1 Technical Notes about PCRE
2 --------------------------
4 These are very rough technical notes that record potentially useful information
10 Many years ago I implemented some regular expression functions to an algorithm
11 suggested by Martin Richards. These were not Unix-like in form, and were quite
12 restricted in what they could do by comparison with Perl. The interesting part
13 about the algorithm was that the amount of space required to hold the compiled
14 form of an expression was known in advance. The code to apply an expression did
15 not operate by backtracking, as the original Henry Spencer code and current
16 Perl code does, but instead checked all possibilities simultaneously by keeping
17 a list of current states and checking all of them as it advanced through the
18 subject string. In the terminology of Jeffrey Friedl's book, it was a "DFA
19 algorithm", though it was not a traditional Finite State Machine (FSM). When
20 the pattern was all used up, all remaining states were possible matches, and
21 the one matching the longest subset of the subject string was chosen. This did
22 not necessarily maximize the individual wild portions of the pattern, as is
23 expected in Unix and Perl-style regular expressions.
28 By contrast, the code originally written by Henry Spencer (which was
29 subsequently heavily modified for Perl) compiles the expression twice: once in
30 a dummy mode in order to find out how much store will be needed, and then for
31 real. (The Perl version probably doesn't do this any more; I'm talking about
32 the original library.) The execution function operates by backtracking and
33 maximizing (or, optionally, minimizing in Perl) the amount of the subject that
34 matches individual wild portions of the pattern. This is an "NFA algorithm" in
37 OK, here's the real stuff
38 -------------------------
40 For the set of functions that form the "basic" PCRE library (which are
41 unrelated to those mentioned above), I tried at first to invent an algorithm
42 that used an amount of store bounded by a multiple of the number of characters
43 in the pattern, to save on compiling time. However, because of the greater
44 complexity in Perl regular expressions, I couldn't do this. In any case, a
45 first pass through the pattern is helpful for other reasons.
47 Computing the memory requirement: how it was
48 --------------------------------------------
50 Up to and including release 6.7, PCRE worked by running a very degenerate first
51 pass to calculate a maximum store size, and then a second pass to do the real
52 compile - which might use a bit less than the predicted amount of memory. The
53 idea was that this would turn out faster than the Henry Spencer code because
54 the first pass is degenerate and the second pass can just store stuff straight
55 into the vector, which it knows is big enough.
57 Computing the memory requirement: how it is
58 -------------------------------------------
60 By the time I was working on a potential 6.8 release, the degenerate first pass
61 had become very complicated and hard to maintain. Indeed one of the early
62 things I did for 6.8 was to fix Yet Another Bug in the memory computation. Then
63 I had a flash of inspiration as to how I could run the real compile function in
64 a "fake" mode that enables it to compute how much memory it would need, while
65 actually only ever using a few hundred bytes of working memory, and without too
66 many tests of the mode that might slow it down. So I re-factored the compiling
67 functions to work this way. This got rid of about 600 lines of source. It
68 should make future maintenance and development easier. As this was such a major
69 change, I never released 6.8, instead upping the number to 7.0 (other quite
70 major changes were also present in the 7.0 release).
72 A side effect of this work was that the previous limit of 200 on the nesting
73 depth of parentheses was removed. However, there is a downside: pcre_compile()
74 runs more slowly than before (30% or more, depending on the pattern) because it
75 is doing a full analysis of the pattern. My hope was that this would not be a
76 big issue, and in the event, nobody has commented on it.
78 Traditional matching function
79 -----------------------------
81 The "traditional", and original, matching function is called pcre_exec(), and
82 it implements an NFA algorithm, similar to the original Henry Spencer algorithm
83 and the way that Perl works. This is not surprising, since it is intended to be
84 as compatible with Perl as possible. This is the function most users of PCRE
85 will use most of the time.
87 Supplementary matching function
88 -------------------------------
90 From PCRE 6.0, there is also a supplementary matching function called
91 pcre_dfa_exec(). This implements a DFA matching algorithm that searches
92 simultaneously for all possible matches that start at one point in the subject
93 string. (Going back to my roots: see Historical Note 1 above.) This function
94 intreprets the same compiled pattern data as pcre_exec(); however, not all the
95 facilities are available, and those that are do not always work in quite the
96 same way. See the user documentation for details.
98 The algorithm that is used for pcre_dfa_exec() is not a traditional FSM,
99 because it may have a number of states active at one time. More work would be
100 needed at compile time to produce a traditional FSM where only one state is
101 ever active at once. I believe some other regex matchers work this way.
