This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
Allow trie use for /iaa matching
[perl5.git] / regcomp.c
CommitLineData
a0d0e21e
LW
1/* regcomp.c
2 */
3
4/*
4ac71550
TC
5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
6 *
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
a0d0e21e
LW
8 */
9
61296642
DM
10/* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
166f8a29 12 * a regular expression.
e4a054ea
DM
13 *
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
166f8a29
DM
18 */
19
a687059c
LW
20/* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
22 */
23
24/* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
27 */
28
e50aee73 29/* The names of the functions have been changed from regcomp and
3b753521 30 * regexec to pregcomp and pregexec in order to avoid conflicts
e50aee73
AD
31 * with the POSIX routines of the same names.
32*/
33
b9d5759e 34#ifdef PERL_EXT_RE_BUILD
54df2634 35#include "re_top.h"
b81d288d 36#endif
56953603 37
a687059c 38/*
e50aee73 39 * pregcomp and pregexec -- regsub and regerror are not used in perl
a687059c
LW
40 *
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
43 *
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
47 *
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
50 * from defects in it.
51 *
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
54 *
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
57 *
58 *
59 **** Alterations to Henry's code are...
60 ****
4bb101f2 61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
1129b882
NC
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
a687059c 64 ****
9ef589d8
LW
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
67
a687059c
LW
68 *
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
72 */
73#include "EXTERN.h"
864dbfa3 74#define PERL_IN_REGCOMP_C
a687059c 75#include "perl.h"
d06ea78c 76
acfe0abc 77#ifndef PERL_IN_XSUB_RE
d06ea78c
GS
78# include "INTERN.h"
79#endif
c277df42
IZ
80
81#define REG_COMP_C
54df2634
NC
82#ifdef PERL_IN_XSUB_RE
83# include "re_comp.h"
40b6423c 84extern const struct regexp_engine my_reg_engine;
54df2634
NC
85#else
86# include "regcomp.h"
87#endif
a687059c 88
04e98a4d 89#include "dquote_static.c"
26faadbd 90#include "charclass_invlists.h"
81e983c1 91#include "inline_invlist.c"
1b0f46bf 92#include "unicode_constants.h"
04e98a4d 93
94dc5c2d 94#define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
26faadbd 95#define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
2c61f163 96#define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
94dc5c2d 97
d4cce5f1 98#ifdef op
11343788 99#undef op
d4cce5f1 100#endif /* op */
11343788 101
fe14fcc3 102#ifdef MSDOS
7e4e8c89 103# if defined(BUGGY_MSC6)
fe14fcc3 104 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
7e4e8c89 105# pragma optimize("a",off)
fe14fcc3 106 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
7e4e8c89
NC
107# pragma optimize("w",on )
108# endif /* BUGGY_MSC6 */
fe14fcc3
LW
109#endif /* MSDOS */
110
a687059c
LW
111#ifndef STATIC
112#define STATIC static
113#endif
114
b1603ef8 115
830247a4 116typedef struct RExC_state_t {
514a91f1
DM
117 U32 flags; /* RXf_* are we folding, multilining? */
118 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 119 char *precomp; /* uncompiled string. */
288b8c02 120 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf
YO
121 regexp *rx; /* perl core regexp structure */
122 regexp_internal *rxi; /* internal data for regexp object pprivate field */
fac92740 123 char *start; /* Start of input for compile */
830247a4
IZ
124 char *end; /* End of input for compile */
125 char *parse; /* Input-scan pointer. */
ea3daa5d 126 SSize_t whilem_seen; /* number of WHILEM in this expr */
fac92740 127 regnode *emit_start; /* Start of emitted-code area */
3b57cd43 128 regnode *emit_bound; /* First regnode outside of the allocated space */
f7c7e32a
DM
129 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
130 implies compiling, so don't emit */
131 regnode emit_dummy; /* placeholder for emit to point to */
830247a4
IZ
132 I32 naughty; /* How bad is this pattern? */
133 I32 sawback; /* Did we see \1, ...? */
134 U32 seen;
ea3daa5d 135 SSize_t size; /* Code size. */
c74340f9
YO
136 I32 npar; /* Capture buffer count, (OPEN). */
137 I32 cpar; /* Capture buffer count, (CLOSE). */
e2e6a0f1 138 I32 nestroot; /* root parens we are in - used by accept */
830247a4
IZ
139 I32 extralen;
140 I32 seen_zerolen;
40d049e4
YO
141 regnode **open_parens; /* pointers to open parens */
142 regnode **close_parens; /* pointers to close parens */
143 regnode *opend; /* END node in program */
02daf0ab
YO
144 I32 utf8; /* whether the pattern is utf8 or not */
145 I32 orig_utf8; /* whether the pattern was originally in utf8 */
146 /* XXX use this for future optimisation of case
147 * where pattern must be upgraded to utf8. */
e40e74fe
KW
148 I32 uni_semantics; /* If a d charset modifier should use unicode
149 rules, even if the pattern is not in
150 utf8 */
81714fb9 151 HV *paren_names; /* Paren names */
1f1031fe 152
40d049e4
YO
153 regnode **recurse; /* Recurse regops */
154 I32 recurse_count; /* Number of recurse regops */
b57e4118 155 I32 in_lookbehind;
4624b182 156 I32 contains_locale;
bb3f3ed2 157 I32 override_recoding;
9d53c457 158 I32 in_multi_char_class;
3d2bd50a 159 struct reg_code_block *code_blocks; /* positions of literal (?{})
68e2671b 160 within pattern */
b1603ef8
DM
161 int num_code_blocks; /* size of code_blocks[] */
162 int code_index; /* next code_blocks[] slot */
830247a4
IZ
163#if ADD_TO_REGEXEC
164 char *starttry; /* -Dr: where regtry was called. */
165#define RExC_starttry (pRExC_state->starttry)
166#endif
d24ca0c5 167 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 168#ifdef DEBUGGING
be8e71aa 169 const char *lastparse;
3dab1dad 170 I32 lastnum;
1f1031fe 171 AV *paren_name_list; /* idx -> name */
3dab1dad
YO
172#define RExC_lastparse (pRExC_state->lastparse)
173#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 174#define RExC_paren_name_list (pRExC_state->paren_name_list)
3dab1dad 175#endif
830247a4
IZ
176} RExC_state_t;
177
e2509266 178#define RExC_flags (pRExC_state->flags)
514a91f1 179#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 180#define RExC_precomp (pRExC_state->precomp)
288b8c02 181#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 182#define RExC_rx (pRExC_state->rx)
f8fc2ecf 183#define RExC_rxi (pRExC_state->rxi)
fac92740 184#define RExC_start (pRExC_state->start)
830247a4
IZ
185#define RExC_end (pRExC_state->end)
186#define RExC_parse (pRExC_state->parse)
187#define RExC_whilem_seen (pRExC_state->whilem_seen)
7122b237
YO
188#ifdef RE_TRACK_PATTERN_OFFSETS
189#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
190#endif
830247a4 191#define RExC_emit (pRExC_state->emit)
f7c7e32a 192#define RExC_emit_dummy (pRExC_state->emit_dummy)
fac92740 193#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 194#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
195#define RExC_naughty (pRExC_state->naughty)
196#define RExC_sawback (pRExC_state->sawback)
197#define RExC_seen (pRExC_state->seen)
198#define RExC_size (pRExC_state->size)
199#define RExC_npar (pRExC_state->npar)
e2e6a0f1 200#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
201#define RExC_extralen (pRExC_state->extralen)
202#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 203#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 204#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 205#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
206#define RExC_open_parens (pRExC_state->open_parens)
207#define RExC_close_parens (pRExC_state->close_parens)
208#define RExC_opend (pRExC_state->opend)
81714fb9 209#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
210#define RExC_recurse (pRExC_state->recurse)
211#define RExC_recurse_count (pRExC_state->recurse_count)
b57e4118 212#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 213#define RExC_contains_locale (pRExC_state->contains_locale)
9d53c457
KW
214#define RExC_override_recoding (pRExC_state->override_recoding)
215#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
830247a4 216
cde0cee5 217
a687059c
LW
218#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
219#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
4d68ffa0 220 ((*s) == '{' && regcurly(s, FALSE)))
a687059c 221
35c8bce7
LW
222#ifdef SPSTART
223#undef SPSTART /* dratted cpp namespace... */
224#endif
a687059c
LW
225/*
226 * Flags to be passed up and down.
227 */
a687059c 228#define WORST 0 /* Worst case. */
a3b492c3 229#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 230
e64f369d 231/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
232 * character. (There needs to be a case: in the switch statement in regexec.c
233 * for any node marked SIMPLE.) Note that this is not the same thing as
234 * REGNODE_SIMPLE */
fda99bee 235#define SIMPLE 0x02
e64f369d 236#define SPSTART 0x04 /* Starts with * or + */
8d9c2815
NC
237#define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
238#define TRYAGAIN 0x10 /* Weeded out a declaration. */
239#define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
a687059c 240
3dab1dad
YO
241#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
242
07be1b83
YO
243/* whether trie related optimizations are enabled */
244#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
245#define TRIE_STUDY_OPT
786e8c11 246#define FULL_TRIE_STUDY
07be1b83
YO
247#define TRIE_STCLASS
248#endif
1de06328
YO
249
250
40d049e4
YO
251
252#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
253#define PBITVAL(paren) (1 << ((paren) & 7))
254#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
255#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
256#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
257
bbd61b5f 258#define REQUIRE_UTF8 STMT_START { \
8d9c2815
NC
259 if (!UTF) { \
260 *flagp = RESTART_UTF8; \
261 return NULL; \
262 } \
bbd61b5f 263 } STMT_END
40d049e4 264
f19b1a63
KW
265/* This converts the named class defined in regcomp.h to its equivalent class
266 * number defined in handy.h. */
267#define namedclass_to_classnum(class) ((int) ((class) / 2))
268#define classnum_to_namedclass(classnum) ((classnum) * 2)
269
1de06328
YO
270/* About scan_data_t.
271
272 During optimisation we recurse through the regexp program performing
273 various inplace (keyhole style) optimisations. In addition study_chunk
274 and scan_commit populate this data structure with information about
275 what strings MUST appear in the pattern. We look for the longest
3b753521 276 string that must appear at a fixed location, and we look for the
1de06328
YO
277 longest string that may appear at a floating location. So for instance
278 in the pattern:
279
280 /FOO[xX]A.*B[xX]BAR/
281
282 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
283 strings (because they follow a .* construct). study_chunk will identify
284 both FOO and BAR as being the longest fixed and floating strings respectively.
285
286 The strings can be composites, for instance
287
288 /(f)(o)(o)/
289
290 will result in a composite fixed substring 'foo'.
291
292 For each string some basic information is maintained:
293
294 - offset or min_offset
295 This is the position the string must appear at, or not before.
296 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
297 characters must match before the string we are searching for.
298 Likewise when combined with minlenp and the length of the string it
1de06328
YO
299 tells us how many characters must appear after the string we have
300 found.
301
302 - max_offset
303 Only used for floating strings. This is the rightmost point that
ea3daa5d 304 the string can appear at. If set to SSize_t_MAX it indicates that the
1de06328
YO
305 string can occur infinitely far to the right.
306
307 - minlenp
2d608413
KW
308 A pointer to the minimum number of characters of the pattern that the
309 string was found inside. This is important as in the case of positive
1de06328
YO
310 lookahead or positive lookbehind we can have multiple patterns
311 involved. Consider
312
313 /(?=FOO).*F/
314
315 The minimum length of the pattern overall is 3, the minimum length
316 of the lookahead part is 3, but the minimum length of the part that
317 will actually match is 1. So 'FOO's minimum length is 3, but the
318 minimum length for the F is 1. This is important as the minimum length
319 is used to determine offsets in front of and behind the string being
320 looked for. Since strings can be composites this is the length of the
486ec47a 321 pattern at the time it was committed with a scan_commit. Note that
1de06328
YO
322 the length is calculated by study_chunk, so that the minimum lengths
323 are not known until the full pattern has been compiled, thus the
324 pointer to the value.
325
326 - lookbehind
327
328 In the case of lookbehind the string being searched for can be
329 offset past the start point of the final matching string.
330 If this value was just blithely removed from the min_offset it would
331 invalidate some of the calculations for how many chars must match
332 before or after (as they are derived from min_offset and minlen and
333 the length of the string being searched for).
334 When the final pattern is compiled and the data is moved from the
335 scan_data_t structure into the regexp structure the information
336 about lookbehind is factored in, with the information that would
337 have been lost precalculated in the end_shift field for the
338 associated string.
339
340 The fields pos_min and pos_delta are used to store the minimum offset
341 and the delta to the maximum offset at the current point in the pattern.
342
343*/
2c2d71f5
JH
344
345typedef struct scan_data_t {
1de06328
YO
346 /*I32 len_min; unused */
347 /*I32 len_delta; unused */
49f55535 348 SSize_t pos_min;
ea3daa5d 349 SSize_t pos_delta;
2c2d71f5 350 SV *last_found;
ea3daa5d 351 SSize_t last_end; /* min value, <0 unless valid. */
49f55535 352 SSize_t last_start_min;
ea3daa5d 353 SSize_t last_start_max;
1de06328
YO
354 SV **longest; /* Either &l_fixed, or &l_float. */
355 SV *longest_fixed; /* longest fixed string found in pattern */
49f55535 356 SSize_t offset_fixed; /* offset where it starts */
ea3daa5d 357 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
1de06328
YO
358 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
359 SV *longest_float; /* longest floating string found in pattern */
49f55535 360 SSize_t offset_float_min; /* earliest point in string it can appear */
ea3daa5d
FC
361 SSize_t offset_float_max; /* latest point in string it can appear */
362 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
49f55535 363 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
2c2d71f5
JH
364 I32 flags;
365 I32 whilem_c;
49f55535 366 SSize_t *last_closep;
653099ff 367 struct regnode_charclass_class *start_class;
2c2d71f5
JH
368} scan_data_t;
369
c02c3054
KW
370/* The below is perhaps overboard, but this allows us to save a test at the
371 * expense of a mask. This is because on both EBCDIC and ASCII machines, 'A'
372 * and 'a' differ by a single bit; the same with the upper and lower case of
373 * all other ASCII-range alphabetics. On ASCII platforms, they are 32 apart;
374 * on EBCDIC, they are 64. This uses an exclusive 'or' to find that bit and
375 * then inverts it to form a mask, with just a single 0, in the bit position
376 * where the upper- and lowercase differ. XXX There are about 40 other
377 * instances in the Perl core where this micro-optimization could be used.
378 * Should decide if maintenance cost is worse, before changing those
379 *
380 * Returns a boolean as to whether or not 'v' is either a lowercase or
381 * uppercase instance of 'c', where 'c' is in [A-Za-z]. If 'c' is a
382 * compile-time constant, the generated code is better than some optimizing
383 * compilers figure out, amounting to a mask and test. The results are
384 * meaningless if 'c' is not one of [A-Za-z] */
385#define isARG2_lower_or_UPPER_ARG1(c, v) \
386 (((v) & ~('A' ^ 'a')) == ((c) & ~('A' ^ 'a')))
387
a687059c 388/*
e50aee73 389 * Forward declarations for pregcomp()'s friends.
a687059c 390 */
a0d0e21e 391
27da23d5 392static const scan_data_t zero_scan_data =
1de06328 393 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
c277df42
IZ
394
395#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
07be1b83
YO
396#define SF_BEFORE_SEOL 0x0001
397#define SF_BEFORE_MEOL 0x0002
c277df42
IZ
398#define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
399#define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
400
09b7f37c
CB
401#ifdef NO_UNARY_PLUS
402# define SF_FIX_SHIFT_EOL (0+2)
403# define SF_FL_SHIFT_EOL (0+4)
404#else
405# define SF_FIX_SHIFT_EOL (+2)
406# define SF_FL_SHIFT_EOL (+4)
407#endif
c277df42
IZ
408
409#define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
410#define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
411
412#define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
413#define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
07be1b83
YO
414#define SF_IS_INF 0x0040
415#define SF_HAS_PAR 0x0080
416#define SF_IN_PAR 0x0100
417#define SF_HAS_EVAL 0x0200
418#define SCF_DO_SUBSTR 0x0400
653099ff
GS
419#define SCF_DO_STCLASS_AND 0x0800
420#define SCF_DO_STCLASS_OR 0x1000
421#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 422#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 423
786e8c11 424#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
e2e6a0f1 425#define SCF_SEEN_ACCEPT 0x8000
688e0391 426#define SCF_TRIE_DOING_RESTUDY 0x10000
07be1b83 427
43fead97 428#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
429
430/* The enums for all these are ordered so things work out correctly */
a62b1201 431#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
cfaf538b 432#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
00b27cfc 433#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
cfaf538b
KW
434#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
435#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
2f7f8cb1 436#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
a725e29c 437#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 438
43fead97 439#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 440
93733859 441#define OOB_NAMEDCLASS -1
b8c5462f 442
8e661ac5
KW
443/* There is no code point that is out-of-bounds, so this is problematic. But
444 * its only current use is to initialize a variable that is always set before
445 * looked at. */
446#define OOB_UNICODE 0xDEADBEEF
447
a0ed51b3
LW
448#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
449#define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
450
8615cb43 451
b45f050a
JF
452/* length of regex to show in messages that don't mark a position within */
453#define RegexLengthToShowInErrorMessages 127
454
455/*
456 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
457 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
458 * op/pragma/warn/regcomp.
