This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c: Split count variable into two: min, max
[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. */
ea3daa5d 126 SSize_t whilem_seen; /* number of WHILEM in this expr */
fac92740 127 regnode *emit_start; /* Start of emitted-code area */
3b57cd43 128 regnode *emit_bound; /* First regnode outside of the allocated space */
f7c7e32a
DM
129 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
130 implies compiling, so don't emit */
131 regnode emit_dummy; /* placeholder for emit to point to */
830247a4
IZ
132 I32 naughty; /* How bad is this pattern? */
133 I32 sawback; /* Did we see \1, ...? */
134 U32 seen;
ea3daa5d 135 SSize_t size; /* Code size. */
c74340f9
YO
136 I32 npar; /* Capture buffer count, (OPEN). */
137 I32 cpar; /* Capture buffer count, (CLOSE). */
e2e6a0f1 138 I32 nestroot; /* root parens we are in - used by accept */
830247a4
IZ
139 I32 extralen;
140 I32 seen_zerolen;
40d049e4
YO
141 regnode **open_parens; /* pointers to open parens */
142 regnode **close_parens; /* pointers to close parens */
143 regnode *opend; /* END node in program */
02daf0ab
YO
144 I32 utf8; /* whether the pattern is utf8 or not */
145 I32 orig_utf8; /* whether the pattern was originally in utf8 */
146 /* XXX use this for future optimisation of case
147 * where pattern must be upgraded to utf8. */
e40e74fe
KW
148 I32 uni_semantics; /* If a d charset modifier should use unicode
149 rules, even if the pattern is not in
150 utf8 */
81714fb9 151 HV *paren_names; /* Paren names */
1f1031fe 152
40d049e4
YO
153 regnode **recurse; /* Recurse regops */
154 I32 recurse_count; /* Number of recurse regops */
b57e4118 155 I32 in_lookbehind;
4624b182 156 I32 contains_locale;
bb3f3ed2 157 I32 override_recoding;
9d53c457 158 I32 in_multi_char_class;
3d2bd50a 159 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 160 within pattern */
b1603ef8
DM
161 int num_code_blocks; /* size of code_blocks[] */
162 int code_index; /* next code_blocks[] slot */
830247a4
IZ
163#if ADD_TO_REGEXEC
164 char *starttry; /* -Dr: where regtry was called. */
165#define RExC_starttry (pRExC_state->starttry)
166#endif
d24ca0c5 167 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 168#ifdef DEBUGGING
be8e71aa 169 const char *lastparse;
3dab1dad 170 I32 lastnum;
1f1031fe 171 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
172#define RExC_lastparse (pRExC_state->lastparse)
173#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 174#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 175#endif
830247a4
IZ
176} RExC_state_t;
177
e2509266 178#define RExC_flags (pRExC_state->flags)
514a91f1 179#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 180#define RExC_precomp (pRExC_state->precomp)
288b8c02 181#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 182#define RExC_rx (pRExC_state->rx)
f8fc2ecf 183#define RExC_rxi (pRExC_state->rxi)
fac92740 184#define RExC_start (pRExC_state->start)
830247a4
IZ
185#define RExC_end (pRExC_state->end)
186#define RExC_parse (pRExC_state->parse)
187#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237
YO
188#ifdef RE_TRACK_PATTERN_OFFSETS
189#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
190#endif
830247a4 191#define RExC_emit (pRExC_state->emit)
f7c7e32a 192#define RExC_emit_dummy (pRExC_state->emit_dummy)
fac92740 193#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 194#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
195#define RExC_naughty (pRExC_state->naughty)
196#define RExC_sawback (pRExC_state->sawback)
197#define RExC_seen (pRExC_state->seen)
198#define RExC_size (pRExC_state->size)
199#define RExC_npar (pRExC_state->npar)
e2e6a0f1 200#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
201#define RExC_extralen (pRExC_state->extralen)
202#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 203#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 204#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 205#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
206#define RExC_open_parens (pRExC_state->open_parens)
207#define RExC_close_parens (pRExC_state->close_parens)
208#define RExC_opend (pRExC_state->opend)
81714fb9 209#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
210#define RExC_recurse (pRExC_state->recurse)
211#define RExC_recurse_count (pRExC_state->recurse_count)
b57e4118 212#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 213#define RExC_contains_locale (pRExC_state->contains_locale)
9d53c457
KW
214#define RExC_override_recoding (pRExC_state->override_recoding)
215#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
830247a4 216
cde0cee5 217
a687059c
LW
218#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
219#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
4d68ffa0 220 ((*s) == '{' && regcurly(s, FALSE)))
a687059c 221
35c8bce7
LW
222#ifdef SPSTART
223#undef SPSTART /* dratted cpp namespace... */
224#endif
a687059c
LW
225/*
226 * Flags to be passed up and down.
227 */
a687059c 228#define WORST 0 /* Worst case. */
a3b492c3 229#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 230
e64f369d 231/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
232 * character. (There needs to be a case: in the switch statement in regexec.c
233 * for any node marked SIMPLE.) Note that this is not the same thing as
234 * REGNODE_SIMPLE */
fda99bee 235#define SIMPLE 0x02
e64f369d 236#define SPSTART 0x04 /* Starts with * or + */
8d9c2815
NC
237#define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
238#define TRYAGAIN 0x10 /* Weeded out a declaration. */
239#define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
a687059c 240
3dab1dad
YO
241#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
242
07be1b83
YO
243/* whether trie related optimizations are enabled */
244#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
245#define TRIE_STUDY_OPT
786e8c11 246#define FULL_TRIE_STUDY
07be1b83
YO
247#define TRIE_STCLASS
248#endif
1de06328
YO
249
250
40d049e4
YO
251
252#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
253#define PBITVAL(paren) (1 << ((paren) & 7))
254#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
255#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
256#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
257
bbd61b5f 258#define REQUIRE_UTF8 STMT_START { \
8d9c2815
NC
259 if (!UTF) { \
260 *flagp = RESTART_UTF8; \
261 return NULL; \
262 } \
bbd61b5f 263 } STMT_END
40d049e4 264
f19b1a63
KW
265/* This converts the named class defined in regcomp.h to its equivalent class
266 * number defined in handy.h. */
267#define namedclass_to_classnum(class) ((int) ((class) / 2))
268#define classnum_to_namedclass(classnum) ((classnum) * 2)
269
1de06328
YO
270/* About scan_data_t.
271
272 During optimisation we recurse through the regexp program performing
273 various inplace (keyhole style) optimisations. In addition study_chunk
274 and scan_commit populate this data structure with information about
275 what strings MUST appear in the pattern. We look for the longest
3b753521 276 string that must appear at a fixed location, and we look for the
1de06328
YO
277 longest string that may appear at a floating location. So for instance
278 in the pattern:
279
280 /FOO[xX]A.*B[xX]BAR/
281
282 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
283 strings (because they follow a .* construct). study_chunk will identify
284 both FOO and BAR as being the longest fixed and floating strings respectively.
285
286 The strings can be composites, for instance
287
288 /(f)(o)(o)/
289
290 will result in a composite fixed substring 'foo'.
291
292 For each string some basic information is maintained:
293
294 - offset or min_offset
295 This is the position the string must appear at, or not before.
296 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
297 characters must match before the string we are searching for.
298 Likewise when combined with minlenp and the length of the string it
1de06328
YO
299 tells us how many characters must appear after the string we have
300 found.
301
302 - max_offset
303 Only used for floating strings. This is the rightmost point that
ea3daa5d 304 the string can appear at. If set to SSize_t_MAX it indicates that the
1de06328
YO
305 string can occur infinitely far to the right.
306
307 - minlenp
2d608413
KW
308 A pointer to the minimum number of characters of the pattern that the
309 string was found inside. This is important as in the case of positive
1de06328
YO
310 lookahead or positive lookbehind we can have multiple patterns
311 involved. Consider
312
313 /(?=FOO).*F/
314
315 The minimum length of the pattern overall is 3, the minimum length
316 of the lookahead part is 3, but the minimum length of the part that
317 will actually match is 1. So 'FOO's minimum length is 3, but the
318 minimum length for the F is 1. This is important as the minimum length
319 is used to determine offsets in front of and behind the string being
320 looked for. Since strings can be composites this is the length of the
486ec47a 321 pattern at the time it was committed with a scan_commit. Note that
1de06328
YO
322 the length is calculated by study_chunk, so that the minimum lengths
323 are not known until the full pattern has been compiled, thus the
324 pointer to the value.
325
326 - lookbehind
327
328 In the case of lookbehind the string being searched for can be
329 offset past the start point of the final matching string.
330 If this value was just blithely removed from the min_offset it would
331 invalidate some of the calculations for how many chars must match
332 before or after (as they are derived from min_offset and minlen and
333 the length of the string being searched for).
334 When the final pattern is compiled and the data is moved from the
335 scan_data_t structure into the regexp structure the information
336 about lookbehind is factored in, with the information that would
337 have been lost precalculated in the end_shift field for the
338 associated string.
339
340 The fields pos_min and pos_delta are used to store the minimum offset
341 and the delta to the maximum offset at the current point in the pattern.
342
343*/
2c2d71f5
JH
344
345typedef struct scan_data_t {
1de06328
YO
346 /*I32 len_min; unused */
347 /*I32 len_delta; unused */
49f55535 348 SSize_t pos_min;
ea3daa5d 349 SSize_t pos_delta;
2c2d71f5 350 SV *last_found;
ea3daa5d 351 SSize_t last_end; /* min value, <0 unless valid. */
49f55535 352 SSize_t last_start_min;
ea3daa5d 353 SSize_t last_start_max;
1de06328
YO
354 SV **longest; /* Either &l_fixed, or &l_float. */
355 SV *longest_fixed; /* longest fixed string found in pattern */
49f55535 356 SSize_t offset_fixed; /* offset where it starts */
ea3daa5d 357 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
358 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
359 SV *longest_float; /* longest floating string found in pattern */
49f55535 360 SSize_t offset_float_min; /* earliest point in string it can appear */
ea3daa5d
FC
361 SSize_t offset_float_max; /* latest point in string it can appear */
362 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
49f55535 363 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
2c2d71f5
JH
364 I32 flags;
365 I32 whilem_c;
49f55535 366 SSize_t *last_closep;
653099ff 367 struct regnode_charclass_class *start_class;
2c2d71f5
JH
368} scan_data_t;
369
c02c3054
KW
370/* The below is perhaps overboard, but this allows us to save a test at the
371 * expense of a mask. This is because on both EBCDIC and ASCII machines, 'A'
372 * and 'a' differ by a single bit; the same with the upper and lower case of
373 * all other ASCII-range alphabetics. On ASCII platforms, they are 32 apart;
374 * on EBCDIC, they are 64. This uses an exclusive 'or' to find that bit and
375 * then inverts it to form a mask, with just a single 0, in the bit position
376 * where the upper- and lowercase differ. XXX There are about 40 other
377 * instances in the Perl core where this micro-optimization could be used.
378 * Should decide if maintenance cost is worse, before changing those
379 *
380 * Returns a boolean as to whether or not 'v' is either a lowercase or
381 * uppercase instance of 'c', where 'c' is in [A-Za-z]. If 'c' is a
382 * compile-time constant, the generated code is better than some optimizing
383 * compilers figure out, amounting to a mask and test. The results are
384 * meaningless if 'c' is not one of [A-Za-z] */
385#define isARG2_lower_or_UPPER_ARG1(c, v) \
386 (((v) & ~('A' ^ 'a')) == ((c) & ~('A' ^ 'a')))
387
a687059c 388/*
e50aee73 389 * Forward declarations for pregcomp()'s friends.
a687059c 390 */
a0d0e21e 391
27da23d5 392static const scan_data_t zero_scan_data =
1de06328 393 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
394
395#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
396#define SF_BEFORE_SEOL 0x0001
397#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
398#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
399#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
400
09b7f37c
CB
401#ifdef NO_UNARY_PLUS
402# define SF_FIX_SHIFT_EOL (0+2)
403# define SF_FL_SHIFT_EOL (0+4)
404#else
405# define SF_FIX_SHIFT_EOL (+2)
406# define SF_FL_SHIFT_EOL (+4)
407#endif
c277df42
IZ
408
409#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
410#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
411
412#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
413#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
414#define SF_IS_INF 0x0040
415#define SF_HAS_PAR 0x0080
416#define SF_IN_PAR 0x0100
417#define SF_HAS_EVAL 0x0200
418#define SCF_DO_SUBSTR 0x0400
653099ff
GS
419#define SCF_DO_STCLASS_AND 0x0800
420#define SCF_DO_STCLASS_OR 0x1000
421#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 422#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 423
786e8c11 424#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
e2e6a0f1 425#define SCF_SEEN_ACCEPT 0x8000
688e0391 426#define SCF_TRIE_DOING_RESTUDY 0x10000
07be1b83 427
43fead97 428#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
429
430/* The enums for all these are ordered so things work out correctly */
a62b1201 431#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
cfaf538b 432#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
00b27cfc 433#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
434#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
435#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1 436#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a725e29c 437#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 438
43fead97 439#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 440
93733859 441#define OOB_NAMEDCLASS -1
b8c5462f 442
8e661ac5
KW
443/* There is no code point that is out-of-bounds, so this is problematic. But
444 * its only current use is to initialize a variable that is always set before
445 * looked at. */
446#define OOB_UNICODE 0xDEADBEEF
447
a0ed51b3
LW
448#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
449#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
450
8615cb43 451
b45f050a
JF
452/* length of regex to show in messages that don't mark a position within */
453#define RegexLengthToShowInErrorMessages 127
454
455/*
456 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
457 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
458 * op/pragma/warn/regcomp.
459 */
7253e4e3
RK
460#define MARKER1 "<-- HERE" /* marker as it appears in the description */
461#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 462
7253e4e3 463#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
464
465/*
466 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
467 * arg. Show regex, up to a maximum length. If it's too long, chop and add
468 * "...".
469 */
58e23c8d 470#define _FAIL(code) STMT_START { \
bfed75c6 471 const char *ellipses = ""; \
ccb2c380
MP
472 IV len = RExC_end - RExC_precomp; \
473 \
474 if (!SIZE_ONLY) \
a5e7bc51 475 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
476 if (len > RegexLengthToShowInErrorMessages) { \
477 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
478 len = RegexLengthToShowInErrorMessages - 10; \
479 ellipses = "..."; \
480 } \
58e23c8d 481 code; \
ccb2c380 482} STMT_END
8615cb43 483
58e23c8d
YO
484#define FAIL(msg) _FAIL( \
485 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
486 msg, (int)len, RExC_precomp, ellipses))
487
488#define FAIL2(msg,arg) _FAIL( \
489 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
490 arg, (int)len, RExC_precomp, ellipses))
491
b45f050a 492/*
b45f050a
JF
493 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
494 */
ccb2c380 495#define Simple_vFAIL(m) STMT_START { \
a28509cc 496 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
497 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
498 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
499} STMT_END
b45f050a
JF
500
501/*
502 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
503 */
ccb2c380
MP
504#define vFAIL(m) STMT_START { \
505 if (!SIZE_ONLY) \
a5e7bc51 506 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
507 Simple_vFAIL(m); \
508} STMT_END
b45f050a
JF
509
510/*
511 * Like Simple_vFAIL(), but accepts two arguments.
