This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c: Remove unreachable code
[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
fab2782b
YO
1469#define TRIE_READ_CHAR STMT_START { \
1470 wordlen++; \
1471 if ( UTF ) { \
1472 /* if it is UTF then it is either already folded, or does not need folding */ \
c80e42f3 1473 uvc = utf8n_to_uvchr( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
fab2782b
YO
1474 } \
1475 else if (folder == PL_fold_latin1) { \
7d006b13
KW
1476 /* This folder implies Unicode rules, which in the range expressible \
1477 * by not UTF is the lower case, with the two exceptions, one of \
1478 * which should have been taken care of before calling this */ \
1479 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
1480 uvc = toLOWER_L1(*uc); \
1481 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
1482 len = 1; \
fab2782b
YO
1483 } else { \
1484 /* raw data, will be folded later if needed */ \
1485 uvc = (U32)*uc; \
1486 len = 1; \
1487 } \
786e8c11
YO
1488} STMT_END
1489
1490
1491
1492#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1493 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1494 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1495 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1496 } \
1497 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1498 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1499 TRIE_LIST_CUR( state )++; \
1500} STMT_END
07be1b83 1501
786e8c11
YO
1502#define TRIE_LIST_NEW(state) STMT_START { \
1503 Newxz( trie->states[ state ].trans.list, \
1504 4, reg_trie_trans_le ); \
1505 TRIE_LIST_CUR( state ) = 1; \
1506 TRIE_LIST_LEN( state ) = 4; \
1507} STMT_END
07be1b83 1508
786e8c11
YO
1509#define TRIE_HANDLE_WORD(state) STMT_START { \
1510 U16 dupe= trie->states[ state ].wordnum; \
1511 regnode * const noper_next = regnext( noper ); \
1512 \
786e8c11
YO
1513 DEBUG_r({ \
1514 /* store the word for dumping */ \
1515 SV* tmp; \
1516 if (OP(noper) != NOTHING) \
740cce10 1517 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1518 else \
740cce10 1519 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1520 av_push( trie_words, tmp ); \
786e8c11
YO
1521 }); \
1522 \
1523 curword++; \
2e64971a
DM
1524 trie->wordinfo[curword].prev = 0; \
1525 trie->wordinfo[curword].len = wordlen; \
1526 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1527 \
1528 if ( noper_next < tail ) { \
1529 if (!trie->jump) \
c944940b 1530 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1531 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1532 if (!jumper) \
1533 jumper = noper_next; \
1534 if (!nextbranch) \
1535 nextbranch= regnext(cur); \
1536 } \
1537 \
1538 if ( dupe ) { \
2e64971a
DM
1539 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1540 /* chain, so that when the bits of chain are later */\
1541 /* linked together, the dups appear in the chain */\
1542 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1543 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1544 } else { \
1545 /* we haven't inserted this word yet. */ \
1546 trie->states[ state ].wordnum = curword; \
1547 } \
1548} STMT_END
07be1b83 1549
3dab1dad 1550
786e8c11
YO
1551#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1552 ( ( base + charid >= ucharcount \
1553 && base + charid < ubound \
1554 && state == trie->trans[ base - ucharcount + charid ].check \
1555 && trie->trans[ base - ucharcount + charid ].next ) \
1556 ? trie->trans[ base - ucharcount + charid ].next \
1557 : ( state==1 ? special : 0 ) \
1558 )
3dab1dad 1559
786e8c11
YO
1560#define MADE_TRIE 1
1561#define MADE_JUMP_TRIE 2
1562#define MADE_EXACT_TRIE 4
3dab1dad 1563
a3621e74 1564STATIC I32
786e8c11 1565S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1566{
27da23d5 1567 dVAR;
a3621e74
YO
1568 /* first pass, loop through and scan words */
1569 reg_trie_data *trie;
55eed653 1570 HV *widecharmap = NULL;
2b8b4781 1571 AV *revcharmap = newAV();
a3621e74 1572 regnode *cur;
9f7f3913 1573 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1574 STRLEN len = 0;
1575 UV uvc = 0;
1576 U16 curword = 0;
1577 U32 next_alloc = 0;
786e8c11
YO
1578 regnode *jumper = NULL;
1579 regnode *nextbranch = NULL;
7f69552c 1580 regnode *convert = NULL;
2e64971a 1581 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1582 /* we just use folder as a flag in utf8 */
1e696034 1583 const U8 * folder = NULL;
a3621e74 1584
2b8b4781
NC
1585#ifdef DEBUGGING
1586 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1587 AV *trie_words = NULL;
1588 /* along with revcharmap, this only used during construction but both are
1589 * useful during debugging so we store them in the struct when debugging.
8e11feef 1590 */
2b8b4781
NC
1591#else
1592 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1593 STRLEN trie_charcount=0;
3dab1dad 1594#endif
2b8b4781 1595 SV *re_trie_maxbuff;
a3621e74 1596 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1597
1598 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1599#ifndef DEBUGGING
1600 PERL_UNUSED_ARG(depth);
1601#endif
a3621e74 1602
1e696034 1603 switch (flags) {
79a81a6e 1604 case EXACT: break;
2f7f8cb1 1605 case EXACTFA:
fab2782b
YO
1606 case EXACTFU_SS:
1607 case EXACTFU_TRICKYFOLD:
1e696034
KW
1608 case EXACTFU: folder = PL_fold_latin1; break;
1609 case EXACTF: folder = PL_fold; break;
1610 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1611 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1612 }
1613
c944940b 1614 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1615 trie->refcount = 1;
3dab1dad 1616 trie->startstate = 1;
786e8c11 1617 trie->wordcount = word_count;
f8fc2ecf 1618 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1619 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1620 if (flags == EXACT)
c944940b 1621 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1622 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1623 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1624
a3621e74 1625 DEBUG_r({
2b8b4781 1626 trie_words = newAV();
a3621e74 1627 });
a3621e74 1628
0111c4fd 1629 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1630 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1631 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1632 }
df826430 1633 DEBUG_TRIE_COMPILE_r({
3dab1dad 1634 PerlIO_printf( Perl_debug_log,
786e8c11 1635 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1636 (int)depth * 2 + 2, "",
1637 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1638 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1639 (int)depth);
3dab1dad 1640 });
7f69552c
YO
1641
1642 /* Find the node we are going to overwrite */
1643 if ( first == startbranch && OP( last ) != BRANCH ) {
1644 /* whole branch chain */
1645 convert = first;
1646 } else {
1647 /* branch sub-chain */
1648 convert = NEXTOPER( first );
1649 }
1650
a3621e74
YO
1651 /* -- First loop and Setup --
1652
1653 We first traverse the branches and scan each word to determine if it
1654 contains widechars, and how many unique chars there are, this is
1655 important as we have to build a table with at least as many columns as we
1656 have unique chars.
1657
1658 We use an array of integers to represent the character codes 0..255
38a44b82 1659 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1660 native representation of the character value as the key and IV's for the
1661 coded index.
1662
1663 *TODO* If we keep track of how many times each character is used we can
1664 remap the columns so that the table compression later on is more
3b753521 1665 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1666 middle and the least common are on the outside. IMO this would be better
1667 than a most to least common mapping as theres a decent chance the most
1668 common letter will share a node with the least common, meaning the node
486ec47a 1669 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1670 case is when we have the least common nodes twice.
1671
1672 */
1673
a3621e74 1674 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1675 regnode *noper = NEXTOPER( cur );
e1ec3a88 1676 const U8 *uc = (U8*)STRING( noper );
df826430 1677 const U8 *e = uc + STR_LEN( noper );
a3621e74 1678 STRLEN foldlen = 0;
07be1b83 1679 U32 wordlen = 0; /* required init */
02daf0ab
YO
1680 STRLEN chars = 0;
1681 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1682
3dab1dad 1683 if (OP(noper) == NOTHING) {
df826430
YO
1684 regnode *noper_next= regnext(noper);
1685 if (noper_next != tail && OP(noper_next) == flags) {
1686 noper = noper_next;
1687 uc= (U8*)STRING(noper);
1688 e= uc + STR_LEN(noper);
1689 trie->minlen= STR_LEN(noper);
1690 } else {
1691 trie->minlen= 0;
1692 continue;
1693 }
3dab1dad 1694 }
df826430 1695
fab2782b 1696 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1697 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1698 regardless of encoding */
fab2782b
YO
1699 if (OP( noper ) == EXACTFU_SS) {
1700 /* false positives are ok, so just set this */
0dc4a61d 1701 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
fab2782b
YO
1702 }
1703 }
a3621e74 1704 for ( ; uc < e ; uc += len ) {
3dab1dad 1705 TRIE_CHARCOUNT(trie)++;
a3621e74 1706 TRIE_READ_CHAR;
3dab1dad 1707 chars++;
a3621e74 1708 if ( uvc < 256 ) {
fab2782b
YO
1709 if ( folder ) {
1710 U8 folded= folder[ (U8) uvc ];
1711 if ( !trie->charmap[ folded ] ) {
1712 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1713 TRIE_STORE_REVCHAR( folded );
1714 }
1715 }
a3621e74
YO
1716 if ( !trie->charmap[ uvc ] ) {
1717 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1718 TRIE_STORE_REVCHAR( uvc );
a3621e74 1719 }
02daf0ab 1720 if ( set_bit ) {
62012aee
KW
1721 /* store the codepoint in the bitmap, and its folded
1722 * equivalent. */
fab2782b 1723 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1724
1725 /* store the folded codepoint */
fab2782b 1726 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1727
1728 if ( !UTF ) {
1729 /* store first byte of utf8 representation of
acdf4139 1730 variant codepoints */
f4cd282c 1731 if (! NATIVE_IS_INVARIANT(uvc)) {
acdf4139 1732 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1733 }
1734 }
02daf0ab
YO
1735 set_bit = 0; /* We've done our bit :-) */
1736 }
a3621e74
YO
1737 } else {
1738 SV** svpp;
55eed653
NC
1739 if ( !widecharmap )
1740 widecharmap = newHV();
a3621e74 1741
55eed653 1742 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1743
1744 if ( !svpp )
e4584336 1745 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1746
1747 if ( !SvTRUE( *svpp ) ) {
1748 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1749 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1750 }
1751 }
1752 }
3dab1dad 1753 if( cur == first ) {
fab2782b
YO
1754 trie->minlen = chars;
1755 trie->maxlen = chars;
3dab1dad 1756 } else if (chars < trie->minlen) {
fab2782b 1757 trie->minlen = chars;
3dab1dad 1758 } else if (chars > trie->maxlen) {
fab2782b
YO
1759 trie->maxlen = chars;
1760 }
1761 if (OP( noper ) == EXACTFU_SS) {
1762 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1763 if (trie->minlen > 1)
1764 trie->minlen= 1;
1765 }
1766 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1767 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1768 * - We assume that any such sequence might match a 2 byte string */
1769 if (trie->minlen > 2 )
1770 trie->minlen= 2;
3dab1dad
YO
1771 }
1772
a3621e74
YO
1773 } /* end first pass */
1774 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1775 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1776 (int)depth * 2 + 2,"",
55eed653 1777 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1778 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1779 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1780 );
a3621e74
YO
1781
1782 /*
1783 We now know what we are dealing with in terms of unique chars and
1784 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1785 representation using a flat table will take. If it's over a reasonable
1786 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1787 conservative but potentially much slower representation using an array
1788 of lists.