459 */
7253e4e3
RK
460#define MARKER1 "<-- HERE" /* marker as it appears in the description */
461#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 462
7253e4e3 463#define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
b45f050a
JF
464
465/*
466 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
467 * arg. Show regex, up to a maximum length. If it's too long, chop and add
468 * "...".
469 */
58e23c8d 470#define _FAIL(code) STMT_START { \
bfed75c6 471 const char *ellipses = ""; \
ccb2c380
MP
472 IV len = RExC_end - RExC_precomp; \
473 \
474 if (!SIZE_ONLY) \
a5e7bc51 475 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
476 if (len > RegexLengthToShowInErrorMessages) { \
477 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
478 len = RegexLengthToShowInErrorMessages - 10; \
479 ellipses = "..."; \
480 } \
58e23c8d 481 code; \
ccb2c380 482} STMT_END
8615cb43 483
58e23c8d
YO
484#define FAIL(msg) _FAIL( \
485 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
486 msg, (int)len, RExC_precomp, ellipses))
487
488#define FAIL2(msg,arg) _FAIL( \
489 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
490 arg, (int)len, RExC_precomp, ellipses))
491
b45f050a 492/*
b45f050a
JF
493 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
494 */
ccb2c380 495#define Simple_vFAIL(m) STMT_START { \
a28509cc 496 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
497 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
498 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
499} STMT_END
b45f050a
JF
500
501/*
502 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
503 */
ccb2c380
MP
504#define vFAIL(m) STMT_START { \
505 if (!SIZE_ONLY) \
a5e7bc51 506 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
507 Simple_vFAIL(m); \
508} STMT_END
b45f050a
JF
509
510/*
511 * Like Simple_vFAIL(), but accepts two arguments.
512 */
ccb2c380 513#define Simple_vFAIL2(m,a1) STMT_START { \
a28509cc 514 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
515 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
516 (int)offset, RExC_precomp, RExC_precomp + offset); \
517} STMT_END
b45f050a
JF
518
519/*
520 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
521 */
ccb2c380
MP
522#define vFAIL2(m,a1) STMT_START { \
523 if (!SIZE_ONLY) \
a5e7bc51 524 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
525 Simple_vFAIL2(m, a1); \
526} STMT_END
b45f050a
JF
527
528
529/*
530 * Like Simple_vFAIL(), but accepts three arguments.
531 */
ccb2c380 532#define Simple_vFAIL3(m, a1, a2) STMT_START { \
a28509cc 533 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
534 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
535 (int)offset, RExC_precomp, RExC_precomp + offset); \
536} STMT_END
b45f050a
JF
537
538/*
539 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
540 */
ccb2c380
MP
541#define vFAIL3(m,a1,a2) STMT_START { \
542 if (!SIZE_ONLY) \
a5e7bc51 543 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
544 Simple_vFAIL3(m, a1, a2); \
545} STMT_END
b45f050a
JF
546
547/*
548 * Like Simple_vFAIL(), but accepts four arguments.
549 */
ccb2c380 550#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
a28509cc 551 const IV offset = RExC_parse - RExC_precomp; \
ccb2c380
MP
552 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
553 (int)offset, RExC_precomp, RExC_precomp + offset); \
554} STMT_END
b45f050a 555
95db3ffa
KW
556#define vFAIL4(m,a1,a2,a3) STMT_START { \
557 if (!SIZE_ONLY) \
558 SAVEFREESV(RExC_rx_sv); \
559 Simple_vFAIL4(m, a1, a2, a3); \
560} STMT_END
561
5e0a247b
KW
562/* m is not necessarily a "literal string", in this macro */
563#define reg_warn_non_literal_string(loc, m) STMT_START { \
564 const IV offset = loc - RExC_precomp; \
565 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
566 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
567} STMT_END
568
668c081a 569#define ckWARNreg(loc,m) STMT_START { \
a28509cc 570 const IV offset = loc - RExC_precomp; \
f10f4c18
NC
571 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
572 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
573} STMT_END
574
0d6106aa
KW
575#define vWARN_dep(loc, m) STMT_START { \
576 const IV offset = loc - RExC_precomp; \
577 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
578 (int)offset, RExC_precomp, RExC_precomp + offset); \
579} STMT_END
580
147508a2
KW
581#define ckWARNdep(loc,m) STMT_START { \
582 const IV offset = loc - RExC_precomp; \
583 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
584 m REPORT_LOCATION, \
585 (int)offset, RExC_precomp, RExC_precomp + offset); \
586} STMT_END
587
668c081a 588#define ckWARNregdep(loc,m) STMT_START { \
a28509cc 589 const IV offset = loc - RExC_precomp; \
d1d15184 590 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
f10f4c18
NC
591 m REPORT_LOCATION, \
592 (int)offset, RExC_precomp, RExC_precomp + offset); \
ccb2c380
MP
593} STMT_END
594
b23eb183 595#define ckWARN2reg_d(loc,m, a1) STMT_START { \
2335b3d3 596 const IV offset = loc - RExC_precomp; \
b23eb183 597 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
2335b3d3
KW
598 m REPORT_LOCATION, \
599 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
600} STMT_END
601
668c081a 602#define ckWARN2reg(loc, m, a1) STMT_START { \
a28509cc 603 const IV offset = loc - RExC_precomp; \
668c081a 604 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
ccb2c380
MP
605 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
606} STMT_END
607
608#define vWARN3(loc, m, a1, a2) STMT_START { \
a28509cc 609 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
610 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
611 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
612} STMT_END
613
668c081a
NC
614#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
615 const IV offset = loc - RExC_precomp; \
616 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
617 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
618} STMT_END
619
ccb2c380 620#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
a28509cc 621 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
622 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
623 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
624} STMT_END
625
668c081a
NC
626#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
627 const IV offset = loc - RExC_precomp; \
628 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
629 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
630} STMT_END
631
ccb2c380 632#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
a28509cc 633 const IV offset = loc - RExC_precomp; \
ccb2c380
MP
634 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
635 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
636} STMT_END
9d1d55b5 637
8615cb43 638
cd439c50 639/* Allow for side effects in s */
ccb2c380
MP
640#define REGC(c,s) STMT_START { \
641 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
642} STMT_END
cd439c50 643
fac92740
MJD
644/* Macros for recording node offsets. 20001227 mjd@plover.com
645 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
646 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
647 * Element 0 holds the number n.
07be1b83 648 * Position is 1 indexed.
fac92740 649 */
7122b237
YO
650#ifndef RE_TRACK_PATTERN_OFFSETS
651#define Set_Node_Offset_To_R(node,byte)
652#define Set_Node_Offset(node,byte)
653#define Set_Cur_Node_Offset
654#define Set_Node_Length_To_R(node,len)
655#define Set_Node_Length(node,len)
6a86c6ad 656#define Set_Node_Cur_Length(node,start)
7122b237
YO
657#define Node_Offset(n)
658#define Node_Length(n)
659#define Set_Node_Offset_Length(node,offset,len)
660#define ProgLen(ri) ri->u.proglen
661#define SetProgLen(ri,x) ri->u.proglen = x
662#else
663#define ProgLen(ri) ri->u.offsets[0]
664#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
665#define Set_Node_Offset_To_R(node,byte) STMT_START { \
666 if (! SIZE_ONLY) { \
667 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 668 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 669 if((node) < 0) { \
551405c4 670 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
ccb2c380
MP
671 } else { \
672 RExC_offsets[2*(node)-1] = (byte); \
673 } \
674 } \
675} STMT_END
676
677#define Set_Node_Offset(node,byte) \
678 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
679#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
680
681#define Set_Node_Length_To_R(node,len) STMT_START { \
682 if (! SIZE_ONLY) { \
683 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 684 __LINE__, (int)(node), (int)(len))); \
ccb2c380 685 if((node) < 0) { \
551405c4 686 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
ccb2c380
MP
687 } else { \
688 RExC_offsets[2*(node)] = (len); \
689 } \
690 } \
691} STMT_END
692
693#define Set_Node_Length(node,len) \
694 Set_Node_Length_To_R((node)-RExC_emit_start, len)
6a86c6ad
NC
695#define Set_Node_Cur_Length(node, start) \
696 Set_Node_Length(node, RExC_parse - start)
fac92740
MJD
697
698/* Get offsets and lengths */
699#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
700#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
701
07be1b83
YO
702#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
703 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
704 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
705} STMT_END
7122b237 706#endif
07be1b83
YO
707
708#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
709#define EXPERIMENTAL_INPLACESCAN
f427392e 710#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 711
304ee84b
YO
712#define DEBUG_STUDYDATA(str,data,depth) \
713DEBUG_OPTIMISE_MORE_r(if(data){ \
1de06328 714 PerlIO_printf(Perl_debug_log, \
304ee84b
YO
715 "%*s" str "Pos:%"IVdf"/%"IVdf \
716 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
1de06328
YO
717 (int)(depth)*2, "", \
718 (IV)((data)->pos_min), \
719 (IV)((data)->pos_delta), \
304ee84b 720 (UV)((data)->flags), \
1de06328 721 (IV)((data)->whilem_c), \
304ee84b
YO
722 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
723 is_inf ? "INF " : "" \
1de06328
YO
724 ); \
725 if ((data)->last_found) \
726 PerlIO_printf(Perl_debug_log, \
727 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
728 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
729 SvPVX_const((data)->last_found), \
730 (IV)((data)->last_end), \
731 (IV)((data)->last_start_min), \
732 (IV)((data)->last_start_max), \
733 ((data)->longest && \
734 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
735 SvPVX_const((data)->longest_fixed), \
736 (IV)((data)->offset_fixed), \
737 ((data)->longest && \
738 (data)->longest==&((data)->longest_float)) ? "*" : "", \
739 SvPVX_const((data)->longest_float), \
740 (IV)((data)->offset_float_min), \
741 (IV)((data)->offset_float_max) \
742 ); \
743 PerlIO_printf(Perl_debug_log,"\n"); \
744});
745
653099ff 746/* Mark that we cannot extend a found fixed substring at this point.
786e8c11 747 Update the longest found anchored substring and the longest found
653099ff
GS
748 floating substrings if needed. */
749
4327152a 750STATIC void
ea3daa5d
FC
751S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
752 SSize_t *minlenp, int is_inf)
c277df42 753{
e1ec3a88
AL
754 const STRLEN l = CHR_SVLEN(data->last_found);
755 const STRLEN old_l = CHR_SVLEN(*data->longest);
1de06328 756 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 757
7918f24d
NC
758 PERL_ARGS_ASSERT_SCAN_COMMIT;
759
c277df42 760 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
6b43b216 761 SvSetMagicSV(*data->longest, data->last_found);
c277df42
IZ
762 if (*data->longest == data->longest_fixed) {
763 data->offset_fixed = l ? data->last_start_min : data->pos_min;
764 if (data->flags & SF_BEFORE_EOL)
b81d288d 765 data->flags
c277df42
IZ
766 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
767 else
768 data->flags &= ~SF_FIX_BEFORE_EOL;
686b73d4 769 data->minlen_fixed=minlenp;
1de06328 770 data->lookbehind_fixed=0;
a0ed51b3 771 }
304ee84b 772 else { /* *data->longest == data->longest_float */
c277df42 773 data->offset_float_min = l ? data->last_start_min : data->pos_min;
b81d288d
AB
774 data->offset_float_max = (l
775 ? data->last_start_max
ea3daa5d
FC
776 : (data->pos_delta == SSize_t_MAX
777 ? SSize_t_MAX
778 : data->pos_min + data->pos_delta));
779 if (is_inf
780 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
781 data->offset_float_max = SSize_t_MAX;
c277df42 782 if (data->flags & SF_BEFORE_EOL)
b81d288d 783 data->flags
c277df42
IZ
784 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
785 else
786 data->flags &= ~SF_FL_BEFORE_EOL;
1de06328
YO
787 data->minlen_float=minlenp;
788 data->lookbehind_float=0;
c277df42
IZ
789 }
790 }
791 SvCUR_set(data->last_found, 0);
0eda9292 792 {
a28509cc 793 SV * const sv = data->last_found;
097eb12c
AL
794 if (SvUTF8(sv) && SvMAGICAL(sv)) {
795 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
796 if (mg)
797 mg->mg_len = 0;
798 }
0eda9292 799 }
c277df42
IZ
800 data->last_end = -1;
801 data->flags &= ~SF_BEFORE_EOL;
bcdf7404 802 DEBUG_STUDYDATA("commit: ",data,0);
c277df42
IZ
803}
804
899d20b9
KW
805/* These macros set, clear and test whether the synthetic start class ('ssc',
806 * given by the parameter) matches an empty string (EOS). This uses the
807 * 'next_off' field in the node, to save a bit in the flags field. The ssc
808 * stands alone, so there is never a next_off, so this field is otherwise
809 * unused. The EOS information is used only for compilation, but theoretically
810 * it could be passed on to the execution code. This could be used to store
811 * more than one bit of information, but only this one is currently used. */
812#define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
813#define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
814#define TEST_SSC_EOS(node) cBOOL((node)->next_off)
815
653099ff
GS
816/* Can match anything (initialization) */
817STATIC void
3fffb88a 818S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 819{
7918f24d
NC
820 PERL_ARGS_ASSERT_CL_ANYTHING;
821
f8bef550 822 ANYOF_BITMAP_SETALL(cl);
899d20b9
KW
823 cl->flags = ANYOF_UNICODE_ALL;
824 SET_SSC_EOS(cl);
3fffb88a
KW
825
826 /* If any portion of the regex is to operate under locale rules,
827 * initialization includes it. The reason this isn't done for all regexes
828 * is that the optimizer was written under the assumption that locale was
829 * all-or-nothing. Given the complexity and lack of documentation in the
830 * optimizer, and that there are inadequate test cases for locale, so many
831 * parts of it may not work properly, it is safest to avoid locale unless
832 * necessary. */
833 if (RExC_contains_locale) {
9d7a1e63 834 ANYOF_CLASS_SETALL(cl); /* /l uses class */
b25a1036 835 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
3fffb88a 836 }
9d7a1e63
KW
837 else {
838 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
839 }
653099ff
GS
840}
841
842/* Can match anything (initialization) */
843STATIC int
5f66b61c 844S_cl_is_anything(const struct regnode_charclass_class *cl)
653099ff
GS
845{
846 int value;
847
7918f24d
NC
848 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
849
3732a889 850 for (value = 0; value < ANYOF_MAX; value += 2)
653099ff
GS
851 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
852 return 1;
1aa99e6b
IH
853 if (!(cl->flags & ANYOF_UNICODE_ALL))
854 return 0;
10edeb5d 855 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
f8bef550 856 return 0;
653099ff
GS
857 return 1;
858}
859
860/* Can match anything (initialization) */
861STATIC void
e755fd73 862S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
653099ff 863{
7918f24d
NC
864 PERL_ARGS_ASSERT_CL_INIT;
865
8ecf7187 866 Zero(cl, 1, struct regnode_charclass_class);
653099ff 867 cl->type = ANYOF;
3fffb88a 868 cl_anything(pRExC_state, cl);
1411dba4 869 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
653099ff
GS
870}
871
1051e1c4 872/* These two functions currently do the exact same thing */
4d5e5e00 873#define cl_init_zero cl_init
653099ff 874
dd58aee1
KW
875/* 'AND' a given class with another one. Can create false positives. 'cl'
876 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
877 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 878STATIC void
5f66b61c 879S_cl_and(struct regnode_charclass_class *cl,
a28509cc 880 const struct regnode_charclass_class *and_with)
653099ff 881{
7918f24d 882 PERL_ARGS_ASSERT_CL_AND;
40d049e4 883
954a2af6 884 assert(PL_regkind[and_with->type] == ANYOF);
1e6ade67 885
c6b76537 886 /* I (khw) am not sure all these restrictions are necessary XXX */
1e6ade67
KW
887 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
888 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
653099ff 889 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
890 && !(and_with->flags & ANYOF_LOC_FOLD)
891 && !(cl->flags & ANYOF_LOC_FOLD)) {
653099ff
GS
892 int i;
893
894 if (and_with->flags & ANYOF_INVERT)
895 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
896 cl->bitmap[i] &= ~and_with->bitmap[i];
897 else
898 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
899 cl->bitmap[i] &= and_with->bitmap[i];
900 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
1aa99e6b 901
c6b76537 902 if (and_with->flags & ANYOF_INVERT) {
8951c461 903
c6b76537
KW
904 /* Here, the and'ed node is inverted. Get the AND of the flags that
905 * aren't affected by the inversion. Those that are affected are
906 * handled individually below */
907 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
908 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
909 cl->flags |= affected_flags;
910
911 /* We currently don't know how to deal with things that aren't in the
912 * bitmap, but we know that the intersection is no greater than what
913 * is already in cl, so let there be false positives that get sorted
914 * out after the synthetic start class succeeds, and the node is
915 * matched for real. */
916
917 /* The inversion of these two flags indicate that the resulting
918 * intersection doesn't have them */
919 if (and_with->flags & ANYOF_UNICODE_ALL) {
4713bfe1
KW
920 cl->flags &= ~ANYOF_UNICODE_ALL;
921 }
c6b76537
KW
922 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
923 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
137165a6 924 }
1aa99e6b 925 }
c6b76537 926 else { /* and'd node is not inverted */
3ad98780
KW
927 U8 outside_bitmap_but_not_utf8; /* Temp variable */
928
137165a6 929 if (! ANYOF_NONBITMAP(and_with)) {
c6b76537
KW
930
931 /* Here 'and_with' doesn't match anything outside the bitmap
932 * (except possibly ANYOF_UNICODE_ALL), which means the
933 * intersection can't either, except for ANYOF_UNICODE_ALL, in
934 * which case we don't know what the intersection is, but it's no
935 * greater than what cl already has, so can just leave it alone,
936 * with possible false positives */
937 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
938 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
871d0d1a 939 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
c6b76537 940 }
137165a6 941 }
c6b76537
KW
942 else if (! ANYOF_NONBITMAP(cl)) {
943
944 /* Here, 'and_with' does match something outside the bitmap, and cl
945 * doesn't have a list of things to match outside the bitmap. If
946 * cl can match all code points above 255, the intersection will
3ad98780
KW
947 * be those above-255 code points that 'and_with' matches. If cl
948 * can't match all Unicode code points, it means that it can't
949 * match anything outside the bitmap (since the 'if' that got us
950 * into this block tested for that), so we leave the bitmap empty.