104 Format of compiled patterns
105 ---------------------------
107 The compiled form of a pattern is a vector of bytes, containing items of
108 variable length. The first byte in an item is an opcode, and the length of the
109 item is either implicit in the opcode or contained in the data bytes that
112 In many cases below LINK_SIZE data values are specified for offsets within the
113 compiled pattern. The default value for LINK_SIZE is 2, but PCRE can be
114 compiled to use 3-byte or 4-byte values for these offsets (impairing the
115 performance). This is necessary only when patterns whose compiled length is
116 greater than 64K are going to be processed. In this description, we assume the
117 "normal" compilation options. Data values that are counts (e.g. for
118 quantifiers) are always just two bytes long.
120 A list of the opcodes follows:
123 Opcodes with no following data
124 ------------------------------
126 These items are all just one byte long
128 OP_END end of pattern
129 OP_ANY match any one character other than newline
130 OP_ALLANY match any one character, including newline
131 OP_ANYBYTE match any single byte, even in UTF-8 mode
132 OP_SOD match start of data: \A
133 OP_SOM, start of match (subject + offset): \G
134 OP_SET_SOM, set start of match (\K)
135 OP_CIRC ^ (start of data, or after \n in multiline)
136 OP_NOT_WORD_BOUNDARY \W
148 OP_EODN match end of data or \n at end: \Z
149 OP_EOD match end of data: \z
150 OP_DOLL $ (end of data, or before \n in multiline)
151 OP_EXTUNI match an extended Unicode character
152 OP_ANYNL match any Unicode newline sequence
154 OP_ACCEPT ) These are Perl 5.10's "backtracking
155 OP_COMMIT ) control verbs". If OP_ACCEPT is inside
156 OP_FAIL ) capturing parentheses, it may be preceded
157 OP_PRUNE ) by one or more OP_CLOSE, followed by a 2-byte
158 OP_SKIP ) number, indicating which parentheses must be
162 Repeating single characters
163 ---------------------------
165 The common repeats (*, +, ?) when applied to a single character use the
178 In ASCII mode, these are two-byte items; in UTF-8 mode, the length is variable.
179 Those with "MIN" in their name are the minimizing versions. Those with "POS" in
180 their names are possessive versions. Each is followed by the character that is
181 to be repeated. Other repeats make use of
188 which are followed by a two-byte count (most significant first) and the
189 repeated character. OP_UPTO matches from 0 to the given number. A repeat with a
190 non-zero minimum and a fixed maximum is coded as an OP_EXACT followed by an
191 OP_UPTO (or OP_MINUPTO or OPT_POSUPTO).
194 Repeating character types
195 -------------------------
197 Repeats of things like \d are done exactly as for single characters, except
198 that instead of a character, the opcode for the type is stored in the data
199 byte. The opcodes are:
216 Match by Unicode property
217 -------------------------
219 OP_PROP and OP_NOTPROP are used for positive and negative matches of a
220 character by testing its Unicode property (the \p and \P escape sequences).
221 Each is followed by two bytes that encode the desired property as a type and a
224 Repeats of these items use the OP_TYPESTAR etc. set of opcodes, followed by
225 three bytes: OP_PROP or OP_NOTPROP and then the desired property type and
229 Matching literal characters
230 ---------------------------
232 The OP_CHAR opcode is followed by a single character that is to be matched
233 casefully. For caseless matching, OP_CHARNC is used. In UTF-8 mode, the
234 character may be more than one byte long. (Earlier versions of PCRE used
235 multi-character strings, but this was changed to allow some new features to be
242 If there is only one character, OP_CHAR or OP_CHARNC is used for a positive
243 class, and OP_NOT for a negative one (that is, for something like [^a]).
244 However, in UTF-8 mode, the use of OP_NOT applies only to characters with
245 values < 128, because OP_NOT is confined to single bytes.
247 Another set of repeating opcodes (OP_NOTSTAR etc.) are used for a repeated,
248 negated, single-character class. The normal ones (OP_STAR etc.) are used for a
249 repeated positive single-character class.
251 When there's more than one character in a class and all the characters are less
252 than 256, OP_CLASS is used for a positive class, and OP_NCLASS for a negative
253 one. In either case, the opcode is followed by a 32-byte bit map containing a 1
254 bit for every character that is acceptable. The bits are counted from the least
255 significant end of each byte.
