• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1Technical Notes about PCRE2
2---------------------------
3
4These are very rough technical notes that record potentially useful information
5about PCRE2 internals. PCRE2 is a library based on the original PCRE library,
6but with a revised (and incompatible) API. To avoid confusion, the original
7library is referred to as PCRE1 below. For information about testing PCRE2, see
8the pcre2test documentation and the comment at the head of the RunTest file.
9
10PCRE1 releases were up to 8.3x when PCRE2 was developed, and later bug fix
11releases remain in the 8.xx series. PCRE2 releases started at 10.00 to avoid
12confusion with PCRE1.
13
14
15Historical note 1
16-----------------
17
18Many years ago I implemented some regular expression functions to an algorithm
19suggested by Martin Richards. The rather simple patterns were not Unix-like in
20form, and were quite restricted in what they could do by comparison with Perl.
21The interesting part about the algorithm was that the amount of space required
22to hold the compiled form of an expression was known in advance. The code to
23apply an expression did not operate by backtracking, as the original Henry
24Spencer code and current PCRE2 and Perl code does, but instead checked all
25possibilities simultaneously by keeping a list of current states and checking
26all of them as it advanced through the subject string. In the terminology of
27Jeffrey Friedl's book, it was a "DFA algorithm", though it was not a
28traditional Finite State Machine (FSM). When the pattern was all used up, all
29remaining states were possible matches, and the one matching the longest subset
30of the subject string was chosen. This did not necessarily maximize the
31individual wild portions of the pattern, as is expected in Unix and Perl-style
32regular expressions.
33
34
35Historical note 2
36-----------------
37
38By contrast, the code originally written by Henry Spencer (which was
39subsequently heavily modified for Perl) compiles the expression twice: once in
40a dummy mode in order to find out how much store will be needed, and then for
41real. (The Perl version probably doesn't do this any more; I'm talking about
42the original library.) The execution function operates by backtracking and
43maximizing (or, optionally, minimizing, in Perl) the amount of the subject that
44matches individual wild portions of the pattern. This is an "NFA algorithm" in
45Friedl's terminology.
46
47
48OK, here's the real stuff
49-------------------------
50
51For the set of functions that formed the original PCRE1 library in 1997 (which
52are unrelated to those mentioned above), I tried at first to invent an
53algorithm that used an amount of store bounded by a multiple of the number of
54characters in the pattern, to save on compiling time. However, because of the
55greater complexity in Perl regular expressions, I couldn't do this, even though
56the then current Perl 5.004 patterns were much simpler than those supported
57nowadays. In any case, a first pass through the pattern is helpful for other
58reasons.
59
60
61Support for 16-bit and 32-bit data strings
62-------------------------------------------
63
64The PCRE2 library can be compiled in any combination of 8-bit, 16-bit or 32-bit
65modes, creating up to three different libraries. In the description that
66follows, the word "short" is used for a 16-bit data quantity, and the phrase
67"code unit" is used for a quantity that is a byte in 8-bit mode, a short in
6816-bit mode and a 32-bit word in 32-bit mode. The names of PCRE2 functions are
69given in generic form, without the _8, _16, or _32 suffix.
70
71
72Computing the memory requirement: how it was
73--------------------------------------------
74
75Up to and including release 6.7, PCRE1 worked by running a very degenerate
76first pass to calculate a maximum memory requirement, and then a second pass to
77do the real compile - which might use a bit less than the predicted amount of
78memory. The idea was that this would turn out faster than the Henry Spencer
79code because the first pass is degenerate and the second pass can just store
80stuff straight into memory, which it knows is big enough.
81
82
83Computing the memory requirement: how it is
84-------------------------------------------
85
86By the time I was working on a potential 6.8 release, the degenerate first pass
87had become very complicated and hard to maintain. Indeed one of the early
88things I did for 6.8 was to fix Yet Another Bug in the memory computation. Then
89I had a flash of inspiration as to how I could run the real compile function in
90a "fake" mode that enables it to compute how much memory it would need, while
91in most cases only ever using a small amount of working memory, and without too
92many tests of the mode that might slow it down. So I refactored the compiling
93functions to work this way. This got rid of about 600 lines of source and made
94further maintenance and development easier. As this was such a major change, I
95never released 6.8, instead upping the number to 7.0 (other quite major changes
96were also present in the 7.0 release).
97
98A side effect of this work was that the previous limit of 200 on the nesting
99depth of parentheses was removed. However, there was a downside: compiling ran
100more slowly than before (30% or more, depending on the pattern) because it now
101did a full analysis of the pattern. My hope was that this would not be a big
102issue, and in the event, nobody has commented on it.