951 */
c6b76537
KW
952 if (cl->flags & ANYOF_UNICODE_ALL) {
953 ARG_SET(cl, ARG(and_with));
3ad98780
KW
954
955 /* and_with's ARG may match things that don't require UTF8.
956 * And now cl's will too, in spite of this being an 'and'. See
957 * the comments below about the kludge */
958 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
c6b76537
KW
959 }
960 }
961 else {
962 /* Here, both 'and_with' and cl match something outside the
963 * bitmap. Currently we do not do the intersection, so just match
964 * whatever cl had at the beginning. */
965 }
966
967
3ad98780
KW
968 /* Take the intersection of the two sets of flags. However, the
969 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
970 * kludge around the fact that this flag is not treated like the others
971 * which are initialized in cl_anything(). The way the optimizer works
972 * is that the synthetic start class (SSC) is initialized to match
973 * anything, and then the first time a real node is encountered, its
974 * values are AND'd with the SSC's with the result being the values of
975 * the real node. However, there are paths through the optimizer where
976 * the AND never gets called, so those initialized bits are set
977 * inappropriately, which is not usually a big deal, as they just cause
978 * false positives in the SSC, which will just mean a probably
979 * imperceptible slow down in execution. However this bit has a
980 * higher false positive consequence in that it can cause utf8.pm,
981 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
982 * bigger slowdown and also causes significant extra memory to be used.
983 * In order to prevent this, the code now takes a different tack. The
984 * bit isn't set unless some part of the regular expression needs it,
985 * but once set it won't get cleared. This means that these extra
986 * modules won't get loaded unless there was some path through the
987 * pattern that would have required them anyway, and so any false
988 * positives that occur by not ANDing them out when they could be
989 * aren't as severe as they would be if we treated this bit like all
990 * the others */
991 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
992 & ANYOF_NONBITMAP_NON_UTF8;
c6b76537 993 cl->flags &= and_with->flags;
3ad98780 994 cl->flags |= outside_bitmap_but_not_utf8;
137165a6 995 }
653099ff
GS
996}
997
dd58aee1
KW
998/* 'OR' a given class with another one. Can create false positives. 'cl'
999 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
1000 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
653099ff 1001STATIC void
3fffb88a 1002S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
653099ff 1003{
7918f24d
NC
1004 PERL_ARGS_ASSERT_CL_OR;
1005
653099ff 1006 if (or_with->flags & ANYOF_INVERT) {
c6b76537
KW
1007
1008 /* Here, the or'd node is to be inverted. This means we take the
1009 * complement of everything not in the bitmap, but currently we don't
1010 * know what that is, so give up and match anything */
1011 if (ANYOF_NONBITMAP(or_with)) {
3fffb88a 1012 cl_anything(pRExC_state, cl);
c6b76537 1013 }
653099ff
GS
1014 /* We do not use
1015 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
1016 * <= (B1 | !B2) | (CL1 | !CL2)
1017 * which is wasteful if CL2 is small, but we ignore CL2:
1018 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
1019 * XXXX Can we handle case-fold? Unclear:
1020 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
1021 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
1022 */
c6b76537 1023 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
1024 && !(or_with->flags & ANYOF_LOC_FOLD)
1025 && !(cl->flags & ANYOF_LOC_FOLD) ) {
653099ff
GS
1026 int i;
1027
1028 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1029 cl->bitmap[i] |= ~or_with->bitmap[i];
1030 } /* XXXX: logic is complicated otherwise */
1031 else {
3fffb88a 1032 cl_anything(pRExC_state, cl);
653099ff 1033 }
c6b76537
KW
1034
1035 /* And, we can just take the union of the flags that aren't affected
1036 * by the inversion */
1037 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
1038
1039 /* For the remaining flags:
1040 ANYOF_UNICODE_ALL and inverted means to not match anything above
1041 255, which means that the union with cl should just be
1042 what cl has in it, so can ignore this flag
1043 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
f1658f69
KW
1044 is (ASCII) 127-255 to match them, but then invert that, so
1045 the union with cl should just be what cl has in it, so can
c6b76537
KW
1046 ignore this flag
1047 */
1048 } else { /* 'or_with' is not inverted */
653099ff
GS
1049 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1050 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
538b546e
KW
1051 && (!(or_with->flags & ANYOF_LOC_FOLD)
1052 || (cl->flags & ANYOF_LOC_FOLD)) ) {
653099ff
GS
1053 int i;
1054
1055 /* OR char bitmap and class bitmap separately */
1056 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1057 cl->bitmap[i] |= or_with->bitmap[i];
c96939e4
KW
1058 if (or_with->flags & ANYOF_CLASS) {
1059 ANYOF_CLASS_OR(or_with, cl);
1060 }
653099ff
GS
1061 }
1062 else { /* XXXX: logic is complicated, leave it along for a moment. */
3fffb88a 1063 cl_anything(pRExC_state, cl);
653099ff 1064 }
9826f543 1065
c6b76537
KW
1066 if (ANYOF_NONBITMAP(or_with)) {
1067
1068 /* Use the added node's outside-the-bit-map match if there isn't a
1069 * conflict. If there is a conflict (both nodes match something
1070 * outside the bitmap, but what they match outside is not the same
1071 * pointer, and hence not easily compared until XXX we extend
1072 * inversion lists this far), give up and allow the start class to
d94b1d13
KW
1073 * match everything outside the bitmap. If that stuff is all above
1074 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
c6b76537
KW
1075 if (! ANYOF_NONBITMAP(cl)) {
1076 ARG_SET(cl, ARG(or_with));
1077 }
1078 else if (ARG(cl) != ARG(or_with)) {
d94b1d13
KW
1079
1080 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1081 cl_anything(pRExC_state, cl);
1082 }
1083 else {
1084 cl->flags |= ANYOF_UNICODE_ALL;
1085 }
c6b76537 1086 }
4c34a693 1087 }
0b9668ee
KW
1088
1089 /* Take the union */
1090 cl->flags |= or_with->flags;
1aa99e6b 1091 }
653099ff
GS
1092}
1093
a3621e74
YO
1094#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1095#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1096#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1097#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1098
3dab1dad
YO
1099
1100#ifdef DEBUGGING
07be1b83 1101/*
2b8b4781
NC
1102 dump_trie(trie,widecharmap,revcharmap)
1103 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1104 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
1105
1106 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
1107 The _interim_ variants are used for debugging the interim
1108 tables that are used to generate the final compressed
1109 representation which is what dump_trie expects.
1110
486ec47a 1111 Part of the reason for their existence is to provide a form
3dab1dad 1112 of documentation as to how the different representations function.
07be1b83
YO
1113
1114*/
3dab1dad
YO
1115
1116/*
3dab1dad
YO
1117 Dumps the final compressed table form of the trie to Perl_debug_log.
1118 Used for debugging make_trie().
1119*/
b9a59e08 1120
3dab1dad 1121STATIC void
2b8b4781
NC
1122S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1123 AV *revcharmap, U32 depth)
3dab1dad
YO
1124{
1125 U32 state;
ab3bbdeb 1126 SV *sv=sv_newmortal();
55eed653 1127 int colwidth= widecharmap ? 6 : 4;
2e64971a 1128 U16 word;
3dab1dad
YO
1129 GET_RE_DEBUG_FLAGS_DECL;
1130
7918f24d 1131 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 1132
3dab1dad
YO
1133 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1134 (int)depth * 2 + 2,"",
1135 "Match","Base","Ofs" );
1136
1137 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 1138 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 1139 if ( tmp ) {
ab3bbdeb
YO
1140 PerlIO_printf( Perl_debug_log, "%*s",
1141 colwidth,
ddc5bc0f 1142 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1143 PL_colors[0], PL_colors[1],
1144 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1145 PERL_PV_ESCAPE_FIRSTCHAR
1146 )
1147 );
3dab1dad
YO
1148 }
1149 }
1150 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1151 (int)depth * 2 + 2,"");
1152
1153 for( state = 0 ; state < trie->uniquecharcount ; state++ )
ab3bbdeb 1154 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
3dab1dad
YO
1155 PerlIO_printf( Perl_debug_log, "\n");
1156
1e2e3d02 1157 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 1158 const U32 base = trie->states[ state ].trans.base;
3dab1dad
YO
1159
1160 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1161
1162 if ( trie->states[ state ].wordnum ) {
1163 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1164 } else {
1165 PerlIO_printf( Perl_debug_log, "%6s", "" );
1166 }
1167
1168 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1169
1170 if ( base ) {
1171 U32 ofs = 0;
1172
1173 while( ( base + ofs < trie->uniquecharcount ) ||
1174 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1175 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1176 ofs++;
1177
1178 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1179
1180 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1181 if ( ( base + ofs >= trie->uniquecharcount ) &&
1182 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1183 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1184 {
ab3bbdeb
YO
1185 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1186 colwidth,
3dab1dad
YO
1187 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1188 } else {
ab3bbdeb 1189 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
3dab1dad
YO
1190 }
1191 }
1192
1193 PerlIO_printf( Perl_debug_log, "]");
1194
1195 }
1196 PerlIO_printf( Perl_debug_log, "\n" );
1197 }
2e64971a
DM
1198 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1199 for (word=1; word <= trie->wordcount; word++) {
1200 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1201 (int)word, (int)(trie->wordinfo[word].prev),
1202 (int)(trie->wordinfo[word].len));
1203 }
1204 PerlIO_printf(Perl_debug_log, "\n" );
3dab1dad
YO
1205}
1206/*
3dab1dad
YO
1207 Dumps a fully constructed but uncompressed trie in list form.
1208 List tries normally only are used for construction when the number of
1209 possible chars (trie->uniquecharcount) is very high.
1210 Used for debugging make_trie().
1211*/
1212STATIC void
55eed653 1213S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1214 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1215 U32 depth)
3dab1dad
YO
1216{
1217 U32 state;
ab3bbdeb 1218 SV *sv=sv_newmortal();
55eed653 1219 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1220 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1221
1222 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1223
3dab1dad 1224 /* print out the table precompression. */
ab3bbdeb
YO
1225 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1226 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1227 "------:-----+-----------------\n" );
3dab1dad
YO
1228
1229 for( state=1 ; state < next_alloc ; state ++ ) {
1230 U16 charid;
1231
ab3bbdeb 1232 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
3dab1dad
YO
1233 (int)depth * 2 + 2,"", (UV)state );
1234 if ( ! trie->states[ state ].wordnum ) {
1235 PerlIO_printf( Perl_debug_log, "%5s| ","");
1236 } else {
1237 PerlIO_printf( Perl_debug_log, "W%4x| ",
1238 trie->states[ state ].wordnum
1239 );
1240 }
1241 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2b8b4781 1242 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb
YO
1243 if ( tmp ) {
1244 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1245 colwidth,
ddc5bc0f 1246 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1247 PL_colors[0], PL_colors[1],
1248 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1249 PERL_PV_ESCAPE_FIRSTCHAR
1250 ) ,
1e2e3d02
YO
1251 TRIE_LIST_ITEM(state,charid).forid,
1252 (UV)TRIE_LIST_ITEM(state,charid).newstate
1253 );
1254 if (!(charid % 10))
664e119d
RGS
1255 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1256 (int)((depth * 2) + 14), "");
1e2e3d02 1257 }
ab3bbdeb
YO
1258 }
1259 PerlIO_printf( Perl_debug_log, "\n");
3dab1dad
YO
1260 }
1261}
1262
1263/*
3dab1dad
YO
1264 Dumps a fully constructed but uncompressed trie in table form.
1265 This is the normal DFA style state transition table, with a few
1266 twists to facilitate compression later.
1267 Used for debugging make_trie().
1268*/
1269STATIC void
55eed653 1270S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
1271 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1272 U32 depth)
3dab1dad
YO
1273{
1274 U32 state;
1275 U16 charid;
ab3bbdeb 1276 SV *sv=sv_newmortal();
55eed653 1277 int colwidth= widecharmap ? 6 : 4;
3dab1dad 1278 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1279
1280 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
3dab1dad
YO
1281
1282 /*
1283 print out the table precompression so that we can do a visual check
1284 that they are identical.
1285 */
1286
1287 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1288
1289 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 1290 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 1291 if ( tmp ) {
ab3bbdeb
YO
1292 PerlIO_printf( Perl_debug_log, "%*s",
1293 colwidth,
ddc5bc0f 1294 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
1295 PL_colors[0], PL_colors[1],
1296 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1297 PERL_PV_ESCAPE_FIRSTCHAR
1298 )
1299 );
3dab1dad
YO
1300 }
1301 }
1302
1303 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1304
1305 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb 1306 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
3dab1dad
YO
1307 }
1308
1309 PerlIO_printf( Perl_debug_log, "\n" );
1310
1311 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1312
1313 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1314 (int)depth * 2 + 2,"",
1315 (UV)TRIE_NODENUM( state ) );
1316
1317 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
1318 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1319 if (v)
1320 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1321 else
1322 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
3dab1dad
YO
1323 }
1324 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1325 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1326 } else {
1327 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1328 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1329 }
1330 }
07be1b83 1331}
3dab1dad
YO
1332
1333#endif
1334
2e64971a 1335
786e8c11
YO
1336/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1337 startbranch: the first branch in the whole branch sequence
1338 first : start branch of sequence of branch-exact nodes.
1339 May be the same as startbranch
1340 last : Thing following the last branch.
1341 May be the same as tail.
1342 tail : item following the branch sequence
1343 count : words in the sequence
1344 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1345 depth : indent depth
3dab1dad 1346
786e8c11 1347Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 1348
786e8c11
YO
1349A trie is an N'ary tree where the branches are determined by digital
1350decomposition of the key. IE, at the root node you look up the 1st character and
1351follow that branch repeat until you find the end of the branches. Nodes can be
1352marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 1353
786e8c11 1354 /he|she|his|hers/
72f13be8 1355
786e8c11
YO
1356would convert into the following structure. Numbers represent states, letters
1357following numbers represent valid transitions on the letter from that state, if
1358the number is in square brackets it represents an accepting state, otherwise it
1359will be in parenthesis.
07be1b83 1360
786e8c11
YO
1361 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1362 | |
1363 | (2)
1364 | |
1365 (1) +-i->(6)-+-s->[7]
1366 |
1367 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 1368
786e8c11
YO
1369 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1370
1371This shows that when matching against the string 'hers' we will begin at state 1
1372read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1373then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1374is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1375single traverse. We store a mapping from accepting to state to which word was
1376matched, and then when we have multiple possibilities we try to complete the
1377rest of the regex in the order in which they occured in the alternation.
