This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
regcomp.c: Rename macros; reword some panic messages
[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. */
51684e3c
KW
136 char *copy_start; /* start of copy of input within
137 constructed parse string */
138 char *copy_start_in_input; /* Position in input string
139 corresponding to copy_start */
ea3daa5d 140 SSize_t whilem_seen; /* number of WHILEM in this expr */
fac92740 141 regnode *emit_start; /* Start of emitted-code area */
538e84ed
KW
142 regnode *emit_bound; /* First regnode outside of the
143 allocated space */
f7c7e32a
DM
144 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
145 implies compiling, so don't emit */
9a81a976
KW
146 regnode_ssc emit_dummy; /* placeholder for emit to point to;
147 large enough for the largest
148 non-EXACTish node, so can use it as
149 scratch in pass1 */
830247a4
IZ
150 I32 naughty; /* How bad is this pattern? */
151 I32 sawback; /* Did we see \1, ...? */
152 U32 seen;
ea3daa5d 153 SSize_t size; /* Code size. */
e7f023ff 154 I32 npar; /* Capture buffer count, (OPEN) plus
538e84ed
KW
155 one. ("par" 0 is the whole
156 pattern)*/
157 I32 nestroot; /* root parens we are in - used by
158 accept */
830247a4
IZ
159 I32 extralen;
160 I32 seen_zerolen;
40d049e4
YO
161 regnode **open_parens; /* pointers to open parens */
162 regnode **close_parens; /* pointers to close parens */
d5a00e4a 163 regnode *end_op; /* END node in program */
02daf0ab
YO
164 I32 utf8; /* whether the pattern is utf8 or not */
165 I32 orig_utf8; /* whether the pattern was originally in utf8 */
166 /* XXX use this for future optimisation of case
167 * where pattern must be upgraded to utf8. */
e40e74fe
KW
168 I32 uni_semantics; /* If a d charset modifier should use unicode
169 rules, even if the pattern is not in
170 utf8 */
81714fb9 171 HV *paren_names; /* Paren names */
538e84ed 172
40d049e4 173 regnode **recurse; /* Recurse regops */
d5a00e4a 174 I32 recurse_count; /* Number of recurse regops we have generated */
4286711a 175 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
538e84ed 176 through */
09a65838 177 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
b57e4118 178 I32 in_lookbehind;
4624b182 179 I32 contains_locale;
bb3f3ed2 180 I32 override_recoding;
b6d67071
KW
181#ifdef EBCDIC
182 I32 recode_x_to_native;
183#endif
9d53c457 184 I32 in_multi_char_class;
1acab4c5 185 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
68e2671b 186 within pattern */
b1603ef8 187 int code_index; /* next code_blocks[] slot */
ee273784 188 SSize_t maxlen; /* mininum possible number of chars in string to match */
3f910e62
YO
189 scan_frame *frame_head;
190 scan_frame *frame_last;
191 U32 frame_count;
7eec73eb 192 AV *warn_text;
dc6d7f5c 193#ifdef ADD_TO_REGEXEC
830247a4
IZ
194 char *starttry; /* -Dr: where regtry was called. */
195#define RExC_starttry (pRExC_state->starttry)
196#endif
d24ca0c5 197 SV *runtime_code_qr; /* qr with the runtime code blocks */
3dab1dad 198#ifdef DEBUGGING
be8e71aa 199 const char *lastparse;
3dab1dad 200 I32 lastnum;
1f1031fe 201 AV *paren_name_list; /* idx -> name */
d9a72fcc 202 U32 study_chunk_recursed_count;
c9f0d54c
YO
203 SV *mysv1;
204 SV *mysv2;
3dab1dad
YO
205#define RExC_lastparse (pRExC_state->lastparse)
206#define RExC_lastnum (pRExC_state->lastnum)
1f1031fe 207#define RExC_paren_name_list (pRExC_state->paren_name_list)
d9a72fcc 208#define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
c9f0d54c
YO
209#define RExC_mysv (pRExC_state->mysv1)
210#define RExC_mysv1 (pRExC_state->mysv1)
211#define RExC_mysv2 (pRExC_state->mysv2)
212
3dab1dad 213#endif
512c0f5a 214 bool seen_unfolded_sharp_s;
911bd04e 215 bool strict;
da7cf1cc 216 bool study_started;
034602eb 217 bool in_script_run;
09b2b2e6 218};
830247a4 219
e2509266 220#define RExC_flags (pRExC_state->flags)
514a91f1 221#define RExC_pm_flags (pRExC_state->pm_flags)
830247a4 222#define RExC_precomp (pRExC_state->precomp)
51684e3c
KW
223#define RExC_copy_start_in_input (pRExC_state->copy_start_in_input)
224#define RExC_copy_start_in_constructed (pRExC_state->copy_start)
711b303b 225#define RExC_precomp_end (pRExC_state->precomp_end)
288b8c02 226#define RExC_rx_sv (pRExC_state->rx_sv)
830247a4 227#define RExC_rx (pRExC_state->rx)
f8fc2ecf 228#define RExC_rxi (pRExC_state->rxi)
fac92740 229#define RExC_start (pRExC_state->start)
830247a4
IZ
230#define RExC_end (pRExC_state->end)
231#define RExC_parse (pRExC_state->parse)
232#define RExC_whilem_seen (pRExC_state->whilem_seen)
512c0f5a
KW
233
234/* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
235 * EXACTF node, hence was parsed under /di rules. If later in the parse,
236 * something forces the pattern into using /ui rules, the sharp s should be
237 * folded into the sequence 'ss', which takes up more space than previously
238 * calculated. This means that the sizing pass needs to be restarted. (The
239 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
240 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
241 * so there is no need to resize [perl #125990]. */
242#define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
243
7122b237 244#ifdef RE_TRACK_PATTERN_OFFSETS
538e84ed
KW
245#define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
246 others */
7122b237 247#endif
830247a4 248#define RExC_emit (pRExC_state->emit)
f7c7e32a 249#define RExC_emit_dummy (pRExC_state->emit_dummy)
fac92740 250#define RExC_emit_start (pRExC_state->emit_start)
3b57cd43 251#define RExC_emit_bound (pRExC_state->emit_bound)
830247a4
IZ
252#define RExC_sawback (pRExC_state->sawback)
253#define RExC_seen (pRExC_state->seen)
254#define RExC_size (pRExC_state->size)
ee273784 255#define RExC_maxlen (pRExC_state->maxlen)
830247a4 256#define RExC_npar (pRExC_state->npar)
e2e6a0f1 257#define RExC_nestroot (pRExC_state->nestroot)
830247a4
IZ
258#define RExC_extralen (pRExC_state->extralen)
259#define RExC_seen_zerolen (pRExC_state->seen_zerolen)
1aa99e6b 260#define RExC_utf8 (pRExC_state->utf8)
e40e74fe 261#define RExC_uni_semantics (pRExC_state->uni_semantics)
02daf0ab 262#define RExC_orig_utf8 (pRExC_state->orig_utf8)
40d049e4
YO
263#define RExC_open_parens (pRExC_state->open_parens)
264#define RExC_close_parens (pRExC_state->close_parens)
5bd2d46e 265#define RExC_end_op (pRExC_state->end_op)
81714fb9 266#define RExC_paren_names (pRExC_state->paren_names)
40d049e4
YO
267#define RExC_recurse (pRExC_state->recurse)
268#define RExC_recurse_count (pRExC_state->recurse_count)
09a65838 269#define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
538e84ed
KW
270#define RExC_study_chunk_recursed_bytes \
271 (pRExC_state->study_chunk_recursed_bytes)
b57e4118 272#define RExC_in_lookbehind (pRExC_state->in_lookbehind)
4624b182 273#define RExC_contains_locale (pRExC_state->contains_locale)
b6d67071
KW
274#ifdef EBCDIC
275# define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
276#endif
9d53c457 277#define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
3f910e62
YO
278#define RExC_frame_head (pRExC_state->frame_head)
279#define RExC_frame_last (pRExC_state->frame_last)
280#define RExC_frame_count (pRExC_state->frame_count)
67cdf558 281#define RExC_strict (pRExC_state->strict)
da7cf1cc 282#define RExC_study_started (pRExC_state->study_started)
7eec73eb 283#define RExC_warn_text (pRExC_state->warn_text)
034602eb 284#define RExC_in_script_run (pRExC_state->in_script_run)
830247a4 285
99807a43
HS
286/* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
287 * a flag to disable back-off on the fixed/floating substrings - if it's
288 * a high complexity pattern we assume the benefit of avoiding a full match
289 * is worth the cost of checking for the substrings even if they rarely help.
290 */
291#define RExC_naughty (pRExC_state->naughty)
292#define TOO_NAUGHTY (10)
293#define MARK_NAUGHTY(add) \
294 if (RExC_naughty < TOO_NAUGHTY) \
295 RExC_naughty += (add)
296#define MARK_NAUGHTY_EXP(exp, add) \
297 if (RExC_naughty < TOO_NAUGHTY) \
298 RExC_naughty += RExC_naughty / (exp) + (add)
cde0cee5 299
a687059c
LW
300#define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
301#define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
412f55bb 302 ((*s) == '{' && regcurly(s)))
a687059c
LW
303
304/*
305 * Flags to be passed up and down.
306 */
a687059c 307#define WORST 0 /* Worst case. */
a3b492c3 308#define HASWIDTH 0x01 /* Known to match non-null strings. */
fda99bee 309
e64f369d 310/* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
2fd92675
KW
311 * character. (There needs to be a case: in the switch statement in regexec.c
312 * for any node marked SIMPLE.) Note that this is not the same thing as
313 * REGNODE_SIMPLE */
fda99bee 314#define SIMPLE 0x02
e64f369d 315#define SPSTART 0x04 /* Starts with * or + */
8d9c2815
NC
316#define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
317#define TRYAGAIN 0x10 /* Weeded out a declaration. */
b97943f7
KW
318#define RESTART_PASS1 0x20 /* Need to restart sizing pass */
319#define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
320 calcuate sizes as UTF-8 */
a687059c 321
3dab1dad
YO
322#define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
323
07be1b83
YO
324/* whether trie related optimizations are enabled */
325#if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
326#define TRIE_STUDY_OPT
786e8c11 327#define FULL_TRIE_STUDY
07be1b83
YO
328#define TRIE_STCLASS
329#endif
1de06328
YO
330
331
40d049e4
YO
332
333#define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
334#define PBITVAL(paren) (1 << ((paren) & 7))
335#define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
336#define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
337#define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
338
82a6ada4 339#define REQUIRE_UTF8(flagp) STMT_START { \
8d9c2815 340 if (!UTF) { \
82a6ada4
KW
341 assert(PASS1); \
342 *flagp = RESTART_PASS1|NEED_UTF8; \
8d9c2815
NC
343 return NULL; \
344 } \
82a6ada4 345 } STMT_END
40d049e4 346
512c0f5a
KW
347/* Change from /d into /u rules, and restart the parse if we've already seen
348 * something whose size would increase as a result, by setting *flagp and
349 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
142ece99 350 * we've changed to /u during the parse. */
512c0f5a
KW
351#define REQUIRE_UNI_RULES(flagp, restart_retval) \
352 STMT_START { \
353 if (DEPENDS_SEMANTICS) { \
354 assert(PASS1); \
355 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
356 RExC_uni_semantics = 1; \
357 if (RExC_seen_unfolded_sharp_s) { \
358 *flagp |= RESTART_PASS1; \
359 return restart_retval; \
360 } \
361 } \
362 } STMT_END
363
fd7fc6ae
YO
364/* Executes a return statement with the value 'X', if 'flags' contains any of
365 * 'RESTART_PASS1', 'NEED_UTF8', or 'extra'. If so, *flagp is set to those
366 * flags */
367#define RETURN_X_ON_RESTART_OR_FLAGS(X, flags, flagp, extra) \
368 STMT_START { \
369 if ((flags) & (RESTART_PASS1|NEED_UTF8|(extra))) { \
370 *(flagp) = (flags) & (RESTART_PASS1|NEED_UTF8|(extra)); \
371 return X; \
372 } \
373 } STMT_END
374
5d555bb8 375#define RETURN_FAIL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
fd7fc6ae
YO
376 RETURN_X_ON_RESTART_OR_FLAGS(NULL,flags,flagp,extra)
377
378#define RETURN_X_ON_RESTART(X, flags,flagp) \
379 RETURN_X_ON_RESTART_OR_FLAGS( X, flags, flagp, 0)
380
381
5d555bb8 382#define RETURN_FAIL_ON_RESTART_FLAGP_OR_FLAGS(flagp,extra) \
fd7fc6ae
YO
383 if (*(flagp) & (RESTART_PASS1|(extra))) return NULL
384
385#define MUST_RESTART(flags) ((flags) & (RESTART_PASS1))
386
5d555bb8 387#define RETURN_FAIL_ON_RESTART(flags,flagp) \
fd7fc6ae 388 RETURN_X_ON_RESTART(NULL, flags,flagp)
5d555bb8
KW
389#define RETURN_FAIL_ON_RESTART_FLAGP(flagp) \
390 RETURN_FAIL_ON_RESTART_FLAGP_OR_FLAGS(flagp,0)
fd7fc6ae 391
f19b1a63
KW
392/* This converts the named class defined in regcomp.h to its equivalent class
393 * number defined in handy.h. */
394#define namedclass_to_classnum(class) ((int) ((class) / 2))
395#define classnum_to_namedclass(classnum) ((classnum) * 2)
396
de92f5e6
KW
397#define _invlist_union_complement_2nd(a, b, output) \
398 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
399#define _invlist_intersection_complement_2nd(a, b, output) \
400 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
401
1de06328
YO
402/* About scan_data_t.
403
404 During optimisation we recurse through the regexp program performing
405 various inplace (keyhole style) optimisations. In addition study_chunk
406 and scan_commit populate this data structure with information about
538e84ed 407 what strings MUST appear in the pattern. We look for the longest
3b753521 408 string that must appear at a fixed location, and we look for the
1de06328
YO
409 longest string that may appear at a floating location. So for instance
410 in the pattern:
538e84ed 411
1de06328 412 /FOO[xX]A.*B[xX]BAR/
538e84ed 413
1de06328
YO
414 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
415 strings (because they follow a .* construct). study_chunk will identify
416 both FOO and BAR as being the longest fixed and floating strings respectively.
538e84ed 417
1de06328 418 The strings can be composites, for instance
538e84ed 419
1de06328 420 /(f)(o)(o)/
538e84ed 421
1de06328 422 will result in a composite fixed substring 'foo'.
538e84ed 423
1de06328 424 For each string some basic information is maintained:
538e84ed 425
6ea77169 426 - min_offset
1de06328
YO
427 This is the position the string must appear at, or not before.
428 It also implicitly (when combined with minlenp) tells us how many
3b753521
FN
429 characters must match before the string we are searching for.
430 Likewise when combined with minlenp and the length of the string it
538e84ed 431 tells us how many characters must appear after the string we have
1de06328 432 found.
538e84ed 433
1de06328
YO
434 - max_offset
435 Only used for floating strings. This is the rightmost point that
ea3daa5d 436 the string can appear at. If set to SSize_t_MAX it indicates that the
1de06328 437 string can occur infinitely far to the right.
6e12f832 438 For fixed strings, it is equal to min_offset.
538e84ed 439
1de06328 440 - minlenp
2d608413
KW
441 A pointer to the minimum number of characters of the pattern that the
442 string was found inside. This is important as in the case of positive
538e84ed 443 lookahead or positive lookbehind we can have multiple patterns
1de06328 444 involved. Consider
538e84ed 445
1de06328 446 /(?=FOO).*F/
538e84ed 447
1de06328
YO
448 The minimum length of the pattern overall is 3, the minimum length
449 of the lookahead part is 3, but the minimum length of the part that
538e84ed 450 will actually match is 1. So 'FOO's minimum length is 3, but the
1de06328 451 minimum length for the F is 1. This is important as the minimum length
538e84ed 452 is used to determine offsets in front of and behind the string being
1de06328 453 looked for. Since strings can be composites this is the length of the
486ec47a 454 pattern at the time it was committed with a scan_commit. Note that
1de06328 455 the length is calculated by study_chunk, so that the minimum lengths
538e84ed 456 are not known until the full pattern has been compiled, thus the
1de06328 457 pointer to the value.
538e84ed 458
1de06328 459 - lookbehind
538e84ed 460
1de06328 461 In the case of lookbehind the string being searched for can be
538e84ed 462 offset past the start point of the final matching string.
