This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
Merge branch 'data-dumper-testing' into blead
[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];
9cf78fa0 1031 ANYOF_CLASS_OR(or_with, cl);
653099ff
GS
1032 }
1033 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 1034 cl_anything(pRExC_state, cl);
653099ff 1035 }
9826f543 1036
c6b76537
KW
1037 if (ANYOF_NONBITMAP(or_with)) {
1038
1039 /* Use the added node's outside-the-bit-map match if there isn't a
1040 * conflict. If there is a conflict (both nodes match something
1041 * outside the bitmap, but what they match outside is not the same
1042 * pointer, and hence not easily compared until XXX we extend
1043 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
1044 * match everything outside the bitmap. If that stuff is all above
1045 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
1046 if (! ANYOF_NONBITMAP(cl)) {
1047 ARG_SET(cl, ARG(or_with));
1048 }
1049 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
1050
1051 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1052 cl_anything(pRExC_state, cl);
1053 }
1054 else {
1055 cl->flags |= ANYOF_UNICODE_ALL;
1056 }
c6b76537 1057 }
4c34a693 1058 }
0b9668ee
KW
1059
1060 /* Take the union */
1061 cl->flags |= or_with->flags;
1aa99e6b 1062 }
653099ff
GS
1063}
1064
a3621e74
YO
1065#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1066#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1067#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1068#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1069
3dab1dad
YO
1070
1071#ifdef DEBUGGING
07be1b83 1072/*
2b8b4781
NC
1073 dump_trie(trie,widecharmap,revcharmap)
1074 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1075 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1076
1077 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1078 The _interim_ variants are used for debugging the interim
1079 tables that are used to generate the final compressed
1080 representation which is what dump_trie expects.
1081
486ec47a 1082 Part of the reason for their existence is to provide a form
3dab1dad 1083 of documentation as to how the different representations function.
07be1b83
YO
1084
1085*/
3dab1dad
YO
1086
1087/*
3dab1dad
YO
1088 Dumps the final compressed table form of the trie to Perl_debug_log.
1089 Used for debugging make_trie().
1090*/
b9a59e08 1091
3dab1dad 1092STATIC void
2b8b4781
NC
1093S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1094 AV *revcharmap, U32 depth)
3dab1dad
YO
1095{
1096 U32 state;
ab3bbdeb 1097 SV *sv=sv_newmortal();
55eed653 1098 int colwidth= widecharmap ? 6 : 4;
2e64971a 1099 U16 word;
3dab1dad
YO
1100 GET_RE_DEBUG_FLAGS_DECL;
1101
7918f24d 1102 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1103
3dab1dad
YO
1104 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1105 (int)depth * 2 + 2,"",
1106 "Match","Base","Ofs" );
1107
1108 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1109 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1110 if ( tmp ) {
ab3bbdeb
YO
1111 PerlIO_printf( Perl_debug_log, "%*s",
1112 colwidth,
ddc5bc0f 1113 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1114 PL_colors[0], PL_colors[1],
1115 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1116 PERL_PV_ESCAPE_FIRSTCHAR
1117 )
1118 );
3dab1dad
YO
1119 }
1120 }
1121 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1122 (int)depth * 2 + 2,"");
1123
1124 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1125 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1126 PerlIO_printf( Perl_debug_log, "\n");
1127
1e2e3d02 1128 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1129 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1130
1131 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1132
1133 if ( trie->states[ state ].wordnum ) {
1134 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1135 } else {
1136 PerlIO_printf( Perl_debug_log, "%6s", "" );
1137 }
1138
1139 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1140
1141 if ( base ) {
1142 U32 ofs = 0;
1143
1144 while( ( base + ofs < trie->uniquecharcount ) ||
1145 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1146 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1147 ofs++;
1148
1149 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1150
1151 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1152 if ( ( base + ofs >= trie->uniquecharcount ) &&
1153 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1154 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1155 {
ab3bbdeb
YO
1156 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1157 colwidth,
3dab1dad
YO
1158 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1159 } else {
ab3bbdeb 1160 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1161 }
1162 }
1163
1164 PerlIO_printf( Perl_debug_log, "]");
1165
1166 }
1167 PerlIO_printf( Perl_debug_log, "\n" );
1168 }
2e64971a
DM
1169 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1170 for (word=1; word <= trie->wordcount; word++) {
1171 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1172 (int)word, (int)(trie->wordinfo[word].prev),
1173 (int)(trie->wordinfo[word].len));
1174 }
1175 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1176}
1177/*
3dab1dad
YO
1178 Dumps a fully constructed but uncompressed trie in list form.
1179 List tries normally only are used for construction when the number of
1180 possible chars (trie->uniquecharcount) is very high.
1181 Used for debugging make_trie().
1182*/
1183STATIC void
55eed653 1184S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1185 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1186 U32 depth)
3dab1dad
YO
1187{
1188 U32 state;
ab3bbdeb 1189 SV *sv=sv_newmortal();
55eed653 1190 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1191 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1192
1193 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1194
3dab1dad 1195 /* print out the table precompression. */
ab3bbdeb
YO
1196 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1197 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1198 "------:-----+-----------------\n" );
3dab1dad
YO
1199
1200 for( state=1 ; state < next_alloc ; state ++ ) {
1201 U16 charid;
1202
ab3bbdeb 1203 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1204 (int)depth * 2 + 2,"", (UV)state );
1205 if ( ! trie->states[ state ].wordnum ) {
1206 PerlIO_printf( Perl_debug_log, "%5s| ","");
1207 } else {
1208 PerlIO_printf( Perl_debug_log, "W%4x| ",
1209 trie->states[ state ].wordnum
1210 );
1211 }
1212 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1213 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1214 if ( tmp ) {
1215 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1216 colwidth,
ddc5bc0f 1217 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1218 PL_colors[0], PL_colors[1],
1219 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1220 PERL_PV_ESCAPE_FIRSTCHAR
1221 ) ,
1e2e3d02
YO
1222 TRIE_LIST_ITEM(state,charid).forid,
1223 (UV)TRIE_LIST_ITEM(state,charid).newstate
1224 );
1225 if (!(charid % 10))
664e119d
RGS
1226 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1227 (int)((depth * 2) + 14), "");
1e2e3d02 1228 }
ab3bbdeb
YO
1229 }
1230 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1231 }
1232}
1233
1234/*
3dab1dad
YO
1235 Dumps a fully constructed but uncompressed trie in table form.
1236 This is the normal DFA style state transition table, with a few
1237 twists to facilitate compression later.
1238 Used for debugging make_trie().
1239*/
1240STATIC void
55eed653 1241S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1242 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1243 U32 depth)
3dab1dad
YO
1244{
1245 U32 state;
1246 U16 charid;
ab3bbdeb 1247 SV *sv=sv_newmortal();
55eed653 1248 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1249 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1250
1251 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1252
1253 /*
1254 print out the table precompression so that we can do a visual check
1255 that they are identical.
1256 */
1257
1258 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1259
1260 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1261 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1262 if ( tmp ) {
ab3bbdeb
YO
1263 PerlIO_printf( Perl_debug_log, "%*s",
1264 colwidth,
ddc5bc0f 1265 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1266 PL_colors[0], PL_colors[1],
1267 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1268 PERL_PV_ESCAPE_FIRSTCHAR
1269 )
1270 );
3dab1dad
YO
1271 }
1272 }
1273
1274 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1275
1276 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1277 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1278 }
1279
1280 PerlIO_printf( Perl_debug_log, "\n" );
1281
1282 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1283
1284 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1285 (int)depth * 2 + 2,"",
1286 (UV)TRIE_NODENUM( state ) );
1287
1288 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1289 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1290 if (v)
1291 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1292 else
1293 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1294 }
1295 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1296 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1297 } else {
1298 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1299 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1300 }
1301 }
07be1b83 1302}
3dab1dad
YO
1303
1304#endif
1305
2e64971a 1306
786e8c11
YO
1307/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1308 startbranch: the first branch in the whole branch sequence
1309 first : start branch of sequence of branch-exact nodes.
1310 May be the same as startbranch
1311 last : Thing following the last branch.
1312 May be the same as tail.
1313 tail : item following the branch sequence
1314 count : words in the sequence
1315 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1316 depth : indent depth
3dab1dad 1317
786e8c11 1318Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1319
786e8c11
YO
1320A trie is an N'ary tree where the branches are determined by digital
1321decomposition of the key. IE, at the root node you look up the 1st character and
1322follow that branch repeat until you find the end of the branches. Nodes can be
1323marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1324
786e8c11 1325 /he|she|his|hers/
72f13be8 1326
786e8c11
YO
1327would convert into the following structure. Numbers represent states, letters
1328following numbers represent valid transitions on the letter from that state, if
1329the number is in square brackets it represents an accepting state, otherwise it
1330will be in parenthesis.
07be1b83 1331
786e8c11
YO
1332 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1333 | |
1334 | (2)
1335 | |
1336 (1) +-i->(6)-+-s->[7]
1337 |
1338 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1339
786e8c11
YO
1340 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1341
1342This shows that when matching against the string 'hers' we will begin at state 1
1343read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1344then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1345is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1346single traverse. We store a mapping from accepting to state to which word was
1347matched, and then when we have multiple possibilities we try to complete the
1348rest of the regex in the order in which they occured in the alternation.
1349
1350The only prior NFA like behaviour that would be changed by the TRIE support is
1351the silent ignoring of duplicate alternations which are of the form:
1352
1353 / (DUPE|DUPE) X? (?{ ... }) Y /x
1354
4b714af6 1355Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1356and without the optimisation. With the optimisations dupes will be silently
486ec47a 1357ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1358the following demonstrates:
1359
1360 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1361
1362which prints out 'word' three times, but
1363
1364 'words'=~/(word|word|word)(?{ print $1 })S/
1365
1366which doesnt print it out at all. This is due to other optimisations kicking in.