1789
1790 At the end we convert both representations into the same compressed
1791 form that will be used in regexec.c for matching with. The latter
1792 is a form that cannot be used to construct with but has memory
1793 properties similar to the list form and access properties similar
1794 to the table form making it both suitable for fast searches and
1795 small enough that its feasable to store for the duration of a program.
1796
1797 See the comment in the code where the compressed table is produced
1798 inplace from the flat tabe representation for an explanation of how
1799 the compression works.
1800
1801 */
1802
1803
2e64971a
DM
1804 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1805 prev_states[1] = 0;
1806
3dab1dad 1807 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1808 /*
1809 Second Pass -- Array Of Lists Representation
1810
1811 Each state will be represented by a list of charid:state records
1812 (reg_trie_trans_le) the first such element holds the CUR and LEN
1813 points of the allocated array. (See defines above).
1814
1815 We build the initial structure using the lists, and then convert
1816 it into the compressed table form which allows faster lookups
1817 (but cant be modified once converted).
a3621e74
YO
1818 */
1819
a3621e74
YO
1820 STRLEN transcount = 1;
1821
1e2e3d02
YO
1822 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1823 "%*sCompiling trie using list compiler\n",
1824 (int)depth * 2 + 2, ""));
686b73d4 1825
c944940b
JH
1826 trie->states = (reg_trie_state *)
1827 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1828 sizeof(reg_trie_state) );
a3621e74
YO
1829 TRIE_LIST_NEW(1);
1830 next_alloc = 2;
1831
1832 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1833
df826430 1834 regnode *noper = NEXTOPER( cur );
c445ea15 1835 U8 *uc = (U8*)STRING( noper );
df826430 1836 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1837 U32 state = 1; /* required init */
1838 U16 charid = 0; /* sanity init */
07be1b83 1839 U32 wordlen = 0; /* required init */
c445ea15 1840
df826430
YO
1841 if (OP(noper) == NOTHING) {
1842 regnode *noper_next= regnext(noper);
1843 if (noper_next != tail && OP(noper_next) == flags) {
1844 noper = noper_next;
1845 uc= (U8*)STRING(noper);
1846 e= uc + STR_LEN(noper);
1847 }
1848 }
1849
3dab1dad 1850 if (OP(noper) != NOTHING) {
786e8c11 1851 for ( ; uc < e ; uc += len ) {
c445ea15 1852
786e8c11 1853 TRIE_READ_CHAR;
c445ea15 1854
786e8c11
YO
1855 if ( uvc < 256 ) {
1856 charid = trie->charmap[ uvc ];
c445ea15 1857 } else {
55eed653 1858 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1859 if ( !svpp ) {
1860 charid = 0;
1861 } else {
1862 charid=(U16)SvIV( *svpp );
1863 }
c445ea15 1864 }
786e8c11
YO
1865 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1866 if ( charid ) {
a3621e74 1867
786e8c11
YO
1868 U16 check;
1869 U32 newstate = 0;
a3621e74 1870
786e8c11
YO
1871 charid--;
1872 if ( !trie->states[ state ].trans.list ) {
1873 TRIE_LIST_NEW( state );
c445ea15 1874 }
786e8c11
YO
1875 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1876 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1877 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1878 break;
1879 }
1880 }
1881 if ( ! newstate ) {
1882 newstate = next_alloc++;
2e64971a 1883 prev_states[newstate] = state;
786e8c11
YO
1884 TRIE_LIST_PUSH( state, charid, newstate );
1885 transcount++;
1886 }
1887 state = newstate;
1888 } else {
1889 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1890 }
a28509cc 1891 }
c445ea15 1892 }
3dab1dad 1893 TRIE_HANDLE_WORD(state);
a3621e74
YO
1894
1895 } /* end second pass */
1896
1e2e3d02
YO
1897 /* next alloc is the NEXT state to be allocated */
1898 trie->statecount = next_alloc;
c944940b
JH
1899 trie->states = (reg_trie_state *)
1900 PerlMemShared_realloc( trie->states,
1901 next_alloc
1902 * sizeof(reg_trie_state) );
a3621e74 1903
3dab1dad 1904 /* and now dump it out before we compress it */
2b8b4781
NC
1905 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1906 revcharmap, next_alloc,
1907 depth+1)
1e2e3d02 1908 );
a3621e74 1909
c944940b
JH
1910 trie->trans = (reg_trie_trans *)
1911 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1912 {
1913 U32 state;
a3621e74
YO
1914 U32 tp = 0;
1915 U32 zp = 0;
1916
1917
1918 for( state=1 ; state < next_alloc ; state ++ ) {
1919 U32 base=0;
1920
1921 /*
1922 DEBUG_TRIE_COMPILE_MORE_r(
1923 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1924 );
1925 */
1926
1927 if (trie->states[state].trans.list) {
1928 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1929 U16 maxid=minid;
a28509cc 1930 U16 idx;
a3621e74
YO
1931
1932 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1933 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1934 if ( forid < minid ) {
1935 minid=forid;
1936 } else if ( forid > maxid ) {
1937 maxid=forid;
1938 }
a3621e74
YO
1939 }
1940 if ( transcount < tp + maxid - minid + 1) {
1941 transcount *= 2;
c944940b
JH
1942 trie->trans = (reg_trie_trans *)
1943 PerlMemShared_realloc( trie->trans,
446bd890
NC
1944 transcount
1945 * sizeof(reg_trie_trans) );
a3621e74
YO
1946 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1947 }
1948 base = trie->uniquecharcount + tp - minid;
1949 if ( maxid == minid ) {
1950 U32 set = 0;
1951 for ( ; zp < tp ; zp++ ) {
1952 if ( ! trie->trans[ zp ].next ) {
1953 base = trie->uniquecharcount + zp - minid;
1954 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1955 trie->trans[ zp ].check = state;
1956 set = 1;
1957 break;
1958 }
1959 }
1960 if ( !set ) {
1961 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1962 trie->trans[ tp ].check = state;
1963 tp++;
1964 zp = tp;
1965 }
1966 } else {
1967 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 1968 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
1969 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1970 trie->trans[ tid ].check = state;
1971 }
1972 tp += ( maxid - minid + 1 );
1973 }
1974 Safefree(trie->states[ state ].trans.list);
1975 }
1976 /*
1977 DEBUG_TRIE_COMPILE_MORE_r(
1978 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1979 );
1980 */
1981 trie->states[ state ].trans.base=base;
1982 }
cc601c31 1983 trie->lasttrans = tp + 1;
a3621e74
YO
1984 }
1985 } else {
1986 /*
1987 Second Pass -- Flat Table Representation.
1988
1989 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1990 We know that we will need Charcount+1 trans at most to store the data
1991 (one row per char at worst case) So we preallocate both structures
1992 assuming worst case.
1993
1994 We then construct the trie using only the .next slots of the entry
1995 structs.
1996
3b753521 1997 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
1998 make compression both faster and easier by keeping track of how many non
1999 zero fields are in the node.
2000
2001 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2002 transition.
2003
2004 There are two terms at use here: state as a TRIE_NODEIDX() which is a
2005 number representing the first entry of the node, and state as a
2006 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
2007 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
2008 are 2 entrys per node. eg:
2009
2010 A B A B
2011 1. 2 4 1. 3 7
2012 2. 0 3 3. 0 5
2013 3. 0 0 5. 0 0
2014 4. 0 0 7. 0 0
2015
2016 The table is internally in the right hand, idx form. However as we also
2017 have to deal with the states array which is indexed by nodenum we have to
2018 use TRIE_NODENUM() to convert.