512 */
ccb2c380 513#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 514 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
515 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
516 (int)offset, RExC_precomp, RExC_precomp + offset); \
517} STMT_END
b45f050a
JF
518
519/*
520 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
521 */
ccb2c380
MP
522#define vFAIL2(m,a1) STMT_START { \
523 if (!SIZE_ONLY) \
a5e7bc51 524 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
525 Simple_vFAIL2(m, a1); \
526} STMT_END
b45f050a
JF
527
528
529/*
530 * Like Simple_vFAIL(), but accepts three arguments.
531 */
ccb2c380 532#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 533 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
534 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
535 (int)offset, RExC_precomp, RExC_precomp + offset); \
536} STMT_END
b45f050a
JF
537
538/*
539 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
540 */
ccb2c380
MP
541#define vFAIL3(m,a1,a2) STMT_START { \
542 if (!SIZE_ONLY) \
a5e7bc51 543 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
544 Simple_vFAIL3(m, a1, a2); \
545} STMT_END
b45f050a
JF
546
547/*
548 * Like Simple_vFAIL(), but accepts four arguments.
549 */
ccb2c380 550#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 551 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
552 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
553 (int)offset, RExC_precomp, RExC_precomp + offset); \
554} STMT_END
b45f050a 555
95db3ffa
KW
556#define vFAIL4(m,a1,a2,a3) STMT_START { \
557 if (!SIZE_ONLY) \
558 SAVEFREESV(RExC_rx_sv); \
559 Simple_vFAIL4(m, a1, a2, a3); \
560} STMT_END
561
5e0a247b
KW
562/* m is not necessarily a "literal string", in this macro */
563#define reg_warn_non_literal_string(loc, m) STMT_START { \
564 const IV offset = loc - RExC_precomp; \
565 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
566 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
567} STMT_END
568
668c081a 569#define ckWARNreg(loc,m) STMT_START { \
a28509cc 570 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
571 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
572 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
573} STMT_END
574
0d6106aa
KW
575#define vWARN_dep(loc, m) STMT_START { \
576 const IV offset = loc - RExC_precomp; \
577 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
578 (int)offset, RExC_precomp, RExC_precomp + offset); \
579} STMT_END
580
147508a2
KW
581#define ckWARNdep(loc,m) STMT_START { \
582 const IV offset = loc - RExC_precomp; \
583 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
584 m REPORT_LOCATION, \
585 (int)offset, RExC_precomp, RExC_precomp + offset); \
586} STMT_END
587
668c081a 588#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 589 const IV offset = loc - RExC_precomp; \
d1d15184 590 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
591 m REPORT_LOCATION, \
592 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
593} STMT_END
594
b23eb183 595#define ckWARN2reg_d(loc,m, a1) STMT_START { \
2335b3d3 596 const IV offset = loc - RExC_precomp; \
b23eb183 597 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
2335b3d3
KW
598 m REPORT_LOCATION, \
599 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
600} STMT_END
601
668c081a 602#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 603 const IV offset = loc - RExC_precomp; \
668c081a 604 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
605 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
606} STMT_END
607
608#define vWARN3(loc, m, a1, a2) 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, (int)offset, RExC_precomp, RExC_precomp + offset); \
612} STMT_END
613
668c081a
NC
614#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
615 const IV offset = loc - RExC_precomp; \
616 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
617 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
618} STMT_END
619
ccb2c380 620#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 621 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
622 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
623 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
624} STMT_END
625
668c081a
NC
626#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
627 const IV offset = loc - RExC_precomp; \
628 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
629 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
630} STMT_END
631
ccb2c380 632#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 633 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
634 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
635 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
636} STMT_END
9d1d55b5 637
8615cb43 638
cd439c50 639/* Allow for side effects in s */
ccb2c380
MP
640#define REGC(c,s) STMT_START { \
641 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
642} STMT_END
cd439c50 643
fac92740
MJD
644/* Macros for recording node offsets. 20001227 mjd@plover.com
645 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
646 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
647 * Element 0 holds the number n.
07be1b83 648 * Position is 1 indexed.
fac92740 649 */
7122b237
YO
650#ifndef RE_TRACK_PATTERN_OFFSETS
651#define Set_Node_Offset_To_R(node,byte)
652#define Set_Node_Offset(node,byte)
653#define Set_Cur_Node_Offset
654#define Set_Node_Length_To_R(node,len)
655#define Set_Node_Length(node,len)
6a86c6ad 656#define Set_Node_Cur_Length(node,start)
7122b237
YO
657#define Node_Offset(n)
658#define Node_Length(n)
659#define Set_Node_Offset_Length(node,offset,len)
660#define ProgLen(ri) ri->u.proglen
661#define SetProgLen(ri,x) ri->u.proglen = x
662#else
663#define ProgLen(ri) ri->u.offsets[0]
664#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
665#define Set_Node_Offset_To_R(node,byte) STMT_START { \
666 if (! SIZE_ONLY) { \
667 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 668 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 669 if((node) < 0) { \
551405c4 670 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
671 } else { \
672 RExC_offsets[2*(node)-1] = (byte); \
673 } \
674 } \
675} STMT_END
676
677#define Set_Node_Offset(node,byte) \
678 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
679#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
680
681#define Set_Node_Length_To_R(node,len) STMT_START { \
682 if (! SIZE_ONLY) { \
683 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 684 __LINE__, (int)(node), (int)(len))); \
ccb2c380 685 if((node) < 0) { \
551405c4 686 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
687 } else { \
688 RExC_offsets[2*(node)] = (len); \
689 } \
690 } \
691} STMT_END
692
693#define Set_Node_Length(node,len) \
694 Set_Node_Length_To_R((node)-RExC_emit_start, len)
6a86c6ad
NC
695#define Set_Node_Cur_Length(node, start) \
696 Set_Node_Length(node, RExC_parse - start)
fac92740
MJD
697
698/* Get offsets and lengths */
699#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
700#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
701
07be1b83
YO
702#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
703 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
704 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
705} STMT_END
7122b237 706#endif
07be1b83
YO
707
708#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
709#define EXPERIMENTAL_INPLACESCAN
f427392e 710#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 711
304ee84b
YO
712#define DEBUG_STUDYDATA(str,data,depth) \
713DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 714 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
715 "%*s" str "Pos:%"IVdf"/%"IVdf \
716 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
717 (int)(depth)*2, "", \
718 (IV)((data)->pos_min), \
719 (IV)((data)->pos_delta), \
304ee84b 720 (UV)((data)->flags), \
1de06328 721 (IV)((data)->whilem_c), \
304ee84b
YO
722 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
723 is_inf ? "INF " : "" \
1de06328
YO
724 ); \
725 if ((data)->last_found) \
726 PerlIO_printf(Perl_debug_log, \
727 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
728 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
729 SvPVX_const((data)->last_found), \
730 (IV)((data)->last_end), \
731 (IV)((data)->last_start_min), \
732 (IV)((data)->last_start_max), \
733 ((data)->longest && \
734 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
735 SvPVX_const((data)->longest_fixed), \
736 (IV)((data)->offset_fixed), \
737 ((data)->longest && \
738 (data)->longest==&((data)->longest_float)) ? "*" : "", \
739 SvPVX_const((data)->longest_float), \
740 (IV)((data)->offset_float_min), \
741 (IV)((data)->offset_float_max) \
742 ); \
743 PerlIO_printf(Perl_debug_log,"\n"); \
744});
745
653099ff 746/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 747 Update the longest found anchored substring and the longest found
653099ff
GS
748 floating substrings if needed. */
749
4327152a 750STATIC void
ea3daa5d
FC
751S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
752 SSize_t *minlenp, int is_inf)
c277df42 753{
e1ec3a88
AL
754 const STRLEN l = CHR_SVLEN(data->last_found);
755 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 756 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 757
7918f24d
NC
758 PERL_ARGS_ASSERT_SCAN_COMMIT;
759
c277df42 760 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 761 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
762 if (*data->longest == data->longest_fixed) {
763 data->offset_fixed = l ? data->last_start_min : data->pos_min;
764 if (data->flags & SF_BEFORE_EOL)
b81d288d 765 data->flags
c277df42
IZ
766 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
767 else
768 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 769 data->minlen_fixed=minlenp;
1de06328 770 data->lookbehind_fixed=0;
a0ed51b3 771 }
304ee84b 772 else { /* *data->longest == data->longest_float */
c277df42 773 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
774 data->offset_float_max = (l
775 ? data->last_start_max
ea3daa5d
FC
776 : (data->pos_delta == SSize_t_MAX
777 ? SSize_t_MAX
778 : data->pos_min + data->pos_delta));
779 if (is_inf
780 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
781 data->offset_float_max = SSize_t_MAX;
c277df42 782 if (data->flags & SF_BEFORE_EOL)
b81d288d 783 data->flags
c277df42
IZ
784 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
785 else
786 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
787 data->minlen_float=minlenp;
788 data->lookbehind_float=0;
c277df42
IZ
789 }
790 }
791 SvCUR_set(data->last_found, 0);
0eda9292 792 {
a28509cc 793 SV * const sv = data->last_found;
097eb12c
AL
794 if (SvUTF8(sv) && SvMAGICAL(sv)) {
795 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
796 if (mg)
797 mg->mg_len = 0;
798 }
0eda9292 799 }
c277df42
IZ
800 data->last_end = -1;
801 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 802 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
803}
804
899d20b9
KW
805/* These macros set, clear and test whether the synthetic start class ('ssc',
806 * given by the parameter) matches an empty string (EOS). This uses the
807 * 'next_off' field in the node, to save a bit in the flags field. The ssc
808 * stands alone, so there is never a next_off, so this field is otherwise
809 * unused. The EOS information is used only for compilation, but theoretically
810 * it could be passed on to the execution code. This could be used to store
811 * more than one bit of information, but only this one is currently used. */
812#define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
813#define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
814#define TEST_SSC_EOS(node) cBOOL((node)->next_off)
815
653099ff
GS
816/* Can match anything (initialization) */
817STATIC void
3fffb88a 818S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 819{
7918f24d
NC
820 PERL_ARGS_ASSERT_CL_ANYTHING;
821
f8bef550 822 ANYOF_BITMAP_SETALL(cl);
899d20b9
KW
823 cl->flags = ANYOF_UNICODE_ALL;
824 SET_SSC_EOS(cl);
3fffb88a
KW
825
826 /* If any portion of the regex is to operate under locale rules,
827 * initialization includes it. The reason this isn't done for all regexes
828 * is that the optimizer was written under the assumption that locale was
829 * all-or-nothing. Given the complexity and lack of documentation in the
830 * optimizer, and that there are inadequate test cases for locale, so many
831 * parts of it may not work properly, it is safest to avoid locale unless
832 * necessary. */
833 if (RExC_contains_locale) {
9d7a1e63 834 ANYOF_CLASS_SETALL(cl); /* /l uses class */
b25a1036 835 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
3fffb88a 836 }
9d7a1e63
KW
837 else {
838 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
839 }
653099ff
GS
840}
841
842/* Can match anything (initialization) */
843STATIC int
5f66b61c 844S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
845{
846 int value;
847
7918f24d
NC
848 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
849
3732a889 850 for (value = 0; value < ANYOF_MAX; value += 2)
653099ff
GS
851 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
852 return 1;
1aa99e6b
IH
853 if (!(cl->flags & ANYOF_UNICODE_ALL))
854 return 0;
10edeb5d 855 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 856 return 0;
653099ff
GS
857 return 1;
858}
859
860/* Can match anything (initialization) */
861STATIC void
e755fd73 862S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 863{
7918f24d
NC
864 PERL_ARGS_ASSERT_CL_INIT;
865
8ecf7187 866 Zero(cl, 1, struct regnode_charclass_class);
653099ff 867 cl->type = ANYOF;
3fffb88a 868 cl_anything(pRExC_state, cl);
1411dba4 869 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
870}
871
1051e1c4 872/* These two functions currently do the exact same thing */
4d5e5e00 873#define cl_init_zero cl_init
653099ff 874
dd58aee1
KW
875/* 'AND' a given class with another one. Can create false positives. 'cl'
876 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
877 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 878STATIC void
5f66b61c 879S_cl_and(struct regnode_charclass_class *cl,
a28509cc 880 const struct regnode_charclass_class *and_with)
653099ff 881{
7918f24d 882 PERL_ARGS_ASSERT_CL_AND;
40d049e4 883
954a2af6 884 assert(PL_regkind[and_with->type] == ANYOF);
1e6ade67 885
c6b76537 886 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
887 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
888 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 889 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
890 && !(and_with->flags & ANYOF_LOC_FOLD)
891 && !(cl->flags & ANYOF_LOC_FOLD)) {
653099ff
GS
892 int i;
893
894 if (and_with->flags & ANYOF_INVERT)
895 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
896 cl->bitmap[i] &= ~and_with->bitmap[i];
897 else
898 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
899 cl->bitmap[i] &= and_with->bitmap[i];
900 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 901
c6b76537 902 if (and_with->flags & ANYOF_INVERT) {
8951c461 903
c6b76537
KW
904 /* Here, the and'ed node is inverted. Get the AND of the flags that
905 * aren't affected by the inversion. Those that are affected are
906 * handled individually below */
907 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
908 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
909 cl->flags |= affected_flags;
910
911 /* We currently don't know how to deal with things that aren't in the
912 * bitmap, but we know that the intersection is no greater than what
913 * is already in cl, so let there be false positives that get sorted
914 * out after the synthetic start class succeeds, and the node is
915 * matched for real. */
916
917 /* The inversion of these two flags indicate that the resulting
918 * intersection doesn't have them */
919 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
920 cl->flags &= ~ANYOF_UNICODE_ALL;
921 }
c6b76537
KW
922 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
923 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 924 }
1aa99e6b 925 }
c6b76537 926 else { /* and'd node is not inverted */
3ad98780
KW
927 U8 outside_bitmap_but_not_utf8; /* Temp variable */
928
137165a6 929 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
930
931 /* Here 'and_with' doesn't match anything outside the bitmap
932 * (except possibly ANYOF_UNICODE_ALL), which means the
933 * intersection can't either, except for ANYOF_UNICODE_ALL, in
934 * which case we don't know what the intersection is, but it's no
935 * greater than what cl already has, so can just leave it alone,
936 * with possible false positives */
937 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
938 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 939 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 940 }
137165a6 941 }
c6b76537
KW
942 else if (! ANYOF_NONBITMAP(cl)) {
943
944 /* Here, 'and_with' does match something outside the bitmap, and cl
945 * doesn't have a list of things to match outside the bitmap. If
946 * cl can match all code points above 255, the intersection will
3ad98780
KW
947 * be those above-255 code points that 'and_with' matches. If cl
948 * can't match all Unicode code points, it means that it can't
949 * match anything outside the bitmap (since the 'if' that got us
950 * into this block tested for that), so we leave the bitmap empty.
951 */
c6b76537
KW
952 if (cl->flags & ANYOF_UNICODE_ALL) {
953 ARG_SET(cl, ARG(and_with));
3ad98780
KW
954
955 /* and_with's ARG may match things that don't require UTF8.