1de06328
YO
463 If this value was just blithely removed from the min_offset it would
464 invalidate some of the calculations for how many chars must match
465 before or after (as they are derived from min_offset and minlen and
538e84ed 466 the length of the string being searched for).
1de06328
YO
467 When the final pattern is compiled and the data is moved from the
468 scan_data_t structure into the regexp structure the information
538e84ed
KW
469 about lookbehind is factored in, with the information that would
470 have been lost precalculated in the end_shift field for the
1de06328
YO
471 associated string.
472
473 The fields pos_min and pos_delta are used to store the minimum offset
538e84ed 474 and the delta to the maximum offset at the current point in the pattern.
1de06328
YO
475
476*/
2c2d71f5 477
cb6aef8b
DM
478struct scan_data_substrs {
479 SV *str; /* longest substring found in pattern */
480 SSize_t min_offset; /* earliest point in string it can appear */
481 SSize_t max_offset; /* latest point in string it can appear */
482 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
483 SSize_t lookbehind; /* is the pos of the string modified by LB */
11683ecb 484 I32 flags; /* per substring SF_* and SCF_* flags */
cb6aef8b
DM
485};
486
2c2d71f5 487typedef struct scan_data_t {
1de06328
YO
488 /*I32 len_min; unused */
489 /*I32 len_delta; unused */
49f55535 490 SSize_t pos_min;
ea3daa5d 491 SSize_t pos_delta;
2c2d71f5 492 SV *last_found;
ea3daa5d 493 SSize_t last_end; /* min value, <0 unless valid. */
49f55535 494 SSize_t last_start_min;
ea3daa5d 495 SSize_t last_start_max;
2099df82
DM
496 U8 cur_is_floating; /* whether the last_* values should be set as
497 * the next fixed (0) or floating (1)
498 * substring */
6ea77169 499
2099df82
DM
500 /* [0] is longest fixed substring so far, [1] is longest float so far */
501 struct scan_data_substrs substrs[2];
6ea77169 502
11683ecb 503 I32 flags; /* common SF_* and SCF_* flags */
2c2d71f5 504 I32 whilem_c;
49f55535 505 SSize_t *last_closep;
b8f7bb16 506 regnode_ssc *start_class;
2c2d71f5
JH
507} scan_data_t;
508
a687059c 509/*
e50aee73 510 * Forward declarations for pregcomp()'s friends.
a687059c 511 */
a0d0e21e 512
6ea77169 513static const scan_data_t zero_scan_data = {
a2ca2017 514 0, 0, NULL, 0, 0, 0, 0,
6ea77169 515 {
11683ecb
DM
516 { NULL, 0, 0, 0, 0, 0 },
517 { NULL, 0, 0, 0, 0, 0 },
6ea77169
DM
518 },
519 0, 0, NULL, NULL
520};
c277df42 521
11683ecb
DM
522/* study flags */
523
07be1b83
YO
524#define SF_BEFORE_SEOL 0x0001
525#define SF_BEFORE_MEOL 0x0002
11683ecb 526#define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
c277df42 527
07be1b83
YO
528#define SF_IS_INF 0x0040
529#define SF_HAS_PAR 0x0080
530#define SF_IN_PAR 0x0100
531#define SF_HAS_EVAL 0x0200
bc604ad8
KW
532
533
534/* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
535 * longest substring in the pattern. When it is not set the optimiser keeps
536 * track of position, but does not keep track of the actual strings seen,
537 *
538 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
539 * /foo/i will not.
540 *
541 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
542 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
543 * turned off because of the alternation (BRANCH). */
07be1b83 544#define SCF_DO_SUBSTR 0x0400
bc604ad8 545
653099ff
GS
546#define SCF_DO_STCLASS_AND 0x0800
547#define SCF_DO_STCLASS_OR 0x1000
548#define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
e1901655 549#define SCF_WHILEM_VISITED_POS 0x2000
c277df42 550
786e8c11 551#define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
538e84ed 552#define SCF_SEEN_ACCEPT 0x8000
688e0391 553#define SCF_TRIE_DOING_RESTUDY 0x10000
4286711a
YO
554#define SCF_IN_DEFINE 0x20000
555
556
557
07be1b83 558
43fead97 559#define UTF cBOOL(RExC_utf8)
00b27cfc
KW
560
561/* The enums for all these are ordered so things work out correctly */
a62b1201 562#define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
538e84ed
KW
563#define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
564 == REGEX_DEPENDS_CHARSET)
00b27cfc 565#define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
538e84ed
KW
566#define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
567 >= REGEX_UNICODE_CHARSET)
568#define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
569 == REGEX_ASCII_RESTRICTED_CHARSET)
570#define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
571 >= REGEX_ASCII_RESTRICTED_CHARSET)
572#define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
573 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
a62b1201 574
43fead97 575#define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
a0ed51b3 576
f2c2a6ab
KW
577/* For programs that want to be strictly Unicode compatible by dying if any
578 * attempt is made to match a non-Unicode code point against a Unicode
579 * property. */
580#define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
581
93733859 582#define OOB_NAMEDCLASS -1
b8c5462f 583
8e661ac5
KW
584/* There is no code point that is out-of-bounds, so this is problematic. But
585 * its only current use is to initialize a variable that is always set before
586 * looked at. */
587#define OOB_UNICODE 0xDEADBEEF
588
a0ed51b3 589#define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
a0ed51b3 590
8615cb43 591
b45f050a
JF
592/* length of regex to show in messages that don't mark a position within */
593#define RegexLengthToShowInErrorMessages 127
594
595/*
596 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
597 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
598 * op/pragma/warn/regcomp.
599 */
7253e4e3
RK
600#define MARKER1 "<-- HERE" /* marker as it appears in the description */
601#define MARKER2 " <-- HERE " /* marker as it appears within the regex */
b81d288d 602
538e84ed 603#define REPORT_LOCATION " in regex; marked by " MARKER1 \
147e3846 604 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
b45f050a 605
285b5ca0
KW
606/* The code in this file in places uses one level of recursion with parsing
607 * rebased to an alternate string constructed by us in memory. This can take
608 * the form of something that is completely different from the input, or
609 * something that uses the input as part of the alternate. In the first case,
610 * there should be no possibility of an error, as we are in complete control of
51684e3c
KW
611 * the alternate string. But in the second case we don't completely control
612 * the input portion, so there may be errors in that. Here's an example:
285b5ca0
KW
613 * /[abc\x{DF}def]/ui
614 * is handled specially because \x{df} folds to a sequence of more than one
51684e3c 615 * character: 'ss'. What is done is to create and parse an alternate string,
285b5ca0
KW
616 * which looks like this:
617 * /(?:\x{DF}|[abc\x{DF}def])/ui
618 * where it uses the input unchanged in the middle of something it constructs,
619 * which is a branch for the DF outside the character class, and clustering
620 * parens around the whole thing. (It knows enough to skip the DF inside the
621 * class while in this substitute parse.) 'abc' and 'def' may have errors that
622 * need to be reported. The general situation looks like this:
623 *
51684e3c 624 * |<------- identical ------>|
285b5ca0 625 * sI tI xI eI
51684e3c 626 * Input: ---------------------------------------------------------------
285b5ca0
KW
627 * Constructed: ---------------------------------------------------
628 * sC tC xC eC EC
51684e3c 629 * |<------- identical ------>|
285b5ca0 630 *
51684e3c
KW
631 * sI..eI is the portion of the input pattern we are concerned with here.
632 * sC..EC is the constructed substitute parse string.
633 * sC..tC is constructed by us
634 * tC..eC is an exact duplicate of the portion of the input pattern tI..eI.
635 * In the diagram, these are vertically aligned.
636 * eC..EC is also constructed by us.
637 * xC is the position in the substitute parse string where we found a
638 * problem.
639 * xI is the position in the original pattern corresponding to xC.
285b5ca0 640 *
51684e3c
KW
641 * We want to display a message showing the real input string. Thus we need to
642 * translate from xC to xI. We know that xC >= tC, since the portion of the
643 * string sC..tC has been constructed by us, and so shouldn't have errors. We
644 * get:
645 * xI = tI + (xC - tC)
285b5ca0 646 *
51684e3c
KW
647 * When the substitute parse is constructed, the code needs to set:
648 * RExC_start (sC)
649 * RExC_end (eC)
650 * RExC_copy_start_in_input (tI)
651 * RExC_copy_start_in_constructed (tC)
652 * and restore them when done.
285b5ca0 653 *
51684e3c
KW
654 * During normal processing of the input pattern, both
655 * 'RExC_copy_start_in_input' and 'RExC_copy_start_in_constructed' are set to
656 * sI, so that xC equals xI.
285b5ca0
KW
657 */
658
51684e3c
KW
659#define sI RExC_precomp
660#define eI RExC_precomp_end
661#define sC RExC_start
662#define eC RExC_end
663#define tI RExC_copy_start_in_input
664#define tC RExC_copy_start_in_constructed
665#define xI(xC) (tI + (xC - tC))
666#define xI_offset(xC) (xI(xC) - sI)
285b5ca0
KW
667
668#define REPORT_LOCATION_ARGS(xC) \
669 UTF8fARG(UTF, \
51684e3c 670 (xI(xC) > eI) /* Don't run off end */ \
285b5ca0 671 ? eC - sC /* Length before the <--HERE */ \
bc929b61
KW
672 : ((xI_offset(xC) >= 0) \
673 ? xI_offset(xC) \
674 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
675 IVdf " trying to output message for " \
676 " pattern %.*s", \
51684e3c 677 __FILE__, __LINE__, (IV) xI_offset(xC), \
bc929b61 678 ((int) (eC - sC)), sC), 0)), \
51684e3c 679 sI), /* The input pattern printed up to the <--HERE */ \
285b5ca0 680 UTF8fARG(UTF, \
51684e3c
KW
681 (xI(xC) > eI) ? 0 : eI - xI(xC), /* Length after <--HERE */ \
682 (xI(xC) > eI) ? eI : xI(xC)) /* pattern after <--HERE */
c1d900c3 683
8a6d8ec6
HS
684/* Used to point after bad bytes for an error message, but avoid skipping
685 * past a nul byte. */
686#define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
687
b45f050a
JF
688/*
689 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
690 * arg. Show regex, up to a maximum length. If it's too long, chop and add
691 * "...".
692 */
58e23c8d 693#define _FAIL(code) STMT_START { \
bfed75c6 694 const char *ellipses = ""; \
711b303b 695 IV len = RExC_precomp_end - RExC_precomp; \
ccb2c380
MP
696 \
697 if (!SIZE_ONLY) \
a5e7bc51 698 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
699 if (len > RegexLengthToShowInErrorMessages) { \
700 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
701 len = RegexLengthToShowInErrorMessages - 10; \
702 ellipses = "..."; \
703 } \
58e23c8d 704 code; \
ccb2c380 705} STMT_END
8615cb43 706
58e23c8d 707#define FAIL(msg) _FAIL( \
147e3846 708 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
c1d900c3 709 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d
YO
710
711#define FAIL2(msg,arg) _FAIL( \
147e3846 712 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
c1d900c3 713 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
58e23c8d 714
b45f050a 715/*
b45f050a
JF
716 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
717 */
ccb2c380 718#define Simple_vFAIL(m) STMT_START { \
ccb2c380 719 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
d528642a 720 m, REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 721} STMT_END
b45f050a
JF
722
723/*
724 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
725 */
ccb2c380
MP
726#define vFAIL(m) STMT_START { \
727 if (!SIZE_ONLY) \
a5e7bc51 728 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
729 Simple_vFAIL(m); \
730} STMT_END
b45f050a
JF
731
732/*
733 * Like Simple_vFAIL(), but accepts two arguments.
734 */
ccb2c380 735#define Simple_vFAIL2(m,a1) STMT_START { \
d528642a
KW
736 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
737 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 738} STMT_END
b45f050a
JF
739
740/*
741 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
742 */
ccb2c380
MP
743#define vFAIL2(m,a1) STMT_START { \
744 if (!SIZE_ONLY) \
a5e7bc51 745 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
746 Simple_vFAIL2(m, a1); \
747} STMT_END
b45f050a
JF
748
749
750/*
751 * Like Simple_vFAIL(), but accepts three arguments.
752 */
ccb2c380 753#define Simple_vFAIL3(m, a1, a2) STMT_START { \
c1d900c3 754 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
d528642a 755 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 756} STMT_END
b45f050a
JF
757
758/*
759 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
760 */
ccb2c380
MP
761#define vFAIL3(m,a1,a2) STMT_START { \
762 if (!SIZE_ONLY) \
a5e7bc51 763 SAVEFREESV(RExC_rx_sv); \
ccb2c380
MP
764 Simple_vFAIL3(m, a1, a2); \
765} STMT_END
b45f050a
JF
766
767/*
768 * Like Simple_vFAIL(), but accepts four arguments.
769 */
ccb2c380 770#define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
d528642a
KW
771 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
772 REPORT_LOCATION_ARGS(RExC_parse)); \
ccb2c380 773} STMT_END
b45f050a 774
95db3ffa
KW
775#define vFAIL4(m,a1,a2,a3) STMT_START { \
776 if (!SIZE_ONLY) \
777 SAVEFREESV(RExC_rx_sv); \
778 Simple_vFAIL4(m, a1, a2, a3); \
779} STMT_END
780
946095af 781/* A specialized version of vFAIL2 that works with UTF8f */
d528642a
KW
782#define vFAIL2utf8f(m, a1) STMT_START { \
783 if (!SIZE_ONLY) \
784 SAVEFREESV(RExC_rx_sv); \
785 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
786 REPORT_LOCATION_ARGS(RExC_parse)); \
946095af
BF
787} STMT_END
788
3ba22297 789#define vFAIL3utf8f(m, a1, a2) STMT_START { \
3ba22297
KW
790 if (!SIZE_ONLY) \
791 SAVEFREESV(RExC_rx_sv); \
792 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
d528642a 793 REPORT_LOCATION_ARGS(RExC_parse)); \
3ba22297
KW
794} STMT_END
795
499333dc
KW
796/* These have asserts in them because of [perl #122671] Many warnings in
797 * regcomp.c can occur twice. If they get output in pass1 and later in that
798 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
799 * would get output again. So they should be output in pass2, and these
800 * asserts make sure new warnings follow that paradigm. */
946095af 801
5e0a247b
KW
802/* m is not necessarily a "literal string", in this macro */
803#define reg_warn_non_literal_string(loc, m) STMT_START { \
d528642a
KW
804 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
805 "%s" REPORT_LOCATION, \
806 m, REPORT_LOCATION_ARGS(loc)); \
5e0a247b
KW
807} STMT_END
808
668c081a 809#define ckWARNreg(loc,m) STMT_START { \
d528642a
KW
810 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
811 m REPORT_LOCATION, \
812 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
813} STMT_END
814
b927b7e9 815#define vWARN(loc, m) STMT_START { \
d528642a
KW
816 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
817 m REPORT_LOCATION, \
818 REPORT_LOCATION_ARGS(loc)); \
b927b7e9
KW
819} STMT_END
820
0d6106aa 821#define vWARN_dep(loc, m) STMT_START { \
d528642a
KW
822 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
823 m REPORT_LOCATION, \
824 REPORT_LOCATION_ARGS(loc)); \
0d6106aa
KW
825} STMT_END
826
147508a2 827#define ckWARNdep(loc,m) STMT_START { \
d528642a
KW
828 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
829 m REPORT_LOCATION, \
830 REPORT_LOCATION_ARGS(loc)); \
147508a2
KW
831} STMT_END
832
d528642a
KW
833#define ckWARNregdep(loc,m) STMT_START { \
834 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
835 WARN_REGEXP), \
836 m REPORT_LOCATION, \
837 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
838} STMT_END
839
d528642a
KW
840#define ckWARN2reg_d(loc,m, a1) STMT_START { \
841 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
842 m REPORT_LOCATION, \
843 a1, REPORT_LOCATION_ARGS(loc)); \
2335b3d3
KW
844} STMT_END
845
d528642a
KW
846#define ckWARN2reg(loc, m, a1) STMT_START { \
847 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
848 m REPORT_LOCATION, \
849 a1, REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
850} STMT_END
851
d528642a
KW
852#define vWARN3(loc, m, a1, a2) STMT_START { \
853 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
854 m REPORT_LOCATION, \
855 a1, a2, REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
856} STMT_END
857
d528642a
KW
858#define ckWARN3reg(loc, m, a1, a2) STMT_START { \
859 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
860 m REPORT_LOCATION, \
861 a1, a2, \
862 REPORT_LOCATION_ARGS(loc)); \
668c081a
NC
863} STMT_END
864
ccb2c380 865#define vWARN4(loc, m, a1, a2, a3) STMT_START { \
d528642a
KW
866 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
867 m REPORT_LOCATION, \
868 a1, a2, a3, \
869 REPORT_LOCATION_ARGS(loc)); \
ccb2c380
MP
870} STMT_END
871
668c081a 872#define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
d528642a
KW
873 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
874 m REPORT_LOCATION, \
875 a1, a2, a3, \
876 REPORT_LOCATION_ARGS(loc)); \
668c081a
NC
877} STMT_END
878
ccb2c380 879#define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
d528642a
KW
880 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
881 m REPORT_LOCATION, \
882 a1, a2, a3, a4, \
883 REPORT_LOCATION_ARGS(loc)); \
ccb2c380 884} STMT_END
9d1d55b5 885
538e84ed 886/* Macros for recording node offsets. 20001227 mjd@plover.com
fac92740
MJD
887 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
888 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
889 * Element 0 holds the number n.