103
104At release 8.34, a limit on the nesting depth of parentheses was re-introduced
105(default 250, settable at build time) so as to put a limit on the amount of
106system stack used by the compile function, which uses recursive function calls
107for nested parenthesized groups. This is a safety feature for environments with
108small stacks where the patterns are provided by users.
109
110
111Yet another pattern scan
112------------------------
113
114History repeated itself for PCRE2 release 10.20. A number of bugs relating to
115named subpatterns had been discovered by fuzzers. Most of these were related to
116the handling of forward references when it was not known if the named group was
117unique. (References to non-unique names use a different opcode and more
118memory.) The use of duplicate group numbers (the (?| facility) also caused
119issues.
120
121To get around these problems I adopted a new approach by adding a third pass
122over the pattern (really a "pre-pass"), which did nothing other than identify
123all the named subpatterns and their corresponding group numbers. This means
124that the actual compile (both the memory-computing dummy run and the real
125compile) has full knowledge of group names and numbers throughout. Several
126dozen lines of messy code were eliminated, though the new pre-pass was not
127short. In particular, parsing and skipping over [] classes is complicated.
128
129While working on 10.22 I realized that I could simplify yet again by moving
130more of the parsing into the pre-pass, thus avoiding doing it in two places, so
131after 10.22 was released, the code underwent yet another big refactoring. This
132is how it is from 10.23 onwards:
133
134The function called parse_regex() scans the pattern characters, parsing them
135into literal data and meta characters. It converts escapes such as \x{123}
136into literals, handles \Q...\E, and skips over comments and non-significant
137white space. The result of the scanning is put into a vector of 32-bit unsigned
138integers. Values less than 0x80000000 are literal data. Higher values represent
139meta-characters. The top 16-bits of such values identify the meta-character,
140and these are given names such as META_CAPTURE. The lower 16-bits are available
141for data, for example, the capturing group number. The only situation in which
142literal data values greater than 0x7fffffff can appear is when the 32-bit
143library is running in non-UTF mode. This is handled by having a special
144meta-character that is followed by the 32-bit data value.
145
146The size of the parsed pattern vector, when auto-callouts are not enabled, is
147bounded by the length of the pattern (with one exception). The code is written
148so that each item in the pattern uses no more vector elements than the number
149of code units in the item itself. The exception is the aforementioned large
15032-bit number handling. For this reason, 32-bit non-UTF patterns are scanned in
151advance to check for such values. When auto-callouts are enabled, the generous
152assumption is made that there will be a callout for each pattern code unit
153(which of course is only actually true if all code units are literals) plus one
154at the end. There is a default parsed pattern vector on the system stack, but
155if this is not big enough, heap memory is used.
156
157As before, the actual compiling function is run twice, the first time to
158determine the amount of memory needed for the final compiled pattern. It
159now processes the parsed pattern vector, not the pattern itself, although some
160of the parsed items refer to strings in the pattern - for example, group
161names. As escapes and comments have already been processed, the code is a bit
162simpler than before.
163
164Most errors can be diagnosed during the parsing scan. For those that cannot
165(for example, "lookbehind assertion is not fixed length"), the parsed code
166contains offsets into the pattern so that the actual compiling code can
167report where errors are.
168
169
170The elements of the parsed pattern vector
171-----------------------------------------
172
173The word "offset" below means a code unit offset into the pattern. When
174PCRE2_SIZE (which is usually size_t) is no bigger than uint32_t, an offset is
175stored in a single parsed pattern element. Otherwise (typically on 64-bit
176systems) it occupies two elements. The following meta items occupy just one
177element, with no data:
178
179META_ACCEPT           (*ACCEPT)
180META_ASTERISK         *
181META_ASTERISK_PLUS    *+
182META_ASTERISK_QUERY   *?
183META_ATOMIC           (?> start of atomic group
184META_CIRCUMFLEX       ^ metacharacter
185META_CLASS            [ start of non-empty class
186META_CLASS_EMPTY      [] empty class - only with PCRE2_ALLOW_EMPTY_CLASS
187META_CLASS_EMPTY_NOT  [^] negative empty class - ditto
188META_CLASS_END        ] end of non-empty class
189META_CLASS_NOT        [^ start non-empty negative class
190META_COMMIT           (*COMMIT)
191META_COND_ASSERT      (?(?assertion)
192META_DOLLAR           $ metacharacter
193META_DOT              . metacharacter
194META_END              End of pattern (this value is 0x80000000)
195META_FAIL             (*FAIL)
196META_KET              ) closing parenthesis
197META_LOOKAHEAD        (?= start of lookahead
198META_LOOKAHEAD_NA     (*napla: start of non-atomic lookahead
199META_LOOKAHEADNOT     (?! start of negative lookahead
200META_NOCAPTURE        (?: no capture parens
201META_PLUS             +
202META_PLUS_PLUS        ++
203META_PLUS_QUERY       +?