2019
2020 */
1e2e3d02
YO
2021 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2022 "%*sCompiling trie using table compiler\n",
2023 (int)depth * 2 + 2, ""));
3dab1dad 2024
c944940b
JH
2025 trie->trans = (reg_trie_trans *)
2026 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2027 * trie->uniquecharcount + 1,
2028 sizeof(reg_trie_trans) );
2029 trie->states = (reg_trie_state *)
2030 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2031 sizeof(reg_trie_state) );
a3621e74
YO
2032 next_alloc = trie->uniquecharcount + 1;
2033
3dab1dad 2034
a3621e74
YO
2035 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2036
df826430 2037 regnode *noper = NEXTOPER( cur );
a28509cc 2038 const U8 *uc = (U8*)STRING( noper );
df826430 2039 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
2040
2041 U32 state = 1; /* required init */
2042
2043 U16 charid = 0; /* sanity init */
2044 U32 accept_state = 0; /* sanity init */
a3621e74 2045
07be1b83 2046 U32 wordlen = 0; /* required init */
a3621e74 2047
df826430
YO
2048 if (OP(noper) == NOTHING) {
2049 regnode *noper_next= regnext(noper);
2050 if (noper_next != tail && OP(noper_next) == flags) {
2051 noper = noper_next;
2052 uc= (U8*)STRING(noper);
2053 e= uc + STR_LEN(noper);
2054 }
2055 }
fab2782b 2056
3dab1dad 2057 if ( OP(noper) != NOTHING ) {
786e8c11 2058 for ( ; uc < e ; uc += len ) {
a3621e74 2059
786e8c11 2060 TRIE_READ_CHAR;
a3621e74 2061
786e8c11
YO
2062 if ( uvc < 256 ) {
2063 charid = trie->charmap[ uvc ];
2064 } else {
55eed653 2065 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2066 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2067 }
786e8c11
YO
2068 if ( charid ) {
2069 charid--;
2070 if ( !trie->trans[ state + charid ].next ) {
2071 trie->trans[ state + charid ].next = next_alloc;
2072 trie->trans[ state ].check++;
2e64971a
DM
2073 prev_states[TRIE_NODENUM(next_alloc)]
2074 = TRIE_NODENUM(state);
786e8c11
YO
2075 next_alloc += trie->uniquecharcount;
2076 }
2077 state = trie->trans[ state + charid ].next;
2078 } else {
2079 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2080 }
2081 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2082 }
a3621e74 2083 }
3dab1dad
YO
2084 accept_state = TRIE_NODENUM( state );
2085 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2086
2087 } /* end second pass */
2088
3dab1dad 2089 /* and now dump it out before we compress it */
2b8b4781
NC
2090 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2091 revcharmap,
2092 next_alloc, depth+1));
a3621e74 2093
a3621e74
YO
2094 {
2095 /*
2096 * Inplace compress the table.*
2097
2098 For sparse data sets the table constructed by the trie algorithm will
2099 be mostly 0/FAIL transitions or to put it another way mostly empty.
2100 (Note that leaf nodes will not contain any transitions.)
2101
2102 This algorithm compresses the tables by eliminating most such
2103 transitions, at the cost of a modest bit of extra work during lookup:
2104
2105 - Each states[] entry contains a .base field which indicates the
2106 index in the state[] array wheres its transition data is stored.
2107
3b753521 2108 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2109
2110 - If .base is nonzero then charid is added to it to find an entry in
2111 the trans array.
2112
2113 -If trans[states[state].base+charid].check!=state then the
2114 transition is taken to be a 0/Fail transition. Thus if there are fail
2115 transitions at the front of the node then the .base offset will point
2116 somewhere inside the previous nodes data (or maybe even into a node
2117 even earlier), but the .check field determines if the transition is
2118 valid.
2119
786e8c11 2120 XXX - wrong maybe?
a3621e74 2121 The following process inplace converts the table to the compressed
3b753521 2122 table: We first do not compress the root node 1,and mark all its
a3621e74 2123 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2124 allows us to do a DFA construction from the compressed table later,
2125 and ensures that any .base pointers we calculate later are greater
2126 than 0.
a3621e74
YO
2127
2128 - We set 'pos' to indicate the first entry of the second node.
2129
2130 - We then iterate over the columns of the node, finding the first and
2131 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2132 and set the .check pointers accordingly, and advance pos
2133 appropriately and repreat for the next node. Note that when we copy
2134 the next pointers we have to convert them from the original
2135 NODEIDX form to NODENUM form as the former is not valid post
2136 compression.
2137
2138 - If a node has no transitions used we mark its base as 0 and do not
2139 advance the pos pointer.
2140
2141 - If a node only has one transition we use a second pointer into the
2142 structure to fill in allocated fail transitions from other states.
2143 This pointer is independent of the main pointer and scans forward
2144 looking for null transitions that are allocated to a state. When it
2145 finds one it writes the single transition into the "hole". If the
786e8c11 2146 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2147
2148 - Once compressed we can Renew/realloc the structures to release the
2149 excess space.
2150
2151 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2152 specifically Fig 3.47 and the associated pseudocode.
2153
2154 demq
2155 */
a3b680e6 2156 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2157 U32 state, charid;
a3621e74 2158 U32 pos = 0, zp=0;
1e2e3d02 2159 trie->statecount = laststate;
a3621e74
YO
2160
2161 for ( state = 1 ; state < laststate ; state++ ) {
2162 U8 flag = 0;
a28509cc
AL
2163 const U32 stateidx = TRIE_NODEIDX( state );
2164 const U32 o_used = trie->trans[ stateidx ].check;
2165 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2166 trie->trans[ stateidx ].check = 0;
2167
2168 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2169 if ( flag || trie->trans[ stateidx + charid ].next ) {
2170 if ( trie->trans[ stateidx + charid ].next ) {
2171 if (o_used == 1) {
2172 for ( ; zp < pos ; zp++ ) {
2173 if ( ! trie->trans[ zp ].next ) {
2174 break;
2175 }
2176 }
2177 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2178 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2179 trie->trans[ zp ].check = state;
2180 if ( ++zp > pos ) pos = zp;
2181 break;
2182 }
2183 used--;
2184 }
2185 if ( !flag ) {
2186 flag = 1;
2187 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2188 }
2189 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2190 trie->trans[ pos ].check = state;
2191 pos++;
2192 }
2193 }
2194 }
cc601c31 2195 trie->lasttrans = pos + 1;
c944940b
JH
2196 trie->states = (reg_trie_state *)
2197 PerlMemShared_realloc( trie->states, laststate
2198 * sizeof(reg_trie_state) );
a3621e74 2199 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2200 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2201 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2202 (int)depth * 2 + 2,"",
2203 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2204 (IV)next_alloc,
2205 (IV)pos,
a3621e74
YO
2206 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2207 );
2208
2209 } /* end table compress */
2210 }
1e2e3d02
YO
2211 DEBUG_TRIE_COMPILE_MORE_r(
2212 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2213 (int)depth * 2 + 2, "",
2214 (UV)trie->statecount,
2215 (UV)trie->lasttrans)
2216 );
cc601c31 2217 /* resize the trans array to remove unused space */
c944940b
JH
2218 trie->trans = (reg_trie_trans *)
2219 PerlMemShared_realloc( trie->trans, trie->lasttrans
2220 * sizeof(reg_trie_trans) );
a3621e74 2221
3b753521 2222 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2223 U8 nodetype =(U8)(flags & 0xFF);
2224 char *str=NULL;
786e8c11 2225
07be1b83 2226#ifdef DEBUGGING
e62cc96a 2227 regnode *optimize = NULL;
7122b237
YO
2228#ifdef RE_TRACK_PATTERN_OFFSETS
2229
b57a0404
JH
2230 U32 mjd_offset = 0;
2231 U32 mjd_nodelen = 0;
7122b237
YO
2232#endif /* RE_TRACK_PATTERN_OFFSETS */
2233#endif /* DEBUGGING */
a3621e74 2234 /*
3dab1dad
YO
2235 This means we convert either the first branch or the first Exact,
2236 depending on whether the thing following (in 'last') is a branch
2237 or not and whther first is the startbranch (ie is it a sub part of
2238 the alternation or is it the whole thing.)
3b753521 2239 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2240 the whole branch sequence, including the first.