07be1b83 890 * Position is 1 indexed.
fac92740 891 */
7122b237
YO
892#ifndef RE_TRACK_PATTERN_OFFSETS
893#define Set_Node_Offset_To_R(node,byte)
894#define Set_Node_Offset(node,byte)
895#define Set_Cur_Node_Offset
896#define Set_Node_Length_To_R(node,len)
897#define Set_Node_Length(node,len)
6a86c6ad 898#define Set_Node_Cur_Length(node,start)
538e84ed
KW
899#define Node_Offset(n)
900#define Node_Length(n)
7122b237
YO
901#define Set_Node_Offset_Length(node,offset,len)
902#define ProgLen(ri) ri->u.proglen
903#define SetProgLen(ri,x) ri->u.proglen = x
904#else
905#define ProgLen(ri) ri->u.offsets[0]
906#define SetProgLen(ri,x) ri->u.offsets[0] = x
ccb2c380
MP
907#define Set_Node_Offset_To_R(node,byte) STMT_START { \
908 if (! SIZE_ONLY) { \
909 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
2a49f0f5 910 __LINE__, (int)(node), (int)(byte))); \
ccb2c380 911 if((node) < 0) { \
538e84ed
KW
912 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
913 (int)(node)); \
ccb2c380
MP
914 } else { \
915 RExC_offsets[2*(node)-1] = (byte); \
916 } \
917 } \
918} STMT_END
919
920#define Set_Node_Offset(node,byte) \
921 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
922#define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
923
924#define Set_Node_Length_To_R(node,len) STMT_START { \
925 if (! SIZE_ONLY) { \
926 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
551405c4 927 __LINE__, (int)(node), (int)(len))); \
ccb2c380 928 if((node) < 0) { \
538e84ed
KW
929 Perl_croak(aTHX_ "value of node is %d in Length macro", \
930 (int)(node)); \
ccb2c380
MP
931 } else { \
932 RExC_offsets[2*(node)] = (len); \
933 } \
934 } \
935} STMT_END
936
937#define Set_Node_Length(node,len) \
938 Set_Node_Length_To_R((node)-RExC_emit_start, len)
6a86c6ad
NC
939#define Set_Node_Cur_Length(node, start) \
940 Set_Node_Length(node, RExC_parse - start)
fac92740
MJD
941
942/* Get offsets and lengths */
943#define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
944#define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
945
07be1b83
YO
946#define Set_Node_Offset_Length(node,offset,len) STMT_START { \
947 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
948 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
949} STMT_END
7122b237 950#endif
07be1b83
YO
951
952#if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
953#define EXPERIMENTAL_INPLACESCAN
f427392e 954#endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
07be1b83 955
cb41e5d6
YO
956#ifdef DEBUGGING
957int
6ad9a8ab 958Perl_re_printf(pTHX_ const char *fmt, ...)
cb41e5d6
YO
959{
960 va_list ap;
961 int result;
962 PerlIO *f= Perl_debug_log;
963 PERL_ARGS_ASSERT_RE_PRINTF;
964 va_start(ap, fmt);
965 result = PerlIO_vprintf(f, fmt, ap);
966 va_end(ap);
967 return result;
968}
969
970int
7b031478 971Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
cb41e5d6
YO
972{
973 va_list ap;
974 int result;
975 PerlIO *f= Perl_debug_log;
976 PERL_ARGS_ASSERT_RE_INDENTF;
977 va_start(ap, depth);
daeb874b 978 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
cb41e5d6
YO
979 result = PerlIO_vprintf(f, fmt, ap);
980 va_end(ap);
981 return result;
982}
983#endif /* DEBUGGING */
984
7b031478 985#define DEBUG_RExC_seen() \
538e84ed 986 DEBUG_OPTIMISE_MORE_r({ \
6ad9a8ab 987 Perl_re_printf( aTHX_ "RExC_seen: "); \
538e84ed 988 \
e384d5c1 989 if (RExC_seen & REG_ZERO_LEN_SEEN) \
6ad9a8ab 990 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
538e84ed 991 \
e384d5c1 992 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
6ad9a8ab 993 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
538e84ed 994 \
e384d5c1 995 if (RExC_seen & REG_GPOS_SEEN) \
6ad9a8ab 996 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
538e84ed 997 \
e384d5c1 998 if (RExC_seen & REG_RECURSE_SEEN) \
6ad9a8ab 999 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
538e84ed 1000 \
7b031478 1001 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
6ad9a8ab 1002 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
538e84ed 1003 \
e384d5c1 1004 if (RExC_seen & REG_VERBARG_SEEN) \
6ad9a8ab 1005 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
538e84ed 1006 \
e384d5c1 1007 if (RExC_seen & REG_CUTGROUP_SEEN) \
6ad9a8ab 1008 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
538e84ed 1009 \
e384d5c1 1010 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
6ad9a8ab 1011 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
538e84ed 1012 \
e384d5c1 1013 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
6ad9a8ab 1014 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
538e84ed 1015 \
7b031478 1016 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
6ad9a8ab 1017 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
ee273784 1018 \
6ad9a8ab 1019 Perl_re_printf( aTHX_ "\n"); \
9e9ecfdf
YO
1020 });
1021
fdfb4f21 1022#define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
6ad9a8ab 1023 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
fdfb4f21 1024
fdfb4f21 1025
f5a36d78
DM
1026#ifdef DEBUGGING
1027static void
1028S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1029 const char *close_str)
1030{
1031 if (!flags)
1032 return;
1033
1034 Perl_re_printf( aTHX_ "%s", open_str);
11683ecb
DM
1035 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1036 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
f5a36d78
DM
1037 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1038 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1039 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1040 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1041 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1042 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1043 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1044 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1045 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1046 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1047 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1048 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1049 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1050 Perl_re_printf( aTHX_ "%s", close_str);
1051}
1052
1053
1054static void
1055S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1056 U32 depth, int is_inf)
1057{
1058 GET_RE_DEBUG_FLAGS_DECL;
1059
1060 DEBUG_OPTIMISE_MORE_r({
1061 if (!data)
1062 return;
1063 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1064 depth,
1065 where,
1066 (IV)data->pos_min,
1067 (IV)data->pos_delta,
1068 (UV)data->flags
1069 );
1070
1071 S_debug_show_study_flags(aTHX_ data->flags," [","]");
fdfb4f21 1072
f5a36d78
DM
1073 Perl_re_printf( aTHX_
1074 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1075 (IV)data->whilem_c,
1076 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1077 is_inf ? "INF " : ""
1078 );
1079
11683ecb 1080 if (data->last_found) {
37b6262f 1081 int i;
11683ecb
DM
1082 Perl_re_printf(aTHX_
1083 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1084 SvPVX_const(data->last_found),
1085 (IV)data->last_end,
1086 (IV)data->last_start_min,
1087 (IV)data->last_start_max
1088 );
1089
2099df82 1090 for (i = 0; i < 2; i++) {
37b6262f
DM
1091 Perl_re_printf(aTHX_
1092 " %s%s: '%s' @ %" IVdf "/%" IVdf,
2099df82
DM
1093 data->cur_is_floating == i ? "*" : "",
1094 i ? "Float" : "Fixed",
37b6262f
DM
1095 SvPVX_const(data->substrs[i].str),
1096 (IV)data->substrs[i].min_offset,
1097 (IV)data->substrs[i].max_offset
1098 );
1099 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1100 }
11683ecb 1101 }
f5a36d78
DM
1102
1103 Perl_re_printf( aTHX_ "\n");
1104 });
1105}
1106
1107
1108static void
1109S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1110 regnode *scan, U32 depth, U32 flags)
1111{
1112 GET_RE_DEBUG_FLAGS_DECL;
1113
1114 DEBUG_OPTIMISE_r({
1115 regnode *Next;
1116
1117 if (!scan)
1118 return;
1119 Next = regnext(scan);
1120 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1121 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1122 depth,
1123 str,
1124 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1125 Next ? (REG_NODE_NUM(Next)) : 0 );
1126 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1127 Perl_re_printf( aTHX_ "\n");
1128 });
1129}
1130
1131
1132# define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1133 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1134
1135# define DEBUG_PEEP(str, scan, depth, flags) \
1136 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1137
1138#else
1139# define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1140# define DEBUG_PEEP(str, scan, depth, flags) NOOP
1141#endif
1de06328 1142
cb41e5d6 1143
c6871b76
KW
1144/* =========================================================
1145 * BEGIN edit_distance stuff.
1146 *
1147 * This calculates how many single character changes of any type are needed to
1148 * transform a string into another one. It is taken from version 3.1 of
1149 *
1150 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1151 */
1152
1153/* Our unsorted dictionary linked list. */
1154/* Note we use UVs, not chars. */
1155
1156struct dictionary{
1157 UV key;
1158 UV value;
1159 struct dictionary* next;
1160};
1161typedef struct dictionary item;
1162
1163
1164PERL_STATIC_INLINE item*
1165push(UV key,item* curr)
1166{
1167 item* head;
d09f14bf 1168 Newx(head, 1, item);
c6871b76
KW
1169 head->key = key;
1170 head->value = 0;
1171 head->next = curr;
1172 return head;
1173}
1174
1175
1176PERL_STATIC_INLINE item*
1177find(item* head, UV key)
1178{
1179 item* iterator = head;
1180 while (iterator){
1181 if (iterator->key == key){
1182 return iterator;
1183 }
1184 iterator = iterator->next;
1185 }
1186
1187 return NULL;
1188}
1189
1190PERL_STATIC_INLINE item*
1191uniquePush(item* head,UV key)
1192{
1193 item* iterator = head;
1194
1195 while (iterator){
1196 if (iterator->key == key) {
1197 return head;
1198 }
1199 iterator = iterator->next;
1200 }
1201
1202 return push(key,head);
1203}
1204
1205PERL_STATIC_INLINE void
1206dict_free(item* head)
1207{
1208 item* iterator = head;
1209
1210 while (iterator) {
1211 item* temp = iterator;
1212 iterator = iterator->next;
1213 Safefree(temp);
1214 }
1215
1216 head = NULL;
1217}
1218
1219/* End of Dictionary Stuff */
1220
1221/* All calculations/work are done here */
1222STATIC int
1223S_edit_distance(const UV* src,
1224 const UV* tgt,
1225 const STRLEN x, /* length of src[] */
1226 const STRLEN y, /* length of tgt[] */
1227 const SSize_t maxDistance
1228)
1229{
1230 item *head = NULL;
1231 UV swapCount,swapScore,targetCharCount,i,j;
1232 UV *scores;
1233 UV score_ceil = x + y;
1234
1235 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1236
1237 /* intialize matrix start values */
d09f14bf 1238 Newx(scores, ( (x + 2) * (y + 2)), UV);
c6871b76
KW
1239 scores[0] = score_ceil;
1240 scores[1 * (y + 2) + 0] = score_ceil;
1241 scores[0 * (y + 2) + 1] = score_ceil;
1242 scores[1 * (y + 2) + 1] = 0;
1243 head = uniquePush(uniquePush(head,src[0]),tgt[0]);
1244
1245 /* work loops */
1246 /* i = src index */
1247 /* j = tgt index */
1248 for (i=1;i<=x;i++) {
1249 if (i < x)
1250 head = uniquePush(head,src[i]);
1251 scores[(i+1) * (y + 2) + 1] = i;
1252 scores[(i+1) * (y + 2) + 0] = score_ceil;
1253 swapCount = 0;
1254
1255 for (j=1;j<=y;j++) {
1256 if (i == 1) {
1257 if(j < y)
1258 head = uniquePush(head,tgt[j]);
1259 scores[1 * (y + 2) + (j + 1)] = j;
1260 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1261 }
1262
1263 targetCharCount = find(head,tgt[j-1])->value;
1264 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1265
1266 if (src[i-1] != tgt[j-1]){
1267 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));
1268 }
1269 else {
1270 swapCount = j;
1271 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1272 }
1273 }
1274
1275 find(head,src[i-1])->value = i;
1276 }
1277
1278 {
1279 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1280 dict_free(head);
1281 Safefree(scores);
1282 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1283 }
1284}
1285
1286/* END of edit_distance() stuff
1287 * ========================================================= */
1288
8e35b056
KW
1289/* is c a control character for which we have a mnemonic? */
1290#define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1291
549b4e78
KW
1292STATIC const char *
1293S_cntrl_to_mnemonic(const U8 c)
1294{
1295 /* Returns the mnemonic string that represents character 'c', if one
1296 * exists; NULL otherwise. The only ones that exist for the purposes of
1297 * this routine are a few control characters */
1298
1299 switch (c) {
1300 case '\a': return "\\a";
1301 case '\b': return "\\b";
1302 case ESC_NATIVE: return "\\e";
1303 case '\f': return "\\f";
1304 case '\n': return "\\n";
1305 case '\r': return "\\r";
1306 case '\t': return "\\t";
1307 }
1308
1309 return NULL;
1310}
1311
653099ff 1312/* Mark that we cannot extend a found fixed substring at this point.
