This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
Fix skipping char in (*sr:...) defn
[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"
afda64e8 84EXTERN_C const struct regexp_engine my_reg_engine;
54df2634
NC
85#else
86# include "regcomp.h"
87#endif
a687059c 88
f7e03a10 89#include "dquote_inline.h"
b992490d 90#include "invlist_inline.h"
1b0f46bf 91#include "unicode_constants.h"
04e98a4d 92
538e84ed
KW
93#define HAS_NONLATIN1_FOLD_CLOSURE(i) \
94 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
f12c0118
KW
95#define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
96 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
26faadbd 97#define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
2c61f163 98#define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
94dc5c2d 99
a687059c
LW
100#ifndef STATIC
101#define STATIC static
102#endif
103
3f910e62
YO
104/* this is a chain of data about sub patterns we are processing that
105 need to be handled separately/specially in study_chunk. Its so
106 we can simulate recursion without losing state. */
107struct scan_frame;
108typedef struct scan_frame {
109 regnode *last_regnode; /* last node to process in this frame */
110 regnode *next_regnode; /* next node to process when last is reached */
111 U32 prev_recursed_depth;
112 I32 stopparen; /* what stopparen do we use */
3f910e62
YO
113
114 struct scan_frame *this_prev_frame; /* this previous frame */
115 struct scan_frame *prev_frame; /* previous frame */
116 struct scan_frame *next_frame; /* next frame */
117} scan_frame;
118
f4ae5a27
KW
119/* Certain characters are output as a sequence with the first being a
120 * backslash. */
eb7486de 121#define isBACKSLASHED_PUNCT(c) strchr("-[]\\^", c)
f4ae5a27
KW
122
123
09b2b2e6 124struct RExC_state_t {
514a91f1
DM
125 U32 flags; /* RXf_* are we folding, multilining? */
126 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
830247a4 127 char *precomp; /* uncompiled string. */
711b303b 128 char *precomp_end; /* pointer to end of uncompiled string. */
288b8c02 129 REGEXP *rx_sv; /* The SV that is the regexp. */
f8fc2ecf 130 regexp *rx; /* perl core regexp structure */
538e84ed
KW
131 regexp_internal *rxi; /* internal data for regexp object
132 pprivate field */
fac92740 133 char *start; /* Start of input for compile */
830247a4
IZ
134 char *end; /* End of input for compile */
135 char *parse; /* Input-scan pointer. */
285b5ca0
KW
136 char *adjusted_start; /* 'start', adjusted. See code use */
137 STRLEN precomp_adj; /* an offset beyond precomp. See code use */
ea3daa5d 138 SSize_t whilem_seen; /* number of WHILEM in this expr */
fac92740 139 regnode *emit_start; /* Start of emitted-code area */
538e84ed
KW
140 regnode *emit_bound; /* First regnode outside of the
141 allocated space */
f7c7e32a
DM
142 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
143 implies compiling, so don't emit */
9a81a976
KW
144 regnode_ssc emit_dummy; /* placeholder for emit to point to;
145 large enough for the largest
146 non-EXACTish node, so can use it as
147 scratch in pass1 */
830247a4
IZ
148 I32 naughty; /* How bad is this pattern? */
149 I32 sawback; /* Did we see \1, ...? */
150 U32 seen;
ea3daa5d 151 SSize_t size; /* Code size. */
e7f023ff 152 I32 npar; /* Capture buffer count, (OPEN) plus
538e84ed
KW
153 one. ("par" 0 is the whole
154 pattern)*/
155 I32 nestroot; /* root parens we are in - used by
156 accept */
830247a4
IZ
157 I32 extralen;
158 I32 seen_zerolen;
40d049e4
YO
159 regnode **open_parens; /* pointers to open parens */
160 regnode **close_parens; /* pointers to close parens */
d5a00e4a 161 regnode *end_op; /* END node in program */
02daf0ab
YO
162 I32 utf8; /* whether the pattern is utf8 or not */
163 I32 orig_utf8; /* whether the pattern was originally in utf8 */
164 /* XXX use this for future optimisation of case
165 * where pattern must be upgraded to utf8. */
e40e74fe
KW
166 I32 uni_semantics; /* If a d charset modifier should use unicode
167 rules, even if the pattern is not in
168 utf8 */
81714fb9 169 HV *paren_names; /* Paren names */
538e84ed 170
40d049e4 171 regnode **recurse; /* Recurse regops */
d5a00e4a 172 I32 recurse_count; /* Number of recurse regops we have generated */
4286711a 173 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
538e84ed 174 through */
09a65838 175 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
b57e4118 176 I32 in_lookbehind;
4624b182 177 I32 contains_locale;
bb3f3ed2 178 I32 override_recoding;
b6d67071
KW
179#ifdef EBCDIC
180 I32 recode_x_to_native;
181#endif
9d53c457 182 I32 in_multi_char_class;
1acab4c5 183 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
68e2671b 184 within pattern */
b1603ef8 185 int code_index; /* next code_blocks[] slot */
ee273784 186 SSize_t maxlen; /* mininum possible number of chars in string to match */
3f910e62
YO
187 scan_frame *frame_head;
188 scan_frame *frame_last;
189 U32 frame_count;
7eec73eb 190 AV *warn_text;
dc6d7f5c 191#ifdef ADD_TO_REGEXEC
830247a4
IZ
192 char *starttry; /* -Dr: where regtry was called. */
193#define RExC_starttry (pRExC_state->starttry)
194#endif
d24ca0c5 195 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 196#ifdef DEBUGGING
be8e71aa 197 const char *lastparse;
3dab1dad 198 I32 lastnum;
1f1031fe 199 AV *paren_name_list; /* idx -> name */
d9a72fcc 200 U32 study_chunk_recursed_count;
c9f0d54c
YO
201 SV *mysv1;
202 SV *mysv2;
3dab1dad
YO
203#define RExC_lastparse (pRExC_state->lastparse)
204#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 205#define RExC_paren_name_list (pRExC_state->paren_name_list)
d9a72fcc 206#define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
c9f0d54c
YO
207#define RExC_mysv (pRExC_state->mysv1)
208#define RExC_mysv1 (pRExC_state->mysv1)
209#define RExC_mysv2 (pRExC_state->mysv2)
210
3dab1dad 211#endif
512c0f5a 212 bool seen_unfolded_sharp_s;
911bd04e 213 bool strict;
da7cf1cc 214 bool study_started;
034602eb 215 bool in_script_run;
09b2b2e6 216};
830247a4 217
e2509266 218#define RExC_flags (pRExC_state->flags)
514a91f1 219#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 220#define RExC_precomp (pRExC_state->precomp)
285b5ca0
KW
221#define RExC_precomp_adj (pRExC_state->precomp_adj)
222#define RExC_adjusted_start (pRExC_state->adjusted_start)
711b303b 223#define RExC_precomp_end (pRExC_state->precomp_end)
288b8c02 224#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 225#define RExC_rx (pRExC_state->rx)
f8fc2ecf 226#define RExC_rxi (pRExC_state->rxi)
fac92740 227#define RExC_start (pRExC_state->start)
830247a4
IZ
228#define RExC_end (pRExC_state->end)
229#define RExC_parse (pRExC_state->parse)
230#define RExC_whilem_seen (pRExC_state->whilem_seen)
512c0f5a
KW
231
232/* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
233 * EXACTF node, hence was parsed under /di rules. If later in the parse,
234 * something forces the pattern into using /ui rules, the sharp s should be
235 * folded into the sequence 'ss', which takes up more space than previously
236 * calculated. This means that the sizing pass needs to be restarted. (The
237 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
238 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
239 * so there is no need to resize [perl #125990]. */
240#define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
241
7122b237 242#ifdef RE_TRACK_PATTERN_OFFSETS
538e84ed
KW
243#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
244 others */
7122b237 245#endif
830247a4 246#define RExC_emit (pRExC_state->emit)
f7c7e32a 247#define RExC_emit_dummy (pRExC_state->emit_dummy)
fac92740 248#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 249#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
250#define RExC_sawback (pRExC_state->sawback)
251#define RExC_seen (pRExC_state->seen)
252#define RExC_size (pRExC_state->size)
ee273784 253#define RExC_maxlen (pRExC_state->maxlen)
830247a4 254#define RExC_npar (pRExC_state->npar)
e2e6a0f1 255#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
256#define RExC_extralen (pRExC_state->extralen)
257#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 258#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 259#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 260#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
261#define RExC_open_parens (pRExC_state->open_parens)
262#define RExC_close_parens (pRExC_state->close_parens)
5bd2d46e 263#define RExC_end_op (pRExC_state->end_op)
81714fb9 264#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
265#define RExC_recurse (pRExC_state->recurse)
266#define RExC_recurse_count (pRExC_state->recurse_count)
09a65838 267#define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
538e84ed
KW
268#define RExC_study_chunk_recursed_bytes \
269 (pRExC_state->study_chunk_recursed_bytes)
b57e4118 270#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 271#define RExC_contains_locale (pRExC_state->contains_locale)
b6d67071
KW
272#ifdef EBCDIC
273# define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
274#endif
9d53c457 275#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
3f910e62
YO
276#define RExC_frame_head (pRExC_state->frame_head)
277#define RExC_frame_last (pRExC_state->frame_last)
278#define RExC_frame_count (pRExC_state->frame_count)
67cdf558 279#define RExC_strict (pRExC_state->strict)
da7cf1cc 280#define RExC_study_started (pRExC_state->study_started)
7eec73eb 281#define RExC_warn_text (pRExC_state->warn_text)
034602eb 282#define RExC_in_script_run (pRExC_state->in_script_run)
830247a4 283
99807a43
HS
284/* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
285 * a flag to disable back-off on the fixed/floating substrings - if it's
286 * a high complexity pattern we assume the benefit of avoiding a full match
287 * is worth the cost of checking for the substrings even if they rarely help.
288 */
289#define RExC_naughty (pRExC_state->naughty)
290#define TOO_NAUGHTY (10)
291#define MARK_NAUGHTY(add) \
292 if (RExC_naughty < TOO_NAUGHTY) \
293 RExC_naughty += (add)
294#define MARK_NAUGHTY_EXP(exp, add) \
295 if (RExC_naughty < TOO_NAUGHTY) \
296 RExC_naughty += RExC_naughty / (exp) + (add)
cde0cee5 297
a687059c
LW
298#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
299#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
412f55bb 300 ((*s) == '{' && regcurly(s)))
a687059c
LW
301
302/*
303 * Flags to be passed up and down.
304 */
a687059c 305#define WORST 0 /* Worst case. */
a3b492c3 306#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 307
e64f369d 308/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
309 * character. (There needs to be a case: in the switch statement in regexec.c
310 * for any node marked SIMPLE.) Note that this is not the same thing as
311 * REGNODE_SIMPLE */
fda99bee 312#define SIMPLE 0x02
e64f369d 313#define SPSTART 0x04 /* Starts with * or + */
8d9c2815
NC
314#define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
315#define TRYAGAIN 0x10 /* Weeded out a declaration. */
b97943f7
KW
316#define RESTART_PASS1 0x20 /* Need to restart sizing pass */
317#define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
318 calcuate sizes as UTF-8 */
a687059c 319
3dab1dad
YO
320#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
321
07be1b83
YO
322/* whether trie related optimizations are enabled */
323#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
324#define TRIE_STUDY_OPT
786e8c11 325#define FULL_TRIE_STUDY
07be1b83
YO
326#define TRIE_STCLASS
327#endif
1de06328
YO
328
329
40d049e4
YO
330
331#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
332#define PBITVAL(paren) (1 << ((paren) & 7))
333#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
334#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
335#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
336
82a6ada4 337#define REQUIRE_UTF8(flagp) STMT_START { \
8d9c2815 338 if (!UTF) { \
82a6ada4
KW
339 assert(PASS1); \
340 *flagp = RESTART_PASS1|NEED_UTF8; \
8d9c2815
NC
341 return NULL; \
342 } \
82a6ada4 343 } STMT_END
40d049e4 344
512c0f5a
KW
345/* Change from /d into /u rules, and restart the parse if we've already seen
346 * something whose size would increase as a result, by setting *flagp and
347 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
142ece99 348 * we've changed to /u during the parse. */
512c0f5a
KW
349#define REQUIRE_UNI_RULES(flagp, restart_retval) \
350 STMT_START { \
351 if (DEPENDS_SEMANTICS) { \
352 assert(PASS1); \
353 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
354 RExC_uni_semantics = 1; \
355 if (RExC_seen_unfolded_sharp_s) { \
356 *flagp |= RESTART_PASS1; \
357 return restart_retval; \
358 } \
359 } \
360 } STMT_END
361
fd7fc6ae
YO
362/* Executes a return statement with the value 'X', if 'flags' contains any of
363 * 'RESTART_PASS1', 'NEED_UTF8', or 'extra'. If so, *flagp is set to those
364 * flags */
365#define RETURN_X_ON_RESTART_OR_FLAGS(X, flags, flagp, extra) \
366 STMT_START { \
367 if ((flags) & (RESTART_PASS1|NEED_UTF8|(extra))) { \
368 *(flagp) = (flags) & (RESTART_PASS1|NEED_UTF8|(extra)); \
369 return X; \
370 } \
371 } STMT_END
372
373#define RETURN_NULL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
374 RETURN_X_ON_RESTART_OR_FLAGS(NULL,flags,flagp,extra)
375
376#define RETURN_X_ON_RESTART(X, flags,flagp) \
377 RETURN_X_ON_RESTART_OR_FLAGS( X, flags, flagp, 0)
378
379
380#define RETURN_NULL_ON_RESTART_FLAGP_OR_FLAGS(flagp,extra) \
381 if (*(flagp) & (RESTART_PASS1|(extra))) return NULL
382
383#define MUST_RESTART(flags) ((flags) & (RESTART_PASS1))
384
385#define RETURN_NULL_ON_RESTART(flags,flagp) \
386 RETURN_X_ON_RESTART(NULL, flags,flagp)
387#define RETURN_NULL_ON_RESTART_FLAGP(flagp) \
388 RETURN_NULL_ON_RESTART_FLAGP_OR_FLAGS(flagp,0)
389
f19b1a63
KW
390/* This converts the named class defined in regcomp.h to its equivalent class
391 * number defined in handy.h. */
392#define namedclass_to_classnum(class) ((int) ((class) / 2))
393#define classnum_to_namedclass(classnum) ((classnum) * 2)
394
de92f5e6
KW
395#define _invlist_union_complement_2nd(a, b, output) \
396 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
397#define _invlist_intersection_complement_2nd(a, b, output) \
398 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
399
1de06328
YO
400/* About scan_data_t.
401
402 During optimisation we recurse through the regexp program performing
403 various inplace (keyhole style) optimisations. In addition study_chunk
404 and scan_commit populate this data structure with information about
538e84ed 405 what strings MUST appear in the pattern. We look for the longest
3b753521 406 string that must appear at a fixed location, and we look for the
1de06328
YO
407 longest string that may appear at a floating location. So for instance
408 in the pattern:
538e84ed 409
1de06328 410 /FOO[xX]A.*B[xX]BAR/
538e84ed 411
1de06328
YO
412 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
413 strings (because they follow a .* construct). study_chunk will identify
414 both FOO and BAR as being the longest fixed and floating strings respectively.
538e84ed 415
1de06328 416 The strings can be composites, for instance
538e84ed 417
1de06328 418 /(f)(o)(o)/
538e84ed 419
1de06328 420 will result in a composite fixed substring 'foo'.
538e84ed 421
1de06328 422 For each string some basic information is maintained:
538e84ed 423
6ea77169 424 - min_offset
1de06328
YO
425 This is the position the string must appear at, or not before.
426 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
427 characters must match before the string we are searching for.
428 Likewise when combined with minlenp and the length of the string it
538e84ed 429 tells us how many characters must appear after the string we have
1de06328 430 found.
538e84ed 431
1de06328
YO
432 - max_offset
433 Only used for floating strings. This is the rightmost point that
ea3daa5d 434 the string can appear at. If set to SSize_t_MAX it indicates that the
1de06328 435 string can occur infinitely far to the right.