a3621e74 2241 */
3dab1dad 2242 /* Find the node we are going to overwrite */
7f69552c 2243 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2244 /* branch sub-chain */
3dab1dad 2245 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2246#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2247 DEBUG_r({
2248 mjd_offset= Node_Offset((convert));
2249 mjd_nodelen= Node_Length((convert));
2250 });
7122b237 2251#endif
7f69552c 2252 /* whole branch chain */
7122b237
YO
2253 }
2254#ifdef RE_TRACK_PATTERN_OFFSETS
2255 else {
7f69552c
YO
2256 DEBUG_r({
2257 const regnode *nop = NEXTOPER( convert );
2258 mjd_offset= Node_Offset((nop));
2259 mjd_nodelen= Node_Length((nop));
2260 });
07be1b83
YO
2261 }
2262 DEBUG_OPTIMISE_r(
2263 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2264 (int)depth * 2 + 2, "",
786e8c11 2265 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2266 );
7122b237 2267#endif
3dab1dad
YO
2268 /* But first we check to see if there is a common prefix we can
2269 split out as an EXACT and put in front of the TRIE node. */
2270 trie->startstate= 1;
55eed653 2271 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2272 U32 state;
1e2e3d02 2273 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2274 U32 ofs = 0;
8e11feef
RGS
2275 I32 idx = -1;
2276 U32 count = 0;
2277 const U32 base = trie->states[ state ].trans.base;
a3621e74 2278
3dab1dad 2279 if ( trie->states[state].wordnum )
8e11feef 2280 count = 1;
a3621e74 2281
8e11feef 2282 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2283 if ( ( base + ofs >= trie->uniquecharcount ) &&
2284 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2285 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2286 {
3dab1dad 2287 if ( ++count > 1 ) {
2b8b4781 2288 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2289 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2290 if ( state == 1 ) break;
3dab1dad
YO
2291 if ( count == 2 ) {
2292 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2293 DEBUG_OPTIMISE_r(
8e11feef
RGS
2294 PerlIO_printf(Perl_debug_log,
2295 "%*sNew Start State=%"UVuf" Class: [",
2296 (int)depth * 2 + 2, "",
786e8c11 2297 (UV)state));
be8e71aa 2298 if (idx >= 0) {
2b8b4781 2299 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2300 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2301
3dab1dad 2302 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2303 if ( folder )
2304 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2305 DEBUG_OPTIMISE_r(
f1f66076 2306 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2307 );
8e11feef
RGS
2308 }
2309 }
2310 TRIE_BITMAP_SET(trie,*ch);
2311 if ( folder )
2312 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2313 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2314 }
2315 idx = ofs;
2316 }
3dab1dad
YO
2317 }
2318 if ( count == 1 ) {
2b8b4781 2319 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2320 STRLEN len;
2321 char *ch = SvPV( *tmp, len );
de734bd5
A
2322 DEBUG_OPTIMISE_r({
2323 SV *sv=sv_newmortal();
8e11feef
RGS
2324 PerlIO_printf( Perl_debug_log,
2325 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2326 (int)depth * 2 + 2, "",
de734bd5
A
2327 (UV)state, (UV)idx,
2328 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2329 PL_colors[0], PL_colors[1],
2330 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2331 PERL_PV_ESCAPE_FIRSTCHAR
2332 )
2333 );
2334 });
3dab1dad
YO
2335 if ( state==1 ) {
2336 OP( convert ) = nodetype;
2337 str=STRING(convert);
2338 STR_LEN(convert)=0;
2339 }
c490c714
YO
2340 STR_LEN(convert) += len;
2341 while (len--)
de734bd5 2342 *str++ = *ch++;
8e11feef 2343 } else {
f9049ba1 2344#ifdef DEBUGGING
8e11feef
RGS
2345 if (state>1)
2346 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2347#endif
8e11feef
RGS
2348 break;
2349 }
2350 }
2e64971a 2351 trie->prefixlen = (state-1);
3dab1dad 2352 if (str) {
8e11feef 2353 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2354 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2355 trie->startstate = state;
07be1b83
YO
2356 trie->minlen -= (state - 1);
2357 trie->maxlen -= (state - 1);
33809eae
JH
2358#ifdef DEBUGGING
2359 /* At least the UNICOS C compiler choked on this
2360 * being argument to DEBUG_r(), so let's just have
2361 * it right here. */
2362 if (
2363#ifdef PERL_EXT_RE_BUILD
2364 1
2365#else
2366 DEBUG_r_TEST
2367#endif
2368 ) {
2369 regnode *fix = convert;
2370 U32 word = trie->wordcount;
2371 mjd_nodelen++;
2372 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2373 while( ++fix < n ) {
2374 Set_Node_Offset_Length(fix, 0, 0);
2375 }
2376 while (word--) {
2377 SV ** const tmp = av_fetch( trie_words, word, 0 );
2378 if (tmp) {
2379 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2380 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2381 else
2382 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2383 }
2384 }
2385 }
2386#endif
8e11feef
RGS
2387 if (trie->maxlen) {
2388 convert = n;
2389 } else {
3dab1dad 2390 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2391 DEBUG_r(optimize= n);
3dab1dad
YO
2392 }
2393 }
2394 }
a5ca303d
YO
2395 if (!jumper)
2396 jumper = last;
3dab1dad 2397 if ( trie->maxlen ) {
8e11feef
RGS
2398 NEXT_OFF( convert ) = (U16)(tail - convert);
2399 ARG_SET( convert, data_slot );
786e8c11
YO
2400 /* Store the offset to the first unabsorbed branch in
2401 jump[0], which is otherwise unused by the jump logic.
2402 We use this when dumping a trie and during optimisation. */
2403 if (trie->jump)
7f69552c 2404 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2405
6c48061a
YO
2406 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2407 * and there is a bitmap
2408 * and the first "jump target" node we found leaves enough room
2409 * then convert the TRIE node into a TRIEC node, with the bitmap
2410 * embedded inline in the opcode - this is hypothetically faster.
2411 */
2412 if ( !trie->states[trie->startstate].wordnum
2413 && trie->bitmap
2414 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2415 {
2416 OP( convert ) = TRIEC;
2417 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2418 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2419 trie->bitmap= NULL;
2420 } else
2421 OP( convert ) = TRIE;
a3621e74 2422
3dab1dad
YO
2423 /* store the type in the flags */
2424 convert->flags = nodetype;
a5ca303d
YO
2425 DEBUG_r({
2426 optimize = convert
2427 + NODE_STEP_REGNODE
2428 + regarglen[ OP( convert ) ];
2429 });
2430 /* XXX We really should free up the resource in trie now,
2431 as we won't use them - (which resources?) dmq */
3dab1dad 2432 }
a3621e74 2433 /* needed for dumping*/
e62cc96a 2434 DEBUG_r(if (optimize) {
07be1b83 2435 regnode *opt = convert;
bcdf7404 2436
e62cc96a 2437 while ( ++opt < optimize) {
07be1b83
YO
2438 Set_Node_Offset_Length(opt,0,0);
2439 }
786e8c11
YO
2440 /*
2441 Try to clean up some of the debris left after the
2442 optimisation.
a3621e74 2443 */
786e8c11 2444 while( optimize < jumper ) {
07be1b83 2445 mjd_nodelen += Node_Length((optimize));
a3621e74 2446 OP( optimize ) = OPTIMIZED;
07be1b83 2447 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2448 optimize++;
2449 }
07be1b83 2450 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2451 });
2452 } /* end node insert */
2e64971a
DM
2453
2454 /* Finish populating the prev field of the wordinfo array. Walk back
2455 * from each accept state until we find another accept state, and if
2456 * so, point the first word's .prev field at the second word. If the
2457 * second already has a .prev field set, stop now. This will be the
2458 * case either if we've already processed that word's accept state,
3b753521
FN
2459 * or that state had multiple words, and the overspill words were
2460 * already linked up earlier.
2e64971a
DM
2461 */
2462 {
2463 U16 word;
2464 U32 state;
2465 U16 prev;
2466
2467 for (word=1; word <= trie->wordcount; word++) {
2468 prev = 0;
2469 if (trie->wordinfo[word].prev)
2470 continue;
2471 state = trie->wordinfo[word].accept;
2472 while (state) {
2473 state = prev_states[state];
2474 if (!state)
2475 break;
2476 prev = trie->states[state].wordnum;
2477 if (prev)
2478 break;
2479 }
2480 trie->wordinfo[word].prev = prev;
2481 }
2482 Safefree(prev_states);
2483 }
2484
2485
2486 /* and now dump out the compressed format */
2487 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2488
55eed653 2489 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2490#ifdef DEBUGGING
2491 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2492 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2493#else
03e70be4 2494 SvREFCNT_dec_NN(revcharmap);
07be1b83 2495#endif
786e8c11
YO
2496 return trie->jump
2497 ? MADE_JUMP_TRIE
2498 : trie->startstate>1
2499 ? MADE_EXACT_TRIE
2500 : MADE_TRIE;
2501}
2502
2503STATIC void
2504S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2505{
3b753521 2506/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2507
2508 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2509 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2510 ISBN 0-201-10088-6
2511
2512 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2513 suffix of the current state's 'word' that is also a proper prefix of another word in our
2514 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2515 the DFA not to have to restart after its tried and failed a word at a given point, it
2516 simply continues as though it had been matching the other word in the first place.
2517 Consider
2518 'abcdgu'=~/abcdefg|cdgu/
2519 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2520 fail, which would bring us to the state representing 'd' in the second word where we would
2521 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2522 */
2523 /* add a fail transition */
3251b653
NC
2524 const U32 trie_offset = ARG(source);
2525 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2526 U32 *q;
2527 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2528 const U32 numstates = trie->statecount;
786e8c11
YO
2529 const U32 ubound = trie->lasttrans + ucharcount;
2530 U32 q_read = 0;
2531 U32 q_write = 0;
2532 U32 charid;
2533 U32 base = trie->states[ 1 ].trans.base;
2534 U32 *fail;
2535 reg_ac_data *aho;
2536 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2537 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2538
2539 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2540#ifndef DEBUGGING
2541 PERL_UNUSED_ARG(depth);
2542#endif
2543
2544
2545 ARG_SET( stclass, data_slot );
c944940b 2546 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2547 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2548 aho->trie=trie_offset;
446bd890
NC
2549 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2550 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2551 Newxz( q, numstates, U32);
c944940b 2552 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2553 aho->refcount = 1;
2554 fail = aho->fail;
2555 /* initialize fail[0..1] to be 1 so that we always have
2556 a valid final fail state */
2557 fail[ 0 ] = fail[ 1 ] = 1;
2558
2559 for ( charid = 0; charid < ucharcount ; charid++ ) {
2560 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2561 if ( newstate ) {
2562 q[ q_write ] = newstate;
2563 /* set to point at the root */
2564 fail[ q[ q_write++ ] ]=1;
2565 }
2566 }
2567 while ( q_read < q_write) {
2568 const U32 cur = q[ q_read++ % numstates ];
2569 base = trie->states[ cur ].trans.base;
2570
2571 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2572 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2573 if (ch_state) {
2574 U32 fail_state = cur;
2575 U32 fail_base;
2576 do {
2577 fail_state = fail[ fail_state ];
2578 fail_base = aho->states[ fail_state ].trans.base;
2579 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2580
2581 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2582 fail[ ch_state ] = fail_state;
2583 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2584 {
2585 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2586 }
2587 q[ q_write++ % numstates] = ch_state;
2588 }
2589 }
2590 }
2591 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2592 when we fail in state 1, this allows us to use the
2593 charclass scan to find a valid start char. This is based on the principle
2594 that theres a good chance the string being searched contains lots of stuff
2595 that cant be a start char.