1378
1379The only prior NFA like behaviour that would be changed by the TRIE support is
1380the silent ignoring of duplicate alternations which are of the form:
1381
1382 / (DUPE|DUPE) X? (?{ ... }) Y /x
1383
4b714af6 1384Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 1385and without the optimisation. With the optimisations dupes will be silently
486ec47a 1386ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
1387the following demonstrates:
1388
1389 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1390
1391which prints out 'word' three times, but
1392
1393 'words'=~/(word|word|word)(?{ print $1 })S/
1394
1395which doesnt print it out at all. This is due to other optimisations kicking in.
1396
1397Example of what happens on a structural level:
1398
486ec47a 1399The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
1400
1401 1: CURLYM[1] {1,32767}(18)
1402 5: BRANCH(8)
1403 6: EXACT <ac>(16)
1404 8: BRANCH(11)
1405 9: EXACT <ad>(16)
1406 11: BRANCH(14)
1407 12: EXACT <ab>(16)
1408 16: SUCCEED(0)
1409 17: NOTHING(18)
1410 18: END(0)
1411
1412This would be optimizable with startbranch=5, first=5, last=16, tail=16
1413and should turn into:
1414
1415 1: CURLYM[1] {1,32767}(18)
1416 5: TRIE(16)
1417 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1418 <ac>
1419 <ad>
1420 <ab>
1421 16: SUCCEED(0)
1422 17: NOTHING(18)
1423 18: END(0)
1424
1425Cases where tail != last would be like /(?foo|bar)baz/:
1426
1427 1: BRANCH(4)
1428 2: EXACT <foo>(8)
1429 4: BRANCH(7)
1430 5: EXACT <bar>(8)
1431 7: TAIL(8)
1432 8: EXACT <baz>(10)
1433 10: END(0)
1434
1435which would be optimizable with startbranch=1, first=1, last=7, tail=8
1436and would end up looking like:
1437
1438 1: TRIE(8)
1439 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1440 <foo>
1441 <bar>
1442 7: TAIL(8)
1443 8: EXACT <baz>(10)
1444 10: END(0)
1445
c80e42f3 1446 d = uvchr_to_utf8_flags(d, uv, 0);
786e8c11
YO
1447
1448is the recommended Unicode-aware way of saying
1449
1450 *(d++) = uv;
1451*/
1452
fab2782b 1453#define TRIE_STORE_REVCHAR(val) \
786e8c11 1454 STMT_START { \
73031816 1455 if (UTF) { \
fab2782b 1456 SV *zlopp = newSV(7); /* XXX: optimize me */ \
88c9ea1e 1457 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
c80e42f3 1458 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
73031816
NC
1459 SvCUR_set(zlopp, kapow - flrbbbbb); \
1460 SvPOK_on(zlopp); \
1461 SvUTF8_on(zlopp); \
1462 av_push(revcharmap, zlopp); \
1463 } else { \
fab2782b 1464 char ooooff = (char)val; \
73031816
NC
1465 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1466 } \
1467 } STMT_END
786e8c11 1468
914a25d5
KW
1469/* This gets the next character from the input, folding it if not already
1470 * folded. */
1471#define TRIE_READ_CHAR STMT_START { \
1472 wordlen++; \
1473 if ( UTF ) { \
1474 /* if it is UTF then it is either already folded, or does not need \
1475 * folding */ \
1476 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1477 } \
1478 else if (folder == PL_fold_latin1) { \
7d006b13
KW
1479 /* This folder implies Unicode rules, which in the range expressible \
1480 * by not UTF is the lower case, with the two exceptions, one of \
1481 * which should have been taken care of before calling this */ \
1482 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
1483 uvc = toLOWER_L1(*uc); \
1484 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
1485 len = 1; \
914a25d5
KW
1486 } else { \
1487 /* raw data, will be folded later if needed */ \
1488 uvc = (U32)*uc; \
1489 len = 1; \
1490 } \
786e8c11
YO
1491} STMT_END
1492
1493
1494
1495#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1496 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
f9003953
NC
1497 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1498 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
786e8c11
YO
1499 } \
1500 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1501 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1502 TRIE_LIST_CUR( state )++; \
1503} STMT_END
07be1b83 1504
786e8c11
YO
1505#define TRIE_LIST_NEW(state) STMT_START { \
1506 Newxz( trie->states[ state ].trans.list, \
1507 4, reg_trie_trans_le ); \
1508 TRIE_LIST_CUR( state ) = 1; \
1509 TRIE_LIST_LEN( state ) = 4; \
1510} STMT_END
07be1b83 1511
786e8c11
YO
1512#define TRIE_HANDLE_WORD(state) STMT_START { \
1513 U16 dupe= trie->states[ state ].wordnum; \
1514 regnode * const noper_next = regnext( noper ); \
1515 \
786e8c11
YO
1516 DEBUG_r({ \
1517 /* store the word for dumping */ \
1518 SV* tmp; \
1519 if (OP(noper) != NOTHING) \
740cce10 1520 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 1521 else \
740cce10 1522 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 1523 av_push( trie_words, tmp ); \
786e8c11
YO
1524 }); \
1525 \
1526 curword++; \
2e64971a
DM
1527 trie->wordinfo[curword].prev = 0; \
1528 trie->wordinfo[curword].len = wordlen; \
1529 trie->wordinfo[curword].accept = state; \
786e8c11
YO
1530 \
1531 if ( noper_next < tail ) { \
1532 if (!trie->jump) \
c944940b 1533 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
7f69552c 1534 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
1535 if (!jumper) \
1536 jumper = noper_next; \
1537 if (!nextbranch) \
1538 nextbranch= regnext(cur); \
1539 } \
1540 \
1541 if ( dupe ) { \
2e64971a
DM
1542 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1543 /* chain, so that when the bits of chain are later */\
1544 /* linked together, the dups appear in the chain */\
1545 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1546 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
1547 } else { \
1548 /* we haven't inserted this word yet. */ \
1549 trie->states[ state ].wordnum = curword; \
1550 } \
1551} STMT_END
07be1b83 1552
3dab1dad 1553
786e8c11
YO
1554#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1555 ( ( base + charid >= ucharcount \
1556 && base + charid < ubound \
1557 && state == trie->trans[ base - ucharcount + charid ].check \
1558 && trie->trans[ base - ucharcount + charid ].next ) \
1559 ? trie->trans[ base - ucharcount + charid ].next \
1560 : ( state==1 ? special : 0 ) \
1561 )
3dab1dad 1562
786e8c11
YO
1563#define MADE_TRIE 1
1564#define MADE_JUMP_TRIE 2
1565#define MADE_EXACT_TRIE 4
3dab1dad 1566
a3621e74 1567STATIC I32
786e8c11 1568S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
a3621e74 1569{
27da23d5 1570 dVAR;
a3621e74
YO
1571 /* first pass, loop through and scan words */
1572 reg_trie_data *trie;
55eed653 1573 HV *widecharmap = NULL;
2b8b4781 1574 AV *revcharmap = newAV();
a3621e74 1575 regnode *cur;
9f7f3913 1576 const U32 uniflags = UTF8_ALLOW_DEFAULT;
a3621e74
YO
1577 STRLEN len = 0;
1578 UV uvc = 0;
1579 U16 curword = 0;
1580 U32 next_alloc = 0;
786e8c11
YO
1581 regnode *jumper = NULL;
1582 regnode *nextbranch = NULL;
7f69552c 1583 regnode *convert = NULL;
2e64971a 1584 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 1585 /* we just use folder as a flag in utf8 */
1e696034 1586 const U8 * folder = NULL;
a3621e74 1587
2b8b4781
NC
1588#ifdef DEBUGGING
1589 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1590 AV *trie_words = NULL;
1591 /* along with revcharmap, this only used during construction but both are
1592 * useful during debugging so we store them in the struct when debugging.
8e11feef 1593 */
2b8b4781
NC
1594#else
1595 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
3dab1dad 1596 STRLEN trie_charcount=0;
3dab1dad 1597#endif
2b8b4781 1598 SV *re_trie_maxbuff;
a3621e74 1599 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
1600
1601 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
1602#ifndef DEBUGGING
1603 PERL_UNUSED_ARG(depth);
1604#endif
a3621e74 1605
1e696034 1606 switch (flags) {
79a81a6e 1607 case EXACT: break;
2f7f8cb1 1608 case EXACTFA:
fab2782b 1609 case EXACTFU_SS:
1e696034
KW
1610 case EXACTFU: folder = PL_fold_latin1; break;
1611 case EXACTF: folder = PL_fold; break;
1612 case EXACTFL: folder = PL_fold_locale; break;
fab2782b 1613 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
1614 }
1615
c944940b 1616 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 1617 trie->refcount = 1;
3dab1dad 1618 trie->startstate = 1;
786e8c11 1619 trie->wordcount = word_count;
f8fc2ecf 1620 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 1621 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
fab2782b 1622 if (flags == EXACT)
c944940b 1623 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
1624 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1625 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1626
a3621e74 1627 DEBUG_r({
2b8b4781 1628 trie_words = newAV();
a3621e74 1629 });
a3621e74 1630
0111c4fd 1631 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
a3621e74 1632 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 1633 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 1634 }
df826430 1635 DEBUG_TRIE_COMPILE_r({
3dab1dad 1636 PerlIO_printf( Perl_debug_log,
786e8c11 1637 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
3dab1dad
YO
1638 (int)depth * 2 + 2, "",
1639 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
786e8c11 1640 REG_NODE_NUM(last), REG_NODE_NUM(tail),
85c3142d 1641 (int)depth);
3dab1dad 1642 });
7f69552c
YO
1643
1644 /* Find the node we are going to overwrite */
1645 if ( first == startbranch && OP( last ) != BRANCH ) {
1646 /* whole branch chain */
1647 convert = first;
1648 } else {
1649 /* branch sub-chain */
1650 convert = NEXTOPER( first );
1651 }
1652
a3621e74
YO
1653 /* -- First loop and Setup --
1654
1655 We first traverse the branches and scan each word to determine if it
1656 contains widechars, and how many unique chars there are, this is
1657 important as we have to build a table with at least as many columns as we
1658 have unique chars.
1659
1660 We use an array of integers to represent the character codes 0..255
38a44b82 1661 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
a3621e74
YO
1662 native representation of the character value as the key and IV's for the
1663 coded index.
1664
1665 *TODO* If we keep track of how many times each character is used we can
1666 remap the columns so that the table compression later on is more
3b753521 1667 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
1668 middle and the least common are on the outside. IMO this would be better
1669 than a most to least common mapping as theres a decent chance the most
1670 common letter will share a node with the least common, meaning the node
486ec47a 1671 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
1672 case is when we have the least common nodes twice.
1673
1674 */
1675
a3621e74 1676 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 1677 regnode *noper = NEXTOPER( cur );
e1ec3a88 1678 const U8 *uc = (U8*)STRING( noper );
df826430 1679 const U8 *e = uc + STR_LEN( noper );
a3621e74 1680 STRLEN foldlen = 0;
07be1b83 1681 U32 wordlen = 0; /* required init */
afa96d92
KW
1682 STRLEN minbytes = 0;
1683 STRLEN maxbytes = 0;
02daf0ab 1684 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
a3621e74 1685
3dab1dad 1686 if (OP(noper) == NOTHING) {
df826430
YO
1687 regnode *noper_next= regnext(noper);
1688 if (noper_next != tail && OP(noper_next) == flags) {
1689 noper = noper_next;
1690 uc= (U8*)STRING(noper);
1691 e= uc + STR_LEN(noper);
1692 trie->minlen= STR_LEN(noper);
1693 } else {
1694 trie->minlen= 0;
1695 continue;
1696 }
3dab1dad 1697 }
df826430 1698
fab2782b 1699 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
1700 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1701 regardless of encoding */
fab2782b
YO
1702 if (OP( noper ) == EXACTFU_SS) {
1703 /* false positives are ok, so just set this */
0dc4a61d 1704 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
fab2782b
YO
1705 }
1706 }
a3621e74 1707 for ( ; uc < e ; uc += len ) {
3dab1dad 1708 TRIE_CHARCOUNT(trie)++;
a3621e74 1709 TRIE_READ_CHAR;
645de4ce
KW
1710
1711 /* Acummulate to the current values, the range in the number of
1712 * bytes that this character could match. The max is presumed to
1713 * be the same as the folded input (which TRIE_READ_CHAR returns),
1714 * except that when this is not in UTF-8, it could be matched
1715 * against a string which is UTF-8, and the variant characters
1716 * could be 2 bytes instead of the 1 here. Likewise, for the
1717 * minimum number of bytes when not folded. When folding, the min
1718 * is assumed to be 1 byte could fold to match the single character
1719 * here, or in the case of a multi-char fold, 1 byte can fold to
1720 * the whole sequence. 'foldlen' is used to denote whether we are
1721 * in such a sequence, skipping the min setting if so. XXX TODO
1722 * Use the exact list of what folds to each character, from
1723 * PL_utf8_foldclosures */
1724 if (UTF) {
1725 maxbytes += UTF8SKIP(uc);
1726 if (! folder) {
1727 /* A non-UTF-8 string could be 1 byte to match our 2 */
1728 minbytes += (UTF8_IS_DOWNGRADEABLE_START(*uc))
1729 ? 1
1730 : UTF8SKIP(uc);
1731 }
1732 else {
1733 if (foldlen) {
1734 foldlen -= UTF8SKIP(uc);
1735 }
1736 else {
1737 foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e);
1738 minbytes++;
1739 }
1740 }
1741 }
1742 else {
1743 maxbytes += (UNI_IS_INVARIANT(*uc))
1744 ? 1
1745 : 2;
1746 if (! folder) {
1747 minbytes++;
1748 }
1749 else {
1750 if (foldlen) {
1751 foldlen--;
1752 }
1753 else {
1754 foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e);
1755 minbytes++;
1756 }
1757 }
1758 }
a3621e74 1759 if ( uvc < 256 ) {
fab2782b
YO
1760 if ( folder ) {
1761 U8 folded= folder[ (U8) uvc ];
1762 if ( !trie->charmap[ folded ] ) {
1763 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1764 TRIE_STORE_REVCHAR( folded );
1765 }
1766 }
a3621e74
YO
1767 if ( !trie->charmap[ uvc ] ) {
1768 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 1769 TRIE_STORE_REVCHAR( uvc );
a3621e74 1770 }
02daf0ab 1771 if ( set_bit ) {
62012aee
KW
1772 /* store the codepoint in the bitmap, and its folded
1773 * equivalent. */
fab2782b 1774 TRIE_BITMAP_SET(trie, uvc);
0921ee73
T
1775
1776 /* store the folded codepoint */
fab2782b 1777 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
0921ee73
T
1778
1779 if ( !UTF ) {
1780 /* store first byte of utf8 representation of
acdf4139 1781 variant codepoints */
f4cd282c 1782 if (! NATIVE_IS_INVARIANT(uvc)) {
acdf4139 1783 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
0921ee73
T
1784 }
1785 }
02daf0ab
YO
1786 set_bit = 0; /* We've done our bit :-) */
1787 }
a3621e74
YO
1788 } else {
1789 SV** svpp;
55eed653
NC
1790 if ( !widecharmap )
1791 widecharmap = newHV();
a3621e74 1792
55eed653 1793 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
1794
1795 if ( !svpp )
e4584336 1796 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
a3621e74
YO
1797
1798 if ( !SvTRUE( *svpp ) ) {
1799 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 1800 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
1801 }
1802 }
1803 }
3dab1dad 1804 if( cur == first ) {
afa96d92
KW
1805 trie->minlen = minbytes;
1806 trie->maxlen = maxbytes;
1807 } else if (minbytes < trie->minlen) {
1808 trie->minlen = minbytes;
1809 } else if (maxbytes > trie->maxlen) {
1810 trie->maxlen = maxbytes;
fab2782b 1811 }
a3621e74
YO
1812 } /* end first pass */
1813 DEBUG_TRIE_COMPILE_r(
3dab1dad
YO
1814 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1815 (int)depth * 2 + 2,"",
55eed653 1816 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
1817 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1818 (int)trie->minlen, (int)trie->maxlen )
a3621e74 1819 );
a3621e74
YO
1820
1821 /*
1822 We now know what we are dealing with in terms of unique chars and
1823 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
1824 representation using a flat table will take. If it's over a reasonable
1825 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
1826 conservative but potentially much slower representation using an array
1827 of lists.
1828
1829 At the end we convert both representations into the same compressed
1830 form that will be used in regexec.c for matching with. The latter
1831 is a form that cannot be used to construct with but has memory
1832 properties similar to the list form and access properties similar
1833 to the table form making it both suitable for fast searches and
1834 small enough that its feasable to store for the duration of a program.
1835
1836 See the comment in the code where the compressed table is produced
1837 inplace from the flat tabe representation for an explanation of how
1838 the compression works.
1839
1840 */
1841
1842
2e64971a
DM
1843 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1844 prev_states[1] = 0;
1845
3dab1dad 1846 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
a3621e74
YO
1847 /*
1848 Second Pass -- Array Of Lists Representation
1849
1850 Each state will be represented by a list of charid:state records
1851 (reg_trie_trans_le) the first such element holds the CUR and LEN
1852 points of the allocated array. (See defines above).
