1 Technical Notes about PCRE
2 --------------------------
4 These are very rough technical notes that record potentially useful information
11 Many years ago I implemented some regular expression functions to an algorithm
12 suggested by Martin Richards. These were not Unix-like in form, and were quite
13 restricted in what they could do by comparison with Perl. The interesting part
14 about the algorithm was that the amount of space required to hold the compiled
15 form of an expression was known in advance. The code to apply an expression did
16 not operate by backtracking, as the original Henry Spencer code and current
17 Perl code does, but instead checked all possibilities simultaneously by keeping
18 a list of current states and checking all of them as it advanced through the
19 subject string. In the terminology of Jeffrey Friedl's book, it was a "DFA
20 algorithm", though it was not a traditional Finite State Machine (FSM). When
21 the pattern was all used up, all remaining states were possible matches, and
22 the one matching the longest subset of the subject string was chosen. This did
23 not necessarily maximize the individual wild portions of the pattern, as is
24 expected in Unix and Perl-style regular expressions.
30 By contrast, the code originally written by Henry Spencer (which was
31 subsequently heavily modified for Perl) compiles the expression twice: once in
32 a dummy mode in order to find out how much store will be needed, and then for
33 real. (The Perl version probably doesn't do this any more; I'm talking about
34 the original library.) The execution function operates by backtracking and
35 maximizing (or, optionally, minimizing in Perl) the amount of the subject that
36 matches individual wild portions of the pattern. This is an "NFA algorithm" in
40 OK, here's the real stuff
41 -------------------------
43 For the set of functions that form the "basic" PCRE library (which are
44 unrelated to those mentioned above), I tried at first to invent an algorithm
45 that used an amount of store bounded by a multiple of the number of characters
46 in the pattern, to save on compiling time. However, because of the greater
47 complexity in Perl regular expressions, I couldn't do this. In any case, a
48 first pass through the pattern is helpful for other reasons.
51 Computing the memory requirement: how it was
52 --------------------------------------------
54 Up to and including release 6.7, PCRE worked by running a very degenerate first
55 pass to calculate a maximum store size, and then a second pass to do the real
56 compile - which might use a bit less than the predicted amount of memory. The
57 idea was that this would turn out faster than the Henry Spencer code because
58 the first pass is degenerate and the second pass can just store stuff straight
59 into the vector, which it knows is big enough.
62 Computing the memory requirement: how it is
63 -------------------------------------------
65 By the time I was working on a potential 6.8 release, the degenerate first pass
66 had become very complicated and hard to maintain. Indeed one of the early
67 things I did for 6.8 was to fix Yet Another Bug in the memory computation. Then
68 I had a flash of inspiration as to how I could run the real compile function in
69 a "fake" mode that enables it to compute how much memory it would need, while
70 actually only ever using a few hundred bytes of working memory, and without too
71 many tests of the mode that might slow it down. So I re-factored the compiling
72 functions to work this way. This got rid of about 600 lines of source. It
73 should make future maintenance and development easier. As this was such a major
74 change, I never released 6.8, instead upping the number to 7.0 (other quite
75 major changes were also present in the 7.0 release).
77 A side effect of this work was that the previous limit of 200 on the nesting
78 depth of parentheses was removed. However, there is a downside: pcre_compile()
79 runs more slowly than before (30% or more, depending on the pattern) because it
80 is doing a full analysis of the pattern. My hope was that this would not be a
81 big issue, and in the event, nobody has commented on it.
84 Traditional matching function
85 -----------------------------
87 The "traditional", and original, matching function is called pcre_exec(), and
88 it implements an NFA algorithm, similar to the original Henry Spencer algorithm
89 and the way that Perl works. This is not surprising, since it is intended to be
90 as compatible with Perl as possible. This is the function most users of PCRE
91 will use most of the time.