6e12f832 436 For fixed strings, it is equal to min_offset.
538e84ed 437
1de06328 438 - minlenp
2d608413
KW
439 A pointer to the minimum number of characters of the pattern that the
440 string was found inside. This is important as in the case of positive
538e84ed 441 lookahead or positive lookbehind we can have multiple patterns
1de06328 442 involved. Consider
538e84ed 443
1de06328 444 /(?=FOO).*F/
538e84ed 445
1de06328
YO
446 The minimum length of the pattern overall is 3, the minimum length
447 of the lookahead part is 3, but the minimum length of the part that
538e84ed 448 will actually match is 1. So 'FOO's minimum length is 3, but the
1de06328 449 minimum length for the F is 1. This is important as the minimum length
538e84ed 450 is used to determine offsets in front of and behind the string being
1de06328 451 looked for. Since strings can be composites this is the length of the
486ec47a 452 pattern at the time it was committed with a scan_commit. Note that
1de06328 453 the length is calculated by study_chunk, so that the minimum lengths
538e84ed 454 are not known until the full pattern has been compiled, thus the
1de06328 455 pointer to the value.
538e84ed 456
1de06328 457 - lookbehind
538e84ed 458
1de06328 459 In the case of lookbehind the string being searched for can be
538e84ed 460 offset past the start point of the final matching string.
1de06328
YO
461 If this value was just blithely removed from the min_offset it would
462 invalidate some of the calculations for how many chars must match
463 before or after (as they are derived from min_offset and minlen and
538e84ed 464 the length of the string being searched for).
1de06328
YO
465 When the final pattern is compiled and the data is moved from the
466 scan_data_t structure into the regexp structure the information
538e84ed
KW
467 about lookbehind is factored in, with the information that would
468 have been lost precalculated in the end_shift field for the
1de06328
YO
469 associated string.
470
471 The fields pos_min and pos_delta are used to store the minimum offset
538e84ed 472 and the delta to the maximum offset at the current point in the pattern.
1de06328
YO
473
474*/
2c2d71f5 475
cb6aef8b
DM
476struct scan_data_substrs {
477 SV *str; /* longest substring found in pattern */
478 SSize_t min_offset; /* earliest point in string it can appear */
479 SSize_t max_offset; /* latest point in string it can appear */
480 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
481 SSize_t lookbehind; /* is the pos of the string modified by LB */
11683ecb 482 I32 flags; /* per substring SF_* and SCF_* flags */
cb6aef8b
DM
483};
484
2c2d71f5 485typedef struct scan_data_t {
1de06328
YO
486 /*I32 len_min; unused */
487 /*I32 len_delta; unused */
49f55535 488 SSize_t pos_min;
ea3daa5d 489 SSize_t pos_delta;
2c2d71f5 490 SV *last_found;
ea3daa5d 491 SSize_t last_end; /* min value, <0 unless valid. */
49f55535 492 SSize_t last_start_min;
ea3daa5d 493 SSize_t last_start_max;
2099df82
DM
494 U8 cur_is_floating; /* whether the last_* values should be set as
495 * the next fixed (0) or floating (1)
496 * substring */
6ea77169 497
2099df82
DM
498 /* [0] is longest fixed substring so far, [1] is longest float so far */
499 struct scan_data_substrs substrs[2];
6ea77169 500
11683ecb 501 I32 flags; /* common SF_* and SCF_* flags */
2c2d71f5 502 I32 whilem_c;
49f55535 503 SSize_t *last_closep;
b8f7bb16 504 regnode_ssc *start_class;
2c2d71f5
JH
505} scan_data_t;
506
a687059c 507/*
e50aee73 508 * Forward declarations for pregcomp()'s friends.
a687059c 509 */
a0d0e21e 510
6ea77169 511static const scan_data_t zero_scan_data = {
a2ca2017 512 0, 0, NULL, 0, 0, 0, 0,
6ea77169 513 {
11683ecb
DM
514 { NULL, 0, 0, 0, 0, 0 },
515 { NULL, 0, 0, 0, 0, 0 },
6ea77169
DM
516 },
517 0, 0, NULL, NULL
518};
c277df42 519
11683ecb
DM
520/* study flags */
521
07be1b83
YO
522#define SF_BEFORE_SEOL 0x0001
523#define SF_BEFORE_MEOL 0x0002
11683ecb 524#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
c277df42 525
07be1b83
YO
526#define SF_IS_INF 0x0040
527#define SF_HAS_PAR 0x0080
528#define SF_IN_PAR 0x0100
529#define SF_HAS_EVAL 0x0200
bc604ad8
KW
530
531
532/* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
533 * longest substring in the pattern. When it is not set the optimiser keeps
534 * track of position, but does not keep track of the actual strings seen,
535 *
536 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
537 * /foo/i will not.
538 *
539 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
540 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
541 * turned off because of the alternation (BRANCH). */
07be1b83 542#define SCF_DO_SUBSTR 0x0400
bc604ad8 543
653099ff
GS
544#define SCF_DO_STCLASS_AND 0x0800
545#define SCF_DO_STCLASS_OR 0x1000
546#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 547#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 548
786e8c11 549#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
538e84ed 550#define SCF_SEEN_ACCEPT 0x8000
688e0391 551#define SCF_TRIE_DOING_RESTUDY 0x10000
4286711a
YO
552#define SCF_IN_DEFINE 0x20000
553
554
555
07be1b83 556
43fead97 557#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
558
559/* The enums for all these are ordered so things work out correctly */
a62b1201 560#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
538e84ed
KW
561#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
562 == REGEX_DEPENDS_CHARSET)
00b27cfc 563#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
538e84ed
KW
564#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
565 >= REGEX_UNICODE_CHARSET)
566#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
567 == REGEX_ASCII_RESTRICTED_CHARSET)
568#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
569 >= REGEX_ASCII_RESTRICTED_CHARSET)
570#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
571 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 572
43fead97 573#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 574
f2c2a6ab
KW
575/* For programs that want to be strictly Unicode compatible by dying if any
576 * attempt is made to match a non-Unicode code point against a Unicode
577 * property. */
578#define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
579
93733859 580#define OOB_NAMEDCLASS -1
b8c5462f 581
8e661ac5
KW
582/* There is no code point that is out-of-bounds, so this is problematic. But
583 * its only current use is to initialize a variable that is always set before
584 * looked at. */
585#define OOB_UNICODE 0xDEADBEEF
586
a0ed51b3 587#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
a0ed51b3 588
8615cb43 589
b45f050a
JF
590/* length of regex to show in messages that don't mark a position within */
591#define RegexLengthToShowInErrorMessages 127
592
593/*
594 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
595 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
596 * op/pragma/warn/regcomp.
597 */
7253e4e3
RK
598#define MARKER1 "<-- HERE" /* marker as it appears in the description */
599#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 600
538e84ed 601#define REPORT_LOCATION " in regex; marked by " MARKER1 \
147e3846 602 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
b45f050a 603
285b5ca0
KW
604/* The code in this file in places uses one level of recursion with parsing
605 * rebased to an alternate string constructed by us in memory. This can take
606 * the form of something that is completely different from the input, or
607 * something that uses the input as part of the alternate. In the first case,
608 * there should be no possibility of an error, as we are in complete control of
609 * the alternate string. But in the second case we don't control the input
610 * portion, so there may be errors in that. Here's an example:
611 * /[abc\x{DF}def]/ui
612 * is handled specially because \x{df} folds to a sequence of more than one
613 * character, 'ss'. What is done is to create and parse an alternate string,
614 * which looks like this:
615 * /(?:\x{DF}|[abc\x{DF}def])/ui
616 * where it uses the input unchanged in the middle of something it constructs,
617 * which is a branch for the DF outside the character class, and clustering
618 * parens around the whole thing. (It knows enough to skip the DF inside the
619 * class while in this substitute parse.) 'abc' and 'def' may have errors that
620 * need to be reported. The general situation looks like this:
621 *
622 * sI tI xI eI
623 * Input: ----------------------------------------------------
624 * Constructed: ---------------------------------------------------
625 * sC tC xC eC EC
626 *
627 * The input string sI..eI is the input pattern. The string sC..EC is the
628 * constructed substitute parse string. The portions sC..tC and eC..EC are
629 * constructed by us. The portion tC..eC is an exact duplicate of the input
630 * pattern tI..eI. In the diagram, these are vertically aligned. Suppose that
631 * while parsing, we find an error at xC. We want to display a message showing
632 * the real input string. Thus we need to find the point xI in it which
633 * corresponds to xC. xC >= tC, since the portion of the string sC..tC has
634 * been constructed by us, and so shouldn't have errors. We get:
635 *
636 * xI = sI + (tI - sI) + (xC - tC)
637 *
638 * and, the offset into sI is:
639 *
640 * (xI - sI) = (tI - sI) + (xC - tC)
641 *
642 * When the substitute is constructed, we save (tI -sI) as RExC_precomp_adj,
643 * and we save tC as RExC_adjusted_start.
903c858a
KW
644 *
645 * During normal processing of the input pattern, everything points to that,
646 * with RExC_precomp_adj set to 0, and RExC_adjusted_start set to sI.
285b5ca0
KW
647 */
648
649#define tI_sI RExC_precomp_adj
650#define tC RExC_adjusted_start
651#define sC RExC_precomp
652#define xI_offset(xC) ((IV) (tI_sI + (xC - tC)))
653#define xI(xC) (sC + xI_offset(xC))
654#define eC RExC_precomp_end
655
656#define REPORT_LOCATION_ARGS(xC) \
657 UTF8fARG(UTF, \
658 (xI(xC) > eC) /* Don't run off end */ \
659 ? eC - sC /* Length before the <--HERE */ \
e0c80300 660 : ( __ASSERT_(xI_offset(xC) >= 0) xI_offset(xC) ), \
285b5ca0
KW
661 sC), /* The input pattern printed up to the <--HERE */ \
662 UTF8fARG(UTF, \
663 (xI(xC) > eC) ? 0 : eC - xI(xC), /* Length after <--HERE */ \
664 (xI(xC) > eC) ? eC : xI(xC)) /* pattern after <--HERE */
c1d900c3 665
8a6d8ec6
HS
666/* Used to point after bad bytes for an error message, but avoid skipping
667 * past a nul byte. */
668#define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
669
b45f050a
JF
670/*
671 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
672 * arg. Show regex, up to a maximum length. If it's too long, chop and add
673 * "...".
674 */
58e23c8d 675#define _FAIL(code) STMT_START { \
bfed75c6 676 const char *ellipses = ""; \
711b303b 677 IV len = RExC_precomp_end - RExC_precomp; \
ccb2c380
MP
678 \
679 if (!SIZE_ONLY) \
a5e7bc51 680 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
681 if (len > RegexLengthToShowInErrorMessages) { \
682 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
683 len = RegexLengthToShowInErrorMessages - 10; \
684 ellipses = "..."; \
685 } \
58e23c8d 686 code; \
ccb2c380 687} STMT_END
8615cb43 688
58e23c8d 689#define FAIL(msg) _FAIL( \
147e3846 690 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
c1d900c3 691 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d
YO
692
693#define FAIL2(msg,arg) _FAIL( \
147e3846 694 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
c1d900c3 695 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d 696
b45f050a 697/*
b45f050a
JF
698 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
699 */
ccb2c380 700#define Simple_vFAIL(m) STMT_START { \
ccb2c380 701 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
d528642a 702 m, REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 703} STMT_END
b45f050a
JF
704
705/*
706 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
707 */
ccb2c380
MP
708#define vFAIL(m) STMT_START { \
709 if (!SIZE_ONLY) \
a5e7bc51 710 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
711 Simple_vFAIL(m); \
712} STMT_END
b45f050a
JF
713
714/*
715 * Like Simple_vFAIL(), but accepts two arguments.
716 */
ccb2c380 717#define Simple_vFAIL2(m,a1) STMT_START { \
d528642a
KW
718 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
719 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 720} STMT_END
b45f050a
JF
721
722/*
723 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
724 */
ccb2c380
MP
725#define vFAIL2(m,a1) STMT_START { \
726 if (!SIZE_ONLY) \
a5e7bc51 727 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
728 Simple_vFAIL2(m, a1); \
729} STMT_END
b45f050a
JF
730
731
732/*
733 * Like Simple_vFAIL(), but accepts three arguments.
734 */
ccb2c380 735#define Simple_vFAIL3(m, a1, a2) STMT_START { \
c1d900c3 736 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
d528642a 737 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 738} STMT_END
b45f050a
JF
739
740/*
741 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
742 */
ccb2c380
MP
743#define vFAIL3(m,a1,a2) STMT_START { \
744 if (!SIZE_ONLY) \
a5e7bc51 745 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
746 Simple_vFAIL3(m, a1, a2); \
747} STMT_END
b45f050a
JF
748
749/*
750 * Like Simple_vFAIL(), but accepts four arguments.
751 */
ccb2c380 752#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
d528642a
KW
753 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
754 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 755} STMT_END
b45f050a 756
95db3ffa
KW
757#define vFAIL4(m,a1,a2,a3) STMT_START { \
758 if (!SIZE_ONLY) \
759 SAVEFREESV(RExC_rx_sv); \
760 Simple_vFAIL4(m, a1, a2, a3); \
761} STMT_END
762
946095af 763/* A specialized version of vFAIL2 that works with UTF8f */
d528642a
KW
764#define vFAIL2utf8f(m, a1) STMT_START { \
765 if (!SIZE_ONLY) \
766 SAVEFREESV(RExC_rx_sv); \
767 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
768 REPORT_LOCATION_ARGS(RExC_parse)); \
946095af
BF
769} STMT_END
770
3ba22297 771#define vFAIL3utf8f(m, a1, a2) STMT_START { \
3ba22297
KW
772 if (!SIZE_ONLY) \
773 SAVEFREESV(RExC_rx_sv); \
774 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
d528642a 775 REPORT_LOCATION_ARGS(RExC_parse)); \
3ba22297
KW
776} STMT_END
777
499333dc
KW
778/* These have asserts in them because of [perl #122671] Many warnings in
779 * regcomp.c can occur twice. If they get output in pass1 and later in that
780 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
781 * would get output again. So they should be output in pass2, and these
782 * asserts make sure new warnings follow that paradigm. */
946095af 783
5e0a247b
KW
784/* m is not necessarily a "literal string", in this macro */
785#define reg_warn_non_literal_string(loc, m) STMT_START { \
d528642a
KW
786 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
787 "%s" REPORT_LOCATION, \
788 m, REPORT_LOCATION_ARGS(loc)); \
5e0a247b
KW
789} STMT_END
790
668c081a 791#define ckWARNreg(loc,m) STMT_START { \
d528642a
KW
792 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
793 m REPORT_LOCATION, \
794 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
795} STMT_END
796
b927b7e9 797#define vWARN(loc, m) STMT_START { \
d528642a
KW
798 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
799 m REPORT_LOCATION, \
800 REPORT_LOCATION_ARGS(loc)); \
b927b7e9
KW
801} STMT_END
802
0d6106aa 803#define vWARN_dep(loc, m) STMT_START { \
d528642a
KW
804 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
805 m REPORT_LOCATION, \
806 REPORT_LOCATION_ARGS(loc)); \
0d6106aa
KW
807} STMT_END
808
147508a2 809#define ckWARNdep(loc,m) STMT_START { \
d528642a
KW
810 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
811 m REPORT_LOCATION, \
812 REPORT_LOCATION_ARGS(loc)); \
147508a2
KW
813} STMT_END
814
d528642a
KW
815#define ckWARNregdep(loc,m) STMT_START { \
816 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
817 WARN_REGEXP), \
818 m REPORT_LOCATION, \
819 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
820} STMT_END
821
d528642a
KW
822#define ckWARN2reg_d(loc,m, a1) STMT_START { \
823 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
824 m REPORT_LOCATION, \
825 a1, REPORT_LOCATION_ARGS(loc)); \
2335b3d3
KW
826} STMT_END
827
d528642a
KW
828#define ckWARN2reg(loc, m, a1) STMT_START { \
829 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
830 m REPORT_LOCATION, \
831 a1, REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
832} STMT_END
833
d528642a
KW
834#define vWARN3(loc, m, a1, a2) STMT_START { \
835 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
836 m REPORT_LOCATION, \
837 a1, a2, REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
838} STMT_END
839
d528642a
KW
840#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
841 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
842 m REPORT_LOCATION, \
843 a1, a2, \
844 REPORT_LOCATION_ARGS(loc)); \
668c081a
NC
845} STMT_END
846
ccb2c380 847#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
d528642a
KW
848 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
849 m REPORT_LOCATION, \
850 a1, a2, a3, \
851 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
852} STMT_END
853
668c081a 854#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
d528642a
KW
855 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
856 m REPORT_LOCATION, \
857 a1, a2, a3, \
858 REPORT_LOCATION_ARGS(loc)); \
668c081a
NC
859} STMT_END
860
ccb2c380 861#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
d528642a
KW
862 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
863 m REPORT_LOCATION, \
864 a1, a2, a3, a4, \
865 REPORT_LOCATION_ARGS(loc)); \
ccb2c380 866} STMT_END
9d1d55b5 867
538e84ed 868/* Macros for recording node offsets. 20001227 mjd@plover.com
fac92740
MJD
869 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
870 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
871 * Element 0 holds the number n.