2596 */
2597 fail[ 0 ] = fail[ 1 ] = 0;
2598 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2599 PerlIO_printf(Perl_debug_log,
2600 "%*sStclass Failtable (%"UVuf" states): 0",
2601 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2602 );
786e8c11
YO
2603 for( q_read=1; q_read<numstates; q_read++ ) {
2604 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2605 }
2606 PerlIO_printf(Perl_debug_log, "\n");
2607 });
2608 Safefree(q);
2609 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2610}
2611
786e8c11 2612
a3621e74 2613/*
5d1c421c
JH
2614 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2615 * These need to be revisited when a newer toolchain becomes available.
2616 */
2617#if defined(__sparc64__) && defined(__GNUC__)
2618# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2619# undef SPARC64_GCC_WORKAROUND
2620# define SPARC64_GCC_WORKAROUND 1
2621# endif
2622#endif
2623
07be1b83 2624#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2625 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2626 SV * const mysv=sv_newmortal(); \
2627 regnode *Next = regnext(scan); \
2628 regprop(RExC_rx, mysv, scan); \
7f69552c 2629 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2630 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2631 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2632 }});
07be1b83 2633
1de06328 2634
bb914485 2635/* The below joins as many adjacent EXACTish nodes as possible into a single
0a982f06
KW
2636 * one. The regop may be changed if the node(s) contain certain sequences that
2637 * require special handling. The joining is only done if:
bb914485
KW
2638 * 1) there is room in the current conglomerated node to entirely contain the
2639 * next one.
2640 * 2) they are the exact same node type
2641 *
87b8b349 2642 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
2643 * these get optimized out
2644 *
0a982f06
KW
2645 * If a node is to match under /i (folded), the number of characters it matches
2646 * can be different than its character length if it contains a multi-character
2647 * fold. *min_subtract is set to the total delta of the input nodes.
bb914485 2648 *
a0c4c608
KW
2649 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2650 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2651 *
bb914485 2652 * This is as good a place as any to discuss the design of handling these
0a982f06
KW
2653 * multi-character fold sequences. It's been wrong in Perl for a very long
2654 * time. There are three code points in Unicode whose multi-character folds
2655 * were long ago discovered to mess things up. The previous designs for
2656 * dealing with these involved assigning a special node for them. This
2657 * approach doesn't work, as evidenced by this example:
a0c4c608 2658 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
0a982f06
KW
2659 * Both these fold to "sss", but if the pattern is parsed to create a node that
2660 * would match just the \xDF, it won't be able to handle the case where a
bb914485
KW
2661 * successful match would have to cross the node's boundary. The new approach
2662 * that hopefully generally solves the problem generates an EXACTFU_SS node
2663 * that is "sss".
2664 *
0a982f06
KW
2665 * It turns out that there are problems with all multi-character folds, and not
2666 * just these three. Now the code is general, for all such cases, but the
2667 * three still have some special handling. The approach taken is:
2668 * 1) This routine examines each EXACTFish node that could contain multi-
2669 * character fold sequences. It returns in *min_subtract how much to
9d071ca8 2670 * subtract from the the actual length of the string to get a real minimum
0a982f06
KW
2671 * match length; it is 0 if there are no multi-char folds. This delta is
2672 * used by the caller to adjust the min length of the match, and the delta
2673 * between min and max, so that the optimizer doesn't reject these
2674 * possibilities based on size constraints.
2675 * 2) Certain of these sequences require special handling by the trie code,
2676 * so, if found, this code changes the joined node type to special ops:
2677 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2678 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2679 * is used for an EXACTFU node that contains at least one "ss" sequence in
2680 * it. For non-UTF-8 patterns and strings, this is the only case where
2681 * there is a possible fold length change. That means that a regular
2682 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2683 * with length changes, and so can be processed faster. regexec.c takes
2684 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2685 * pre-folded by regcomp.c. This saves effort in regex matching.
87b8b349 2686 * However, the pre-folding isn't done for non-UTF8 patterns because the
0a982f06
KW
2687 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2688 * down by forcing the pattern into UTF8 unless necessary. Also what
2689 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2690 * possibilities for the non-UTF8 patterns are quite simple, except for
2691 * the sharp s. All the ones that don't involve a UTF-8 target string are
2692 * members of a fold-pair, and arrays are set up for all of them so that
2693 * the other member of the pair can be found quickly. Code elsewhere in
2694 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2695 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2696 * described in the next item.
1ca267a5
KW
2697 * 4) A problem remains for the sharp s in EXACTF and EXACTFA nodes when the
2698 * pattern isn't in UTF-8. (BTW, there cannot be an EXACTF node with a
2699 * UTF-8 pattern.) An assumption that the optimizer part of regexec.c
2700 * (probably unwittingly, in Perl_regexec_flags()) makes is that a
2701 * character in the pattern corresponds to at most a single character in
2702 * the target string. (And I do mean character, and not byte here, unlike
2703 * other parts of the documentation that have never been updated to
2704 * account for multibyte Unicode.) sharp s in EXACTF nodes can match the
2705 * two character string 'ss'; in EXACTFA nodes it can match
2706 * "\x{17F}\x{17F}". These violate the assumption, and they are the only
2707 * instances where it is violated. I'm reluctant to try to change the
2708 * assumption, as the code involved is impenetrable to me (khw), so
2709 * instead the code here punts. This routine examines (when the pattern
2710 * isn't UTF-8) EXACTF and EXACTFA nodes for the sharp s, and returns a
2711 * boolean indicating whether or not the node contains a sharp s. When it
2712 * is true, the caller sets a flag that later causes the optimizer in this
2713 * file to not set values for the floating and fixed string lengths, and
2714 * thus avoids the optimizer code in regexec.c that makes the invalid
2715 * assumption. Thus, there is no optimization based on string lengths for
2716 * non-UTF8-pattern EXACTF and EXACTFA nodes that contain the sharp s.
2717 * (The reason the assumption is wrong only in these two cases is that all
2718 * other non-UTF-8 folds are 1-1; and, for UTF-8 patterns, we pre-fold all
2719 * other folds to their expanded versions. We can't prefold sharp s to
2720 * 'ss' in EXACTF nodes because we don't know at compile time if it
2721 * actually matches 'ss' or not. It will match iff the target string is
2722 * in UTF-8, unlike the EXACTFU nodes, where it always matches; and
2723 * EXACTFA and EXACTFL where it never does. In an EXACTFA node in a UTF-8
2724 * pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the problem;
2725 * but in a non-UTF8 pattern, folding it to that above-Latin1 string would
2726 * require the pattern to be forced into UTF-8, the overhead of which we
2727 * want to avoid.)
