state);
a3621e74
YO
1880
1881 } /* end second pass */
1882
1e2e3d02
YO
1883 /* next alloc is the NEXT state to be allocated */
1884 trie->statecount = next_alloc;
c944940b
JH
1885 trie->states = (reg_trie_state *)
1886 PerlMemShared_realloc( trie->states,
1887 next_alloc
1888 * sizeof(reg_trie_state) );
a3621e74 1889
3dab1dad 1890 /* and now dump it out before we compress it */
2b8b4781
NC
1891 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1892 revcharmap, next_alloc,
1893 depth+1)
1e2e3d02 1894 );
a3621e74 1895
c944940b
JH
1896 trie->trans = (reg_trie_trans *)
1897 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1898 {
1899 U32 state;
a3621e74
YO
1900 U32 tp = 0;
1901 U32 zp = 0;
1902
1903
1904 for( state=1 ; state < next_alloc ; state ++ ) {
1905 U32 base=0;
1906
1907 /*
1908 DEBUG_TRIE_COMPILE_MORE_r(
1909 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1910 );
1911 */
1912
1913 if (trie->states[state].trans.list) {
1914 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1915 U16 maxid=minid;
a28509cc 1916 U16 idx;
a3621e74
YO
1917
1918 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1919 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1920 if ( forid < minid ) {
1921 minid=forid;
1922 } else if ( forid > maxid ) {
1923 maxid=forid;
1924 }
a3621e74
YO
1925 }
1926 if ( transcount < tp + maxid - minid + 1) {
1927 transcount *= 2;
c944940b
JH
1928 trie->trans = (reg_trie_trans *)
1929 PerlMemShared_realloc( trie->trans,
446bd890
NC
1930 transcount
1931 * sizeof(reg_trie_trans) );
a3621e74
YO
1932 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1933 }
1934 base = trie->uniquecharcount + tp - minid;
1935 if ( maxid == minid ) {
1936 U32 set = 0;
1937 for ( ; zp < tp ; zp++ ) {
1938 if ( ! trie->trans[ zp ].next ) {
1939 base = trie->uniquecharcount + zp - minid;
1940 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1941 trie->trans[ zp ].check = state;
1942 set = 1;
1943 break;
1944 }
1945 }
1946 if ( !set ) {
1947 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1948 trie->trans[ tp ].check = state;
1949 tp++;
1950 zp = tp;
1951 }
1952 } else {
1953 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1954 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1955 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1956 trie->trans[ tid ].check = state;
1957 }
1958 tp += ( maxid - minid + 1 );
1959 }
1960 Safefree(trie->states[ state ].trans.list);
1961 }
1962 /*
1963 DEBUG_TRIE_COMPILE_MORE_r(
1964 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1965 );
1966 */
1967 trie->states[ state ].trans.base=base;
1968 }
cc601c31 1969 trie->lasttrans = tp + 1;
a3621e74
YO
1970 }
1971 } else {
1972 /*
1973 Second Pass -- Flat Table Representation.
1974
1975 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1976 We know that we will need Charcount+1 trans at most to store the data
1977 (one row per char at worst case) So we preallocate both structures
1978 assuming worst case.
1979
1980 We then construct the trie using only the .next slots of the entry
1981 structs.
1982
3b753521 1983 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1984 make compression both faster and easier by keeping track of how many non
1985 zero fields are in the node.
1986
1987 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1988 transition.
1989
1990 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1991 number representing the first entry of the node, and state as a
1992 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1993 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1994 are 2 entrys per node. eg:
1995
1996 A B A B
1997 1. 2 4 1. 3 7
1998 2. 0 3 3. 0 5
1999 3. 0 0 5. 0 0
2000 4. 0 0 7. 0 0
2001
2002 The table is internally in the right hand, idx form. However as we also
2003 have to deal with the states array which is indexed by nodenum we have to
2004 use TRIE_NODENUM() to convert.
2005
2006 */
1e2e3d02
YO
2007 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2008 "%*sCompiling trie using table compiler\n",
2009 (int)depth * 2 + 2, ""));
3dab1dad 2010
c944940b
JH
2011 trie->trans = (reg_trie_trans *)
2012 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2013 * trie->uniquecharcount + 1,
2014 sizeof(reg_trie_trans) );
2015 trie->states = (reg_trie_state *)
2016 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2017 sizeof(reg_trie_state) );
a3621e74
YO
2018 next_alloc = trie->uniquecharcount + 1;
2019
3dab1dad 2020
a3621e74
YO
2021 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2022
df826430 2023 regnode *noper = NEXTOPER( cur );
a28509cc 2024 const U8 *uc = (U8*)STRING( noper );
df826430 2025 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
2026
2027 U32 state = 1; /* required init */
2028
2029 U16 charid = 0; /* sanity init */
2030 U32 accept_state = 0; /* sanity init */
2031 U8 *scan = (U8*)NULL; /* sanity init */
2032
2033 STRLEN foldlen = 0; /* required init */
07be1b83 2034 U32 wordlen = 0; /* required init */
fab2782b 2035 STRLEN skiplen = 0;
a3621e74
YO
2036 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2037
df826430
YO
2038 if (OP(noper) == NOTHING) {
2039 regnode *noper_next= regnext(noper);
2040 if (noper_next != tail && OP(noper_next) == flags) {
2041 noper = noper_next;
2042 uc= (U8*)STRING(noper);
2043 e= uc + STR_LEN(noper);
2044 }
2045 }
fab2782b 2046
3dab1dad 2047 if ( OP(noper) != NOTHING ) {
786e8c11 2048 for ( ; uc < e ; uc += len ) {
a3621e74 2049
786e8c11 2050 TRIE_READ_CHAR;
a3621e74 2051
786e8c11
YO
2052 if ( uvc < 256 ) {
2053 charid = trie->charmap[ uvc ];
2054 } else {
55eed653 2055 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2056 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2057 }
786e8c11
YO
2058 if ( charid ) {
2059 charid--;
2060 if ( !trie->trans[ state + charid ].next ) {
2061 trie->trans[ state + charid ].next = next_alloc;
2062 trie->trans[ state ].check++;
2e64971a
DM
2063 prev_states[TRIE_NODENUM(next_alloc)]
2064 = TRIE_NODENUM(state);
786e8c11
YO
2065 next_alloc += trie->uniquecharcount;
2066 }
2067 state = trie->trans[ state + charid ].next;
2068 } else {
2069 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2070 }
2071 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2072 }
a3621e74 2073 }
3dab1dad
YO
2074 accept_state = TRIE_NODENUM( state );
2075 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2076
2077 } /* end second pass */
2078
3dab1dad 2079 /* and now dump it out before we compress it */
2b8b4781
NC
2080 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2081 revcharmap,
2082 next_alloc, depth+1));
a3621e74 2083
a3621e74
YO
2084 {
2085 /*
2086 * Inplace compress the table.*
2087
2088 For sparse data sets the table constructed by the trie algorithm will
2089 be mostly 0/FAIL transitions or to put it another way mostly empty.
2090 (Note that leaf nodes will not contain any transitions.)
2091
2092 This algorithm compresses the tables by eliminating most such
2093 transitions, at the cost of a modest bit of extra work during lookup:
2094
2095 - Each states[] entry contains a .base field which indicates the
2096 index in the state[] array wheres its transition data is stored.
2097
3b753521 2098 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2099
2100 - If .base is nonzero then charid is added to it to find an entry in
2101 the trans array.
2102
2103 -If trans[states[state].base+charid].check!=state then the
2104 transition is taken to be a 0/Fail transition. Thus if there are fail
2105 transitions at the front of the node then the .base offset will point
2106 somewhere inside the previous nodes data (or maybe even into a node
2107 even earlier), but the .check field determines if the transition is
2108 valid.