07be1b83 872 * Position is 1 indexed.
fac92740 873 */
7122b237
YO
874#ifndef RE_TRACK_PATTERN_OFFSETS
875#define Set_Node_Offset_To_R(node,byte)
876#define Set_Node_Offset(node,byte)
877#define Set_Cur_Node_Offset
878#define Set_Node_Length_To_R(node,len)
879#define Set_Node_Length(node,len)
6a86c6ad 880#define Set_Node_Cur_Length(node,start)
538e84ed
KW
881#define Node_Offset(n)
882#define Node_Length(n)
7122b237
YO
883#define Set_Node_Offset_Length(node,offset,len)
884#define ProgLen(ri) ri->u.proglen
885#define SetProgLen(ri,x) ri->u.proglen = x
886#else
887#define ProgLen(ri) ri->u.offsets[0]
888#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
889#define Set_Node_Offset_To_R(node,byte) STMT_START { \
890 if (! SIZE_ONLY) { \
891 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 892 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 893 if((node) < 0) { \
538e84ed
KW
894 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
895 (int)(node)); \
ccb2c380
MP
896 } else { \
897 RExC_offsets[2*(node)-1] = (byte); \
898 } \
899 } \
900} STMT_END
901
902#define Set_Node_Offset(node,byte) \
903 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
904#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
905
906#define Set_Node_Length_To_R(node,len) STMT_START { \
907 if (! SIZE_ONLY) { \
908 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 909 __LINE__, (int)(node), (int)(len))); \
ccb2c380 910 if((node) < 0) { \
538e84ed
KW
911 Perl_croak(aTHX_ "value of node is %d in Length macro", \
912 (int)(node)); \
ccb2c380
MP
913 } else { \
914 RExC_offsets[2*(node)] = (len); \
915 } \
916 } \
917} STMT_END
918
919#define Set_Node_Length(node,len) \
920 Set_Node_Length_To_R((node)-RExC_emit_start, len)
6a86c6ad
NC
921#define Set_Node_Cur_Length(node, start) \
922 Set_Node_Length(node, RExC_parse - start)
fac92740
MJD
923
924/* Get offsets and lengths */
925#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
926#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
927
07be1b83
YO
928#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
929 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
930 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
931} STMT_END
7122b237 932#endif
07be1b83
YO
933
934#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
935#define EXPERIMENTAL_INPLACESCAN
f427392e 936#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 937
cb41e5d6
YO
938#ifdef DEBUGGING
939int
6ad9a8ab 940Perl_re_printf(pTHX_ const char *fmt, ...)
cb41e5d6
YO
941{
942 va_list ap;
943 int result;
944 PerlIO *f= Perl_debug_log;
945 PERL_ARGS_ASSERT_RE_PRINTF;
946 va_start(ap, fmt);
947 result = PerlIO_vprintf(f, fmt, ap);
948 va_end(ap);
949 return result;
950}
951
952int
7b031478 953Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
cb41e5d6
YO
954{
955 va_list ap;
956 int result;
957 PerlIO *f= Perl_debug_log;
958 PERL_ARGS_ASSERT_RE_INDENTF;
959 va_start(ap, depth);
daeb874b 960 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
cb41e5d6
YO
961 result = PerlIO_vprintf(f, fmt, ap);
962 va_end(ap);
963 return result;
964}
965#endif /* DEBUGGING */
966
7b031478 967#define DEBUG_RExC_seen() \
538e84ed 968 DEBUG_OPTIMISE_MORE_r({ \
6ad9a8ab 969 Perl_re_printf( aTHX_ "RExC_seen: "); \
538e84ed 970 \
e384d5c1 971 if (RExC_seen & REG_ZERO_LEN_SEEN) \
6ad9a8ab 972 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
538e84ed 973 \
e384d5c1 974 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
6ad9a8ab 975 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
538e84ed 976 \
e384d5c1 977 if (RExC_seen & REG_GPOS_SEEN) \
6ad9a8ab 978 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
538e84ed 979 \
e384d5c1 980 if (RExC_seen & REG_RECURSE_SEEN) \
6ad9a8ab 981 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
538e84ed 982 \
7b031478 983 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
6ad9a8ab 984 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
538e84ed 985 \
e384d5c1 986 if (RExC_seen & REG_VERBARG_SEEN) \
6ad9a8ab 987 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
538e84ed 988 \
e384d5c1 989 if (RExC_seen & REG_CUTGROUP_SEEN) \
6ad9a8ab 990 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
538e84ed 991 \
e384d5c1 992 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
6ad9a8ab 993 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
538e84ed 994 \
e384d5c1 995 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
6ad9a8ab 996 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
538e84ed 997 \
7b031478 998 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
6ad9a8ab 999 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
ee273784 1000 \
6ad9a8ab 1001 Perl_re_printf( aTHX_ "\n"); \
9e9ecfdf
YO
1002 });
1003
fdfb4f21 1004#define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
6ad9a8ab 1005 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
fdfb4f21 1006
fdfb4f21 1007
f5a36d78
DM
1008#ifdef DEBUGGING
1009static void
1010S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1011 const char *close_str)
1012{
1013 if (!flags)
1014 return;
1015
1016 Perl_re_printf( aTHX_ "%s", open_str);
11683ecb
DM
1017 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1018 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
f5a36d78
DM
1019 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1020 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1021 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1022 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1023 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1024 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1025 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1026 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1027 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1028 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1029 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1030 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1031 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1032 Perl_re_printf( aTHX_ "%s", close_str);
1033}
1034
1035
1036static void
1037S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1038 U32 depth, int is_inf)
1039{
1040 GET_RE_DEBUG_FLAGS_DECL;
1041
1042 DEBUG_OPTIMISE_MORE_r({
1043 if (!data)
1044 return;
1045 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1046 depth,
1047 where,
1048 (IV)data->pos_min,
1049 (IV)data->pos_delta,
1050 (UV)data->flags
1051 );
1052
1053 S_debug_show_study_flags(aTHX_ data->flags," [","]");
fdfb4f21 1054
f5a36d78
DM
1055 Perl_re_printf( aTHX_
1056 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1057 (IV)data->whilem_c,
1058 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1059 is_inf ? "INF " : ""
1060 );
1061
11683ecb 1062 if (data->last_found) {
37b6262f 1063 int i;
11683ecb
DM
1064 Perl_re_printf(aTHX_
1065 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1066 SvPVX_const(data->last_found),
1067 (IV)data->last_end,
1068 (IV)data->last_start_min,
1069 (IV)data->last_start_max
1070 );
1071
2099df82 1072 for (i = 0; i < 2; i++) {
37b6262f
DM
1073 Perl_re_printf(aTHX_
1074 " %s%s: '%s' @ %" IVdf "/%" IVdf,
2099df82
DM
1075 data->cur_is_floating == i ? "*" : "",
1076 i ? "Float" : "Fixed",
37b6262f
DM
1077 SvPVX_const(data->substrs[i].str),
1078 (IV)data->substrs[i].min_offset,
1079 (IV)data->substrs[i].max_offset
1080 );
1081 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1082 }
11683ecb 1083 }
f5a36d78
DM
1084
1085 Perl_re_printf( aTHX_ "\n");
1086 });
1087}
1088
1089
1090static void
1091S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1092 regnode *scan, U32 depth, U32 flags)
1093{
1094 GET_RE_DEBUG_FLAGS_DECL;
1095
1096 DEBUG_OPTIMISE_r({
1097 regnode *Next;
1098
1099 if (!scan)
1100 return;
1101 Next = regnext(scan);
1102 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1103 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1104 depth,
1105 str,
1106 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1107 Next ? (REG_NODE_NUM(Next)) : 0 );
1108 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1109 Perl_re_printf( aTHX_ "\n");
1110 });
1111}
1112
1113
1114# define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1115 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1116
1117# define DEBUG_PEEP(str, scan, depth, flags) \
1118 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1119
1120#else
1121# define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1122# define DEBUG_PEEP(str, scan, depth, flags) NOOP
1123#endif
1de06328 1124
cb41e5d6 1125
c6871b76
KW
1126/* =========================================================
1127 * BEGIN edit_distance stuff.
1128 *
1129 * This calculates how many single character changes of any type are needed to
1130 * transform a string into another one. It is taken from version 3.1 of
1131 *
1132 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1133 */
1134
1135/* Our unsorted dictionary linked list. */
1136/* Note we use UVs, not chars. */
1137
1138struct dictionary{
1139 UV key;
1140 UV value;
1141 struct dictionary* next;
1142};
1143typedef struct dictionary item;
1144
1145
1146PERL_STATIC_INLINE item*
1147push(UV key,item* curr)
1148{
1149 item* head;
d09f14bf 1150 Newx(head, 1, item);
c6871b76
KW
1151 head->key = key;
1152 head->value = 0;
1153 head->next = curr;
1154 return head;
1155}
1156
1157
1158PERL_STATIC_INLINE item*
1159find(item* head, UV key)
1160{
1161 item* iterator = head;
1162 while (iterator){
1163 if (iterator->key == key){
1164 return iterator;
1165 }
1166 iterator = iterator->next;
1167 }
1168
1169 return NULL;
1170}
1171
1172PERL_STATIC_INLINE item*
1173uniquePush(item* head,UV key)
1174{
1175 item* iterator = head;
1176
1177 while (iterator){
1178 if (iterator->key == key) {
1179 return head;
1180 }
1181 iterator = iterator->next;
1182 }
1183
1184 return push(key,head);
1185}
1186
1187PERL_STATIC_INLINE void
1188dict_free(item* head)
1189{
1190 item* iterator = head;
1191
1192 while (iterator) {
1193 item* temp = iterator;
1194 iterator = iterator->next;
1195 Safefree(temp);
1196 }
1197
1198 head = NULL;
1199}
1200
1201/* End of Dictionary Stuff */
1202
1203/* All calculations/work are done here */
1204STATIC int
1205S_edit_distance(const UV* src,
1206 const UV* tgt,
1207 const STRLEN x, /* length of src[] */
1208 const STRLEN y, /* length of tgt[] */
1209 const SSize_t maxDistance
1210)
1211{
1212 item *head = NULL;
1213 UV swapCount,swapScore,targetCharCount,i,j;
1214 UV *scores;
1215 UV score_ceil = x + y;
1216
1217 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1218
1219 /* intialize matrix start values */
d09f14bf 1220 Newx(scores, ( (x + 2) * (y + 2)), UV);
c6871b76
KW
1221 scores[0] = score_ceil;
1222 scores[1 * (y + 2) + 0] = score_ceil;
1223 scores[0 * (y + 2) + 1] = score_ceil;
1224 scores[1 * (y + 2) + 1] = 0;
1225 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1226
1227 /* work loops */
1228 /* i = src index */
1229 /* j = tgt index */
1230 for (i=1;i<=x;i++) {
1231 if (i < x)
1232 head = uniquePush(head,src[i]);
1233 scores[(i+1) * (y + 2) + 1] = i;
1234 scores[(i+1) * (y + 2) + 0] = score_ceil;
1235 swapCount = 0;
1236
1237 for (j=1;j<=y;j++) {
1238 if (i == 1) {
1239 if(j < y)
1240 head = uniquePush(head,tgt[j]);
1241 scores[1 * (y + 2) + (j + 1)] = j;
1242 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1243 }
1244
1245 targetCharCount = find(head,tgt[j-1])->value;
1246 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1247
1248 if (src[i-1] != tgt[j-1]){
1249 scores[(i+1) * (y + 2) + (j + 1)] = MIN(swapScore,(MIN(scores[i * (y + 2) + j], MIN(scores[(i+1) * (y + 2) + j], scores[i * (y + 2) + (j + 1)])) + 1));
1250 }
1251 else {
1252 swapCount = j;
1253 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1254 }
1255 }
1256
1257 find(head,src[i-1])->value = i;
1258 }
1259
1260 {
1261 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1262 dict_free(head);
1263 Safefree(scores);
1264 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1265 }
1266}
1267
1268/* END of edit_distance() stuff
1269 * ========================================================= */
1270
8e35b056
KW
1271/* is c a control character for which we have a mnemonic? */
1272#define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1273
549b4e78
KW
1274STATIC const char *
1275S_cntrl_to_mnemonic(const U8 c)
1276{
1277 /* Returns the mnemonic string that represents character 'c', if one
1278 * exists; NULL otherwise. The only ones that exist for the purposes of
1279 * this routine are a few control characters */
1280
1281 switch (c) {
1282 case '\a': return "\\a";
1283 case '\b': return "\\b";
1284 case ESC_NATIVE: return "\\e";
1285 case '\f': return "\\f";
1286 case '\n': return "\\n";
1287 case '\r': return "\\r";
1288 case '\t': return "\\t";
1289 }
1290
1291 return NULL;
1292}
1293
653099ff 1294/* Mark that we cannot extend a found fixed substring at this point.