94 Supplementary matching function
95 -------------------------------
97 From PCRE 6.0, there is also a supplementary matching function called
98 pcre_dfa_exec(). This implements a DFA matching algorithm that searches
99 simultaneously for all possible matches that start at one point in the subject
100 string. (Going back to my roots: see Historical Note 1 above.) This function
101 intreprets the same compiled pattern data as pcre_exec(); however, not all the
102 facilities are available, and those that are do not always work in quite the
103 same way. See the user documentation for details.
105 The algorithm that is used for pcre_dfa_exec() is not a traditional FSM,
106 because it may have a number of states active at one time. More work would be
107 needed at compile time to produce a traditional FSM where only one state is
108 ever active at once. I believe some other regex matchers work this way.
111 Format of compiled patterns
112 ---------------------------
114 The compiled form of a pattern is a vector of bytes, containing items of
115 variable length. The first byte in an item is an opcode, and the length of the
116 item is either implicit in the opcode or contained in the data bytes that
119 In many cases below LINK_SIZE data values are specified for offsets within the
120 compiled pattern. The default value for LINK_SIZE is 2, but PCRE can be
121 compiled to use 3-byte or 4-byte values for these offsets (impairing the
122 performance). This is necessary only when patterns whose compiled length is
123 greater than 64K are going to be processed. In this description, we assume the
124 "normal" compilation options. Data values that are counts (e.g. for
125 quantifiers) are always just two bytes long.
127 A list of the opcodes follows:
129 Opcodes with no following data
130 ------------------------------
132 These items are all just one byte long
134 OP_END end of pattern
135 OP_ANY match any one character other than newline
136 OP_ALLANY match any one character, including newline
137 OP_ANYBYTE match any single byte, even in UTF-8 mode
138 OP_SOD match start of data: \A
139 OP_SOM, start of match (subject + offset): \G
140 OP_SET_SOM, set start of match (\K)
141 OP_CIRC ^ (start of data, or after \n in multiline)
142 OP_NOT_WORD_BOUNDARY \W
154 OP_EODN match end of data or \n at end: \Z
155 OP_EOD match end of data: \z
156 OP_DOLL $ (end of data, or before \n in multiline)
157 OP_EXTUNI match an extended Unicode character
158 OP_ANYNL match any Unicode newline sequence
160 OP_ACCEPT ) These are Perl 5.10's "backtracking control
161 OP_COMMIT ) verbs". If OP_ACCEPT is inside capturing
162 OP_FAIL ) parentheses, it may be preceded by one or more
163 OP_PRUNE ) OP_CLOSE, followed by a 2-byte number,
164 OP_SKIP ) indicating which parentheses must be closed.
167 Backtracking control verbs with data
168 ------------------------------------
170 OP_THEN is followed by a LINK_SIZE offset, which is the distance back to the
171 start of the current branch.
173 OP_MARK is followed by the mark name, preceded by a one-byte length, and
174 followed by a binary zero. For (*PRUNE), (*SKIP), and (*THEN) with arguments,
175 the opcodes OP_PRUNE_ARG, OP_SKIP_ARG, and OP_THEN_ARG are used. For the first
176 two, the name follows immediately; for OP_THEN_ARG, it follows the LINK_SIZE
180 Repeating single characters
181 ---------------------------
183 The common repeats (*, +, ?) when applied to a single character use the
196 In ASCII mode, these are two-byte items; in UTF-8 mode, the length is variable.
197 Those with "MIN" in their name are the minimizing versions. Those with "POS" in
198 their names are possessive versions. Each is followed by the character that is
199 to be repeated. Other repeats make use of
206 which are followed by a two-byte count (most significant first) and the
207 repeated character. OP_UPTO matches from 0 to the given number. A repeat with a
208 non-zero minimum and a fixed maximum is coded as an OP_EXACT followed by an
209 OP_UPTO (or OP_MINUPTO or OPT_POSUPTO).
212 Repeating character types
213 -------------------------
215 Repeats of things like \d are done exactly as for single characters, except
216 that instead of a character, the opcode for the type is stored in the data
217 byte. The opcodes are:
234 Match by Unicode property
235 -------------------------
237 OP_PROP and OP_NOTPROP are used for positive and negative matches of a
238 character by testing its Unicode property (the \p and \P escape sequences).