2109
786e8c11 2110 XXX - wrong maybe?
a3621e74 2111 The following process inplace converts the table to the compressed
3b753521 2112 table: We first do not compress the root node 1,and mark all its
a3621e74 2113 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2114 allows us to do a DFA construction from the compressed table later,
2115 and ensures that any .base pointers we calculate later are greater
2116 than 0.
a3621e74
YO
2117
2118 - We set 'pos' to indicate the first entry of the second node.
2119
2120 - We then iterate over the columns of the node, finding the first and
2121 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2122 and set the .check pointers accordingly, and advance pos
2123 appropriately and repreat for the next node. Note that when we copy
2124 the next pointers we have to convert them from the original
2125 NODEIDX form to NODENUM form as the former is not valid post
2126 compression.
2127
2128 - If a node has no transitions used we mark its base as 0 and do not
2129 advance the pos pointer.
2130
2131 - If a node only has one transition we use a second pointer into the
2132 structure to fill in allocated fail transitions from other states.
2133 This pointer is independent of the main pointer and scans forward
2134 looking for null transitions that are allocated to a state. When it
2135 finds one it writes the single transition into the "hole". If the
786e8c11 2136 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2137
2138 - Once compressed we can Renew/realloc the structures to release the
2139 excess space.
2140
2141 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2142 specifically Fig 3.47 and the associated pseudocode.
2143
2144 demq
2145 */
a3b680e6 2146 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2147 U32 state, charid;
a3621e74 2148 U32 pos = 0, zp=0;
1e2e3d02 2149 trie->statecount = laststate;
a3621e74
YO
2150
2151 for ( state = 1 ; state < laststate ; state++ ) {
2152 U8 flag = 0;
a28509cc
AL
2153 const U32 stateidx = TRIE_NODEIDX( state );
2154 const U32 o_used = trie->trans[ stateidx ].check;
2155 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2156 trie->trans[ stateidx ].check = 0;
2157
2158 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2159 if ( flag || trie->trans[ stateidx + charid ].next ) {
2160 if ( trie->trans[ stateidx + charid ].next ) {
2161 if (o_used == 1) {
2162 for ( ; zp < pos ; zp++ ) {
2163 if ( ! trie->trans[ zp ].next ) {
2164 break;
2165 }
2166 }
2167 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2168 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2169 trie->trans[ zp ].check = state;
2170 if ( ++zp > pos ) pos = zp;
2171 break;
2172 }
2173 used--;
2174 }
2175 if ( !flag ) {
2176 flag = 1;
2177 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2178 }
2179 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2180 trie->trans[ pos ].check = state;
2181 pos++;
2182 }
2183 }
2184 }
cc601c31 2185 trie->lasttrans = pos + 1;
c944940b
JH
2186 trie->states = (reg_trie_state *)
2187 PerlMemShared_realloc( trie->states, laststate
2188 * sizeof(reg_trie_state) );
a3621e74 2189 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2190 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2191 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2192 (int)depth * 2 + 2,"",
2193 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2194 (IV)next_alloc,
2195 (IV)pos,
a3621e74
YO
2196 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2197 );
2198
2199 } /* end table compress */
2200 }
1e2e3d02
YO
2201 DEBUG_TRIE_COMPILE_MORE_r(
2202 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2203 (int)depth * 2 + 2, "",
2204 (UV)trie->statecount,
2205 (UV)trie->lasttrans)
2206 );
cc601c31 2207 /* resize the trans array to remove unused space */
c944940b
JH
2208 trie->trans = (reg_trie_trans *)
2209 PerlMemShared_realloc( trie->trans, trie->lasttrans
2210 * sizeof(reg_trie_trans) );
a3621e74 2211
3b753521 2212 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2213 U8 nodetype =(U8)(flags & 0xFF);
2214 char *str=NULL;
786e8c11 2215
07be1b83 2216#ifdef DEBUGGING
e62cc96a 2217 regnode *optimize = NULL;
7122b237
YO
2218#ifdef RE_TRACK_PATTERN_OFFSETS
2219
b57a0404
JH
2220 U32 mjd_offset = 0;
2221 U32 mjd_nodelen = 0;
7122b237
YO
2222#endif /* RE_TRACK_PATTERN_OFFSETS */
2223#endif /* DEBUGGING */
a3621e74 2224 /*
3dab1dad
YO
2225 This means we convert either the first branch or the first Exact,
2226 depending on whether the thing following (in 'last') is a branch
2227 or not and whther first is the startbranch (ie is it a sub part of
2228 the alternation or is it the whole thing.)
3b753521 2229 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2230 the whole branch sequence, including the first.
a3621e74 2231 */
3dab1dad 2232 /* Find the node we are going to overwrite */
7f69552c 2233 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2234 /* branch sub-chain */
3dab1dad 2235 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2236#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2237 DEBUG_r({
2238 mjd_offset= Node_Offset((convert));
2239 mjd_nodelen= Node_Length((convert));
2240 });
7122b237 2241#endif
7f69552c 2242 /* whole branch chain */
7122b237
YO
2243 }
2244#ifdef RE_TRACK_PATTERN_OFFSETS
2245 else {
7f69552c
YO
2246 DEBUG_r({
2247 const regnode *nop = NEXTOPER( convert );
2248 mjd_offset= Node_Offset((nop));
2249 mjd_nodelen= Node_Length((nop));
2250 });
07be1b83
YO
2251 }
2252 DEBUG_OPTIMISE_r(
2253 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2254 (int)depth * 2 + 2, "",
786e8c11 2255 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2256 );
7122b237 2257#endif
3dab1dad
YO
2258 /* But first we check to see if there is a common prefix we can
2259 split out as an EXACT and put in front of the TRIE node. */
2260 trie->startstate= 1;
55eed653 2261 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2262 U32 state;
1e2e3d02 2263 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2264 U32 ofs = 0;
8e11feef
RGS
2265 I32 idx = -1;
2266 U32 count = 0;
2267 const U32 base = trie->states[ state ].trans.base;
a3621e74 2268
3dab1dad 2269 if ( trie->states[state].wordnum )
8e11feef 2270 count = 1;
a3621e74 2271
8e11feef 2272 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2273 if ( ( base + ofs >= trie->uniquecharcount ) &&
2274 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2275 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2276 {
3dab1dad 2277 if ( ++count > 1 ) {
2b8b4781 2278 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2279 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2280 if ( state == 1 ) break;
3dab1dad
YO
2281 if ( count == 2 ) {
2282 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2283 DEBUG_OPTIMISE_r(
8e11feef
RGS
2284 PerlIO_printf(Perl_debug_log,
2285 "%*sNew Start State=%"UVuf" Class: [",
2286 (int)depth * 2 + 2, "",
786e8c11 2287 (UV)state));
be8e71aa 2288 if (idx >= 0) {
2b8b4781 2289 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2290 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2291
3dab1dad 2292 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2293 if ( folder )
2294 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2295 DEBUG_OPTIMISE_r(
f1f66076 2296 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2297 );
8e11feef
RGS
2298 }
2299 }
2300 TRIE_BITMAP_SET(trie,*ch);
2301 if ( folder )
2302 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2303 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2304 }
2305 idx = ofs;
2306 }
3dab1dad
YO
2307 }
2308 if ( count == 1 ) {
2b8b4781 2309 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2310 STRLEN len;
2311 char *ch = SvPV( *tmp, len );
de734bd5
A
2312 DEBUG_OPTIMISE_r({
2313 SV *sv=sv_newmortal();
8e11feef
RGS
2314 PerlIO_printf( Perl_debug_log,
2315 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2316 (int)depth * 2 + 2, "",
de734bd5
A
2317 (UV)state, (UV)idx,
2318 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2319 PL_colors[0], PL_colors[1],
2320 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2321 PERL_PV_ESCAPE_FIRSTCHAR
2322 )
2323 );
2324 });
3dab1dad
YO
2325 if ( state==1 ) {
2326 OP( convert ) = nodetype;
2327 str=STRING(convert);
2328 STR_LEN(convert)=0;
2329 }
c490c714
YO
2330 STR_LEN(convert) += len;
2331 while (len--)
de734bd5 2332 *str++ = *ch++;
8e11feef 2333 } else {
f9049ba1 2334#ifdef DEBUGGING
8e11feef
RGS
2335 if (state>1)
2336 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2337#endif
8e11feef
RGS
2338 break;
2339 }
2340 }
2e64971a 2341 trie->prefixlen = (state-1);
3dab1dad 2342 if (str) {
8e11feef 2343 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2344 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2345 trie->startstate = state;
07be1b83
YO
2346 trie->minlen -= (state - 1);
2347 trie->maxlen -= (state - 1);
33809eae
JH
2348#ifdef DEBUGGING
2349 /* At least the UNICOS C compiler choked on this
2350 * being argument to DEBUG_r(), so let's just have
2351 * it right here. */
2352 if (
2353#ifdef PERL_EXT_RE_BUILD
2354 1
2355#else
2356 DEBUG_r_TEST
2357#endif
2358 ) {
2359 regnode *fix = convert;
2360 U32 word = trie->wordcount;
2361 mjd_nodelen++;
2362 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2363 while( ++fix < n ) {
2364 Set_Node_Offset_Length(fix, 0, 0);
2365 }
2366 while (word--) {
2367 SV ** const tmp = av_fetch( trie_words, word, 0 );
2368 if (tmp) {
2369 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2370 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2371 else
2372 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2373 }
2374 }
2375 }
2376#endif
8e11feef
RGS
2377 if (trie->maxlen) {
2378 convert = n;
2379 } else {
3dab1dad 2380 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2381 DEBUG_r(optimize= n);
3dab1dad
YO
2382 }
2383 }
2384 }
a5ca303d
YO
2385 if (!jumper)
2386 jumper = last;
3dab1dad 2387 if ( trie->maxlen ) {
8e11feef
RGS
2388 NEXT_OFF( convert ) = (U16)(tail - convert);
2389 ARG_SET( convert, data_slot );
786e8c11
YO
2390 /* Store the offset to the first unabsorbed branch in
2391 jump[0], which is otherwise unused by the jump logic.