37b6262f 1295 Update the longest found anchored substring or the longest found
653099ff
GS
1296 floating substrings if needed. */
1297
4327152a 1298STATIC void
ea3daa5d
FC
1299S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1300 SSize_t *minlenp, int is_inf)
c277df42 1301{
e1ec3a88 1302 const STRLEN l = CHR_SVLEN(data->last_found);
2099df82 1303 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
a2ca2017 1304 const STRLEN old_l = CHR_SVLEN(longest_sv);
1de06328 1305 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 1306
7918f24d
NC
1307 PERL_ARGS_ASSERT_SCAN_COMMIT;
1308
c277df42 1309 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
2099df82 1310 const U8 i = data->cur_is_floating;
a2ca2017 1311 SvSetMagicSV(longest_sv, data->last_found);
37b6262f
DM
1312 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1313
2099df82
DM
1314 if (!i) /* fixed */
1315 data->substrs[0].max_offset = data->substrs[0].min_offset;
1316 else { /* float */
1317 data->substrs[1].max_offset = (l
646e8787
DM
1318 ? data->last_start_max
1319 : (data->pos_delta > SSize_t_MAX - data->pos_min
ea3daa5d
FC
1320 ? SSize_t_MAX
1321 : data->pos_min + data->pos_delta));
1322 if (is_inf
2099df82
DM
1323 || (STRLEN)data->substrs[1].max_offset > (STRLEN)SSize_t_MAX)
1324 data->substrs[1].max_offset = SSize_t_MAX;
37b6262f 1325 }
11683ecb 1326
37b6262f
DM
1327 if (data->flags & SF_BEFORE_EOL)
1328 data->substrs[i].flags |= (data->flags & SF_BEFORE_EOL);
1329 else
1330 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1331 data->substrs[i].minlenp = minlenp;
1332 data->substrs[i].lookbehind = 0;
c277df42 1333 }
37b6262f 1334
c277df42 1335 SvCUR_set(data->last_found, 0);
0eda9292 1336 {
a28509cc 1337 SV * const sv = data->last_found;
097eb12c
AL
1338 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1339 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1340 if (mg)
1341 mg->mg_len = 0;
1342 }
0eda9292 1343 }
c277df42
IZ
1344 data->last_end = -1;
1345 data->flags &= ~SF_BEFORE_EOL;
f5a36d78 1346 DEBUG_STUDYDATA("commit", data, 0, is_inf);
c277df42
IZ
1347}
1348
cdd87c1d
KW
1349/* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1350 * list that describes which code points it matches */
1351
653099ff 1352STATIC void
3420edd7 1353S_ssc_anything(pTHX_ regnode_ssc *ssc)
653099ff 1354{
cdd87c1d
KW
1355 /* Set the SSC 'ssc' to match an empty string or any code point */
1356
557bd3fb 1357 PERL_ARGS_ASSERT_SSC_ANYTHING;
7918f24d 1358
71068078 1359 assert(is_ANYOF_SYNTHETIC(ssc));
3fffb88a 1360
0854ea0b
KW
1361 /* mortalize so won't leak */
1362 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
93e92956 1363 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
653099ff
GS
1364}
1365
653099ff 1366STATIC int
dc3bf405 1367S_ssc_is_anything(const regnode_ssc *ssc)
653099ff 1368{
c144baaa
KW
1369 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1370 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1371 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1372 * in any way, so there's no point in using it */
cdd87c1d
KW
1373
1374 UV start, end;
1375 bool ret;
653099ff 1376
557bd3fb 1377 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
7918f24d 1378
71068078 1379 assert(is_ANYOF_SYNTHETIC(ssc));
cdd87c1d 1380
93e92956 1381 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
cdd87c1d
KW
1382 return FALSE;
1383 }
1384
1385 /* See if the list consists solely of the range 0 - Infinity */
1386 invlist_iterinit(ssc->invlist);
1387 ret = invlist_iternext(ssc->invlist, &start, &end)
1388 && start == 0
1389 && end == UV_MAX;
1390
1391 invlist_iterfinish(ssc->invlist);
1392
1393 if (ret) {
1394 return TRUE;
1395 }
1396
1397 /* If e.g., both \w and \W are set, matches everything */
e0e1be5f 1398 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1399 int i;
1400 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1401 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1402 return TRUE;
1403 }
1404 }
1405 }
1406
1407 return FALSE;
653099ff
GS
1408}
1409
653099ff 1410STATIC void
cdd87c1d 1411S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
653099ff 1412{
cdd87c1d
KW
1413 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1414 * string, any code point, or any posix class under locale */
1415
557bd3fb 1416 PERL_ARGS_ASSERT_SSC_INIT;
7918f24d 1417
557bd3fb 1418 Zero(ssc, 1, regnode_ssc);
71068078 1419 set_ANYOF_SYNTHETIC(ssc);
93e92956 1420 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
3420edd7 1421 ssc_anything(ssc);
cdd87c1d 1422
2f306ab9
KW
1423 /* If any portion of the regex is to operate under locale rules that aren't
1424 * fully known at compile time, initialization includes it. The reason
1425 * this isn't done for all regexes is that the optimizer was written under
1426 * the assumption that locale was all-or-nothing. Given the complexity and
1427 * lack of documentation in the optimizer, and that there are inadequate
1428 * test cases for locale, many parts of it may not work properly, it is
1429 * safest to avoid locale unless necessary. */
cdd87c1d
KW
1430 if (RExC_contains_locale) {
1431 ANYOF_POSIXL_SETALL(ssc);
cdd87c1d
KW
1432 }
1433 else {
1434 ANYOF_POSIXL_ZERO(ssc);
1435 }
653099ff
GS
1436}
1437
b423522f 1438STATIC int
dc3bf405
BF
1439S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1440 const regnode_ssc *ssc)
b423522f
KW
1441{
1442 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1443 * to the list of code points matched, and locale posix classes; hence does
1444 * not check its flags) */
1445
1446 UV start, end;
1447 bool ret;
1448
1449 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1450
71068078 1451 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1452
1453 invlist_iterinit(ssc->invlist);
1454 ret = invlist_iternext(ssc->invlist, &start, &end)
1455 && start == 0
1456 && end == UV_MAX;
1457
1458 invlist_iterfinish(ssc->invlist);
1459
1460 if (! ret) {
1461 return FALSE;
1462 }
1463
e0e1be5f 1464 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
31f05a37 1465 return FALSE;
b423522f
KW
1466 }
1467
1468 return TRUE;
1469}
1470
1471STATIC SV*
1472S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
5c0f85ef 1473 const regnode_charclass* const node)
b423522f
KW
1474{
1475 /* Returns a mortal inversion list defining which code points are matched
1476 * by 'node', which is of type ANYOF. Handles complementing the result if
1477 * appropriate. If some code points aren't knowable at this time, the
31f05a37
KW
1478 * returned list must, and will, contain every code point that is a
1479 * possibility. */
b423522f 1480
e2506fa7 1481 SV* invlist = NULL;
1ee208c4 1482 SV* only_utf8_locale_invlist = NULL;
b423522f
KW
1483 unsigned int i;
1484 const U32 n = ARG(node);
31f05a37 1485 bool new_node_has_latin1 = FALSE;
b423522f
KW
1486
1487 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1488
1489 /* Look at the data structure created by S_set_ANYOF_arg() */
93e92956 1490 if (n != ANYOF_ONLY_HAS_BITMAP) {
b423522f
KW
1491 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1492 AV * const av = MUTABLE_AV(SvRV(rv));
1493 SV **const ary = AvARRAY(av);
1494 assert(RExC_rxi->data->what[n] == 's');
1495
1496 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1497 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1498 }
1499 else if (ary[0] && ary[0] != &PL_sv_undef) {
1500
1501 /* Here, no compile-time swash, and there are things that won't be
1502 * known until runtime -- we have to assume it could be anything */
e2506fa7 1503 invlist = sv_2mortal(_new_invlist(1));
b423522f
KW
1504 return _add_range_to_invlist(invlist, 0, UV_MAX);
1505 }
1ee208c4 1506 else if (ary[3] && ary[3] != &PL_sv_undef) {
b423522f
KW
1507
1508 /* Here no compile-time swash, and no run-time only data. Use the
1509 * node's inversion list */
1ee208c4
KW
1510 invlist = sv_2mortal(invlist_clone(ary[3]));
1511 }
1512
1513 /* Get the code points valid only under UTF-8 locales */
037715a6 1514 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1ee208c4
KW
1515 && ary[2] && ary[2] != &PL_sv_undef)
1516 {
1517 only_utf8_locale_invlist = ary[2];
b423522f
KW
1518 }
1519 }
1520
e2506fa7
KW
1521 if (! invlist) {
1522 invlist = sv_2mortal(_new_invlist(0));
1523 }
1524
dcb20b36
KW
1525 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1526 * code points, and an inversion list for the others, but if there are code
1527 * points that should match only conditionally on the target string being
1528 * UTF-8, those are placed in the inversion list, and not the bitmap.
1529 * Since there are circumstances under which they could match, they are
1530 * included in the SSC. But if the ANYOF node is to be inverted, we have
1531 * to exclude them here, so that when we invert below, the end result
1532 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1533 * have to do this here before we add the unconditionally matched code
1534 * points */
b423522f
KW
1535 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1536 _invlist_intersection_complement_2nd(invlist,
1537 PL_UpperLatin1,
1538 &invlist);
1539 }
1540
1541 /* Add in the points from the bit map */
dcb20b36 1542 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
b423522f 1543 if (ANYOF_BITMAP_TEST(node, i)) {
6f8848d5
TC
1544 unsigned int start = i++;
1545
1546 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1547 /* empty */
1548 }
1549 invlist = _add_range_to_invlist(invlist, start, i-1);
31f05a37 1550 new_node_has_latin1 = TRUE;
b423522f
KW
1551 }
1552 }
1553
1554 /* If this can match all upper Latin1 code points, have to add them
ac33c516
KW
1555 * as well. But don't add them if inverting, as when that gets done below,
1556 * it would exclude all these characters, including the ones it shouldn't
1557 * that were added just above */
1558 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
f240c685
KW
1559 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1560 {
b423522f
KW
1561 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1562 }
1563
1564 /* Similarly for these */
93e92956 1565 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
e0a1ff7a 1566 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
b423522f
KW
1567 }
1568
1569 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1570 _invlist_invert(invlist);
1571 }
037715a6 1572 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
31f05a37
KW
1573
1574 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1575 * locale. We can skip this if there are no 0-255 at all. */
1576 _invlist_union(invlist, PL_Latin1, &invlist);
1577 }
1578
1ee208c4
KW
1579 /* Similarly add the UTF-8 locale possible matches. These have to be
1580 * deferred until after the non-UTF-8 locale ones are taken care of just
1581 * above, or it leads to wrong results under ANYOF_INVERT */
1582 if (only_utf8_locale_invlist) {
31f05a37 1583 _invlist_union_maybe_complement_2nd(invlist,
1ee208c4 1584 only_utf8_locale_invlist,
31f05a37
KW
1585 ANYOF_FLAGS(node) & ANYOF_INVERT,
1586 &invlist);
1587 }
b423522f
KW
1588
1589 return invlist;
1590}
1591
1051e1c4 1592/* These two functions currently do the exact same thing */
557bd3fb 1593#define ssc_init_zero ssc_init
653099ff 1594
cdd87c1d
KW
1595#define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1596#define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1597
557bd3fb 1598/* 'AND' a given class with another one. Can create false positives. 'ssc'
93e92956
KW
1599 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1600 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
cdd87c1d 1601
653099ff 1602STATIC void
b423522f 1603S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1604 const regnode_charclass *and_with)
653099ff 1605{
cdd87c1d
KW
1606 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1607 * another SSC or a regular ANYOF class. Can create false positives. */
40d049e4 1608
a0dd4231
KW
1609 SV* anded_cp_list;
1610 U8 anded_flags;
1e6ade67 1611
cdd87c1d 1612 PERL_ARGS_ASSERT_SSC_AND;
653099ff 1613
71068078 1614 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1615
1616 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1617 * the code point inversion list and just the relevant flags */
71068078 1618 if (is_ANYOF_SYNTHETIC(and_with)) {
7dcac5f6 1619 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
a0dd4231 1620 anded_flags = ANYOF_FLAGS(and_with);
e9b08962
KW
1621
1622 /* XXX This is a kludge around what appears to be deficiencies in the
1623 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1624 * there are paths through the optimizer where it doesn't get weeded
1625 * out when it should. And if we don't make some extra provision for
1626 * it like the code just below, it doesn't get added when it should.
1627 * This solution is to add it only when AND'ing, which is here, and
1628 * only when what is being AND'ed is the pristine, original node
1629 * matching anything. Thus it is like adding it to ssc_anything() but
1630 * only when the result is to be AND'ed. Probably the same solution
1631 * could be adopted for the same problem we have with /l matching,
1632 * which is solved differently in S_ssc_init(), and that would lead to
1633 * fewer false positives than that solution has. But if this solution
1634 * creates bugs, the consequences are only that a warning isn't raised
1635 * that should be; while the consequences for having /l bugs is
1636 * incorrect matches */
7dcac5f6 1637 if (ssc_is_anything((regnode_ssc *)and_with)) {
f240c685 1638 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
e9b08962 1639 }
a0dd4231
KW
1640 }
1641 else {
5c0f85ef 1642 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
f240c685
KW
1643 if (OP(and_with) == ANYOFD) {
1644 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1645 }
1646 else {
1647 anded_flags = ANYOF_FLAGS(and_with)
1648 &( ANYOF_COMMON_FLAGS
108316fb
KW
1649 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1650 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
d1c40ef5
KW
1651 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1652 anded_flags &=
1653 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1654 }
f240c685 1655 }
a0dd4231
KW
1656 }
1657
1658 ANYOF_FLAGS(ssc) &= anded_flags;
cdd87c1d
KW
1659
1660 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1661 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1662 * 'and_with' may be inverted. When not inverted, we have the situation of
1663 * computing:
1664 * (C1 | P1) & (C2 | P2)
1665 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1666 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1667 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1668 * <= ((C1 & C2) | P1 | P2)
1669 * Alternatively, the last few steps could be:
1670 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1671 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1672 * <= (C1 | C2 | (P1 & P2))
1673 * We favor the second approach if either P1 or P2 is non-empty. This is
1674 * because these components are a barrier to doing optimizations, as what
1675 * they match cannot be known until the moment of matching as they are
1676 * dependent on the current locale, 'AND"ing them likely will reduce or
1677 * eliminate them.
1678 * But we can do better if we know that C1,P1 are in their initial state (a
1679 * frequent occurrence), each matching everything:
1680 * (<everything>) & (C2 | P2) = C2 | P2
1681 * Similarly, if C2,P2 are in their initial state (again a frequent
1682 * occurrence), the result is a no-op
1683 * (C1 | P1) & (<everything>) = C1 | P1
1684 *
1685 * Inverted, we have
1686 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1687 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1688 * <= (C1 & ~C2) | (P1 & ~P2)
1689 * */
1aa99e6b 1690
a0dd4231 1691 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
71068078 1692 && ! is_ANYOF_SYNTHETIC(and_with))
a0dd4231 1693 {
cdd87c1d 1694 unsigned int i;
8951c461 1695
cdd87c1d
KW
1696 ssc_intersection(ssc,
1697 anded_cp_list,
1698 FALSE /* Has already been inverted */
1699 );
c6b76537 1700
cdd87c1d
KW
1701 /* If either P1 or P2 is empty, the intersection will be also; can skip
1702 * the loop */
93e92956 1703 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
cdd87c1d
KW
1704 ANYOF_POSIXL_ZERO(ssc);
1705 }
e0e1be5f 1706 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1707
1708 /* Note that the Posix class component P from 'and_with' actually
1709 * looks like:
1710 * P = Pa | Pb | ... | Pn
1711 * where each component is one posix class, such as in [\w\s].
1712 * Thus
1713 * ~P = ~(Pa | Pb | ... | Pn)
1714 * = ~Pa & ~Pb & ... & ~Pn
1715 * <= ~Pa | ~Pb | ... | ~Pn
1716 * The last is something we can easily calculate, but unfortunately
1717 * is likely to have many false positives. We could do better
1718 * in some (but certainly not all) instances if two classes in
1719 * P have known relationships. For example
1720 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1721 * So
1722 * :lower: & :print: = :lower:
1723 * And similarly for classes that must be disjoint. For example,
1724 * since \s and \w can have no elements in common based on rules in
1725 * the POSIX standard,
1726 * \w & ^\S = nothing
1727 * Unfortunately, some vendor locales do not meet the Posix
1728 * standard, in particular almost everything by Microsoft.