37b6262f 1313 Update the longest found anchored substring or the longest found
653099ff
GS
1314 floating substrings if needed. */
1315
4327152a 1316STATIC void
ea3daa5d
FC
1317S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1318 SSize_t *minlenp, int is_inf)
c277df42 1319{
e1ec3a88 1320 const STRLEN l = CHR_SVLEN(data->last_found);
2099df82 1321 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
a2ca2017 1322 const STRLEN old_l = CHR_SVLEN(longest_sv);
1de06328 1323 GET_RE_DEBUG_FLAGS_DECL;
b81d288d 1324
7918f24d
NC
1325 PERL_ARGS_ASSERT_SCAN_COMMIT;
1326
c277df42 1327 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
2099df82 1328 const U8 i = data->cur_is_floating;
a2ca2017 1329 SvSetMagicSV(longest_sv, data->last_found);
37b6262f
DM
1330 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1331
2099df82
DM
1332 if (!i) /* fixed */
1333 data->substrs[0].max_offset = data->substrs[0].min_offset;
1334 else { /* float */
1335 data->substrs[1].max_offset = (l
646e8787
DM
1336 ? data->last_start_max
1337 : (data->pos_delta > SSize_t_MAX - data->pos_min
ea3daa5d
FC
1338 ? SSize_t_MAX
1339 : data->pos_min + data->pos_delta));
1340 if (is_inf
2099df82
DM
1341 || (STRLEN)data->substrs[1].max_offset > (STRLEN)SSize_t_MAX)
1342 data->substrs[1].max_offset = SSize_t_MAX;
37b6262f 1343 }
11683ecb 1344
37b6262f
DM
1345 if (data->flags & SF_BEFORE_EOL)
1346 data->substrs[i].flags |= (data->flags & SF_BEFORE_EOL);
1347 else
1348 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1349 data->substrs[i].minlenp = minlenp;
1350 data->substrs[i].lookbehind = 0;
c277df42 1351 }
37b6262f 1352
c277df42 1353 SvCUR_set(data->last_found, 0);
0eda9292 1354 {
a28509cc 1355 SV * const sv = data->last_found;
097eb12c
AL
1356 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1357 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1358 if (mg)
1359 mg->mg_len = 0;
1360 }
0eda9292 1361 }
c277df42
IZ
1362 data->last_end = -1;
1363 data->flags &= ~SF_BEFORE_EOL;
f5a36d78 1364 DEBUG_STUDYDATA("commit", data, 0, is_inf);
c277df42
IZ
1365}
1366
cdd87c1d
KW
1367/* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1368 * list that describes which code points it matches */
1369
653099ff 1370STATIC void
3420edd7 1371S_ssc_anything(pTHX_ regnode_ssc *ssc)
653099ff 1372{
cdd87c1d
KW
1373 /* Set the SSC 'ssc' to match an empty string or any code point */
1374
557bd3fb 1375 PERL_ARGS_ASSERT_SSC_ANYTHING;
7918f24d 1376
71068078 1377 assert(is_ANYOF_SYNTHETIC(ssc));
3fffb88a 1378
0854ea0b
KW
1379 /* mortalize so won't leak */
1380 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
93e92956 1381 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
653099ff
GS
1382}
1383
653099ff 1384STATIC int
dc3bf405 1385S_ssc_is_anything(const regnode_ssc *ssc)
653099ff 1386{
c144baaa
KW
1387 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1388 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1389 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1390 * in any way, so there's no point in using it */
cdd87c1d
KW
1391
1392 UV start, end;
1393 bool ret;
653099ff 1394
557bd3fb 1395 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
7918f24d 1396
71068078 1397 assert(is_ANYOF_SYNTHETIC(ssc));
cdd87c1d 1398
93e92956 1399 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
cdd87c1d
KW
1400 return FALSE;
1401 }
1402
1403 /* See if the list consists solely of the range 0 - Infinity */
1404 invlist_iterinit(ssc->invlist);
1405 ret = invlist_iternext(ssc->invlist, &start, &end)
1406 && start == 0
1407 && end == UV_MAX;
1408
1409 invlist_iterfinish(ssc->invlist);
1410
1411 if (ret) {
1412 return TRUE;
1413 }
1414
1415 /* If e.g., both \w and \W are set, matches everything */
e0e1be5f 1416 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1417 int i;
1418 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1419 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1420 return TRUE;
1421 }
1422 }
1423 }
1424
1425 return FALSE;
653099ff
GS
1426}
1427
653099ff 1428STATIC void
cdd87c1d 1429S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
653099ff 1430{
cdd87c1d
KW
1431 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1432 * string, any code point, or any posix class under locale */
1433
557bd3fb 1434 PERL_ARGS_ASSERT_SSC_INIT;
7918f24d 1435
557bd3fb 1436 Zero(ssc, 1, regnode_ssc);
71068078 1437 set_ANYOF_SYNTHETIC(ssc);
93e92956 1438 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
3420edd7 1439 ssc_anything(ssc);
cdd87c1d 1440
2f306ab9
KW
1441 /* If any portion of the regex is to operate under locale rules that aren't
1442 * fully known at compile time, initialization includes it. The reason
1443 * this isn't done for all regexes is that the optimizer was written under
1444 * the assumption that locale was all-or-nothing. Given the complexity and
1445 * lack of documentation in the optimizer, and that there are inadequate
1446 * test cases for locale, many parts of it may not work properly, it is
1447 * safest to avoid locale unless necessary. */
cdd87c1d
KW
1448 if (RExC_contains_locale) {
1449 ANYOF_POSIXL_SETALL(ssc);
cdd87c1d
KW
1450 }
1451 else {
1452 ANYOF_POSIXL_ZERO(ssc);
1453 }
653099ff
GS
1454}
1455
b423522f 1456STATIC int
dc3bf405
BF
1457S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1458 const regnode_ssc *ssc)
b423522f
KW
1459{
1460 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1461 * to the list of code points matched, and locale posix classes; hence does
1462 * not check its flags) */
1463
1464 UV start, end;
1465 bool ret;
1466
1467 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1468
71068078 1469 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1470
1471 invlist_iterinit(ssc->invlist);
1472 ret = invlist_iternext(ssc->invlist, &start, &end)
1473 && start == 0
1474 && end == UV_MAX;
1475
1476 invlist_iterfinish(ssc->invlist);
1477
1478 if (! ret) {
1479 return FALSE;
1480 }
1481
e0e1be5f 1482 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
31f05a37 1483 return FALSE;
b423522f
KW
1484 }
1485
1486 return TRUE;
1487}
1488
1489STATIC SV*
1490S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
5c0f85ef 1491 const regnode_charclass* const node)
b423522f
KW
1492{
1493 /* Returns a mortal inversion list defining which code points are matched
1494 * by 'node', which is of type ANYOF. Handles complementing the result if
1495 * appropriate. If some code points aren't knowable at this time, the
31f05a37
KW
1496 * returned list must, and will, contain every code point that is a
1497 * possibility. */
b423522f 1498
e2506fa7 1499 SV* invlist = NULL;
1ee208c4 1500 SV* only_utf8_locale_invlist = NULL;
b423522f
KW
1501 unsigned int i;
1502 const U32 n = ARG(node);
31f05a37 1503 bool new_node_has_latin1 = FALSE;
b423522f
KW
1504
1505 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1506
1507 /* Look at the data structure created by S_set_ANYOF_arg() */
93e92956 1508 if (n != ANYOF_ONLY_HAS_BITMAP) {
b423522f
KW
1509 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1510 AV * const av = MUTABLE_AV(SvRV(rv));
1511 SV **const ary = AvARRAY(av);
1512 assert(RExC_rxi->data->what[n] == 's');
1513
1514 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
28118b9c 1515 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1]), NULL));
b423522f
KW
1516 }
1517 else if (ary[0] && ary[0] != &PL_sv_undef) {
1518
1519 /* Here, no compile-time swash, and there are things that won't be
1520 * known until runtime -- we have to assume it could be anything */
e2506fa7 1521 invlist = sv_2mortal(_new_invlist(1));
b423522f
KW
1522 return _add_range_to_invlist(invlist, 0, UV_MAX);
1523 }
1ee208c4 1524 else if (ary[3] && ary[3] != &PL_sv_undef) {
b423522f
KW
1525
1526 /* Here no compile-time swash, and no run-time only data. Use the
1527 * node's inversion list */
28118b9c 1528 invlist = sv_2mortal(invlist_clone(ary[3], NULL));
1ee208c4
KW
1529 }
1530
1531 /* Get the code points valid only under UTF-8 locales */
037715a6 1532 if ((ANYOF_FLAGS(node) & ANYOFL_FOLD)
1ee208c4
KW
1533 && ary[2] && ary[2] != &PL_sv_undef)
1534 {
1535 only_utf8_locale_invlist = ary[2];
b423522f
KW
1536 }
1537 }
1538
e2506fa7
KW
1539 if (! invlist) {
1540 invlist = sv_2mortal(_new_invlist(0));
1541 }
1542
dcb20b36
KW
1543 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1544 * code points, and an inversion list for the others, but if there are code
1545 * points that should match only conditionally on the target string being
1546 * UTF-8, those are placed in the inversion list, and not the bitmap.
1547 * Since there are circumstances under which they could match, they are
1548 * included in the SSC. But if the ANYOF node is to be inverted, we have
1549 * to exclude them here, so that when we invert below, the end result
1550 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1551 * have to do this here before we add the unconditionally matched code
1552 * points */
b423522f
KW
1553 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1554 _invlist_intersection_complement_2nd(invlist,
1555 PL_UpperLatin1,
1556 &invlist);
1557 }
1558
1559 /* Add in the points from the bit map */
dcb20b36 1560 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
b423522f 1561 if (ANYOF_BITMAP_TEST(node, i)) {
6f8848d5
TC
1562 unsigned int start = i++;
1563
1564 for (; i < NUM_ANYOF_CODE_POINTS && ANYOF_BITMAP_TEST(node, i); ++i) {
1565 /* empty */
1566 }
1567 invlist = _add_range_to_invlist(invlist, start, i-1);
31f05a37 1568 new_node_has_latin1 = TRUE;
b423522f
KW
1569 }
1570 }
1571
1572 /* If this can match all upper Latin1 code points, have to add them
ac33c516
KW
1573 * as well. But don't add them if inverting, as when that gets done below,
1574 * it would exclude all these characters, including the ones it shouldn't
1575 * that were added just above */
1576 if (! (ANYOF_FLAGS(node) & ANYOF_INVERT) && OP(node) == ANYOFD
f240c685
KW
1577 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1578 {
b423522f
KW
1579 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1580 }
1581
1582 /* Similarly for these */
93e92956 1583 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
e0a1ff7a 1584 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
b423522f
KW
1585 }
1586
1587 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1588 _invlist_invert(invlist);
1589 }
037715a6 1590 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOFL_FOLD) {
31f05a37
KW
1591
1592 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1593 * locale. We can skip this if there are no 0-255 at all. */
1594 _invlist_union(invlist, PL_Latin1, &invlist);
1595 }
1596
1ee208c4
KW
1597 /* Similarly add the UTF-8 locale possible matches. These have to be
1598 * deferred until after the non-UTF-8 locale ones are taken care of just
1599 * above, or it leads to wrong results under ANYOF_INVERT */
1600 if (only_utf8_locale_invlist) {
31f05a37 1601 _invlist_union_maybe_complement_2nd(invlist,
1ee208c4 1602 only_utf8_locale_invlist,
31f05a37
KW
1603 ANYOF_FLAGS(node) & ANYOF_INVERT,
1604 &invlist);
1605 }
b423522f
KW
1606
1607 return invlist;
1608}
1609
1051e1c4 1610/* These two functions currently do the exact same thing */
557bd3fb 1611#define ssc_init_zero ssc_init
653099ff 1612
cdd87c1d
KW
1613#define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1614#define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1615
557bd3fb 1616/* 'AND' a given class with another one. Can create false positives. 'ssc'
93e92956
KW
1617 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1618 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
cdd87c1d 1619
653099ff 1620STATIC void
b423522f 1621S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1622 const regnode_charclass *and_with)
653099ff 1623{
cdd87c1d
KW
1624 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1625 * another SSC or a regular ANYOF class. Can create false positives. */
40d049e4 1626
a0dd4231
KW
1627 SV* anded_cp_list;
1628 U8 anded_flags;
1e6ade67 1629
cdd87c1d 1630 PERL_ARGS_ASSERT_SSC_AND;
653099ff 1631
71068078 1632 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1633
1634 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1635 * the code point inversion list and just the relevant flags */
71068078 1636 if (is_ANYOF_SYNTHETIC(and_with)) {
7dcac5f6 1637 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
a0dd4231 1638 anded_flags = ANYOF_FLAGS(and_with);
e9b08962
KW
1639
1640 /* XXX This is a kludge around what appears to be deficiencies in the
1641 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1642 * there are paths through the optimizer where it doesn't get weeded
1643 * out when it should. And if we don't make some extra provision for
1644 * it like the code just below, it doesn't get added when it should.
1645 * This solution is to add it only when AND'ing, which is here, and
1646 * only when what is being AND'ed is the pristine, original node
1647 * matching anything. Thus it is like adding it to ssc_anything() but
1648 * only when the result is to be AND'ed. Probably the same solution
1649 * could be adopted for the same problem we have with /l matching,
1650 * which is solved differently in S_ssc_init(), and that would lead to
1651 * fewer false positives than that solution has. But if this solution
1652 * creates bugs, the consequences are only that a warning isn't raised
1653 * that should be; while the consequences for having /l bugs is
1654 * incorrect matches */
7dcac5f6 1655 if (ssc_is_anything((regnode_ssc *)and_with)) {
f240c685 1656 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
e9b08962 1657 }
a0dd4231
KW
1658 }
1659 else {
5c0f85ef 1660 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
f240c685
KW
1661 if (OP(and_with) == ANYOFD) {
1662 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1663 }
1664 else {
1665 anded_flags = ANYOF_FLAGS(and_with)
1666 &( ANYOF_COMMON_FLAGS
108316fb
KW
1667 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1668 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
d1c40ef5
KW
1669 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(and_with))) {
1670 anded_flags &=
1671 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1672 }
f240c685 1673 }
a0dd4231
KW
1674 }
1675
1676 ANYOF_FLAGS(ssc) &= anded_flags;
cdd87c1d
KW
1677
1678 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1679 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1680 * 'and_with' may be inverted. When not inverted, we have the situation of
1681 * computing:
1682 * (C1 | P1) & (C2 | P2)
1683 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1684 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1685 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1686 * <= ((C1 & C2) | P1 | P2)
1687 * Alternatively, the last few steps could be:
1688 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1689 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1690 * <= (C1 | C2 | (P1 & P2))
1691 * We favor the second approach if either P1 or P2 is non-empty. This is
1692 * because these components are a barrier to doing optimizations, as what
1693 * they match cannot be known until the moment of matching as they are
1694 * dependent on the current locale, 'AND"ing them likely will reduce or
1695 * eliminate them.
1696 * But we can do better if we know that C1,P1 are in their initial state (a
1697 * frequent occurrence), each matching everything:
1698 * (<everything>) & (C2 | P2) = C2 | P2
1699 * Similarly, if C2,P2 are in their initial state (again a frequent
1700 * occurrence), the result is a no-op
1701 * (C1 | P1) & (<everything>) = C1 | P1
1702 *
1703 * Inverted, we have
1704 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1705 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1706 * <= (C1 & ~C2) | (P1 & ~P2)
1707 * */
1aa99e6b 1708
a0dd4231 1709 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
71068078 1710 && ! is_ANYOF_SYNTHETIC(and_with))
a0dd4231 1711 {
cdd87c1d 1712 unsigned int i;
8951c461 1713
cdd87c1d
KW
1714 ssc_intersection(ssc,
1715 anded_cp_list,
1716 FALSE /* Has already been inverted */
1717 );
c6b76537 1718
cdd87c1d
KW
1719 /* If either P1 or P2 is empty, the intersection will be also; can skip
1720 * the loop */
93e92956 1721 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
cdd87c1d
KW
1722 ANYOF_POSIXL_ZERO(ssc);
1723 }
e0e1be5f 1724 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1725
1726 /* Note that the Posix class component P from 'and_with' actually
1727 * looks like:
1728 * P = Pa | Pb | ... | Pn
1729 * where each component is one posix class, such as in [\w\s].
1730 * Thus
1731 * ~P = ~(Pa | Pb | ... | Pn)
1732 * = ~Pa & ~Pb & ... & ~Pn
1733 * <= ~Pa | ~Pb | ... | ~Pn
1734 * The last is something we can easily calculate, but unfortunately
1735 * is likely to have many false positives. We could do better
1736 * in some (but certainly not all) instances if two classes in
1737 * P have known relationships. For example
1738 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1739 * So
1740 * :lower: & :print: = :lower:
1741 * And similarly for classes that must be disjoint. For example,
1742 * since \s and \w can have no elements in common based on rules in
1743 * the POSIX standard,
1744 * \w & ^\S = nothing
1745 * Unfortunately, some vendor locales do not meet the Posix
1746 * standard, in particular almost everything by Microsoft.