2392 We use this when dumping a trie and during optimisation. */
2393 if (trie->jump)
7f69552c 2394 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2395
6c48061a
YO
2396 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2397 * and there is a bitmap
2398 * and the first "jump target" node we found leaves enough room
2399 * then convert the TRIE node into a TRIEC node, with the bitmap
2400 * embedded inline in the opcode - this is hypothetically faster.
2401 */
2402 if ( !trie->states[trie->startstate].wordnum
2403 && trie->bitmap
2404 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2405 {
2406 OP( convert ) = TRIEC;
2407 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2408 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2409 trie->bitmap= NULL;
2410 } else
2411 OP( convert ) = TRIE;
a3621e74 2412
3dab1dad
YO
2413 /* store the type in the flags */
2414 convert->flags = nodetype;
a5ca303d
YO
2415 DEBUG_r({
2416 optimize = convert
2417 + NODE_STEP_REGNODE
2418 + regarglen[ OP( convert ) ];
2419 });
2420 /* XXX We really should free up the resource in trie now,
2421 as we won't use them - (which resources?) dmq */
3dab1dad 2422 }
a3621e74 2423 /* needed for dumping*/
e62cc96a 2424 DEBUG_r(if (optimize) {
07be1b83 2425 regnode *opt = convert;
bcdf7404 2426
e62cc96a 2427 while ( ++opt < optimize) {
07be1b83
YO
2428 Set_Node_Offset_Length(opt,0,0);
2429 }
786e8c11
YO
2430 /*
2431 Try to clean up some of the debris left after the
2432 optimisation.
a3621e74 2433 */
786e8c11 2434 while( optimize < jumper ) {
07be1b83 2435 mjd_nodelen += Node_Length((optimize));
a3621e74 2436 OP( optimize ) = OPTIMIZED;
07be1b83 2437 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2438 optimize++;
2439 }
07be1b83 2440 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2441 });
2442 } /* end node insert */
2e64971a
DM
2443
2444 /* Finish populating the prev field of the wordinfo array. Walk back
2445 * from each accept state until we find another accept state, and if
2446 * so, point the first word's .prev field at the second word. If the
2447 * second already has a .prev field set, stop now. This will be the
2448 * case either if we've already processed that word's accept state,
3b753521
FN
2449 * or that state had multiple words, and the overspill words were
2450 * already linked up earlier.
2e64971a
DM
2451 */
2452 {
2453 U16 word;
2454 U32 state;
2455 U16 prev;
2456
2457 for (word=1; word <= trie->wordcount; word++) {
2458 prev = 0;
2459 if (trie->wordinfo[word].prev)
2460 continue;
2461 state = trie->wordinfo[word].accept;
2462 while (state) {
2463 state = prev_states[state];
2464 if (!state)
2465 break;
2466 prev = trie->states[state].wordnum;
2467 if (prev)
2468 break;
2469 }
2470 trie->wordinfo[word].prev = prev;
2471 }
2472 Safefree(prev_states);
2473 }
2474
2475
2476 /* and now dump out the compressed format */
2477 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2478
55eed653 2479 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2480#ifdef DEBUGGING
2481 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2482 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2483#else
03e70be4 2484 SvREFCNT_dec_NN(revcharmap);
07be1b83 2485#endif
786e8c11
YO
2486 return trie->jump
2487 ? MADE_JUMP_TRIE
2488 : trie->startstate>1
2489 ? MADE_EXACT_TRIE
2490 : MADE_TRIE;
2491}
2492
2493STATIC void
2494S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2495{
3b753521 2496/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2497
2498 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2499 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2500 ISBN 0-201-10088-6
2501
2502 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2503 suffix of the current state's 'word' that is also a proper prefix of another word in our
2504 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2505 the DFA not to have to restart after its tried and failed a word at a given point, it
2506 simply continues as though it had been matching the other word in the first place.
2507 Consider
2508 'abcdgu'=~/abcdefg|cdgu/
2509 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2510 fail, which would bring us to the state representing 'd' in the second word where we would
2511 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2512 */
2513 /* add a fail transition */
3251b653
NC
2514 const U32 trie_offset = ARG(source);
2515 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2516 U32 *q;
2517 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2518 const U32 numstates = trie->statecount;
786e8c11
YO
2519 const U32 ubound = trie->lasttrans + ucharcount;
2520 U32 q_read = 0;
2521 U32 q_write = 0;
2522 U32 charid;
2523 U32 base = trie->states[ 1 ].trans.base;
2524 U32 *fail;
2525 reg_ac_data *aho;
2526 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2527 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2528
2529 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2530#ifndef DEBUGGING
2531 PERL_UNUSED_ARG(depth);
2532#endif
2533
2534
2535 ARG_SET( stclass, data_slot );
c944940b 2536 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2537 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2538 aho->trie=trie_offset;
446bd890
NC
2539 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2540 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2541 Newxz( q, numstates, U32);
c944940b 2542 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2543 aho->refcount = 1;
2544 fail = aho->fail;
2545 /* initialize fail[0..1] to be 1 so that we always have
2546 a valid final fail state */
2547 fail[ 0 ] = fail[ 1 ] = 1;
2548
2549 for ( charid = 0; charid < ucharcount ; charid++ ) {
2550 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2551 if ( newstate ) {
2552 q[ q_write ] = newstate;
2553 /* set to point at the root */
2554 fail[ q[ q_write++ ] ]=1;
2555 }
2556 }
2557 while ( q_read < q_write) {
2558 const U32 cur = q[ q_read++ % numstates ];
2559 base = trie->states[ cur ].trans.base;
2560
2561 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2562 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2563 if (ch_state) {
2564 U32 fail_state = cur;
2565 U32 fail_base;
2566 do {
2567 fail_state = fail[ fail_state ];
2568 fail_base = aho->states[ fail_state ].trans.base;
2569 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2570
2571 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2572 fail[ ch_state ] = fail_state;
2573 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2574 {
2575 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2576 }
2577 q[ q_write++ % numstates] = ch_state;
2578 }
2579 }
2580 }
2581 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2582 when we fail in state 1, this allows us to use the
2583 charclass scan to find a valid start char. This is based on the principle
2584 that theres a good chance the string being searched contains lots of stuff
2585 that cant be a start char.