956 * And now cl's will too, in spite of this being an 'and'. See
957 * the comments below about the kludge */
958 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
959 }
960 }
961 else {
962 /* Here, both 'and_with' and cl match something outside the
963 * bitmap. Currently we do not do the intersection, so just match
964 * whatever cl had at the beginning. */
965 }
966
967
3ad98780
KW
968 /* Take the intersection of the two sets of flags. However, the
969 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
970 * kludge around the fact that this flag is not treated like the others
971 * which are initialized in cl_anything(). The way the optimizer works
972 * is that the synthetic start class (SSC) is initialized to match
973 * anything, and then the first time a real node is encountered, its
974 * values are AND'd with the SSC's with the result being the values of
975 * the real node. However, there are paths through the optimizer where
976 * the AND never gets called, so those initialized bits are set
977 * inappropriately, which is not usually a big deal, as they just cause
978 * false positives in the SSC, which will just mean a probably
979 * imperceptible slow down in execution. However this bit has a
980 * higher false positive consequence in that it can cause utf8.pm,
981 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
982 * bigger slowdown and also causes significant extra memory to be used.
983 * In order to prevent this, the code now takes a different tack. The
984 * bit isn't set unless some part of the regular expression needs it,
985 * but once set it won't get cleared. This means that these extra
986 * modules won't get loaded unless there was some path through the
987 * pattern that would have required them anyway, and so any false
988 * positives that occur by not ANDing them out when they could be
989 * aren't as severe as they would be if we treated this bit like all
990 * the others */
991 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
992 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 993 cl->flags &= and_with->flags;
3ad98780 994 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 995 }
653099ff
GS
996}
997
dd58aee1
KW
998/* 'OR' a given class with another one. Can create false positives. 'cl'
999 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
1000 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 1001STATIC void
3fffb88a 1002S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 1003{
7918f24d
NC
1004 PERL_ARGS_ASSERT_CL_OR;
1005
653099ff 1006 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
1007
1008 /* Here, the or'd node is to be inverted. This means we take the
1009 * complement of everything not in the bitmap, but currently we don't
1010 * know what that is, so give up and match anything */
1011 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 1012 cl_anything(pRExC_state, cl);
c6b76537 1013 }
653099ff
GS
1014 /* We do not use
1015 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
1016 * <= (B1 | !B2) | (CL1 | !CL2)
1017 * which is wasteful if CL2 is small, but we ignore CL2:
1018 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
1019 * XXXX Can we handle case-fold? Unclear:
1020 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
1021 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
1022 */
c6b76537 1023 else 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 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1029 cl->bitmap[i] |= ~or_with->bitmap[i];
1030 } /* XXXX: logic is complicated otherwise */
1031 else {
3fffb88a 1032 cl_anything(pRExC_state, cl);
653099ff 1033 }
c6b76537
KW
1034
1035 /* And, we can just take the union of the flags that aren't affected
1036 * by the inversion */
1037 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
1038
1039 /* For the remaining flags:
1040 ANYOF_UNICODE_ALL and inverted means to not match anything above
1041 255, which means that the union with cl should just be
1042 what cl has in it, so can ignore this flag
1043 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
f1658f69
KW
1044 is (ASCII) 127-255 to match them, but then invert that, so
1045 the union with cl should just be what cl has in it, so can
c6b76537
KW
1046 ignore this flag
1047 */
1048 } else { /* 'or_with' is not inverted */
653099ff
GS
1049 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1050 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
1051 && (!(or_with->flags & ANYOF_LOC_FOLD)
1052 || (cl->flags & ANYOF_LOC_FOLD)) ) {
653099ff
GS
1053 int i;
1054
1055 /* OR char bitmap and class bitmap separately */
1056 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1057 cl->bitmap[i] |= or_with->bitmap[i];
c96939e4
KW
1058 if (or_with->flags & ANYOF_CLASS) {
1059 ANYOF_CLASS_OR(or_with, cl);
1060 }
653099ff
GS
1061 }
1062 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 1063 cl_anything(pRExC_state, cl);
653099ff 1064 }
9826f543 1065
c6b76537
KW
1066 if (ANYOF_NONBITMAP(or_with)) {
1067
1068 /* Use the added node's outside-the-bit-map match if there isn't a
1069 * conflict. If there is a conflict (both nodes match something
1070 * outside the bitmap, but what they match outside is not the same
1071 * pointer, and hence not easily compared until XXX we extend
1072 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
1073 * match everything outside the bitmap. If that stuff is all above
1074 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
1075 if (! ANYOF_NONBITMAP(cl)) {
1076 ARG_SET(cl, ARG(or_with));
1077 }
1078 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
1079
1080 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1081 cl_anything(pRExC_state, cl);
1082 }
1083 else {
1084 cl->flags |= ANYOF_UNICODE_ALL;
1085 }
c6b76537 1086 }
4c34a693 1087 }
0b9668ee
KW
1088
1089 /* Take the union */
1090 cl->flags |= or_with->flags;
1aa99e6b 1091 }
653099ff
GS
1092}
1093
a3621e74
YO
1094#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1095#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1096#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1097#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1098
3dab1dad
YO
1099
1100#ifdef DEBUGGING
07be1b83 1101/*
2b8b4781
NC
1102 dump_trie(trie,widecharmap,revcharmap)
1103 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1104 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1105
1106 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1107 The _interim_ variants are used for debugging the interim
1108 tables that are used to generate the final compressed
1109 representation which is what dump_trie expects.
1110
486ec47a 1111 Part of the reason for their existence is to provide a form
3dab1dad 1112 of documentation as to how the different representations function.
07be1b83
YO
1113
1114*/
3dab1dad
YO
1115
1116/*
3dab1dad
YO
1117 Dumps the final compressed table form of the trie to Perl_debug_log.
1118 Used for debugging make_trie().
1119*/
b9a59e08 1120
3dab1dad 1121STATIC void
2b8b4781
NC
1122S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1123 AV *revcharmap, U32 depth)
3dab1dad
YO
1124{
1125 U32 state;
ab3bbdeb 1126 SV *sv=sv_newmortal();
55eed653 1127 int colwidth= widecharmap ? 6 : 4;
2e64971a 1128 U16 word;
3dab1dad
YO
1129 GET_RE_DEBUG_FLAGS_DECL;
1130
7918f24d 1131 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1132
3dab1dad
YO
1133 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1134 (int)depth * 2 + 2,"",
1135 "Match","Base","Ofs" );
1136
1137 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1138 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1139 if ( tmp ) {
ab3bbdeb
YO
1140 PerlIO_printf( Perl_debug_log, "%*s",
1141 colwidth,
ddc5bc0f 1142 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1143 PL_colors[0], PL_colors[1],
1144 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1145 PERL_PV_ESCAPE_FIRSTCHAR
1146 )
1147 );
3dab1dad
YO
1148 }
1149 }
1150 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1151 (int)depth * 2 + 2,"");
1152
1153 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1154 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1155 PerlIO_printf( Perl_debug_log, "\n");
1156
1e2e3d02 1157 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1158 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1159
1160 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1161
1162 if ( trie->states[ state ].wordnum ) {
1163 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1164 } else {
1165 PerlIO_printf( Perl_debug_log, "%6s", "" );
1166 }
1167
1168 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1169
1170 if ( base ) {
1171 U32 ofs = 0;
1172
1173 while( ( base + ofs < trie->uniquecharcount ) ||
1174 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1175 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1176 ofs++;
1177
1178 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1179
1180 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1181 if ( ( base + ofs >= trie->uniquecharcount ) &&
1182 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1183 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1184 {
ab3bbdeb
YO
1185 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1186 colwidth,
3dab1dad
YO
1187 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1188 } else {
ab3bbdeb 1189 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1190 }
1191 }
1192
1193 PerlIO_printf( Perl_debug_log, "]");
1194
1195 }
1196 PerlIO_printf( Perl_debug_log, "\n" );
1197 }
2e64971a
DM
1198 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1199 for (word=1; word <= trie->wordcount; word++) {
1200 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1201 (int)word, (int)(trie->wordinfo[word].prev),
1202 (int)(trie->wordinfo[word].len));
1203 }
1204 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1205}
1206/*
3dab1dad
YO
1207 Dumps a fully constructed but uncompressed trie in list form.
1208 List tries normally only are used for construction when the number of
1209 possible chars (trie->uniquecharcount) is very high.
1210 Used for debugging make_trie().
1211*/
1212STATIC void
55eed653 1213S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1214 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1215 U32 depth)
3dab1dad
YO
1216{
1217 U32 state;
ab3bbdeb 1218 SV *sv=sv_newmortal();
55eed653 1219 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1220 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1221
1222 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1223
3dab1dad 1224 /* print out the table precompression. */
ab3bbdeb
YO
1225 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1226 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1227 "------:-----+-----------------\n" );
3dab1dad
YO
1228
1229 for( state=1 ; state < next_alloc ; state ++ ) {
1230 U16 charid;
1231
ab3bbdeb 1232 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1233 (int)depth * 2 + 2,"", (UV)state );
1234 if ( ! trie->states[ state ].wordnum ) {
1235 PerlIO_printf( Perl_debug_log, "%5s| ","");
1236 } else {
1237 PerlIO_printf( Perl_debug_log, "W%4x| ",
1238 trie->states[ state ].wordnum
1239 );
1240 }
1241 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1242 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1243 if ( tmp ) {
1244 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1245 colwidth,
ddc5bc0f 1246 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1247 PL_colors[0], PL_colors[1],
1248 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1249 PERL_PV_ESCAPE_FIRSTCHAR
1250 ) ,
1e2e3d02
YO
1251 TRIE_LIST_ITEM(state,charid).forid,
1252 (UV)TRIE_LIST_ITEM(state,charid).newstate
1253 );
1254 if (!(charid % 10))
664e119d
RGS
1255 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1256 (int)((depth * 2) + 14), "");
1e2e3d02 1257 }
ab3bbdeb
YO
1258 }
1259 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1260 }
1261}
1262
1263/*
3dab1dad
YO
1264 Dumps a fully constructed but uncompressed trie in table form.
1265 This is the normal DFA style state transition table, with a few
1266 twists to facilitate compression later.
1267 Used for debugging make_trie().
1268*/
1269STATIC void
55eed653 1270S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1271 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1272 U32 depth)
3dab1dad
YO
1273{
1274 U32 state;
1275 U16 charid;
ab3bbdeb 1276 SV *sv=sv_newmortal();
55eed653 1277 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1278 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1279
1280 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1281
1282 /*
1283 print out the table precompression so that we can do a visual check
1284 that they are identical.
1285 */
1286
1287 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1288
1289 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1290 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1291 if ( tmp ) {
ab3bbdeb
YO
1292 PerlIO_printf( Perl_debug_log, "%*s",
1293 colwidth,
ddc5bc0f 1294 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1295 PL_colors[0], PL_colors[1],
1296 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1297 PERL_PV_ESCAPE_FIRSTCHAR
1298 )
1299 );
3dab1dad
YO
1300 }
1301 }
1302
1303 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1304
1305 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1306 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1307 }
1308
1309 PerlIO_printf( Perl_debug_log, "\n" );
1310
1311 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1312
1313 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1314 (int)depth * 2 + 2,"",
1315 (UV)TRIE_NODENUM( state ) );
1316
1317 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1318 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1319 if (v)
1320 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1321 else
1322 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1323 }
1324 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1325 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1326 } else {
1327 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1328 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1329 }
1330 }
07be1b83 1331}
3dab1dad
YO
1332
1333#endif
1334
2e64971a 1335
786e8c11
YO
1336/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1337 startbranch: the first branch in the whole branch sequence
1338 first : start branch of sequence of branch-exact nodes.
1339 May be the same as startbranch
1340 last : Thing following the last branch.
1341 May be the same as tail.
1342 tail : item following the branch sequence
1343 count : words in the sequence
1344 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1345 depth : indent depth
3dab1dad 1346
786e8c11 1347Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1348
786e8c11
YO
1349A trie is an N'ary tree where the branches are determined by digital
1350decomposition of the key. IE, at the root node you look up the 1st character and
1351follow that branch repeat until you find the end of the branches. Nodes can be
1352marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1353
786e8c11 1354 /he|she|his|hers/
72f13be8 1355
786e8c11
YO
1356would convert into the following structure. Numbers represent states, letters
1357following numbers represent valid transitions on the letter from that state, if
1358the number is in square brackets it represents an accepting state, otherwise it
1359will be in parenthesis.
07be1b83 1360
786e8c11
YO
1361 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1362 | |
1363 | (2)
1364 | |
1365 (1) +-i->(6)-+-s->[7]
1366 |
1367 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1368
786e8c11
YO
1369 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1370
1371This shows that when matching against the string 'hers' we will begin at state 1
1372read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1373then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1374is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1375single traverse. We store a mapping from accepting to state to which word was
1376matched, and then when we have multiple possibilities we try to complete the
1377rest of the regex in the order in which they occured in the alternation.
1378
1379The only prior NFA like behaviour that would be changed by the TRIE support is
1380the silent ignoring of duplicate alternations which are of the form:
1381
1382 / (DUPE|DUPE) X? (?{ ... }) Y /x
1383
4b714af6 1384Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1385and without the optimisation. With the optimisations dupes will be silently
486ec47a 1386ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1387the following demonstrates:
1388
1389 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1390
1391which prints out 'word' three times, but
1392
1393 'words'=~/(word|word|word)(?{ print $1 })S/
1394
1395which doesnt print it out at all. This is due to other optimisations kicking in.