1747 * The loop below just changes e.g., \w into \W and vice versa */
1748
1ee208c4 1749 regnode_charclass_posixl temp;
cdd87c1d
KW
1750 int add = 1; /* To calculate the index of the complement */
1751
b1234259 1752 Zero(&temp, 1, regnode_charclass_posixl);
cdd87c1d
KW
1753 ANYOF_POSIXL_ZERO(&temp);
1754 for (i = 0; i < ANYOF_MAX; i++) {
1755 assert(i % 2 != 0
7dcac5f6
KW
1756 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1757 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
cdd87c1d 1758
7dcac5f6 1759 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
cdd87c1d
KW
1760 ANYOF_POSIXL_SET(&temp, i + add);
1761 }
1762 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1763 }
1764 ANYOF_POSIXL_AND(&temp, ssc);
c6b76537 1765
cdd87c1d
KW
1766 } /* else ssc already has no posixes */
1767 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1768 in its initial state */
71068078 1769 else if (! is_ANYOF_SYNTHETIC(and_with)
7dcac5f6 1770 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
cdd87c1d
KW
1771 {
1772 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1773 * copy it over 'ssc' */
1774 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
71068078 1775 if (is_ANYOF_SYNTHETIC(and_with)) {
cdd87c1d
KW
1776 StructCopy(and_with, ssc, regnode_ssc);
1777 }
1778 else {
1779 ssc->invlist = anded_cp_list;
1780 ANYOF_POSIXL_ZERO(ssc);
93e92956 1781 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1782 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
cdd87c1d
KW
1783 }
1784 }
1785 }
e0e1be5f 1786 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
93e92956 1787 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
cdd87c1d
KW
1788 {
1789 /* One or the other of P1, P2 is non-empty. */
93e92956 1790 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1ea8b7fe
KW
1791 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1792 }
cdd87c1d
KW
1793 ssc_union(ssc, anded_cp_list, FALSE);
1794 }
1795 else { /* P1 = P2 = empty */
1796 ssc_intersection(ssc, anded_cp_list, FALSE);
1797 }
137165a6 1798 }
653099ff
GS
1799}
1800
653099ff 1801STATIC void
cdd87c1d 1802S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
7dcac5f6 1803 const regnode_charclass *or_with)
653099ff 1804{
cdd87c1d
KW
1805 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1806 * another SSC or a regular ANYOF class. Can create false positives if
1807 * 'or_with' is to be inverted. */
7918f24d 1808
a0dd4231
KW
1809 SV* ored_cp_list;
1810 U8 ored_flags;
c6b76537 1811
cdd87c1d 1812 PERL_ARGS_ASSERT_SSC_OR;
c6b76537 1813
71068078 1814 assert(is_ANYOF_SYNTHETIC(ssc));
a0dd4231
KW
1815
1816 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1817 * the code point inversion list and just the relevant flags */
71068078 1818 if (is_ANYOF_SYNTHETIC(or_with)) {
7dcac5f6 1819 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
a0dd4231
KW
1820 ored_flags = ANYOF_FLAGS(or_with);
1821 }
1822 else {
5c0f85ef 1823 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
eff8b7dc 1824 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
f240c685
KW
1825 if (OP(or_with) != ANYOFD) {
1826 ored_flags
1827 |= ANYOF_FLAGS(or_with)
108316fb
KW
1828 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1829 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
d1c40ef5
KW
1830 if (ANYOFL_UTF8_LOCALE_REQD(ANYOF_FLAGS(or_with))) {
1831 ored_flags |=
1832 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1833 }
f240c685 1834 }
a0dd4231
KW
1835 }
1836
1837 ANYOF_FLAGS(ssc) |= ored_flags;
cdd87c1d
KW
1838
1839 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1840 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1841 * 'or_with' may be inverted. When not inverted, we have the simple
1842 * situation of computing:
1843 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1844 * If P1|P2 yields a situation with both a class and its complement are
1845 * set, like having both \w and \W, this matches all code points, and we
1846 * can delete these from the P component of the ssc going forward. XXX We
1847 * might be able to delete all the P components, but I (khw) am not certain
1848 * about this, and it is better to be safe.
1849 *
1850 * Inverted, we have
1851 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1852 * <= (C1 | P1) | ~C2
1853 * <= (C1 | ~C2) | P1
1854 * (which results in actually simpler code than the non-inverted case)
1855 * */
9826f543 1856
a0dd4231 1857 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
71068078 1858 && ! is_ANYOF_SYNTHETIC(or_with))
a0dd4231 1859 {
cdd87c1d 1860 /* We ignore P2, leaving P1 going forward */
1ea8b7fe 1861 } /* else Not inverted */
93e92956 1862 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
7dcac5f6 1863 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
e0e1be5f 1864 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
cdd87c1d
KW
1865 unsigned int i;
1866 for (i = 0; i < ANYOF_MAX; i += 2) {
1867 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1868 {
1869 ssc_match_all_cp(ssc);
1870 ANYOF_POSIXL_CLEAR(ssc, i);
1871 ANYOF_POSIXL_CLEAR(ssc, i+1);
cdd87c1d
KW
1872 }
1873 }
1874 }
1aa99e6b 1875 }
cdd87c1d
KW
1876
1877 ssc_union(ssc,
1878 ored_cp_list,
1879 FALSE /* Already has been inverted */
1880 );
653099ff
GS
1881}
1882
b423522f
KW
1883PERL_STATIC_INLINE void
1884S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1885{
1886 PERL_ARGS_ASSERT_SSC_UNION;
1887
71068078 1888 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1889
1890 _invlist_union_maybe_complement_2nd(ssc->invlist,
1891 invlist,
1892 invert2nd,
1893 &ssc->invlist);
1894}
1895
1896PERL_STATIC_INLINE void
1897S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1898 SV* const invlist,
1899 const bool invert2nd)
1900{
1901 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1902
71068078 1903 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1904
1905 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1906 invlist,
1907 invert2nd,
1908 &ssc->invlist);
1909}
1910
1911PERL_STATIC_INLINE void
1912S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1913{
1914 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1915
71068078 1916 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1917
1918 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1919}
1920
1921PERL_STATIC_INLINE void
1922S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1923{
1924 /* AND just the single code point 'cp' into the SSC 'ssc' */
1925
1926 SV* cp_list = _new_invlist(2);
1927
1928 PERL_ARGS_ASSERT_SSC_CP_AND;
1929
71068078 1930 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1931
1932 cp_list = add_cp_to_invlist(cp_list, cp);
1933 ssc_intersection(ssc, cp_list,
1934 FALSE /* Not inverted */
1935 );
1936 SvREFCNT_dec_NN(cp_list);
1937}
1938
1939PERL_STATIC_INLINE void
dc3bf405 1940S_ssc_clear_locale(regnode_ssc *ssc)
b423522f
KW
1941{
1942 /* Set the SSC 'ssc' to not match any locale things */
b423522f
KW
1943 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1944
71068078 1945 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f
KW
1946
1947 ANYOF_POSIXL_ZERO(ssc);
1948 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1949}
1950
b35552de
KW
1951#define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1952
1953STATIC bool
1954S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1955{
1956 /* The synthetic start class is used to hopefully quickly winnow down
1957 * places where a pattern could start a match in the target string. If it
1958 * doesn't really narrow things down that much, there isn't much point to
1959 * having the overhead of using it. This function uses some very crude
1960 * heuristics to decide if to use the ssc or not.
1961 *
1962 * It returns TRUE if 'ssc' rules out more than half what it considers to
1963 * be the "likely" possible matches, but of course it doesn't know what the
1964 * actual things being matched are going to be; these are only guesses
1965 *
1966 * For /l matches, it assumes that the only likely matches are going to be
1967 * in the 0-255 range, uniformly distributed, so half of that is 127
1968 * For /a and /d matches, it assumes that the likely matches will be just
1969 * the ASCII range, so half of that is 63
1970 * For /u and there isn't anything matching above the Latin1 range, it
1971 * assumes that that is the only range likely to be matched, and uses
1972 * half that as the cut-off: 127. If anything matches above Latin1,
1973 * it assumes that all of Unicode could match (uniformly), except for
1974 * non-Unicode code points and things in the General Category "Other"
1975 * (unassigned, private use, surrogates, controls and formats). This
1976 * is a much large number. */
1977
b35552de
KW
1978 U32 count = 0; /* Running total of number of code points matched by
1979 'ssc' */
1980 UV start, end; /* Start and end points of current range in inversion
1981 list */
72400949
KW
1982 const U32 max_code_points = (LOC)
1983 ? 256
1984 : (( ! UNI_SEMANTICS
1985 || invlist_highest(ssc->invlist) < 256)
1986 ? 128
1987 : NON_OTHER_COUNT);
1988 const U32 max_match = max_code_points / 2;
b35552de
KW
1989
1990 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1991
1992 invlist_iterinit(ssc->invlist);
1993 while (invlist_iternext(ssc->invlist, &start, &end)) {
72400949
KW
1994 if (start >= max_code_points) {
1995 break;
b35552de 1996 }
72400949 1997 end = MIN(end, max_code_points - 1);
b35552de 1998 count += end - start + 1;
72400949 1999 if (count >= max_match) {
b35552de
KW
2000 invlist_iterfinish(ssc->invlist);
2001 return FALSE;
2002 }
2003 }
2004
2005 return TRUE;
2006}
2007
2008
b423522f
KW
2009STATIC void
2010S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2011{
2012 /* The inversion list in the SSC is marked mortal; now we need a more
2013 * permanent copy, which is stored the same way that is done in a regular
dcb20b36
KW
2014 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2015 * map */
b423522f 2016
28118b9c 2017 SV* invlist = invlist_clone(ssc->invlist, NULL);
b423522f
KW
2018
2019 PERL_ARGS_ASSERT_SSC_FINALIZE;
2020
71068078 2021 assert(is_ANYOF_SYNTHETIC(ssc));
b423522f 2022
a0dd4231 2023 /* The code in this file assumes that all but these flags aren't relevant
93e92956
KW
2024 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2025 * by the time we reach here */
f240c685
KW
2026 assert(! (ANYOF_FLAGS(ssc)
2027 & ~( ANYOF_COMMON_FLAGS
108316fb
KW
2028 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2029 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
a0dd4231 2030
b423522f
KW
2031 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2032
1ee208c4
KW
2033 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
2034 NULL, NULL, NULL, FALSE);
b423522f 2035
85c8e306
KW
2036 /* Make sure is clone-safe */
2037 ssc->invlist = NULL;
2038
e0e1be5f 2039 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
93e92956 2040 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
e0e1be5f 2041 }
1462525b 2042
b2e90ddf
KW
2043 if (RExC_contains_locale) {
2044 OP(ssc) = ANYOFL;
2045 }
2046
1462525b 2047 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
b423522f
KW
2048}
2049
a3621e74
YO
2050#define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2051#define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2052#define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
538e84ed
KW
2053#define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2054 ? (TRIE_LIST_CUR( idx ) - 1) \
2055 : 0 )
a3621e74 2056
3dab1dad
YO
2057
2058#ifdef DEBUGGING
07be1b83 2059/*
2b8b4781
NC
2060 dump_trie(trie,widecharmap,revcharmap)
2061 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2062 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
3dab1dad
YO
2063
2064 These routines dump out a trie in a somewhat readable format.
07be1b83
YO
2065 The _interim_ variants are used for debugging the interim
2066 tables that are used to generate the final compressed
2067 representation which is what dump_trie expects.
2068
486ec47a 2069 Part of the reason for their existence is to provide a form
3dab1dad 2070 of documentation as to how the different representations function.
07be1b83
YO
2071
2072*/
3dab1dad
YO
2073
2074/*
3dab1dad
YO
2075 Dumps the final compressed table form of the trie to Perl_debug_log.
2076 Used for debugging make_trie().
2077*/
b9a59e08 2078
3dab1dad 2079STATIC void
2b8b4781
NC
2080S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2081 AV *revcharmap, U32 depth)
3dab1dad
YO
2082{
2083 U32 state;
ab3bbdeb 2084 SV *sv=sv_newmortal();
55eed653 2085 int colwidth= widecharmap ? 6 : 4;
2e64971a 2086 U16 word;
3dab1dad
YO
2087 GET_RE_DEBUG_FLAGS_DECL;
2088
7918f24d 2089 PERL_ARGS_ASSERT_DUMP_TRIE;
ab3bbdeb 2090
6ad9a8ab 2091 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
1e37780e 2092 depth+1, "Match","Base","Ofs" );
3dab1dad
YO
2093
2094 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2b8b4781 2095 SV ** const tmp = av_fetch( revcharmap, state, 0);
3dab1dad 2096 if ( tmp ) {
6ad9a8ab 2097 Perl_re_printf( aTHX_ "%*s",
ab3bbdeb 2098 colwidth,
538e84ed 2099 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
2100 PL_colors[0], PL_colors[1],
2101 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
2102 PERL_PV_ESCAPE_FIRSTCHAR
2103 )
ab3bbdeb 2104 );
3dab1dad
YO
2105 }
2106 }
1e37780e
YO
2107 Perl_re_printf( aTHX_ "\n");
2108 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
3dab1dad
YO
2109
2110 for( state = 0 ; state < trie->uniquecharcount ; state++ )
6ad9a8ab
YO
2111 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2112 Perl_re_printf( aTHX_ "\n");
3dab1dad 2113
1e2e3d02 2114 for( state = 1 ; state < trie->statecount ; state++ ) {
be8e71aa 2115 const U32 base = trie->states[ state ].trans.base;
3dab1dad 2116
147e3846 2117 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
3dab1dad
YO
2118
2119 if ( trie->states[ state ].wordnum ) {
1e37780e 2120 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
3dab1dad 2121 } else {
6ad9a8ab 2122 Perl_re_printf( aTHX_ "%6s", "" );
3dab1dad
YO
2123 }
2124
147e3846 2125 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
3dab1dad
YO
2126
2127 if ( base ) {
2128 U32 ofs = 0;
2129
2130 while( ( base + ofs < trie->uniquecharcount ) ||
2131 ( base + ofs - trie->uniquecharcount < trie->lasttrans
538e84ed
KW
2132 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2133 != state))
3dab1dad
YO
2134 ofs++;
2135
147e3846 2136 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
3dab1dad
YO
2137
2138 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
538e84ed
KW
2139 if ( ( base + ofs >= trie->uniquecharcount )
2140 && ( base + ofs - trie->uniquecharcount
2141 < trie->lasttrans )
2142 && trie->trans[ base + ofs
2143 - trie->uniquecharcount ].check == state )
3dab1dad 2144 {
147e3846 2145 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
1e37780e
YO
2146 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2147 );
3dab1dad 2148 } else {
6ad9a8ab 2149 Perl_re_printf( aTHX_ "%*s",colwidth," ." );
3dab1dad
YO
2150 }
2151 }
2152
6ad9a8ab 2153 Perl_re_printf( aTHX_ "]");
3dab1dad
YO
2154
2155 }
6ad9a8ab 2156 Perl_re_printf( aTHX_ "\n" );
3dab1dad 2157 }
6ad9a8ab 2158 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
cb41e5d6 2159 depth);
2e64971a 2160 for (word=1; word <= trie->wordcount; word++) {
6ad9a8ab 2161 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2e64971a
DM
2162 (int)word, (int)(trie->wordinfo[word].prev),
2163 (int)(trie->wordinfo[word].len));
2164 }
6ad9a8ab 2165 Perl_re_printf( aTHX_ "\n" );
538e84ed 2166}
3dab1dad 2167/*
3dab1dad 2168 Dumps a fully constructed but uncompressed trie in list form.
538e84ed 2169 List tries normally only are used for construction when the number of
3dab1dad
YO
2170 possible chars (trie->uniquecharcount) is very high.
2171 Used for debugging make_trie().
2172*/
2173STATIC void
55eed653 2174S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
2175 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2176 U32 depth)
3dab1dad
YO
2177{
2178 U32 state;
ab3bbdeb 2179 SV *sv=sv_newmortal();
55eed653 2180 int colwidth= widecharmap ? 6 : 4;
3dab1dad 2181 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2182
2183 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2184
3dab1dad 2185 /* print out the table precompression. */
6ad9a8ab 2186 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
cb41e5d6 2187 depth+1 );
6ad9a8ab 2188 Perl_re_indentf( aTHX_ "%s",
cb41e5d6 2189 depth+1, "------:-----+-----------------\n" );
538e84ed 2190
3dab1dad
YO
2191 for( state=1 ; state < next_alloc ; state ++ ) {
2192 U16 charid;
538e84ed 2193
147e3846 2194 Perl_re_indentf( aTHX_ " %4" UVXf " :",
cb41e5d6 2195 depth+1, (UV)state );
3dab1dad 2196 if ( ! trie->states[ state ].wordnum ) {
6ad9a8ab 2197 Perl_re_printf( aTHX_ "%5s| ","");
3dab1dad 2198 } else {
6ad9a8ab 2199 Perl_re_printf( aTHX_ "W%4x| ",
3dab1dad
YO
2200 trie->states[ state ].wordnum
2201 );
2202 }
2203 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
538e84ed
KW
2204 SV ** const tmp = av_fetch( revcharmap,
2205 TRIE_LIST_ITEM(state,charid).forid, 0);
ab3bbdeb 2206 if ( tmp ) {
147e3846 2207 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
ab3bbdeb 2208 colwidth,
538e84ed
KW
2209 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2210 colwidth,
2211 PL_colors[0], PL_colors[1],
2212 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2213 | PERL_PV_ESCAPE_FIRSTCHAR
ab3bbdeb 2214 ) ,
1e2e3d02
YO
2215 TRIE_LIST_ITEM(state,charid).forid,
2216 (UV)TRIE_LIST_ITEM(state,charid).newstate
2217 );
538e84ed 2218 if (!(charid % 10))
6ad9a8ab 2219 Perl_re_printf( aTHX_ "\n%*s| ",
664e119d 2220 (int)((depth * 2) + 14), "");
1e2e3d02 2221 }
ab3bbdeb 2222 }
6ad9a8ab 2223 Perl_re_printf( aTHX_ "\n");
3dab1dad 2224 }
538e84ed 2225}
3dab1dad
YO
2226
2227/*
3dab1dad 2228 Dumps a fully constructed but uncompressed trie in table form.