bb914485 2728 */
1de06328 2729
9d071ca8 2730#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2731 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2732 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2733
be8e71aa 2734STATIC U32
9d071ca8 2735S_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
2736 /* Merge several consecutive EXACTish nodes into one. */
2737 regnode *n = regnext(scan);
2738 U32 stringok = 1;
2739 regnode *next = scan + NODE_SZ_STR(scan);
2740 U32 merged = 0;
2741 U32 stopnow = 0;
2742#ifdef DEBUGGING
2743 regnode *stop = scan;
72f13be8 2744 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2745#else
d47053eb
RGS
2746 PERL_UNUSED_ARG(depth);
2747#endif
7918f24d
NC
2748
2749 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2750#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2751 PERL_UNUSED_ARG(flags);
2752 PERL_UNUSED_ARG(val);
07be1b83 2753#endif
07be1b83 2754 DEBUG_PEEP("join",scan,depth);
bb914485 2755
3f410cf6
KW
2756 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2757 * EXACT ones that are mergeable to the current one. */
2758 while (n
2759 && (PL_regkind[OP(n)] == NOTHING
2760 || (stringok && OP(n) == OP(scan)))
07be1b83 2761 && NEXT_OFF(n)
3f410cf6
KW
2762 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2763 {
07be1b83
YO
2764
2765 if (OP(n) == TAIL || n > next)
2766 stringok = 0;
2767 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2768 DEBUG_PEEP("skip:",n,depth);
2769 NEXT_OFF(scan) += NEXT_OFF(n);
2770 next = n + NODE_STEP_REGNODE;
2771#ifdef DEBUGGING
2772 if (stringok)
2773 stop = n;
2774#endif
2775 n = regnext(n);
2776 }
2777 else if (stringok) {
786e8c11 2778 const unsigned int oldl = STR_LEN(scan);
07be1b83 2779 regnode * const nnext = regnext(n);
b2230d39 2780
87b8b349
KW
2781 /* XXX I (khw) kind of doubt that this works on platforms where
2782 * U8_MAX is above 255 because of lots of other assumptions */
79a81a6e 2783 /* Don't join if the sum can't fit into a single node */
b2230d39
KW
2784 if (oldl + STR_LEN(n) > U8_MAX)
2785 break;
07be1b83
YO
2786
2787 DEBUG_PEEP("merg",n,depth);
07be1b83 2788 merged++;
b2230d39 2789
07be1b83
YO
2790 NEXT_OFF(scan) += NEXT_OFF(n);
2791 STR_LEN(scan) += STR_LEN(n);
2792 next = n + NODE_SZ_STR(n);
2793 /* Now we can overwrite *n : */
2794 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2795#ifdef DEBUGGING
2796 stop = next - 1;
2797#endif
2798 n = nnext;
2799 if (stopnow) break;
2800 }
2801
d47053eb
RGS
2802#ifdef EXPERIMENTAL_INPLACESCAN
2803 if (flags && !NEXT_OFF(n)) {
2804 DEBUG_PEEP("atch", val, depth);
2805 if (reg_off_by_arg[OP(n)]) {
2806 ARG_SET(n, val - n);
2807 }
2808 else {
2809 NEXT_OFF(n) = val - n;
2810 }
2811 stopnow = 1;
2812 }
07be1b83
YO
2813#endif
2814 }
2c2b7f86 2815
9d071ca8 2816 *min_subtract = 0;
f758bddf 2817 *has_exactf_sharp_s = FALSE;
f646642f 2818
3f410cf6
KW
2819 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2820 * can now analyze for sequences of problematic code points. (Prior to
2821 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2822 * hence missed). The sequences only happen in folding, hence for any
2823 * non-EXACT EXACTish node */
86d6fcad 2824 if (OP(scan) != EXACT) {
0a982f06
KW
2825 const U8 * const s0 = (U8*) STRING(scan);
2826 const U8 * s = s0;
2827 const U8 * const s_end = s0 + STR_LEN(scan);
f758bddf
KW
2828
2829 /* One pass is made over the node's string looking for all the
2830 * possibilities. to avoid some tests in the loop, there are two main
2831 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2832 * non-UTF-8 */
2833 if (UTF) {
86d6fcad 2834
0a982f06
KW
2835 /* Examine the string for a multi-character fold sequence. UTF-8
2836 * patterns have all characters pre-folded by the time this code is
2837 * executed */
2838 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2839 length sequence we are looking for is 2 */
86d6fcad 2840 {
0a982f06
KW
2841 int count = 0;
2842 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2843 if (! len) { /* Not a multi-char fold: get next char */
2844 s += UTF8SKIP(s);
2845 continue;
2846 }
bb914485 2847
0a982f06
KW
2848 /* Nodes with 'ss' require special handling, except for EXACTFL
2849 * and EXACTFA for which there is no multi-char fold to this */
2850 if (len == 2 && *s == 's' && *(s+1) == 's'
2851 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2852 {
2853 count = 2;
2854 OP(scan) = EXACTFU_SS;
2855 s += 2;
2856 }
accd961f
KW
2857 else if (len == 6 /* len is the same in both ASCII and EBCDIC
2858 for these */
0a982f06
KW
2859 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2860 COMBINING_DIAERESIS_UTF8
2861 COMBINING_ACUTE_ACCENT_UTF8,
2862 6)
2863 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2864 COMBINING_DIAERESIS_UTF8
2865 COMBINING_ACUTE_ACCENT_UTF8,
2866 6)))
2867 {
2868 count = 3;
2869
2870 /* These two folds require special handling by trie's, so
2871 * change the node type to indicate this. If EXACTFA and
2872 * EXACTFL were ever to be handled by trie's, this would
2873 * have to be changed. If this node has already been
2874 * changed to EXACTFU_SS in this loop, leave it as is. (I
2875 * (khw) think it doesn't matter in regexec.c for UTF
2876 * patterns, but no need to change it */
2877 if (OP(scan) == EXACTFU) {
2878 OP(scan) = EXACTFU_TRICKYFOLD;
2879 }
2880 s += 6;
2881 }
2882 else { /* Here is a generic multi-char fold. */
2883 const U8* multi_end = s + len;
2884
2885 /* Count how many characters in it. In the case of /l and
2886 * /aa, no folds which contain ASCII code points are
2887 * allowed, so check for those, and skip if found. (In
2888 * EXACTFL, no folds are allowed to any Latin1 code point,
2889 * not just ASCII. But there aren't any of these
2890 * currently, nor ever likely, so don't take the time to
2891 * test for them. The code that generates the
2892 * is_MULTI_foo() macros croaks should one actually get put
2893 * into Unicode .) */
2894 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2895 count = utf8_length(s, multi_end);
2896 s = multi_end;
2897 }
2898 else {
2899 while (s < multi_end) {
2900 if (isASCII(*s)) {
2901 s++;
2902 goto next_iteration;
2903 }
2904 else {
2905 s += UTF8SKIP(s);
2906 }
2907 count++;
2908 }
2909 }
2910 }
f758bddf 2911
0a982f06
KW
2912 /* The delta is how long the sequence is minus 1 (1 is how long
2913 * the character that folds to the sequence is) */
2914 *min_subtract += count - 1;
2915 next_iteration: ;
bb914485
KW
2916 }
2917 }
1ca267a5
KW
2918 else if (OP(scan) == EXACTFA) {
2919
2920 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
2921 * fold to the ASCII range (and there are no existing ones in the
2922 * upper latin1 range). But, as outlined in the comments preceding
2923 * this function, we need to flag any occurrences of the sharp s */
2924 while (s < s_end) {
2925 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
2926 *has_exactf_sharp_s = TRUE;
2927 break;
2928 }
2929 s++;
2930 continue;
2931 }
2932 }
2933 else if (OP(scan) != EXACTFL) {
2934
2935 /* Non-UTF-8 pattern, not EXACTFA nor EXACTFL node. Look for the
2936 * multi-char folds that are all Latin1. (This code knows that
2937 * there are no current multi-char folds possible with EXACTFL,
2938 * relying on fold_grind.t to catch any errors if the very unlikely
2939 * event happens that some get added in future Unicode versions.)
2940 * As explained in the comments preceding this function, we look
2941 * also for the sharp s in EXACTF nodes; it can be in the final
0a982f06
KW
2942 * position. Otherwise we can stop looking 1 byte earlier because
2943 * have to find at least two characters for a multi-fold */
f758bddf
KW
2944 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2945
0a982f06 2946 while (s < upper) {
40b1ba4f 2947 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
0a982f06
KW
2948 if (! len) { /* Not a multi-char fold. */
2949 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2950 {
2951 *has_exactf_sharp_s = TRUE;
2952 }
2953 s++;
2954 continue;
2955 }
2956
2957 if (len == 2
c02c3054
KW
2958 && isARG2_lower_or_UPPER_ARG1('s', *s)
2959 && isARG2_lower_or_UPPER_ARG1('s', *(s+1)))
0a982f06
KW
2960 {
2961
2962 /* EXACTF nodes need to know that the minimum length
2963 * changed so that a sharp s in the string can match this
2964 * ss in the pattern, but they remain EXACTF nodes, as they
2965 * won't match this unless the target string is is UTF-8,
2966 * which we don't know until runtime */
2967 if (OP(scan) != EXACTF) {
2968 OP(scan) = EXACTFU_SS;
2969 }
86d6fcad 2970 }
0a982f06
KW
2971
2972 *min_subtract += len - 1;
2973 s += len;
86d6fcad
KW
2974 }
2975 }
07be1b83 2976 }
3f410cf6 2977
07be1b83 2978#ifdef DEBUGGING
bb789b09
DM
2979 /* Allow dumping but overwriting the collection of skipped
2980 * ops and/or strings with fake optimized ops */
07be1b83
YO
2981 n = scan + NODE_SZ_STR(scan);
2982 while (n <= stop) {
bb789b09
DM
2983 OP(n) = OPTIMIZED;
2984 FLAGS(n) = 0;
2985 NEXT_OFF(n) = 0;
07be1b83
YO
2986 n++;
2987 }
2988#endif
2989 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2990 return stopnow;
2991}
2992
486ec47a 2993/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
2994 Finds fixed substrings. */
2995
a0288114 2996/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
2997 to the position after last scanned or to NULL. */
2998
40d049e4
YO
2999#define INIT_AND_WITHP \
3000 assert(!and_withp); \
3001 Newx(and_withp,1,struct regnode_charclass_class); \
3002 SAVEFREEPV(and_withp)
07be1b83 3003
b515a41d 3004/* this is a chain of data about sub patterns we are processing that
486ec47a 3005 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
3006 we can simulate recursion without losing state. */
3007struct scan_frame;
3008typedef struct scan_frame {
3009 regnode *last; /* last node to process in this frame */
3010 regnode *next; /* next node to process when last is reached */
3011 struct scan_frame *prev; /*previous frame*/
3012 I32 stop; /* what stopparen do we use */
3013} scan_frame;
3014
304ee84b
YO
3015