204META_PRUNE            (*PRUNE) - no argument
205META_QUERY            ?
206META_QUERY_PLUS       ?+
207META_QUERY_QUERY      ??
208META_RANGE_ESCAPED    hyphen in class range with at least one escape
209META_RANGE_LITERAL    hyphen in class range defined literally
210META_SKIP             (*SKIP) - no argument
211META_THEN             (*THEN) - no argument
212
213The two RANGE values occur only in character classes. They are positioned
214between two literals that define the start and end of the range. In an EBCDIC
215evironment it is necessary to know whether either of the range values was
216specified as an escape. In an ASCII/Unicode environment the distinction is not
217relevant.
218
219The following have data in the lower 16 bits, and may be followed by other data
220elements:
221
222META_ALT              | alternation
223META_BACKREF          back reference
224META_CAPTURE          start of capturing group
225META_ESCAPE           non-literal escape sequence
226META_RECURSE          recursion call
227
228If the data for META_ALT is non-zero, it is inside a lookbehind, and the data
229is the length of its branch, for which OP_REVERSE must be generated.
230
231META_BACKREF, META_CAPTURE, and META_RECURSE have the capture group number as
232their data in the lower 16 bits of the element.
233
234META_BACKREF is followed by an offset if the back reference group number is 10
235or more. The offsets of the first ocurrences of references to groups whose
236numbers are less than 10 are put in cb->small_ref_offset[] (only the first
237occurrence is useful). On 64-bit systems this avoids using more than two parsed
238pattern elements for items such as \3. The offset is used when an error occurs
239because the reference is to a non-existent group.
240
241META_RECURSE is always followed by an offset, for use in error messages.
242
243META_ESCAPE has an ESC_xxx value as its data. For ESC_P and ESC_p, the next
244element contains the 16-bit type and data property values, packed together.
245ESC_g and ESC_k are used only for named references - numerical ones are turned
246into META_RECURSE or META_BACKREF as appropriate. ESC_g and ESC_k are followed
247by a length and an offset into the pattern to specify the name.
248
249The following have one data item that follows in the next vector element:
250
251META_BIGVALUE         Next is a literal >= META_END
252META_OPTIONS          (?i) and friends (data is new option bits)
253META_POSIX            POSIX class item (data identifies the class)
254META_POSIX_NEG        negative POSIX class item (ditto)
255
256The following are followed by a length element, then a number of character code
257values (which should match with the length):
258
259META_MARK             (*MARK:xxxx)
260META_COMMIT_ARG       )*COMMIT:xxxx)
261META_PRUNE_ARG        (*PRUNE:xxx)
262META_SKIP_ARG         (*SKIP:xxxx)
263META_THEN_ARG         (*THEN:xxxx)
264
265The following are followed by a length element, then an offset in the pattern
266that identifies the name:
267
268META_COND_NAME        (?(<name>) or (?('name') or (?(name)
269META_COND_RNAME       (?(R&name)
270META_COND_RNUMBER     (?(Rdigits)
271META_RECURSE_BYNAME   (?&name)
272META_BACKREF_BYNAME   \k'name'
273
274META_COND_RNUMBER is used for names that start with R and continue with digits,
275because this is an ambiguous case. It could be a back reference to a group with
276that name, or it could be a recursion test on a numbered group.
277
278This one is followed by an offset, for use in error messages, then a number:
279
280META_COND_NUMBER       (?([+-]digits)
281
282The following is followed just by an offset, for use in error messages:
283
284META_COND_DEFINE      (?(DEFINE)
285
286The following are also followed just by an offset, but also the lower 16 bits
287of the main word contain the length of the first branch of the lookbehind
288group; this is used when generating OP_REVERSE for that branch.
289
290META_LOOKBEHIND       (?<=      start of lookbehind
291META_LOOKBEHIND_NA    (*naplb:  start of non-atomic lookbehind
292META_LOOKBEHINDNOT    (?<!      start of negative lookbehind
293
294The following are followed by two elements, the minimum and maximum. Repeat
295values are limited to 65535 (MAX_REPEAT). A maximum value of "unlimited" is
296represented by UNLIMITED_REPEAT, which is bigger than MAX_REPEAT:
297
298META_MINMAX           {n,m}  repeat
299META_MINMAX_PLUS      {n,m}+ repeat
300META_MINMAX_QUERY     {n,m}? repeat
301
302This one is followed by three elements. The first is 0 for '>' and 1 for '>=';
303the next two are the major and minor numbers:
304
305META_COND_VERSION     (?(VERSION<op>x.y)
306
307Callouts are converted into one of two items:
308
309META_CALLOUT_NUMBER   (?C with numerical argument
310META_CALLOUT_STRING   (?C with string argument
311
312In both cases, the next two elements contain the offset and length of the next
313item in the pattern. Then there is either one callout number, or a length and
314an offset for the string argument. The length includes both delimiters.