1396
1397Example of what happens on a structural level:
1398
486ec47a 1399The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1400
1401 1: CURLYM[1] {1,32767}(18)
1402 5: BRANCH(8)
1403 6: EXACT <ac>(16)
1404 8: BRANCH(11)
1405 9: EXACT <ad>(16)
1406 11: BRANCH(14)
1407 12: EXACT <ab>(16)
1408 16: SUCCEED(0)
1409 17: NOTHING(18)
1410 18: END(0)
1411
1412This would be optimizable with startbranch=5, first=5, last=16, tail=16
1413and should turn into:
1414
1415 1: CURLYM[1] {1,32767}(18)
1416 5: TRIE(16)
1417 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1418 <ac>
1419 <ad>
1420 <ab>
1421 16: SUCCEED(0)
1422 17: NOTHING(18)
1423 18: END(0)
1424
1425Cases where tail != last would be like /(?foo|bar)baz/:
1426
1427 1: BRANCH(4)
1428 2: EXACT <foo>(8)
1429 4: BRANCH(7)
1430 5: EXACT <bar>(8)
1431 7: TAIL(8)
1432 8: EXACT <baz>(10)
1433 10: END(0)
1434
1435which would be optimizable with startbranch=1, first=1, last=7, tail=8
1436and would end up looking like:
1437
1438 1: TRIE(8)
1439 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1440 <foo>
1441 <bar>
1442 7: TAIL(8)
1443 8: EXACT <baz>(10)
1444 10: END(0)
1445
c80e42f3 1446 d = uvchr_to_utf8_flags(d, uv, 0);
786e8c11
YO
1447
1448is the recommended Unicode-aware way of saying
1449
1450 *(d++) = uv;
1451*/
1452
fab2782b 1453#define TRIE_STORE_REVCHAR(val) \
786e8c11 1454 STMT_START { \
73031816 1455 if (UTF) { \
fab2782b 1456 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1457 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
c80e42f3 1458 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
73031816
NC
1459 SvCUR_set(zlopp, kapow - flrbbbbb); \
1460 SvPOK_on(zlopp); \
1461 SvUTF8_on(zlopp); \
1462 av_push(revcharmap, zlopp); \
1463 } else { \
fab2782b 1464 char ooooff = (char)val; \
73031816
NC
1465 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1466 } \
1467 } STMT_END
786e8c11 1468
914a25d5
KW
1469/* This gets the next character from the input, folding it if not already
1470 * folded. */
1471#define TRIE_READ_CHAR STMT_START { \
1472 wordlen++; \
1473 if ( UTF ) { \
1474 /* if it is UTF then it is either already folded, or does not need \
1475 * folding */ \
1476 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1477 } \
1478 else if (folder == PL_fold_latin1) { \
7d006b13
KW
1479 /* This folder implies Unicode rules, which in the range expressible \
1480 * by not UTF is the lower case, with the two exceptions, one of \
1481 * which should have been taken care of before calling this */ \
1482 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
1483 uvc = toLOWER_L1(*uc); \
1484 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
1485 len = 1; \
914a25d5
KW
1486 } else { \
1487 /* raw data, will be folded later if needed */ \
1488 uvc = (U32)*uc; \
1489 len = 1; \
1490 } \
786e8c11
YO
1491} STMT_END
1492
1493
1494
1495#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1496 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1497 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1498 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1499 } \
1500 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1501 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1502 TRIE_LIST_CUR( state )++; \
1503} STMT_END
07be1b83 1504
786e8c11
YO
1505#define TRIE_LIST_NEW(state) STMT_START { \
1506 Newxz( trie->states[ state ].trans.list, \
1507 4, reg_trie_trans_le ); \
1508 TRIE_LIST_CUR( state ) = 1; \
1509 TRIE_LIST_LEN( state ) = 4; \
1510} STMT_END
07be1b83 1511
786e8c11
YO
1512#define TRIE_HANDLE_WORD(state) STMT_START { \
1513 U16 dupe= trie->states[ state ].wordnum; \
1514 regnode * const noper_next = regnext( noper ); \
1515 \
786e8c11
YO
1516 DEBUG_r({ \
1517 /* store the word for dumping */ \
1518 SV* tmp; \
1519 if (OP(noper) != NOTHING) \
740cce10 1520 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1521 else \
740cce10 1522 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1523 av_push( trie_words, tmp ); \
786e8c11
YO
1524 }); \
1525 \
1526 curword++; \
2e64971a
DM
1527 trie->wordinfo[curword].prev = 0; \
1528 trie->wordinfo[curword].len = wordlen; \
1529 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1530 \
1531 if ( noper_next < tail ) { \
1532 if (!trie->jump) \
c944940b 1533 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1534 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1535 if (!jumper) \
1536 jumper = noper_next; \
1537 if (!nextbranch) \
1538 nextbranch= regnext(cur); \
1539 } \
1540 \
1541 if ( dupe ) { \
2e64971a
DM
1542 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1543 /* chain, so that when the bits of chain are later */\
1544 /* linked together, the dups appear in the chain */\
1545 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1546 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1547 } else { \
1548 /* we haven't inserted this word yet. */ \
1549 trie->states[ state ].wordnum = curword; \
1550 } \
1551} STMT_END
07be1b83 1552
3dab1dad 1553
786e8c11
YO
1554#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1555 ( ( base + charid >= ucharcount \
1556 && base + charid < ubound \
1557 && state == trie->trans[ base - ucharcount + charid ].check \
1558 && trie->trans[ base - ucharcount + charid ].next ) \
1559 ? trie->trans[ base - ucharcount + charid ].next \
1560 : ( state==1 ? special : 0 ) \
1561 )
3dab1dad 1562
786e8c11
YO
1563#define MADE_TRIE 1
1564#define MADE_JUMP_TRIE 2
1565#define MADE_EXACT_TRIE 4
3dab1dad 1566
a3621e74 1567STATIC I32
786e8c11 1568S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1569{
27da23d5 1570 dVAR;
a3621e74
YO
1571 /* first pass, loop through and scan words */
1572 reg_trie_data *trie;
55eed653 1573 HV *widecharmap = NULL;
2b8b4781 1574 AV *revcharmap = newAV();
a3621e74 1575 regnode *cur;
9f7f3913 1576 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1577 STRLEN len = 0;
1578 UV uvc = 0;
1579 U16 curword = 0;
1580 U32 next_alloc = 0;
786e8c11
YO
1581 regnode *jumper = NULL;
1582 regnode *nextbranch = NULL;
7f69552c 1583 regnode *convert = NULL;
2e64971a 1584 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1585 /* we just use folder as a flag in utf8 */
1e696034 1586 const U8 * folder = NULL;
a3621e74 1587
2b8b4781
NC
1588#ifdef DEBUGGING
1589 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1590 AV *trie_words = NULL;
1591 /* along with revcharmap, this only used during construction but both are
1592 * useful during debugging so we store them in the struct when debugging.
8e11feef 1593 */
2b8b4781
NC
1594#else
1595 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1596 STRLEN trie_charcount=0;
3dab1dad 1597#endif
2b8b4781 1598 SV *re_trie_maxbuff;
a3621e74 1599 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1600
1601 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1602#ifndef DEBUGGING
1603 PERL_UNUSED_ARG(depth);
1604#endif
a3621e74 1605
1e696034 1606 switch (flags) {
79a81a6e 1607 case EXACT: break;
2f7f8cb1 1608 case EXACTFA:
fab2782b
YO
1609 case EXACTFU_SS:
1610 case EXACTFU_TRICKYFOLD:
1e696034
KW
1611 case EXACTFU: folder = PL_fold_latin1; break;
1612 case EXACTF: folder = PL_fold; break;
1613 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1614 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1615 }
1616
c944940b 1617 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1618 trie->refcount = 1;
3dab1dad 1619 trie->startstate = 1;
786e8c11 1620 trie->wordcount = word_count;
f8fc2ecf 1621 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1622 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1623 if (flags == EXACT)
c944940b 1624 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1625 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1626 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1627
a3621e74 1628 DEBUG_r({
2b8b4781 1629 trie_words = newAV();
a3621e74 1630 });
a3621e74 1631
0111c4fd 1632 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1633 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1634 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1635 }
df826430 1636 DEBUG_TRIE_COMPILE_r({
3dab1dad 1637 PerlIO_printf( Perl_debug_log,
786e8c11 1638 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1639 (int)depth * 2 + 2, "",
1640 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1641 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1642 (int)depth);
3dab1dad 1643 });
7f69552c
YO
1644
1645 /* Find the node we are going to overwrite */
1646 if ( first == startbranch && OP( last ) != BRANCH ) {
1647 /* whole branch chain */
1648 convert = first;
1649 } else {
1650 /* branch sub-chain */
1651 convert = NEXTOPER( first );
1652 }
1653
a3621e74
YO
1654 /* -- First loop and Setup --
1655
1656 We first traverse the branches and scan each word to determine if it
1657 contains widechars, and how many unique chars there are, this is
1658 important as we have to build a table with at least as many columns as we
1659 have unique chars.
1660
1661 We use an array of integers to represent the character codes 0..255
38a44b82 1662 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1663 native representation of the character value as the key and IV's for the
1664 coded index.
1665
1666 *TODO* If we keep track of how many times each character is used we can
1667 remap the columns so that the table compression later on is more
3b753521 1668 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1669 middle and the least common are on the outside. IMO this would be better
1670 than a most to least common mapping as theres a decent chance the most
1671 common letter will share a node with the least common, meaning the node
486ec47a 1672 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1673 case is when we have the least common nodes twice.
1674
1675 */
1676
a3621e74 1677 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1678 regnode *noper = NEXTOPER( cur );
e1ec3a88 1679 const U8 *uc = (U8*)STRING( noper );
df826430 1680 const U8 *e = uc + STR_LEN( noper );
a3621e74 1681 STRLEN foldlen = 0;
07be1b83 1682 U32 wordlen = 0; /* required init */
afa96d92
KW
1683 STRLEN minbytes = 0;
1684 STRLEN maxbytes = 0;
02daf0ab 1685 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1686
3dab1dad 1687 if (OP(noper) == NOTHING) {
df826430
YO
1688 regnode *noper_next= regnext(noper);
1689 if (noper_next != tail && OP(noper_next) == flags) {
1690 noper = noper_next;
1691 uc= (U8*)STRING(noper);
1692 e= uc + STR_LEN(noper);
1693 trie->minlen= STR_LEN(noper);
1694 } else {
1695 trie->minlen= 0;
1696 continue;
1697 }
3dab1dad 1698 }
df826430 1699
fab2782b 1700 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1701 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1702 regardless of encoding */
fab2782b
YO
1703 if (OP( noper ) == EXACTFU_SS) {
1704 /* false positives are ok, so just set this */
0dc4a61d 1705 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
fab2782b
YO
1706 }
1707 }
a3621e74 1708 for ( ; uc < e ; uc += len ) {
3dab1dad 1709 TRIE_CHARCOUNT(trie)++;
a3621e74 1710 TRIE_READ_CHAR;
afa96d92
KW
1711 minbytes++;
1712 maxbytes++;
a3621e74 1713 if ( uvc < 256 ) {
fab2782b
YO
1714 if ( folder ) {
1715 U8 folded= folder[ (U8) uvc ];
1716 if ( !trie->charmap[ folded ] ) {
1717 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1718 TRIE_STORE_REVCHAR( folded );
1719 }
1720 }
a3621e74
YO
1721 if ( !trie->charmap[ uvc ] ) {
1722 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1723 TRIE_STORE_REVCHAR( uvc );
a3621e74 1724 }
02daf0ab 1725 if ( set_bit ) {
62012aee
KW
1726 /* store the codepoint in the bitmap, and its folded
1727 * equivalent. */
fab2782b 1728 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1729
1730 /* store the folded codepoint */
fab2782b 1731 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1732
1733 if ( !UTF ) {
1734 /* store first byte of utf8 representation of
acdf4139 1735 variant codepoints */
f4cd282c 1736 if (! NATIVE_IS_INVARIANT(uvc)) {
acdf4139 1737 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1738 }
1739 }
02daf0ab
YO
1740 set_bit = 0; /* We've done our bit :-) */
1741 }
a3621e74
YO
1742 } else {
1743 SV** svpp;
55eed653
NC
1744 if ( !widecharmap )
1745 widecharmap = newHV();
a3621e74 1746
55eed653 1747 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1748
1749 if ( !svpp )
e4584336 1750 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1751
1752 if ( !SvTRUE( *svpp ) ) {
1753 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1754 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1755 }
1756 }
1757 }
3dab1dad 1758 if( cur == first ) {
afa96d92
KW
1759 trie->minlen = minbytes;
1760 trie->maxlen = maxbytes;
1761 } else if (minbytes < trie->minlen) {
1762 trie->minlen = minbytes;
1763 } else if (maxbytes > trie->maxlen) {
1764 trie->maxlen = maxbytes;
fab2782b
YO
1765 }
1766 if (OP( noper ) == EXACTFU_SS) {
1767 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1768 if (trie->minlen > 1)
1769 trie->minlen= 1;
1770 }
1771 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1772 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1773 * - We assume that any such sequence might match a 2 byte string */
1774 if (trie->minlen > 2 )
1775 trie->minlen= 2;
3dab1dad
YO
1776 }
1777
a3621e74
YO
1778 } /* end first pass */
1779 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1780 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1781 (int)depth * 2 + 2,"",
55eed653 1782 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1783 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1784 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1785 );
a3621e74
YO
1786
1787 /*
1788 We now know what we are dealing with in terms of unique chars and
1789 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1790 representation using a flat table will take. If it's over a reasonable
1791 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1792 conservative but potentially much slower representation using an array
1793 of lists.
1794
1795 At the end we convert both representations into the same compressed
1796 form that will be used in regexec.c for matching with. The latter
1797 is a form that cannot be used to construct with but has memory
1798 properties similar to the list form and access properties similar
1799 to the table form making it both suitable for fast searches and
1800 small enough that its feasable to store for the duration of a program.
1801
1802 See the comment in the code where the compressed table is produced
1803 inplace from the flat tabe representation for an explanation of how
1804 the compression works.
1805
1806 */
1807
1808
2e64971a
DM
1809 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1810 prev_states[1] = 0;
1811
3dab1dad 1812 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1813 /*
1814 Second Pass -- Array Of Lists Representation
1815
1816 Each state will be represented by a list of charid:state records
1817 (reg_trie_trans_le) the first such element holds the CUR and LEN
1818 points of the allocated array. (See defines above).
1819
1820 We build the initial structure using the lists, and then convert
1821 it into the compressed table form which allows faster lookups
1822 (but cant be modified once converted).
a3621e74
YO
1823 */
1824
a3621e74
YO
1825 STRLEN transcount = 1;
1826
1e2e3d02
YO
1827 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1828 "%*sCompiling trie using list compiler\n",
1829 (int)depth * 2 + 2, ""));
686b73d4 1830
c944940b
JH
1831 trie->states = (reg_trie_state *)
1832 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1833 sizeof(reg_trie_state) );
a3621e74
YO
1834 TRIE_LIST_NEW(1);
1835 next_alloc = 2;
1836
1837 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1838
df826430 1839 regnode *noper = NEXTOPER( cur );
c445ea15 1840 U8 *uc = (U8*)STRING( noper );
df826430 1841 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1842 U32 state = 1; /* required init */
1843 U16 charid = 0; /* sanity init */
07be1b83 1844 U32 wordlen = 0; /* required init */
c445ea15 1845
df826430
YO
1846 if (OP(noper) == NOTHING) {
1847 regnode *noper_next= regnext(noper);
1848 if (noper_next != tail && OP(noper_next) == flags) {
1849 noper = noper_next;
1850 uc= (U8*)STRING(noper);
1851 e= uc + STR_LEN(noper);
1852 }
1853 }
1854
3dab1dad 1855 if (OP(noper) != NOTHING) {
786e8c11 1856 for ( ; uc < e ; uc += len ) {
c445ea15 1857
786e8c11 1858 TRIE_READ_CHAR;
c445ea15 1859
786e8c11
YO
1860 if ( uvc < 256 ) {
1861 charid = trie->charmap[ uvc ];
c445ea15 1862 } else {
55eed653 1863 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1864 if ( !svpp ) {
1865 charid = 0;
1866 } else {
1867 charid=(U16)SvIV( *svpp );
1868 }
c445ea15 1869 }
786e8c11
YO
1870 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1871 if ( charid ) {
a3621e74 1872
786e8c11
YO
1873 U16 check;
1874 U32 newstate = 0;
a3621e74 1875
786e8c11
YO
1876 charid--;
1877 if ( !trie->states[ state ].trans.list ) {
1878 TRIE_LIST_NEW( state );
c445ea15 1879 }
786e8c11
YO
1880 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1881 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1882 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1883 break;
1884 }
1885 }
1886 if ( ! newstate ) {
1887 newstate = next_alloc++;
2e64971a 1888 prev_states[newstate] = state;
786e8c11
YO
1889 TRIE_LIST_PUSH( state, charid, newstate );
1890 transcount++;
1891 }
1892 state = newstate;
1893 } else {
1894 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1895 }
a28509cc 1896 }
c445ea15 1897 }
3dab1dad 1898 TRIE_HANDLE_WORD(state);
a3621e74
YO
1899
1900 } /* end second pass */
1901
1e2e3d02
YO
1902 /* next alloc is the NEXT state to be allocated */
1903 trie->statecount = next_alloc;
c944940b
JH
1904 trie->states = (reg_trie_state *)
1905 PerlMemShared_realloc( trie->states,
1906 next_alloc
1907 * sizeof(reg_trie_state) );
a3621e74 1908
3dab1dad 1909 /* and now dump it out before we compress it */
2b8b4781
NC
1910 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1911 revcharmap, next_alloc,
1912 depth+1)
1e2e3d02 1913 );
a3621e74 1914
c944940b
JH
1915 trie->trans = (reg_trie_trans *)
1916 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1917 {
1918 U32 state;
a3621e74
YO
1919 U32 tp = 0;
1920 U32 zp = 0;
1921
1922
1923 for( state=1 ; state < next_alloc ; state ++ ) {
1924 U32 base=0;
1925
1926 /*
1927 DEBUG_TRIE_COMPILE_MORE_r(
1928 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1929 );
1930 */
1931
1932 if (trie->states[state].trans.list) {
1933 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1934 U16 maxid=minid;
a28509cc 1935 U16 idx;
a3621e74
YO
1936
1937 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1938 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1939 if ( forid < minid ) {
1940 minid=forid;
1941 } else if ( forid > maxid ) {
1942 maxid=forid;
1943 }
a3621e74
YO
1944 }
1945 if ( transcount < tp + maxid - minid + 1) {
1946 transcount *= 2;
c944940b
JH
1947 trie->trans = (reg_trie_trans *)
1948 PerlMemShared_realloc( trie->trans,
446bd890
NC
1949 transcount
1950 * sizeof(reg_trie_trans) );
a3621e74
YO
1951 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1952 }
1953 base = trie->uniquecharcount + tp - minid;
1954 if ( maxid == minid ) {
1955 U32 set = 0;
1956 for ( ; zp < tp ; zp++ ) {
1957 if ( ! trie->trans[ zp ].next ) {
1958 base = trie->uniquecharcount + zp - minid;
1959 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1960 trie->trans[ zp ].check = state;
1961 set = 1;
1962 break;
1963 }
1964 }
1965 if ( !set ) {
1966 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1967 trie->trans[ tp ].check = state;
1968 tp++;
1969 zp = tp;
1970 }
1971 } else {
1972 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1973 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1974 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1975 trie->trans[ tid ].check = state;
1976 }
1977 tp += ( maxid - minid + 1 );
1978 }
1979 Safefree(trie->states[ state ].trans.list);
1980 }
1981 /*
1982 DEBUG_TRIE_COMPILE_MORE_r(
1983 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1984 );
1985 */
1986 trie->states[ state ].trans.base=base;
1987 }
cc601c31 1988 trie->lasttrans = tp + 1;