239 Each is followed by two bytes that encode the desired property as a type and a
242 Repeats of these items use the OP_TYPESTAR etc. set of opcodes, followed by
243 three bytes: OP_PROP or OP_NOTPROP and then the desired property type and
247 Matching literal characters
248 ---------------------------
250 The OP_CHAR opcode is followed by a single character that is to be matched
251 casefully. For caseless matching, OP_CHARNC is used. In UTF-8 mode, the
252 character may be more than one byte long. (Earlier versions of PCRE used
253 multi-character strings, but this was changed to allow some new features to be
260 If there is only one character, OP_CHAR or OP_CHARNC is used for a positive
261 class, and OP_NOT for a negative one (that is, for something like [^a]).
262 However, in UTF-8 mode, the use of OP_NOT applies only to characters with
263 values < 128, because OP_NOT is confined to single bytes.
265 Another set of repeating opcodes (OP_NOTSTAR etc.) are used for a repeated,
266 negated, single-character class. The normal ones (OP_STAR etc.) are used for a
267 repeated positive single-character class.
269 When there's more than one character in a class and all the characters are less
270 than 256, OP_CLASS is used for a positive class, and OP_NCLASS for a negative
271 one. In either case, the opcode is followed by a 32-byte bit map containing a 1
272 bit for every character that is acceptable. The bits are counted from the least
273 significant end of each byte.
275 The reason for having both OP_CLASS and OP_NCLASS is so that, in UTF-8 mode,
276 subject characters with values greater than 256 can be handled correctly. For
277 OP_CLASS they don't match, whereas for OP_NCLASS they do.
279 For classes containing characters with values > 255, OP_XCLASS is used. It
280 optionally uses a bit map (if any characters lie within it), followed by a list
281 of pairs and single characters. There is a flag character than indicates
282 whether it's a positive or a negative class.
288 OP_REF is followed by two bytes containing the reference number.
291 Repeating character classes and back references
292 -----------------------------------------------
294 Single-character classes are handled specially (see above). This section
295 applies to OP_CLASS and OP_REF. In both cases, the repeat information follows
296 the base item. The matching code looks at the following opcode to see if it is
308 All but the last two are just single-byte items. The others are followed by
309 four bytes of data, comprising the minimum and maximum repeat counts. There are
310 no special possessive opcodes for these repeats; a possessive repeat is
311 compiled into an atomic group.
314 Brackets and alternation
315 ------------------------
317 A pair of non-capturing (round) brackets is wrapped round each expression at
318 compile time, so alternation always happens in the context of brackets.
320 [Note for North Americans: "bracket" to some English speakers, including
321 myself, can be round, square, curly, or pointy. Hence this usage.]
323 Non-capturing brackets use the opcode OP_BRA. Originally PCRE was limited to 99
324 capturing brackets and it used a different opcode for each one. From release
325 3.5, the limit was removed by putting the bracket number into the data for
326 higher-numbered brackets. From release 7.0 all capturing brackets are handled
327 this way, using the single opcode OP_CBRA.
329 A bracket opcode is followed by LINK_SIZE bytes which give the offset to the
330 next alternative OP_ALT or, if there aren't any branches, to the matching
331 OP_KET opcode. Each OP_ALT is followed by LINK_SIZE bytes giving the offset to
332 the next one, or to the OP_KET opcode. For capturing brackets, the bracket
333 number immediately follows the offset, always as a 2-byte item.
335 OP_KET is used for subpatterns that do not repeat indefinitely, while
336 OP_KETRMIN and OP_KETRMAX are used for indefinite repetitions, minimally or
337 maximally respectively. All three are followed by LINK_SIZE bytes giving (as a
338 positive number) the offset back to the matching bracket opcode.
340 If a subpattern is quantified such that it is permitted to match zero times, it
341 is preceded by one of OP_BRAZERO, OP_BRAMINZERO, or OP_SKIPZERO. These are
342 single-byte opcodes that tell the matcher that skipping the following
343 subpattern entirely is a valid branch. In the case of the first two, not
344 skipping the pattern is also valid (greedy and non-greedy). The third is used
345 when a pattern has the quantifier {0,0}. It cannot be entirely discarded,
346 because it may be called as a subroutine from elsewhere in the regex.