315
316
317Traditional matching function
318-----------------------------
319
320The "traditional", and original, matching function is called pcre2_match(), and
321it implements an NFA algorithm, similar to the original Henry Spencer algorithm
322and the way that Perl works. This is not surprising, since it is intended to be
323as compatible with Perl as possible. This is the function most users of PCRE2
324will use most of the time. If PCRE2 is compiled with just-in-time (JIT)
325support, and studying a compiled pattern with JIT is successful, the JIT code
326is run instead of the normal pcre2_match() code, but the result is the same.
327
328
329Supplementary matching function
330-------------------------------
331
332There is also a supplementary matching function called pcre2_dfa_match(). This
333implements a DFA matching algorithm that searches simultaneously for all
334possible matches that start at one point in the subject string. (Going back to
335my roots: see Historical Note 1 above.) This function intreprets the same
336compiled pattern data as pcre2_match(); however, not all the facilities are
337available, and those that are do not always work in quite the same way. See the
338user documentation for details.
339
340The algorithm that is used for pcre2_dfa_match() is not a traditional FSM,
341because it may have a number of states active at one time. More work would be
342needed at compile time to produce a traditional FSM where only one state is
343ever active at once. I believe some other regex matchers work this way. JIT
344support is not available for this kind of matching.
345
346
347Changeable options
348------------------
349
350The /i, /m, or /s options (PCRE2_CASELESS, PCRE2_MULTILINE, PCRE2_DOTALL, and
351others) may be changed in the middle of patterns by items such as (?i). Their
352processing is handled entirely at compile time by generating different opcodes
353for the different settings. The runtime functions do not need to keep track of
354an option's state.
355
356PCRE2_DUPNAMES, PCRE2_EXTENDED, PCRE2_EXTENDED_MORE, and PCRE2_NO_AUTO_CAPTURE
357are tracked and processed during the parsing pre-pass. The others are handled
358from META_OPTIONS items during the main compile phase.
359
360
361Format of compiled patterns
362---------------------------
363
364The compiled form of a pattern is a vector of unsigned code units (bytes in
3658-bit mode, shorts in 16-bit mode, 32-bit words in 32-bit mode), containing
366items of variable length. The first code unit in an item contains an opcode,
367and the length of the item is either implicit in the opcode or contained in the
368data that follows it.
369
370In many cases listed below, LINK_SIZE data values are specified for offsets
371within the compiled pattern. LINK_SIZE always specifies a number of bytes. The
372default value for LINK_SIZE is 2, except for the 32-bit library, where it can
373only be 4. The 8-bit library can be compiled to used 3-byte or 4-byte values,
374and the 16-bit library can be compiled to use 4-byte values, though this
375impairs performance. Specifing a LINK_SIZE larger than 2 for these libraries is
376necessary only when patterns whose compiled length is greater than 65535 code
377units are going to be processed. When a LINK_SIZE value uses more than one code
378unit, the most significant unit is first.
379
380In this description, we assume the "normal" compilation options. Data values
381that are counts (e.g. quantifiers) are always two bytes long in 8-bit mode
382(most significant byte first), and one code unit in 16-bit and 32-bit modes.
383
384
385Opcodes with no following data
386------------------------------
387
388These items are all just one unit long:
389
390  OP_END                 end of pattern
391  OP_ANY                 match any one character other than newline
392  OP_ALLANY              match any one character, including newline
393  OP_ANYBYTE             match any single code unit, even in UTF-8/16 mode
394  OP_SOD                 match start of data: \A
395  OP_SOM,                start of match (subject + offset): \G
396  OP_SET_SOM,            set start of match (\K)
397  OP_CIRC                ^ (start of data)
398  OP_CIRCM               ^ multiline mode (start of data or after newline)
399  OP_NOT_WORD_BOUNDARY   \W
400  OP_WORD_BOUNDARY       \w
401  OP_NOT_DIGIT           \D
402  OP_DIGIT               \d
403  OP_NOT_HSPACE          \H
404  OP_HSPACE              \h
405  OP_NOT_WHITESPACE      \S
406  OP_WHITESPACE          \s
407  OP_NOT_VSPACE          \V
408  OP_VSPACE              \v
409  OP_NOT_WORDCHAR        \W
410  OP_WORDCHAR            \w
411  OP_EODN                match end of data or newline at end: \Z
412  OP_EOD                 match end of data: \z
413  OP_DOLL                $ (end of data, or before final newline)
414  OP_DOLLM               $ multiline mode (end of data or before newline)
415  OP_EXTUNI              match an extended Unicode grapheme cluster
416  OP_ANYNL               match any Unicode newline sequence
417
418  OP_ASSERT_ACCEPT       )
419  OP_ACCEPT              ) These are Perl 5.10's "backtracking control
420  OP_COMMIT              ) verbs". If OP_ACCEPT is inside capturing
421  OP_FAIL                ) parentheses, it may be preceded by one or more
422  OP_PRUNE               ) OP_CLOSE, each followed by a number that
423  OP_SKIP                ) indicates which parentheses must be closed.
424  OP_THEN                )
425
426OP_ASSERT_ACCEPT is used when (*ACCEPT) is encountered within an assertion.