3016#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
3017
49f55535 3018STATIC SSize_t
40d049e4 3019S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
ea3daa5d 3020 SSize_t *minlenp, SSize_t *deltap,
40d049e4
YO
3021 regnode *last,
3022 scan_data_t *data,
3023 I32 stopparen,
3024 U8* recursed,
3025 struct regnode_charclass_class *and_withp,
3026 U32 flags, U32 depth)
c277df42
IZ
3027 /* scanp: Start here (read-write). */
3028 /* deltap: Write maxlen-minlen here. */
3029 /* last: Stop before this one. */
40d049e4
YO
3030 /* data: string data about the pattern */
3031 /* stopparen: treat close N as END */
3032 /* recursed: which subroutines have we recursed into */
3033 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 3034{
97aff369 3035 dVAR;
ea3daa5d
FC
3036 /* There must be at least this number of characters to match */
3037 SSize_t min = 0;
2d608413 3038 I32 pars = 0, code;
c277df42 3039 regnode *scan = *scanp, *next;
ea3daa5d 3040 SSize_t delta = 0;
c277df42 3041 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 3042 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
3043 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3044 scan_data_t data_fake;
a3621e74 3045 SV *re_trie_maxbuff = NULL;
786e8c11 3046 regnode *first_non_open = scan;
ea3daa5d 3047 SSize_t stopmin = SSize_t_MAX;
8aa23a47 3048 scan_frame *frame = NULL;
a3621e74 3049 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 3050
7918f24d
NC
3051 PERL_ARGS_ASSERT_STUDY_CHUNK;
3052
13a24bad 3053#ifdef DEBUGGING
40d049e4 3054 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 3055#endif
40d049e4 3056
786e8c11 3057 if ( depth == 0 ) {
40d049e4 3058 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
3059 first_non_open=regnext(first_non_open);
3060 }
3061
b81d288d 3062
8aa23a47
YO
3063 fake_study_recurse:
3064 while ( scan && OP(scan) != END && scan < last ){
2d608413
KW
3065 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3066 node length to get a real minimum (because
3067 the folded version may be shorter) */
f758bddf 3068 bool has_exactf_sharp_s = FALSE;
8aa23a47 3069 /* Peephole optimizer: */
304ee84b 3070 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 3071 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
3072
3073 /* Its not clear to khw or hv why this is done here, and not in the
3074 * clauses that deal with EXACT nodes. khw's guess is that it's
3075 * because of a previous design */
9d071ca8 3076 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
3077
3078 /* Follow the next-chain of the current node and optimize
3079 away all the NOTHINGs from it. */
3080 if (OP(scan) != CURLYX) {
3081 const int max = (reg_off_by_arg[OP(scan)]
3082 ? I32_MAX
3083 /* I32 may be smaller than U16 on CRAYs! */
3084 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3085 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3086 int noff;
3087 regnode *n = scan;
686b73d4 3088
8aa23a47
YO
3089 /* Skip NOTHING and LONGJMP. */
3090 while ((n = regnext(n))
3091 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3092 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3093 && off + noff < max)
3094 off += noff;
3095 if (reg_off_by_arg[OP(scan)])
3096 ARG(scan) = off;
3097 else
3098 NEXT_OFF(scan) = off;
3099 }
a3621e74 3100
c277df42 3101
8aa23a47
YO
3102
3103 /* The principal pseudo-switch. Cannot be a switch, since we
3104 look into several different things. */
3105 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3106 || OP(scan) == IFTHEN) {
3107 next = regnext(scan);
3108 code = OP(scan);
3109 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3110
8aa23a47
YO
3111 if (OP(next) == code || code == IFTHEN) {
3112 /* NOTE - There is similar code to this block below for handling
3113 TRIE nodes on a re-study. If you change stuff here check there
3114 too. */
49f55535 3115 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
8aa23a47
YO
3116 struct regnode_charclass_class accum;
3117 regnode * const startbranch=scan;
686b73d4 3118
8aa23a47 3119 if (flags & SCF_DO_SUBSTR)
304ee84b 3120 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3121 if (flags & SCF_DO_STCLASS)
e755fd73 3122 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3123
3124 while (OP(scan) == code) {
49f55535
FC
3125 SSize_t deltanext, minnext, fake;
3126 I32 f = 0;
8aa23a47
YO
3127 struct regnode_charclass_class this_class;
3128
3129 num++;
3130 data_fake.flags = 0;
3131 if (data) {
3132 data_fake.whilem_c = data->whilem_c;
3133 data_fake.last_closep = data->last_closep;
3134 }
3135 else
3136 data_fake.last_closep = &fake;
58e23c8d
YO
3137
3138 data_fake.pos_delta = delta;
8aa23a47
YO
3139 next = regnext(scan);
3140 scan = NEXTOPER(scan);
3141 if (code != BRANCH)
c277df42 3142 scan = NEXTOPER(scan);
8aa23a47 3143 if (flags & SCF_DO_STCLASS) {
e755fd73 3144 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3145 data_fake.start_class = &this_class;
3146 f = SCF_DO_STCLASS_AND;
58e23c8d 3147 }
8aa23a47
YO
3148 if (flags & SCF_WHILEM_VISITED_POS)
3149 f |= SCF_WHILEM_VISITED_POS;
3150
3151 /* we suppose the run is continuous, last=next...*/
3152 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3153 next, &data_fake,
3154 stopparen, recursed, NULL, f,depth+1);
3155 if (min1 > minnext)
3156 min1 = minnext;
ea3daa5d 3157 if (deltanext == SSize_t_MAX) {
8aa23a47 3158 is_inf = is_inf_internal = 1;
ea3daa5d 3159 max1 = SSize_t_MAX;
9b139d09
GG
3160 } else if (max1 < minnext + deltanext)
3161 max1 = minnext + deltanext;
8aa23a47
YO
3162 scan = next;
3163 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3164 pars++;
3165 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3166 if ( stopmin > minnext)
3167 stopmin = min + min1;
3168 flags &= ~SCF_DO_SUBSTR;
3169 if (data)
3170 data->flags |= SCF_SEEN_ACCEPT;
3171 }
3172 if (data) {
3173 if (data_fake.flags & SF_HAS_EVAL)
3174 data->flags |= SF_HAS_EVAL;
3175 data->whilem_c = data_fake.whilem_c;
3dab1dad 3176 }
8aa23a47 3177 if (flags & SCF_DO_STCLASS)
3fffb88a 3178 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3179 }
3180 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3181 min1 = 0;
3182 if (flags & SCF_DO_SUBSTR) {
3183 data->pos_min += min1;
ea3daa5d
FC
3184 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3185 data->pos_delta = SSize_t_MAX;
9b139d09
GG
3186 else
3187 data->pos_delta += max1 - min1;
8aa23a47
YO
3188 if (max1 != min1 || is_inf)
3189 data->longest = &(data->longest_float);
3190 }
3191 min += min1;
ea3daa5d
FC
3192 if (delta == SSize_t_MAX
3193 || SSize_t_MAX - delta - (max1 - min1) < 0)
3194 delta = SSize_t_MAX;
9b139d09
GG
3195 else
3196 delta += max1 - min1;
8aa23a47 3197 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3198 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3199 if (min1) {
3200 cl_and(data->start_class, and_withp);
3201 flags &= ~SCF_DO_STCLASS;
653099ff 3202 }
8aa23a47
YO
3203 }
3204 else if (flags & SCF_DO_STCLASS_AND) {
3205 if (min1) {
3206 cl_and(data->start_class, &accum);
3207 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3208 }
8aa23a47
YO
3209 else {
3210 /* Switch to OR mode: cache the old value of
3211 * data->start_class */
3212 INIT_AND_WITHP;
3213 StructCopy(data->start_class, and_withp,
3214 struct regnode_charclass_class);
3215 flags &= ~SCF_DO_STCLASS_AND;
3216 StructCopy(&accum, data->start_class,
3217 struct regnode_charclass_class);
3218 flags |= SCF_DO_STCLASS_OR;
899d20b9 3219 SET_SSC_EOS(data->start_class);
de0c8cb8 3220 }
8aa23a47 3221 }
a3621e74 3222
8aa23a47
YO
3223 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3224 /* demq.
a3621e74 3225
8aa23a47
YO
3226 Assuming this was/is a branch we are dealing with: 'scan' now
3227 points at the item that follows the branch sequence, whatever
3228 it is. We now start at the beginning of the sequence and look
3229 for subsequences of
a3621e74 3230
8aa23a47
YO
3231 BRANCH->EXACT=>x1
3232 BRANCH->EXACT=>x2
3233 tail
a3621e74 3234
8aa23a47 3235 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3236
486ec47a 3237 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3238 element into a trie and then add the subsequent branch exact
3239 strings to the trie.
a3621e74 3240
8aa23a47 3241 We have two cases
a3621e74 3242
3b753521 3243 1. patterns where the whole set of branches can be converted.
a3621e74 3244
8aa23a47 3245 2. patterns where only a subset can be converted.
a3621e74 3246
8aa23a47
YO
3247 In case 1 we can replace the whole set with a single regop
3248 for the trie. In case 2 we need to keep the start and end
3b753521 3249 branches so
a3621e74 3250
8aa23a47
YO
3251 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3252 becomes BRANCH TRIE; BRANCH X;
786e8c11 3253
8aa23a47
YO
3254 There is an additional case, that being where there is a
3255 common prefix, which gets split out into an EXACT like node
3256 preceding the TRIE node.
a3621e74 3257
8aa23a47
YO
3258 If x(1..n)==tail then we can do a simple trie, if not we make
3259 a "jump" trie, such that when we match the appropriate word
486ec47a 3260 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3261 a nested if into a case structure of sorts.
b515a41d 3262
8aa23a47 3263 */
686b73d4 3264
8aa23a47
YO
3265 int made=0;
3266 if (!re_trie_maxbuff) {
3267 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3268 if (!SvIOK(re_trie_maxbuff))
3269 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3270 }
3271 if ( SvIV(re_trie_maxbuff)>=0 ) {
3272 regnode *cur;
3273 regnode *first = (regnode *)NULL;
3274 regnode *last = (regnode *)NULL;
3275 regnode *tail = scan;
fab2782b 3276 U8 trietype = 0;
8aa23a47 3277 U32 count=0;
a3621e74
YO
3278
3279#ifdef DEBUGGING
8aa23a47 3280 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3281#endif
8aa23a47
YO
3282 /* var tail is used because there may be a TAIL
3283 regop in the way. Ie, the exacts will point to the
3284 thing following the TAIL, but the last branch will
3285 point at the TAIL. So we advance tail. If we
3286 have nested (?:) we may have to move through several
3287 tails.