2586 */
2587 fail[ 0 ] = fail[ 1 ] = 0;
2588 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2589 PerlIO_printf(Perl_debug_log,
2590 "%*sStclass Failtable (%"UVuf" states): 0",
2591 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2592 );
786e8c11
YO
2593 for( q_read=1; q_read<numstates; q_read++ ) {
2594 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2595 }
2596 PerlIO_printf(Perl_debug_log, "\n");
2597 });
2598 Safefree(q);
2599 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2600}
2601
786e8c11 2602
a3621e74 2603/*
5d1c421c
JH
2604 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2605 * These need to be revisited when a newer toolchain becomes available.
2606 */
2607#if defined(__sparc64__) && defined(__GNUC__)
2608# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2609# undef SPARC64_GCC_WORKAROUND
2610# define SPARC64_GCC_WORKAROUND 1
2611# endif
2612#endif
2613
07be1b83 2614#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2615 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2616 SV * const mysv=sv_newmortal(); \
2617 regnode *Next = regnext(scan); \
2618 regprop(RExC_rx, mysv, scan); \
7f69552c 2619 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2620 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2621 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2622 }});
07be1b83 2623
1de06328 2624
bb914485 2625/* The below joins as many adjacent EXACTish nodes as possible into a single
0a982f06
KW
2626 * one. The regop may be changed if the node(s) contain certain sequences that
2627 * require special handling. The joining is only done if:
bb914485
KW
2628 * 1) there is room in the current conglomerated node to entirely contain the
2629 * next one.
2630 * 2) they are the exact same node type
2631 *
87b8b349 2632 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
2633 * these get optimized out
2634 *
0a982f06
KW
2635 * If a node is to match under /i (folded), the number of characters it matches
2636 * can be different than its character length if it contains a multi-character
2637 * fold. *min_subtract is set to the total delta of the input nodes.
bb914485 2638 *
a0c4c608
KW
2639 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2640 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2641 *
bb914485 2642 * This is as good a place as any to discuss the design of handling these
0a982f06
KW
2643 * multi-character fold sequences. It's been wrong in Perl for a very long
2644 * time. There are three code points in Unicode whose multi-character folds
2645 * were long ago discovered to mess things up. The previous designs for
2646 * dealing with these involved assigning a special node for them. This
2647 * approach doesn't work, as evidenced by this example:
a0c4c608 2648 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
0a982f06
KW
2649 * Both these fold to "sss", but if the pattern is parsed to create a node that
2650 * would match just the \xDF, it won't be able to handle the case where a
bb914485
KW
2651 * successful match would have to cross the node's boundary. The new approach
2652 * that hopefully generally solves the problem generates an EXACTFU_SS node
2653 * that is "sss".
2654 *
0a982f06
KW
2655 * It turns out that there are problems with all multi-character folds, and not
2656 * just these three. Now the code is general, for all such cases, but the
2657 * three still have some special handling. The approach taken is:
2658 * 1) This routine examines each EXACTFish node that could contain multi-
2659 * character fold sequences. It returns in *min_subtract how much to
9d071ca8 2660 * subtract from the the actual length of the string to get a real minimum
0a982f06
KW
2661 * match length; it is 0 if there are no multi-char folds. This delta is
2662 * used by the caller to adjust the min length of the match, and the delta
2663 * between min and max, so that the optimizer doesn't reject these
2664 * possibilities based on size constraints.
2665 * 2) Certain of these sequences require special handling by the trie code,
2666 * so, if found, this code changes the joined node type to special ops:
2667 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2668 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2669 * is used for an EXACTFU node that contains at least one "ss" sequence in
2670 * it. For non-UTF-8 patterns and strings, this is the only case where
2671 * there is a possible fold length change. That means that a regular
2672 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2673 * with length changes, and so can be processed faster. regexec.c takes
2674 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2675 * pre-folded by regcomp.c. This saves effort in regex matching.
87b8b349 2676 * However, the pre-folding isn't done for non-UTF8 patterns because the
0a982f06
KW
2677 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2678 * down by forcing the pattern into UTF8 unless necessary. Also what
2679 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2680 * possibilities for the non-UTF8 patterns are quite simple, except for
2681 * the sharp s. All the ones that don't involve a UTF-8 target string are
2682 * members of a fold-pair, and arrays are set up for all of them so that
2683 * the other member of the pair can be found quickly. Code elsewhere in
2684 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2685 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2686 * described in the next item.
bb914485
KW
2687 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2688 * 'ss' or not is not knowable at compile time. It will match iff the
2689 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2690 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
6012a526 2691 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
bb914485
KW
2692 * described in item 3). An assumption that the optimizer part of
2693 * regexec.c (probably unwittingly) makes is that a character in the
2694 * pattern corresponds to at most a single character in the target string.
2695 * (And I do mean character, and not byte here, unlike other parts of the
2696 * documentation that have never been updated to account for multibyte
2697 * Unicode.) This assumption is wrong only in this case, as all other
2698 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2699 * virtue of having this file pre-fold UTF-8 patterns. I'm
2700 * reluctant to try to change this assumption, so instead the code punts.
9d071ca8
KW
2701 * This routine examines EXACTF nodes for the sharp s, and returns a
2702 * boolean indicating whether or not the node is an EXACTF node that
2703 * contains a sharp s. When it is true, the caller sets a flag that later
2704 * causes the optimizer in this file to not set values for the floating
2705 * and fixed string lengths, and thus avoids the optimizer code in
2706 * regexec.c that makes the invalid assumption. Thus, there is no
2707 * optimization based on string lengths for EXACTF nodes that contain the
2708 * sharp s. This only happens for /id rules (which means the pattern
2709 * isn't in UTF-8).