1367
1368Example of what happens on a structural level:
1369
486ec47a 1370The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1371
1372 1: CURLYM[1] {1,32767}(18)
1373 5: BRANCH(8)
1374 6: EXACT <ac>(16)
1375 8: BRANCH(11)
1376 9: EXACT <ad>(16)
1377 11: BRANCH(14)
1378 12: EXACT <ab>(16)
1379 16: SUCCEED(0)
1380 17: NOTHING(18)
1381 18: END(0)
1382
1383This would be optimizable with startbranch=5, first=5, last=16, tail=16
1384and should turn into:
1385
1386 1: CURLYM[1] {1,32767}(18)
1387 5: TRIE(16)
1388 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1389 <ac>
1390 <ad>
1391 <ab>
1392 16: SUCCEED(0)
1393 17: NOTHING(18)
1394 18: END(0)
1395
1396Cases where tail != last would be like /(?foo|bar)baz/:
1397
1398 1: BRANCH(4)
1399 2: EXACT <foo>(8)
1400 4: BRANCH(7)
1401 5: EXACT <bar>(8)
1402 7: TAIL(8)
1403 8: EXACT <baz>(10)
1404 10: END(0)
1405
1406which would be optimizable with startbranch=1, first=1, last=7, tail=8
1407and would end up looking like:
1408
1409 1: TRIE(8)
1410 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1411 <foo>
1412 <bar>
1413 7: TAIL(8)
1414 8: EXACT <baz>(10)
1415 10: END(0)
1416
1417 d = uvuni_to_utf8_flags(d, uv, 0);
1418
1419is the recommended Unicode-aware way of saying
1420
1421 *(d++) = uv;
1422*/
1423
fab2782b 1424#define TRIE_STORE_REVCHAR(val) \
786e8c11 1425 STMT_START { \
73031816 1426 if (UTF) { \
fab2782b 1427 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1428 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
fab2782b 1429 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
73031816
NC
1430 SvCUR_set(zlopp, kapow - flrbbbbb); \
1431 SvPOK_on(zlopp); \
1432 SvUTF8_on(zlopp); \
1433 av_push(revcharmap, zlopp); \
1434 } else { \
fab2782b 1435 char ooooff = (char)val; \
73031816
NC
1436 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1437 } \
1438 } STMT_END
786e8c11 1439
fab2782b
YO
1440#define TRIE_READ_CHAR STMT_START { \
1441 wordlen++; \
1442 if ( UTF ) { \
1443 /* if it is UTF then it is either already folded, or does not need folding */ \
1444 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1445 } \
1446 else if (folder == PL_fold_latin1) { \
1447 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1448 if ( foldlen > 0 ) { \
1449 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1450 foldlen -= len; \
1451 scan += len; \
1452 len = 0; \
1453 } else { \
1454 len = 1; \
1455 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1456 skiplen = UNISKIP(uvc); \
1457 foldlen -= skiplen; \
1458 scan = foldbuf + skiplen; \
1459 } \
1460 } else { \
1461 /* raw data, will be folded later if needed */ \
1462 uvc = (U32)*uc; \
1463 len = 1; \
1464 } \
786e8c11
YO
1465} STMT_END
1466
1467
1468
1469#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1470 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1471 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1472 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1473 } \
1474 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1475 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1476 TRIE_LIST_CUR( state )++; \
1477} STMT_END
07be1b83 1478
786e8c11
YO
1479#define TRIE_LIST_NEW(state) STMT_START { \
1480 Newxz( trie->states[ state ].trans.list, \
1481 4, reg_trie_trans_le ); \
1482 TRIE_LIST_CUR( state ) = 1; \
1483 TRIE_LIST_LEN( state ) = 4; \
1484} STMT_END
07be1b83 1485
786e8c11
YO
1486#define TRIE_HANDLE_WORD(state) STMT_START { \
1487 U16 dupe= trie->states[ state ].wordnum; \
1488 regnode * const noper_next = regnext( noper ); \
1489 \
786e8c11
YO
1490 DEBUG_r({ \
1491 /* store the word for dumping */ \
1492 SV* tmp; \
1493 if (OP(noper) != NOTHING) \
740cce10 1494 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1495 else \
740cce10 1496 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1497 av_push( trie_words, tmp ); \
786e8c11
YO
1498 }); \
1499 \
1500 curword++; \
2e64971a
DM
1501 trie->wordinfo[curword].prev = 0; \
1502 trie->wordinfo[curword].len = wordlen; \
1503 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1504 \
1505 if ( noper_next < tail ) { \
1506 if (!trie->jump) \
c944940b 1507 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1508 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1509 if (!jumper) \
1510 jumper = noper_next; \
1511 if (!nextbranch) \
1512 nextbranch= regnext(cur); \
1513 } \
1514 \
1515 if ( dupe ) { \
2e64971a
DM
1516 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1517 /* chain, so that when the bits of chain are later */\
1518 /* linked together, the dups appear in the chain */\
1519 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1520 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1521 } else { \
1522 /* we haven't inserted this word yet. */ \
1523 trie->states[ state ].wordnum = curword; \
1524 } \
1525} STMT_END
07be1b83 1526
3dab1dad 1527
786e8c11
YO
1528#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1529 ( ( base + charid >= ucharcount \
1530 && base + charid < ubound \
1531 && state == trie->trans[ base - ucharcount + charid ].check \
1532 && trie->trans[ base - ucharcount + charid ].next ) \
1533 ? trie->trans[ base - ucharcount + charid ].next \
1534 : ( state==1 ? special : 0 ) \
1535 )
3dab1dad 1536
786e8c11
YO
1537#define MADE_TRIE 1
1538#define MADE_JUMP_TRIE 2
1539#define MADE_EXACT_TRIE 4
3dab1dad 1540
a3621e74 1541STATIC I32
786e8c11 1542S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1543{
27da23d5 1544 dVAR;
a3621e74
YO
1545 /* first pass, loop through and scan words */
1546 reg_trie_data *trie;
55eed653 1547 HV *widecharmap = NULL;
2b8b4781 1548 AV *revcharmap = newAV();
a3621e74 1549 regnode *cur;
9f7f3913 1550 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1551 STRLEN len = 0;
1552 UV uvc = 0;
1553 U16 curword = 0;
1554 U32 next_alloc = 0;
786e8c11
YO
1555 regnode *jumper = NULL;
1556 regnode *nextbranch = NULL;
7f69552c 1557 regnode *convert = NULL;
2e64971a 1558 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1559 /* we just use folder as a flag in utf8 */
1e696034 1560 const U8 * folder = NULL;
a3621e74 1561
2b8b4781
NC
1562#ifdef DEBUGGING
1563 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1564 AV *trie_words = NULL;
1565 /* along with revcharmap, this only used during construction but both are
1566 * useful during debugging so we store them in the struct when debugging.
8e11feef 1567 */
2b8b4781
NC
1568#else
1569 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1570 STRLEN trie_charcount=0;
3dab1dad 1571#endif
2b8b4781 1572 SV *re_trie_maxbuff;
a3621e74 1573 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1574
1575 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1576#ifndef DEBUGGING
1577 PERL_UNUSED_ARG(depth);
1578#endif
a3621e74 1579
1e696034 1580 switch (flags) {
79a81a6e 1581 case EXACT: break;
2f7f8cb1 1582 case EXACTFA:
fab2782b
YO
1583 case EXACTFU_SS:
1584 case EXACTFU_TRICKYFOLD:
1e696034
KW
1585 case EXACTFU: folder = PL_fold_latin1; break;
1586 case EXACTF: folder = PL_fold; break;
1587 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1588 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1589 }
1590
c944940b 1591 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1592 trie->refcount = 1;
3dab1dad 1593 trie->startstate = 1;
786e8c11 1594 trie->wordcount = word_count;
f8fc2ecf 1595 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1596 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1597 if (flags == EXACT)
c944940b 1598 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1599 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1600 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1601
a3621e74 1602 DEBUG_r({
2b8b4781 1603 trie_words = newAV();
a3621e74 1604 });
a3621e74 1605
0111c4fd 1606 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1607 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1608 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1609 }
df826430 1610 DEBUG_TRIE_COMPILE_r({
3dab1dad 1611 PerlIO_printf( Perl_debug_log,
786e8c11 1612 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1613 (int)depth * 2 + 2, "",
1614 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1615 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1616 (int)depth);
3dab1dad 1617 });
7f69552c
YO
1618
1619 /* Find the node we are going to overwrite */
1620 if ( first == startbranch && OP( last ) != BRANCH ) {
1621 /* whole branch chain */
1622 convert = first;
1623 } else {
1624 /* branch sub-chain */
1625 convert = NEXTOPER( first );
1626 }
1627
a3621e74
YO
1628 /* -- First loop and Setup --
1629
1630 We first traverse the branches and scan each word to determine if it
1631 contains widechars, and how many unique chars there are, this is
1632 important as we have to build a table with at least as many columns as we
1633 have unique chars.
1634
1635 We use an array of integers to represent the character codes 0..255
38a44b82 1636 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1637 native representation of the character value as the key and IV's for the
1638 coded index.
1639
1640 *TODO* If we keep track of how many times each character is used we can
1641 remap the columns so that the table compression later on is more
3b753521 1642 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1643 middle and the least common are on the outside. IMO this would be better
1644 than a most to least common mapping as theres a decent chance the most
1645 common letter will share a node with the least common, meaning the node
486ec47a 1646 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1647 case is when we have the least common nodes twice.
1648
1649 */
1650
a3621e74 1651 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1652 regnode *noper = NEXTOPER( cur );
e1ec3a88 1653 const U8 *uc = (U8*)STRING( noper );
df826430 1654 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
1655 STRLEN foldlen = 0;
1656 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1657 STRLEN skiplen = 0;
2af232bd 1658 const U8 *scan = (U8*)NULL;
07be1b83 1659 U32 wordlen = 0; /* required init */
02daf0ab
YO
1660 STRLEN chars = 0;
1661 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1662
3dab1dad 1663 if (OP(noper) == NOTHING) {
df826430
YO
1664 regnode *noper_next= regnext(noper);
1665 if (noper_next != tail && OP(noper_next) == flags) {
1666 noper = noper_next;
1667 uc= (U8*)STRING(noper);
1668 e= uc + STR_LEN(noper);
1669 trie->minlen= STR_LEN(noper);
1670 } else {
1671 trie->minlen= 0;
1672 continue;
1673 }
3dab1dad 1674 }
df826430 1675
fab2782b 1676 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1677 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1678 regardless of encoding */
fab2782b
YO
1679 if (OP( noper ) == EXACTFU_SS) {
1680 /* false positives are ok, so just set this */
1681 TRIE_BITMAP_SET(trie,0xDF);
1682 }
1683 }
a3621e74 1684 for ( ; uc < e ; uc += len ) {
3dab1dad 1685 TRIE_CHARCOUNT(trie)++;
a3621e74 1686 TRIE_READ_CHAR;
3dab1dad 1687 chars++;
a3621e74 1688 if ( uvc < 256 ) {
fab2782b
YO
1689 if ( folder ) {
1690 U8 folded= folder[ (U8) uvc ];
1691 if ( !trie->charmap[ folded ] ) {
1692 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1693 TRIE_STORE_REVCHAR( folded );
1694 }
1695 }
a3621e74
YO
1696 if ( !trie->charmap[ uvc ] ) {
1697 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1698 TRIE_STORE_REVCHAR( uvc );
a3621e74 1699 }
02daf0ab 1700 if ( set_bit ) {
62012aee
KW
1701 /* store the codepoint in the bitmap, and its folded
1702 * equivalent. */
fab2782b 1703 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1704
1705 /* store the folded codepoint */
fab2782b 1706 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1707
1708 if ( !UTF ) {
1709 /* store first byte of utf8 representation of
acdf4139
KW
1710 variant codepoints */
1711 if (! UNI_IS_INVARIANT(uvc)) {
1712 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1713 }
1714 }
02daf0ab
YO
1715 set_bit = 0; /* We've done our bit :-) */
1716 }
a3621e74
YO
1717 } else {
1718 SV** svpp;
55eed653
NC
1719 if ( !widecharmap )
1720 widecharmap = newHV();
a3621e74 1721
55eed653 1722 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1723
1724 if ( !svpp )
e4584336 1725 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1726
1727 if ( !SvTRUE( *svpp ) ) {
1728 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1729 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1730 }
1731 }
1732 }
3dab1dad 1733 if( cur == first ) {
fab2782b
YO
1734 trie->minlen = chars;
1735 trie->maxlen = chars;
3dab1dad 1736 } else if (chars < trie->minlen) {
fab2782b 1737 trie->minlen = chars;
3dab1dad 1738 } else if (chars > trie->maxlen) {
fab2782b
YO
1739 trie->maxlen = chars;
1740 }
1741 if (OP( noper ) == EXACTFU_SS) {
1742 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1743 if (trie->minlen > 1)
1744 trie->minlen= 1;
1745 }
1746 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1747 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1748 * - We assume that any such sequence might match a 2 byte string */
1749 if (trie->minlen > 2 )
1750 trie->minlen= 2;
3dab1dad
YO
1751 }
1752
a3621e74
YO
1753 } /* end first pass */
1754 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1755 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1756 (int)depth * 2 + 2,"",
55eed653 1757 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1758 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1759 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1760 );
a3621e74
YO
1761
1762 /*
1763 We now know what we are dealing with in terms of unique chars and
1764 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1765 representation using a flat table will take. If it's over a reasonable
1766 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1767 conservative but potentially much slower representation using an array
1768 of lists.