538e84ed
KW
2229 This is the normal DFA style state transition table, with a few
2230 twists to facilitate compression later.
3dab1dad
YO
2231 Used for debugging make_trie().
2232*/
2233STATIC void
55eed653 2234S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2b8b4781
NC
2235 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2236 U32 depth)
3dab1dad
YO
2237{
2238 U32 state;
2239 U16 charid;
ab3bbdeb 2240 SV *sv=sv_newmortal();
55eed653 2241 int colwidth= widecharmap ? 6 : 4;
3dab1dad 2242 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2243
2244 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
538e84ed 2245
3dab1dad
YO
2246 /*
2247 print out the table precompression so that we can do a visual check
2248 that they are identical.
2249 */
538e84ed 2250
6ad9a8ab 2251 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
3dab1dad
YO
2252
2253 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2b8b4781 2254 SV ** const tmp = av_fetch( revcharmap, charid, 0);
3dab1dad 2255 if ( tmp ) {
6ad9a8ab 2256 Perl_re_printf( aTHX_ "%*s",
ab3bbdeb 2257 colwidth,
538e84ed 2258 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
ab3bbdeb
YO
2259 PL_colors[0], PL_colors[1],
2260 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed
KW
2261 PERL_PV_ESCAPE_FIRSTCHAR
2262 )
ab3bbdeb 2263 );
3dab1dad
YO
2264 }
2265 }
2266
4aaafc03
YO
2267 Perl_re_printf( aTHX_ "\n");
2268 Perl_re_indentf( aTHX_ "State+-", depth+1 );
3dab1dad
YO
2269
2270 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
6ad9a8ab 2271 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
3dab1dad
YO
2272 }
2273
6ad9a8ab 2274 Perl_re_printf( aTHX_ "\n" );
3dab1dad
YO
2275
2276 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2277
147e3846 2278 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
cb41e5d6 2279 depth+1,
3dab1dad
YO
2280 (UV)TRIE_NODENUM( state ) );
2281
2282 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
ab3bbdeb
YO
2283 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2284 if (v)
147e3846 2285 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
ab3bbdeb 2286 else
6ad9a8ab 2287 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
3dab1dad
YO
2288 }
2289 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
147e3846 2290 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
538e84ed 2291 (UV)trie->trans[ state ].check );
3dab1dad 2292 } else {
147e3846 2293 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
538e84ed 2294 (UV)trie->trans[ state ].check,
3dab1dad
YO
2295 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2296 }
2297 }
07be1b83 2298}
3dab1dad
YO
2299
2300#endif
2301
2e64971a 2302
786e8c11
YO
2303/* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2304 startbranch: the first branch in the whole branch sequence
2305 first : start branch of sequence of branch-exact nodes.
2306 May be the same as startbranch
2307 last : Thing following the last branch.
2308 May be the same as tail.
2309 tail : item following the branch sequence
2310 count : words in the sequence
a4525e78 2311 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
786e8c11 2312 depth : indent depth
3dab1dad 2313
786e8c11 2314Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
07be1b83 2315
786e8c11
YO
2316A trie is an N'ary tree where the branches are determined by digital
2317decomposition of the key. IE, at the root node you look up the 1st character and
2318follow that branch repeat until you find the end of the branches. Nodes can be
2319marked as "accepting" meaning they represent a complete word. Eg:
07be1b83 2320
786e8c11 2321 /he|she|his|hers/
72f13be8 2322
786e8c11
YO
2323would convert into the following structure. Numbers represent states, letters
2324following numbers represent valid transitions on the letter from that state, if
2325the number is in square brackets it represents an accepting state, otherwise it
2326will be in parenthesis.
07be1b83 2327
786e8c11
YO
2328 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2329 | |
2330 | (2)
2331 | |
2332 (1) +-i->(6)-+-s->[7]
2333 |
2334 +-s->(3)-+-h->(4)-+-e->[5]
07be1b83 2335
786e8c11
YO
2336 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2337
2338This shows that when matching against the string 'hers' we will begin at state 1
2339read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2340then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2341is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2342single traverse. We store a mapping from accepting to state to which word was
2343matched, and then when we have multiple possibilities we try to complete the
b8fda935 2344rest of the regex in the order in which they occurred in the alternation.
786e8c11
YO
2345
2346The only prior NFA like behaviour that would be changed by the TRIE support is
2347the silent ignoring of duplicate alternations which are of the form:
2348
2349 / (DUPE|DUPE) X? (?{ ... }) Y /x
2350
4b714af6 2351Thus EVAL blocks following a trie may be called a different number of times with
786e8c11 2352and without the optimisation. With the optimisations dupes will be silently
486ec47a 2353ignored. This inconsistent behaviour of EVAL type nodes is well established as
786e8c11
YO
2354the following demonstrates:
2355
2356 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2357
2358which prints out 'word' three times, but
2359
2360 'words'=~/(word|word|word)(?{ print $1 })S/
2361
2362which doesnt print it out at all. This is due to other optimisations kicking in.
2363
2364Example of what happens on a structural level:
2365
486ec47a 2366The regexp /(ac|ad|ab)+/ will produce the following debug output:
786e8c11
YO
2367
2368 1: CURLYM[1] {1,32767}(18)
2369 5: BRANCH(8)
2370 6: EXACT <ac>(16)
2371 8: BRANCH(11)
2372 9: EXACT <ad>(16)
2373 11: BRANCH(14)
2374 12: EXACT <ab>(16)
2375 16: SUCCEED(0)
2376 17: NOTHING(18)
2377 18: END(0)
2378
2379This would be optimizable with startbranch=5, first=5, last=16, tail=16
2380and should turn into:
2381
2382 1: CURLYM[1] {1,32767}(18)
2383 5: TRIE(16)
2384 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2385 <ac>
2386 <ad>
2387 <ab>
2388 16: SUCCEED(0)
2389 17: NOTHING(18)
2390 18: END(0)
2391
2392Cases where tail != last would be like /(?foo|bar)baz/:
2393
2394 1: BRANCH(4)
2395 2: EXACT <foo>(8)
2396 4: BRANCH(7)
2397 5: EXACT <bar>(8)
2398 7: TAIL(8)
2399 8: EXACT <baz>(10)
2400 10: END(0)
2401
2402which would be optimizable with startbranch=1, first=1, last=7, tail=8
2403and would end up looking like:
2404
2405 1: TRIE(8)
2406 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2407 <foo>
2408 <bar>
2409 7: TAIL(8)
2410 8: EXACT <baz>(10)
2411 10: END(0)
2412
c80e42f3 2413 d = uvchr_to_utf8_flags(d, uv, 0);
786e8c11
YO
2414
2415is the recommended Unicode-aware way of saying
2416
2417 *(d++) = uv;
2418*/
2419
fab2782b 2420#define TRIE_STORE_REVCHAR(val) \
786e8c11 2421 STMT_START { \
73031816 2422 if (UTF) { \
668fcfea 2423 SV *zlopp = newSV(UTF8_MAXBYTES); \
88c9ea1e 2424 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
c80e42f3 2425 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
73031816
NC
2426 SvCUR_set(zlopp, kapow - flrbbbbb); \
2427 SvPOK_on(zlopp); \
2428 SvUTF8_on(zlopp); \
2429 av_push(revcharmap, zlopp); \
2430 } else { \
fab2782b 2431 char ooooff = (char)val; \
73031816
NC
2432 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2433 } \
2434 } STMT_END
786e8c11 2435
914a25d5
KW
2436/* This gets the next character from the input, folding it if not already
2437 * folded. */
2438#define TRIE_READ_CHAR STMT_START { \
2439 wordlen++; \
2440 if ( UTF ) { \
2441 /* if it is UTF then it is either already folded, or does not need \
2442 * folding */ \
1c1d615a 2443 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
914a25d5
KW
2444 } \
2445 else if (folder == PL_fold_latin1) { \
7d006b13
KW
2446 /* This folder implies Unicode rules, which in the range expressible \
2447 * by not UTF is the lower case, with the two exceptions, one of \
2448 * which should have been taken care of before calling this */ \
2449 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2450 uvc = toLOWER_L1(*uc); \
2451 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2452 len = 1; \
914a25d5
KW
2453 } else { \
2454 /* raw data, will be folded later if needed */ \
2455 uvc = (U32)*uc; \
2456 len = 1; \
2457 } \
786e8c11
YO
2458} STMT_END
2459
2460
2461
2462#define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2463 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
00195859 2464 U32 ging = TRIE_LIST_LEN( state ) * 2; \
f9003953 2465 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
00195859 2466 TRIE_LIST_LEN( state ) = ging; \
786e8c11
YO
2467 } \
2468 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2469 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2470 TRIE_LIST_CUR( state )++; \
2471} STMT_END
07be1b83 2472
786e8c11 2473#define TRIE_LIST_NEW(state) STMT_START { \
d09f14bf 2474 Newx( trie->states[ state ].trans.list, \
786e8c11
YO
2475 4, reg_trie_trans_le ); \
2476 TRIE_LIST_CUR( state ) = 1; \
2477 TRIE_LIST_LEN( state ) = 4; \
2478} STMT_END
07be1b83 2479
786e8c11
YO
2480#define TRIE_HANDLE_WORD(state) STMT_START { \
2481 U16 dupe= trie->states[ state ].wordnum; \
2482 regnode * const noper_next = regnext( noper ); \
2483 \
786e8c11
YO
2484 DEBUG_r({ \
2485 /* store the word for dumping */ \
2486 SV* tmp; \
2487 if (OP(noper) != NOTHING) \
740cce10 2488 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
786e8c11 2489 else \
740cce10 2490 tmp = newSVpvn_utf8( "", 0, UTF ); \
2b8b4781 2491 av_push( trie_words, tmp ); \
786e8c11
YO
2492 }); \
2493 \
2494 curword++; \
2e64971a
DM
2495 trie->wordinfo[curword].prev = 0; \
2496 trie->wordinfo[curword].len = wordlen; \
2497 trie->wordinfo[curword].accept = state; \
786e8c11
YO
2498 \
2499 if ( noper_next < tail ) { \
2500 if (!trie->jump) \
538e84ed
KW
2501 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2502 sizeof(U16) ); \
7f69552c 2503 trie->jump[curword] = (U16)(noper_next - convert); \
786e8c11
YO
2504 if (!jumper) \
2505 jumper = noper_next; \
2506 if (!nextbranch) \
2507 nextbranch= regnext(cur); \
2508 } \
2509 \
2510 if ( dupe ) { \
2e64971a
DM
2511 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2512 /* chain, so that when the bits of chain are later */\
2513 /* linked together, the dups appear in the chain */\
2514 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2515 trie->wordinfo[dupe].prev = curword; \
786e8c11
YO
2516 } else { \
2517 /* we haven't inserted this word yet. */ \
2518 trie->states[ state ].wordnum = curword; \
2519 } \
2520} STMT_END
07be1b83 2521
3dab1dad 2522
786e8c11
YO
2523#define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2524 ( ( base + charid >= ucharcount \
2525 && base + charid < ubound \
2526 && state == trie->trans[ base - ucharcount + charid ].check \
2527 && trie->trans[ base - ucharcount + charid ].next ) \
2528 ? trie->trans[ base - ucharcount + charid ].next \
2529 : ( state==1 ? special : 0 ) \
2530 )
3dab1dad 2531
8bcafbf4
YO
2532#define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2533STMT_START { \
2534 TRIE_BITMAP_SET(trie, uvc); \
2535 /* store the folded codepoint */ \
2536 if ( folder ) \
2537 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2538 \
2539 if ( !UTF ) { \
2540 /* store first byte of utf8 representation of */ \
2541 /* variant codepoints */ \
2542 if (! UVCHR_IS_INVARIANT(uvc)) { \
2543 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2544 } \
2545 } \
2546} STMT_END
786e8c11
YO
2547#define MADE_TRIE 1
2548#define MADE_JUMP_TRIE 2
2549#define MADE_EXACT_TRIE 4
3dab1dad 2550
a3621e74 2551STATIC I32
538e84ed
KW
2552S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2553 regnode *first, regnode *last, regnode *tail,
2554 U32 word_count, U32 flags, U32 depth)
a3621e74
YO
2555{
2556 /* first pass, loop through and scan words */
2557 reg_trie_data *trie;
55eed653 2558 HV *widecharmap = NULL;
2b8b4781 2559 AV *revcharmap = newAV();
a3621e74 2560 regnode *cur;
a3621e74
YO
2561 STRLEN len = 0;
2562 UV uvc = 0;
2563 U16 curword = 0;
2564 U32 next_alloc = 0;
786e8c11
YO
2565 regnode *jumper = NULL;
2566 regnode *nextbranch = NULL;
7f69552c 2567 regnode *convert = NULL;
2e64971a 2568 U32 *prev_states; /* temp array mapping each state to previous one */
a3621e74 2569 /* we just use folder as a flag in utf8 */
1e696034 2570 const U8 * folder = NULL;
a3621e74 2571
3a611511
YO
2572 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2573 * which stands for one trie structure, one hash, optionally followed
2574 * by two arrays */
2b8b4781 2575#ifdef DEBUGGING
3a611511 2576 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2b8b4781
NC
2577 AV *trie_words = NULL;
2578 /* along with revcharmap, this only used during construction but both are
2579 * useful during debugging so we store them in the struct when debugging.
8e11feef 2580 */
2b8b4781 2581#else
cf78de0b 2582 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
3dab1dad 2583 STRLEN trie_charcount=0;
3dab1dad 2584#endif
2b8b4781 2585 SV *re_trie_maxbuff;
a3621e74 2586 GET_RE_DEBUG_FLAGS_DECL;
7918f24d
NC
2587
2588 PERL_ARGS_ASSERT_MAKE_TRIE;
72f13be8
YO
2589#ifndef DEBUGGING
2590 PERL_UNUSED_ARG(depth);
2591#endif
a3621e74 2592
1e696034 2593 switch (flags) {
a4525e78 2594 case EXACT: case EXACTL: break;
89829bb5 2595 case EXACTFAA:
fab2782b 2596 case EXACTFU_SS:
a4525e78
KW
2597 case EXACTFU:
2598 case EXACTFLU8: folder = PL_fold_latin1; break;
1e696034 2599 case EXACTF: folder = PL_fold; break;
fab2782b 2600 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1e696034
KW
2601 }
2602
c944940b 2603 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
a3621e74 2604 trie->refcount = 1;
3dab1dad 2605 trie->startstate = 1;
786e8c11 2606 trie->wordcount = word_count;
f8fc2ecf 2607 RExC_rxi->data->data[ data_slot ] = (void*)trie;
c944940b 2608 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
a4525e78 2609 if (flags == EXACT || flags == EXACTL)
c944940b 2610 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2e64971a
DM
2611 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2612 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2613
a3621e74 2614 DEBUG_r({
2b8b4781 2615 trie_words = newAV();
a3621e74 2616 });
a3621e74 2617
0111c4fd 2618 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
316ebaf2 2619 assert(re_trie_maxbuff);
a3621e74 2620 if (!SvIOK(re_trie_maxbuff)) {
0111c4fd 2621 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
a3621e74 2622 }
df826430 2623 DEBUG_TRIE_COMPILE_r({
6ad9a8ab 2624 Perl_re_indentf( aTHX_
cb41e5d6
YO
2625 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2626 depth+1,
538e84ed
KW
2627 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2628 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
3dab1dad 2629 });
538e84ed 2630
7f69552c
YO
2631 /* Find the node we are going to overwrite */
2632 if ( first == startbranch && OP( last ) != BRANCH ) {
2633 /* whole branch chain */
2634 convert = first;
2635 } else {
2636 /* branch sub-chain */
2637 convert = NEXTOPER( first );
2638 }
538e84ed 2639
a3621e74
YO
2640 /* -- First loop and Setup --
2641
2642 We first traverse the branches and scan each word to determine if it
2643 contains widechars, and how many unique chars there are, this is
2644 important as we have to build a table with at least as many columns as we
2645 have unique chars.