1853
1854 We build the initial structure using the lists, and then convert
1855 it into the compressed table form which allows faster lookups
1856 (but cant be modified once converted).
a3621e74
YO
1857 */
1858
a3621e74
YO
1859 STRLEN transcount = 1;
1860
1e2e3d02
YO
1861 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1862 "%*sCompiling trie using list compiler\n",
1863 (int)depth * 2 + 2, ""));
686b73d4 1864
c944940b
JH
1865 trie->states = (reg_trie_state *)
1866 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1867 sizeof(reg_trie_state) );
a3621e74
YO
1868 TRIE_LIST_NEW(1);
1869 next_alloc = 2;
1870
1871 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1872
df826430 1873 regnode *noper = NEXTOPER( cur );
c445ea15 1874 U8 *uc = (U8*)STRING( noper );
df826430 1875 const U8 *e = uc + STR_LEN( noper );
c445ea15
AL
1876 U32 state = 1; /* required init */
1877 U16 charid = 0; /* sanity init */
07be1b83 1878 U32 wordlen = 0; /* required init */
c445ea15 1879
df826430
YO
1880 if (OP(noper) == NOTHING) {
1881 regnode *noper_next= regnext(noper);
1882 if (noper_next != tail && OP(noper_next) == flags) {
1883 noper = noper_next;
1884 uc= (U8*)STRING(noper);
1885 e= uc + STR_LEN(noper);
1886 }
1887 }
1888
3dab1dad 1889 if (OP(noper) != NOTHING) {
786e8c11 1890 for ( ; uc < e ; uc += len ) {
c445ea15 1891
786e8c11 1892 TRIE_READ_CHAR;
c445ea15 1893
786e8c11
YO
1894 if ( uvc < 256 ) {
1895 charid = trie->charmap[ uvc ];
c445ea15 1896 } else {
55eed653 1897 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11
YO
1898 if ( !svpp ) {
1899 charid = 0;
1900 } else {
1901 charid=(U16)SvIV( *svpp );
1902 }
c445ea15 1903 }
786e8c11
YO
1904 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1905 if ( charid ) {
a3621e74 1906
786e8c11
YO
1907 U16 check;
1908 U32 newstate = 0;
a3621e74 1909
786e8c11
YO
1910 charid--;
1911 if ( !trie->states[ state ].trans.list ) {
1912 TRIE_LIST_NEW( state );
c445ea15 1913 }
786e8c11
YO
1914 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1915 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1916 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1917 break;
1918 }
1919 }
1920 if ( ! newstate ) {
1921 newstate = next_alloc++;
2e64971a 1922 prev_states[newstate] = state;
786e8c11
YO
1923 TRIE_LIST_PUSH( state, charid, newstate );
1924 transcount++;
1925 }
1926 state = newstate;
1927 } else {
1928 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
c445ea15 1929 }
a28509cc 1930 }
c445ea15 1931 }
3dab1dad 1932 TRIE_HANDLE_WORD(state);
a3621e74
YO
1933
1934 } /* end second pass */
1935
1e2e3d02
YO
1936 /* next alloc is the NEXT state to be allocated */
1937 trie->statecount = next_alloc;
c944940b
JH
1938 trie->states = (reg_trie_state *)
1939 PerlMemShared_realloc( trie->states,
1940 next_alloc
1941 * sizeof(reg_trie_state) );
a3621e74 1942
3dab1dad 1943 /* and now dump it out before we compress it */
2b8b4781
NC
1944 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1945 revcharmap, next_alloc,
1946 depth+1)
1e2e3d02 1947 );
a3621e74 1948
c944940b
JH
1949 trie->trans = (reg_trie_trans *)
1950 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
1951 {
1952 U32 state;
a3621e74
YO
1953 U32 tp = 0;
1954 U32 zp = 0;
1955
1956
1957 for( state=1 ; state < next_alloc ; state ++ ) {
1958 U32 base=0;
1959
1960 /*
1961 DEBUG_TRIE_COMPILE_MORE_r(
1962 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1963 );
1964 */
1965
1966 if (trie->states[state].trans.list) {
1967 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1968 U16 maxid=minid;
a28509cc 1969 U16 idx;
a3621e74
YO
1970
1971 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
1972 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1973 if ( forid < minid ) {
1974 minid=forid;
1975 } else if ( forid > maxid ) {
1976 maxid=forid;
1977 }
a3621e74
YO
1978 }
1979 if ( transcount < tp + maxid - minid + 1) {
1980 transcount *= 2;
c944940b
JH
1981 trie->trans = (reg_trie_trans *)
1982 PerlMemShared_realloc( trie->trans,
446bd890
NC
1983 transcount
1984 * sizeof(reg_trie_trans) );
a3621e74
YO
1985 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1986 }
1987 base = trie->uniquecharcount + tp - minid;
1988 if ( maxid == minid ) {
1989 U32 set = 0;
1990 for ( ; zp < tp ; zp++ ) {
1991 if ( ! trie->trans[ zp ].next ) {
1992 base = trie->uniquecharcount + zp - minid;
1993 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1994 trie->trans[ zp ].check = state;
1995 set = 1;
1996 break;
1997 }
1998 }
1999 if ( !set ) {
2000 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
2001 trie->trans[ tp ].check = state;
2002 tp++;
2003 zp = tp;
2004 }
2005 } else {
2006 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15 2007 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
a3621e74
YO
2008 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
2009 trie->trans[ tid ].check = state;
2010 }
2011 tp += ( maxid - minid + 1 );
2012 }
2013 Safefree(trie->states[ state ].trans.list);
2014 }
2015 /*
2016 DEBUG_TRIE_COMPILE_MORE_r(
2017 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2018 );
2019 */
2020 trie->states[ state ].trans.base=base;
2021 }
cc601c31 2022 trie->lasttrans = tp + 1;
a3621e74
YO
2023 }
2024 } else {
2025 /*
2026 Second Pass -- Flat Table Representation.
2027
2028 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
2029 We know that we will need Charcount+1 trans at most to store the data
2030 (one row per char at worst case) So we preallocate both structures
2031 assuming worst case.
2032
2033 We then construct the trie using only the .next slots of the entry
2034 structs.
2035
3b753521 2036 We use the .check field of the first entry of the node temporarily to
a3621e74
YO
2037 make compression both faster and easier by keeping track of how many non
2038 zero fields are in the node.
2039
2040 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2041 transition.
2042
2043 There are two terms at use here: state as a TRIE_NODEIDX() which is a
2044 number representing the first entry of the node, and state as a
2045 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
2046 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
2047 are 2 entrys per node. eg:
2048
2049 A B A B
2050 1. 2 4 1. 3 7
2051 2. 0 3 3. 0 5
2052 3. 0 0 5. 0 0
2053 4. 0 0 7. 0 0
2054
2055 The table is internally in the right hand, idx form. However as we also
2056 have to deal with the states array which is indexed by nodenum we have to
2057 use TRIE_NODENUM() to convert.
2058
2059 */
1e2e3d02
YO
2060 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2061 "%*sCompiling trie using table compiler\n",
2062 (int)depth * 2 + 2, ""));
3dab1dad 2063
c944940b
JH
2064 trie->trans = (reg_trie_trans *)
2065 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2066 * trie->uniquecharcount + 1,
2067 sizeof(reg_trie_trans) );
2068 trie->states = (reg_trie_state *)
2069 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2070 sizeof(reg_trie_state) );
a3621e74
YO
2071 next_alloc = trie->uniquecharcount + 1;
2072
3dab1dad 2073
a3621e74
YO
2074 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2075
df826430 2076 regnode *noper = NEXTOPER( cur );
a28509cc 2077 const U8 *uc = (U8*)STRING( noper );
df826430 2078 const U8 *e = uc + STR_LEN( noper );
a3621e74
YO
2079
2080 U32 state = 1; /* required init */
2081
2082 U16 charid = 0; /* sanity init */
2083 U32 accept_state = 0; /* sanity init */
a3621e74 2084
07be1b83 2085 U32 wordlen = 0; /* required init */
a3621e74 2086
df826430
YO
2087 if (OP(noper) == NOTHING) {
2088 regnode *noper_next= regnext(noper);
2089 if (noper_next != tail && OP(noper_next) == flags) {
2090 noper = noper_next;
2091 uc= (U8*)STRING(noper);
2092 e= uc + STR_LEN(noper);
2093 }
2094 }
fab2782b 2095
3dab1dad 2096 if ( OP(noper) != NOTHING ) {
786e8c11 2097 for ( ; uc < e ; uc += len ) {
a3621e74 2098
786e8c11 2099 TRIE_READ_CHAR;
a3621e74 2100
786e8c11
YO
2101 if ( uvc < 256 ) {
2102 charid = trie->charmap[ uvc ];
2103 } else {
55eed653 2104 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
786e8c11 2105 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 2106 }
786e8c11
YO
2107 if ( charid ) {
2108 charid--;
2109 if ( !trie->trans[ state + charid ].next ) {
2110 trie->trans[ state + charid ].next = next_alloc;
2111 trie->trans[ state ].check++;
2e64971a
DM
2112 prev_states[TRIE_NODENUM(next_alloc)]
2113 = TRIE_NODENUM(state);
786e8c11
YO
2114 next_alloc += trie->uniquecharcount;
2115 }
2116 state = trie->trans[ state + charid ].next;
2117 } else {
2118 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2119 }
2120 /* charid is now 0 if we dont know the char read, or nonzero if we do */
a3621e74 2121 }
a3621e74 2122 }
3dab1dad
YO
2123 accept_state = TRIE_NODENUM( state );
2124 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
2125
2126 } /* end second pass */
2127
3dab1dad 2128 /* and now dump it out before we compress it */
2b8b4781
NC
2129 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2130 revcharmap,
2131 next_alloc, depth+1));
a3621e74 2132
a3621e74
YO
2133 {
2134 /*
2135 * Inplace compress the table.*
2136
2137 For sparse data sets the table constructed by the trie algorithm will
2138 be mostly 0/FAIL transitions or to put it another way mostly empty.
2139 (Note that leaf nodes will not contain any transitions.)
2140
2141 This algorithm compresses the tables by eliminating most such
2142 transitions, at the cost of a modest bit of extra work during lookup:
2143
2144 - Each states[] entry contains a .base field which indicates the
2145 index in the state[] array wheres its transition data is stored.
2146
3b753521 2147 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
2148
2149 - If .base is nonzero then charid is added to it to find an entry in
2150 the trans array.
2151
2152 -If trans[states[state].base+charid].check!=state then the
2153 transition is taken to be a 0/Fail transition. Thus if there are fail
2154 transitions at the front of the node then the .base offset will point
2155 somewhere inside the previous nodes data (or maybe even into a node
2156 even earlier), but the .check field determines if the transition is
2157 valid.
2158
786e8c11 2159 XXX - wrong maybe?
a3621e74 2160 The following process inplace converts the table to the compressed
3b753521 2161 table: We first do not compress the root node 1,and mark all its
a3621e74 2162 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
2163 allows us to do a DFA construction from the compressed table later,
2164 and ensures that any .base pointers we calculate later are greater
2165 than 0.
a3621e74
YO
2166
2167 - We set 'pos' to indicate the first entry of the second node.
2168
2169 - We then iterate over the columns of the node, finding the first and
2170 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2171 and set the .check pointers accordingly, and advance pos
2172 appropriately and repreat for the next node. Note that when we copy
2173 the next pointers we have to convert them from the original
2174 NODEIDX form to NODENUM form as the former is not valid post
2175 compression.
2176
2177 - If a node has no transitions used we mark its base as 0 and do not
2178 advance the pos pointer.
2179
2180 - If a node only has one transition we use a second pointer into the
2181 structure to fill in allocated fail transitions from other states.
2182 This pointer is independent of the main pointer and scans forward
2183 looking for null transitions that are allocated to a state. When it
2184 finds one it writes the single transition into the "hole". If the
786e8c11 2185 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
2186
2187 - Once compressed we can Renew/realloc the structures to release the
2188 excess space.
2189
2190 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2191 specifically Fig 3.47 and the associated pseudocode.
2192
2193 demq
2194 */
a3b680e6 2195 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 2196 U32 state, charid;
a3621e74 2197 U32 pos = 0, zp=0;
1e2e3d02 2198 trie->statecount = laststate;
a3621e74
YO
2199
2200 for ( state = 1 ; state < laststate ; state++ ) {
2201 U8 flag = 0;
a28509cc
AL
2202 const U32 stateidx = TRIE_NODEIDX( state );
2203 const U32 o_used = trie->trans[ stateidx ].check;
2204 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
2205 trie->trans[ stateidx ].check = 0;
2206
2207 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2208 if ( flag || trie->trans[ stateidx + charid ].next ) {
2209 if ( trie->trans[ stateidx + charid ].next ) {
2210 if (o_used == 1) {
2211 for ( ; zp < pos ; zp++ ) {
2212 if ( ! trie->trans[ zp ].next ) {
2213 break;
2214 }
2215 }
2216 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2217 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2218 trie->trans[ zp ].check = state;
2219 if ( ++zp > pos ) pos = zp;
2220 break;
2221 }
2222 used--;
2223 }
2224 if ( !flag ) {
2225 flag = 1;
2226 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2227 }
2228 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2229 trie->trans[ pos ].check = state;
2230 pos++;
2231 }
2232 }
2233 }
cc601c31 2234 trie->lasttrans = pos + 1;
c944940b
JH
2235 trie->states = (reg_trie_state *)
2236 PerlMemShared_realloc( trie->states, laststate
2237 * sizeof(reg_trie_state) );
a3621e74 2238 DEBUG_TRIE_COMPILE_MORE_r(
e4584336 2239 PerlIO_printf( Perl_debug_log,
3dab1dad
YO
2240 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2241 (int)depth * 2 + 2,"",
2242 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
5d7488b2
AL
2243 (IV)next_alloc,
2244 (IV)pos,
a3621e74
YO
2245 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2246 );
2247
2248 } /* end table compress */
2249 }
1e2e3d02
YO
2250 DEBUG_TRIE_COMPILE_MORE_r(
2251 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2252 (int)depth * 2 + 2, "",
2253 (UV)trie->statecount,
2254 (UV)trie->lasttrans)
2255 );
cc601c31 2256 /* resize the trans array to remove unused space */
c944940b
JH
2257 trie->trans = (reg_trie_trans *)
2258 PerlMemShared_realloc( trie->trans, trie->lasttrans
2259 * sizeof(reg_trie_trans) );
a3621e74 2260
3b753521 2261 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
2262 U8 nodetype =(U8)(flags & 0xFF);
2263 char *str=NULL;
786e8c11 2264
07be1b83 2265#ifdef DEBUGGING
e62cc96a 2266 regnode *optimize = NULL;
7122b237
YO
2267#ifdef RE_TRACK_PATTERN_OFFSETS
2268
b57a0404
JH
2269 U32 mjd_offset = 0;
2270 U32 mjd_nodelen = 0;
7122b237
YO
2271#endif /* RE_TRACK_PATTERN_OFFSETS */
2272#endif /* DEBUGGING */
a3621e74 2273 /*
3dab1dad
YO
2274 This means we convert either the first branch or the first Exact,
2275 depending on whether the thing following (in 'last') is a branch
2276 or not and whther first is the startbranch (ie is it a sub part of
2277 the alternation or is it the whole thing.)