1769
1770 At the end we convert both representations into the same compressed
1771 form that will be used in regexec.c for matching with. The latter
1772 is a form that cannot be used to construct with but has memory
1773 properties similar to the list form and access properties similar
1774 to the table form making it both suitable for fast searches and
1775 small enough that its feasable to store for the duration of a program.
1776
1777 See the comment in the code where the compressed table is produced
1778 inplace from the flat tabe representation for an explanation of how
1779 the compression works.
1780
1781 */
1782
1783
2e64971a
DM
1784 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1785 prev_states[1] = 0;
1786
3dab1dad 1787 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1788 /*
1789 Second Pass -- Array Of Lists Representation
1790
1791 Each state will be represented by a list of charid:state records
1792 (reg_trie_trans_le) the first such element holds the CUR and LEN
1793 points of the allocated array. (See defines above).
1794
1795 We build the initial structure using the lists, and then convert
1796 it into the compressed table form which allows faster lookups
1797 (but cant be modified once converted).
a3621e74
YO
1798 */
1799
a3621e74
YO
1800 STRLEN transcount = 1;
1801
1e2e3d02
YO
1802 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1803 "%*sCompiling trie using list compiler\n",
1804 (int)depth * 2 + 2, ""));
686b73d4 1805
c944940b
JH
1806 trie->states = (reg_trie_state *)
1807 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1808 sizeof(reg_trie_state) );
a3621e74
YO
1809 TRIE_LIST_NEW(1);
1810 next_alloc = 2;
1811
1812 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1813
df826430 1814 regnode *noper = NEXTOPER( cur );
c445ea15 1815 U8 *uc = (U8*)STRING( noper );
df826430 1816 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1817 U32 state = 1; /* required init */
1818 U16 charid = 0; /* sanity init */
1819 U8 *scan = (U8*)NULL; /* sanity init */
1820 STRLEN foldlen = 0; /* required init */
07be1b83 1821 U32 wordlen = 0; /* required init */
c445ea15 1822 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
fab2782b 1823 STRLEN skiplen = 0;
c445ea15 1824
df826430
YO
1825 if (OP(noper) == NOTHING) {
1826 regnode *noper_next= regnext(noper);
1827 if (noper_next != tail && OP(noper_next) == flags) {
1828 noper = noper_next;
1829 uc= (U8*)STRING(noper);
1830 e= uc + STR_LEN(noper);
1831 }
1832 }
1833
3dab1dad 1834 if (OP(noper) != NOTHING) {
786e8c11 1835 for ( ; uc < e ; uc += len ) {
c445ea15 1836
786e8c11 1837 TRIE_READ_CHAR;
c445ea15 1838
786e8c11
YO
1839 if ( uvc < 256 ) {
1840 charid = trie->charmap[ uvc ];
c445ea15 1841 } else {
55eed653 1842 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1843 if ( !svpp ) {
1844 charid = 0;
1845 } else {
1846 charid=(U16)SvIV( *svpp );
1847 }
c445ea15 1848 }
786e8c11
YO
1849 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1850 if ( charid ) {
a3621e74 1851
786e8c11
YO
1852 U16 check;
1853 U32 newstate = 0;
a3621e74 1854
786e8c11
YO
1855 charid--;
1856 if ( !trie->states[ state ].trans.list ) {
1857 TRIE_LIST_NEW( state );
c445ea15 1858 }
786e8c11
YO
1859 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1860 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1861 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1862 break;
1863 }
1864 }
1865 if ( ! newstate ) {
1866 newstate = next_alloc++;
2e64971a 1867 prev_states[newstate] = state;
786e8c11
YO
1868 TRIE_LIST_PUSH( state, charid, newstate );
1869 transcount++;
1870 }
1871 state = newstate;
1872 } else {
1873 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1874 }
a28509cc 1875 }
c445ea15 1876 }
3dab1dad 1877 TRIE_HANDLE_WORD(state);
a3621e74
YO
1878
1879 } /* end second pass */
1880
1e2e3d02
YO
1881 /* next alloc is the NEXT state to be allocated */
1882 trie->statecount = next_alloc;
c944940b
JH
1883 trie->states = (reg_trie_state *)
1884 PerlMemShared_realloc( trie->states,
1885 next_alloc
1886 * sizeof(reg_trie_state) );
a3621e74 1887
3dab1dad 1888 /* and now dump it out before we compress it */
2b8b4781
NC
1889 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1890 revcharmap, next_alloc,
1891 depth+1)
1e2e3d02 1892 );
a3621e74 1893
c944940b
JH
1894 trie->trans = (reg_trie_trans *)
1895 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1896 {
1897 U32 state;
a3621e74
YO
1898 U32 tp = 0;
1899 U32 zp = 0;
1900
1901
1902 for( state=1 ; state < next_alloc ; state ++ ) {
1903 U32 base=0;
1904
1905 /*
1906 DEBUG_TRIE_COMPILE_MORE_r(
1907 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1908 );
1909 */
1910
1911 if (trie->states[state].trans.list) {
1912 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1913 U16 maxid=minid;
a28509cc 1914 U16 idx;
a3621e74
YO
1915
1916 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1917 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1918 if ( forid < minid ) {
1919 minid=forid;
1920 } else if ( forid > maxid ) {
1921 maxid=forid;
1922 }
a3621e74
YO
1923 }
1924 if ( transcount < tp + maxid - minid + 1) {
1925 transcount *= 2;
c944940b
JH
1926 trie->trans = (reg_trie_trans *)
1927 PerlMemShared_realloc( trie->trans,
446bd890
NC
1928 transcount
1929 * sizeof(reg_trie_trans) );
a3621e74
YO
1930 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1931 }
1932 base = trie->uniquecharcount + tp - minid;
1933 if ( maxid == minid ) {
1934 U32 set = 0;
1935 for ( ; zp < tp ; zp++ ) {
1936 if ( ! trie->trans[ zp ].next ) {
1937 base = trie->uniquecharcount + zp - minid;
1938 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1939 trie->trans[ zp ].check = state;
1940 set = 1;
1941 break;
1942 }
1943 }
1944 if ( !set ) {
1945 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1946 trie->trans[ tp ].check = state;
1947 tp++;
1948 zp = tp;
1949 }
1950 } else {
1951 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1952 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1953 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1954 trie->trans[ tid ].check = state;
1955 }
1956 tp += ( maxid - minid + 1 );
1957 }
1958 Safefree(trie->states[ state ].trans.list);
1959 }
1960 /*
1961 DEBUG_TRIE_COMPILE_MORE_r(
1962 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1963 );
1964 */
1965 trie->states[ state ].trans.base=base;
1966 }
cc601c31 1967 trie->lasttrans = tp + 1;
a3621e74
YO
1968 }
1969 } else {
1970 /*
1971 Second Pass -- Flat Table Representation.
1972
1973 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1974 We know that we will need Charcount+1 trans at most to store the data
1975 (one row per char at worst case) So we preallocate both structures
1976 assuming worst case.
1977
1978 We then construct the trie using only the .next slots of the entry
1979 structs.
1980
3b753521 1981 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1982 make compression both faster and easier by keeping track of how many non
1983 zero fields are in the node.
1984
1985 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1986 transition.
1987
1988 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1989 number representing the first entry of the node, and state as a
1990 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1991 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1992 are 2 entrys per node. eg:
1993
1994 A B A B
1995 1. 2 4 1. 3 7
1996 2. 0 3 3. 0 5
1997 3. 0 0 5. 0 0
1998 4. 0 0 7. 0 0
1999
2000 The table is internally in the right hand, idx form. However as we also
2001 have to deal with the states array which is indexed by nodenum we have to
2002 use TRIE_NODENUM() to convert.
2003
2004 */
1e2e3d02
YO
2005 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2006 "%*sCompiling trie using table compiler\n",
2007 (int)depth * 2 + 2, ""));
3dab1dad 2008
c944940b
JH
2009 trie->trans = (reg_trie_trans *)
2010 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2011 * trie->uniquecharcount + 1,
2012 sizeof(reg_trie_trans) );
2013 trie->states = (reg_trie_state *)
2014 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2015 sizeof(reg_trie_state) );
a3621e74
YO
2016 next_alloc = trie->uniquecharcount + 1;
2017
3dab1dad 2018
a3621e74
YO
2019 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2020
df826430 2021 regnode *noper = NEXTOPER( cur );
a28509cc 2022 const U8 *uc = (U8*)STRING( noper );
df826430 2023 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
2024
2025 U32 state = 1; /* required init */
2026
2027 U16 charid = 0; /* sanity init */
2028 U32 accept_state = 0; /* sanity init */
2029 U8 *scan = (U8*)NULL; /* sanity init */
2030
2031 STRLEN foldlen = 0; /* required init */
07be1b83 2032 U32 wordlen = 0; /* required init */
fab2782b 2033 STRLEN skiplen = 0;
a3621e74
YO
2034 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2035
df826430
YO
2036 if (OP(noper) == NOTHING) {
2037 regnode *noper_next= regnext(noper);
2038 if (noper_next != tail && OP(noper_next) == flags) {
2039 noper = noper_next;
2040 uc= (U8*)STRING(noper);
2041 e= uc + STR_LEN(noper);
2042 }
2043 }
fab2782b 2044
3dab1dad 2045 if ( OP(noper) != NOTHING ) {
786e8c11 2046 for ( ; uc < e ; uc += len ) {
a3621e74 2047
786e8c11 2048 TRIE_READ_CHAR;
a3621e74 2049
786e8c11
YO
2050 if ( uvc < 256 ) {
2051 charid = trie->charmap[ uvc ];
2052 } else {
55eed653 2053 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2054 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2055 }
786e8c11
YO
2056 if ( charid ) {
2057 charid--;
2058 if ( !trie->trans[ state + charid ].next ) {
2059 trie->trans[ state + charid ].next = next_alloc;
2060 trie->trans[ state ].check++;
2e64971a
DM
2061 prev_states[TRIE_NODENUM(next_alloc)]
2062 = TRIE_NODENUM(state);
786e8c11
YO
2063 next_alloc += trie->uniquecharcount;
2064 }
2065 state = trie->trans[ state + charid ].next;
2066 } else {
2067 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2068 }
2069 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2070 }
a3621e74 2071 }
3dab1dad
YO
2072 accept_state = TRIE_NODENUM( state );
2073 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2074
2075 } /* end second pass */
2076
3dab1dad 2077 /* and now dump it out before we compress it */
2b8b4781
NC
2078 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2079 revcharmap,
2080 next_alloc, depth+1));
a3621e74 2081
a3621e74
YO
2082 {
2083 /*
2084 * Inplace compress the table.*
2085
2086 For sparse data sets the table constructed by the trie algorithm will
2087 be mostly 0/FAIL transitions or to put it another way mostly empty.