bb914485 2710 */
1de06328 2711
9d071ca8 2712#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2713 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2714 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2715
be8e71aa 2716STATIC U32
9d071ca8 2717S_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
2718 /* Merge several consecutive EXACTish nodes into one. */
2719 regnode *n = regnext(scan);
2720 U32 stringok = 1;
2721 regnode *next = scan + NODE_SZ_STR(scan);
2722 U32 merged = 0;
2723 U32 stopnow = 0;
2724#ifdef DEBUGGING
2725 regnode *stop = scan;
72f13be8 2726 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2727#else
d47053eb
RGS
2728 PERL_UNUSED_ARG(depth);
2729#endif
7918f24d
NC
2730
2731 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2732#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2733 PERL_UNUSED_ARG(flags);
2734 PERL_UNUSED_ARG(val);
07be1b83 2735#endif
07be1b83 2736 DEBUG_PEEP("join",scan,depth);
bb914485 2737
3f410cf6
KW
2738 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2739 * EXACT ones that are mergeable to the current one. */
2740 while (n
2741 && (PL_regkind[OP(n)] == NOTHING
2742 || (stringok && OP(n) == OP(scan)))
07be1b83 2743 && NEXT_OFF(n)
3f410cf6
KW
2744 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2745 {
07be1b83
YO
2746
2747 if (OP(n) == TAIL || n > next)
2748 stringok = 0;
2749 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2750 DEBUG_PEEP("skip:",n,depth);
2751 NEXT_OFF(scan) += NEXT_OFF(n);
2752 next = n + NODE_STEP_REGNODE;
2753#ifdef DEBUGGING
2754 if (stringok)
2755 stop = n;
2756#endif
2757 n = regnext(n);
2758 }
2759 else if (stringok) {
786e8c11 2760 const unsigned int oldl = STR_LEN(scan);
07be1b83 2761 regnode * const nnext = regnext(n);
b2230d39 2762
87b8b349
KW
2763 /* XXX I (khw) kind of doubt that this works on platforms where
2764 * U8_MAX is above 255 because of lots of other assumptions */
79a81a6e 2765 /* Don't join if the sum can't fit into a single node */
b2230d39
KW
2766 if (oldl + STR_LEN(n) > U8_MAX)
2767 break;
07be1b83
YO
2768
2769 DEBUG_PEEP("merg",n,depth);
07be1b83 2770 merged++;
b2230d39 2771
07be1b83
YO
2772 NEXT_OFF(scan) += NEXT_OFF(n);
2773 STR_LEN(scan) += STR_LEN(n);
2774 next = n + NODE_SZ_STR(n);
2775 /* Now we can overwrite *n : */
2776 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2777#ifdef DEBUGGING
2778 stop = next - 1;
2779#endif
2780 n = nnext;
2781 if (stopnow) break;
2782 }
2783
d47053eb
RGS
2784#ifdef EXPERIMENTAL_INPLACESCAN
2785 if (flags && !NEXT_OFF(n)) {
2786 DEBUG_PEEP("atch", val, depth);
2787 if (reg_off_by_arg[OP(n)]) {
2788 ARG_SET(n, val - n);
2789 }
2790 else {
2791 NEXT_OFF(n) = val - n;
2792 }
2793 stopnow = 1;
2794 }
07be1b83
YO
2795#endif
2796 }
2c2b7f86 2797
9d071ca8 2798 *min_subtract = 0;
f758bddf 2799 *has_exactf_sharp_s = FALSE;
f646642f 2800
3f410cf6
KW
2801 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2802 * can now analyze for sequences of problematic code points. (Prior to
2803 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2804 * hence missed). The sequences only happen in folding, hence for any
2805 * non-EXACT EXACTish node */
86d6fcad 2806 if (OP(scan) != EXACT) {
0a982f06
KW
2807 const U8 * const s0 = (U8*) STRING(scan);
2808 const U8 * s = s0;
2809 const U8 * const s_end = s0 + STR_LEN(scan);
f758bddf
KW
2810
2811 /* One pass is made over the node's string looking for all the
2812 * possibilities. to avoid some tests in the loop, there are two main
2813 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2814 * non-UTF-8 */
2815 if (UTF) {
86d6fcad 2816
0a982f06
KW
2817 /* Examine the string for a multi-character fold sequence. UTF-8
2818 * patterns have all characters pre-folded by the time this code is
2819 * executed */
2820 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2821 length sequence we are looking for is 2 */
86d6fcad 2822 {
0a982f06
KW
2823 int count = 0;
2824 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2825 if (! len) { /* Not a multi-char fold: get next char */
2826 s += UTF8SKIP(s);
2827 continue;
2828 }
bb914485 2829
0a982f06
KW
2830 /* Nodes with 'ss' require special handling, except for EXACTFL
2831 * and EXACTFA for which there is no multi-char fold to this */
2832 if (len == 2 && *s == 's' && *(s+1) == 's'
2833 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2834 {
2835 count = 2;
2836 OP(scan) = EXACTFU_SS;
2837 s += 2;
2838 }
2839 else if (len == 6 /* len is the same in both ASCII and EBCDIC for these */
2840 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2841 COMBINING_DIAERESIS_UTF8
2842 COMBINING_ACUTE_ACCENT_UTF8,
2843 6)
2844 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2845 COMBINING_DIAERESIS_UTF8
2846 COMBINING_ACUTE_ACCENT_UTF8,
2847 6)))
2848 {
2849 count = 3;
2850
2851 /* These two folds require special handling by trie's, so
2852 * change the node type to indicate this. If EXACTFA and
2853 * EXACTFL were ever to be handled by trie's, this would
2854 * have to be changed. If this node has already been
2855 * changed to EXACTFU_SS in this loop, leave it as is. (I
2856 * (khw) think it doesn't matter in regexec.c for UTF
2857 * patterns, but no need to change it */
2858 if (OP(scan) == EXACTFU) {
2859 OP(scan) = EXACTFU_TRICKYFOLD;
2860 }
2861 s += 6;
2862 }
2863 else { /* Here is a generic multi-char fold. */
2864 const U8* multi_end = s + len;
2865
2866 /* Count how many characters in it. In the case of /l and
2867 * /aa, no folds which contain ASCII code points are
2868 * allowed, so check for those, and skip if found. (In
2869 * EXACTFL, no folds are allowed to any Latin1 code point,
2870 * not just ASCII. But there aren't any of these
2871 * currently, nor ever likely, so don't take the time to
2872 * test for them. The code that generates the
2873 * is_MULTI_foo() macros croaks should one actually get put
2874 * into Unicode .) */
2875 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2876 count = utf8_length(s, multi_end);
2877 s = multi_end;
2878 }
2879 else {
2880 while (s < multi_end) {
2881 if (isASCII(*s)) {
2882 s++;
2883 goto next_iteration;
2884 }
2885 else {
2886 s += UTF8SKIP(s);
2887 }
2888 count++;
2889 }
2890 }
2891 }
f758bddf 2892
0a982f06
KW
2893 /* The delta is how long the sequence is minus 1 (1 is how long
2894 * the character that folds to the sequence is) */
2895 *min_subtract += count - 1;
2896 next_iteration: ;
bb914485
KW
2897 }
2898 }
f758bddf 2899 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
bb914485 2900
0a982f06
KW
2901 /* Here, the pattern is not UTF-8. Look for the multi-char folds
2902 * that are all ASCII. As in the above case, EXACTFL and EXACTFA
2903 * nodes can't have multi-char folds to this range (and there are
2db97d41 2904 * no existing ones in the upper latin1 range). In the EXACTF
0a982f06
KW
2905 * case we look also for the sharp s, which can be in the final
2906 * position. Otherwise we can stop looking 1 byte earlier because
2907 * have to find at least two characters for a multi-fold */
f758bddf
KW
2908 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2909
0a982f06
KW
2910 /* The below is perhaps overboard, but this allows us to save a
2911 * test each time through the loop at the expense of a mask. This
2912 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2913 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2914 * are 64. This uses an exclusive 'or' to find that bit and then
2915 * inverts it to form a mask, with just a single 0, in the bit
2916 * position where 'S' and 's' differ. */
2917 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2918 const U8 s_masked = 's' & S_or_s_mask;
2919
2920 while (s < upper) {
40b1ba4f 2921 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
0a982f06
KW
2922 if (! len) { /* Not a multi-char fold. */
2923 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2924 {
2925 *has_exactf_sharp_s = TRUE;
2926 }
2927 s++;
2928 continue;
2929 }
2930
2931 if (len == 2
2932 && ((*s & S_or_s_mask) == s_masked)
2933 && ((*(s+1) & S_or_s_mask) == s_masked))
2934 {
2935
2936 /* EXACTF nodes need to know that the minimum length
2937 * changed so that a sharp s in the string can match this
2938 * ss in the pattern, but they remain EXACTF nodes, as they
2939 * won't match this unless the target string is is UTF-8,
2940 * which we don't know until runtime */
2941 if (OP(scan) != EXACTF) {
2942 OP(scan) = EXACTFU_SS;
2943 }
86d6fcad 2944 }
0a982f06
KW
2945
2946 *min_subtract += len - 1;
2947 s += len;