427This ends the assertion, not the entire pattern match. The assertion (?!) is
428always optimized to OP_FAIL.
429
430OP_ALLANY is used for '.' when PCRE2_DOTALL is set. It is also used for \C in
431non-UTF modes and in UTF-32 mode (since one code unit still equals one
432character). Another use is for [^] when empty classes are permitted
433(PCRE2_ALLOW_EMPTY_CLASS is set).
434
435
436Backtracking control verbs
437--------------------------
438
439Verbs with no arguments generate opcodes with no following data (as listed
440in the section above).
441
442(*MARK:NAME) generates OP_MARK followed by the mark name, preceded by a
443length in one code unit, and followed by a binary zero. The name length is
444limited by the size of the code unit.
445
446(*ACCEPT:NAME) and (*FAIL:NAME) are compiled as (*MARK:NAME)(*ACCEPT) and
447(*MARK:NAME)(*FAIL) respectively.
448
449For (*COMMIT:NAME), (*PRUNE:NAME), (*SKIP:NAME), and (*THEN:NAME), the opcodes
450OP_COMMIT_ARG, OP_PRUNE_ARG, OP_SKIP_ARG, and OP_THEN_ARG are used, with the
451name following in the same format as for OP_MARK.
452
453
454Matching literal characters
455---------------------------
456
457The OP_CHAR opcode is followed by a single character that is to be matched
458casefully. For caseless matching of characters that have at most two
459case-equivalent code points, OP_CHARI is used. In UTF-8 or UTF-16 modes, the
460character may be more than one code unit long. In UTF-32 mode, characters are
461always exactly one code unit long.
462
463If there is only one character in a character class, OP_CHAR or OP_CHARI is
464used for a positive class, and OP_NOT or OP_NOTI for a negative one (that is,
465for something like [^a]).
466
467Caseless matching (positive or negative) of characters that have more than two
468case-equivalent code points (which is possible only in UTF mode) is handled by
469compiling a Unicode property item (see below), with the pseudo-property
470PT_CLIST. The value of this property is an offset in a vector called
471"ucd_caseless_sets" which identifies the start of a short list of equivalent
472characters, terminated by the value NOTACHAR (0xffffffff).
473
474
475Repeating single characters
476---------------------------
477
478The common repeats (*, +, ?), when applied to a single character, use the
479following opcodes, which come in caseful and caseless versions:
480
481  Caseful         Caseless
482  OP_STAR         OP_STARI
483  OP_MINSTAR      OP_MINSTARI
484  OP_POSSTAR      OP_POSSTARI
485  OP_PLUS         OP_PLUSI
486  OP_MINPLUS      OP_MINPLUSI
487  OP_POSPLUS      OP_POSPLUSI
488  OP_QUERY        OP_QUERYI
489  OP_MINQUERY     OP_MINQUERYI
490  OP_POSQUERY     OP_POSQUERYI
491
492Each opcode is followed by the character that is to be repeated. In ASCII or
493UTF-32 modes, these are two-code-unit items; in UTF-8 or UTF-16 modes, the
494length is variable. Those with "MIN" in their names are the minimizing
495versions. Those with "POS" in their names are possessive versions. Other kinds
496of repeat make use of these opcodes:
497
498  Caseful         Caseless
499  OP_UPTO         OP_UPTOI
500  OP_MINUPTO      OP_MINUPTOI
501  OP_POSUPTO      OP_POSUPTOI
502  OP_EXACT        OP_EXACTI
503
504Each of these is followed by a count and then the repeated character. The count
505is two bytes long in 8-bit mode (most significant byte first), or one code unit
506in 16-bit and 32-bit modes.
507
508OP_UPTO matches from 0 to the given number. A repeat with a non-zero minimum
509and a fixed maximum is coded as an OP_EXACT followed by an OP_UPTO (or
510OP_MINUPTO or OPT_POSUPTO).
511
512Another set of matching repeating opcodes (called OP_NOTSTAR, OP_NOTSTARI,
513etc.) are used for repeated, negated, single-character classes such as [^a]*.
514The normal single-character opcodes (OP_STAR, etc.) are used for repeated
515positive single-character classes.