2088 (Note that leaf nodes will not contain any transitions.)
2089
2090 This algorithm compresses the tables by eliminating most such
2091 transitions, at the cost of a modest bit of extra work during lookup:
2092
2093 - Each states[] entry contains a .base field which indicates the
2094 index in the state[] array wheres its transition data is stored.
2095
3b753521 2096 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2097
2098 - If .base is nonzero then charid is added to it to find an entry in
2099 the trans array.
2100
2101 -If trans[states[state].base+charid].check!=state then the
2102 transition is taken to be a 0/Fail transition. Thus if there are fail
2103 transitions at the front of the node then the .base offset will point
2104 somewhere inside the previous nodes data (or maybe even into a node
2105 even earlier), but the .check field determines if the transition is
2106 valid.
2107
786e8c11 2108 XXX - wrong maybe?
a3621e74 2109 The following process inplace converts the table to the compressed
3b753521 2110 table: We first do not compress the root node 1,and mark all its
a3621e74 2111 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2112 allows us to do a DFA construction from the compressed table later,
2113 and ensures that any .base pointers we calculate later are greater
2114 than 0.
a3621e74
YO
2115
2116 - We set 'pos' to indicate the first entry of the second node.
2117
2118 - We then iterate over the columns of the node, finding the first and
2119 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2120 and set the .check pointers accordingly, and advance pos
2121 appropriately and repreat for the next node. Note that when we copy
2122 the next pointers we have to convert them from the original
2123 NODEIDX form to NODENUM form as the former is not valid post
2124 compression.
2125
2126 - If a node has no transitions used we mark its base as 0 and do not
2127 advance the pos pointer.
2128
2129 - If a node only has one transition we use a second pointer into the
2130 structure to fill in allocated fail transitions from other states.
2131 This pointer is independent of the main pointer and scans forward
2132 looking for null transitions that are allocated to a state. When it
2133 finds one it writes the single transition into the "hole". If the
786e8c11 2134 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2135
2136 - Once compressed we can Renew/realloc the structures to release the
2137 excess space.
2138
2139 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2140 specifically Fig 3.47 and the associated pseudocode.
2141
2142 demq
2143 */
a3b680e6 2144 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2145 U32 state, charid;
a3621e74 2146 U32 pos = 0, zp=0;
1e2e3d02 2147 trie->statecount = laststate;
a3621e74
YO
2148
2149 for ( state = 1 ; state < laststate ; state++ ) {
2150 U8 flag = 0;
a28509cc
AL
2151 const U32 stateidx = TRIE_NODEIDX( state );
2152 const U32 o_used = trie->trans[ stateidx ].check;
2153 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2154 trie->trans[ stateidx ].check = 0;
2155
2156 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2157 if ( flag || trie->trans[ stateidx + charid ].next ) {
2158 if ( trie->trans[ stateidx + charid ].next ) {
2159 if (o_used == 1) {
2160 for ( ; zp < pos ; zp++ ) {
2161 if ( ! trie->trans[ zp ].next ) {
2162 break;
2163 }
2164 }
2165 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2166 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2167 trie->trans[ zp ].check = state;
2168 if ( ++zp > pos ) pos = zp;
2169 break;
2170 }
2171 used--;
2172 }
2173 if ( !flag ) {
2174 flag = 1;
2175 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2176 }
2177 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2178 trie->trans[ pos ].check = state;
2179 pos++;
2180 }
2181 }
2182 }
cc601c31 2183 trie->lasttrans = pos + 1;
c944940b
JH
2184 trie->states = (reg_trie_state *)
2185 PerlMemShared_realloc( trie->states, laststate
2186 * sizeof(reg_trie_state) );
a3621e74 2187 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2188 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2189 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2190 (int)depth * 2 + 2,"",
2191 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2192 (IV)next_alloc,
2193 (IV)pos,
a3621e74
YO
2194 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2195 );
2196
2197 } /* end table compress */
2198 }
1e2e3d02
YO
2199 DEBUG_TRIE_COMPILE_MORE_r(
2200 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2201 (int)depth * 2 + 2, "",
2202 (UV)trie->statecount,
2203 (UV)trie->lasttrans)
2204 );
cc601c31 2205 /* resize the trans array to remove unused space */
c944940b
JH
2206 trie->trans = (reg_trie_trans *)
2207 PerlMemShared_realloc( trie->trans, trie->lasttrans
2208 * sizeof(reg_trie_trans) );
a3621e74 2209
3b753521 2210 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2211 U8 nodetype =(U8)(flags & 0xFF);
2212 char *str=NULL;
786e8c11 2213
07be1b83 2214#ifdef DEBUGGING
e62cc96a 2215 regnode *optimize = NULL;
7122b237
YO
2216#ifdef RE_TRACK_PATTERN_OFFSETS
2217
b57a0404
JH
2218 U32 mjd_offset = 0;
2219 U32 mjd_nodelen = 0;
7122b237
YO
2220#endif /* RE_TRACK_PATTERN_OFFSETS */
2221#endif /* DEBUGGING */
a3621e74 2222 /*
3dab1dad
YO
2223 This means we convert either the first branch or the first Exact,
2224 depending on whether the thing following (in 'last') is a branch
2225 or not and whther first is the startbranch (ie is it a sub part of
2226 the alternation or is it the whole thing.)
3b753521 2227 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2228 the whole branch sequence, including the first.
a3621e74 2229 */
3dab1dad 2230 /* Find the node we are going to overwrite */
7f69552c 2231 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2232 /* branch sub-chain */
3dab1dad 2233 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2234#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2235 DEBUG_r({
2236 mjd_offset= Node_Offset((convert));
2237 mjd_nodelen= Node_Length((convert));
2238 });
7122b237 2239#endif
7f69552c 2240 /* whole branch chain */
7122b237
YO
2241 }
2242#ifdef RE_TRACK_PATTERN_OFFSETS
2243 else {
7f69552c
YO
2244 DEBUG_r({
2245 const regnode *nop = NEXTOPER( convert );
2246 mjd_offset= Node_Offset((nop));
2247 mjd_nodelen= Node_Length((nop));
2248 });
07be1b83
YO
2249 }
2250 DEBUG_OPTIMISE_r(
2251 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2252 (int)depth * 2 + 2, "",
786e8c11 2253 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2254 );
7122b237 2255#endif
3dab1dad
YO
2256 /* But first we check to see if there is a common prefix we can
2257 split out as an EXACT and put in front of the TRIE node. */
2258 trie->startstate= 1;
55eed653 2259 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2260 U32 state;
1e2e3d02 2261 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2262 U32 ofs = 0;
8e11feef
RGS
2263 I32 idx = -1;
2264 U32 count = 0;
2265 const U32 base = trie->states[ state ].trans.base;
a3621e74 2266
3dab1dad 2267 if ( trie->states[state].wordnum )
8e11feef 2268 count = 1;
a3621e74 2269
8e11feef 2270 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2271 if ( ( base + ofs >= trie->uniquecharcount ) &&
2272 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2273 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2274 {
3dab1dad 2275 if ( ++count > 1 ) {
2b8b4781 2276 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2277 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2278 if ( state == 1 ) break;
3dab1dad
YO
2279 if ( count == 2 ) {
2280 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2281 DEBUG_OPTIMISE_r(
8e11feef
RGS
2282 PerlIO_printf(Perl_debug_log,
2283 "%*sNew Start State=%"UVuf" Class: [",
2284 (int)depth * 2 + 2, "",
786e8c11 2285 (UV)state));
be8e71aa 2286 if (idx >= 0) {
2b8b4781 2287 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2288 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2289
3dab1dad 2290 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2291 if ( folder )
2292 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2293 DEBUG_OPTIMISE_r(
f1f66076 2294 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2295 );
8e11feef
RGS
2296 }
2297 }
2298 TRIE_BITMAP_SET(trie,*ch);
2299 if ( folder )
2300 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2301 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2302 }
2303 idx = ofs;
2304 }
3dab1dad
YO
2305 }
2306 if ( count == 1 ) {
2b8b4781 2307 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2308 STRLEN len;
2309 char *ch = SvPV( *tmp, len );
de734bd5
A
2310 DEBUG_OPTIMISE_r({
2311 SV *sv=sv_newmortal();
8e11feef
RGS
2312 PerlIO_printf( Perl_debug_log,
2313 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2314 (int)depth * 2 + 2, "",
de734bd5
A
2315 (UV)state, (UV)idx,
2316 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2317 PL_colors[0], PL_colors[1],
2318 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2319 PERL_PV_ESCAPE_FIRSTCHAR
2320 )
2321 );
2322 });
3dab1dad
YO
2323 if ( state==1 ) {
2324 OP( convert ) = nodetype;
2325 str=STRING(convert);
2326 STR_LEN(convert)=0;
2327 }
c490c714
YO
2328 STR_LEN(convert) += len;
2329 while (len--)
de734bd5 2330 *str++ = *ch++;
8e11feef 2331 } else {
f9049ba1 2332#ifdef DEBUGGING
8e11feef
RGS
2333 if (state>1)
2334 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2335#endif
8e11feef
RGS
2336 break;
2337 }
2338 }
2e64971a 2339 trie->prefixlen = (state-1);
3dab1dad 2340 if (str) {
8e11feef 2341 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2342 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2343 trie->startstate = state;
07be1b83
YO
2344 trie->minlen -= (state - 1);
2345 trie->maxlen -= (state - 1);
33809eae
JH
2346#ifdef DEBUGGING
2347 /* At least the UNICOS C compiler choked on this
2348 * being argument to DEBUG_r(), so let's just have
2349 * it right here. */
2350 if (
2351#ifdef PERL_EXT_RE_BUILD
2352 1
2353#else
2354 DEBUG_r_TEST
2355#endif
2356 ) {
2357 regnode *fix = convert;
2358 U32 word = trie->wordcount;
2359 mjd_nodelen++;
2360 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2361 while( ++fix < n ) {
2362 Set_Node_Offset_Length(fix, 0, 0);
2363 }
2364 while (word--) {
2365 SV ** const tmp = av_fetch( trie_words, word, 0 );
2366 if (tmp) {
2367 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2368 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2369 else
2370 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2371 }
2372 }
2373 }
2374#endif
8e11feef
RGS
2375 if (trie->maxlen) {
2376 convert = n;
2377 } else {
3dab1dad 2378 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2379 DEBUG_r(optimize= n);
3dab1dad
YO
2380 }
2381 }
2382 }
a5ca303d
YO
2383 if (!jumper)
2384 jumper = last;
3dab1dad 2385 if ( trie->maxlen ) {
8e11feef
RGS
2386 NEXT_OFF( convert ) = (U16)(tail - convert);
2387 ARG_SET( convert, data_slot );
786e8c11
YO
2388 /* Store the offset to the first unabsorbed branch in
2389 jump[0], which is otherwise unused by the jump logic.