1729 * The loop below just changes e.g., \w into \W and vice versa */
1730
1ee208c4 1731 regnode_charclass_posixl temp;
cdd87c1d
KW
1732 int add = 1; /* To calculate the index of the complement */
1733
b1234259 1734 Zero(&temp, 1, regnode_charclass_posixl);
cdd87c1d
KW
1735 ANYOF_POSIXL_ZERO(&temp);
1736 for (i = 0; i < ANYOF_MAX; i++) {
1737 assert(i % 2 != 0
7dcac5f6
KW
1738 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1739 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
cdd87c1d 1740
7dcac5f6 1741 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
cdd87c1d
KW
1742 ANYOF_POSIXL_SET(&temp, i + add);
1743 }
1744 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1745 }
1746 ANYOF_POSIXL_AND(&temp, ssc);
c6b76537 1747
cdd87c1d
KW
1748 } /* else ssc already has no posixes */
1749 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1750 in its initial state */
71068078 1751 else if (! is_ANYOF_SYNTHETIC(and_with)
7dcac5f6 1752 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
cdd87c1d
KW
1753 {
1754 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1755 * copy it over 'ssc' */
1756 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
71068078 1757 if (is_ANYOF_SYNTHETIC(and_with)) {
cdd87c1d
KW
1758 StructCopy(and_with, ssc, regnode_ssc);
1759 }
1760 else {
1761 ssc->invlist = anded_cp_list;
1762 ANYOF_POSIXL_ZERO(ssc);
93e92956 1763 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1764 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
cdd87c1d
KW
1765 }
1766 }
1767 }
e0e1be5f 1768 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
93e92956 1769 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
cdd87c1d
KW
1770 {
1771 /* One or the other of P1, P2 is non-empty. */
93e92956 1772 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1ea8b7fe
KW
1773 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1774 }
cdd87c1d
KW
1775 ssc_union(ssc, anded_cp_list, FALSE);
1776 }
1777 else { /* P1 = P2 = empty */
1778 ssc_intersection(ssc, anded_cp_list, FALSE);
1779 }
137165a6 1780 }
653099ff
GS
1781}
1782
653099ff 1783STATIC void
cdd87c1d 1784S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1785 const regnode_charclass *or_with)
653099ff 1786{
cdd87c1d
KW
1787 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1788 * another SSC or a regular ANYOF class. Can create false positives if
1789 * 'or_with' is to be inverted. */
7918f24d 1790
a0dd4231
KW
1791 SV* ored_cp_list;
1792 U8 ored_flags;
c6b76537 1793
cdd87c1d 1794 PERL_ARGS_ASSERT_SSC_OR;
c6b76537 1795
71068078 1796 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1797
1798 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1799 * the code point inversion list and just the relevant flags */
71068078 1800 if (is_ANYOF_SYNTHETIC(or_with)) {
7dcac5f6 1801 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
a0dd4231
KW
1802 ored_flags = ANYOF_FLAGS(or_with);
1803 }
1804 else {
5c0f85ef 1805 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
eff8b7dc 1806 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
f240c685
KW
1807 if (OP(or_with) != ANYOFD) {
1808 ored_flags
1809 |= ANYOF_FLAGS(or_with)
108316fb
KW
1810 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1811 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
d1c40ef5
KW
1812 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1813 ored_flags |=
1814 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1815 }
f240c685 1816 }
a0dd4231
KW
1817 }
1818
1819 ANYOF_FLAGS(ssc) |= ored_flags;
cdd87c1d
KW
1820
1821 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1822 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1823 * 'or_with' may be inverted. When not inverted, we have the simple
1824 * situation of computing:
1825 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1826 * If P1|P2 yields a situation with both a class and its complement are
1827 * set, like having both \w and \W, this matches all code points, and we
1828 * can delete these from the P component of the ssc going forward. XXX We
1829 * might be able to delete all the P components, but I (khw) am not certain
1830 * about this, and it is better to be safe.
1831 *
1832 * Inverted, we have
1833 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1834 * <= (C1 | P1) | ~C2
1835 * <= (C1 | ~C2) | P1
1836 * (which results in actually simpler code than the non-inverted case)
1837 * */
9826f543 1838
a0dd4231 1839 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
71068078 1840 && ! is_ANYOF_SYNTHETIC(or_with))
a0dd4231 1841 {
cdd87c1d 1842 /* We ignore P2, leaving P1 going forward */
1ea8b7fe 1843 } /* else Not inverted */
93e92956 1844 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1845 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
e0e1be5f 1846 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1847 unsigned int i;
1848 for (i = 0; i < ANYOF_MAX; i += 2) {
1849 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1850 {
1851 ssc_match_all_cp(ssc);
1852 ANYOF_POSIXL_CLEAR(ssc, i);
1853 ANYOF_POSIXL_CLEAR(ssc, i+1);
cdd87c1d
KW
1854 }
1855 }
1856 }
1aa99e6b 1857 }
cdd87c1d
KW
1858
1859 ssc_union(ssc,
1860 ored_cp_list,
1861 FALSE /* Already has been inverted */
1862 );
653099ff
GS
1863}
1864
b423522f
KW
1865PERL_STATIC_INLINE void
1866S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1867{
1868 PERL_ARGS_ASSERT_SSC_UNION;
1869
71068078 1870 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1871
1872 _invlist_union_maybe_complement_2nd(ssc->invlist,
1873 invlist,
1874 invert2nd,
1875 &ssc->invlist);
1876}
1877
1878PERL_STATIC_INLINE void
1879S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1880 SV* const invlist,
1881 const bool invert2nd)
1882{
1883 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1884
71068078 1885 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1886
1887 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1888 invlist,
1889 invert2nd,
1890 &ssc->invlist);
1891}
1892
1893PERL_STATIC_INLINE void
1894S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1895{
1896 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1897
71068078 1898 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1899
1900 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1901}
1902
1903PERL_STATIC_INLINE void
1904S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1905{
1906 /* AND just the single code point 'cp' into the SSC 'ssc' */
1907
1908 SV* cp_list = _new_invlist(2);
1909
1910 PERL_ARGS_ASSERT_SSC_CP_AND;
1911
71068078 1912 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1913
1914 cp_list = add_cp_to_invlist(cp_list, cp);
1915 ssc_intersection(ssc, cp_list,
1916 FALSE /* Not inverted */
1917 );
1918 SvREFCNT_dec_NN(cp_list);
1919}
1920
1921PERL_STATIC_INLINE void
dc3bf405 1922S_ssc_clear_locale(regnode_ssc *ssc)
b423522f
KW
1923{
1924 /* Set the SSC 'ssc' to not match any locale things */
b423522f
KW
1925 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1926
71068078 1927 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1928
1929 ANYOF_POSIXL_ZERO(ssc);
1930 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1931}
1932
b35552de
KW
1933#define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1934
1935STATIC bool
1936S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1937{
1938 /* The synthetic start class is used to hopefully quickly winnow down
1939 * places where a pattern could start a match in the target string. If it
1940 * doesn't really narrow things down that much, there isn't much point to
1941 * having the overhead of using it. This function uses some very crude
1942 * heuristics to decide if to use the ssc or not.
1943 *
1944 * It returns TRUE if 'ssc' rules out more than half what it considers to
1945 * be the "likely" possible matches, but of course it doesn't know what the
1946 * actual things being matched are going to be; these are only guesses
1947 *
1948 * For /l matches, it assumes that the only likely matches are going to be
1949 * in the 0-255 range, uniformly distributed, so half of that is 127
1950 * For /a and /d matches, it assumes that the likely matches will be just
1951 * the ASCII range, so half of that is 63
1952 * For /u and there isn't anything matching above the Latin1 range, it
1953 * assumes that that is the only range likely to be matched, and uses
1954 * half that as the cut-off: 127. If anything matches above Latin1,
1955 * it assumes that all of Unicode could match (uniformly), except for
1956 * non-Unicode code points and things in the General Category "Other"
1957 * (unassigned, private use, surrogates, controls and formats). This
1958 * is a much large number. */
1959
b35552de
KW
1960 U32 count = 0; /* Running total of number of code points matched by
1961 'ssc' */
1962 UV start, end; /* Start and end points of current range in inversion
1963 list */
72400949
KW
1964 const U32 max_code_points = (LOC)
1965 ? 256
1966 : (( ! UNI_SEMANTICS
1967 || invlist_highest(ssc->invlist) < 256)
1968 ? 128
1969 : NON_OTHER_COUNT);
1970 const U32 max_match = max_code_points / 2;
b35552de
KW
1971
1972 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1973
1974 invlist_iterinit(ssc->invlist);
1975 while (invlist_iternext(ssc->invlist, &start, &end)) {
72400949
KW
1976 if (start >= max_code_points) {
1977 break;
b35552de 1978 }
72400949 1979 end = MIN(end, max_code_points - 1);
b35552de 1980 count += end - start + 1;
72400949 1981 if (count >= max_match) {
b35552de
KW
1982 invlist_iterfinish(ssc->invlist);
1983 return FALSE;
1984 }
1985 }
1986
1987 return TRUE;
1988}
1989
1990
b423522f
KW
1991STATIC void
1992S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1993{
1994 /* The inversion list in the SSC is marked mortal; now we need a more
1995 * permanent copy, which is stored the same way that is done in a regular
dcb20b36
KW
1996 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1997 * map */
b423522f
KW
1998
1999 SV* invlist = invlist_clone(ssc->invlist);
2000
2001 PERL_ARGS_ASSERT_SSC_FINALIZE;
2002
71068078 2003 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f 2004
a0dd4231 2005 /* The code in this file assumes that all but these flags aren't relevant
93e92956
KW
2006 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2007 * by the time we reach here */
f240c685
KW
2008 assert(! (ANYOF_FLAGS(ssc)
2009 & ~( ANYOF_COMMON_FLAGS
108316fb
KW
2010 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2011 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
a0dd4231 2012
b423522f
KW
2013 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2014
1ee208c4
KW
2015 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
2016 NULL, NULL, NULL, FALSE);
b423522f 2017
85c8e306
KW
2018 /* Make sure is clone-safe */
2019 ssc->invlist = NULL;
2020
e0e1be5f 2021 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
93e92956 2022 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
e0e1be5f 2023 }
1462525b 2024
b2e90ddf
KW
2025 if (RExC_contains_locale) {
2026 OP(ssc) = ANYOFL;
2027 }
2028
1462525b 2029 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
b423522f
KW
2030}
2031
a3621e74
YO
2032#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2033#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2034#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
538e84ed
KW
2035#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2036 ? (TRIE_LIST_CUR( idx ) - 1) \
2037 : 0 )
a3621e74 2038
3dab1dad
YO
2039
2040#ifdef DEBUGGING
07be1b83 2041/*
2b8b4781
NC
2042 dump_trie(trie,widecharmap,revcharmap)
2043 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2044 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
2045
2046 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
2047 The _interim_ variants are used for debugging the interim
2048 tables that are used to generate the final compressed
2049 representation which is what dump_trie expects.
2050
486ec47a 2051 Part of the reason for their existence is to provide a form
3dab1dad 2052 of documentation as to how the different representations function.
07be1b83
YO
2053
2054*/
3dab1dad
YO
2055
2056/*
3dab1dad
YO
2057 Dumps the final compressed table form of the trie to Perl_debug_log.
2058 Used for debugging make_trie().
2059*/
b9a59e08 2060
3dab1dad 2061STATIC void
2b8b4781
NC
2062S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2063 AV *revcharmap, U32 depth)
3dab1dad
YO
2064{
2065 U32 state;
ab3bbdeb 2066 SV *sv=sv_newmortal();
55eed653 2067 int colwidth= widecharmap ? 6 : 4;
2e64971a 2068 U16 word;
3dab1dad
YO
2069 GET_RE_DEBUG_FLAGS_DECL;
2070
7918f24d 2071 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 2072
6ad9a8ab 2073 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
1e37780e 2074 depth+1, "Match","Base","Ofs" );
3dab1dad
YO
2075
2076 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 2077 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 2078 if ( tmp ) {
6ad9a8ab 2079 Perl_re_printf( aTHX_ "%*s",
ab3bbdeb 2080 colwidth,
538e84ed 2081 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
2082 PL_colors[0], PL_colors[1],
2083 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
2084 PERL_PV_ESCAPE_FIRSTCHAR
2085 )
ab3bbdeb 2086 );
3dab1dad
YO
2087 }
2088 }
1e37780e
YO
2089 Perl_re_printf( aTHX_ "\n");
2090 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
3dab1dad
YO
2091
2092 for( state = 0 ; state < trie->uniquecharcount ; state++ )
6ad9a8ab
YO
2093 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2094 Perl_re_printf( aTHX_ "\n");
3dab1dad 2095
1e2e3d02 2096 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 2097 const U32 base = trie->states[ state ].trans.base;
3dab1dad 2098
147e3846 2099 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
3dab1dad
YO
2100
2101 if ( trie->states[ state ].wordnum ) {
1e37780e 2102 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
3dab1dad 2103 } else {
6ad9a8ab 2104 Perl_re_printf( aTHX_ "%6s", "" );
3dab1dad
YO
2105 }
2106
147e3846 2107 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
3dab1dad
YO
2108
2109 if ( base ) {
2110 U32 ofs = 0;
2111
2112 while( ( base + ofs < trie->uniquecharcount ) ||
2113 ( base + ofs - trie->uniquecharcount < trie->lasttrans
538e84ed
KW
2114 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2115 != state))
3dab1dad
YO
2116 ofs++;
2117
147e3846 2118 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
3dab1dad
YO
2119
2120 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
538e84ed
KW
2121 if ( ( base + ofs >= trie->uniquecharcount )
2122 && ( base + ofs - trie->uniquecharcount
2123 < trie->lasttrans )
2124 && trie->trans[ base + ofs
2125 - trie->uniquecharcount ].check == state )
3dab1dad 2126 {
147e3846 2127 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
1e37780e
YO
2128 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2129 );
3dab1dad 2130 } else {
6ad9a8ab 2131 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
3dab1dad
YO
2132 }
2133 }
2134
6ad9a8ab 2135 Perl_re_printf( aTHX_ "]");
3dab1dad
YO
2136
2137 }
6ad9a8ab 2138 Perl_re_printf( aTHX_ "\n" );
3dab1dad 2139 }
6ad9a8ab 2140 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
cb41e5d6 2141 depth);
2e64971a 2142 for (word=1; word <= trie->wordcount; word++) {
6ad9a8ab 2143 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2e64971a
DM
2144 (int)word, (int)(trie->wordinfo[word].prev),
2145 (int)(trie->wordinfo[word].len));
2146 }
6ad9a8ab 2147 Perl_re_printf( aTHX_ "\n" );
538e84ed 2148}
3dab1dad 2149/*
3dab1dad 2150 Dumps a fully constructed but uncompressed trie in list form.
538e84ed 2151 List tries normally only are used for construction when the number of
3dab1dad
YO
2152 possible chars (trie->uniquecharcount) is very high.
2153 Used for debugging make_trie().
2154*/
2155STATIC void
55eed653 2156S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
2157 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2158 U32 depth)
3dab1dad
YO
2159{
2160 U32 state;
ab3bbdeb 2161 SV *sv=sv_newmortal();
55eed653 2162 int colwidth= widecharmap ? 6 : 4;
3dab1dad 2163 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2164
2165 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2166
3dab1dad 2167 /* print out the table precompression. */
6ad9a8ab 2168 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
cb41e5d6 2169 depth+1 );
6ad9a8ab 2170 Perl_re_indentf( aTHX_ "%s",
cb41e5d6 2171 depth+1, "------:-----+-----------------\n" );
538e84ed 2172
3dab1dad
YO
2173 for( state=1 ; state < next_alloc ; state ++ ) {
2174 U16 charid;
538e84ed 2175
147e3846 2176 Perl_re_indentf( aTHX_ " %4" UVXf " :",
cb41e5d6 2177 depth+1, (UV)state );
3dab1dad 2178 if ( ! trie->states[ state ].wordnum ) {
6ad9a8ab 2179 Perl_re_printf( aTHX_ "%5s| ","");
3dab1dad 2180 } else {
6ad9a8ab 2181 Perl_re_printf( aTHX_ "W%4x| ",
3dab1dad
YO
2182 trie->states[ state ].wordnum
2183 );
2184 }
2185 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
538e84ed
KW
2186 SV ** const tmp = av_fetch( revcharmap,
2187 TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb 2188 if ( tmp ) {
147e3846 2189 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
ab3bbdeb 2190 colwidth,
538e84ed
KW
2191 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2192 colwidth,
2193 PL_colors[0], PL_colors[1],
2194 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2195 | PERL_PV_ESCAPE_FIRSTCHAR
ab3bbdeb 2196 ) ,
1e2e3d02
YO
2197 TRIE_LIST_ITEM(state,charid).forid,
2198 (UV)TRIE_LIST_ITEM(state,charid).newstate
2199 );
538e84ed 2200 if (!(charid % 10))
6ad9a8ab 2201 Perl_re_printf( aTHX_ "\n%*s| ",
664e119d 2202 (int)((depth * 2) + 14), "");
1e2e3d02 2203 }
ab3bbdeb 2204 }
6ad9a8ab 2205 Perl_re_printf( aTHX_ "\n");
3dab1dad 2206 }
538e84ed 2207}
3dab1dad
YO
2208
2209/*
3dab1dad 2210 Dumps a fully constructed but uncompressed trie in table form.