516
517
518Repeating character types
519-------------------------
520
521Repeats of things like \d are done exactly as for single characters, except
522that instead of a character, the opcode for the type (e.g. OP_DIGIT) is stored
523in the next code unit. The opcodes are:
524
525  OP_TYPESTAR
526  OP_TYPEMINSTAR
527  OP_TYPEPOSSTAR
528  OP_TYPEPLUS
529  OP_TYPEMINPLUS
530  OP_TYPEPOSPLUS
531  OP_TYPEQUERY
532  OP_TYPEMINQUERY
533  OP_TYPEPOSQUERY
534  OP_TYPEUPTO
535  OP_TYPEMINUPTO
536  OP_TYPEPOSUPTO
537  OP_TYPEEXACT
538
539
540Match by Unicode property
541-------------------------
542
543OP_PROP and OP_NOTPROP are used for positive and negative matches of a
544character by testing its Unicode property (the \p and \P escape sequences).
545Each is followed by two code units that encode the desired property as a type
546and a value. The types are a set of #defines of the form PT_xxx, and the values
547are enumerations of the form ucp_xx, defined in the pcre2_ucp.h source file.
548The value is relevant only for PT_GC (General Category), PT_PC (Particular
549Category), PT_SC (Script), and the pseudo-property PT_CLIST, which is used to
550identify a list of case-equivalent characters when there are three or more.
551
552Repeats of these items use the OP_TYPESTAR etc. set of opcodes, followed by
553three code units: OP_PROP or OP_NOTPROP, and then the desired property type and
554value.
555
556
557Character classes
558-----------------
559
560If there is only one character in a class, OP_CHAR or OP_CHARI is used for a
561positive class, and OP_NOT or OP_NOTI for a negative one (that is, for
562something like [^a]), except when caselessly matching a character that has more
563than two case-equivalent code points (which can happen only in UTF mode). In
564this case a Unicode property item is used, as described above in "Matching
565literal characters".
566
567A set of repeating opcodes (called OP_NOTSTAR etc.) are used for repeated,
568negated, single-character classes. The normal single-character opcodes
569(OP_STAR, etc.) are used for repeated positive single-character classes.
570
571When there is more than one character in a class, and all the code points are
572less than 256, OP_CLASS is used for a positive class, and OP_NCLASS for a
573negative one. In either case, the opcode is followed by a 32-byte (16-short,
5748-word) bit map containing a 1 bit for every character that is acceptable. The
575bits are counted from the least significant end of each unit. In caseless mode,
576bits for both cases are set.
577
578The reason for having both OP_CLASS and OP_NCLASS is so that, in UTF-8 and
57916-bit and 32-bit modes, subject characters with values greater than 255 can be
580handled correctly. For OP_CLASS they do not match, whereas for OP_NCLASS they
581do.
582
583For classes containing characters with values greater than 255 or that contain
584\p or \P, OP_XCLASS is used. It optionally uses a bit map if any acceptable
585code points are less than 256, followed by a list of pairs (for a range) and/or
586single characters and/or properties. In caseless mode, all equivalent
587characters are explicitly listed.
588
589OP_XCLASS is followed by a LINK_SIZE value containing the total length of the
590opcode and its data. This is followed by a code unit containing flag bits:
591XCL_NOT indicates that this is a negative class, and XCL_MAP indicates that a
592bit map is present. There follows the bit map, if XCL_MAP is set, and then a
593sequence of items coded as follows:
594
595  XCL_END      marks the end of the list
596  XCL_SINGLE   one character follows
597  XCL_RANGE    two characters follow
598  XCL_PROP     a Unicode property (type, value) follows
599  XCL_NOTPROP  a Unicode property (type, value) follows
600
601If a range starts with a code point less than 256 and ends with one greater
602than 255, it is split into two ranges, with characters less than 256 being
603indicated in the bit map, and the rest with XCL_RANGE.
604
605When XCL_NOT is set, the bit map, if present, contains bits for characters that
606are allowed (exactly as for OP_NCLASS), but the list of items that follow it
607specifies characters and properties that are not allowed.
608
609
610Back references
611---------------
612
613OP_REF (caseful) or OP_REFI (caseless) is followed by a count containing the
614reference number when the reference is to a unique capturing group (either by
615number or by name). When named groups are used, there may be more than one
616group with the same name. In this case, a reference to such a group by name
617generates OP_DNREF or OP_DNREFI. These are followed by two counts: the index
618(not the byte offset) in the group name table of the first entry for the
619required name, followed by the number of groups with the same name. The
620matching code can then search for the first one that is set.