86d6fcad
KW
2948 }
2949 }
07be1b83 2950 }
3f410cf6 2951
07be1b83 2952#ifdef DEBUGGING
bb789b09
DM
2953 /* Allow dumping but overwriting the collection of skipped
2954 * ops and/or strings with fake optimized ops */
07be1b83
YO
2955 n = scan + NODE_SZ_STR(scan);
2956 while (n <= stop) {
bb789b09
DM
2957 OP(n) = OPTIMIZED;
2958 FLAGS(n) = 0;
2959 NEXT_OFF(n) = 0;
07be1b83
YO
2960 n++;
2961 }
2962#endif
2963 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2964 return stopnow;
2965}
2966
486ec47a 2967/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2968 Finds fixed substrings. */
2969
a0288114 2970/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
2971 to the position after last scanned or to NULL. */
2972
40d049e4
YO
2973#define INIT_AND_WITHP \
2974 assert(!and_withp); \
2975 Newx(and_withp,1,struct regnode_charclass_class); \
2976 SAVEFREEPV(and_withp)
07be1b83 2977
b515a41d 2978/* this is a chain of data about sub patterns we are processing that
486ec47a 2979 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
2980 we can simulate recursion without losing state. */
2981struct scan_frame;
2982typedef struct scan_frame {
2983 regnode *last; /* last node to process in this frame */
2984 regnode *next; /* next node to process when last is reached */
2985 struct scan_frame *prev; /*previous frame*/
2986 I32 stop; /* what stopparen do we use */
2987} scan_frame;
2988
304ee84b
YO
2989
2990#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2991
76e3520e 2992STATIC I32
40d049e4 2993S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 2994 I32 *minlenp, I32 *deltap,
40d049e4
YO
2995 regnode *last,
2996 scan_data_t *data,
2997 I32 stopparen,
2998 U8* recursed,
2999 struct regnode_charclass_class *and_withp,
3000 U32 flags, U32 depth)
c277df42
IZ
3001 /* scanp: Start here (read-write). */
3002 /* deltap: Write maxlen-minlen here. */
3003 /* last: Stop before this one. */
40d049e4
YO
3004 /* data: string data about the pattern */
3005 /* stopparen: treat close N as END */
3006 /* recursed: which subroutines have we recursed into */
3007 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 3008{
97aff369 3009 dVAR;
2d608413
KW
3010 I32 min = 0; /* There must be at least this number of characters to match */
3011 I32 pars = 0, code;
c277df42
IZ
3012 regnode *scan = *scanp, *next;
3013 I32 delta = 0;
3014 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 3015 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
3016 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3017 scan_data_t data_fake;
a3621e74 3018 SV *re_trie_maxbuff = NULL;
786e8c11 3019 regnode *first_non_open = scan;
e2e6a0f1 3020 I32 stopmin = I32_MAX;
8aa23a47 3021 scan_frame *frame = NULL;
a3621e74 3022 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 3023
7918f24d
NC
3024 PERL_ARGS_ASSERT_STUDY_CHUNK;
3025
13a24bad 3026#ifdef DEBUGGING
40d049e4 3027 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 3028#endif
40d049e4 3029
786e8c11 3030 if ( depth == 0 ) {
40d049e4 3031 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
3032 first_non_open=regnext(first_non_open);
3033 }
3034
b81d288d 3035
8aa23a47
YO
3036 fake_study_recurse:
3037 while ( scan && OP(scan) != END && scan < last ){
2d608413
KW
3038 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3039 node length to get a real minimum (because
3040 the folded version may be shorter) */
f758bddf 3041 bool has_exactf_sharp_s = FALSE;
8aa23a47 3042 /* Peephole optimizer: */
304ee84b 3043 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 3044 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
3045
3046 /* Its not clear to khw or hv why this is done here, and not in the
3047 * clauses that deal with EXACT nodes. khw's guess is that it's
3048 * because of a previous design */
9d071ca8 3049 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
3050
3051 /* Follow the next-chain of the current node and optimize
3052 away all the NOTHINGs from it. */
3053 if (OP(scan) != CURLYX) {
3054 const int max = (reg_off_by_arg[OP(scan)]
3055 ? I32_MAX
3056 /* I32 may be smaller than U16 on CRAYs! */
3057 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3058 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3059 int noff;
3060 regnode *n = scan;
686b73d4 3061
8aa23a47
YO
3062 /* Skip NOTHING and LONGJMP. */
3063 while ((n = regnext(n))
3064 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3065 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3066 && off + noff < max)
3067 off += noff;
3068 if (reg_off_by_arg[OP(scan)])
3069 ARG(scan) = off;
3070 else
3071 NEXT_OFF(scan) = off;
3072 }
a3621e74 3073
c277df42 3074
8aa23a47
YO
3075
3076 /* The principal pseudo-switch. Cannot be a switch, since we
3077 look into several different things. */
3078 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3079 || OP(scan) == IFTHEN) {
3080 next = regnext(scan);
3081 code = OP(scan);
3082 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3083
8aa23a47
YO
3084 if (OP(next) == code || code == IFTHEN) {
3085 /* NOTE - There is similar code to this block below for handling
3086 TRIE nodes on a re-study. If you change stuff here check there
3087 too. */
3088 I32 max1 = 0, min1 = I32_MAX, num = 0;
3089 struct regnode_charclass_class accum;
3090 regnode * const startbranch=scan;
686b73d4 3091
8aa23a47 3092 if (flags & SCF_DO_SUBSTR)
304ee84b 3093 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3094 if (flags & SCF_DO_STCLASS)
e755fd73 3095 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3096
3097 while (OP(scan) == code) {
3098 I32 deltanext, minnext, f = 0, fake;
3099 struct regnode_charclass_class this_class;
3100
3101 num++;
3102 data_fake.flags = 0;
3103 if (data) {
3104 data_fake.whilem_c = data->whilem_c;
3105 data_fake.last_closep = data->last_closep;
3106 }
3107 else
3108 data_fake.last_closep = &fake;
58e23c8d
YO
3109
3110 data_fake.pos_delta = delta;
8aa23a47
YO
3111 next = regnext(scan);
3112 scan = NEXTOPER(scan);
3113 if (code != BRANCH)
c277df42 3114 scan = NEXTOPER(scan);
8aa23a47 3115 if (flags & SCF_DO_STCLASS) {
e755fd73 3116 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3117 data_fake.start_class = &this_class;
3118 f = SCF_DO_STCLASS_AND;
58e23c8d 3119 }
8aa23a47
YO
3120 if (flags & SCF_WHILEM_VISITED_POS)
3121 f |= SCF_WHILEM_VISITED_POS;
3122
3123 /* we suppose the run is continuous, last=next...*/
3124 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3125 next, &data_fake,
3126 stopparen, recursed, NULL, f,depth+1);
3127 if (min1 > minnext)
3128 min1 = minnext;
3129 if (max1 < minnext + deltanext)
3130 max1 = minnext + deltanext;
3131 if (deltanext == I32_MAX)
3132 is_inf = is_inf_internal = 1;
3133 scan = next;
3134 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3135 pars++;
3136 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3137 if ( stopmin > minnext)
3138 stopmin = min + min1;
3139 flags &= ~SCF_DO_SUBSTR;
3140 if (data)
3141 data->flags |= SCF_SEEN_ACCEPT;
3142 }
3143 if (data) {
3144 if (data_fake.flags & SF_HAS_EVAL)
3145 data->flags |= SF_HAS_EVAL;
3146 data->whilem_c = data_fake.whilem_c;
3dab1dad 3147 }
8aa23a47 3148 if (flags & SCF_DO_STCLASS)
3fffb88a 3149 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3150 }
3151 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3152 min1 = 0;
3153 if (flags & SCF_DO_SUBSTR) {
3154 data->pos_min += min1;
3155 data->pos_delta += max1 - min1;
3156 if (max1 != min1 || is_inf)
3157 data->longest = &(data->longest_float);
3158 }
3159 min += min1;
3160 delta += max1 - min1;
3161 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3162 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3163 if (min1) {
3164 cl_and(data->start_class, and_withp);
3165 flags &= ~SCF_DO_STCLASS;
653099ff 3166 }
8aa23a47
YO
3167 }
3168 else if (flags & SCF_DO_STCLASS_AND) {
3169 if (min1) {
3170 cl_and(data->start_class, &accum);
3171 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3172 }
8aa23a47
YO
3173 else {
3174 /* Switch to OR mode: cache the old value of
3175 * data->start_class */
3176 INIT_AND_WITHP;
3177 StructCopy(data->start_class, and_withp,
3178 struct regnode_charclass_class);
3179 flags &= ~SCF_DO_STCLASS_AND;
3180 StructCopy(&accum, data->start_class,
3181 struct regnode_charclass_class);
3182 flags |= SCF_DO_STCLASS_OR;
899d20b9 3183 SET_SSC_EOS(data->start_class);
de0c8cb8 3184 }
8aa23a47 3185 }
a3621e74 3186
8aa23a47
YO
3187 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3188 /* demq.