538e84ed
KW
2211 This is the normal DFA style state transition table, with a few
2212 twists to facilitate compression later.
3dab1dad
YO
2213 Used for debugging make_trie().
2214*/
2215STATIC void
55eed653 2216S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
2217 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2218 U32 depth)
3dab1dad
YO
2219{
2220 U32 state;
2221 U16 charid;
ab3bbdeb 2222 SV *sv=sv_newmortal();
55eed653 2223 int colwidth= widecharmap ? 6 : 4;
3dab1dad 2224 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2225
2226 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
538e84ed 2227
3dab1dad
YO
2228 /*
2229 print out the table precompression so that we can do a visual check
2230 that they are identical.
2231 */
538e84ed 2232
6ad9a8ab 2233 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
3dab1dad
YO
2234
2235 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 2236 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 2237 if ( tmp ) {
6ad9a8ab 2238 Perl_re_printf( aTHX_ "%*s",
ab3bbdeb 2239 colwidth,
538e84ed 2240 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
2241 PL_colors[0], PL_colors[1],
2242 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
2243 PERL_PV_ESCAPE_FIRSTCHAR
2244 )
ab3bbdeb 2245 );
3dab1dad
YO
2246 }
2247 }
2248
4aaafc03
YO
2249 Perl_re_printf( aTHX_ "\n");
2250 Perl_re_indentf( aTHX_ "State+-", depth+1 );
3dab1dad
YO
2251
2252 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
6ad9a8ab 2253 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
3dab1dad
YO
2254 }
2255
6ad9a8ab 2256 Perl_re_printf( aTHX_ "\n" );
3dab1dad
YO
2257
2258 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2259
147e3846 2260 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
cb41e5d6 2261 depth+1,
3dab1dad
YO
2262 (UV)TRIE_NODENUM( state ) );
2263
2264 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
2265 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2266 if (v)
147e3846 2267 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
ab3bbdeb 2268 else
6ad9a8ab 2269 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
3dab1dad
YO
2270 }
2271 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
147e3846 2272 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
538e84ed 2273 (UV)trie->trans[ state ].check );
3dab1dad 2274 } else {
147e3846 2275 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
538e84ed 2276 (UV)trie->trans[ state ].check,
3dab1dad
YO
2277 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2278 }
2279 }
07be1b83 2280}
3dab1dad
YO
2281
2282#endif
2283
2e64971a 2284
786e8c11
YO
2285/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2286 startbranch: the first branch in the whole branch sequence
2287 first : start branch of sequence of branch-exact nodes.
2288 May be the same as startbranch
2289 last : Thing following the last branch.
2290 May be the same as tail.
2291 tail : item following the branch sequence
2292 count : words in the sequence
a4525e78 2293 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
786e8c11 2294 depth : indent depth
3dab1dad 2295
786e8c11 2296Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 2297
786e8c11
YO
2298A trie is an N'ary tree where the branches are determined by digital
2299decomposition of the key. IE, at the root node you look up the 1st character and
2300follow that branch repeat until you find the end of the branches. Nodes can be
2301marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 2302
786e8c11 2303 /he|she|his|hers/
72f13be8 2304
786e8c11
YO
2305would convert into the following structure. Numbers represent states, letters
2306following numbers represent valid transitions on the letter from that state, if
2307the number is in square brackets it represents an accepting state, otherwise it
2308will be in parenthesis.
07be1b83 2309
786e8c11
YO
2310 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2311 | |
2312 | (2)
2313 | |
2314 (1) +-i->(6)-+-s->[7]
2315 |
2316 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 2317
786e8c11
YO
2318 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2319
2320This shows that when matching against the string 'hers' we will begin at state 1
2321read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2322then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2323is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2324single traverse. We store a mapping from accepting to state to which word was
2325matched, and then when we have multiple possibilities we try to complete the
b8fda935 2326rest of the regex in the order in which they occurred in the alternation.
786e8c11
YO
2327
2328The only prior NFA like behaviour that would be changed by the TRIE support is
2329the silent ignoring of duplicate alternations which are of the form:
2330
2331 / (DUPE|DUPE) X? (?{ ... }) Y /x
2332
4b714af6 2333Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 2334and without the optimisation. With the optimisations dupes will be silently
486ec47a 2335ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
2336the following demonstrates:
2337
2338 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2339
2340which prints out 'word' three times, but
2341
2342 'words'=~/(word|word|word)(?{ print $1 })S/
2343
2344which doesnt print it out at all. This is due to other optimisations kicking in.
2345
2346Example of what happens on a structural level:
2347
486ec47a 2348The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
2349
2350 1: CURLYM[1] {1,32767}(18)
2351 5: BRANCH(8)
2352 6: EXACT <ac>(16)
2353 8: BRANCH(11)
2354 9: EXACT <ad>(16)
2355 11: BRANCH(14)
2356 12: EXACT <ab>(16)
2357 16: SUCCEED(0)
2358 17: NOTHING(18)
2359 18: END(0)
2360
2361This would be optimizable with startbranch=5, first=5, last=16, tail=16
2362and should turn into:
2363
2364 1: CURLYM[1] {1,32767}(18)
2365 5: TRIE(16)
2366 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2367 <ac>
2368 <ad>
2369 <ab>
2370 16: SUCCEED(0)
2371 17: NOTHING(18)
2372 18: END(0)
2373
2374Cases where tail != last would be like /(?foo|bar)baz/:
2375
2376 1: BRANCH(4)
2377 2: EXACT <foo>(8)
2378 4: BRANCH(7)
2379 5: EXACT <bar>(8)
2380 7: TAIL(8)
2381 8: EXACT <baz>(10)
2382 10: END(0)
2383
2384which would be optimizable with startbranch=1, first=1, last=7, tail=8
2385and would end up looking like:
2386
2387 1: TRIE(8)
2388 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2389 <foo>
2390 <bar>
2391 7: TAIL(8)
2392 8: EXACT <baz>(10)
2393 10: END(0)
2394
c80e42f3 2395 d = uvchr_to_utf8_flags(d, uv, 0);
786e8c11
YO
2396
2397is the recommended Unicode-aware way of saying
2398
2399 *(d++) = uv;
2400*/
2401
fab2782b 2402#define TRIE_STORE_REVCHAR(val) \
786e8c11 2403 STMT_START { \
73031816 2404 if (UTF) { \
668fcfea 2405 SV *zlopp = newSV(UTF8_MAXBYTES); \
88c9ea1e 2406 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
c80e42f3 2407 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
73031816
NC
2408 SvCUR_set(zlopp, kapow - flrbbbbb); \
2409 SvPOK_on(zlopp); \
2410 SvUTF8_on(zlopp); \
2411 av_push(revcharmap, zlopp); \
2412 } else { \
fab2782b 2413 char ooooff = (char)val; \
73031816
NC
2414 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2415 } \
2416 } STMT_END
786e8c11 2417
914a25d5
KW
2418/* This gets the next character from the input, folding it if not already
2419 * folded. */
2420#define TRIE_READ_CHAR STMT_START { \
2421 wordlen++; \
2422 if ( UTF ) { \
2423 /* if it is UTF then it is either already folded, or does not need \
2424 * folding */ \
1c1d615a 2425 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
914a25d5
KW
2426 } \
2427 else if (folder == PL_fold_latin1) { \
7d006b13
KW
2428 /* This folder implies Unicode rules, which in the range expressible \
2429 * by not UTF is the lower case, with the two exceptions, one of \
2430 * which should have been taken care of before calling this */ \
2431 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2432 uvc = toLOWER_L1(*uc); \
2433 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2434 len = 1; \
914a25d5
KW
2435 } else { \
2436 /* raw data, will be folded later if needed */ \
2437 uvc = (U32)*uc; \
2438 len = 1; \
2439 } \
786e8c11
YO
2440} STMT_END
2441
2442
2443
2444#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2445 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
00195859 2446 U32 ging = TRIE_LIST_LEN( state ) * 2; \
f9003953 2447 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
00195859 2448 TRIE_LIST_LEN( state ) = ging; \
786e8c11
YO
2449 } \
2450 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2451 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2452 TRIE_LIST_CUR( state )++; \
2453} STMT_END
07be1b83 2454
786e8c11 2455#define TRIE_LIST_NEW(state) STMT_START { \
d09f14bf 2456 Newx( trie->states[ state ].trans.list, \
786e8c11
YO
2457 4, reg_trie_trans_le ); \
2458 TRIE_LIST_CUR( state ) = 1; \
2459 TRIE_LIST_LEN( state ) = 4; \
2460} STMT_END
07be1b83 2461
786e8c11
YO
2462#define TRIE_HANDLE_WORD(state) STMT_START { \
2463 U16 dupe= trie->states[ state ].wordnum; \
2464 regnode * const noper_next = regnext( noper ); \
2465 \
786e8c11
YO
2466 DEBUG_r({ \
2467 /* store the word for dumping */ \
2468 SV* tmp; \
2469 if (OP(noper) != NOTHING) \
740cce10 2470 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 2471 else \
740cce10 2472 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 2473 av_push( trie_words, tmp ); \
786e8c11
YO
2474 }); \
2475 \
2476 curword++; \
2e64971a
DM
2477 trie->wordinfo[curword].prev = 0; \
2478 trie->wordinfo[curword].len = wordlen; \
2479 trie->wordinfo[curword].accept = state; \
786e8c11
YO
2480 \
2481 if ( noper_next < tail ) { \
2482 if (!trie->jump) \
538e84ed
KW
2483 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2484 sizeof(U16) ); \
7f69552c 2485 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
2486 if (!jumper) \
2487 jumper = noper_next; \
2488 if (!nextbranch) \
2489 nextbranch= regnext(cur); \
2490 } \
2491 \
2492 if ( dupe ) { \
2e64971a
DM
2493 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2494 /* chain, so that when the bits of chain are later */\
2495 /* linked together, the dups appear in the chain */\
2496 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2497 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
2498 } else { \
2499 /* we haven't inserted this word yet. */ \
2500 trie->states[ state ].wordnum = curword; \
2501 } \
2502} STMT_END
07be1b83 2503
3dab1dad 2504
786e8c11
YO
2505#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2506 ( ( base + charid >= ucharcount \
2507 && base + charid < ubound \
2508 && state == trie->trans[ base - ucharcount + charid ].check \
2509 && trie->trans[ base - ucharcount + charid ].next ) \
2510 ? trie->trans[ base - ucharcount + charid ].next \
2511 : ( state==1 ? special : 0 ) \
2512 )
3dab1dad 2513
8bcafbf4
YO
2514#define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2515STMT_START { \
2516 TRIE_BITMAP_SET(trie, uvc); \
2517 /* store the folded codepoint */ \
2518 if ( folder ) \
2519 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2520 \
2521 if ( !UTF ) { \
2522 /* store first byte of utf8 representation of */ \
2523 /* variant codepoints */ \
2524 if (! UVCHR_IS_INVARIANT(uvc)) { \
2525 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2526 } \
2527 } \
2528} STMT_END
786e8c11
YO
2529#define MADE_TRIE 1
2530#define MADE_JUMP_TRIE 2
2531#define MADE_EXACT_TRIE 4
3dab1dad 2532
a3621e74 2533STATIC I32
538e84ed
KW
2534S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2535 regnode *first, regnode *last, regnode *tail,
2536 U32 word_count, U32 flags, U32 depth)
a3621e74
YO
2537{
2538 /* first pass, loop through and scan words */
2539 reg_trie_data *trie;
55eed653 2540 HV *widecharmap = NULL;
2b8b4781 2541 AV *revcharmap = newAV();
a3621e74 2542 regnode *cur;
a3621e74
YO
2543 STRLEN len = 0;
2544 UV uvc = 0;
2545 U16 curword = 0;
2546 U32 next_alloc = 0;
786e8c11
YO
2547 regnode *jumper = NULL;
2548 regnode *nextbranch = NULL;
7f69552c 2549 regnode *convert = NULL;
2e64971a 2550 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 2551 /* we just use folder as a flag in utf8 */
1e696034 2552 const U8 * folder = NULL;
a3621e74 2553
3a611511
YO
2554 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2555 * which stands for one trie structure, one hash, optionally followed
2556 * by two arrays */
2b8b4781 2557#ifdef DEBUGGING
3a611511 2558 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2b8b4781
NC
2559 AV *trie_words = NULL;
2560 /* along with revcharmap, this only used during construction but both are
2561 * useful during debugging so we store them in the struct when debugging.
8e11feef 2562 */
2b8b4781 2563#else
cf78de0b 2564 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
3dab1dad 2565 STRLEN trie_charcount=0;
3dab1dad 2566#endif
2b8b4781 2567 SV *re_trie_maxbuff;
a3621e74 2568 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2569
2570 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
2571#ifndef DEBUGGING
2572 PERL_UNUSED_ARG(depth);
2573#endif
a3621e74 2574
1e696034 2575 switch (flags) {
a4525e78 2576 case EXACT: case EXACTL: break;
89829bb5 2577 case EXACTFAA:
fab2782b 2578 case EXACTFU_SS:
a4525e78
KW
2579 case EXACTFU:
2580 case EXACTFLU8: folder = PL_fold_latin1; break;
1e696034 2581 case EXACTF: folder = PL_fold; break;
fab2782b 2582 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
2583 }
2584
c944940b 2585 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 2586 trie->refcount = 1;
3dab1dad 2587 trie->startstate = 1;
786e8c11 2588 trie->wordcount = word_count;
f8fc2ecf 2589 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 2590 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
a4525e78 2591 if (flags == EXACT || flags == EXACTL)
c944940b 2592 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
2593 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2594 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2595
a3621e74 2596 DEBUG_r({
2b8b4781 2597 trie_words = newAV();
a3621e74 2598 });
a3621e74 2599
0111c4fd 2600 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
316ebaf2 2601 assert(re_trie_maxbuff);
a3621e74 2602 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 2603 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 2604 }
df826430 2605 DEBUG_TRIE_COMPILE_r({
6ad9a8ab 2606 Perl_re_indentf( aTHX_
cb41e5d6
YO
2607 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2608 depth+1,
538e84ed
KW
2609 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2610 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
3dab1dad 2611 });
538e84ed 2612
7f69552c
YO
2613 /* Find the node we are going to overwrite */
2614 if ( first == startbranch && OP( last ) != BRANCH ) {
2615 /* whole branch chain */
2616 convert = first;
2617 } else {
2618 /* branch sub-chain */
2619 convert = NEXTOPER( first );
2620 }
538e84ed 2621
a3621e74
YO
2622 /* -- First loop and Setup --
2623
2624 We first traverse the branches and scan each word to determine if it
2625 contains widechars, and how many unique chars there are, this is
2626 important as we have to build a table with at least as many columns as we
2627 have unique chars.
2628
2629 We use an array of integers to represent the character codes 0..255
538e84ed
KW
2630 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2631 the native representation of the character value as the key and IV's for
2632 the coded index.
a3621e74
YO
2633
2634 *TODO* If we keep track of how many times each character is used we can
2635 remap the columns so that the table compression later on is more
3b753521 2636 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
2637 middle and the least common are on the outside. IMO this would be better
2638 than a most to least common mapping as theres a decent chance the most
2639 common letter will share a node with the least common, meaning the node
486ec47a 2640 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
2641 case is when we have the least common nodes twice.