2646
2647 We use an array of integers to represent the character codes 0..255
538e84ed
KW
2648 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2649 the native representation of the character value as the key and IV's for
2650 the coded index.
a3621e74
YO
2651
2652 *TODO* If we keep track of how many times each character is used we can
2653 remap the columns so that the table compression later on is more
3b753521 2654 efficient in terms of memory by ensuring the most common value is in the
a3621e74
YO
2655 middle and the least common are on the outside. IMO this would be better
2656 than a most to least common mapping as theres a decent chance the most
2657 common letter will share a node with the least common, meaning the node
486ec47a 2658 will not be compressible. With a middle is most common approach the worst
a3621e74
YO
2659 case is when we have the least common nodes twice.
2660
2661 */
2662
a3621e74 2663 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
df826430 2664 regnode *noper = NEXTOPER( cur );
944e05e3
YO
2665 const U8 *uc;
2666 const U8 *e;
bc031a7d 2667 int foldlen = 0;
07be1b83 2668 U32 wordlen = 0; /* required init */
bc031a7d
KW
2669 STRLEN minchars = 0;
2670 STRLEN maxchars = 0;
538e84ed
KW
2671 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2672 bitmap?*/
a3621e74 2673
3dab1dad 2674 if (OP(noper) == NOTHING) {
20ed8c88
YO
2675 /* skip past a NOTHING at the start of an alternation
2676 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2677 */
df826430 2678 regnode *noper_next= regnext(noper);
944e05e3
YO
2679 if (noper_next < tail)
2680 noper= noper_next;
2681 }
2682
dca5fc4c
YO
2683 if ( noper < tail &&
2684 (
2685 OP(noper) == flags ||
2686 (
2687 flags == EXACTFU &&
2688 OP(noper) == EXACTFU_SS
2689 )
2690 )
2691 ) {
944e05e3
YO
2692 uc= (U8*)STRING(noper);
2693 e= uc + STR_LEN(noper);
2694 } else {
2695 trie->minlen= 0;
2696 continue;
3dab1dad 2697 }
df826430 2698
944e05e3 2699
fab2782b 2700 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
02daf0ab
YO
2701 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2702 regardless of encoding */
fab2782b
YO
2703 if (OP( noper ) == EXACTFU_SS) {
2704 /* false positives are ok, so just set this */
0dc4a61d 2705 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
fab2782b
YO
2706 }
2707 }
dca5fc4c 2708
bc031a7d
KW
2709 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2710 branch */
3dab1dad 2711 TRIE_CHARCOUNT(trie)++;
a3621e74 2712 TRIE_READ_CHAR;
645de4ce 2713
bc031a7d
KW
2714 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2715 * is in effect. Under /i, this character can match itself, or
2716 * anything that folds to it. If not under /i, it can match just
2717 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2718 * all fold to k, and all are single characters. But some folds
2719 * expand to more than one character, so for example LATIN SMALL
2720 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2721 * the string beginning at 'uc' is 'ffi', it could be matched by
2722 * three characters, or just by the one ligature character. (It
2723 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2724 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2725 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2726 * match.) The trie needs to know the minimum and maximum number
2727 * of characters that could match so that it can use size alone to
2728 * quickly reject many match attempts. The max is simple: it is
2729 * the number of folded characters in this branch (since a fold is
2730 * never shorter than what folds to it. */
2731
2732 maxchars++;
2733
2734 /* And the min is equal to the max if not under /i (indicated by
2735 * 'folder' being NULL), or there are no multi-character folds. If
2736 * there is a multi-character fold, the min is incremented just
2737 * once, for the character that folds to the sequence. Each
2738 * character in the sequence needs to be added to the list below of
2739 * characters in the trie, but we count only the first towards the
2740 * min number of characters needed. This is done through the
2741 * variable 'foldlen', which is returned by the macros that look
2742 * for these sequences as the number of bytes the sequence
2743 * occupies. Each time through the loop, we decrement 'foldlen' by
2744 * how many bytes the current char occupies. Only when it reaches
2745 * 0 do we increment 'minchars' or look for another multi-character
2746 * sequence. */
2747 if (folder == NULL) {
2748 minchars++;
2749 }
2750 else if (foldlen > 0) {
2751 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
645de4ce
KW
2752 }
2753 else {
bc031a7d
KW
2754 minchars++;
2755
2756 /* See if *uc is the beginning of a multi-character fold. If
2757 * so, we decrement the length remaining to look at, to account
2758 * for the current character this iteration. (We can use 'uc'
2759 * instead of the fold returned by TRIE_READ_CHAR because for
2760 * non-UTF, the latin1_safe macro is smart enough to account
2761 * for all the unfolded characters, and because for UTF, the
2762 * string will already have been folded earlier in the
2763 * compilation process */
2764 if (UTF) {
2765 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2766 foldlen -= UTF8SKIP(uc);
645de4ce
KW
2767 }
2768 }
bc031a7d
KW
2769 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2770 foldlen--;
2771 }
645de4ce 2772 }
bc031a7d
KW
2773
2774 /* The current character (and any potential folds) should be added
2775 * to the possible matching characters for this position in this
2776 * branch */
a3621e74 2777 if ( uvc < 256 ) {
fab2782b
YO
2778 if ( folder ) {
2779 U8 folded= folder[ (U8) uvc ];
2780 if ( !trie->charmap[ folded ] ) {
2781 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2782 TRIE_STORE_REVCHAR( folded );
2783 }
2784 }
a3621e74
YO
2785 if ( !trie->charmap[ uvc ] ) {
2786 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
fab2782b 2787 TRIE_STORE_REVCHAR( uvc );
a3621e74 2788 }
02daf0ab 2789 if ( set_bit ) {
62012aee
KW
2790 /* store the codepoint in the bitmap, and its folded
2791 * equivalent. */
8bcafbf4 2792 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
02daf0ab
YO
2793 set_bit = 0; /* We've done our bit :-) */
2794 }
a3621e74 2795 } else {
bc031a7d
KW
2796
2797 /* XXX We could come up with the list of code points that fold
2798 * to this using PL_utf8_foldclosures, except not for
2799 * multi-char folds, as there may be multiple combinations
2800 * there that could work, which needs to wait until runtime to
2801 * resolve (The comment about LIGATURE FFI above is such an
2802 * example */
2803
a3621e74 2804 SV** svpp;
55eed653
NC
2805 if ( !widecharmap )
2806 widecharmap = newHV();
a3621e74 2807
55eed653 2808 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
a3621e74
YO
2809
2810 if ( !svpp )
147e3846 2811 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
a3621e74
YO
2812
2813 if ( !SvTRUE( *svpp ) ) {
2814 sv_setiv( *svpp, ++trie->uniquecharcount );
fab2782b 2815 TRIE_STORE_REVCHAR(uvc);
a3621e74
YO
2816 }
2817 }
bc031a7d
KW
2818 } /* end loop through characters in this branch of the trie */
2819
2820 /* We take the min and max for this branch and combine to find the min
2821 * and max for all branches processed so far */
3dab1dad 2822 if( cur == first ) {
bc031a7d
KW
2823 trie->minlen = minchars;
2824 trie->maxlen = maxchars;
2825 } else if (minchars < trie->minlen) {
2826 trie->minlen = minchars;
2827 } else if (maxchars > trie->maxlen) {
2828 trie->maxlen = maxchars;
fab2782b 2829 }
a3621e74
YO
2830 } /* end first pass */
2831 DEBUG_TRIE_COMPILE_r(
6ad9a8ab 2832 Perl_re_indentf( aTHX_
cb41e5d6
YO
2833 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2834 depth+1,
55eed653 2835 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
be8e71aa
YO
2836 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2837 (int)trie->minlen, (int)trie->maxlen )
a3621e74 2838 );
a3621e74
YO
2839
2840 /*
2841 We now know what we are dealing with in terms of unique chars and
2842 string sizes so we can calculate how much memory a naive
0111c4fd
RGS
2843 representation using a flat table will take. If it's over a reasonable
2844 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
a3621e74
YO
2845 conservative but potentially much slower representation using an array
2846 of lists.
2847
2848 At the end we convert both representations into the same compressed
2849 form that will be used in regexec.c for matching with. The latter
2850 is a form that cannot be used to construct with but has memory
2851 properties similar to the list form and access properties similar
2852 to the table form making it both suitable for fast searches and
2853 small enough that its feasable to store for the duration of a program.
2854
2855 See the comment in the code where the compressed table is produced
2856 inplace from the flat tabe representation for an explanation of how
2857 the compression works.
2858
2859 */
2860
2861
2e64971a
DM
2862 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2863 prev_states[1] = 0;
2864
538e84ed
KW
2865 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2866 > SvIV(re_trie_maxbuff) )
2867 {
a3621e74
YO
2868 /*
2869 Second Pass -- Array Of Lists Representation
2870
2871 Each state will be represented by a list of charid:state records
2872 (reg_trie_trans_le) the first such element holds the CUR and LEN
2873 points of the allocated array. (See defines above).
2874
2875 We build the initial structure using the lists, and then convert
2876 it into the compressed table form which allows faster lookups
2877 (but cant be modified once converted).
a3621e74
YO
2878 */
2879
a3621e74
YO
2880 STRLEN transcount = 1;
2881
6ad9a8ab 2882 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
cb41e5d6 2883 depth+1));
686b73d4 2884
c944940b
JH
2885 trie->states = (reg_trie_state *)
2886 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2887 sizeof(reg_trie_state) );
a3621e74
YO
2888 TRIE_LIST_NEW(1);
2889 next_alloc = 2;
2890
2891 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2892
df826430 2893 regnode *noper = NEXTOPER( cur );
c445ea15
AL
2894 U32 state = 1; /* required init */
2895 U16 charid = 0; /* sanity init */
07be1b83 2896 U32 wordlen = 0; /* required init */
c445ea15 2897
df826430
YO
2898 if (OP(noper) == NOTHING) {
2899 regnode *noper_next= regnext(noper);
944e05e3
YO
2900 if (noper_next < tail)
2901 noper= noper_next;
df826430
YO
2902 }
2903
944e05e3
YO
2904 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
2905 const U8 *uc= (U8*)STRING(noper);
2906 const U8 *e= uc + STR_LEN(noper);
2907
786e8c11 2908 for ( ; uc < e ; uc += len ) {
c445ea15 2909
786e8c11 2910 TRIE_READ_CHAR;
c445ea15 2911
786e8c11
YO
2912 if ( uvc < 256 ) {
2913 charid = trie->charmap[ uvc ];
c445ea15 2914 } else {
538e84ed
KW
2915 SV** const svpp = hv_fetch( widecharmap,
2916 (char*)&uvc,
2917 sizeof( UV ),
2918 0);
786e8c11
YO
2919 if ( !svpp ) {
2920 charid = 0;
2921 } else {
2922 charid=(U16)SvIV( *svpp );
2923 }
c445ea15 2924 }
538e84ed
KW
2925 /* charid is now 0 if we dont know the char read, or
2926 * nonzero if we do */
786e8c11 2927 if ( charid ) {
a3621e74 2928
786e8c11
YO
2929 U16 check;
2930 U32 newstate = 0;
a3621e74 2931
786e8c11
YO
2932 charid--;
2933 if ( !trie->states[ state ].trans.list ) {
2934 TRIE_LIST_NEW( state );
c445ea15 2935 }
538e84ed
KW
2936 for ( check = 1;
2937 check <= TRIE_LIST_USED( state );
2938 check++ )
2939 {
2940 if ( TRIE_LIST_ITEM( state, check ).forid
2941 == charid )
2942 {
786e8c11
YO
2943 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2944 break;
2945 }
2946 }
2947 if ( ! newstate ) {
2948 newstate = next_alloc++;
2e64971a 2949 prev_states[newstate] = state;
786e8c11
YO
2950 TRIE_LIST_PUSH( state, charid, newstate );
2951 transcount++;
2952 }
2953 state = newstate;
2954 } else {
147e3846 2955 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
c445ea15 2956 }
a28509cc 2957 }
c445ea15 2958 }
3dab1dad 2959 TRIE_HANDLE_WORD(state);
a3621e74
YO
2960
2961 } /* end second pass */
2962
1e2e3d02 2963 /* next alloc is the NEXT state to be allocated */
538e84ed 2964 trie->statecount = next_alloc;
c944940b
JH
2965 trie->states = (reg_trie_state *)
2966 PerlMemShared_realloc( trie->states,
2967 next_alloc
2968 * sizeof(reg_trie_state) );
a3621e74 2969
3dab1dad 2970 /* and now dump it out before we compress it */
2b8b4781
NC
2971 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2972 revcharmap, next_alloc,
2973 depth+1)
1e2e3d02 2974 );
a3621e74 2975
c944940b
JH
2976 trie->trans = (reg_trie_trans *)
2977 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
a3621e74
YO
2978 {
2979 U32 state;
a3621e74
YO
2980 U32 tp = 0;
2981 U32 zp = 0;
2982
2983
2984 for( state=1 ; state < next_alloc ; state ++ ) {
2985 U32 base=0;
2986
2987 /*
2988 DEBUG_TRIE_COMPILE_MORE_r(
6ad9a8ab 2989 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
a3621e74
YO
2990 );
2991 */
2992
2993 if (trie->states[state].trans.list) {
2994 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2995 U16 maxid=minid;
a28509cc 2996 U16 idx;
a3621e74
YO
2997
2998 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
c445ea15
AL
2999 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3000 if ( forid < minid ) {
3001 minid=forid;
3002 } else if ( forid > maxid ) {
3003 maxid=forid;
3004 }
a3621e74
YO
3005 }
3006 if ( transcount < tp + maxid - minid + 1) {
3007 transcount *= 2;
c944940b
JH
3008 trie->trans = (reg_trie_trans *)
3009 PerlMemShared_realloc( trie->trans,
446bd890
NC
3010 transcount
3011 * sizeof(reg_trie_trans) );
538e84ed
KW
3012 Zero( trie->trans + (transcount / 2),
3013 transcount / 2,
3014 reg_trie_trans );
a3621e74
YO
3015 }
3016 base = trie->uniquecharcount + tp - minid;
3017 if ( maxid == minid ) {
3018 U32 set = 0;
3019 for ( ; zp < tp ; zp++ ) {
3020 if ( ! trie->trans[ zp ].next ) {
3021 base = trie->uniquecharcount + zp - minid;
538e84ed
KW
3022 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3023 1).newstate;
a3621e74
YO
3024 trie->trans[ zp ].check = state;
3025 set = 1;
3026 break;
3027 }
3028 }
3029 if ( !set ) {
538e84ed
KW
3030 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3031 1).newstate;
a3621e74
YO
3032 trie->trans[ tp ].check = state;
3033 tp++;
3034 zp = tp;
3035 }
3036 } else {
3037 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
538e84ed
KW
3038 const U32 tid = base
3039 - trie->uniquecharcount
3040 + TRIE_LIST_ITEM( state, idx ).forid;
3041 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3042 idx ).newstate;
a3621e74
YO
3043 trie->trans[ tid ].check = state;
3044 }
3045 tp += ( maxid - minid + 1 );
3046 }
3047 Safefree(trie->states[ state ].trans.list);
3048 }
3049 /*
3050 DEBUG_TRIE_COMPILE_MORE_r(
6ad9a8ab 3051 Perl_re_printf( aTHX_ " base: %d\n",base);
a3621e74
YO
3052 );
3053 */
3054 trie->states[ state ].trans.base=base;
3055 }
cc601c31 3056 trie->lasttrans = tp + 1;
a3621e74
YO
3057 }
3058 } else {
3059 /*
3060 Second Pass -- Flat Table Representation.
3061
b423522f
KW
3062 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3063 each. We know that we will need Charcount+1 trans at most to store
3064 the data (one row per char at worst case) So we preallocate both
3065 structures assuming worst case.
a3621e74
YO
3066
3067 We then construct the trie using only the .next slots of the entry
3068 structs.
3069
b423522f
KW
3070 We use the .check field of the first entry of the node temporarily
3071 to make compression both faster and easier by keeping track of how
3072 many non zero fields are in the node.
a3621e74
YO
3073
3074 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3075 transition.