a3621e74 3189
8aa23a47
YO
3190 Assuming this was/is a branch we are dealing with: 'scan' now
3191 points at the item that follows the branch sequence, whatever
3192 it is. We now start at the beginning of the sequence and look
3193 for subsequences of
a3621e74 3194
8aa23a47
YO
3195 BRANCH->EXACT=>x1
3196 BRANCH->EXACT=>x2
3197 tail
a3621e74 3198
8aa23a47 3199 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3200
486ec47a 3201 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3202 element into a trie and then add the subsequent branch exact
3203 strings to the trie.
a3621e74 3204
8aa23a47 3205 We have two cases
a3621e74 3206
3b753521 3207 1. patterns where the whole set of branches can be converted.
a3621e74 3208
8aa23a47 3209 2. patterns where only a subset can be converted.
a3621e74 3210
8aa23a47
YO
3211 In case 1 we can replace the whole set with a single regop
3212 for the trie. In case 2 we need to keep the start and end
3b753521 3213 branches so
a3621e74 3214
8aa23a47
YO
3215 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3216 becomes BRANCH TRIE; BRANCH X;
786e8c11 3217
8aa23a47
YO
3218 There is an additional case, that being where there is a
3219 common prefix, which gets split out into an EXACT like node
3220 preceding the TRIE node.
a3621e74 3221
8aa23a47
YO
3222 If x(1..n)==tail then we can do a simple trie, if not we make
3223 a "jump" trie, such that when we match the appropriate word
486ec47a 3224 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3225 a nested if into a case structure of sorts.
b515a41d 3226
8aa23a47 3227 */
686b73d4 3228
8aa23a47
YO
3229 int made=0;
3230 if (!re_trie_maxbuff) {
3231 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3232 if (!SvIOK(re_trie_maxbuff))
3233 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3234 }
3235 if ( SvIV(re_trie_maxbuff)>=0 ) {
3236 regnode *cur;
3237 regnode *first = (regnode *)NULL;
3238 regnode *last = (regnode *)NULL;
3239 regnode *tail = scan;
fab2782b 3240 U8 trietype = 0;
8aa23a47 3241 U32 count=0;
a3621e74
YO
3242
3243#ifdef DEBUGGING
8aa23a47 3244 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3245#endif
8aa23a47
YO
3246 /* var tail is used because there may be a TAIL
3247 regop in the way. Ie, the exacts will point to the
3248 thing following the TAIL, but the last branch will
3249 point at the TAIL. So we advance tail. If we
3250 have nested (?:) we may have to move through several
3251 tails.
3252 */
3253
3254 while ( OP( tail ) == TAIL ) {
3255 /* this is the TAIL generated by (?:) */
3256 tail = regnext( tail );
3257 }
a3621e74 3258
8aa23a47 3259
df826430 3260 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3261 regprop(RExC_rx, mysv, tail );
3262 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3263 (int)depth * 2 + 2, "",
3264 "Looking for TRIE'able sequences. Tail node is: ",
3265 SvPV_nolen_const( mysv )
3266 );
3267 });
3268
3269 /*
3270
fab2782b
YO
3271 Step through the branches
3272 cur represents each branch,
3273 noper is the first thing to be matched as part of that branch
3274 noper_next is the regnext() of that node.
3275
3276 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3277 via a "jump trie" but we also support building with NOJUMPTRIE,
3278 which restricts the trie logic to structures like /FOO|BAR/.
3279
3280 If noper is a trieable nodetype then the branch is a possible optimization
3281 target. If we are building under NOJUMPTRIE then we require that noper_next
3282 is the same as scan (our current position in the regex program).
3283
3284 Once we have two or more consecutive such branches we can create a
3285 trie of the EXACT's contents and stitch it in place into the program.
3286
3287 If the sequence represents all of the branches in the alternation we
3288 replace the entire thing with a single TRIE node.
3289
3290 Otherwise when it is a subsequence we need to stitch it in place and
3291 replace only the relevant branches. This means the first branch has
3292 to remain as it is used by the alternation logic, and its next pointer,
3293 and needs to be repointed at the item on the branch chain following
3294 the last branch we have optimized away.
3295
3296 This could be either a BRANCH, in which case the subsequence is internal,
3297 or it could be the item following the branch sequence in which case the
3298 subsequence is at the end (which does not necessarily mean the first node
3299 is the start of the alternation).
3300
3301 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3302
3303 optype | trietype
3304 ----------------+-----------
3305 NOTHING | NOTHING
3306 EXACT | EXACT
3307 EXACTFU | EXACTFU
3308 EXACTFU_SS | EXACTFU
3309 EXACTFU_TRICKYFOLD | EXACTFU
3310 EXACTFA | 0
3311
8aa23a47
YO
3312
3313 */
fab2782b
YO
3314#define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3315 ( EXACT == (X) ) ? EXACT : \
3316 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3317 0 )
8aa23a47
YO
3318
3319 /* dont use tail as the end marker for this traverse */
3320 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3321 regnode * const noper = NEXTOPER( cur );
fab2782b
YO
3322 U8 noper_type = OP( noper );
3323 U8 noper_trietype = TRIE_TYPE( noper_type );
b515a41d 3324#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3325 regnode * const noper_next = regnext( noper );
df826430
YO
3326 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3327 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
b515a41d
YO
3328#endif
3329
df826430 3330 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3331 regprop(RExC_rx, mysv, cur);
3332 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3333 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3334
3335 regprop(RExC_rx, mysv, noper);
3336 PerlIO_printf( Perl_debug_log, " -> %s",
3337 SvPV_nolen_const(mysv));
3338
3339 if ( noper_next ) {
3340 regprop(RExC_rx, mysv, noper_next );
3341 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3342 SvPV_nolen_const(mysv));
3343 }
df826430
YO
3344 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3345 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3346 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3347 );
8aa23a47 3348 });
fab2782b
YO
3349
3350 /* Is noper a trieable nodetype that can be merged with the
3351 * current trie (if there is one)? */
3352 if ( noper_trietype
3353 &&
3354 (
df826430
YO
3355 ( noper_trietype == NOTHING)
3356 || ( trietype == NOTHING )
a40630bf 3357 || ( trietype == noper_trietype )
fab2782b 3358 )
786e8c11 3359#ifdef NOJUMPTRIE
8aa23a47 3360 && noper_next == tail
786e8c11 3361#endif
8aa23a47
YO
3362 && count < U16_MAX)
3363 {
fab2782b
YO
3364 /* Handle mergable triable node
3365 * Either we are the first node in a new trieable sequence,
3366 * in which case we do some bookkeeping, otherwise we update
3367 * the end pointer. */
fab2782b 3368 if ( !first ) {
3b6759a6 3369 first = cur;
df826430
YO
3370 if ( noper_trietype == NOTHING ) {
3371#if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3372 regnode * const noper_next = regnext( noper );