2642
2643 */
2644
a3621e74 2645 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 2646 regnode *noper = NEXTOPER( cur );
944e05e3
YO
2647 const U8 *uc;
2648 const U8 *e;
bc031a7d 2649 int foldlen = 0;
07be1b83 2650 U32 wordlen = 0; /* required init */
bc031a7d
KW
2651 STRLEN minchars = 0;
2652 STRLEN maxchars = 0;
538e84ed
KW
2653 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2654 bitmap?*/
a3621e74 2655
3dab1dad 2656 if (OP(noper) == NOTHING) {
20ed8c88
YO
2657 /* skip past a NOTHING at the start of an alternation
2658 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2659 */
df826430 2660 regnode *noper_next= regnext(noper);
944e05e3
YO
2661 if (noper_next < tail)
2662 noper= noper_next;
2663 }
2664
dca5fc4c
YO
2665 if ( noper < tail &&
2666 (
2667 OP(noper) == flags ||
2668 (
2669 flags == EXACTFU &&
2670 OP(noper) == EXACTFU_SS
2671 )
2672 )
2673 ) {
944e05e3
YO
2674 uc= (U8*)STRING(noper);
2675 e= uc + STR_LEN(noper);
2676 } else {
2677 trie->minlen= 0;
2678 continue;
3dab1dad 2679 }
df826430 2680
944e05e3 2681
fab2782b 2682 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
2683 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2684 regardless of encoding */
fab2782b
YO
2685 if (OP( noper ) == EXACTFU_SS) {
2686 /* false positives are ok, so just set this */
0dc4a61d 2687 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
fab2782b
YO
2688 }
2689 }
dca5fc4c 2690
bc031a7d
KW
2691 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2692 branch */
3dab1dad 2693 TRIE_CHARCOUNT(trie)++;
a3621e74 2694 TRIE_READ_CHAR;
645de4ce 2695
bc031a7d
KW
2696 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2697 * is in effect. Under /i, this character can match itself, or
2698 * anything that folds to it. If not under /i, it can match just
2699 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2700 * all fold to k, and all are single characters. But some folds
2701 * expand to more than one character, so for example LATIN SMALL
2702 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2703 * the string beginning at 'uc' is 'ffi', it could be matched by
2704 * three characters, or just by the one ligature character. (It
2705 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2706 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2707 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2708 * match.) The trie needs to know the minimum and maximum number
2709 * of characters that could match so that it can use size alone to
2710 * quickly reject many match attempts. The max is simple: it is
2711 * the number of folded characters in this branch (since a fold is
2712 * never shorter than what folds to it. */
2713
2714 maxchars++;
2715
2716 /* And the min is equal to the max if not under /i (indicated by
2717 * 'folder' being NULL), or there are no multi-character folds. If
2718 * there is a multi-character fold, the min is incremented just
2719 * once, for the character that folds to the sequence. Each
2720 * character in the sequence needs to be added to the list below of
2721 * characters in the trie, but we count only the first towards the
2722 * min number of characters needed. This is done through the
2723 * variable 'foldlen', which is returned by the macros that look
2724 * for these sequences as the number of bytes the sequence
2725 * occupies. Each time through the loop, we decrement 'foldlen' by
2726 * how many bytes the current char occupies. Only when it reaches
2727 * 0 do we increment 'minchars' or look for another multi-character
2728 * sequence. */
2729 if (folder == NULL) {
2730 minchars++;
2731 }
2732 else if (foldlen > 0) {
2733 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
645de4ce
KW
2734 }
2735 else {
bc031a7d
KW
2736 minchars++;
2737
2738 /* See if *uc is the beginning of a multi-character fold. If
2739 * so, we decrement the length remaining to look at, to account
2740 * for the current character this iteration. (We can use 'uc'
2741 * instead of the fold returned by TRIE_READ_CHAR because for
2742 * non-UTF, the latin1_safe macro is smart enough to account
2743 * for all the unfolded characters, and because for UTF, the
2744 * string will already have been folded earlier in the
2745 * compilation process */
2746 if (UTF) {
2747 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2748 foldlen -= UTF8SKIP(uc);
645de4ce
KW
2749 }
2750 }
bc031a7d
KW
2751 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2752 foldlen--;
2753 }
645de4ce 2754 }
bc031a7d
KW
2755
2756 /* The current character (and any potential folds) should be added
2757 * to the possible matching characters for this position in this
2758 * branch */
a3621e74 2759 if ( uvc < 256 ) {
fab2782b
YO
2760 if ( folder ) {
2761 U8 folded= folder[ (U8) uvc ];
2762 if ( !trie->charmap[ folded ] ) {
2763 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2764 TRIE_STORE_REVCHAR( folded );
2765 }
2766 }
a3621e74
YO
2767 if ( !trie->charmap[ uvc ] ) {
2768 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 2769 TRIE_STORE_REVCHAR( uvc );
a3621e74 2770 }
02daf0ab 2771 if ( set_bit ) {
62012aee
KW
2772 /* store the codepoint in the bitmap, and its folded
2773 * equivalent. */
8bcafbf4 2774 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
02daf0ab
YO
2775 set_bit = 0; /* We've done our bit :-) */
2776 }
a3621e74 2777 } else {
bc031a7d
KW
2778
2779 /* XXX We could come up with the list of code points that fold
2780 * to this using PL_utf8_foldclosures, except not for
2781 * multi-char folds, as there may be multiple combinations
2782 * there that could work, which needs to wait until runtime to
2783 * resolve (The comment about LIGATURE FFI above is such an
2784 * example */
2785
a3621e74 2786 SV** svpp;
55eed653
NC
2787 if ( !widecharmap )
2788 widecharmap = newHV();
a3621e74 2789
55eed653 2790 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
2791
2792 if ( !svpp )
147e3846 2793 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
a3621e74
YO
2794
2795 if ( !SvTRUE( *svpp ) ) {
2796 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 2797 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
2798 }
2799 }
bc031a7d
KW
2800 } /* end loop through characters in this branch of the trie */
2801
2802 /* We take the min and max for this branch and combine to find the min
2803 * and max for all branches processed so far */
3dab1dad 2804 if( cur == first ) {
bc031a7d
KW
2805 trie->minlen = minchars;
2806 trie->maxlen = maxchars;
2807 } else if (minchars < trie->minlen) {
2808 trie->minlen = minchars;
2809 } else if (maxchars > trie->maxlen) {
2810 trie->maxlen = maxchars;
fab2782b 2811 }
a3621e74
YO
2812 } /* end first pass */
2813 DEBUG_TRIE_COMPILE_r(
6ad9a8ab 2814 Perl_re_indentf( aTHX_
cb41e5d6
YO
2815 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2816 depth+1,
55eed653 2817 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
2818 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2819 (int)trie->minlen, (int)trie->maxlen )
a3621e74 2820 );
a3621e74
YO
2821
2822 /*
2823 We now know what we are dealing with in terms of unique chars and
2824 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
2825 representation using a flat table will take. If it's over a reasonable
2826 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
2827 conservative but potentially much slower representation using an array
2828 of lists.
2829
2830 At the end we convert both representations into the same compressed
2831 form that will be used in regexec.c for matching with. The latter
2832 is a form that cannot be used to construct with but has memory
2833 properties similar to the list form and access properties similar
2834 to the table form making it both suitable for fast searches and
2835 small enough that its feasable to store for the duration of a program.
2836
2837 See the comment in the code where the compressed table is produced
2838 inplace from the flat tabe representation for an explanation of how
2839 the compression works.
2840
2841 */
2842
2843
2e64971a
DM
2844 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2845 prev_states[1] = 0;
2846
538e84ed
KW
2847 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2848 > SvIV(re_trie_maxbuff) )
2849 {
a3621e74
YO
2850 /*
2851 Second Pass -- Array Of Lists Representation
2852
2853 Each state will be represented by a list of charid:state records
2854 (reg_trie_trans_le) the first such element holds the CUR and LEN
2855 points of the allocated array. (See defines above).
2856
2857 We build the initial structure using the lists, and then convert
2858 it into the compressed table form which allows faster lookups
2859 (but cant be modified once converted).
a3621e74
YO
2860 */
2861
a3621e74
YO
2862 STRLEN transcount = 1;
2863
6ad9a8ab 2864 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
cb41e5d6 2865 depth+1));
686b73d4 2866
c944940b
JH
2867 trie->states = (reg_trie_state *)
2868 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2869 sizeof(reg_trie_state) );
a3621e74
YO
2870 TRIE_LIST_NEW(1);
2871 next_alloc = 2;
2872
2873 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2874
df826430 2875 regnode *noper = NEXTOPER( cur );
c445ea15
AL
2876 U32 state = 1; /* required init */
2877 U16 charid = 0; /* sanity init */
07be1b83 2878 U32 wordlen = 0; /* required init */
c445ea15 2879
df826430
YO
2880 if (OP(noper) == NOTHING) {
2881 regnode *noper_next= regnext(noper);
944e05e3
YO
2882 if (noper_next < tail)
2883 noper= noper_next;
df826430
YO
2884 }
2885
944e05e3
YO
2886 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2887 const U8 *uc= (U8*)STRING(noper);
2888 const U8 *e= uc + STR_LEN(noper);
2889
786e8c11 2890 for ( ; uc < e ; uc += len ) {
c445ea15 2891
786e8c11 2892 TRIE_READ_CHAR;
c445ea15 2893
786e8c11
YO
2894 if ( uvc < 256 ) {
2895 charid = trie->charmap[ uvc ];
c445ea15 2896 } else {
538e84ed
KW
2897 SV** const svpp = hv_fetch( widecharmap,
2898 (char*)&uvc,
2899 sizeof( UV ),
2900 0);
786e8c11
YO
2901 if ( !svpp ) {
2902 charid = 0;
2903 } else {
2904 charid=(U16)SvIV( *svpp );
2905 }
c445ea15 2906 }
538e84ed
KW
2907 /* charid is now 0 if we dont know the char read, or
2908 * nonzero if we do */
786e8c11 2909 if ( charid ) {
a3621e74 2910
786e8c11
YO
2911 U16 check;
2912 U32 newstate = 0;
a3621e74 2913
786e8c11
YO
2914 charid--;
2915 if ( !trie->states[ state ].trans.list ) {
2916 TRIE_LIST_NEW( state );
c445ea15 2917 }
538e84ed
KW
2918 for ( check = 1;
2919 check <= TRIE_LIST_USED( state );
2920 check++ )
2921 {
2922 if ( TRIE_LIST_ITEM( state, check ).forid
2923 == charid )
2924 {
786e8c11
YO
2925 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2926 break;
2927 }
2928 }
2929 if ( ! newstate ) {
2930 newstate = next_alloc++;
2e64971a 2931 prev_states[newstate] = state;
786e8c11
YO
2932 TRIE_LIST_PUSH( state, charid, newstate );
2933 transcount++;
2934 }
2935 state = newstate;
2936 } else {
147e3846 2937 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
c445ea15 2938 }
a28509cc 2939 }
c445ea15 2940 }
3dab1dad 2941 TRIE_HANDLE_WORD(state);
a3621e74
YO
2942
2943 } /* end second pass */
2944
1e2e3d02 2945 /* next alloc is the NEXT state to be allocated */
538e84ed 2946 trie->statecount = next_alloc;
c944940b
JH
2947 trie->states = (reg_trie_state *)
2948 PerlMemShared_realloc( trie->states,
2949 next_alloc
2950 * sizeof(reg_trie_state) );
a3621e74 2951
3dab1dad 2952 /* and now dump it out before we compress it */
2b8b4781
NC
2953 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2954 revcharmap, next_alloc,
2955 depth+1)
1e2e3d02 2956 );
a3621e74 2957
c944940b
JH
2958 trie->trans = (reg_trie_trans *)
2959 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
2960 {
2961 U32 state;
a3621e74
YO
2962 U32 tp = 0;
2963 U32 zp = 0;
2964
2965
2966 for( state=1 ; state < next_alloc ; state ++ ) {
2967 U32 base=0;
2968
2969 /*
2970 DEBUG_TRIE_COMPILE_MORE_r(
6ad9a8ab 2971 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
a3621e74
YO
2972 );
2973 */
2974
2975 if (trie->states[state].trans.list) {
2976 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2977 U16 maxid=minid;
a28509cc 2978 U16 idx;
a3621e74
YO
2979
2980 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
2981 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2982 if ( forid < minid ) {
2983 minid=forid;
2984 } else if ( forid > maxid ) {
2985 maxid=forid;
2986 }
a3621e74
YO
2987 }
2988 if ( transcount < tp + maxid - minid + 1) {
2989 transcount *= 2;
c944940b
JH
2990 trie->trans = (reg_trie_trans *)
2991 PerlMemShared_realloc( trie->trans,
446bd890
NC
2992 transcount
2993 * sizeof(reg_trie_trans) );
538e84ed
KW
2994 Zero( trie->trans + (transcount / 2),
2995 transcount / 2,
2996 reg_trie_trans );
a3621e74
YO
2997 }
2998 base = trie->uniquecharcount + tp - minid;
2999 if ( maxid == minid ) {
3000 U32 set = 0;
3001 for ( ; zp < tp ; zp++ ) {
3002 if ( ! trie->trans[ zp ].next ) {
3003 base = trie->uniquecharcount + zp - minid;
538e84ed
KW
3004 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3005 1).newstate;
a3621e74
YO
3006 trie->trans[ zp ].check = state;
3007 set = 1;
3008 break;
3009 }
3010 }
3011 if ( !set ) {
538e84ed
KW
3012 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3013 1).newstate;
a3621e74
YO
3014 trie->trans[ tp ].check = state;
3015 tp++;
3016 zp = tp;
3017 }
3018 } else {
3019 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
538e84ed
KW
3020 const U32 tid = base
3021 - trie->uniquecharcount
3022 + TRIE_LIST_ITEM( state, idx ).forid;
3023 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3024 idx ).newstate;
a3621e74
YO
3025 trie->trans[ tid ].check = state;
3026 }
3027 tp += ( maxid - minid + 1 );
3028 }
3029 Safefree(trie->states[ state ].trans.list);
3030 }
3031 /*
3032 DEBUG_TRIE_COMPILE_MORE_r(
6ad9a8ab 3033 Perl_re_printf( aTHX_ " base: %d\n",base);
a3621e74
YO
3034 );
3035 */
3036 trie->states[ state ].trans.base=base;
3037 }
cc601c31 3038 trie->lasttrans = tp + 1;
a3621e74
YO
3039 }
3040 } else {
3041 /*
3042 Second Pass -- Flat Table Representation.
3043
b423522f
KW
3044 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3045 each. We know that we will need Charcount+1 trans at most to store
3046 the data (one row per char at worst case) So we preallocate both
3047 structures assuming worst case.
a3621e74
YO
3048
3049 We then construct the trie using only the .next slots of the entry
3050 structs.
3051
b423522f
KW
3052 We use the .check field of the first entry of the node temporarily
3053 to make compression both faster and easier by keeping track of how
3054 many non zero fields are in the node.
a3621e74
YO
3055
3056 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3057 transition.