3b753521 2278 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 2279 the whole branch sequence, including the first.
a3621e74 2280 */
3dab1dad 2281 /* Find the node we are going to overwrite */
7f69552c 2282 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 2283 /* branch sub-chain */
3dab1dad 2284 NEXT_OFF( first ) = (U16)(last - first);
7122b237 2285#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
2286 DEBUG_r({
2287 mjd_offset= Node_Offset((convert));
2288 mjd_nodelen= Node_Length((convert));
2289 });
7122b237 2290#endif
7f69552c 2291 /* whole branch chain */
7122b237
YO
2292 }
2293#ifdef RE_TRACK_PATTERN_OFFSETS
2294 else {
7f69552c
YO
2295 DEBUG_r({
2296 const regnode *nop = NEXTOPER( convert );
2297 mjd_offset= Node_Offset((nop));
2298 mjd_nodelen= Node_Length((nop));
2299 });
07be1b83
YO
2300 }
2301 DEBUG_OPTIMISE_r(
2302 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2303 (int)depth * 2 + 2, "",
786e8c11 2304 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 2305 );
7122b237 2306#endif
3dab1dad
YO
2307 /* But first we check to see if there is a common prefix we can
2308 split out as an EXACT and put in front of the TRIE node. */
2309 trie->startstate= 1;
55eed653 2310 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3dab1dad 2311 U32 state;
1e2e3d02 2312 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 2313 U32 ofs = 0;
8e11feef
RGS
2314 I32 idx = -1;
2315 U32 count = 0;
2316 const U32 base = trie->states[ state ].trans.base;
a3621e74 2317
3dab1dad 2318 if ( trie->states[state].wordnum )
8e11feef 2319 count = 1;
a3621e74 2320
8e11feef 2321 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
2322 if ( ( base + ofs >= trie->uniquecharcount ) &&
2323 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
2324 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2325 {
3dab1dad 2326 if ( ++count > 1 ) {
2b8b4781 2327 SV **tmp = av_fetch( revcharmap, ofs, 0);
07be1b83 2328 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2329 if ( state == 1 ) break;
3dab1dad
YO
2330 if ( count == 2 ) {
2331 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2332 DEBUG_OPTIMISE_r(
8e11feef
RGS
2333 PerlIO_printf(Perl_debug_log,
2334 "%*sNew Start State=%"UVuf" Class: [",
2335 (int)depth * 2 + 2, "",
786e8c11 2336 (UV)state));
be8e71aa 2337 if (idx >= 0) {
2b8b4781 2338 SV ** const tmp = av_fetch( revcharmap, idx, 0);
be8e71aa 2339 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 2340
3dab1dad 2341 TRIE_BITMAP_SET(trie,*ch);
8e11feef
RGS
2342 if ( folder )
2343 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3dab1dad 2344 DEBUG_OPTIMISE_r(
f1f66076 2345 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3dab1dad 2346 );
8e11feef
RGS
2347 }
2348 }
2349 TRIE_BITMAP_SET(trie,*ch);
2350 if ( folder )
2351 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2352 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2353 }
2354 idx = ofs;
2355 }
3dab1dad
YO
2356 }
2357 if ( count == 1 ) {
2b8b4781 2358 SV **tmp = av_fetch( revcharmap, idx, 0);
c490c714
YO
2359 STRLEN len;
2360 char *ch = SvPV( *tmp, len );
de734bd5
A
2361 DEBUG_OPTIMISE_r({
2362 SV *sv=sv_newmortal();
8e11feef
RGS
2363 PerlIO_printf( Perl_debug_log,
2364 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2365 (int)depth * 2 + 2, "",
de734bd5
A
2366 (UV)state, (UV)idx,
2367 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2368 PL_colors[0], PL_colors[1],
2369 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2370 PERL_PV_ESCAPE_FIRSTCHAR
2371 )
2372 );
2373 });
3dab1dad
YO
2374 if ( state==1 ) {
2375 OP( convert ) = nodetype;
2376 str=STRING(convert);
2377 STR_LEN(convert)=0;
2378 }
c490c714
YO
2379 STR_LEN(convert) += len;
2380 while (len--)
de734bd5 2381 *str++ = *ch++;
8e11feef 2382 } else {
f9049ba1 2383#ifdef DEBUGGING
8e11feef
RGS
2384 if (state>1)
2385 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
f9049ba1 2386#endif
8e11feef
RGS
2387 break;
2388 }
2389 }
2e64971a 2390 trie->prefixlen = (state-1);
3dab1dad 2391 if (str) {
8e11feef 2392 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 2393 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 2394 trie->startstate = state;
07be1b83
YO
2395 trie->minlen -= (state - 1);
2396 trie->maxlen -= (state - 1);
33809eae
JH
2397#ifdef DEBUGGING
2398 /* At least the UNICOS C compiler choked on this
2399 * being argument to DEBUG_r(), so let's just have
2400 * it right here. */
2401 if (
2402#ifdef PERL_EXT_RE_BUILD
2403 1
2404#else
2405 DEBUG_r_TEST
2406#endif
2407 ) {
2408 regnode *fix = convert;
2409 U32 word = trie->wordcount;
2410 mjd_nodelen++;
2411 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2412 while( ++fix < n ) {
2413 Set_Node_Offset_Length(fix, 0, 0);
2414 }
2415 while (word--) {
2416 SV ** const tmp = av_fetch( trie_words, word, 0 );
2417 if (tmp) {
2418 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2419 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2420 else
2421 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2422 }
2423 }
2424 }
2425#endif
8e11feef
RGS
2426 if (trie->maxlen) {
2427 convert = n;
2428 } else {
3dab1dad 2429 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 2430 DEBUG_r(optimize= n);
3dab1dad
YO
2431 }
2432 }
2433 }
a5ca303d
YO
2434 if (!jumper)
2435 jumper = last;
3dab1dad 2436 if ( trie->maxlen ) {
8e11feef
RGS
2437 NEXT_OFF( convert ) = (U16)(tail - convert);
2438 ARG_SET( convert, data_slot );
786e8c11
YO
2439 /* Store the offset to the first unabsorbed branch in
2440 jump[0], which is otherwise unused by the jump logic.
2441 We use this when dumping a trie and during optimisation. */
2442 if (trie->jump)
7f69552c 2443 trie->jump[0] = (U16)(nextbranch - convert);
a5ca303d 2444
6c48061a
YO
2445 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2446 * and there is a bitmap
2447 * and the first "jump target" node we found leaves enough room
2448 * then convert the TRIE node into a TRIEC node, with the bitmap
2449 * embedded inline in the opcode - this is hypothetically faster.
2450 */
2451 if ( !trie->states[trie->startstate].wordnum
2452 && trie->bitmap
2453 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
786e8c11
YO
2454 {
2455 OP( convert ) = TRIEC;
2456 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
446bd890 2457 PerlMemShared_free(trie->bitmap);
786e8c11
YO
2458 trie->bitmap= NULL;
2459 } else
2460 OP( convert ) = TRIE;
a3621e74 2461
3dab1dad
YO
2462 /* store the type in the flags */
2463 convert->flags = nodetype;
a5ca303d
YO
2464 DEBUG_r({
2465 optimize = convert
2466 + NODE_STEP_REGNODE
2467 + regarglen[ OP( convert ) ];
2468 });
2469 /* XXX We really should free up the resource in trie now,
2470 as we won't use them - (which resources?) dmq */
3dab1dad 2471 }
a3621e74 2472 /* needed for dumping*/
e62cc96a 2473 DEBUG_r(if (optimize) {
07be1b83 2474 regnode *opt = convert;
bcdf7404 2475
e62cc96a 2476 while ( ++opt < optimize) {
07be1b83
YO
2477 Set_Node_Offset_Length(opt,0,0);
2478 }
786e8c11
YO
2479 /*
2480 Try to clean up some of the debris left after the
2481 optimisation.
a3621e74 2482 */
786e8c11 2483 while( optimize < jumper ) {
07be1b83 2484 mjd_nodelen += Node_Length((optimize));
a3621e74 2485 OP( optimize ) = OPTIMIZED;
07be1b83 2486 Set_Node_Offset_Length(optimize,0,0);
a3621e74
YO
2487 optimize++;
2488 }
07be1b83 2489 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
a3621e74
YO
2490 });
2491 } /* end node insert */
2e64971a
DM
2492
2493 /* Finish populating the prev field of the wordinfo array. Walk back
2494 * from each accept state until we find another accept state, and if
2495 * so, point the first word's .prev field at the second word. If the
2496 * second already has a .prev field set, stop now. This will be the
2497 * case either if we've already processed that word's accept state,
3b753521
FN
2498 * or that state had multiple words, and the overspill words were
2499 * already linked up earlier.
2e64971a
DM
2500 */
2501 {
2502 U16 word;
2503 U32 state;
2504 U16 prev;
2505
2506 for (word=1; word <= trie->wordcount; word++) {
2507 prev = 0;
2508 if (trie->wordinfo[word].prev)
2509 continue;
2510 state = trie->wordinfo[word].accept;
2511 while (state) {
2512 state = prev_states[state];
2513 if (!state)
2514 break;
2515 prev = trie->states[state].wordnum;
2516 if (prev)
2517 break;
2518 }
2519 trie->wordinfo[word].prev = prev;
2520 }
2521 Safefree(prev_states);
2522 }
2523
2524
2525 /* and now dump out the compressed format */
2526 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2527
55eed653 2528 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2b8b4781
NC
2529#ifdef DEBUGGING
2530 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2531 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2532#else
03e70be4 2533 SvREFCNT_dec_NN(revcharmap);
07be1b83 2534#endif
786e8c11
YO
2535 return trie->jump
2536 ? MADE_JUMP_TRIE
2537 : trie->startstate>1
2538 ? MADE_EXACT_TRIE
2539 : MADE_TRIE;
2540}
2541
2542STATIC void
2543S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2544{
3b753521 2545/* The Trie is constructed and compressed now so we can build a fail array if it's needed
786e8c11
YO
2546
2547 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2548 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2549 ISBN 0-201-10088-6
2550
2551 We find the fail state for each state in the trie, this state is the longest proper
3b753521
FN
2552 suffix of the current state's 'word' that is also a proper prefix of another word in our
2553 trie. State 1 represents the word '' and is thus the default fail state. This allows
786e8c11
YO
2554 the DFA not to have to restart after its tried and failed a word at a given point, it
2555 simply continues as though it had been matching the other word in the first place.
2556 Consider
2557 'abcdgu'=~/abcdefg|cdgu/
2558 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
3b753521
FN
2559 fail, which would bring us to the state representing 'd' in the second word where we would
2560 try 'g' and succeed, proceeding to match 'cdgu'.
786e8c11
YO
2561 */
2562 /* add a fail transition */
3251b653
NC
2563 const U32 trie_offset = ARG(source);
2564 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
786e8c11
YO
2565 U32 *q;
2566 const U32 ucharcount = trie->uniquecharcount;
1e2e3d02 2567 const U32 numstates = trie->statecount;
786e8c11
YO
2568 const U32 ubound = trie->lasttrans + ucharcount;
2569 U32 q_read = 0;
2570 U32 q_write = 0;
2571 U32 charid;
2572 U32 base = trie->states[ 1 ].trans.base;
2573 U32 *fail;
2574 reg_ac_data *aho;
2575 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2576 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2577
2578 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
786e8c11
YO
2579#ifndef DEBUGGING
2580 PERL_UNUSED_ARG(depth);
2581#endif
2582
2583
2584 ARG_SET( stclass, data_slot );
c944940b 2585 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
f8fc2ecf 2586 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3251b653 2587 aho->trie=trie_offset;
446bd890
NC
2588 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2589 Copy( trie->states, aho->states, numstates, reg_trie_state );
786e8c11 2590 Newxz( q, numstates, U32);
c944940b 2591 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
786e8c11
YO
2592 aho->refcount = 1;
2593 fail = aho->fail;
2594 /* initialize fail[0..1] to be 1 so that we always have
2595 a valid final fail state */
2596 fail[ 0 ] = fail[ 1 ] = 1;
2597
2598 for ( charid = 0; charid < ucharcount ; charid++ ) {
2599 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2600 if ( newstate ) {
2601 q[ q_write ] = newstate;
2602 /* set to point at the root */
2603 fail[ q[ q_write++ ] ]=1;
2604 }
2605 }
2606 while ( q_read < q_write) {
2607 const U32 cur = q[ q_read++ % numstates ];
2608 base = trie->states[ cur ].trans.base;
2609
2610 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2611 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2612 if (ch_state) {
2613 U32 fail_state = cur;
2614 U32 fail_base;
2615 do {
2616 fail_state = fail[ fail_state ];
2617 fail_base = aho->states[ fail_state ].trans.base;
2618 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2619
2620 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2621 fail[ ch_state ] = fail_state;
2622 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2623 {
2624 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2625 }
2626 q[ q_write++ % numstates] = ch_state;
2627 }
2628 }
2629 }
2630 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2631 when we fail in state 1, this allows us to use the
2632 charclass scan to find a valid start char. This is based on the principle
2633 that theres a good chance the string being searched contains lots of stuff
2634 that cant be a start char.
2635 */
2636 fail[ 0 ] = fail[ 1 ] = 0;
2637 DEBUG_TRIE_COMPILE_r({
6d99fb9b
JH
2638 PerlIO_printf(Perl_debug_log,
2639 "%*sStclass Failtable (%"UVuf" states): 0",
2640 (int)(depth * 2), "", (UV)numstates
1e2e3d02 2641 );
786e8c11
YO
2642 for( q_read=1; q_read<numstates; q_read++ ) {
2643 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2644 }
2645 PerlIO_printf(Perl_debug_log, "\n");
2646 });
2647 Safefree(q);
2648 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
a3621e74
YO
2649}
2650
786e8c11 2651
a3621e74 2652/*
5d1c421c
JH
2653 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2654 * These need to be revisited when a newer toolchain becomes available.
2655 */
2656#if defined(__sparc64__) && defined(__GNUC__)
2657# if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2658# undef SPARC64_GCC_WORKAROUND
2659# define SPARC64_GCC_WORKAROUND 1
2660# endif
2661#endif
2662
07be1b83 2663#define DEBUG_PEEP(str,scan,depth) \
b515a41d 2664 DEBUG_OPTIMISE_r({if (scan){ \
07be1b83
YO
2665 SV * const mysv=sv_newmortal(); \
2666 regnode *Next = regnext(scan); \
2667 regprop(RExC_rx, mysv, scan); \
7f69552c 2668 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
07be1b83
YO
2669 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2670 Next ? (REG_NODE_NUM(Next)) : 0 ); \
b515a41d 2671 }});
07be1b83 2672
1de06328 2673
bb914485 2674/* The below joins as many adjacent EXACTish nodes as possible into a single
0a982f06
KW
2675 * one. The regop may be changed if the node(s) contain certain sequences that
2676 * require special handling. The joining is only done if:
bb914485
KW
2677 * 1) there is room in the current conglomerated node to entirely contain the
2678 * next one.
2679 * 2) they are the exact same node type
2680 *
87b8b349 2681 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
bb914485
KW
2682 * these get optimized out
2683 *
0a982f06
KW
2684 * If a node is to match under /i (folded), the number of characters it matches
2685 * can be different than its character length if it contains a multi-character
2686 * fold. *min_subtract is set to the total delta of the input nodes.
bb914485 2687 *
a0c4c608
KW
2688 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2689 * and contains LATIN SMALL LETTER SHARP S
f758bddf 2690 *
bb914485 2691 * This is as good a place as any to discuss the design of handling these
0a982f06
KW
2692 * multi-character fold sequences. It's been wrong in Perl for a very long
2693 * time. There are three code points in Unicode whose multi-character folds
2694 * were long ago discovered to mess things up. The previous designs for
2695 * dealing with these involved assigning a special node for them. This
2696 * approach doesn't work, as evidenced by this example:
a0c4c608 2697 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
0a982f06
KW
2698 * Both these fold to "sss", but if the pattern is parsed to create a node that
2699 * would match just the \xDF, it won't be able to handle the case where a
bb914485
KW
2700 * successful match would have to cross the node's boundary. The new approach
2701 * that hopefully generally solves the problem generates an EXACTFU_SS node
2702 * that is "sss".
2703 *
0a982f06 2704 * It turns out that there are problems with all multi-character folds, and not
cb117658
KW
2705 * just these three. Now the code is general, for all such cases. The
2706 * approach taken is:
0a982f06
KW
2707 * 1) This routine examines each EXACTFish node that could contain multi-
2708 * character fold sequences. It returns in *min_subtract how much to
9d071ca8 2709 * subtract from the the actual length of the string to get a real minimum
0a982f06
KW
2710 * match length; it is 0 if there are no multi-char folds. This delta is
2711 * used by the caller to adjust the min length of the match, and the delta
2712 * between min and max, so that the optimizer doesn't reject these
2713 * possibilities based on size constraints.
cb117658 2714 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
0a982f06
KW
2715 * is used for an EXACTFU node that contains at least one "ss" sequence in
2716 * it. For non-UTF-8 patterns and strings, this is the only case where
2717 * there is a possible fold length change. That means that a regular
2718 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2719 * with length changes, and so can be processed faster. regexec.c takes
2720 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2721 * pre-folded by regcomp.c. This saves effort in regex matching.
87b8b349 2722 * However, the pre-folding isn't done for non-UTF8 patterns because the
0a982f06
KW
2723 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2724 * down by forcing the pattern into UTF8 unless necessary. Also what
2725 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2726 * possibilities for the non-UTF8 patterns are quite simple, except for
2727 * the sharp s. All the ones that don't involve a UTF-8 target string are
2728 * members of a fold-pair, and arrays are set up for all of them so that
2729 * the other member of the pair can be found quickly. Code elsewhere in
2730 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2731 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2732 * described in the next item.