3076
b423522f
KW
3077 There are two terms at use here: state as a TRIE_NODEIDX() which is
3078 a number representing the first entry of the node, and state as a
3079 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3080 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3081 if there are 2 entrys per node. eg:
a3621e74
YO
3082
3083 A B A B
3084 1. 2 4 1. 3 7
3085 2. 0 3 3. 0 5
3086 3. 0 0 5. 0 0
3087 4. 0 0 7. 0 0
3088
b423522f
KW
3089 The table is internally in the right hand, idx form. However as we
3090 also have to deal with the states array which is indexed by nodenum
3091 we have to use TRIE_NODENUM() to convert.
a3621e74
YO
3092
3093 */
6ad9a8ab 3094 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
cb41e5d6 3095 depth+1));
3dab1dad 3096
c944940b
JH
3097 trie->trans = (reg_trie_trans *)
3098 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3099 * trie->uniquecharcount + 1,
3100 sizeof(reg_trie_trans) );
3101 trie->states = (reg_trie_state *)
3102 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3103 sizeof(reg_trie_state) );
a3621e74
YO
3104 next_alloc = trie->uniquecharcount + 1;
3105
3dab1dad 3106
a3621e74
YO
3107 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3108
df826430 3109 regnode *noper = NEXTOPER( cur );
a3621e74
YO
3110
3111 U32 state = 1; /* required init */
3112
3113 U16 charid = 0; /* sanity init */
3114 U32 accept_state = 0; /* sanity init */
a3621e74 3115
07be1b83 3116 U32 wordlen = 0; /* required init */
a3621e74 3117
df826430
YO
3118 if (OP(noper) == NOTHING) {
3119 regnode *noper_next= regnext(noper);
944e05e3
YO
3120 if (noper_next < tail)
3121 noper= noper_next;
df826430 3122 }
fab2782b 3123
944e05e3
YO
3124 if ( noper < tail && ( OP(noper) == flags || ( flags == EXACTFU && OP(noper) == EXACTFU_SS ) ) ) {
3125 const U8 *uc= (U8*)STRING(noper);
3126 const U8 *e= uc + STR_LEN(noper);
3127
786e8c11 3128 for ( ; uc < e ; uc += len ) {
a3621e74 3129
786e8c11 3130 TRIE_READ_CHAR;
a3621e74 3131
786e8c11
YO
3132 if ( uvc < 256 ) {
3133 charid = trie->charmap[ uvc ];
3134 } else {
538e84ed
KW
3135 SV* const * const svpp = hv_fetch( widecharmap,
3136 (char*)&uvc,
3137 sizeof( UV ),
3138 0);
786e8c11 3139 charid = svpp ? (U16)SvIV(*svpp) : 0;
a3621e74 3140 }
786e8c11
YO
3141 if ( charid ) {
3142 charid--;
3143 if ( !trie->trans[ state + charid ].next ) {
3144 trie->trans[ state + charid ].next = next_alloc;
3145 trie->trans[ state ].check++;
2e64971a
DM
3146 prev_states[TRIE_NODENUM(next_alloc)]
3147 = TRIE_NODENUM(state);
786e8c11
YO
3148 next_alloc += trie->uniquecharcount;
3149 }
3150 state = trie->trans[ state + charid ].next;
3151 } else {
147e3846 3152 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
786e8c11 3153 }
538e84ed
KW
3154 /* charid is now 0 if we dont know the char read, or
3155 * nonzero if we do */
a3621e74 3156 }
a3621e74 3157 }
3dab1dad
YO
3158 accept_state = TRIE_NODENUM( state );
3159 TRIE_HANDLE_WORD(accept_state);
a3621e74
YO
3160
3161 } /* end second pass */
3162
3dab1dad 3163 /* and now dump it out before we compress it */
2b8b4781
NC
3164 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3165 revcharmap,
3166 next_alloc, depth+1));
a3621e74 3167
a3621e74
YO
3168 {
3169 /*
3170 * Inplace compress the table.*
3171
3172 For sparse data sets the table constructed by the trie algorithm will
3173 be mostly 0/FAIL transitions or to put it another way mostly empty.
3174 (Note that leaf nodes will not contain any transitions.)
3175
3176 This algorithm compresses the tables by eliminating most such
3177 transitions, at the cost of a modest bit of extra work during lookup:
3178
3179 - Each states[] entry contains a .base field which indicates the
3180 index in the state[] array wheres its transition data is stored.
3181
3b753521 3182 - If .base is 0 there are no valid transitions from that node.
a3621e74
YO
3183
3184 - If .base is nonzero then charid is added to it to find an entry in
3185 the trans array.
3186
3187 -If trans[states[state].base+charid].check!=state then the
3188 transition is taken to be a 0/Fail transition. Thus if there are fail
3189 transitions at the front of the node then the .base offset will point
3190 somewhere inside the previous nodes data (or maybe even into a node
3191 even earlier), but the .check field determines if the transition is
3192 valid.
3193
786e8c11 3194 XXX - wrong maybe?
a3621e74 3195 The following process inplace converts the table to the compressed
3b753521 3196 table: We first do not compress the root node 1,and mark all its
a3621e74 3197 .check pointers as 1 and set its .base pointer as 1 as well. This
3b753521
FN
3198 allows us to do a DFA construction from the compressed table later,
3199 and ensures that any .base pointers we calculate later are greater
3200 than 0.
a3621e74
YO
3201
3202 - We set 'pos' to indicate the first entry of the second node.
3203
3204 - We then iterate over the columns of the node, finding the first and
3205 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3206 and set the .check pointers accordingly, and advance pos
3207 appropriately and repreat for the next node. Note that when we copy
3208 the next pointers we have to convert them from the original
3209 NODEIDX form to NODENUM form as the former is not valid post
3210 compression.
3211
3212 - If a node has no transitions used we mark its base as 0 and do not
3213 advance the pos pointer.
3214
3215 - If a node only has one transition we use a second pointer into the
3216 structure to fill in allocated fail transitions from other states.
3217 This pointer is independent of the main pointer and scans forward
3218 looking for null transitions that are allocated to a state. When it
3219 finds one it writes the single transition into the "hole". If the
786e8c11 3220 pointer doesnt find one the single transition is appended as normal.
a3621e74
YO
3221
3222 - Once compressed we can Renew/realloc the structures to release the
3223 excess space.
3224
3225 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3226 specifically Fig 3.47 and the associated pseudocode.
3227
3228 demq
3229 */
a3b680e6 3230 const U32 laststate = TRIE_NODENUM( next_alloc );
a28509cc 3231 U32 state, charid;
a3621e74 3232 U32 pos = 0, zp=0;
1e2e3d02 3233 trie->statecount = laststate;
a3621e74
YO
3234
3235 for ( state = 1 ; state < laststate ; state++ ) {
3236 U8 flag = 0;
a28509cc
AL
3237 const U32 stateidx = TRIE_NODEIDX( state );
3238 const U32 o_used = trie->trans[ stateidx ].check;
3239 U32 used = trie->trans[ stateidx ].check;
a3621e74
YO
3240 trie->trans[ stateidx ].check = 0;
3241
538e84ed
KW
3242 for ( charid = 0;
3243 used && charid < trie->uniquecharcount;
3244 charid++ )
3245 {
a3621e74
YO
3246 if ( flag || trie->trans[ stateidx + charid ].next ) {
3247 if ( trie->trans[ stateidx + charid ].next ) {
3248 if (o_used == 1) {
3249 for ( ; zp < pos ; zp++ ) {
3250 if ( ! trie->trans[ zp ].next ) {
3251 break;
3252 }
3253 }
538e84ed
KW
3254 trie->states[ state ].trans.base
3255 = zp
3256 + trie->uniquecharcount
3257 - charid ;
3258 trie->trans[ zp ].next
3259 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3260 + charid ].next );
a3621e74
YO
3261 trie->trans[ zp ].check = state;
3262 if ( ++zp > pos ) pos = zp;
3263 break;
3264 }
3265 used--;
3266 }
3267 if ( !flag ) {
3268 flag = 1;
538e84ed
KW
3269 trie->states[ state ].trans.base
3270 = pos + trie->uniquecharcount - charid ;
a3621e74 3271 }
538e84ed
KW
3272 trie->trans[ pos ].next
3273 = SAFE_TRIE_NODENUM(
3274 trie->trans[ stateidx + charid ].next );
a3621e74
YO
3275 trie->trans[ pos ].check = state;
3276 pos++;
3277 }
3278 }
3279 }
cc601c31 3280 trie->lasttrans = pos + 1;
c944940b
JH
3281 trie->states = (reg_trie_state *)
3282 PerlMemShared_realloc( trie->states, laststate
3283 * sizeof(reg_trie_state) );
a3621e74 3284 DEBUG_TRIE_COMPILE_MORE_r(
147e3846 3285 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
cb41e5d6 3286 depth+1,
538e84ed
KW
3287 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3288 + 1 ),
3289 (IV)next_alloc,
3290 (IV)pos,
3291 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
a3621e74
YO
3292 );
3293
3294 } /* end table compress */
3295 }
1e2e3d02 3296 DEBUG_TRIE_COMPILE_MORE_r(
147e3846 3297 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
cb41e5d6 3298 depth+1,
1e2e3d02
YO
3299 (UV)trie->statecount,
3300 (UV)trie->lasttrans)
3301 );
cc601c31 3302 /* resize the trans array to remove unused space */
c944940b
JH
3303 trie->trans = (reg_trie_trans *)
3304 PerlMemShared_realloc( trie->trans, trie->lasttrans
3305 * sizeof(reg_trie_trans) );
a3621e74 3306
538e84ed 3307 { /* Modify the program and insert the new TRIE node */
3dab1dad
YO
3308 U8 nodetype =(U8)(flags & 0xFF);
3309 char *str=NULL;
538e84ed 3310
07be1b83 3311#ifdef DEBUGGING
e62cc96a 3312 regnode *optimize = NULL;
7122b237
YO
3313#ifdef RE_TRACK_PATTERN_OFFSETS
3314
b57a0404
JH
3315 U32 mjd_offset = 0;
3316 U32 mjd_nodelen = 0;
7122b237
YO
3317#endif /* RE_TRACK_PATTERN_OFFSETS */
3318#endif /* DEBUGGING */
a3621e74 3319 /*
3dab1dad
YO
3320 This means we convert either the first branch or the first Exact,
3321 depending on whether the thing following (in 'last') is a branch
3322 or not and whther first is the startbranch (ie is it a sub part of
3323 the alternation or is it the whole thing.)
3b753521 3324 Assuming its a sub part we convert the EXACT otherwise we convert
3dab1dad 3325 the whole branch sequence, including the first.
a3621e74 3326 */
3dab1dad 3327 /* Find the node we are going to overwrite */
7f69552c 3328 if ( first != startbranch || OP( last ) == BRANCH ) {
07be1b83 3329 /* branch sub-chain */
3dab1dad 3330 NEXT_OFF( first ) = (U16)(last - first);
7122b237 3331#ifdef RE_TRACK_PATTERN_OFFSETS
07be1b83
YO
3332 DEBUG_r({
3333 mjd_offset= Node_Offset((convert));
3334 mjd_nodelen= Node_Length((convert));
3335 });
7122b237 3336#endif
7f69552c 3337 /* whole branch chain */
7122b237
YO
3338 }
3339#ifdef RE_TRACK_PATTERN_OFFSETS
3340 else {
7f69552c
YO
3341 DEBUG_r({
3342 const regnode *nop = NEXTOPER( convert );
3343 mjd_offset= Node_Offset((nop));
3344 mjd_nodelen= Node_Length((nop));
3345 });
07be1b83
YO
3346 }
3347 DEBUG_OPTIMISE_r(
147e3846 3348 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
cb41e5d6 3349 depth+1,
786e8c11 3350 (UV)mjd_offset, (UV)mjd_nodelen)
07be1b83 3351 );
7122b237 3352#endif
538e84ed 3353 /* But first we check to see if there is a common prefix we can
3dab1dad
YO
3354 split out as an EXACT and put in front of the TRIE node. */
3355 trie->startstate= 1;
55eed653 3356 if ( trie->bitmap && !widecharmap && !trie->jump ) {
5ee57374
YO
3357 /* we want to find the first state that has more than
3358 * one transition, if that state is not the first state
3359 * then we have a common prefix which we can remove.
3360 */
3dab1dad 3361 U32 state;
1e2e3d02 3362 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
a3621e74 3363 U32 ofs = 0;
ee0dfd0b
YO
3364 I32 first_ofs = -1; /* keeps track of the ofs of the first
3365 transition, -1 means none */
8e11feef
RGS
3366 U32 count = 0;
3367 const U32 base = trie->states[ state ].trans.base;
a3621e74 3368
5ee57374 3369 /* does this state terminate an alternation? */
3dab1dad 3370 if ( trie->states[state].wordnum )
8e11feef 3371 count = 1;
a3621e74 3372
8e11feef 3373 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
cc601c31
YO
3374 if ( ( base + ofs >= trie->uniquecharcount ) &&
3375 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
a3621e74
YO
3376 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3377 {
3dab1dad 3378 if ( ++count > 1 ) {
5ee57374 3379 /* we have more than one transition */
d3d91c7c
YO
3380 SV **tmp;
3381 U8 *ch;
5ee57374
YO
3382 /* if this is the first state there is no common prefix
3383 * to extract, so we can exit */
8e11feef 3384 if ( state == 1 ) break;
d3d91c7c
YO
3385 tmp = av_fetch( revcharmap, ofs, 0);
3386 ch = (U8*)SvPV_nolen_const( *tmp );
3387
5ee57374
YO
3388 /* if we are on count 2 then we need to initialize the
3389 * bitmap, and store the previous char if there was one
3390 * in it*/
3dab1dad 3391 if ( count == 2 ) {
5ee57374 3392 /* clear the bitmap */
3dab1dad
YO
3393 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3394 DEBUG_OPTIMISE_r(
147e3846 3395 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
cb41e5d6 3396 depth+1,
786e8c11 3397 (UV)state));
ee0dfd0b
YO
3398 if (first_ofs >= 0) {
3399 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
be8e71aa 3400 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
8e11feef 3401
8bcafbf4 3402 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
3dab1dad 3403 DEBUG_OPTIMISE_r(
6ad9a8ab 3404 Perl_re_printf( aTHX_ "%s", (char*)ch)
3dab1dad 3405 );
8e11feef
RGS
3406 }
3407 }
8bcafbf4
YO
3408 /* store the current firstchar in the bitmap */
3409 TRIE_BITMAP_SET_FOLDED(trie,*ch,folder);
6ad9a8ab 3410 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
8e11feef 3411 }
ee0dfd0b 3412 first_ofs = ofs;
8e11feef 3413 }
3dab1dad
YO
3414 }
3415 if ( count == 1 ) {
5ee57374
YO
3416 /* This state has only one transition, its transition is part
3417 * of a common prefix - we need to concatenate the char it
3418 * represents to what we have so far. */
ee0dfd0b 3419 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
c490c714
YO
3420 STRLEN len;
3421 char *ch = SvPV( *tmp, len );
de734bd5
A
3422 DEBUG_OPTIMISE_r({
3423 SV *sv=sv_newmortal();
147e3846 3424 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
cb41e5d6 3425 depth+1,
ee0dfd0b 3426 (UV)state, (UV)first_ofs,
538e84ed 3427 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
de734bd5
A
3428 PL_colors[0], PL_colors[1],
3429 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
538e84ed 3430 PERL_PV_ESCAPE_FIRSTCHAR
de734bd5
A
3431 )
3432 );
3433 });
3dab1dad
YO
3434 if ( state==1 ) {
3435 OP( convert ) = nodetype;
3436 str=STRING(convert);
3437 STR_LEN(convert)=0;
3438 }
c490c714
YO
3439 STR_LEN(convert) += len;
3440 while (len--)
de734bd5 3441 *str++ = *ch++;
8e11feef 3442 } else {
538e84ed 3443#ifdef DEBUGGING
8e11feef 3444 if (state>1)
6ad9a8ab 3445 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
f9049ba1 3446#endif
8e11feef
RGS
3447 break;
3448 }
3449 }
2e64971a 3450 trie->prefixlen = (state-1);
3dab1dad 3451 if (str) {
8e11feef 3452 regnode *n = convert+NODE_SZ_STR(convert);
07be1b83 3453 NEXT_OFF(convert) = NODE_SZ_STR(convert);
8e11feef 3454 trie->startstate = state;
07be1b83
YO
3455 trie->minlen -= (state - 1);
3456 trie->maxlen -= (state - 1);
33809eae
JH
3457#ifdef DEBUGGING
3458 /* At least the UNICOS C compiler choked on this
3459 * being argument to DEBUG_r(), so let's just have
3460 * it right here. */
3461 if (
3462#ifdef PERL_EXT_RE_BUILD
3463 1
3464#else
3465 DEBUG_r_TEST
3466#endif
3467 ) {
3468 regnode *fix = convert;
3469 U32 word = trie->wordcount;
9984212a 3470#ifdef RE_TRACK_PATTERN_OFFSETS
33809eae 3471 mjd_nodelen++;