3058
b423522f
KW
3059 There are two terms at use here: state as a TRIE_NODEIDX() which is
3060 a number representing the first entry of the node, and state as a
3061 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3062 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3063 if there are 2 entrys per node. eg:
a3621e74
YO
3064
3065 A B A B
3066 1. 2 4 1. 3 7
3067 2. 0 3 3. 0 5
3068 3. 0 0 5. 0 0
3069 4. 0 0 7. 0 0
3070
b423522f
KW
3071 The table is internally in the right hand, idx form. However as we
3072 also have to deal with the states array which is indexed by nodenum
3073 we have to use TRIE_NODENUM() to convert.
a3621e74
YO
3074
3075 */
6ad9a8ab 3076 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
cb41e5d6 3077 depth+1));
3dab1dad 3078
c944940b
JH
3079 trie->trans = (reg_trie_trans *)
3080 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3081 * trie->uniquecharcount + 1,
3082 sizeof(reg_trie_trans) );
3083 trie->states = (reg_trie_state *)
3084 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3085 sizeof(reg_trie_state) );
a3621e74
YO
3086 next_alloc = trie->uniquecharcount + 1;
3087
3dab1dad 3088
a3621e74
YO
3089 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3090
df826430 3091 regnode *noper = NEXTOPER( cur );
a3621e74
YO
3092
3093 U32 state = 1; /* required init */
3094
3095 U16 charid = 0; /* sanity init */
3096 U32 accept_state = 0; /* sanity init */
a3621e74 3097
07be1b83 3098 U32 wordlen = 0; /* required init */
a3621e74 3099
df826430
YO
3100 if (OP(noper) == NOTHING) {
3101 regnode *noper_next= regnext(noper);
944e05e3
YO
3102 if (noper_next < tail)
3103 noper= noper_next;
df826430 3104 }
fab2782b 3105
944e05e3
YO
3106 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3107 const U8 *uc= (U8*)STRING(noper);
3108 const U8 *e= uc + STR_LEN(noper);
3109
786e8c11 3110 for ( ; uc < e ; uc += len ) {
a3621e74 3111
786e8c11 3112 TRIE_READ_CHAR;
a3621e74 3113
786e8c11
YO
3114 if ( uvc < 256 ) {
3115 charid = trie->charmap[ uvc ];
3116 } else {
538e84ed
KW
3117 SV* const * const svpp = hv_fetch( widecharmap,
3118 (char*)&uvc,
3119 sizeof( UV ),
3120 0);
786e8c11 3121 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 3122 }
786e8c11
YO
3123 if ( charid ) {
3124 charid--;
3125 if ( !trie->trans[ state + charid ].next ) {
3126 trie->trans[ state + charid ].next = next_alloc;
3127 trie->trans[ state ].check++;
2e64971a
DM
3128 prev_states[TRIE_NODENUM(next_alloc)]
3129 = TRIE_NODENUM(state);
786e8c11
YO
3130 next_alloc += trie->uniquecharcount;
3131 }
3132 state = trie->trans[ state + charid ].next;
3133 } else {
147e3846 3134 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
786e8c11 3135 }
538e84ed
KW
3136 /* charid is now 0 if we dont know the char read, or
3137 * nonzero if we do */
a3621e74 3138 }
a3621e74 3139 }
3dab1dad
YO
3140 accept_state = TRIE_NODENUM( state );
3141 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
3142
3143 } /* end second pass */
3144
3dab1dad 3145 /* and now dump it out before we compress it */
2b8b4781
NC
3146 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3147 revcharmap,
3148 next_alloc, depth+1));
a3621e74 3149
a3621e74
YO
3150 {
3151 /*
3152 * Inplace compress the table.*
3153
3154 For sparse data sets the table constructed by the trie algorithm will
3155 be mostly 0/FAIL transitions or to put it another way mostly empty.
3156 (Note that leaf nodes will not contain any transitions.)
3157
3158 This algorithm compresses the tables by eliminating most such
3159 transitions, at the cost of a modest bit of extra work during lookup:
3160
3161 - Each states[] entry contains a .base field which indicates the
3162 index in the state[] array wheres its transition data is stored.
3163
3b753521 3164 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
3165
3166 - If .base is nonzero then charid is added to it to find an entry in
3167 the trans array.
3168
3169 -If trans[states[state].base+charid].check!=state then the
3170 transition is taken to be a 0/Fail transition. Thus if there are fail
3171 transitions at the front of the node then the .base offset will point
3172 somewhere inside the previous nodes data (or maybe even into a node
3173 even earlier), but the .check field determines if the transition is
3174 valid.
3175
786e8c11 3176 XXX - wrong maybe?
a3621e74 3177 The following process inplace converts the table to the compressed
3b753521 3178 table: We first do not compress the root node 1,and mark all its
a3621e74 3179 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
3180 allows us to do a DFA construction from the compressed table later,
3181 and ensures that any .base pointers we calculate later are greater
3182 than 0.
a3621e74
YO
3183
3184 - We set 'pos' to indicate the first entry of the second node.
3185
3186 - We then iterate over the columns of the node, finding the first and
3187 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3188 and set the .check pointers accordingly, and advance pos
3189 appropriately and repreat for the next node. Note that when we copy
3190 the next pointers we have to convert them from the original
3191 NODEIDX form to NODENUM form as the former is not valid post
3192 compression.
3193
3194 - If a node has no transitions used we mark its base as 0 and do not
3195 advance the pos pointer.
3196
3197 - If a node only has one transition we use a second pointer into the
3198 structure to fill in allocated fail transitions from other states.
3199 This pointer is independent of the main pointer and scans forward
3200 looking for null transitions that are allocated to a state. When it
3201 finds one it writes the single transition into the "hole". If the
786e8c11 3202 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
3203
3204 - Once compressed we can Renew/realloc the structures to release the
3205 excess space.
3206
3207 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3208 specifically Fig 3.47 and the associated pseudocode.
3209
3210 demq
3211 */
a3b680e6 3212 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 3213 U32 state, charid;
a3621e74 3214 U32 pos = 0, zp=0;
1e2e3d02 3215 trie->statecount = laststate;
a3621e74
YO
3216
3217 for ( state = 1 ; state < laststate ; state++ ) {
3218 U8 flag = 0;
a28509cc
AL
3219 const U32 stateidx = TRIE_NODEIDX( state );
3220 const U32 o_used = trie->trans[ stateidx ].check;
3221 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
3222 trie->trans[ stateidx ].check = 0;
3223
538e84ed
KW
3224 for ( charid = 0;
3225 used && charid < trie->uniquecharcount;
3226 charid++ )
3227 {
a3621e74
YO
3228 if ( flag || trie->trans[ stateidx + charid ].next ) {
3229 if ( trie->trans[ stateidx + charid ].next ) {
3230 if (o_used == 1) {
3231 for ( ; zp < pos ; zp++ ) {
3232 if ( ! trie->trans[ zp ].next ) {
3233 break;
3234 }
3235 }
538e84ed
KW
3236 trie->states[ state ].trans.base
3237 = zp
3238 + trie->uniquecharcount
3239 - charid ;
3240 trie->trans[ zp ].next
3241 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3242 + charid ].next );
a3621e74
YO
3243 trie->trans[ zp ].check = state;
3244 if ( ++zp > pos ) pos = zp;
3245 break;
3246 }
3247 used--;
3248 }
3249 if ( !flag ) {
3250 flag = 1;
538e84ed
KW
3251 trie->states[ state ].trans.base
3252 = pos + trie->uniquecharcount - charid ;
a3621e74 3253 }
538e84ed
KW
3254 trie->trans[ pos ].next
3255 = SAFE_TRIE_NODENUM(
3256 trie->trans[ stateidx + charid ].next );
a3621e74
YO
3257 trie->trans[ pos ].check = state;
3258 pos++;
3259 }
3260 }
3261 }
cc601c31 3262 trie->lasttrans = pos + 1;
c944940b
JH
3263 trie->states = (reg_trie_state *)
3264 PerlMemShared_realloc( trie->states, laststate
3265 * sizeof(reg_trie_state) );
a3621e74 3266 DEBUG_TRIE_COMPILE_MORE_r(
147e3846 3267 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
cb41e5d6 3268 depth+1,
538e84ed
KW
3269 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3270 + 1 ),
3271 (IV)next_alloc,
3272 (IV)pos,
3273 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
a3621e74
YO
3274 );
3275
3276 } /* end table compress */
3277 }
1e2e3d02 3278 DEBUG_TRIE_COMPILE_MORE_r(
147e3846 3279 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
cb41e5d6 3280 depth+1,
1e2e3d02
YO
3281 (UV)trie->statecount,
3282 (UV)trie->lasttrans)
3283 );
cc601c31 3284 /* resize the trans array to remove unused space */
c944940b
JH
3285 trie->trans = (reg_trie_trans *)
3286 PerlMemShared_realloc( trie->trans, trie->lasttrans
3287 * sizeof(reg_trie_trans) );
a3621e74 3288
538e84ed 3289 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
3290 U8 nodetype =(U8)(flags & 0xFF);
3291 char *str=NULL;
538e84ed 3292
07be1b83 3293#ifdef DEBUGGING
e62cc96a 3294 regnode *optimize = NULL;
7122b237
YO
3295#ifdef RE_TRACK_PATTERN_OFFSETS
3296
b57a0404
JH
3297 U32 mjd_offset = 0;
3298 U32 mjd_nodelen = 0;
7122b237
YO
3299#endif /* RE_TRACK_PATTERN_OFFSETS */
3300#endif /* DEBUGGING */
a3621e74 3301 /*
3dab1dad
YO
3302 This means we convert either the first branch or the first Exact,
3303 depending on whether the thing following (in 'last') is a branch
3304 or not and whther first is the startbranch (ie is it a sub part of
3305 the alternation or is it the whole thing.)
3b753521 3306 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 3307 the whole branch sequence, including the first.
a3621e74 3308 */
3dab1dad 3309 /* Find the node we are going to overwrite */
7f69552c 3310 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 3311 /* branch sub-chain */
3dab1dad 3312 NEXT_OFF( first ) = (U16)(last - first);
7122b237 3313#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
3314 DEBUG_r({
3315 mjd_offset= Node_Offset((convert));
3316 mjd_nodelen= Node_Length((convert));
3317 });
7122b237 3318#endif
7f69552c 3319 /* whole branch chain */
7122b237
YO
3320 }
3321#ifdef RE_TRACK_PATTERN_OFFSETS
3322 else {
7f69552c
YO
3323 DEBUG_r({
3324 const regnode *nop = NEXTOPER( convert );
3325 mjd_offset= Node_Offset((nop));
3326 mjd_nodelen= Node_Length((nop));
3327 });
07be1b83
YO
3328 }
3329 DEBUG_OPTIMISE_r(
147e3846 3330 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
cb41e5d6 3331 depth+1,
786e8c11 3332 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 3333 );
7122b237 3334#endif
538e84ed 3335 /* But first we check to see if there is a common prefix we can
3dab1dad
YO
3336 split out as an EXACT and put in front of the TRIE node. */
3337 trie->startstate= 1;
55eed653 3338 if ( trie->bitmap && !widecharmap && !trie->jump ) {
5ee57374
YO
3339 /* we want to find the first state that has more than
3340 * one transition, if that state is not the first state
3341 * then we have a common prefix which we can remove.
3342 */
3dab1dad 3343 U32 state;
1e2e3d02 3344 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 3345 U32 ofs = 0;
ee0dfd0b
YO
3346 I32 first_ofs = -1; /* keeps track of the ofs of the first
3347 transition, -1 means none */
8e11feef
RGS
3348 U32 count = 0;
3349 const U32 base = trie->states[ state ].trans.base;
a3621e74 3350
5ee57374 3351 /* does this state terminate an alternation? */
3dab1dad 3352 if ( trie->states[state].wordnum )
8e11feef 3353 count = 1;
a3621e74 3354
8e11feef 3355 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
3356 if ( ( base + ofs >= trie->uniquecharcount ) &&
3357 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
3358 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3359 {
3dab1dad 3360 if ( ++count > 1 ) {
5ee57374 3361 /* we have more than one transition */
d3d91c7c
YO
3362 SV **tmp;
3363 U8 *ch;
5ee57374
YO
3364 /* if this is the first state there is no common prefix
3365 * to extract, so we can exit */
8e11feef 3366 if ( state == 1 ) break;
d3d91c7c
YO
3367 tmp = av_fetch( revcharmap, ofs, 0);
3368 ch = (U8*)SvPV_nolen_const( *tmp );
3369
5ee57374
YO
3370 /* if we are on count 2 then we need to initialize the
3371 * bitmap, and store the previous char if there was one
3372 * in it*/
3dab1dad 3373 if ( count == 2 ) {
5ee57374 3374 /* clear the bitmap */
3dab1dad
YO
3375 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3376 DEBUG_OPTIMISE_r(
147e3846 3377 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
cb41e5d6 3378 depth+1,
786e8c11 3379 (UV)state));
ee0dfd0b
YO
3380 if (first_ofs >= 0) {
3381 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
be8e71aa 3382 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 3383
8bcafbf4 3384 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3dab1dad 3385 DEBUG_OPTIMISE_r(
6ad9a8ab 3386 Perl_re_printf( aTHX_ "%s", (char*)ch)
3dab1dad 3387 );
8e11feef
RGS
3388 }
3389 }
8bcafbf4
YO
3390 /* store the current firstchar in the bitmap */
3391 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
6ad9a8ab 3392 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
8e11feef 3393 }
ee0dfd0b 3394 first_ofs = ofs;
8e11feef 3395 }
3dab1dad
YO
3396 }
3397 if ( count == 1 ) {
5ee57374
YO
3398 /* This state has only one transition, its transition is part
3399 * of a common prefix - we need to concatenate the char it
3400 * represents to what we have so far. */
ee0dfd0b 3401 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
c490c714
YO
3402 STRLEN len;
3403 char *ch = SvPV( *tmp, len );
de734bd5
A
3404 DEBUG_OPTIMISE_r({
3405 SV *sv=sv_newmortal();
147e3846 3406 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
cb41e5d6 3407 depth+1,
ee0dfd0b 3408 (UV)state, (UV)first_ofs,
538e84ed 3409 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
de734bd5
A
3410 PL_colors[0], PL_colors[1],
3411 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed 3412 PERL_PV_ESCAPE_FIRSTCHAR
de734bd5
A
3413 )
3414 );
3415 });
3dab1dad
YO
3416 if ( state==1 ) {
3417 OP( convert ) = nodetype;
3418 str=STRING(convert);
3419 STR_LEN(convert)=0;
3420 }
c490c714
YO
3421 STR_LEN(convert) += len;
3422 while (len--)
de734bd5 3423 *str++ = *ch++;
8e11feef 3424 } else {
538e84ed 3425#ifdef DEBUGGING
8e11feef 3426 if (state>1)
6ad9a8ab 3427 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
f9049ba1 3428#endif
8e11feef
RGS
3429 break;
3430 }
3431 }
2e64971a 3432 trie->prefixlen = (state-1);
3dab1dad 3433 if (str) {
8e11feef 3434 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 3435 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 3436 trie->startstate = state;
07be1b83
YO
3437 trie->minlen -= (state - 1);
3438 trie->maxlen -= (state - 1);
33809eae
JH
3439#ifdef DEBUGGING
3440 /* At least the UNICOS C compiler choked on this
3441 * being argument to DEBUG_r(), so let's just have
3442 * it right here. */
3443 if (
3444#ifdef PERL_EXT_RE_BUILD
3445 1
3446#else
3447 DEBUG_r_TEST
3448#endif
3449 ) {
3450 regnode *fix = convert;
3451 U32 word = trie->wordcount;
3452 mjd_nodelen++;
3453 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3454 while( ++fix < n ) {
3455 Set_Node_Offset_Length(fix, 0, 0);
3456 }
3457 while (word--) {
3458 SV ** const tmp = av_fetch( trie_words, word, 0 );
3459 if (tmp) {
3460 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3461 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3462 else
3463 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3464 }
3465 }
3466 }
3467#endif
8e11feef
RGS
3468 if (trie->maxlen) {
3469 convert = n;
3470 } else {
3dab1dad 3471 NEXT_OFF(convert) = (U16)(tail - convert);
a5ca303d 3472 DEBUG_r(optimize= n);
3dab1dad
YO
3473 }
3474 }
3475 }
538e84ed
KW
3476 if (!jumper)
3477 jumper = last;
3dab1dad 3478 if ( trie->maxlen ) {
8e11feef
RGS
3479 NEXT_OFF( convert ) = (U16)(tail - convert);
3480 ARG_SET( convert, data_slot );
538e84ed
KW
3481 /* Store the offset to the first unabsorbed branch in
3482 jump[0], which is otherwise unused by the jump logic.
786e8c11 3483 We use this when dumping a trie and during optimisation. */
538e84ed 3484 if (trie->jump)
7f69552c 3485 trie->jump[0] = (U16)(nextbranch - convert);
538e84ed 3486
6c48061a
YO
3487 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3488 * and there is a bitmap
3489 * and the first "jump target" node we found leaves enough room
3490 * then convert the TRIE node into a TRIEC node, with the bitmap
3491 * embedded inline in the opcode - this is hypothetically faster.
3492 */
3493 if ( !trie->states[trie->startstate].wordnum