3288 */
3289
3290 while ( OP( tail ) == TAIL ) {
3291 /* this is the TAIL generated by (?:) */
3292 tail = regnext( tail );
3293 }
a3621e74 3294
8aa23a47 3295
df826430 3296 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3297 regprop(RExC_rx, mysv, tail );
3298 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3299 (int)depth * 2 + 2, "",
3300 "Looking for TRIE'able sequences. Tail node is: ",
3301 SvPV_nolen_const( mysv )
3302 );
3303 });
3304
3305 /*
3306
fab2782b
YO
3307 Step through the branches
3308 cur represents each branch,
3309 noper is the first thing to be matched as part of that branch
3310 noper_next is the regnext() of that node.
3311
3312 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3313 via a "jump trie" but we also support building with NOJUMPTRIE,
3314 which restricts the trie logic to structures like /FOO|BAR/.
3315
3316 If noper is a trieable nodetype then the branch is a possible optimization
3317 target. If we are building under NOJUMPTRIE then we require that noper_next
3318 is the same as scan (our current position in the regex program).
3319
3320 Once we have two or more consecutive such branches we can create a
3321 trie of the EXACT's contents and stitch it in place into the program.
3322
3323 If the sequence represents all of the branches in the alternation we
3324 replace the entire thing with a single TRIE node.
3325
3326 Otherwise when it is a subsequence we need to stitch it in place and
3327 replace only the relevant branches. This means the first branch has
3328 to remain as it is used by the alternation logic, and its next pointer,
3329 and needs to be repointed at the item on the branch chain following
3330 the last branch we have optimized away.
3331
3332 This could be either a BRANCH, in which case the subsequence is internal,
3333 or it could be the item following the branch sequence in which case the
3334 subsequence is at the end (which does not necessarily mean the first node
3335 is the start of the alternation).
3336
3337 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3338
3339 optype | trietype
3340 ----------------+-----------
3341 NOTHING | NOTHING
3342 EXACT | EXACT
3343 EXACTFU | EXACTFU
3344 EXACTFU_SS | EXACTFU
3345 EXACTFU_TRICKYFOLD | EXACTFU
3346 EXACTFA | 0
3347
8aa23a47
YO
3348
3349 */
fab2782b
YO
3350#define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3351 ( EXACT == (X) ) ? EXACT : \
3352 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3353 0 )
8aa23a47
YO
3354
3355 /* dont use tail as the end marker for this traverse */
3356 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3357 regnode * const noper = NEXTOPER( cur );
fab2782b
YO
3358 U8 noper_type = OP( noper );
3359 U8 noper_trietype = TRIE_TYPE( noper_type );
b515a41d 3360#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3361 regnode * const noper_next = regnext( noper );
df826430
YO
3362 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3363 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
b515a41d
YO
3364#endif
3365
df826430 3366 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3367 regprop(RExC_rx, mysv, cur);
3368 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3369 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3370
3371 regprop(RExC_rx, mysv, noper);
3372 PerlIO_printf( Perl_debug_log, " -> %s",
3373 SvPV_nolen_const(mysv));
3374
3375 if ( noper_next ) {
3376 regprop(RExC_rx, mysv, noper_next );
3377 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3378 SvPV_nolen_const(mysv));
3379 }
df826430
YO
3380 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3381 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3382 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3383 );
8aa23a47 3384 });
fab2782b
YO
3385
3386 /* Is noper a trieable nodetype that can be merged with the
3387 * current trie (if there is one)? */
3388 if ( noper_trietype
3389 &&
3390 (
df826430
YO
3391 ( noper_trietype == NOTHING)
3392 || ( trietype == NOTHING )
a40630bf 3393 || ( trietype == noper_trietype )
fab2782b 3394 )
786e8c11 3395#ifdef NOJUMPTRIE
8aa23a47 3396 && noper_next == tail
786e8c11 3397#endif
8aa23a47
YO
3398 && count < U16_MAX)
3399 {
fab2782b
YO
3400 /* Handle mergable triable node
3401 * Either we are the first node in a new trieable sequence,
3402 * in which case we do some bookkeeping, otherwise we update
3403 * the end pointer. */
fab2782b 3404 if ( !first ) {
3b6759a6 3405 first = cur;
df826430
YO
3406 if ( noper_trietype == NOTHING ) {
3407#if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3408 regnode * const noper_next = regnext( noper );
3b6759a6 3409 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
df826430
YO
3410 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3411#endif
3412
190c1910 3413 if ( noper_next_trietype ) {
df826430 3414 trietype = noper_next_trietype;
190c1910
YO
3415 } else if (noper_next_type) {
3416 /* a NOTHING regop is 1 regop wide. We need at least two
3417 * for a trie so we can't merge this in */
3418 first = NULL;
3419 }
3420 } else {
3421 trietype = noper_trietype;
3b6759a6 3422 }
8aa23a47 3423 } else {
fab2782b
YO
3424 if ( trietype == NOTHING )
3425 trietype = noper_trietype;
8aa23a47
YO
3426 last = cur;
3427 }
df826430
YO
3428 if (first)
3429 count++;
fab2782b
YO
3430 } /* end handle mergable triable node */
3431 else {
3432 /* handle unmergable node -
3433 * noper may either be a triable node which can not be tried
3434 * together with the current trie, or a non triable node */
729aaeb5
YO
3435 if ( last ) {
3436 /* If last is set and trietype is not NOTHING then we have found
3437 * at least two triable branch sequences in a row of a similar
3438 * trietype so we can turn them into a trie. If/when we
3439 * allow NOTHING to start a trie sequence this condition will be
3440 * required, and it isn't expensive so we leave it in for now. */
e6351b37 3441 if ( trietype && trietype != NOTHING )
729aaeb5
YO
3442 make_trie( pRExC_state,
3443 startbranch, first, cur, tail, count,
3444 trietype, depth+1 );
fab2782b 3445 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
8aa23a47 3446 }
fab2782b 3447 if ( noper_trietype
786e8c11 3448#ifdef NOJUMPTRIE
8aa23a47 3449 && noper_next == tail
786e8c11 3450#endif
8aa23a47 3451 ){
fab2782b 3452 /* noper is triable, so we can start a new trie sequence */
8aa23a47
YO
3453 count = 1;
3454 first = cur;
fab2782b
YO
3455 trietype = noper_trietype;
3456 } else if (first) {
3457 /* if we already saw a first but the current node is not triable then we have
3458 * to reset the first information. */
8aa23a47
YO
3459 count = 0;
3460 first = NULL;
fab2782b 3461 trietype = 0;
8aa23a47 3462 }
fab2782b
YO
3463 } /* end handle unmergable node */
3464 } /* loop over branches */
df826430 3465 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3466 regprop(RExC_rx, mysv, cur);
3467 PerlIO_printf( Perl_debug_log,
3468 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3469 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3470
3471 });
e6351b37 3472 if ( last && trietype ) {
3b6759a6
YO
3473 if ( trietype != NOTHING ) {
3474 /* the last branch of the sequence was part of a trie,
3475 * so we have to construct it here outside of the loop
3476 */
3477 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
686b73d4 3478#ifdef TRIE_STUDY_OPT
3b6759a6
YO
3479 if ( ((made == MADE_EXACT_TRIE &&
3480 startbranch == first)
3481 || ( first_non_open == first )) &&
3482 depth==0 ) {
3483 flags |= SCF_TRIE_RESTUDY;
3484 if ( startbranch == first
3485 && scan == tail )
3486 {
3487 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3488 }
8aa23a47 3489 }
3dab1dad 3490#endif
3b6759a6
YO
3491 } else {
3492 /* at this point we know whatever we have is a NOTHING sequence/branch
3493 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3494 */
3495 if ( startbranch == first ) {
3496 regnode *opt;
3497 /* the entire thing is a NOTHING sequence, something like this:
3498 * (?:|) So we can turn it into a plain NOTHING op. */
3499 DEBUG_TRIE_COMPILE_r({
3500 regprop(RExC_rx, mysv, cur);
3501 PerlIO_printf( Perl_debug_log,
3502 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3503 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3504
3505 });
3506 OP(startbranch)= NOTHING;
3507 NEXT_OFF(startbranch)= tail - startbranch;
3508 for ( opt= startbranch + 1; opt < tail ; opt++ )
3509 OP(opt)= OPTIMIZED;
3510 }
3511 }
fab2782b
YO
3512 } /* end if ( last) */
3513 } /* TRIE_MAXBUF is non zero */
8aa23a47
YO
3514
3515 } /* do trie */
3516
653099ff 3517 }
8aa23a47
YO
3518 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3519 scan = NEXTOPER(NEXTOPER(scan));
3520 } else /* single branch is optimized. */
3521 scan = NEXTOPER(scan);
3522 continue;
3523 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3524 scan_frame *newframe = NULL;
3525 I32 paren;
3526 regnode *start;
3527 regnode *end;
3528
3529 if (OP(scan) != SUSPEND) {
3530 /* set the pointer */
3531 if (OP(scan) == GOSUB) {
3532 paren = ARG(scan);
3533 RExC_recurse[ARG2L(scan)] = scan;
3534 start = RExC_open_parens[paren-1];
3535 end = RExC_close_parens[paren-1];
3536 } else {
3537 paren = 0;
f8fc2ecf 3538 start = RExC_rxi->program + 1;
8aa23a47
YO
3539 end = RExC_opend;
3540 }
3541 if (!recursed) {
3542 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3543 SAVEFREEPV(recursed);
3544 }
3545 if (!PAREN_TEST(recursed,paren+1)) {
3546 PAREN_SET(recursed,paren+1);
3547 Newx(newframe,1,scan_frame);
3548 } else {
3549 if (flags & SCF_DO_SUBSTR) {
304ee84b 3550 SCAN_COMMIT(pRExC_state,data,minlenp);
8aa23a47
YO
3551 data->longest = &(data->longest_float);
3552 }
3553 is_inf = is_inf_internal = 1;