a3621e74
YO
1989 }
1990 } else {
1991 /*
1992 Second Pass -- Flat Table Representation.
1993
1994 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1995 We know that we will need Charcount+1 trans at most to store the data
1996 (one row per char at worst case) So we preallocate both structures
1997 assuming worst case.
1998
1999 We then construct the trie using only the .next slots of the entry
2000 structs.
2001
3b753521 2002 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
2003 make compression both faster and easier by keeping track of how many non
2004 zero fields are in the node.
2005
2006 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2007 transition.
2008
2009 There are two terms at use here: state as a TRIE_NODEIDX() which is a
2010 number representing the first entry of the node, and state as a
2011 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
2012 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
2013 are 2 entrys per node. eg:
2014
2015 A B A B
2016 1. 2 4 1. 3 7
2017 2. 0 3 3. 0 5
2018 3. 0 0 5. 0 0
2019 4. 0 0 7. 0 0
2020
2021 The table is internally in the right hand, idx form. However as we also
2022 have to deal with the states array which is indexed by nodenum we have to
2023 use TRIE_NODENUM() to convert.
2024
2025 */
1e2e3d02
YO
2026 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2027 "%*sCompiling trie using table compiler\n",
2028 (int)depth * 2 + 2, ""));
3dab1dad 2029
c944940b
JH
2030 trie->trans = (reg_trie_trans *)
2031 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2032 * trie->uniquecharcount + 1,
2033 sizeof(reg_trie_trans) );
2034 trie->states = (reg_trie_state *)
2035 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2036 sizeof(reg_trie_state) );
a3621e74
YO
2037 next_alloc = trie->uniquecharcount + 1;
2038
3dab1dad 2039
a3621e74
YO
2040 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2041
df826430 2042 regnode *noper = NEXTOPER( cur );
a28509cc 2043 const U8 *uc = (U8*)STRING( noper );
df826430 2044 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
2045
2046 U32 state = 1; /* required init */
2047
2048 U16 charid = 0; /* sanity init */
2049 U32 accept_state = 0; /* sanity init */
a3621e74 2050
07be1b83 2051 U32 wordlen = 0; /* required init */
a3621e74 2052
df826430
YO
2053 if (OP(noper) == NOTHING) {
2054 regnode *noper_next= regnext(noper);
2055 if (noper_next != tail && OP(noper_next) == flags) {
2056 noper = noper_next;
2057 uc= (U8*)STRING(noper);
2058 e= uc + STR_LEN(noper);
2059 }
2060 }
fab2782b 2061
3dab1dad 2062 if ( OP(noper) != NOTHING ) {
786e8c11 2063 for ( ; uc < e ; uc += len ) {
a3621e74 2064
786e8c11 2065 TRIE_READ_CHAR;
a3621e74 2066
786e8c11
YO
2067 if ( uvc < 256 ) {
2068 charid = trie->charmap[ uvc ];
2069 } else {
55eed653 2070 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2071 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2072 }
786e8c11
YO
2073 if ( charid ) {
2074 charid--;
2075 if ( !trie->trans[ state + charid ].next ) {
2076 trie->trans[ state + charid ].next = next_alloc;
2077 trie->trans[ state ].check++;
2e64971a
DM
2078 prev_states[TRIE_NODENUM(next_alloc)]
2079 = TRIE_NODENUM(state);
786e8c11
YO
2080 next_alloc += trie->uniquecharcount;
2081 }
2082 state = trie->trans[ state + charid ].next;
2083 } else {
2084 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2085 }
2086 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2087 }
a3621e74 2088 }
3dab1dad
YO
2089 accept_state = TRIE_NODENUM( state );
2090 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2091
2092 } /* end second pass */
2093
3dab1dad 2094 /* and now dump it out before we compress it */
2b8b4781
NC
2095 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2096 revcharmap,
2097 next_alloc, depth+1));
a3621e74 2098
a3621e74
YO
2099 {
2100 /*
2101 * Inplace compress the table.*
2102
2103 For sparse data sets the table constructed by the trie algorithm will
2104 be mostly 0/FAIL transitions or to put it another way mostly empty.
2105 (Note that leaf nodes will not contain any transitions.)
2106
2107 This algorithm compresses the tables by eliminating most such
2108 transitions, at the cost of a modest bit of extra work during lookup:
2109
2110 - Each states[] entry contains a .base field which indicates the
2111 index in the state[] array wheres its transition data is stored.
2112
3b753521 2113 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2114
2115 - If .base is nonzero then charid is added to it to find an entry in
2116 the trans array.
2117
2118 -If trans[states[state].base+charid].check!=state then the
2119 transition is taken to be a 0/Fail transition. Thus if there are fail
2120 transitions at the front of the node then the .base offset will point
2121 somewhere inside the previous nodes data (or maybe even into a node
2122 even earlier), but the .check field determines if the transition is
2123 valid.
2124
786e8c11 2125 XXX - wrong maybe?
a3621e74 2126 The following process inplace converts the table to the compressed
3b753521 2127 table: We first do not compress the root node 1,and mark all its
a3621e74 2128 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2129 allows us to do a DFA construction from the compressed table later,
2130 and ensures that any .base pointers we calculate later are greater
2131 than 0.
a3621e74
YO
2132
2133 - We set 'pos' to indicate the first entry of the second node.
2134
2135 - We then iterate over the columns of the node, finding the first and
2136 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2137 and set the .check pointers accordingly, and advance pos
2138 appropriately and repreat for the next node. Note that when we copy
2139 the next pointers we have to convert them from the original
2140 NODEIDX form to NODENUM form as the former is not valid post
2141 compression.
2142
2143 - If a node has no transitions used we mark its base as 0 and do not
2144 advance the pos pointer.
2145
2146 - If a node only has one transition we use a second pointer into the
2147 structure to fill in allocated fail transitions from other states.
2148 This pointer is independent of the main pointer and scans forward
2149 looking for null transitions that are allocated to a state. When it
2150 finds one it writes the single transition into the "hole". If the
786e8c11 2151 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2152
2153 - Once compressed we can Renew/realloc the structures to release the
2154 excess space.
2155
2156 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2157 specifically Fig 3.47 and the associated pseudocode.
2158
2159 demq
2160 */
a3b680e6 2161 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2162 U32 state, charid;
a3621e74 2163 U32 pos = 0, zp=0;
1e2e3d02 2164 trie->statecount = laststate;
a3621e74
YO
2165
2166 for ( state = 1 ; state < laststate ; state++ ) {
2167 U8 flag = 0;
a28509cc
AL
2168 const U32 stateidx = TRIE_NODEIDX( state );
2169 const U32 o_used = trie->trans[ stateidx ].check;
2170 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2171 trie->trans[ stateidx ].check = 0;
2172
2173 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2174 if ( flag || trie->trans[ stateidx + charid ].next ) {
2175 if ( trie->trans[ stateidx + charid ].next ) {
2176 if (o_used == 1) {
2177 for ( ; zp < pos ; zp++ ) {
2178 if ( ! trie->trans[ zp ].next ) {
2179 break;
2180 }
2181 }
2182 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2183 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2184 trie->trans[ zp ].check = state;
2185 if ( ++zp > pos ) pos = zp;
2186 break;
2187 }
2188 used--;
2189 }
2190 if ( !flag ) {
2191 flag = 1;
2192 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2193 }
2194 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2195 trie->trans[ pos ].check = state;
2196 pos++;
2197 }
2198 }
2199 }
cc601c31 2200 trie->lasttrans = pos + 1;
c944940b
JH
2201 trie->states = (reg_trie_state *)
2202 PerlMemShared_realloc( trie->states, laststate
2203 * sizeof(reg_trie_state) );
a3621e74 2204 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2205 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2206 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2207 (int)depth * 2 + 2,"",
2208 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2209 (IV)next_alloc,
2210 (IV)pos,
a3621e74
YO
2211 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2212 );
2213
2214 } /* end table compress */
2215 }
1e2e3d02
YO
2216 DEBUG_TRIE_COMPILE_MORE_r(
2217 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2218 (int)depth * 2 + 2, "",
2219 (UV)trie->statecount,
2220 (UV)trie->lasttrans)
2221 );
cc601c31 2222 /* resize the trans array to remove unused space */
c944940b
JH
2223 trie->trans = (reg_trie_trans *)
2224 PerlMemShared_realloc( trie->trans, trie->lasttrans
2225 * sizeof(reg_trie_trans) );
a3621e74 2226
3b753521 2227 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2228 U8 nodetype =(U8)(flags & 0xFF);
2229 char *str=NULL;
786e8c11 2230
07be1b83 2231#ifdef DEBUGGING
e62cc96a 2232 regnode *optimize = NULL;
7122b237
YO
2233#ifdef RE_TRACK_PATTERN_OFFSETS
2234
b57a0404
JH
2235 U32 mjd_offset = 0;
2236 U32 mjd_nodelen = 0;
7122b237
YO
2237#endif /* RE_TRACK_PATTERN_OFFSETS */
2238#endif /* DEBUGGING */
a3621e74 2239 /*
3dab1dad
YO
2240 This means we convert either the first branch or the first Exact,
2241 depending on whether the thing following (in 'last') is a branch
2242 or not and whther first is the startbranch (ie is it a sub part of
2243 the alternation or is it the whole thing.)
3b753521 2244 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2245 the whole branch sequence, including the first.
a3621e74 2246 */
3dab1dad 2247 /* Find the node we are going to overwrite */
7f69552c 2248 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2249 /* branch sub-chain */
3dab1dad 2250 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2251#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2252 DEBUG_r({
2253 mjd_offset= Node_Offset((convert));
2254 mjd_nodelen= Node_Length((convert));
2255 });
7122b237 2256#endif
7f69552c 2257 /* whole branch chain */
7122b237
YO
2258 }
2259#ifdef RE_TRACK_PATTERN_OFFSETS
2260 else {
7f69552c
YO
2261 DEBUG_r({
2262 const regnode *nop = NEXTOPER( convert );
2263 mjd_offset= Node_Offset((nop));
2264 mjd_nodelen= Node_Length((nop));
2265 });
07be1b83
YO
2266 }
2267 DEBUG_OPTIMISE_r(
2268 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2269 (int)depth * 2 + 2, "",
786e8c11 2270 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2271 );
7122b237 2272#endif
3dab1dad
YO
2273 /* But first we check to see if there is a common prefix we can
2274 split out as an EXACT and put in front of the TRIE node. */
2275 trie->startstate= 1;
55eed653 2276 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2277 U32 state;
1e2e3d02 2278 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2279 U32 ofs = 0;
8e11feef
RGS
2280 I32 idx = -1;
2281 U32 count = 0;
2282 const U32 base = trie->states[ state ].trans.base;
a3621e74 2283
3dab1dad 2284 if ( trie->states[state].wordnum )
8e11feef 2285 count = 1;
a3621e74 2286
8e11feef 2287 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2288 if ( ( base + ofs >= trie->uniquecharcount ) &&
2289 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2290 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2291 {
3dab1dad 2292 if ( ++count > 1 ) {
2b8b4781 2293 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2294 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2295 if ( state == 1 ) break;
3dab1dad
YO
2296 if ( count == 2 ) {
2297 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2298 DEBUG_OPTIMISE_r(
8e11feef
RGS
2299 PerlIO_printf(Perl_debug_log,
2300 "%*sNew Start State=%"UVuf" Class: [",
2301 (int)depth * 2 + 2, "",
786e8c11 2302 (UV)state));
be8e71aa 2303 if (idx >= 0) {
2b8b4781 2304 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2305 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2306
3dab1dad 2307 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2308 if ( folder )
2309 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2310 DEBUG_OPTIMISE_r(
f1f66076 2311 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2312 );
8e11feef
RGS
2313 }
2314 }
2315 TRIE_BITMAP_SET(trie,*ch);
2316 if ( folder )
2317 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2318 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2319 }
2320 idx = ofs;
2321 }
3dab1dad
YO
2322 }
2323 if ( count == 1 ) {
2b8b4781 2324 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2325 STRLEN len;
2326 char *ch = SvPV( *tmp, len );
de734bd5
A
2327 DEBUG_OPTIMISE_r({
2328 SV *sv=sv_newmortal();
8e11feef
RGS
2329 PerlIO_printf( Perl_debug_log,
2330 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2331 (int)depth * 2 + 2, "",
de734bd5
A
2332 (UV)state, (UV)idx,
2333 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2334 PL_colors[0], PL_colors[1],
2335 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2336 PERL_PV_ESCAPE_FIRSTCHAR
2337 )
2338 );
2339 });
3dab1dad
YO
2340 if ( state==1 ) {
2341 OP( convert ) = nodetype;
2342 str=STRING(convert);
2343 STR_LEN(convert)=0;
2344 }
c490c714
YO
2345 STR_LEN(convert) += len;
2346 while (len--)
de734bd5 2347 *str++ = *ch++;
8e11feef 2348 } else {
f9049ba1 2349#ifdef DEBUGGING
8e11feef
RGS
2350 if (state>1)
2351 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2352#endif
8e11feef
RGS
2353 break;
2354 }
2355 }
2e64971a 2356 trie->prefixlen = (state-1);
3dab1dad 2357 if (str) {
8e11feef 2358 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2359 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2360 trie->startstate = state;
07be1b83
YO
2361 trie->minlen -= (state - 1);
2362 trie->maxlen -= (state - 1);
33809eae
JH
2363#ifdef DEBUGGING
2364 /* At least the UNICOS C compiler choked on this
2365 * being argument to DEBUG_r(), so let's just have
2366 * it right here. */
2367 if (
2368#ifdef PERL_EXT_RE_BUILD
2369 1
2370#else
2371 DEBUG_r_TEST
2372#endif
2373 ) {
2374 regnode *fix = convert;
2375 U32 word = trie->wordcount;
2376 mjd_nodelen++;
2377 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2378 while( ++fix < n ) {
2379 Set_Node_Offset_Length(fix, 0, 0);
2380 }
2381 while (word--) {
2382 SV ** const tmp = av_fetch( trie_words, word, 0 );
2383 if (tmp) {
2384 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2385 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2386 else
2387 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2388 }
2389 }
2390 }
2391#endif
8e11feef
RGS
2392 if (trie->maxlen) {
2393 convert = n;
2394 } else {
3dab1dad 2395 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2396 DEBUG_r(optimize= n);
3dab1dad
YO
2397 }
2398 }
2399 }
a5ca303d
YO
2400 if (!jumper)
2401 jumper = last;
3dab1dad 2402 if ( trie->maxlen ) {
8e11feef
RGS
2403 NEXT_OFF( convert ) = (U16)(tail - convert);
2404 ARG_SET( convert, data_slot );
786e8c11
YO
2405 /* Store the offset to the first unabsorbed branch in
2406 jump[0], which is otherwise unused by the jump logic.