098b07d5 2733 * 3) A problem remains for the sharp s in EXACTF and EXACTFA nodes when the
1ca267a5
KW
2734 * pattern isn't in UTF-8. (BTW, there cannot be an EXACTF node with a
2735 * UTF-8 pattern.) An assumption that the optimizer part of regexec.c
2736 * (probably unwittingly, in Perl_regexec_flags()) makes is that a
2737 * character in the pattern corresponds to at most a single character in
2738 * the target string. (And I do mean character, and not byte here, unlike
2739 * other parts of the documentation that have never been updated to
2740 * account for multibyte Unicode.) sharp s in EXACTF nodes can match the
2741 * two character string 'ss'; in EXACTFA nodes it can match
2742 * "\x{17F}\x{17F}". These violate the assumption, and they are the only
2743 * instances where it is violated. I'm reluctant to try to change the
2744 * assumption, as the code involved is impenetrable to me (khw), so
2745 * instead the code here punts. This routine examines (when the pattern
2746 * isn't UTF-8) EXACTF and EXACTFA nodes for the sharp s, and returns a
2747 * boolean indicating whether or not the node contains a sharp s. When it
2748 * is true, the caller sets a flag that later causes the optimizer in this
2749 * file to not set values for the floating and fixed string lengths, and
2750 * thus avoids the optimizer code in regexec.c that makes the invalid
2751 * assumption. Thus, there is no optimization based on string lengths for
2752 * non-UTF8-pattern EXACTF and EXACTFA nodes that contain the sharp s.
2753 * (The reason the assumption is wrong only in these two cases is that all
2754 * other non-UTF-8 folds are 1-1; and, for UTF-8 patterns, we pre-fold all
2755 * other folds to their expanded versions. We can't prefold sharp s to
2756 * 'ss' in EXACTF nodes because we don't know at compile time if it
2757 * actually matches 'ss' or not. It will match iff the target string is
2758 * in UTF-8, unlike the EXACTFU nodes, where it always matches; and
2759 * EXACTFA and EXACTFL where it never does. In an EXACTFA node in a UTF-8
2760 * pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the problem;
2761 * but in a non-UTF8 pattern, folding it to that above-Latin1 string would
2762 * require the pattern to be forced into UTF-8, the overhead of which we
2763 * want to avoid.)
098b07d5
KW
2764 *
2765 * Similarly, the code that generates tries doesn't currently handle
2766 * not-already-folded multi-char folds, and it looks like a pain to change
2767 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
2768 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
2769 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
2770 * using /iaa matching will be doing so almost entirely with ASCII
2771 * strings, so this should rarely be encountered in practice */
1de06328 2772
9d071ca8 2773#define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
07be1b83 2774 if (PL_regkind[OP(scan)] == EXACT) \
9d071ca8 2775 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
07be1b83 2776
be8e71aa 2777STATIC U32
9d071ca8 2778S_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
2779 /* Merge several consecutive EXACTish nodes into one. */
2780 regnode *n = regnext(scan);
2781 U32 stringok = 1;
2782 regnode *next = scan + NODE_SZ_STR(scan);
2783 U32 merged = 0;
2784 U32 stopnow = 0;
2785#ifdef DEBUGGING
2786 regnode *stop = scan;
72f13be8 2787 GET_RE_DEBUG_FLAGS_DECL;
f9049ba1 2788#else
d47053eb
RGS
2789 PERL_UNUSED_ARG(depth);
2790#endif
7918f24d
NC
2791
2792 PERL_ARGS_ASSERT_JOIN_EXACT;
d47053eb 2793#ifndef EXPERIMENTAL_INPLACESCAN
f9049ba1
SP
2794 PERL_UNUSED_ARG(flags);
2795 PERL_UNUSED_ARG(val);
07be1b83 2796#endif
07be1b83 2797 DEBUG_PEEP("join",scan,depth);
bb914485 2798
3f410cf6
KW
2799 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2800 * EXACT ones that are mergeable to the current one. */
2801 while (n
2802 && (PL_regkind[OP(n)] == NOTHING
2803 || (stringok && OP(n) == OP(scan)))
07be1b83 2804 && NEXT_OFF(n)
3f410cf6
KW
2805 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2806 {
07be1b83
YO
2807
2808 if (OP(n) == TAIL || n > next)
2809 stringok = 0;
2810 if (PL_regkind[OP(n)] == NOTHING) {
07be1b83
YO
2811 DEBUG_PEEP("skip:",n,depth);
2812 NEXT_OFF(scan) += NEXT_OFF(n);
2813 next = n + NODE_STEP_REGNODE;
2814#ifdef DEBUGGING
2815 if (stringok)
2816 stop = n;
2817#endif
2818 n = regnext(n);
2819 }
2820 else if (stringok) {
786e8c11 2821 const unsigned int oldl = STR_LEN(scan);
07be1b83 2822 regnode * const nnext = regnext(n);
b2230d39 2823
87b8b349
KW
2824 /* XXX I (khw) kind of doubt that this works on platforms where
2825 * U8_MAX is above 255 because of lots of other assumptions */
79a81a6e 2826 /* Don't join if the sum can't fit into a single node */
b2230d39
KW
2827 if (oldl + STR_LEN(n) > U8_MAX)
2828 break;
07be1b83
YO
2829
2830 DEBUG_PEEP("merg",n,depth);
07be1b83 2831 merged++;
b2230d39 2832
07be1b83
YO
2833 NEXT_OFF(scan) += NEXT_OFF(n);
2834 STR_LEN(scan) += STR_LEN(n);
2835 next = n + NODE_SZ_STR(n);
2836 /* Now we can overwrite *n : */
2837 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2838#ifdef DEBUGGING
2839 stop = next - 1;
2840#endif
2841 n = nnext;
2842 if (stopnow) break;
2843 }
2844
d47053eb
RGS
2845#ifdef EXPERIMENTAL_INPLACESCAN
2846 if (flags && !NEXT_OFF(n)) {
2847 DEBUG_PEEP("atch", val, depth);
2848 if (reg_off_by_arg[OP(n)]) {
2849 ARG_SET(n, val - n);
2850 }
2851 else {
2852 NEXT_OFF(n) = val - n;
2853 }
2854 stopnow = 1;
2855 }
07be1b83
YO
2856#endif
2857 }
2c2b7f86 2858
9d071ca8 2859 *min_subtract = 0;
f758bddf 2860 *has_exactf_sharp_s = FALSE;
f646642f 2861
3f410cf6
KW
2862 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2863 * can now analyze for sequences of problematic code points. (Prior to
2864 * this final joining, sequences could have been split over boundaries, and
a0c4c608
KW
2865 * hence missed). The sequences only happen in folding, hence for any
2866 * non-EXACT EXACTish node */
86d6fcad 2867 if (OP(scan) != EXACT) {
0a982f06
KW
2868 const U8 * const s0 = (U8*) STRING(scan);
2869 const U8 * s = s0;
2870 const U8 * const s_end = s0 + STR_LEN(scan);
f758bddf
KW
2871
2872 /* One pass is made over the node's string looking for all the
2873 * possibilities. to avoid some tests in the loop, there are two main
2874 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2875 * non-UTF-8 */
2876 if (UTF) {
86d6fcad 2877
0a982f06
KW
2878 /* Examine the string for a multi-character fold sequence. UTF-8
2879 * patterns have all characters pre-folded by the time this code is
2880 * executed */
2881 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2882 length sequence we are looking for is 2 */
86d6fcad 2883 {
0a982f06
KW
2884 int count = 0;
2885 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2886 if (! len) { /* Not a multi-char fold: get next char */
2887 s += UTF8SKIP(s);
2888 continue;
2889 }
bb914485 2890
0a982f06 2891 /* Nodes with 'ss' require special handling, except for EXACTFL
098b07d5
KW
2892 * and EXACTFA-ish for which there is no multi-char fold to
2893 * this */
0a982f06 2894 if (len == 2 && *s == 's' && *(s+1) == 's'
098b07d5
KW
2895 && OP(scan) != EXACTFL
2896 && OP(scan) != EXACTFA
2897 && OP(scan) != EXACTFA_NO_TRIE)
0a982f06
KW
2898 {
2899 count = 2;
2900 OP(scan) = EXACTFU_SS;
2901 s += 2;
2902 }
0a982f06
KW
2903 else { /* Here is a generic multi-char fold. */
2904 const U8* multi_end = s + len;
2905
2906 /* Count how many characters in it. In the case of /l and
2907 * /aa, no folds which contain ASCII code points are
2908 * allowed, so check for those, and skip if found. (In
2909 * EXACTFL, no folds are allowed to any Latin1 code point,
2910 * not just ASCII. But there aren't any of these
2911 * currently, nor ever likely, so don't take the time to
2912 * test for them. The code that generates the
2913 * is_MULTI_foo() macros croaks should one actually get put
2914 * into Unicode .) */
098b07d5
KW
2915 if (OP(scan) != EXACTFL
2916 && OP(scan) != EXACTFA
2917 && OP(scan) != EXACTFA_NO_TRIE)
2918 {
0a982f06
KW
2919 count = utf8_length(s, multi_end);
2920 s = multi_end;
2921 }
2922 else {
2923 while (s < multi_end) {
2924 if (isASCII(*s)) {
2925 s++;
2926 goto next_iteration;
2927 }
2928 else {
2929 s += UTF8SKIP(s);
2930 }
2931 count++;
2932 }
2933 }
2934 }
f758bddf 2935
0a982f06
KW
2936 /* The delta is how long the sequence is minus 1 (1 is how long
2937 * the character that folds to the sequence is) */
2938 *min_subtract += count - 1;
2939 next_iteration: ;
bb914485
KW
2940 }
2941 }
1ca267a5
KW
2942 else if (OP(scan) == EXACTFA) {
2943
2944 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
2945 * fold to the ASCII range (and there are no existing ones in the
2946 * upper latin1 range). But, as outlined in the comments preceding
098b07d5
KW
2947 * this function, we need to flag any occurrences of the sharp s.
2948 * This character forbids trie formation (because of added
2949 * complexity) */
1ca267a5
KW
2950 while (s < s_end) {
2951 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
098b07d5 2952 OP(scan) = EXACTFA_NO_TRIE;
1ca267a5
KW
2953 *has_exactf_sharp_s = TRUE;
2954 break;
2955 }
2956 s++;
2957 continue;
2958 }
2959 }
2960 else if (OP(scan) != EXACTFL) {
2961
2962 /* Non-UTF-8 pattern, not EXACTFA nor EXACTFL node. Look for the
2963 * multi-char folds that are all Latin1. (This code knows that
2964 * there are no current multi-char folds possible with EXACTFL,
2965 * relying on fold_grind.t to catch any errors if the very unlikely
2966 * event happens that some get added in future Unicode versions.)
2967 * As explained in the comments preceding this function, we look
2968 * also for the sharp s in EXACTF nodes; it can be in the final
0a982f06
KW
2969 * position. Otherwise we can stop looking 1 byte earlier because
2970 * have to find at least two characters for a multi-fold */
f758bddf
KW
2971 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2972
0a982f06 2973 while (s < upper) {
40b1ba4f 2974 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
0a982f06
KW
2975 if (! len) { /* Not a multi-char fold. */
2976 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2977 {
2978 *has_exactf_sharp_s = TRUE;
2979 }
2980 s++;
2981 continue;
2982 }
2983
2984 if (len == 2
c02c3054
KW
2985 && isARG2_lower_or_UPPER_ARG1('s', *s)
2986 && isARG2_lower_or_UPPER_ARG1('s', *(s+1)))
0a982f06
KW
2987 {
2988
2989 /* EXACTF nodes need to know that the minimum length
2990 * changed so that a sharp s in the string can match this
2991 * ss in the pattern, but they remain EXACTF nodes, as they
2992 * won't match this unless the target string is is UTF-8,
2993 * which we don't know until runtime */
2994 if (OP(scan) != EXACTF) {
2995 OP(scan) = EXACTFU_SS;
2996 }
86d6fcad 2997 }
0a982f06
KW
2998
2999 *min_subtract += len - 1;
3000 s += len;
86d6fcad
KW
3001 }
3002 }
07be1b83 3003 }
3f410cf6 3004
07be1b83 3005#ifdef DEBUGGING
bb789b09
DM
3006 /* Allow dumping but overwriting the collection of skipped
3007 * ops and/or strings with fake optimized ops */
07be1b83
YO
3008 n = scan + NODE_SZ_STR(scan);
3009 while (n <= stop) {
bb789b09
DM
3010 OP(n) = OPTIMIZED;
3011 FLAGS(n) = 0;
3012 NEXT_OFF(n) = 0;
07be1b83
YO
3013 n++;
3014 }
3015#endif
3016 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3017 return stopnow;
3018}
3019
486ec47a 3020/* REx optimizer. Converts nodes into quicker variants "in place".