348 A subpattern with an indefinite maximum repetition is replicated in the
349 compiled data its minimum number of times (or once with OP_BRAZERO if the
350 minimum is zero), with the final copy terminating with OP_KETRMIN or OP_KETRMAX
353 A subpattern with a bounded maximum repetition is replicated in a nested
354 fashion up to the maximum number of times, with OP_BRAZERO or OP_BRAMINZERO
355 before each replication after the minimum, so that, for example, (abc){2,5} is
356 compiled as (abc)(abc)((abc)((abc)(abc)?)?)?, except that each bracketed group
359 When a repeated subpattern has an unbounded upper limit, it is checked to see
360 whether it could match an empty string. If this is the case, the opcode in the
361 final replication is changed to OP_SBRA or OP_SCBRA. This tells the matcher
362 that it needs to check for matching an empty string when it hits OP_KETRMIN or
363 OP_KETRMAX, and if so, to break the loop.
369 Forward assertions are just like other subpatterns, but starting with one of
370 the opcodes OP_ASSERT or OP_ASSERT_NOT. Backward assertions use the opcodes
371 OP_ASSERTBACK and OP_ASSERTBACK_NOT, and the first opcode inside the assertion
372 is OP_REVERSE, followed by a two byte count of the number of characters to move
373 back the pointer in the subject string. When operating in UTF-8 mode, the count
374 is a character count rather than a byte count. A separate count is present in
375 each alternative of a lookbehind assertion, allowing them to have different
379 Once-only (atomic) subpatterns
380 ------------------------------
382 These are also just like other subpatterns, but they start with the opcode
383 OP_ONCE. The check for matching an empty string in an unbounded repeat is
384 handled entirely at runtime, so there is just this one opcode.
387 Conditional subpatterns
388 -----------------------
390 These are like other subpatterns, but they start with the opcode OP_COND, or
391 OP_SCOND for one that might match an empty string in an unbounded repeat. If
392 the condition is a back reference, this is stored at the start of the
393 subpattern using the opcode OP_CREF followed by two bytes containing the
394 reference number. OP_NCREF is used instead if the reference was generated by
395 name (so that the runtime code knows to check for duplicate names).
397 If the condition is "in recursion" (coded as "(?(R)"), or "in recursion of
398 group x" (coded as "(?(Rx)"), the group number is stored at the start of the
399 subpattern using the opcode OP_RREF or OP_NRREF (cf OP_NCREF), and a value of
400 zero for "the whole pattern". For a DEFINE condition, just the single byte
401 OP_DEF is used (it has no associated data). Otherwise, a conditional subpattern
402 always starts with one of the assertions.
408 Recursion either matches the current regex, or some subexpression. The opcode
409 OP_RECURSE is followed by an value which is the offset to the starting bracket
410 from the start of the whole pattern. From release 6.5, OP_RECURSE is
411 automatically wrapped inside OP_ONCE brackets (because otherwise some patterns
412 broke it). OP_RECURSE is also used for "subroutine" calls, even though they
413 are not strictly a recursion.
419 OP_CALLOUT is followed by one byte of data that holds a callout number in the
420 range 0 to 254 for manual callouts, or 255 for an automatic callout. In both
421 cases there follows a two-byte value giving the offset in the pattern to the
422 start of the following item, and another two-byte item giving the length of the
429 If any of the /i, /m, or /s options are changed within a pattern, an OP_OPT
430 opcode is compiled, followed by one byte containing the new settings of these
431 flags. If there are several alternatives, there is an occurrence of OP_OPT at
432 the start of all those following the first options change, to set appropriate
433 options for the start of the alternative. Immediately after the end of the
434 group there is another such item to reset the flags to their previous values. A
435 change of flag right at the very start of the pattern can be handled entirely
436 at compile time, and so does not cause anything to be put into the compiled