2390 We use this when dumping a trie and during optimisation. */
2391 if (trie->jump)
7f69552c 2392 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2393
6c48061a
YO
2394 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2395 * and there is a bitmap
2396 * and the first "jump target" node we found leaves enough room
2397 * then convert the TRIE node into a TRIEC node, with the bitmap
2398 * embedded inline in the opcode - this is hypothetically faster.
2399 */
2400 if ( !trie->states[trie->startstate].wordnum
2401 && trie->bitmap
2402 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2403 {
2404 OP( convert ) = TRIEC;
2405 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2406 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2407 trie->bitmap= NULL;
2408 } else
2409 OP( convert ) = TRIE;
a3621e74 2410
3dab1dad
YO
2411 /* store the type in the flags */
2412 convert->flags = nodetype;
a5ca303d
YO
2413 DEBUG_r({
2414 optimize = convert
2415 + NODE_STEP_REGNODE
2416 + regarglen[ OP( convert ) ];
2417 });
2418 /* XXX We really should free up the resource in trie now,
2419 as we won't use them - (which resources?) dmq */
3dab1dad 2420 }
a3621e74 2421 /* needed for dumping*/
e62cc96a 2422 DEBUG_r(if (optimize) {
07be1b83 2423 regnode *opt = convert;
bcdf7404 2424
e62cc96a 2425 while ( ++opt < optimize) {
07be1b83
YO
2426 Set_Node_Offset_Length(opt,0,0);
2427 }
786e8c11
YO
2428 /*
2429 Try to clean up some of the debris left after the
2430 optimisation.
a3621e74 2431 */
786e8c11 2432 while( optimize < jumper ) {
07be1b83 2433 mjd_nodelen += Node_Length((optimize));
a3621e74 2434 OP( optimize ) = OPTIMIZED;
07be1b83 2435 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2436 optimize++;
2437 }
07be1b83 2438 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2439 });
2440 } /* end node insert */
2e64971a
DM
2441
2442 /* Finish populating the prev field of the wordinfo array. Walk back
2443 * from each accept state until we find another accept state, and if
2444 * so, point the first word's .prev field at the second word. If the
2445 * second already has a .prev field set, stop now. This will be the
2446 * case either if we've already processed that word's accept state,
3b753521
FN
2447 * or that state had multiple words, and the overspill words were
2448 * already linked up earlier.
2e64971a
DM
2449 */
2450 {
2451 U16 word;
2452 U32 state;
2453 U16 prev;
2454
2455 for (word=1; word <= trie->wordcount; word++) {
2456 prev = 0;
2457 if (trie->wordinfo[word].prev)
2458 continue;
2459 state = trie->wordinfo[word].accept;
2460 while (state) {
2461 state = prev_states[state];
2462 if (!state)
2463 break;
2464 prev = trie->states[state].wordnum;
2465 if (prev)
2466 break;
2467 }
2468 trie->wordinfo[word].prev = prev;
2469 }
2470 Safefree(prev_states);
2471 }
2472
2473
2474 /* and now dump out the compressed format */
2475 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2476
55eed653 2477 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2478#ifdef DEBUGGING
2479 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2480 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2481#else
03e70be4 2482 SvREFCNT_dec_NN(revcharmap);
07be1b83 2483#endif
786e8c11
YO
2484 return trie->jump
2485 ? MADE_JUMP_TRIE
2486 : trie->startstate>1
2487 ? MADE_EXACT_TRIE
2488 : MADE_TRIE;
2489}
2490
2491STATIC void
2492S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2493{
3b753521 2494/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2495
2496 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2497 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2498 ISBN 0-201-10088-6
2499
2500 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2501 suffix of the current state's 'word' that is also a proper prefix of another word in our
2502 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2503 the DFA not to have to restart after its tried and failed a word at a given point, it
2504 simply continues as though it had been matching the other word in the first place.
2505 Consider
2506 'abcdgu'=~/abcdefg|cdgu/
2507 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2508 fail, which would bring us to the state representing 'd' in the second word where we would
2509 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2510 */
2511 /* add a fail transition */
3251b653
NC
2512 const U32 trie_offset = ARG(source);
2513 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2514 U32 *q;
2515 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2516 const U32 numstates = trie->statecount;
786e8c11
YO
2517 const U32 ubound = trie->lasttrans + ucharcount;
2518 U32 q_read = 0;
2519 U32 q_write = 0;
2520 U32 charid;
2521 U32 base = trie->states[ 1 ].trans.base;
2522 U32 *fail;
2523 reg_ac_data *aho;
2524 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2525 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2526
2527 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2528#ifndef DEBUGGING
2529 PERL_UNUSED_ARG(depth);
2530#endif
2531
2532
2533 ARG_SET( stclass, data_slot );
c944940b 2534 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2535 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2536 aho->trie=trie_offset;
446bd890
NC
2537 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2538 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2539 Newxz( q, numstates, U32);
c944940b 2540 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2541 aho->refcount = 1;
2542 fail = aho->fail;
2543 /* initialize fail[0..1] to be 1 so that we always have
2544 a valid final fail state */
2545 fail[ 0 ] = fail[ 1 ] = 1;
2546
2547 for ( charid = 0; charid < ucharcount ; charid++ ) {
2548 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2549 if ( newstate ) {
2550 q[ q_write ] = newstate;
2551 /* set to point at the root */
2552 fail[ q[ q_write++ ] ]=1;
2553 }
2554 }
2555 while ( q_read < q_write) {
2556 const U32 cur = q[ q_read++ % numstates ];
2557 base = trie->states[ cur ].trans.base;
2558
2559 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2560 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2561 if (ch_state) {
2562 U32 fail_state = cur;
2563 U32 fail_base;
2564 do {
2565 fail_state = fail[ fail_state ];
2566 fail_base = aho->states[ fail_state ].trans.base;
2567 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2568
2569 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2570 fail[ ch_state ] = fail_state;
2571 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2572 {
2573 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2574 }
2575 q[ q_write++ % numstates] = ch_state;
2576 }
2577 }
2578 }
2579 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2580 when we fail in state 1, this allows us to use the
2581 charclass scan to find a valid start char. This is based on the principle
2582 that theres a good chance the string being searched contains lots of stuff
2583 that cant be a start char.
2584 */
2585 fail[ 0 ] = fail[ 1 ] = 0;
2586 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2587 PerlIO_printf(Perl_debug_log,
2588 "%*sStclass Failtable (%"UVuf" states): 0",
2589 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2590 );
786e8c11
YO
2591 for( q_read=1; q_read<numstates; q_read++ ) {
2592 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2593 }
2594 PerlIO_printf(Perl_debug_log, "\n");
2595 });
2596 Safefree(q);
2597 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2598}
2599
786e8c11 2600
a3621e74 2601/*
5d1c421c
JH
2602 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2603 * These need to be revisited when a newer toolchain becomes available.
2604 */
2605#if defined(__sparc64__) && defined(__GNUC__)
2606# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2607# undef SPARC64_GCC_WORKAROUND
2608# define SPARC64_GCC_WORKAROUND 1
2609# endif
2610#endif
2611
07be1b83 2612#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2613 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2614 SV * const mysv=sv_newmortal(); \
2615 regnode *Next = regnext(scan); \
2616 regprop(RExC_rx, mysv, scan); \
7f69552c 2617 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2618 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2619 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2620 }});
07be1b83 2621
1de06328 2622
bb914485 2623/* The below joins as many adjacent EXACTish nodes as possible into a single
0a982f06
KW
2624 * one. The regop may be changed if the node(s) contain certain sequences that
2625 * require special handling. The joining is only done if:
bb914485
KW
2626 * 1) there is room in the current conglomerated node to entirely contain the
2627 * next one.
2628 * 2) they are the exact same node type
2629 *
87b8b349 2630 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
2631 * these get optimized out
2632 *
0a982f06
KW
2633 * If a node is to match under /i (folded), the number of characters it matches
2634 * can be different than its character length if it contains a multi-character
2635 * fold. *min_subtract is set to the total delta of the input nodes.
bb914485 2636 *
a0c4c608
KW
2637 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2638 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2639 *
bb914485 2640 * This is as good a place as any to discuss the design of handling these
0a982f06
KW
2641 * multi-character fold sequences. It's been wrong in Perl for a very long
2642 * time. There are three code points in Unicode whose multi-character folds
2643 * were long ago discovered to mess things up. The previous designs for
2644 * dealing with these involved assigning a special node for them. This
2645 * approach doesn't work, as evidenced by this example:
a0c4c608 2646 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
0a982f06
KW
2647 * Both these fold to "sss", but if the pattern is parsed to create a node that
2648 * would match just the \xDF, it won't be able to handle the case where a
bb914485
KW
2649 * successful match would have to cross the node's boundary. The new approach
2650 * that hopefully generally solves the problem generates an EXACTFU_SS node
2651 * that is "sss".
2652 *
0a982f06
KW
2653 * It turns out that there are problems with all multi-character folds, and not
2654 * just these three. Now the code is general, for all such cases, but the
2655 * three still have some special handling. The approach taken is:
2656 * 1) This routine examines each EXACTFish node that could contain multi-
2657 * character fold sequences. It returns in *min_subtract how much to
9d071ca8 2658 * subtract from the the actual length of the string to get a real minimum
0a982f06
KW
2659 * match length; it is 0 if there are no multi-char folds. This delta is
2660 * used by the caller to adjust the min length of the match, and the delta
2661 * between min and max, so that the optimizer doesn't reject these
2662 * possibilities based on size constraints.
2663 * 2) Certain of these sequences require special handling by the trie code,
2664 * so, if found, this code changes the joined node type to special ops:
2665 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2666 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2667 * is used for an EXACTFU node that contains at least one "ss" sequence in
2668 * it. For non-UTF-8 patterns and strings, this is the only case where
2669 * there is a possible fold length change. That means that a regular
2670 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2671 * with length changes, and so can be processed faster. regexec.c takes
2672 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2673 * pre-folded by regcomp.c. This saves effort in regex matching.
87b8b349 2674 * However, the pre-folding isn't done for non-UTF8 patterns because the
0a982f06
KW
2675 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2676 * down by forcing the pattern into UTF8 unless necessary. Also what
2677 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2678 * possibilities for the non-UTF8 patterns are quite simple, except for
2679 * the sharp s. All the ones that don't involve a UTF-8 target string are
2680 * members of a fold-pair, and arrays are set up for all of them so that
2681 * the other member of the pair can be found quickly. Code elsewhere in
2682 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2683 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2684 * described in the next item.
bb914485
KW
2685 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2686 * 'ss' or not is not knowable at compile time. It will match iff the
2687 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2688 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
6012a526 2689 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
bb914485
KW
2690 * described in item 3). An assumption that the optimizer part of
2691 * regexec.c (probably unwittingly) makes is that a character in the
2692 * pattern corresponds to at most a single character in the target string.
2693 * (And I do mean character, and not byte here, unlike other parts of the
2694 * documentation that have never been updated to account for multibyte
2695 * Unicode.) This assumption is wrong only in this case, as all other
2696 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2697 * virtue of having this file pre-fold UTF-8 patterns. I'm
2698 * reluctant to try to change this assumption, so instead the code punts.