2407 We use this when dumping a trie and during optimisation. */
2408 if (trie->jump)
7f69552c 2409 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2410
6c48061a
YO
2411 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2412 * and there is a bitmap
2413 * and the first "jump target" node we found leaves enough room
2414 * then convert the TRIE node into a TRIEC node, with the bitmap
2415 * embedded inline in the opcode - this is hypothetically faster.
2416 */
2417 if ( !trie->states[trie->startstate].wordnum
2418 && trie->bitmap
2419 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2420 {
2421 OP( convert ) = TRIEC;
2422 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2423 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2424 trie->bitmap= NULL;
2425 } else
2426 OP( convert ) = TRIE;
a3621e74 2427
3dab1dad
YO
2428 /* store the type in the flags */
2429 convert->flags = nodetype;
a5ca303d
YO
2430 DEBUG_r({
2431 optimize = convert
2432 + NODE_STEP_REGNODE
2433 + regarglen[ OP( convert ) ];
2434 });
2435 /* XXX We really should free up the resource in trie now,
2436 as we won't use them - (which resources?) dmq */
3dab1dad 2437 }
a3621e74 2438 /* needed for dumping*/
e62cc96a 2439 DEBUG_r(if (optimize) {
07be1b83 2440 regnode *opt = convert;
bcdf7404 2441
e62cc96a 2442 while ( ++opt < optimize) {
07be1b83
YO
2443 Set_Node_Offset_Length(opt,0,0);
2444 }
786e8c11
YO
2445 /*
2446 Try to clean up some of the debris left after the
2447 optimisation.
a3621e74 2448 */
786e8c11 2449 while( optimize < jumper ) {
07be1b83 2450 mjd_nodelen += Node_Length((optimize));
a3621e74 2451 OP( optimize ) = OPTIMIZED;
07be1b83 2452 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2453 optimize++;
2454 }
07be1b83 2455 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2456 });
2457 } /* end node insert */
2e64971a
DM
2458
2459 /* Finish populating the prev field of the wordinfo array. Walk back
2460 * from each accept state until we find another accept state, and if
2461 * so, point the first word's .prev field at the second word. If the
2462 * second already has a .prev field set, stop now. This will be the
2463 * case either if we've already processed that word's accept state,
3b753521
FN
2464 * or that state had multiple words, and the overspill words were
2465 * already linked up earlier.
2e64971a
DM
2466 */
2467 {
2468 U16 word;
2469 U32 state;
2470 U16 prev;
2471
2472 for (word=1; word <= trie->wordcount; word++) {
2473 prev = 0;
2474 if (trie->wordinfo[word].prev)
2475 continue;
2476 state = trie->wordinfo[word].accept;
2477 while (state) {
2478 state = prev_states[state];
2479 if (!state)
2480 break;
2481 prev = trie->states[state].wordnum;
2482 if (prev)
2483 break;
2484 }
2485 trie->wordinfo[word].prev = prev;
2486 }
2487 Safefree(prev_states);
2488 }
2489
2490
2491 /* and now dump out the compressed format */
2492 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2493
55eed653 2494 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2495#ifdef DEBUGGING
2496 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2497 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2498#else
03e70be4 2499 SvREFCNT_dec_NN(revcharmap);
07be1b83 2500#endif
786e8c11
YO
2501 return trie->jump
2502 ? MADE_JUMP_TRIE
2503 : trie->startstate>1
2504 ? MADE_EXACT_TRIE
2505 : MADE_TRIE;
2506}
2507
2508STATIC void
2509S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2510{
3b753521 2511/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2512
2513 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2514 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2515 ISBN 0-201-10088-6
2516
2517 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2518 suffix of the current state's 'word' that is also a proper prefix of another word in our
2519 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2520 the DFA not to have to restart after its tried and failed a word at a given point, it
2521 simply continues as though it had been matching the other word in the first place.
2522 Consider
2523 'abcdgu'=~/abcdefg|cdgu/
2524 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2525 fail, which would bring us to the state representing 'd' in the second word where we would
2526 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2527 */
2528 /* add a fail transition */
3251b653
NC
2529 const U32 trie_offset = ARG(source);
2530 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2531 U32 *q;
2532 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2533 const U32 numstates = trie->statecount;
786e8c11
YO
2534 const U32 ubound = trie->lasttrans + ucharcount;
2535 U32 q_read = 0;
2536 U32 q_write = 0;
2537 U32 charid;
2538 U32 base = trie->states[ 1 ].trans.base;
2539 U32 *fail;
2540 reg_ac_data *aho;
2541 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2542 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2543
2544 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2545#ifndef DEBUGGING
2546 PERL_UNUSED_ARG(depth);
2547#endif
2548
2549
2550 ARG_SET( stclass, data_slot );
c944940b 2551 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2552 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2553 aho->trie=trie_offset;
446bd890
NC
2554 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2555 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2556 Newxz( q, numstates, U32);
c944940b 2557 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2558 aho->refcount = 1;
2559 fail = aho->fail;
2560 /* initialize fail[0..1] to be 1 so that we always have
2561 a valid final fail state */
2562 fail[ 0 ] = fail[ 1 ] = 1;
2563
2564 for ( charid = 0; charid < ucharcount ; charid++ ) {
2565 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2566 if ( newstate ) {
2567 q[ q_write ] = newstate;
2568 /* set to point at the root */
2569 fail[ q[ q_write++ ] ]=1;
2570 }
2571 }
2572 while ( q_read < q_write) {
2573 const U32 cur = q[ q_read++ % numstates ];
2574 base = trie->states[ cur ].trans.base;
2575
2576 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2577 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2578 if (ch_state) {
2579 U32 fail_state = cur;
2580 U32 fail_base;
2581 do {
2582 fail_state = fail[ fail_state ];
2583 fail_base = aho->states[ fail_state ].trans.base;
2584 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2585
2586 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2587 fail[ ch_state ] = fail_state;
2588 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2589 {
2590 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2591 }
2592 q[ q_write++ % numstates] = ch_state;
2593 }
2594 }
2595 }
2596 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2597 when we fail in state 1, this allows us to use the
2598 charclass scan to find a valid start char. This is based on the principle
2599 that theres a good chance the string being searched contains lots of stuff
2600 that cant be a start char.
2601 */
2602 fail[ 0 ] = fail[ 1 ] = 0;
2603 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2604 PerlIO_printf(Perl_debug_log,
2605 "%*sStclass Failtable (%"UVuf" states): 0",
2606 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2607 );
786e8c11
YO
2608 for( q_read=1; q_read<numstates; q_read++ ) {
2609 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2610 }
2611 PerlIO_printf(Perl_debug_log, "\n");
2612 });
2613 Safefree(q);
2614 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2615}
2616
786e8c11 2617
a3621e74 2618/*
5d1c421c
JH
2619 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2620 * These need to be revisited when a newer toolchain becomes available.
2621 */
2622#if defined(__sparc64__) && defined(__GNUC__)
2623# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2624# undef SPARC64_GCC_WORKAROUND
2625# define SPARC64_GCC_WORKAROUND 1
2626# endif
2627#endif
2628
07be1b83 2629#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2630 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2631 SV * const mysv=sv_newmortal(); \
2632 regnode *Next = regnext(scan); \
2633 regprop(RExC_rx, mysv, scan); \
7f69552c 2634 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2635 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2636 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2637 }});
07be1b83 2638
1de06328 2639
bb914485 2640/* The below joins as many adjacent EXACTish nodes as possible into a single
0a982f06
KW
2641 * one. The regop may be changed if the node(s) contain certain sequences that
2642 * require special handling. The joining is only done if:
bb914485
KW
2643 * 1) there is room in the current conglomerated node to entirely contain the
2644 * next one.
2645 * 2) they are the exact same node type
2646 *
87b8b349 2647 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
2648 * these get optimized out
2649 *
0a982f06
KW
2650 * If a node is to match under /i (folded), the number of characters it matches
2651 * can be different than its character length if it contains a multi-character
2652 * fold. *min_subtract is set to the total delta of the input nodes.
bb914485 2653 *
a0c4c608
KW
2654 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2655 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2656 *
bb914485 2657 * This is as good a place as any to discuss the design of handling these
0a982f06
KW
2658 * multi-character fold sequences. It's been wrong in Perl for a very long
2659 * time. There are three code points in Unicode whose multi-character folds
2660 * were long ago discovered to mess things up. The previous designs for
2661 * dealing with these involved assigning a special node for them. This
2662 * approach doesn't work, as evidenced by this example:
a0c4c608 2663 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
0a982f06
KW
2664 * Both these fold to "sss", but if the pattern is parsed to create a node that
2665 * would match just the \xDF, it won't be able to handle the case where a
bb914485
KW
2666 * successful match would have to cross the node's boundary. The new approach
2667 * that hopefully generally solves the problem generates an EXACTFU_SS node
2668 * that is "sss".
2669 *
0a982f06
KW
2670 * It turns out that there are problems with all multi-character folds, and not
2671 * just these three. Now the code is general, for all such cases, but the
2672 * three still have some special handling. The approach taken is:
2673 * 1) This routine examines each EXACTFish node that could contain multi-
2674 * character fold sequences. It returns in *min_subtract how much to
9d071ca8 2675 * subtract from the the actual length of the string to get a real minimum
0a982f06
KW
2676 * match length; it is 0 if there are no multi-char folds. This delta is
2677 * used by the caller to adjust the min length of the match, and the delta
2678 * between min and max, so that the optimizer doesn't reject these
2679 * possibilities based on size constraints.
2680 * 2) Certain of these sequences require special handling by the trie code,
2681 * so, if found, this code changes the joined node type to special ops:
2682 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2683 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2684 * is used for an EXACTFU node that contains at least one "ss" sequence in
2685 * it. For non-UTF-8 patterns and strings, this is the only case where
2686 * there is a possible fold length change. That means that a regular
2687 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2688 * with length changes, and so can be processed faster. regexec.c takes
2689 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2690 * pre-folded by regcomp.c. This saves effort in regex matching.
87b8b349 2691 * However, the pre-folding isn't done for non-UTF8 patterns because the
0a982f06
KW
2692 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2693 * down by forcing the pattern into UTF8 unless necessary. Also what
2694 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2695 * possibilities for the non-UTF8 patterns are quite simple, except for
2696 * the sharp s. All the ones that don't involve a UTF-8 target string are
2697 * members of a fold-pair, and arrays are set up for all of them so that
2698 * the other member of the pair can be found quickly. Code elsewhere in
2699 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2700 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2701 * described in the next item.
1ca267a5
KW
2702 * 4) A problem remains for the sharp s in EXACTF and EXACTFA nodes when the
2703 * pattern isn't in UTF-8. (BTW, there cannot be an EXACTF node with a
2704 * UTF-8 pattern.) An assumption that the optimizer part of regexec.c
2705 * (probably unwittingly, in Perl_regexec_flags()) makes is that a
2706 * character in the pattern corresponds to at most a single character in
2707 * the target string. (And I do mean character, and not byte here, unlike
2708 * other parts of the documentation that have never been updated to
2709 * account for multibyte Unicode.) sharp s in EXACTF nodes can match the
2710 * two character string 'ss'; in EXACTFA nodes it can match
2711 * "\x{17F}\x{17F}". These violate the assumption, and they are the only
2712 * instances where it is violated. I'm reluctant to try to change the
2713 * assumption, as the code involved is impenetrable to me (khw), so
2714 * instead the code here punts. This routine examines (when the pattern
2715 * isn't UTF-8) EXACTF and EXACTFA nodes for the sharp s, and returns a
2716 * boolean indicating whether or not the node contains a sharp s. When it
2717 * is true, the caller sets a flag that later causes the optimizer in this
2718 * file to not set values for the floating and fixed string lengths, and
2719 * thus avoids the optimizer code in regexec.c that makes the invalid
2720 * assumption. Thus, there is no optimization based on string lengths for
2721 * non-UTF8-pattern EXACTF and EXACTFA nodes that contain the sharp s.