653099ff
GS
3021 Finds fixed substrings. */
3022
a0288114 3023/* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
c277df42
IZ
3024 to the position after last scanned or to NULL. */
3025
40d049e4
YO
3026#define INIT_AND_WITHP \
3027 assert(!and_withp); \
3028 Newx(and_withp,1,struct regnode_charclass_class); \
3029 SAVEFREEPV(and_withp)
07be1b83 3030
b515a41d 3031/* this is a chain of data about sub patterns we are processing that
486ec47a 3032 need to be handled separately/specially in study_chunk. Its so
b515a41d
YO
3033 we can simulate recursion without losing state. */
3034struct scan_frame;
3035typedef struct scan_frame {
3036 regnode *last; /* last node to process in this frame */
3037 regnode *next; /* next node to process when last is reached */
3038 struct scan_frame *prev; /*previous frame*/
3039 I32 stop; /* what stopparen do we use */
3040} scan_frame;
3041
304ee84b
YO
3042
3043#define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
3044
49f55535 3045STATIC SSize_t
40d049e4 3046S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
ea3daa5d 3047 SSize_t *minlenp, SSize_t *deltap,
40d049e4
YO
3048 regnode *last,
3049 scan_data_t *data,
3050 I32 stopparen,
3051 U8* recursed,
3052 struct regnode_charclass_class *and_withp,
3053 U32 flags, U32 depth)
c277df42
IZ
3054 /* scanp: Start here (read-write). */
3055 /* deltap: Write maxlen-minlen here. */
3056 /* last: Stop before this one. */
40d049e4
YO
3057 /* data: string data about the pattern */
3058 /* stopparen: treat close N as END */
3059 /* recursed: which subroutines have we recursed into */
3060 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
c277df42 3061{
97aff369 3062 dVAR;
ea3daa5d
FC
3063 /* There must be at least this number of characters to match */
3064 SSize_t min = 0;
2d608413 3065 I32 pars = 0, code;
c277df42 3066 regnode *scan = *scanp, *next;
ea3daa5d 3067 SSize_t delta = 0;
c277df42 3068 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
aca2d497 3069 int is_inf_internal = 0; /* The studied chunk is infinite */
c277df42
IZ
3070 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3071 scan_data_t data_fake;
a3621e74 3072 SV *re_trie_maxbuff = NULL;
786e8c11 3073 regnode *first_non_open = scan;
ea3daa5d 3074 SSize_t stopmin = SSize_t_MAX;
8aa23a47 3075 scan_frame *frame = NULL;
a3621e74 3076 GET_RE_DEBUG_FLAGS_DECL;
8aa23a47 3077
7918f24d
NC
3078 PERL_ARGS_ASSERT_STUDY_CHUNK;
3079
13a24bad 3080#ifdef DEBUGGING
40d049e4 3081 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
13a24bad 3082#endif
40d049e4 3083
786e8c11 3084 if ( depth == 0 ) {
40d049e4 3085 while (first_non_open && OP(first_non_open) == OPEN)
786e8c11
YO
3086 first_non_open=regnext(first_non_open);
3087 }
3088
b81d288d 3089
8aa23a47
YO
3090 fake_study_recurse:
3091 while ( scan && OP(scan) != END && scan < last ){
2d608413
KW
3092 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3093 node length to get a real minimum (because
3094 the folded version may be shorter) */
f758bddf 3095 bool has_exactf_sharp_s = FALSE;
8aa23a47 3096 /* Peephole optimizer: */
304ee84b 3097 DEBUG_STUDYDATA("Peep:", data,depth);
8aa23a47 3098 DEBUG_PEEP("Peep",scan,depth);
a0c4c608
KW
3099
3100 /* Its not clear to khw or hv why this is done here, and not in the
3101 * clauses that deal with EXACT nodes. khw's guess is that it's
3102 * because of a previous design */
9d071ca8 3103 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
8aa23a47
YO
3104
3105 /* Follow the next-chain of the current node and optimize
3106 away all the NOTHINGs from it. */
3107 if (OP(scan) != CURLYX) {
3108 const int max = (reg_off_by_arg[OP(scan)]
3109 ? I32_MAX
3110 /* I32 may be smaller than U16 on CRAYs! */
3111 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3112 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3113 int noff;
3114 regnode *n = scan;
686b73d4 3115
8aa23a47
YO
3116 /* Skip NOTHING and LONGJMP. */
3117 while ((n = regnext(n))
3118 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3119 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3120 && off + noff < max)
3121 off += noff;
3122 if (reg_off_by_arg[OP(scan)])
3123 ARG(scan) = off;
3124 else
3125 NEXT_OFF(scan) = off;
3126 }
a3621e74 3127
c277df42 3128
8aa23a47
YO
3129
3130 /* The principal pseudo-switch. Cannot be a switch, since we
3131 look into several different things. */
3132 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3133 || OP(scan) == IFTHEN) {
3134 next = regnext(scan);
3135 code = OP(scan);
3136 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
686b73d4 3137
8aa23a47
YO
3138 if (OP(next) == code || code == IFTHEN) {
3139 /* NOTE - There is similar code to this block below for handling
3140 TRIE nodes on a re-study. If you change stuff here check there
3141 too. */
49f55535 3142 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
8aa23a47
YO
3143 struct regnode_charclass_class accum;
3144 regnode * const startbranch=scan;
686b73d4 3145
8aa23a47 3146 if (flags & SCF_DO_SUBSTR)
304ee84b 3147 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
8aa23a47 3148 if (flags & SCF_DO_STCLASS)
e755fd73 3149 cl_init_zero(pRExC_state, &accum);
8aa23a47
YO
3150
3151 while (OP(scan) == code) {
49f55535
FC
3152 SSize_t deltanext, minnext, fake;
3153 I32 f = 0;
8aa23a47
YO
3154 struct regnode_charclass_class this_class;
3155
3156 num++;
3157 data_fake.flags = 0;
3158 if (data) {
3159 data_fake.whilem_c = data->whilem_c;
3160 data_fake.last_closep = data->last_closep;
3161 }
3162 else
3163 data_fake.last_closep = &fake;
58e23c8d
YO
3164
3165 data_fake.pos_delta = delta;
8aa23a47
YO
3166 next = regnext(scan);
3167 scan = NEXTOPER(scan);
3168 if (code != BRANCH)
c277df42 3169 scan = NEXTOPER(scan);
8aa23a47 3170 if (flags & SCF_DO_STCLASS) {
e755fd73 3171 cl_init(pRExC_state, &this_class);
8aa23a47
YO
3172 data_fake.start_class = &this_class;
3173 f = SCF_DO_STCLASS_AND;
58e23c8d 3174 }
8aa23a47
YO
3175 if (flags & SCF_WHILEM_VISITED_POS)
3176 f |= SCF_WHILEM_VISITED_POS;
3177
3178 /* we suppose the run is continuous, last=next...*/
3179 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3180 next, &data_fake,
3181 stopparen, recursed, NULL, f,depth+1);
3182 if (min1 > minnext)
3183 min1 = minnext;
ea3daa5d 3184 if (deltanext == SSize_t_MAX) {
8aa23a47 3185 is_inf = is_inf_internal = 1;
ea3daa5d 3186 max1 = SSize_t_MAX;
9b139d09
GG
3187 } else if (max1 < minnext + deltanext)
3188 max1 = minnext + deltanext;
8aa23a47
YO
3189 scan = next;
3190 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3191 pars++;
3192 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3193 if ( stopmin > minnext)
3194 stopmin = min + min1;
3195 flags &= ~SCF_DO_SUBSTR;
3196 if (data)
3197 data->flags |= SCF_SEEN_ACCEPT;
3198 }
3199 if (data) {
3200 if (data_fake.flags & SF_HAS_EVAL)
3201 data->flags |= SF_HAS_EVAL;
3202 data->whilem_c = data_fake.whilem_c;
3dab1dad 3203 }
8aa23a47 3204 if (flags & SCF_DO_STCLASS)
3fffb88a 3205 cl_or(pRExC_state, &accum, &this_class);
8aa23a47
YO
3206 }
3207 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3208 min1 = 0;
3209 if (flags & SCF_DO_SUBSTR) {
3210 data->pos_min += min1;
ea3daa5d
FC
3211 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3212 data->pos_delta = SSize_t_MAX;
9b139d09
GG
3213 else
3214 data->pos_delta += max1 - min1;
8aa23a47
YO
3215 if (max1 != min1 || is_inf)
3216 data->longest = &(data->longest_float);
3217 }
3218 min += min1;
ea3daa5d
FC
3219 if (delta == SSize_t_MAX
3220 || SSize_t_MAX - delta - (max1 - min1) < 0)
3221 delta = SSize_t_MAX;
9b139d09
GG
3222 else
3223 delta += max1 - min1;
8aa23a47 3224 if (flags & SCF_DO_STCLASS_OR) {
3fffb88a 3225 cl_or(pRExC_state, data->start_class, &accum);
8aa23a47
YO
3226 if (min1) {
3227 cl_and(data->start_class, and_withp);
3228 flags &= ~SCF_DO_STCLASS;
653099ff 3229 }
8aa23a47
YO
3230 }
3231 else if (flags & SCF_DO_STCLASS_AND) {
3232 if (min1) {
3233 cl_and(data->start_class, &accum);
3234 flags &= ~SCF_DO_STCLASS;
de0c8cb8 3235 }
8aa23a47
YO
3236 else {
3237 /* Switch to OR mode: cache the old value of
3238 * data->start_class */
3239 INIT_AND_WITHP;
3240 StructCopy(data->start_class, and_withp,
3241 struct regnode_charclass_class);
3242 flags &= ~SCF_DO_STCLASS_AND;
3243 StructCopy(&accum, data->start_class,
3244 struct regnode_charclass_class);
3245 flags |= SCF_DO_STCLASS_OR;
899d20b9 3246 SET_SSC_EOS(data->start_class);
de0c8cb8 3247 }
8aa23a47 3248 }
a3621e74 3249
8aa23a47
YO
3250 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3251 /* demq.
a3621e74 3252
8aa23a47
YO
3253 Assuming this was/is a branch we are dealing with: 'scan' now
3254 points at the item that follows the branch sequence, whatever
3255 it is. We now start at the beginning of the sequence and look
3256 for subsequences of
a3621e74 3257
8aa23a47
YO
3258 BRANCH->EXACT=>x1
3259 BRANCH->EXACT=>x2
3260 tail
a3621e74 3261
8aa23a47 3262 which would be constructed from a pattern like /A|LIST|OF|WORDS/
a3621e74 3263
486ec47a 3264 If we can find such a subsequence we need to turn the first
8aa23a47
YO
3265 element into a trie and then add the subsequent branch exact
3266 strings to the trie.
a3621e74 3267
8aa23a47 3268 We have two cases
a3621e74 3269
3b753521 3270 1. patterns where the whole set of branches can be converted.
a3621e74 3271
8aa23a47 3272 2. patterns where only a subset can be converted.
a3621e74 3273
8aa23a47
YO
3274 In case 1 we can replace the whole set with a single regop
3275 for the trie. In case 2 we need to keep the start and end
3b753521 3276 branches so
a3621e74 3277
8aa23a47
YO
3278 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3279 becomes BRANCH TRIE; BRANCH X;
786e8c11 3280
8aa23a47
YO
3281 There is an additional case, that being where there is a
3282 common prefix, which gets split out into an EXACT like node
3283 preceding the TRIE node.
a3621e74 3284
8aa23a47
YO
3285 If x(1..n)==tail then we can do a simple trie, if not we make
3286 a "jump" trie, such that when we match the appropriate word
486ec47a 3287 we "jump" to the appropriate tail node. Essentially we turn
8aa23a47 3288 a nested if into a case structure of sorts.
b515a41d 3289
8aa23a47 3290 */
686b73d4 3291
8aa23a47
YO
3292 int made=0;
3293 if (!re_trie_maxbuff) {
3294 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3295 if (!SvIOK(re_trie_maxbuff))
3296 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3297 }
3298 if ( SvIV(re_trie_maxbuff)>=0 ) {
3299 regnode *cur;
3300 regnode *first = (regnode *)NULL;
3301 regnode *last = (regnode *)NULL;
3302 regnode *tail = scan;
fab2782b 3303 U8 trietype = 0;
8aa23a47 3304 U32 count=0;
a3621e74
YO
3305
3306#ifdef DEBUGGING
8aa23a47 3307 SV * const mysv = sv_newmortal(); /* for dumping */
a3621e74 3308#endif
8aa23a47
YO
3309 /* var tail is used because there may be a TAIL
3310 regop in the way. Ie, the exacts will point to the
3311 thing following the TAIL, but the last branch will
3312 point at the TAIL. So we advance tail. If we
3313 have nested (?:) we may have to move through several
3314 tails.
3315 */
3316
3317 while ( OP( tail ) == TAIL ) {
3318 /* this is the TAIL generated by (?:) */
3319 tail = regnext( tail );
3320 }
a3621e74 3321
8aa23a47 3322
df826430 3323 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3324 regprop(RExC_rx, mysv, tail );
3325 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3326 (int)depth * 2 + 2, "",
3327 "Looking for TRIE'able sequences. Tail node is: ",
3328 SvPV_nolen_const( mysv )
3329 );
3330 });
3331
3332 /*
3333
fab2782b
YO
3334 Step through the branches
3335 cur represents each branch,
3336 noper is the first thing to be matched as part of that branch
3337 noper_next is the regnext() of that node.
3338
3339 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3340 via a "jump trie" but we also support building with NOJUMPTRIE,
3341 which restricts the trie logic to structures like /FOO|BAR/.
3342
3343 If noper is a trieable nodetype then the branch is a possible optimization
3344 target. If we are building under NOJUMPTRIE then we require that noper_next
3345 is the same as scan (our current position in the regex program).
3346
3347 Once we have two or more consecutive such branches we can create a
3348 trie of the EXACT's contents and stitch it in place into the program.
3349
3350 If the sequence represents all of the branches in the alternation we
3351 replace the entire thing with a single TRIE node.
3352
3353 Otherwise when it is a subsequence we need to stitch it in place and
3354 replace only the relevant branches. This means the first branch has
3355 to remain as it is used by the alternation logic, and its next pointer,
3356 and needs to be repointed at the item on the branch chain following
3357 the last branch we have optimized away.
3358
3359 This could be either a BRANCH, in which case the subsequence is internal,
3360 or it could be the item following the branch sequence in which case the
3361 subsequence is at the end (which does not necessarily mean the first node
3362 is the start of the alternation).
3363
3364 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3365
3366 optype | trietype
3367 ----------------+-----------
3368 NOTHING | NOTHING
3369 EXACT | EXACT
3370 EXACTFU | EXACTFU
3371 EXACTFU_SS | EXACTFU
098b07d5 3372 EXACTFA | EXACTFA
fab2782b 3373
8aa23a47
YO
3374
3375 */
fab2782b
YO
3376#define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3377 ( EXACT == (X) ) ? EXACT : \
cb117658 3378 ( EXACTFU == (X) || EXACTFU_SS == (X) ) ? EXACTFU : \
098b07d5 3379 ( EXACTFA == (X) ) ? EXACTFA : \
fab2782b 3380 0 )
8aa23a47
YO
3381
3382 /* dont use tail as the end marker for this traverse */
3383 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3384 regnode * const noper = NEXTOPER( cur );
fab2782b
YO
3385 U8 noper_type = OP( noper );
3386 U8 noper_trietype = TRIE_TYPE( noper_type );
b515a41d 3387#if defined(DEBUGGING) || defined(NOJUMPTRIE)
8aa23a47 3388 regnode * const noper_next = regnext( noper );
df826430
YO
3389 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3390 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
b515a41d
YO
3391#endif
3392
df826430 3393 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3394 regprop(RExC_rx, mysv, cur);
3395 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3396 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3397
3398 regprop(RExC_rx, mysv, noper);
3399 PerlIO_printf( Perl_debug_log, " -> %s",
3400 SvPV_nolen_const(mysv));
3401
3402 if ( noper_next ) {
3403 regprop(RExC_rx, mysv, noper_next );
3404 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3405 SvPV_nolen_const(mysv));
3406 }
df826430
YO
3407 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3408 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3409 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3410 );
8aa23a47 3411 });
fab2782b
YO
3412
3413 /* Is noper a trieable nodetype that can be merged with the
3414 * current trie (if there is one)? */
3415 if ( noper_trietype
3416 &&
3417 (
df826430
YO
3418 ( noper_trietype == NOTHING)
3419 || ( trietype == NOTHING )
a40630bf 3420 || ( trietype == noper_trietype )
fab2782b 3421 )
786e8c11 3422#ifdef NOJUMPTRIE
8aa23a47 3423 && noper_next == tail
786e8c11 3424#endif
8aa23a47
YO
3425 && count < U16_MAX)
3426 {
fab2782b
YO
3427 /* Handle mergable triable node
3428 * Either we are the first node in a new trieable sequence,
3429 * in which case we do some bookkeeping, otherwise we update
3430 * the end pointer. */
fab2782b 3431 if ( !first ) {
3b6759a6 3432 first = cur;
df826430
YO
3433 if ( noper_trietype == NOTHING ) {
3434#if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3435 regnode * const noper_next = regnext( noper );
3b6759a6 3436 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
df826430
YO
3437 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3438#endif
3439
190c1910 3440 if ( noper_next_trietype ) {
df826430 3441 trietype = noper_next_trietype;
190c1910
YO
3442 } else if (noper_next_type) {
3443 /* a NOTHING regop is 1 regop wide. We need at least two
3444 * for a trie so we can't merge this in */
3445 first = NULL;
3446 }
3447 } else {
3448 trietype = noper_trietype;
3b6759a6 3449 }
8aa23a47 3450 } else {
fab2782b
YO
3451 if ( trietype == NOTHING )
3452 trietype = noper_trietype;
8aa23a47
YO
3453 last = cur;
3454 }
df826430
YO
3455 if (first)
3456 count++;
fab2782b
YO
3457 } /* end handle mergable triable node */
3458 else {
3459 /* handle unmergable node -
3460 * noper may either be a triable node which can not be tried
3461 * together with the current trie, or a non triable node */
729aaeb5
YO
3462 if ( last ) {
3463 /* If last is set and trietype is not NOTHING then we have found
3464 * at least two triable branch sequences in a row of a similar
3465 * trietype so we can turn them into a trie. If/when we
3466 * allow NOTHING to start a trie sequence this condition will be
3467 * required, and it isn't expensive so we leave it in for now. */
e6351b37 3468 if ( trietype && trietype != NOTHING )
729aaeb5
YO
3469 make_trie( pRExC_state,
3470 startbranch, first, cur, tail, count,
3471 trietype, depth+1 );
fab2782b 3472 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
8aa23a47 3473 }
fab2782b 3474 if ( noper_trietype
786e8c11 3475#ifdef NOJUMPTRIE
8aa23a47 3476 && noper_next == tail
786e8c11 3477#endif
8aa23a47 3478 ){
fab2782b 3479 /* noper is triable, so we can start a new trie sequence */
8aa23a47
YO
3480 count = 1;
3481 first = cur;
fab2782b
YO
3482 trietype = noper_trietype;
3483 } else if (first) {
3484 /* if we already saw a first but the current node is not triable then we have
3485 * to reset the first information. */
8aa23a47
YO
3486 count = 0;
3487 first = NULL;
fab2782b 3488 trietype = 0;
8aa23a47 3489 }
fab2782b
YO
3490 } /* end handle unmergable node */
3491 } /* loop over branches */
df826430 3492 DEBUG_TRIE_COMPILE_r({
8aa23a47
YO
3493 regprop(RExC_rx, mysv, cur);
3494 PerlIO_printf( Perl_debug_log,
3495 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3496 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3497
3498 });
e6351b37 3499 if ( last && trietype ) {
3b6759a6
YO
3500 if ( trietype != NOTHING ) {
3501 /* the last branch of the sequence was part of a trie,
3502 * so we have to construct it here outside of the loop
3503 */
3504 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
686b73d4 3505#ifdef TRIE_STUDY_OPT
3b6759a6
YO
3506 if ( ((made == MADE_EXACT_TRIE &&
3507 startbranch == first)
3508 || ( first_non_open == first )) &&
3509 depth==0 ) {
3510 flags |= SCF_TRIE_RESTUDY;
3511 if ( startbranch == first
3512 && scan == tail )
3513 {
3514 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3515 }
8aa23a47 3516 }
3dab1dad 3517#endif
3b6759a6
YO
3518 } else {
3519 /* at this point we know whatever we have is a NOTHING sequence/branch
3520 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3521 */
3522 if ( startbranch == first ) {
3523 regnode *opt;
3524 /* the entire thing is a NOTHING sequence, something like this:
3525 * (?:|) So we can turn it into a plain NOTHING op. */
3526 DEBUG_TRIE_COMPILE_r({
3527 regprop(RExC_rx, mysv, cur);
3528 PerlIO_printf( Perl_debug_log,
3529 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3530 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3531
3532 });
3533 OP(startbranch)= NOTHING;
3534 NEXT_OFF(startbranch)= tail - startbranch;
3535 for ( opt= startbranch + 1; opt < tail ; opt++ )
3536 OP(opt)= OPTIMIZED;
3537 }
3538 }
fab2782b
YO
3539 } /* end if ( last) */
3540 } /* TRIE_MAXBUF is non zero */
8aa23a47
YO
3541
3542 } /* do trie */
3543
653099ff 3544 }
8aa23a47
YO
3545 else if ( code == BRANCHJ ) { /* single branch is optimized. */