9d071ca8
KW
2699 * This routine examines EXACTF nodes for the sharp s, and returns a
2700 * boolean indicating whether or not the node is an EXACTF node that
2701 * contains a sharp s. When it is true, the caller sets a flag that later
2702 * causes the optimizer in this file to not set values for the floating
2703 * and fixed string lengths, and thus avoids the optimizer code in
2704 * regexec.c that makes the invalid assumption. Thus, there is no
2705 * optimization based on string lengths for EXACTF nodes that contain the
2706 * sharp s. This only happens for /id rules (which means the pattern
2707 * isn't in UTF-8).
bb914485 2708 */
1de06328 2709
9d071ca8 2710#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2711 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2712 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2713
be8e71aa 2714STATIC U32
9d071ca8 2715S_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
2716 /* Merge several consecutive EXACTish nodes into one. */
2717 regnode *n = regnext(scan);
2718 U32 stringok = 1;
2719 regnode *next = scan + NODE_SZ_STR(scan);
2720 U32 merged = 0;
2721 U32 stopnow = 0;
2722#ifdef DEBUGGING
2723 regnode *stop = scan;
72f13be8 2724 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2725#else
d47053eb
RGS
2726 PERL_UNUSED_ARG(depth);
2727#endif
7918f24d
NC
2728
2729 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2730#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2731 PERL_UNUSED_ARG(flags);
2732 PERL_UNUSED_ARG(val);
07be1b83 2733#endif
07be1b83 2734 DEBUG_PEEP("join",scan,depth);
bb914485 2735
3f410cf6
KW
2736 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2737 * EXACT ones that are mergeable to the current one. */
2738 while (n
2739 && (PL_regkind[OP(n)] == NOTHING
2740 || (stringok && OP(n) == OP(scan)))
07be1b83 2741 && NEXT_OFF(n)
3f410cf6
KW
2742 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2743 {
07be1b83
YO
2744
2745 if (OP(n) == TAIL || n > next)
2746 stringok = 0;
2747 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2748 DEBUG_PEEP("skip:",n,depth);
2749 NEXT_OFF(scan) += NEXT_OFF(n);
2750 next = n + NODE_STEP_REGNODE;
2751#ifdef DEBUGGING
2752 if (stringok)
2753 stop = n;
2754#endif
2755 n = regnext(n);
2756 }
2757 else if (stringok) {
786e8c11 2758 const unsigned int oldl = STR_LEN(scan);
07be1b83 2759 regnode * const nnext = regnext(n);
b2230d39 2760
87b8b349
KW
2761 /* XXX I (khw) kind of doubt that this works on platforms where
2762 * U8_MAX is above 255 because of lots of other assumptions */
79a81a6e 2763 /* Don't join if the sum can't fit into a single node */
b2230d39
KW
2764 if (oldl + STR_LEN(n) > U8_MAX)
2765 break;
07be1b83
YO
2766
2767 DEBUG_PEEP("merg",n,depth);
07be1b83 2768 merged++;
b2230d39 2769
07be1b83
YO
2770 NEXT_OFF(scan) += NEXT_OFF(n);
2771 STR_LEN(scan) += STR_LEN(n);
2772 next = n + NODE_SZ_STR(n);
2773 /* Now we can overwrite *n : */
2774 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2775#ifdef DEBUGGING
2776 stop = next - 1;
2777#endif
2778 n = nnext;
2779 if (stopnow) break;
2780 }
2781
d47053eb
RGS
2782#ifdef EXPERIMENTAL_INPLACESCAN
2783 if (flags && !NEXT_OFF(n)) {
2784 DEBUG_PEEP("atch", val, depth);
2785 if (reg_off_by_arg[OP(n)]) {
2786 ARG_SET(n, val - n);
2787 }
2788 else {
2789 NEXT_OFF(n) = val - n;
2790 }
2791 stopnow = 1;
2792 }
07be1b83
YO
2793#endif
2794 }
2c2b7f86 2795
9d071ca8 2796 *min_subtract = 0;
f758bddf 2797 *has_exactf_sharp_s = FALSE;
f646642f 2798
3f410cf6
KW
2799 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2800 * can now analyze for sequences of problematic code points. (Prior to
2801 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2802 * hence missed). The sequences only happen in folding, hence for any
2803 * non-EXACT EXACTish node */
86d6fcad 2804 if (OP(scan) != EXACT) {
0a982f06
KW
2805 const U8 * const s0 = (U8*) STRING(scan);
2806 const U8 * s = s0;
2807 const U8 * const s_end = s0 + STR_LEN(scan);
f758bddf
KW
2808
2809 /* One pass is made over the node's string looking for all the
2810 * possibilities. to avoid some tests in the loop, there are two main
2811 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2812 * non-UTF-8 */
2813 if (UTF) {
86d6fcad 2814
0a982f06
KW
2815 /* Examine the string for a multi-character fold sequence. UTF-8
2816 * patterns have all characters pre-folded by the time this code is
2817 * executed */
2818 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2819 length sequence we are looking for is 2 */
86d6fcad 2820 {
0a982f06
KW
2821 int count = 0;
2822 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2823 if (! len) { /* Not a multi-char fold: get next char */
2824 s += UTF8SKIP(s);
2825 continue;
2826 }
bb914485 2827
0a982f06
KW
2828 /* Nodes with 'ss' require special handling, except for EXACTFL
2829 * and EXACTFA for which there is no multi-char fold to this */
2830 if (len == 2 && *s == 's' && *(s+1) == 's'
2831 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2832 {
2833 count = 2;
2834 OP(scan) = EXACTFU_SS;
2835 s += 2;
2836 }
2837 else if (len == 6 /* len is the same in both ASCII and EBCDIC for these */
2838 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2839 COMBINING_DIAERESIS_UTF8
2840 COMBINING_ACUTE_ACCENT_UTF8,
2841 6)
2842 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2843 COMBINING_DIAERESIS_UTF8
2844 COMBINING_ACUTE_ACCENT_UTF8,
2845 6)))
2846 {
2847 count = 3;
2848
2849 /* These two folds require special handling by trie's, so
2850 * change the node type to indicate this. If EXACTFA and
2851 * EXACTFL were ever to be handled by trie's, this would
2852 * have to be changed. If this node has already been
2853 * changed to EXACTFU_SS in this loop, leave it as is. (I
2854 * (khw) think it doesn't matter in regexec.c for UTF
2855 * patterns, but no need to change it */
2856 if (OP(scan) == EXACTFU) {
2857 OP(scan) = EXACTFU_TRICKYFOLD;
2858 }
2859 s += 6;
2860 }
2861 else { /* Here is a generic multi-char fold. */
2862 const U8* multi_end = s + len;
2863
2864 /* Count how many characters in it. In the case of /l and
2865 * /aa, no folds which contain ASCII code points are
2866 * allowed, so check for those, and skip if found. (In
2867 * EXACTFL, no folds are allowed to any Latin1 code point,
2868 * not just ASCII. But there aren't any of these
2869 * currently, nor ever likely, so don't take the time to
2870 * test for them. The code that generates the
2871 * is_MULTI_foo() macros croaks should one actually get put
2872 * into Unicode .) */
2873 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2874 count = utf8_length(s, multi_end);
2875 s = multi_end;
2876 }
2877 else {
2878 while (s < multi_end) {
2879 if (isASCII(*s)) {
2880 s++;
2881 goto next_iteration;
2882 }
2883 else {
2884 s += UTF8SKIP(s);
2885 }
2886 count++;
2887 }
2888 }
2889 }
f758bddf 2890
0a982f06
KW
2891 /* The delta is how long the sequence is minus 1 (1 is how long
2892 * the character that folds to the sequence is) */
2893 *min_subtract += count - 1;
2894 next_iteration: ;
bb914485
KW
2895 }
2896 }
f758bddf 2897 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
bb914485 2898
0a982f06
KW
2899 /* Here, the pattern is not UTF-8. Look for the multi-char folds
2900 * that are all ASCII. As in the above case, EXACTFL and EXACTFA
2901 * nodes can't have multi-char folds to this range (and there are
2db97d41 2902 * no existing ones in the upper latin1 range). In the EXACTF
0a982f06
KW
2903 * case we look also for the sharp s, which can be in the final
2904 * position. Otherwise we can stop looking 1 byte earlier because
2905 * have to find at least two characters for a multi-fold */
f758bddf
KW
2906 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2907
0a982f06
KW
2908 /* The below is perhaps overboard, but this allows us to save a
2909 * test each time through the loop at the expense of a mask. This
2910 * is because on both EBCDIC and ASCII machines, 'S' and 's' differ
2911 * by a single bit. On ASCII they are 32 apart; on EBCDIC, they
2912 * are 64. This uses an exclusive 'or' to find that bit and then
2913 * inverts it to form a mask, with just a single 0, in the bit
2914 * position where 'S' and 's' differ. */
2915 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2916 const U8 s_masked = 's' & S_or_s_mask;
2917
2918 while (s < upper) {
40b1ba4f 2919 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
0a982f06
KW
2920 if (! len) { /* Not a multi-char fold. */
2921 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2922 {
2923 *has_exactf_sharp_s = TRUE;
2924 }
2925 s++;
2926 continue;
2927 }
2928
2929 if (len == 2
2930 && ((*s & S_or_s_mask) == s_masked)
2931 && ((*(s+1) & S_or_s_mask) == s_masked))
2932 {
2933
2934 /* EXACTF nodes need to know that the minimum length
2935 * changed so that a sharp s in the string can match this
2936 * ss in the pattern, but they remain EXACTF nodes, as they
2937 * won't match this unless the target string is is UTF-8,
2938 * which we don't know until runtime */
2939 if (OP(scan) != EXACTF) {
2940 OP(scan) = EXACTFU_SS;
2941 }
86d6fcad 2942 }
0a982f06
KW
2943
2944 *min_subtract += len - 1;
2945 s += len;
86d6fcad
KW
2946 }
2947 }
07be1b83 2948 }
3f410cf6 2949
07be1b83 2950#ifdef DEBUGGING
bb789b09
DM
2951 /* Allow dumping but overwriting the collection of skipped
2952 * ops and/or strings with fake optimized ops */
07be1b83
YO
2953 n = scan + NODE_SZ_STR(scan);
2954 while (n <= stop) {
bb789b09
DM
2955 OP(n) = OPTIMIZED;
2956 FLAGS(n) = 0;
2957 NEXT_OFF(n) = 0;
07be1b83
YO
2958 n++;
2959 }
2960#endif
2961 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2962 return stopnow;
2963}
2964
486ec47a 2965/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2966 Finds fixed substrings. */
2967
a0288114 2968/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
2969 to the position after last scanned or to NULL. */
2970
40d049e4
YO
2971#define INIT_AND_WITHP \
2972 assert(!and_withp); \
2973 Newx(and_withp,1,struct regnode_charclass_class); \
2974 SAVEFREEPV(and_withp)
07be1b83 2975
b515a41d 2976/* this is a chain of data about sub patterns we are processing that
486ec47a 2977 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
2978 we can simulate recursion without losing state. */
2979struct scan_frame;
2980typedef struct scan_frame {
2981 regnode *last; /* last node to process in this frame */
2982 regnode *next; /* next node to process when last is reached */
2983 struct scan_frame *prev; /*previous frame*/
2984 I32 stop; /* what stopparen do we use */
2985} scan_frame;
2986
304ee84b
YO
2987