621
622
623Repeating character classes and back references
624-----------------------------------------------
625
626Single-character classes are handled specially (see above). This section
627applies to other classes and also to back references. In both cases, the repeat
628information follows the base item. The matching code looks at the following
629opcode to see if it is one of these:
630
631  OP_CRSTAR
632  OP_CRMINSTAR
633  OP_CRPOSSTAR
634  OP_CRPLUS
635  OP_CRMINPLUS
636  OP_CRPOSPLUS
637  OP_CRQUERY
638  OP_CRMINQUERY
639  OP_CRPOSQUERY
640  OP_CRRANGE
641  OP_CRMINRANGE
642  OP_CRPOSRANGE
643
644All but the last three are single-code-unit items, with no data. The range
645opcodes are followed by the minimum and maximum repeat counts.
646
647
648Brackets and alternation
649------------------------
650
651A pair of non-capturing round brackets is wrapped round each expression at
652compile time, so alternation always happens in the context of brackets.
653
654[Note for North Americans: "bracket" to some English speakers, including
655myself, can be round, square, curly, or pointy. Hence this usage rather than
656"parentheses".]
657
658Non-capturing brackets use the opcode OP_BRA, capturing brackets use OP_CBRA. A
659bracket opcode is followed by a LINK_SIZE value which gives the offset to the
660next alternative OP_ALT or, if there aren't any branches, to the terminating
661opcode. Each OP_ALT is followed by a LINK_SIZE value giving the offset to the
662next one, or to the final opcode. For capturing brackets, the bracket number is
663a count that immediately follows the offset.
664
665There are several opcodes that mark the end of a subpattern group. OP_KET is
666used for subpatterns that do not repeat indefinitely, OP_KETRMIN and
667OP_KETRMAX are used for indefinite repetitions, minimally or maximally
668respectively, and OP_KETRPOS for possessive repetitions (see below for more
669details). All four are followed by a LINK_SIZE value giving (as a positive
670number) the offset back to the matching bracket opcode.
671
672If a subpattern is quantified such that it is permitted to match zero times, it
673is preceded by one of OP_BRAZERO, OP_BRAMINZERO, or OP_SKIPZERO. These are
674single-unit opcodes that tell the matcher that skipping the following
675subpattern entirely is a valid match. In the case of the first two, not
676skipping the pattern is also valid (greedy and non-greedy). The third is used
677when a pattern has the quantifier {0,0}. It cannot be entirely discarded,
678because it may be called as a subroutine from elsewhere in the pattern.
679
680A subpattern with an indefinite maximum repetition is replicated in the
681compiled data its minimum number of times (or once with OP_BRAZERO if the
682minimum is zero), with the final copy terminating with OP_KETRMIN or OP_KETRMAX
683as appropriate.
684
685A subpattern with a bounded maximum repetition is replicated in a nested
686fashion up to the maximum number of times, with OP_BRAZERO or OP_BRAMINZERO
687before each replication after the minimum, so that, for example, (abc){2,5} is
688compiled as (abc)(abc)((abc)((abc)(abc)?)?)?, except that each bracketed group
689has the same number.
690
691When a repeated subpattern has an unbounded upper limit, it is checked to see
692whether it could match an empty string. If this is the case, the opcode in the
693final replication is changed to OP_SBRA or OP_SCBRA. This tells the matcher
694that it needs to check for matching an empty string when it hits OP_KETRMIN or
695OP_KETRMAX, and if so, to break the loop.
696
697
698Possessive brackets
699-------------------
700
701When a repeated group (capturing or non-capturing) is marked as possessive by
702the "+" notation, e.g. (abc)++, different opcodes are used. Their names all
703have POS on the end, e.g. OP_BRAPOS instead of OP_BRA and OP_SCBRAPOS instead
704of OP_SCBRA. The end of such a group is marked by OP_KETRPOS. If the minimum
705repetition is zero, the group is preceded by OP_BRAPOSZERO.
706
707
708Once-only (atomic) groups
709-------------------------
710
711These are just like other subpatterns, but they start with the opcode OP_ONCE.
712The check for matching an empty string in an unbounded repeat is handled
713entirely at runtime, so there is just this one opcode for atomic groups.
714
715
716Assertions
717----------
718
719Forward assertions are also just like other subpatterns, but starting with one
720of the opcodes OP_ASSERT, OP_ASSERT_NA (non-atomic assertion), or
721OP_ASSERT_NOT. Backward assertions use the opcodes OP_ASSERTBACK,
722OP_ASSERTBACK_NA, and OP_ASSERTBACK_NOT, and the first opcode inside the
723assertion is OP_REVERSE, followed by a count of the number of characters to
724move back the pointer in the subject string. In ASCII or UTF-32 mode, the count
725is also the number of code units, but in UTF-8/16 mode each character may
726occupy more than one code unit. A separate count is present in each alternative
727of a lookbehind assertion, allowing each branch to have a different (but fixed)
728length.
729
730
731Conditional subpatterns
732-----------------------
733
734These are like other subpatterns, but they start with the opcode OP_COND, or
735OP_SCOND for one that might match an empty string in an unbounded repeat.