3b6759a6 3373 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
df826430
YO
3374 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3375#endif
3376
190c1910 3377 if ( noper_next_trietype ) {
df826430 3378 trietype = noper_next_trietype;
190c1910
YO
3379 } else if (noper_next_type) {
3380 /* a NOTHING regop is 1 regop wide. We need at least two
3381 * for a trie so we can't merge this in */
3382 first = NULL;
3383 }
3384 } else {
3385 trietype = noper_trietype;
3b6759a6 3386 }
8aa23a47 3387 } else {
fab2782b
YO
3388 if ( trietype == NOTHING )
3389 trietype = noper_trietype;
8aa23a47
YO
3390 last = cur;
3391 }
df826430
YO
3392 if (first)
3393 count++;
fab2782b
YO
3394 } /* end handle mergable triable node */
3395 else {
3396 /* handle unmergable node -
3397 * noper may either be a triable node which can not be tried
3398 * together with the current trie, or a non triable node */
729aaeb5
YO
3399 if ( last ) {
3400 /* If last is set and trietype is not NOTHING then we have found
3401 * at least two triable branch sequences in a row of a similar
3402 * trietype so we can turn them into a trie. If/when we
3403 * allow NOTHING to start a trie sequence this condition will be
3404 * required, and it isn't expensive so we leave it in for now. */
e6351b37 3405 if ( trietype && trietype != NOTHING )
729aaeb5
YO
3406 make_trie( pRExC_state,
3407 startbranch, first, cur, tail, count,
3408 trietype, depth+1 );
fab2782b 3409 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
8aa23a47 3410 }
fab2782b 3411 if ( noper_trietype
786e8c11 3412#ifdef NOJUMPTRIE
8aa23a47 3413 && noper_next == tail
786e8c11 3414#endif
8aa23a47 3415 ){
fab2782b 3416 /* noper is triable, so we can start a new trie sequence */
8aa23a47
YO
3417 count = 1;
3418 first = cur;
fab2782b
YO
3419 trietype = noper_trietype;
3420 } else if (first) {
3421 /* if we already saw a first but the current node is not triable then we have
3422 * to reset the first information. */
8aa23a47
YO
3423 count = 0;
3424 first = NULL;
fab2782b 3425 trietype = 0;
8aa23a47 3426 }
fab2782b
YO
3427 } /* end handle unmergable node */
3428 } /* loop over branches */
df826430 3429 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3430 regprop(RExC_rx, mysv, cur);
3431 PerlIO_printf( Perl_debug_log,
3432 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3433 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3434
3435 });
e6351b37 3436 if ( last && trietype ) {
3b6759a6
YO
3437 if ( trietype != NOTHING ) {
3438 /* the last branch of the sequence was part of a trie,
3439 * so we have to construct it here outside of the loop
3440 */
3441 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
686b73d4 3442#ifdef TRIE_STUDY_OPT
3b6759a6
YO
3443 if ( ((made == MADE_EXACT_TRIE &&
3444 startbranch == first)
3445 || ( first_non_open == first )) &&
3446 depth==0 ) {
3447 flags |= SCF_TRIE_RESTUDY;
3448 if ( startbranch == first
3449 && scan == tail )
3450 {
3451 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3452 }
8aa23a47 3453 }
3dab1dad 3454#endif
3b6759a6
YO
3455 } else {
3456 /* at this point we know whatever we have is a NOTHING sequence/branch
3457 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3458 */
3459 if ( startbranch == first ) {
3460 regnode *opt;
3461 /* the entire thing is a NOTHING sequence, something like this:
3462 * (?:|) So we can turn it into a plain NOTHING op. */
3463 DEBUG_TRIE_COMPILE_r({
3464 regprop(RExC_rx, mysv, cur);
3465 PerlIO_printf( Perl_debug_log,
3466 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3467 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3468
3469 });
3470 OP(startbranch)= NOTHING;
3471 NEXT_OFF(startbranch)= tail - startbranch;
3472 for ( opt= startbranch + 1; opt < tail ; opt++ )
3473 OP(opt)= OPTIMIZED;
3474 }
3475 }
fab2782b
YO
3476 } /* end if ( last) */
3477 } /* TRIE_MAXBUF is non zero */
8aa23a47
YO
3478
3479 } /* do trie */
3480
653099ff 3481 }
8aa23a47
YO
3482 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3483 scan = NEXTOPER(NEXTOPER(scan));
3484 } else /* single branch is optimized. */
3485 scan = NEXTOPER(scan);
3486 continue;
3487 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3488 scan_frame *newframe = NULL;
3489 I32 paren;
3490 regnode *start;
3491 regnode *end;
3492
3493 if (OP(scan) != SUSPEND) {
3494 /* set the pointer */
3495 if (OP(scan) == GOSUB) {
3496 paren = ARG(scan);
3497 RExC_recurse[ARG2L(scan)] = scan;
3498 start = RExC_open_parens[paren-1];
3499 end = RExC_close_parens[paren-1];
3500 } else {
3501 paren = 0;
f8fc2ecf 3502 start = RExC_rxi->program + 1;
8aa23a47
YO
3503 end = RExC_opend;
3504 }
3505 if (!recursed) {
3506 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3507 SAVEFREEPV(recursed);
3508 }
3509 if (!PAREN_TEST(recursed,paren+1)) {
3510 PAREN_SET(recursed,paren+1);
3511 Newx(newframe,1,scan_frame);
3512 } else {
3513 if (flags & SCF_DO_SUBSTR) {
304ee84b 3514 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3515 data->longest = &(data->longest_float);
3516 }
3517 is_inf = is_inf_internal = 1;
3518 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3519 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3520 flags &= ~SCF_DO_STCLASS;
3521 }
3522 } else {
3523 Newx(newframe,1,scan_frame);
3524 paren = stopparen;
3525 start = scan+2;
3526 end = regnext(scan);
3527 }
3528 if (newframe) {
3529 assert(start);
3530 assert(end);
3531 SAVEFREEPV(newframe);
3532 newframe->next = regnext(scan);
3533 newframe->last = last;
3534 newframe->stop = stopparen;
3535 newframe->prev = frame;
3536
3537 frame = newframe;
3538 scan = start;
3539 stopparen = paren;
3540 last = end;
3541
3542 continue;
3543 }
3544 }
3545 else if (OP(scan) == EXACT) {
3546 I32 l = STR_LEN(scan);
3547 UV uc;
3548 if (UTF) {
3549 const U8 * const s = (U8*)STRING(scan);
4b88fb76 3550 uc = utf8_to_uvchr_buf(s, s + l, NULL);
8aa23a47 3551 l = utf8_length(s, s + l);
8aa23a47
YO
3552 } else {
3553 uc = *((U8*)STRING(scan));
3554 }
3555 min += l;
3556 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3557 /* The code below prefers earlier match for fixed
3558 offset, later match for variable offset. */
3559 if (data->last_end == -1) { /* Update the start info. */
3560 data->last_start_min = data->pos_min;
3561 data->last_start_max = is_inf
3562 ? I32_MAX : data->pos_min + data->pos_delta;
b515a41d 3563 }
8aa23a47
YO
3564 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3565 if (UTF)
3566 SvUTF8_on(data->last_found);
3567 {
3568 SV * const sv = data->last_found;
3569 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3570 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3571 if (mg && mg->mg_len >= 0)
3572 mg->mg_len += utf8_length((U8*)STRING(scan),
3573 (U8*)STRING(scan)+STR_LEN(scan));
b515a41d 3574 }
8aa23a47
YO
3575 data->last_end = data->pos_min + l;
3576 data->pos_min += l; /* As in the first entry. */
3577 data->flags &= ~SF_BEFORE_EOL;
3578 }
3579 if (flags & SCF_DO_STCLASS_AND) {
3580 /* Check wheth