2988#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2989
76e3520e 2990STATIC I32
40d049e4 2991S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
1de06328 2992 I32 *minlenp, I32 *deltap,
40d049e4
YO
2993 regnode *last,
2994 scan_data_t *data,
2995 I32 stopparen,
2996 U8* recursed,
2997 struct regnode_charclass_class *and_withp,
2998 U32 flags, U32 depth)
c277df42
IZ
2999 /* scanp: Start here (read-write). */
3000 /* deltap: Write maxlen-minlen here. */
3001 /* last: Stop before this one. */
40d049e4
YO
3002 /* data: string data about the pattern */
3003 /* stopparen: treat close N as END */
3004 /* recursed: which subroutines have we recursed into */
3005 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 3006{
97aff369 3007 dVAR;
2d608413
KW
3008 I32 min = 0; /* There must be at least this number of characters to match */
3009 I32 pars = 0, code;
c277df42
IZ
3010 regnode *scan = *scanp, *next;
3011 I32 delta = 0;
3012 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 3013 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
3014 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3015 scan_data_t data_fake;
a3621e74 3016 SV *re_trie_maxbuff = NULL;
786e8c11 3017 regnode *first_non_open = scan;
e2e6a0f1 3018 I32 stopmin = I32_MAX;
8aa23a47 3019 scan_frame *frame = NULL;
a3621e74 3020 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 3021
7918f24d
NC
3022 PERL_ARGS_ASSERT_STUDY_CHUNK;
3023
13a24bad 3024#ifdef DEBUGGING
40d049e4 3025 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 3026#endif
40d049e4 3027
786e8c11 3028 if ( depth == 0 ) {
40d049e4 3029 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
3030 first_non_open=regnext(first_non_open);
3031 }
3032
b81d288d 3033
8aa23a47
YO
3034 fake_study_recurse:
3035 while ( scan && OP(scan) != END && scan < last ){
2d608413
KW
3036 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3037 node length to get a real minimum (because
3038 the folded version may be shorter) */
f758bddf 3039 bool has_exactf_sharp_s = FALSE;
8aa23a47 3040 /* Peephole optimizer: */
304ee84b 3041 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 3042 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
3043
3044 /* Its not clear to khw or hv why this is done here, and not in the
3045 * clauses that deal with EXACT nodes. khw's guess is that it's
3046 * because of a previous design */
9d071ca8 3047 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
3048
3049 /* Follow the next-chain of the current node and optimize
3050 away all the NOTHINGs from it. */
3051 if (OP(scan) != CURLYX) {
3052 const int max = (reg_off_by_arg[OP(scan)]
3053 ? I32_MAX
3054 /* I32 may be smaller than U16 on CRAYs! */
3055 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3056 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3057 int noff;
3058 regnode *n = scan;
686b73d4 3059
8aa23a47
YO
3060 /* Skip NOTHING and LONGJMP. */
3061 while ((n = regnext(n))
3062 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3063 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3064 && off + noff < max)
3065 off += noff;
3066 if (reg_off_by_arg[OP(scan)])
3067 ARG(scan) = off;
3068 else
3069 NEXT_OFF(scan) = off;
3070 }
a3621e74 3071
c277df42 3072
8aa23a47
YO
3073
3074 /* The principal pseudo-switch. Cannot be a switch, since we
3075 look into several different things. */
3076 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3077 || OP(scan) == IFTHEN) {
3078 next = regnext(scan);
3079 code = OP(scan);
3080 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3081
8aa23a47
YO
3082 if (OP(next) == code || code == IFTHEN) {
3083 /* NOTE - There is similar code to this block below for handling
3084 TRIE nodes on a re-study. If you change stuff here check there
3085 too. */
3086 I32 max1 = 0, min1 = I32_MAX, num = 0;
3087 struct regnode_charclass_class accum;
3088 regnode * const startbranch=scan;
686b73d4 3089
8aa23a47 3090 if (flags & SCF_DO_SUBSTR)
304ee84b 3091 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3092 if (flags & SCF_DO_STCLASS)
e755fd73 3093 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3094
3095 while (OP(scan) == code) {
3096 I32 deltanext, minnext, f = 0, fake;
3097 struct regnode_charclass_class this_class;
3098
3099 num++;
3100 data_fake.flags = 0;
3101 if (data) {
3102 data_fake.whilem_c = data->whilem_c;
3103 data_fake.last_closep = data->last_closep;
3104 }
3105 else
3106 data_fake.last_closep = &fake;
58e23c8d
YO
3107
3108 data_fake.pos_delta = delta;
8aa23a47
YO
3109 next = regnext(scan);
3110 scan = NEXTOPER(scan);
3111 if (code != BRANCH)
c277df42 3112 scan = NEXTOPER(scan);
8aa23a47 3113 if (flags & SCF_DO_STCLASS) {
e755fd73 3114 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3115 data_fake.start_class = &this_class;
3116 f = SCF_DO_STCLASS_AND;
58e23c8d 3117 }
8aa23a47
YO
3118 if (flags & SCF_WHILEM_VISITED_POS)
3119 f |= SCF_WHILEM_VISITED_POS;
3120
3121 /* we suppose the run is continuous, last=next...*/
3122 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3123 next, &data_fake,
3124 stopparen, recursed, NULL, f,depth+1);
3125 if (min1 > minnext)
3126 min1 = minnext;
3127 if (max1 < minnext + deltanext)
3128 max1 = minnext + deltanext;
3129 if (deltanext == I32_MAX)
3130 is_inf = is_inf_internal = 1;
3131 scan = next;
3132 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3133 pars++;
3134 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3135 if ( stopmin > minnext)
3136 stopmin = min + min1;
3137 flags &= ~SCF_DO_SUBSTR;
3138 if (data)
3139 data->flags |= SCF_SEEN_ACCEPT;
3140 }
3141 if (data) {
3142 if (data_fake.flags & SF_HAS_EVAL)
3143 data->flags |= SF_HAS_EVAL;
3144 data->whilem_c = data_fake.whilem_c;
3dab1dad 3145 }
8aa23a47 3146 if (flags & SCF_DO_STCLASS)
3fffb88a 3147 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3148 }
3149 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3150 min1 = 0;
3151 if (flags & SCF_DO_SUBSTR) {
3152 data->pos_min += min1;
3153 data->pos_delta += max1 - min1;
3154 if (max1 != min1 || is_inf)
3155 data->longest = &(data->longest_float);
3156 }
3157 min += min1;
3158 delta += max1 - min1;
3159 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3160 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3161 if (min1) {
3162 cl_and(data->start_class, and_withp);
3163 flags &= ~SCF_DO_STCLASS;
653099ff 3164 }
8aa23a47
YO
3165 }
3166 else if (flags & SCF_DO_STCLASS_AND) {
3167 if (min1) {
3168 cl_and(data->start_class, &accum);
3169 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3170 }
8aa23a47
YO
3171 else {
3172 /* Switch to OR mode: cache the old value of
3173 * data->start_class */
3174 INIT_AND_WITHP;
3175 StructCopy(data->start_class, and_withp,
3176 struct regnode_charclass_class);
3177 flags &= ~SCF_DO_STCLASS_AND;
3178 StructCopy(&accum, data->start_class,
3179 struct regnode_charclass_class);
3180 flags |= SCF_DO_STCLASS_OR;
899d20b9 3181 SET_SSC_EOS(data->start_class);
de0c8cb8 3182 }
8aa23a47 3183 }
a3621e74 3184
8aa23a47
YO
3185 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3186 /* demq.
a3621e74 3187
8aa23a47
YO
3188 Assuming this was/is a branch we are dealing with: 'scan' now
3189 points at the item that follows the branch sequence, whatever
3190 it is. We now start at the beginning of the sequence and look
3191 for subsequences of
a3621e74 3192
8aa23a47
YO
3193 BRANCH->EXACT=>x1
3194 BRANCH->EXACT=>x2
3195 tail
a3621e74 3196
8aa23a47 3197 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3198
486ec47a 3199 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3200 element into a trie and then add the subsequent branch exact
3201 strings to the trie.
a3621e74 3202
8aa23a47 3203 We have two cases
a3621e74 3204
3b753521 3205 1. patterns where the whole set of branches can be converted.
a3621e74 3206
8aa23a47 3207 2. patterns where only a subset can be converted.
a3621e74 3208
8aa23a47
YO
3209 In case 1 we can replace the whole set with a single regop
3210 for the trie. In case 2 we need to keep the start and end
3b753521 3211 branches so
a3621e74 3212
8aa23a47
YO
3213 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3214 becomes BRANCH TRIE; BRANCH X;
786e8c11 3215
8aa23a47
YO
3216 There is an additional case, that being where there is a
3217 common prefix, which gets split out into an EXACT like node
3218 preceding the TRIE node.
a3621e74 3219
8aa23a47
YO
3220 If x(1..n)==tail then we can do a simple trie, if not we make
3221 a "jump" trie, such that when we match the appropriate word
486ec47a 3222 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3223 a nested if into a case structure of sorts.
b515a41d 3224
8aa23a47 3225 */
686b73d4 3226
8aa23a47
YO
3227 int made=0;
3228 if (!re_trie_maxbuff) {
3229 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3230 if (!SvIOK(re_trie_maxbuff))
3231 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3232 }
3233 if ( SvIV(re_trie_maxbuff)>=0 ) {
3234 regnode *cur;
3235 regnode *first = (regnode *)NULL;
3236 regnode *last = (regnode *)NULL;
3237 regnode *tail = scan;
fab2782b 3238 U8 trietype = 0;
8aa23a47 3239 U32 count=0;
a3621e74
YO
3240
3241#ifdef DEBUGGING
8aa23a47 3242 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3243#endif
8aa23a47
YO
3244 /* var tail is used because there may be a TAIL
3245 regop in the way. Ie, the exacts will point to the
3246 thing following the TAIL, but the last branch will
3247 point at the TAIL. So we advance tail. If we
3248 have nested (?:) we may have to move through several
3249 tails.
3250 */
3251
3252 while ( OP( tail ) == TAIL ) {
3253 /* this is the TAIL generated by (?:) */
3254 tail = regnext( tail );
3255 }
a3621e74 3256
8aa23a47 3257
df826430 3258 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3259 regprop(RExC_rx, mysv, tail );
3260 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3261 (int)depth * 2 + 2, "",
3262 "Looking for TRIE'able sequences. Tail node is: ",
3263 SvPV_nolen_const( mysv )
3264 );
3265 });
3266
3267 /*
3268
fab2782b
YO
3269 Step through the branches
3270 cur represents each branch,
3271 noper is the first thing to be matched as part of that branch
3272 noper_next is the regnext() of that node.
3273
3274 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3275 via a "jump trie" but we also support building with NOJUMPTRIE,
3276 which restricts the trie logic to structures like /FOO|BAR/.
3277
3278 If noper is a trieable nodetype then the branch is a possible optimization
3279 target. If we are building under NOJUMPTRIE then we require that noper_next
3280 is the same as scan (our current position in the regex program).