736
737If the condition is a back reference, this is stored at the start of the
738subpattern using the opcode OP_CREF followed by a count containing the
739reference number, provided that the reference is to a unique capturing group.
740If the reference was by name and there is more than one group with that name,
741OP_DNCREF is used instead. It is followed by two counts: the index in the group
742names table, and the number of groups with the same name. The allows the
743matcher to check if any group with the given name is set.
744
745If the condition is "in recursion" (coded as "(?(R)"), or "in recursion of
746group x" (coded as "(?(Rx)"), the group number is stored at the start of the
747subpattern using the opcode OP_RREF (with a value of RREF_ANY (0xffff) for "the
748whole pattern") or OP_DNRREF (with data as for OP_DNCREF).
749
750For a DEFINE condition, OP_FALSE is used (with no associated data). During
751compilation, however, a DEFINE condition is coded as OP_DEFINE so that, when
752the conditional group is complete, there can be a check to ensure that it
753contains only one top-level branch. Once this has happened, the opcode is
754changed to OP_FALSE, so the matcher never sees OP_DEFINE.
755
756There is a special PCRE2-specific condition of the form (VERSION[>]=x.y), which
757tests the PCRE2 version number. This compiles into one of the opcodes OP_TRUE
758or OP_FALSE.
759
760If a condition is not a back reference, recursion test, DEFINE, or VERSION, it
761must start with a parenthesized atomic assertion, whose opcode normally
762immediately follows OP_COND or OP_SCOND. However, if automatic callouts are
763enabled, a callout is inserted immediately before the assertion. It is also
764possible to insert a manual callout at this point. Only assertion conditions
765may have callouts preceding the condition.
766
767A condition that is the negative assertion (?!) is optimized to OP_FAIL in all
768parts of the pattern, so this is another opcode that may appear as a condition.
769It is treated the same as OP_FALSE.
770
771
772Recursion
773---------
774
775Recursion either matches the current pattern, or some subexpression. The opcode
776OP_RECURSE is followed by a LINK_SIZE value that is the offset to the starting
777bracket from the start of the whole pattern. OP_RECURSE is also used for
778"subroutine" calls, even though they are not strictly a recursion. Up till
779release 10.30 recursions were treated as atomic groups, making them
780incompatible with Perl (but PCRE had them well before Perl did). From 10.30,
781backtracking into recursions is supported.
782
783Repeated recursions used to be wrapped inside OP_ONCE brackets, which not only
784forced no backtracking, but also allowed repetition to be handled as for other
785bracketed groups. From 10.30 onwards, repeated recursions are duplicated for
786their minimum repetitions, and then wrapped in non-capturing brackets for the
787remainder. For example, (?1){3} is treated as (?1)(?1)(?1), and (?1){2,4} is
788treated as (?1)(?1)(?:(?1)){0,2}.
789
790
791Callouts
792--------
793
794A callout may have either a numerical argument or a string argument. These use
795OP_CALLOUT or OP_CALLOUT_STR, respectively. In each case these are followed by
796two LINK_SIZE values giving the offset in the pattern string to the start of
797the following item, and another count giving the length of this item. These
798values make it possible for pcre2test to output useful tracing information
799using callouts.
800
801In the case of a numeric callout, after these two values there is a single code
802unit containing the callout number, in the range 0-255, with 255 being used for
803callouts that are automatically inserted as a result of the PCRE2_AUTO_CALLOUT
804option. Thus, this opcode item is of fixed length:
805
806  [OP_CALLOUT] [PATTERN_OFFSET] [PATTERN_LENGTH] [NUMBER]
807
808For callouts with string arguments, OP_CALLOUT_STR has three more data items:
809a LINK_SIZE value giving the complete length of the entire opcode item, a
810LINK_SIZE item containing the offset within the pattern string to the start of
811the string argument, and the string itself, preceded by its starting delimiter
812and followed by a binary zero. When a callout function is called, a pointer to
813the actual string is passed, but the delimiter can be accessed as string[-1] if
814the application needs it. In the 8-bit library, the callout in /X(?C'abc')Y/ is
815compiled as the following bytes (decimal numbers represent binary values):
816
817  [OP_CALLOUT_STR]  [0] [10]  [0] [1]  [0] [14]  [0] [5] ['] [a] [b] [c] [0]
818                    --------  -------  --------  -------
819                       |         |        |         |
820                       ------- LINK_SIZE items ------
821
822Opcode table checking
823---------------------
824
825The last opcode that is defined in pcre2_internal.h is OP_TABLE_LENGTH. This is
826not a real opcode, but is used to check at compile time that tables indexed by
827opcode are the correct length, in order to catch updating errors.
828
829Philip Hazel
83012 July 2019
831