2722 * (The reason the assumption is wrong only in these two cases is that all
2723 * other non-UTF-8 folds are 1-1; and, for UTF-8 patterns, we pre-fold all
2724 * other folds to their expanded versions. We can't prefold sharp s to
2725 * 'ss' in EXACTF nodes because we don't know at compile time if it
2726 * actually matches 'ss' or not. It will match iff the target string is
2727 * in UTF-8, unlike the EXACTFU nodes, where it always matches; and
2728 * EXACTFA and EXACTFL where it never does. In an EXACTFA node in a UTF-8
2729 * pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the problem;
2730 * but in a non-UTF8 pattern, folding it to that above-Latin1 string would
2731 * require the pattern to be forced into UTF-8, the overhead of which we
2732 * want to avoid.)
bb914485 2733 */
1de06328 2734
9d071ca8 2735#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2736 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2737 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2738
be8e71aa 2739STATIC U32
9d071ca8 2740S_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
2741 /* Merge several consecutive EXACTish nodes into one. */
2742 regnode *n = regnext(scan);
2743 U32 stringok = 1;
2744 regnode *next = scan + NODE_SZ_STR(scan);
2745 U32 merged = 0;
2746 U32 stopnow = 0;
2747#ifdef DEBUGGING
2748 regnode *stop = scan;
72f13be8 2749 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2750#else
d47053eb
RGS
2751 PERL_UNUSED_ARG(depth);
2752#endif
7918f24d
NC
2753
2754 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2755#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2756 PERL_UNUSED_ARG(flags);
2757 PERL_UNUSED_ARG(val);
07be1b83 2758#endif
07be1b83 2759 DEBUG_PEEP("join",scan,depth);
bb914485 2760
3f410cf6
KW
2761 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2762 * EXACT ones that are mergeable to the current one. */
2763 while (n
2764 && (PL_regkind[OP(n)] == NOTHING
2765 || (stringok && OP(n) == OP(scan)))
07be1b83 2766 && NEXT_OFF(n)
3f410cf6
KW
2767 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2768 {
07be1b83
YO
2769
2770 if (OP(n) == TAIL || n > next)
2771 stringok = 0;
2772 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2773 DEBUG_PEEP("skip:",n,depth);
2774 NEXT_OFF(scan) += NEXT_OFF(n);
2775 next = n + NODE_STEP_REGNODE;
2776#ifdef DEBUGGING
2777 if (stringok)
2778 stop = n;
2779#endif
2780 n = regnext(n);
2781 }
2782 else if (stringok) {
786e8c11 2783 const unsigned int oldl = STR_LEN(scan);
07be1b83 2784 regnode * const nnext = regnext(n);
b2230d39 2785
87b8b349
KW
2786 /* XXX I (khw) kind of doubt that this works on platforms where
2787 * U8_MAX is above 255 because of lots of other assumptions */
79a81a6e 2788 /* Don't join if the sum can't fit into a single node */
b2230d39
KW
2789 if (oldl + STR_LEN(n) > U8_MAX)
2790 break;
07be1b83
YO
2791
2792 DEBUG_PEEP("merg",n,depth);
07be1b83 2793 merged++;
b2230d39 2794
07be1b83
YO
2795 NEXT_OFF(scan) += NEXT_OFF(n);
2796 STR_LEN(scan) += STR_LEN(n);
2797 next = n + NODE_SZ_STR(n);
2798 /* Now we can overwrite *n : */
2799 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2800#ifdef DEBUGGING
2801 stop = next - 1;
2802#endif
2803 n = nnext;
2804 if (stopnow) break;
2805 }
2806
d47053eb
RGS
2807#ifdef EXPERIMENTAL_INPLACESCAN
2808 if (flags && !NEXT_OFF(n)) {
2809 DEBUG_PEEP("atch", val, depth);
2810 if (reg_off_by_arg[OP(n)]) {
2811 ARG_SET(n, val - n);
2812 }
2813 else {
2814 NEXT_OFF(n) = val - n;
2815 }
2816 stopnow = 1;
2817 }
07be1b83
YO
2818#endif
2819 }
2c2b7f86 2820
9d071ca8 2821 *min_subtract = 0;
f758bddf 2822 *has_exactf_sharp_s = FALSE;
f646642f 2823
3f410cf6
KW
2824 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2825 * can now analyze for sequences of problematic code points. (Prior to
2826 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2827 * hence missed). The sequences only happen in folding, hence for any
2828 * non-EXACT EXACTish node */
86d6fcad 2829 if (OP(scan) != EXACT) {
0a982f06
KW
2830 const U8 * const s0 = (U8*) STRING(scan);
2831 const U8 * s = s0;
2832 const U8 * const s_end = s0 + STR_LEN(scan);
f758bddf
KW
2833
2834 /* One pass is made over the node's string looking for all the
2835 * possibilities. to avoid some tests in the loop, there are two main
2836 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2837 * non-UTF-8 */
2838 if (UTF) {
86d6fcad 2839
0a982f06
KW
2840 /* Examine the string for a multi-character fold sequence. UTF-8
2841 * patterns have all characters pre-folded by the time this code is
2842 * executed */
2843 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2844 length sequence we are looking for is 2 */
86d6fcad 2845 {
0a982f06
KW
2846 int count = 0;
2847 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2848 if (! len) { /* Not a multi-char fold: get next char */
2849 s += UTF8SKIP(s);
2850 continue;
2851 }
bb914485 2852
0a982f06
KW
2853 /* Nodes with 'ss' require special handling, except for EXACTFL
2854 * and EXACTFA for which there is no multi-char fold to this */
2855 if (len == 2 && *s == 's' && *(s+1) == 's'
2856 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2857 {
2858 count = 2;
2859 OP(scan) = EXACTFU_SS;
2860 s += 2;
2861 }
accd961f
KW
2862 else if (len == 6 /* len is the same in both ASCII and EBCDIC
2863 for these */
0a982f06
KW
2864 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2865 COMBINING_DIAERESIS_UTF8
2866 COMBINING_ACUTE_ACCENT_UTF8,
2867 6)
2868 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2869 COMBINING_DIAERESIS_UTF8
2870 COMBINING_ACUTE_ACCENT_UTF8,
2871 6)))
2872 {
2873 count = 3;
2874
2875 /* These two folds require special handling by trie's, so
2876 * change the node type to indicate this. If EXACTFA and
2877 * EXACTFL were ever to be handled by trie's, this would
2878 * have to be changed. If this node has already been
2879 * changed to EXACTFU_SS in this loop, leave it as is. (I
2880 * (khw) think it doesn't matter in regexec.c for UTF
2881 * patterns, but no need to change it */
2882 if (OP(scan) == EXACTFU) {
2883 OP(scan) = EXACTFU_TRICKYFOLD;
2884 }
2885 s += 6;
2886 }
2887 else { /* Here is a generic multi-char fold. */
2888 const U8* multi_end = s + len;
2889
2890 /* Count how many characters in it. In the case of /l and
2891 * /aa, no folds which contain ASCII code points are
2892 * allowed, so check for those, and skip if found. (In
2893 * EXACTFL, no folds are allowed to any Latin1 code point,
2894 * not just ASCII. But there aren't any of these
2895 * currently, nor ever likely, so don't take the time to
2896 * test for them. The code that generates the
2897 * is_MULTI_foo() macros croaks should one actually get put
2898 * into Unicode .) */
2899 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2900 count = utf8_length(s, multi_end);
2901 s = multi_end;
2902 }
2903 else {
2904 while (s < multi_end) {
2905 if (isASCII(*s)) {
2906 s++;
2907 goto next_iteration;
2908 }
2909 else {
2910 s += UTF8SKIP(s);
2911 }
2912 count++;
2913 }
2914 }
2915 }
f758bddf 2916
0a982f06
KW
2917 /* The delta is how long the sequence is minus 1 (1 is how long
2918 * the character that folds to the sequence is) */
2919 *min_subtract += count - 1;
2920 next_iteration: ;
bb914485
KW
2921 }
2922 }
1ca267a5
KW
2923 else if (OP(scan) == EXACTFA) {
2924
2925 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
2926 * fold to the ASCII range (and there are no existing ones in the
2927 * upper latin1 range). But, as outlined in the comments preceding
2928 * this function, we need to flag any occurrences of the sharp s */
2929 while (s < s_end) {
2930 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
2931 *has_exactf_sharp_s = TRUE;
2932 break;
2933 }
2934 s++;
2935 continue;
2936 }
2937 }
2938 else if (OP(scan) != EXACTFL) {
2939
2940 /* Non-UTF-8 pattern, not EXACTFA nor EXACTFL node. Look for the
2941 * multi-char folds that are all Latin1. (This code knows that
2942 * there are no current multi-char folds possible with EXACTFL,
2943 * relying on fold_grind.t to catch any errors if the very unlikely
2944 * event happens that some get added in future Unicode versions.)
2945 * As explained in the comments preceding this function, we look
2946 * also for the sharp s in EXACTF nodes; it can be in the final
0a982f06
KW
2947 * position. Otherwise we can stop looking 1 byte earlier because
2948 * have to find at least two characters for a multi-fold */
f758bddf
KW
2949 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2950
0a982f06 2951 while (s < upper) {
40b1ba4f 2952 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
0a982f06
KW
2953 if (! len) { /* Not a multi-char fold. */
2954 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2955 {
2956 *has_exactf_sharp_s = TRUE;
2957 }
2958 s++;
2959 continue;
2960 }
2961
2962 if (len == 2
c02c3054
KW
2963 && isARG2_lower_or_UPPER_ARG1('s', *s)
2964 && isARG2_lower_or_UPPER_ARG1('s', *(s+1)))
0a982f06
KW
2965 {
2966
2967 /* EXACTF nodes need to know that the minimum length
2968 * changed so that a sharp s in the string can match this
2969 * ss in the pattern, but they remain EXACTF nodes, as they
2970 * won't match this unless the target string is is UTF-8,
2971 * which we don't know until runtime */
2972 if (OP(scan) != EXACTF) {
2973 OP(scan) = EXACTFU_SS;
2974 }
86d6fcad 2975 }
0a982f06
KW
2976
2977 *min_subtract += len - 1;
2978 s += len;
86d6fcad
KW
2979 }
2980 }
07be1b83 2981 }
3f410cf6 2982
07be1b83 2983#ifdef DEBUGGING
bb789b09
DM
2984 /* Allow dumping but overwriting the collection of skipped
2985 * ops and/or strings with fake optimized ops */
07be1b83
YO
2986 n = scan + NODE_SZ_STR(scan);
2987 while (n <= stop) {
bb789b09
DM
2988 OP(n) = OPTIMIZED;
2989 FLAGS(n) = 0;
2990 NEXT_OFF(n) = 0;
07be1b83
YO
2991 n++;
2992 }
2993#endif
2994 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2995 return stopnow;
2996}
2997
486ec47a 2998/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2999 Finds fixed substrings. */
3000
a0288114 3001/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
3002 to the position after last scanned or to NULL. */
3003
40d049e4
YO
3004#define INIT_AND_WITHP \
3005 assert(!and_withp); \
3006 Newx(and_withp,1,struct regnode_charclass_class); \
3007 SAVEFREEPV(and_withp)
07be1b83 3008
b515a41d 3009/* this is a chain of data about sub patterns we are processing that
486ec47a 3010 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
3011 we can simulate recursion without losing state. */
3012struct scan_frame;
3013typedef struct scan_frame {
3014 regnode *last; /* last node to process in this frame */
3015 regnode *next; /* next node to process when last is reached */
3016 struct scan_frame *prev; /*previous frame*/
3017 I32 stop; /* what stopparen do we use */
3018} scan_frame;
3019
304ee84b
YO
3020
3021#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
3022
49f55535 3023STATIC SSize_t
40d049e4 3024S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
ea3daa5d 3025 SSize_t *minlenp, SSize_t *deltap,
40d049e4
YO
3026 regnode *last,
3027 scan_data_t *data,
3028 I32 stopparen,
3029 U8* recursed,
3030 struct regnode_charclass_class *and_withp,
3031 U32 flags, U32 depth)
c277df42
IZ
3032 /* scanp: Start here (read-write). */
3033 /* deltap: Write maxlen-minlen here. */
3034 /* last: Stop before this one. */
40d049e4
YO
3035 /* data: string data about the pattern */
3036 /* stopparen: treat close N as END */
3037 /* recursed: which subroutines have we recursed into */
3038 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 3039{
97aff369 3040 dVAR;
ea3daa5d
FC
3041 /* There must be at least this number of characters to match */
3042 SSize_t min = 0;
2d608413 3043 I32 pars = 0, code;
c277df42 3044 regnode *scan = *scanp, *next;
ea3daa5d 3045 SSize_t delta = 0;
c277df42 3046 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 3047 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
3048 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3049 scan_data_t data_fake;
a3621e74 3050 SV *re_trie_maxbuff = NULL;
786e8c11 3051 regnode *first_non_open = scan;
ea3daa5d 3052 SSize_t stopmin = SSize_t_MAX;
8aa23a47 3053 scan_frame *frame = NULL;
a3621e74 3054 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 3055
7918f24d
NC
3056 PERL_ARGS_ASSERT_STUDY_CHUNK;
3057
13a24bad 3058#ifdef DEBUGGING
40d049e4 3059 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 3060#endif
40d049e4 3061
786e8c11 3062 if ( depth == 0 ) {
40d049e4 3063 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
3064 first_non_open=regnext(first_non_open);
3065 }
3066
b81d288d 3067
8aa23a47
YO
3068 fake_study_recurse:
3069 while ( scan && OP(scan) != END && scan < last ){
2d608413
KW
3070 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3071 node length to get a real minimum (because
3072 the folded version may be shorter) */
f758bddf 3073 bool has_exactf_sharp_s = FALSE;
8aa23a47 3074 /* Peephole optimizer: */
304ee84b 3075 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 3076 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
3077
3078 /* Its not clear to khw or hv why this is done here, and not in the
3079 * clauses that deal with EXACT nodes. khw's guess is that it's
3080 * because of a previous design */
9d071ca8 3081 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
3082
3083 /* Follow the next-chain of the current node and optimize
3084 away all the NOTHINGs from it. */
3085 if (OP(scan) != CURLYX) {
3086 const int max = (reg_off_by_arg[OP(scan)]
3087 ? I32_MAX
3088 /* I32 may be smaller than U16 on CRAYs! */
3089 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3090 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3091 int noff;
3092 regnode *n = scan;
686b73d4 3093
8aa23a47
YO
3094 /* Skip NOTHING and LONGJMP. */
3095 while ((n = regnext(n))
3096 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3097 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3098 && off + noff < max)
3099 off += noff;
3100 if (reg_off_by_arg[OP(scan)])
3101 ARG(scan) = off;
3102 else
3103 NEXT_OFF(scan) = off;
3104 }
a3621e74 3105
c277df42 3106
8aa23a47
YO
3107
3108 /* The principal pseudo-switch. Cannot be a switch, since we
3109 look into several different things. */
3110 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3111 || OP(scan) == IFTHEN) {
3112 next = regnext(scan);
3113 code = OP(scan);
3114 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3115
8aa23a47
YO
3116 if (OP(next) == code || code == IFTHEN) {
3117 /* NOTE - There is similar code to this block below for handling
3118 TRIE nodes on a re-study. If you change stuff here check there
3119 too. */
49f55535 3120 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
8aa23a47
YO
3121 struct regnode_charclass_class accum;
3122 regnode * const startbranch=scan;
686b73d4 3123
8aa23a47 3124 if (flags & SCF_DO_SUBSTR)
304ee84b 3125 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3126 if (flags & SCF_DO_STCLASS)
e755fd73 3127 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3128
3129 while (OP(scan) == code) {
49f55535
FC
3130 SSize_t deltanext, minnext, fake;
3131 I32 f = 0;
8aa23a47
YO
3132 struct regnode_charclass_class this_class;
3133
3134 num++;
3135 data_fake.flags = 0;
3136 if (data) {
3137 data_fake.whilem_c = data->whilem_c;
3138 data_fake.last_closep = data->last_closep;
3139 }
3140 else
3141 data_fake.last_closep = &fake;
58e23c8d
YO
3142
3143 data_fake.pos_delta = delta;
8aa23a47
YO
3144 next = regnext(scan);
3145 scan = NEXTOPER(scan);
3146 if (code != BRANCH)
c277df42 3147 scan = NEXTOPER(scan);
8aa23a47 3148 if (flags & SCF_DO_STCLASS) {
e755fd73 3149 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3150 data_fake.start_class = &this_class;
3151 f = SCF_DO_STCLASS_AND;
58e23c8d 3152 }
8aa23a47
YO
3153 if (flags & SCF_WHILEM_VISITED_POS)
3154 f |= SCF_WHILEM_VISITED_POS;
3155
3156 /* we suppose the run is continuous, last=next...*/
3157 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3158 next, &data_fake,
3159 stopparen, recursed, NULL, f,depth+1);
3160 if (min1 > minnext)
3161 min1 = minnext;
ea3daa5d 3162 if (deltanext == SSize_t_MAX) {
8aa23a47 3163 is_inf = is_inf_internal = 1;
ea3daa5d 3164 max1 = SSize_t_MAX;
9b139d09
GG
3165 } else if (max1 < minnext + deltanext)
3166 max1 = minnext + deltanext;
8aa23a47
YO
3167 scan = next;
3168 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3169 pars++;
3170 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3171 if ( stopmin > minnext)
3172 stopmin = min + min1;
3173 flags &= ~SCF_DO_SUBSTR;
3174 if (data)
3175 data->flags |= SCF_SEEN_ACCEPT;
3176 }
3177 if (data) {
3178 if (data_fake.flags & SF_HAS_EVAL)
3179 data->flags |= SF_HAS_EVAL;
3180 data->whilem_c = data_fake.whilem_c;
3dab1dad 3181 }
8aa23a47 3182 if (flags & SCF_DO_STCLASS)
3fffb88a 3183 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3184 }
3185 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3186 min1 = 0;
3187 if (flags & SCF_DO_SUBSTR) {
3188 data->pos_min += min1;
ea3daa5d
FC
3189 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3190 data->pos_delta = SSize_t_MAX;
9b139d09
GG
3191 else
3192 data->pos_delta += max1 - min1;
8aa23a47
YO
3193 if (max1 != min1 || is_inf)
3194 data->longest = &(data->longest_float);
3195 }
3196 min += min1;
ea3daa5d
FC
3197 if (delta == SSize_t_MAX
3198 || SSize_t_MAX - delta - (max1 - min1) < 0)
3199 delta = SSize_t_MAX;
9b139d09
GG
3200 else
3201 delta += max1 - min1;
8aa23a47 3202 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3203 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3204 if (min1) {
3205 cl_and(data->start_class, and_withp);
3206 flags &= ~SCF_DO_STCLASS;
653099ff 3207 }
8aa23a47
YO
3208 }
3209 else if (flags & SCF_DO_STCLASS_AND) {
3210 if (min1) {
3211 cl_and(data->start_class, &accum);
3212 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3213 }
8aa23a47
YO
3214 else {
3215 /* Switch to OR mode: cache the old value of
3216 * data->start_class */
3217 INIT_AND_WITHP;
3218 StructCopy(data->start_class, and_withp,
3219 struct regnode_charclass_class);
3220 flags &= ~SCF_DO_STCLASS_AND;
3221 StructCopy(&accum, data->start_class,
3222 struct regnode_charclass_class);
3223 flags |= SCF_DO_STCLASS_OR;
899d20b9 3224 SET_SSC_EOS(data->start_class);
de0c8cb8 3225 }
8aa23a47 3226 }
a3621e74 3227
8aa23a47
YO
3228 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3229 /* demq.