3281
3282 Once we have two or more consecutive such branches we can create a
3283 trie of the EXACT's contents and stitch it in place into the program.
3284
3285 If the sequence represents all of the branches in the alternation we
3286 replace the entire thing with a single TRIE node.
3287
3288 Otherwise when it is a subsequence we need to stitch it in place and
3289 replace only the relevant branches. This means the first branch has
3290 to remain as it is used by the alternation logic, and its next pointer,
3291 and needs to be repointed at the item on the branch chain following
3292 the last branch we have optimized away.
3293
3294 This could be either a BRANCH, in which case the subsequence is internal,
3295 or it could be the item following the branch sequence in which case the
3296 subsequence is at the end (which does not necessarily mean the first node
3297 is the start of the alternation).
3298
3299 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3300
3301 optype | trietype
3302 ----------------+-----------
3303 NOTHING | NOTHING
3304 EXACT | EXACT
3305 EXACTFU | EXACTFU
3306 EXACTFU_SS | EXACTFU
3307 EXACTFU_TRICKYFOLD | EXACTFU
3308 EXACTFA | 0
3309
8aa23a47
YO
3310
3311 */
fab2782b
YO
3312#define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3313 ( EXACT == (X) ) ? EXACT : \
3314 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3315 0 )
8aa23a47
YO
3316
3317 /* dont use tail as the end marker for this traverse */
3318 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3319 regnode * const noper = NEXTOPER( cur );
fab2782b
YO
3320 U8 noper_type = OP( noper );
3321 U8 noper_trietype = TRIE_TYPE( noper_type );
b515a41d 3322#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3323 regnode * const noper_next = regnext( noper );
df826430
YO
3324 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3325 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
b515a41d
YO
3326#endif
3327
df826430 3328 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3329 regprop(RExC_rx, mysv, cur);
3330 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3331 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3332
3333 regprop(RExC_rx, mysv, noper);
3334 PerlIO_printf( Perl_debug_log, " -> %s",
3335 SvPV_nolen_const(mysv));
3336
3337 if ( noper_next ) {
3338 regprop(RExC_rx, mysv, noper_next );
3339 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3340 SvPV_nolen_const(mysv));
3341 }
df826430
YO
3342 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3343 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3344 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3345 );
8aa23a47 3346 });
fab2782b
YO
3347
3348 /* Is noper a trieable nodetype that can be merged with the
3349 * current trie (if there is one)? */
3350 if ( noper_trietype
3351 &&
3352 (
df826430
YO
3353 ( noper_trietype == NOTHING)
3354 || ( trietype == NOTHING )
a40630bf 3355 || ( trietype == noper_trietype )
fab2782b 3356 )
786e8c11 3357#ifdef NOJUMPTRIE
8aa23a47 3358 && noper_next == tail
786e8c11 3359#endif
8aa23a47
YO
3360 && count < U16_MAX)
3361 {
fab2782b
YO
3362 /* Handle mergable triable node
3363 * Either we are the first node in a new trieable sequence,
3364 * in which case we do some bookkeeping, otherwise we update
3365 * the end pointer. */
fab2782b 3366 if ( !first ) {
3b6759a6 3367 first = cur;
df826430
YO
3368 if ( noper_trietype == NOTHING ) {
3369#if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3370 regnode * const noper_next = regnext( noper );
3b6759a6 3371 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
df826430
YO
3372 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3373#endif
3374
190c1910 3375 if ( noper_next_trietype ) {
df826430 3376 trietype = noper_next_trietype;
190c1910
YO
3377 } else if (noper_next_type) {
3378 /* a NOTHING regop is 1 regop wide. We need at least two
3379 * for a trie so we can't merge this in */
3380 first = NULL;
3381 }
3382 } else {
3383 trietype = noper_trietype;
3b6759a6 3384 }
8aa23a47 3385 } else {
fab2782b
YO
3386 if ( trietype == NOTHING )
3387 trietype = noper_trietype;
8aa23a47
YO
3388 last = cur;
3389 }
df826430
YO
3390 if (first)
3391 count++;
fab2782b
YO
3392 } /* end handle mergable triable node */
3393 else {
3394 /* handle unmergable node -
3395 * noper may either be a triable node which can not be tried
3396 * together with the current trie, or a non triable node */
729aaeb5
YO
3397 if ( last ) {
3398 /* If last is set and trietype is not NOTHING then we have found
3399 * at least two triable branch sequences in a row of a similar
3400 * trietype so we can turn them into a trie. If/when we
3401 * allow NOTHING to start a trie sequence this condition will be
3402 * required, and it isn't expensive so we leave it in for now. */
e6351b37 3403 if ( trietype && trietype != NOTHING )
729aaeb5
YO
3404 make_trie( pRExC_state,
3405 startbranch, first, cur, tail, count,
3406 trietype, depth+1 );
fab2782b 3407 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
8aa23a47 3408 }
fab2782b 3409 if ( noper_trietype
786e8c11 3410#ifdef NOJUMPTRIE
8aa23a47 3411 && noper_next == tail
786e8c11 3412#endif
8aa23a47 3413 ){
fab2782b 3414 /* noper is triable, so we can start a new trie sequence */
8aa23a47
YO
3415 count = 1;
3416 first = cur;
fab2782b
YO
3417 trietype = noper_trietype;
3418 } else if (first) {
3419 /* if we already saw a first but the current node is not triable then we have
3420 * to reset the first information. */
8aa23a47
YO
3421 count = 0;
3422 first = NULL;
fab2782b 3423 trietype = 0;
8aa23a47 3424 }
fab2782b
YO
3425 } /* end handle unmergable node */
3426 } /* loop over branches */
df826430 3427 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3428 regprop(RExC_rx, mysv, cur);
3429 PerlIO_printf( Perl_debug_log,
3430 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3431 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3432
3433 });
e6351b37 3434 if ( last && trietype ) {
3b6759a6
YO
3435 if ( trietype != NOTHING ) {
3436 /* the last branch of the sequence was part of a trie,
3437 * so we have to construct it here outside of the loop
3438 */
3439 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
686b73d4 3440#ifdef TRIE_STUDY_OPT
3b6759a6
YO
3441 if ( ((made == MADE_EXACT_TRIE &&
3442 startbranch == first)
3443 || ( first_non_open == first )) &&
3444 depth==0 ) {
3445 flags |= SCF_TRIE_RESTUDY;
3446 if ( startbranch == first
3447 && scan == tail )
3448 {
3449 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3450 }
8aa23a47 3451 }
3dab1dad 3452#endif
3b6759a6
YO
3453 } else {
3454 /* at this point we know whatever we have is a NOTHING sequence/branch
3455 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3456 */
3457 if ( startbranch == first ) {
3458 regnode *opt;
3459 /* the entire thing is a NOTHING sequence, something like this:
3460 * (?:|) So we can turn it into a plain NOTHING op. */
3461 DEBUG_TRIE_COMPILE_r({
3462 regprop(RExC_rx, mysv, cur);
3463 PerlIO_printf( Perl_debug_log,
3464 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3465 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3466
3467 });
3468 OP(startbranch)= NOTHING;
3469 NEXT_OFF(startbranch)= tail - startbranch;
3470 for ( opt= startbranch + 1; opt < tail ; opt++ )
3471 OP(opt)= OPTIMIZED;
3472 }
3473 }
fab2782b
YO
3474 } /* end if ( last) */
3475 } /* TRIE_MAXBUF is non zero */
8aa23a47
YO
3476
3477 } /* do trie */
3478
653099ff 3479 }
8aa23a47
YO
3480 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3481 scan = NEXTOPER(NEXTOPER(scan));
3482 } else /* single branch is optimized. */
3483 scan = NEXTOPER(scan);
3484 continue;
3485 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3486 scan_frame *newframe = NULL;
3487 I32 paren;
3488 regnode *start;
3489 regnode *end;
3490
3491 if (OP(scan) != SUSPEND) {
3492 /* set the pointer */
3493 if (OP(scan) == GOSUB) {
3494 paren = ARG(scan);
3495 RExC_recurse[ARG2L(scan)] = scan;
3496 start = RExC_open_parens[paren-1];
3497 end = RExC_close_parens[paren-1];
3498 } else {
3499 paren = 0;
f8fc2ecf 3500 start = RExC_rxi->program + 1;
8aa23a47
YO
3501 end = RExC_opend;
3502 }
3503 if (!recursed) {
3504 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3505 SAVEFREEPV(recursed);
3506 }
3507 if (!PAREN_TEST(recursed,paren+1)) {
3508 PAREN_SET(recursed,paren+1);
3509 Newx(newframe,1,scan_frame);
3510 } else {
3511 if (flags & SCF_DO_SUBSTR) {
304ee84b 3512 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3513 data->longest = &(data->longest_float);
3514 }
3515 is_inf = is_inf_internal = 1;
3516 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3fffb88a 3517 cl_anything(pRExC_state, data->start_class);
8aa23a47
YO
3518 flags &= ~SCF_DO_STCLASS;
3519 }
3520 } else {
3521 Newx(newframe,1,scan_frame);
3522 paren = stopparen;
3523 start = scan+2;
3524 end = regnext(scan);
3525 }
3526 if (newframe) {
3527 assert(start);
3528 assert(end);
3529 SAVEFREEPV(newframe);
3530 newframe->next = regnext(scan);
3531 newframe->last = last;
3532 newframe->stop = stopparen;
3533 newframe->prev = frame;
3534
3535 frame = newframe;
3536 scan = start;
3537 stopparen = paren;
3538 last = end;
3539
3540 continue;
3541 }
3542 }
3543 else if (OP(scan) == EXACT) {
3544 I32 l = STR_LEN(scan);
3545 UV uc;
3546 if (UTF) {
3547 const U8 * const s = (U8*)STRING(scan);
4b88fb76 3548 uc = utf8_to_uvchr_buf(s, s + l, NULL);
8aa23a47 3549 l = utf8_length(s, s + l);
8aa23a47
YO
3550 } else {
3551 uc = *((U8*)STRING(scan));
3552 }
3553 min += l;
3554 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3555 /* The code below prefers earlier match for fixed
3556 offset, later match for variable offset. */
3557 if (data->last_end == -1) { /* Update the start info. */
3558 data->last_start_min = data->pos_min;
3559 data->last_start_max = is_inf
3560 ? I32_MAX : data->pos_min + data->pos_delta;
b515a41d 3561 }
8aa23a47
YO
3562 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3563 if (UTF)
3564 SvUTF8_on(data->last_found);
3565 {
3566 SV * const sv = data->last_found;
3567 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3568 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3569 if (mg && mg->mg_len >= 0)
3570 mg->mg_len += utf8_length((U8*)STRING(scan),
3571 (U8*)STRING(scan)+STR_LEN(scan));
b515a41d 3572 }
8aa23a47
YO
3573 data->last_end = data->pos_min + l;
3574 data->pos_min += l; /* As in the first entry. */
3575 data->flags &= ~SF_BEFORE_EOL;
3576 }
3577 if (flags & SCF_DO_STCLASS_AND) {
3578 /* Check whether it is compatible with what we know already! */
3579 int compat = 1;
3580