a3621e74 3230
8aa23a47
YO
3231 Assuming this was/is a branch we are dealing with: 'scan' now
3232 points at the item that follows the branch sequence, whatever
3233 it is. We now start at the beginning of the sequence and look
3234 for subsequences of
a3621e74 3235
8aa23a47
YO
3236 BRANCH->EXACT=>x1
3237 BRANCH->EXACT=>x2
3238 tail
a3621e74 3239
8aa23a47 3240 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3241
486ec47a 3242 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3243 element into a trie and then add the subsequent branch exact
3244 strings to the trie.
a3621e74 3245
8aa23a47 3246 We have two cases
a3621e74 3247
3b753521 3248 1. patterns where the whole set of branches can be converted.
a3621e74 3249
8aa23a47 3250 2. patterns where only a subset can be converted.
a3621e74 3251
8aa23a47
YO
3252 In case 1 we can replace the whole set with a single regop
3253 for the trie. In case 2 we need to keep the start and end
3b753521 3254 branches so
a3621e74 3255
8aa23a47
YO
3256 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3257 becomes BRANCH TRIE; BRANCH X;
786e8c11 3258
8aa23a47
YO
3259 There is an additional case, that being where there is a
3260 common prefix, which gets split out into an EXACT like node
3261 preceding the TRIE node.
a3621e74 3262
8aa23a47
YO
3263 If x(1..n)==tail then we can do a simple trie, if not we make
3264 a "jump" trie, such that when we match the appropriate word
486ec47a 3265 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3266 a nested if into a case structure of sorts.
b515a41d 3267
8aa23a47 3268 */
686b73d4 3269
8aa23a47
YO
3270 int made=0;
3271 if (!re_trie_maxbuff) {
3272 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3273 if (!SvIOK(re_trie_maxbuff))
3274 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3275 }
3276 if ( SvIV(re_trie_maxbuff)>=0 ) {
3277 regnode *cur;
3278 regnode *first = (regnode *)NULL;
3279 regnode *last = (regnode *)NULL;
3280 regnode *tail = scan;
fab2782b 3281 U8 trietype = 0;
8aa23a47 3282 U32 count=0;
a3621e74
YO
3283
3284#ifdef DEBUGGING
8aa23a47 3285 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3286#endif
8aa23a47
YO
3287 /* var tail is used because there may be a TAIL
3288 regop in the way. Ie, the exacts will point to the
3289 thing following the TAIL, but the last branch will
3290 point at the TAIL. So we advance tail. If we
3291 have nested (?:) we may have to move through several
3292 tails.
3293 */
3294
3295 while ( OP( tail ) == TAIL ) {
3296 /* this is the TAIL generated by (?:) */
3297 tail = regnext( tail );
3298 }
a3621e74 3299
8aa23a47 3300
df826430 3301 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3302 regprop(RExC_rx, mysv, tail );
3303 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3304 (int)depth * 2 + 2, "",
3305 "Looking for TRIE'able sequences. Tail node is: ",
3306 SvPV_nolen_const( mysv )
3307 );
3308 });
3309
3310 /*
3311
fab2782b
YO
3312 Step through the branches
3313 cur represents each branch,
3314 noper is the first thing to be matched as part of that branch
3315 noper_next is the regnext() of that node.
3316
3317 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3318 via a "jump trie" but we also support building with NOJUMPTRIE,
3319 which restricts the trie logic to structures like /FOO|BAR/.
3320
3321 If noper is a trieable nodetype then the branch is a possible optimization
3322 target. If we are building under NOJUMPTRIE then we require that noper_next
3323 is the same as scan (our current position in the regex program).
3324
3325 Once we have two or more consecutive such branches we can create a
3326 trie of the EXACT's contents and stitch it in place into the program.
3327
3328 If the sequence represents all of the branches in the alternation we
3329 replace the entire thing with a single TRIE node.
3330
3331 Otherwise when it is a subsequence we need to stitch it in place and
3332 replace only the relevant branches. This means the first branch has
3333 to remain as it is used by the alternation logic, and its next pointer,
3334 and needs to be repointed at the item on the branch chain following
3335 the last branch we have optimized away.
3336
3337 This could be either a BRANCH, in which case the subsequence is internal,
3338 or it could be the item following the branch sequence in which case the
3339 subsequence is at the end (which does not necessarily mean the first node
3340 is the start of the alternation).
3341
3342 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3343
3344 optype | trietype
3345 ----------------+-----------
3346 NOTHING | NOTHING
3347 EXACT | EXACT
3348 EXACTFU | EXACTFU
3349 EXACTFU_SS | EXACTFU
3350 EXACTFU_TRICKYFOLD | EXACTFU
3351 EXACTFA | 0
3352
8aa23a47
YO
3353
3354 */
fab2782b
YO
3355#define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3356 ( EXACT == (X) ) ? EXACT : \
3357 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3358 0 )
8aa23a47
YO
3359
3360 /* dont use tail as the end marker for this traverse */
3361 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3362 regnode * const noper = NEXTOPER( cur );
fab2782b
YO
3363 U8 noper_type = OP( noper );
3364 U8 noper_trietype = TRIE_TYPE( noper_type );
b515a41d 3365#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3366 regnode * const noper_next = regnext( noper );
df826430
YO
3367 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3368 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
b515a41d
YO
3369#endif
3370
df826430 3371 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3372 regprop(RExC_rx, mysv, cur);
3373 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3374 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3375
3376 regprop(RExC_rx, mysv, noper);
3377 PerlIO_printf( Perl_debug_log, " -> %s",
3378 SvPV_nolen_const(mysv));
3379
3380 if ( noper_next ) {
3381 regprop(RExC_rx, mysv, noper_next );
3382 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3383 SvPV_nolen_const(mysv));
3384 }
df826430
YO
3385 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3386 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3387 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3388 );
8aa23a47 3389 });
fab2782b
YO
3390
3391 /* Is noper a trieable nodetype that can be merged with the
3392 * current trie (if there is one)? */
3393 if ( noper_trietype
3394 &&
3395 (
df826430
YO
3396 ( noper_trietype == NOTHING)
3397 || ( trietype == NOTHING )
a40630bf 3398 || ( trietype == noper_trietype )
fab2782b 3399 )
786e8c11 3400#ifdef NOJUMPTRIE
8aa23a47 3401 && noper_next == tail
786e8c11 3402#endif
8aa23a47
YO
3403 && count < U16_MAX)
3404 {
fab2782b
YO
3405 /* Handle mergable triable node
3406 * Either we are the first node in a new trieable sequence,
3407 * in which case we do some bookkeeping, otherwise we update
3408 * the end pointer. */
fab2782b 3409 if ( !first ) {
3b6759a6 3410 first = cur;
df826430
YO
3411 if ( noper_trietype == NOTHING ) {
3412#if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3413 regnode * const noper_next = regnext( noper );
3b6759a6 3414 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
df826430
YO
3415 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3416#endif
3417
190c1910 3418 if ( noper_next_trietype ) {
df826430 3419 trietype = noper_next_trietype;
190c1910
YO
3420 } else if (noper_next_type) {
3421 /* a NOTHING regop is 1 regop wide. We need at least two
3422 * for a trie so we can't merge this in */
3423 first = NULL;
3424 }
3425 } else {
3426 trietype = noper_trietype;
3b6759a6 3427 }
8aa23a47 3428 } else {
fab2782b
YO
3429 if ( trietype == NOTHING )
3430 trietype = noper_trietype;
8aa23a47
YO
3431 last = cur;
3432 }
df826430
YO
3433 if (first)
3434 count++;
fab2782b
YO
3435 } /* end handle mergable triable node */
3436 else {
3437 /* handle unmergable node -
3438 * noper may either be a triable node which can not be tried
3439 * together with the current trie, or a non triable node */
729aaeb5
YO
3440 if ( last ) {
3441 /* If last is set and trietype is not NOTHING then we have found
3442 * at least two triable branch sequences in a row of a similar
3443 * trietype so we can turn them into a trie. If/when we
3444 * allow NOTHING to start a trie sequence this condition will be
3445 * required, and it isn't expensive so we leave it in for now. */
e6351b37 3446 if ( trietype && trietype != NOTHING )
729aaeb5
YO
3447 make_trie( pRExC_state,
3448 startbranch, first, cur, tail, count,
3449 trietype, depth+1 );
fab2782b 3450 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
8aa23a47 3451 }
fab2782b 3452 if ( noper_trietype
786e8c11 3453#ifdef NOJUMPTRIE
8aa23a47 3454 && noper_next == tail
786e8c11 3455#endif
8aa23a47 3456 ){
fab2782b 3457 /* noper is triable, so we can start a new trie sequence */
8aa23a47
YO
3458 count = 1;
3459 first = cur;
fab2782b
YO
3460 trietype = noper_trietype;
3461 } else if (first) {
3462 /* if we already saw a first but the current node is not triable then we have
3463 * to reset the first information. */
8aa23a47
YO
3464 count = 0;
3465 first = NULL;
fab2782b 3466 trietype = 0;
8aa23a47 3467 }
fab2782b
YO
3468 } /* end handle unmergable node */
3469 } /* loop over branches */
df826430 3470 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3471 regprop(RExC_rx, mysv, cur);
3472 PerlIO_printf( Perl_debug_log,
3473 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3474 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3475
3476 });
e6351b37 3477 if ( last && trietype ) {
3b6759a6
YO
3478 if ( trietype != NOTHING ) {
3479 /* the last branch of the sequence was part of a trie,
3480 * so we have to construct it here outside of the loop
3481 */
3482 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
686b73d4 3483#ifdef TRIE_STUDY_OPT
3b6759a6
YO
3484 if ( ((made == MADE_EXACT_TRIE &&
3485 startbranch == first)
3486 || ( first_non_open == first )) &&
3487 depth==0 ) {
3488 flags |= SCF_TRIE_RESTUDY;
3489 if ( startbranch == first
3490 && scan == tail )
3491 {
3492 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3493 }
8aa23a47 3494 }
3dab1dad 3495#endif
3b6759a6
YO
3496 } else {
3497 /* at this point we know whatever we have is a NOTHING sequence/branch
3498 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3499 */
3500 if ( startbranch == first ) {
3501 regnode *opt;
3502 /* the entire thing is a NOTHING sequence, something like this:
3503 * (?:|) So we can turn it into a plain NOTHING op. */
3504 DEBUG_TRIE_COMPILE_r({
3505 regprop(RExC_rx, mysv, cur);
3506 PerlIO_printf( Perl_debug_log,
3507 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3508 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3509
3510 });
3511 OP(startbranch)= NOTHING;
3512 NEXT_OFF(startbranch)= tail - startbranch;
3513 for ( opt= startbranch + 1; opt < tail ; opt++ )
3514 OP(opt)= OPTIMIZED;
3515 }
3516 }
fab2782b
YO
3517 } /* end if ( last) */
3518 } /* TRIE_MAXBUF is non zero */
8aa23a47
YO
3519
3520 } /* do trie */
3521
653099ff 3522 }
8aa23a47
YO
3523 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3524 scan = NEXTOPER(NEXTOPER(scan));
3525 } else /* single branch is optimized. */
3526 scan = NEXTOPER(scan);
3527 continue;
3528 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3529 scan_frame *newframe = NULL;
3530 I32 paren;
3531 regnode *start;
3532 regnode *end;
3533
3534 if (OP(scan) != SUSPEND) {
3535 /* set the pointer */
3536 if (OP(scan) == GOSUB) {
3537 paren = ARG(scan);
3538 RExC_recurse[ARG2L(scan)] = scan;
3539 start = RExC_open_parens[paren-1];
3540 end = RExC_close_parens[paren-1];
3541 } else {
3542 paren = 0;
f8fc2ecf 3543 start = RExC_rxi->program + 1;
8aa23a47
YO
3544 end = RExC_opend;
3545 }
3546 if (!recursed) {
3547 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3548 SAVEFREEPV(recursed);
3549 }
3550 if (!PAREN_TEST(recursed,paren+1)) {
3551 PAREN_SET(recursed,paren+1);
3552 Newx(newframe,1,scan_frame);
3553 } else {
3554 if (flags & SCF_DO_SUBSTR) {
304ee84b 3555 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3556 data->longest = &(data->longest_float);