257 The reason for having both OP_CLASS and OP_NCLASS is so that, in UTF-8 mode,
258 subject characters with values greater than 256 can be handled correctly. For
259 OP_CLASS they don't match, whereas for OP_NCLASS they do.
261 For classes containing characters with values > 255, OP_XCLASS is used. It
262 optionally uses a bit map (if any characters lie within it), followed by a list
263 of pairs and single characters. There is a flag character than indicates
264 whether it's a positive or a negative class.
270 OP_REF is followed by two bytes containing the reference number.
273 Repeating character classes and back references
274 -----------------------------------------------
276 Single-character classes are handled specially (see above). This section
277 applies to OP_CLASS and OP_REF. In both cases, the repeat information follows
278 the base item. The matching code looks at the following opcode to see if it is
290 All but the last two are just single-byte items. The others are followed by
291 four bytes of data, comprising the minimum and maximum repeat counts. There are
292 no special possessive opcodes for these repeats; a possessive repeat is
293 compiled into an atomic group.
296 Brackets and alternation
297 ------------------------
299 A pair of non-capturing (round) brackets is wrapped round each expression at
300 compile time, so alternation always happens in the context of brackets.
302 [Note for North Americans: "bracket" to some English speakers, including
303 myself, can be round, square, curly, or pointy. Hence this usage.]
305 Non-capturing brackets use the opcode OP_BRA. Originally PCRE was limited to 99
306 capturing brackets and it used a different opcode for each one. From release
307 3.5, the limit was removed by putting the bracket number into the data for
308 higher-numbered brackets. From release 7.0 all capturing brackets are handled
309 this way, using the single opcode OP_CBRA.
311 A bracket opcode is followed by LINK_SIZE bytes which give the offset to the
312 next alternative OP_ALT or, if there aren't any branches, to the matching
313 OP_KET opcode. Each OP_ALT is followed by LINK_SIZE bytes giving the offset to
314 the next one, or to the OP_KET opcode. For capturing brackets, the bracket
315 number immediately follows the offset, always as a 2-byte item.
317 OP_KET is used for subpatterns that do not repeat indefinitely, while
318 OP_KETRMIN and OP_KETRMAX are used for indefinite repetitions, minimally or
319 maximally respectively. All three are followed by LINK_SIZE bytes giving (as a
320 positive number) the offset back to the matching bracket opcode.
322 If a subpattern is quantified such that it is permitted to match zero times, it
323 is preceded by one of OP_BRAZERO, OP_BRAMINZERO, or OP_SKIPZERO. These are
324 single-byte opcodes that tell the matcher that skipping the following
325 subpattern entirely is a valid branch. In the case of the first two, not
326 skipping the pattern is also valid (greedy and non-greedy). The third is used
327 when a pattern has the quantifier {0,0}. It cannot be entirely discarded,
328 because it may be called as a subroutine from elsewhere in the regex.
330 A subpattern with an indefinite maximum repetition is replicated in the
331 compiled data its minimum number of times (or once with OP_BRAZERO if the
332 minimum is zero), with the final copy terminating with OP_KETRMIN or OP_KETRMAX
335 A subpattern with a bounded maximum repetition is replicated in a nested
336 fashion up to the maximum number of times, with OP_BRAZERO or OP_BRAMINZERO
337 before each replication after the minimum, so that, for example, (abc){2,5} is
338 compiled as (abc)(abc)((abc)((abc)(abc)?)?)?, except that each bracketed group
341 When a repeated subpattern has an unbounded upper limit, it is checked to see
342 whether it could match an empty string. If this is the case, the opcode in the
343 final replication is changed to OP_SBRA or OP_SCBRA. This tells the matcher
344 that it needs to check for matching an empty string when it hits OP_KETRMIN or
345 OP_KETRMAX, and if so, to break the loop.
351 Forward assertions are just like other subpatterns, but starting with one of
352 the opcodes OP_ASSERT or OP_ASSERT_NOT. Backward assertions use the opcodes
353 OP_ASSERTBACK and OP_ASSERTBACK_NOT, and the first opcode inside the assertion
354 is OP_REVERSE, followed by a two byte count of the number of characters to move
355 back the pointer in the subject string. When operating in UTF-8 mode, the count
356 is a character count rather than a byte count. A separate count is present in
357 each alternative of a lookbehind assertion, allowing them to have different
361 Once-only (atomic) subpatterns
362 ------------------------------
364 These are also just like other subpatterns, but they start with the opcode
365 OP_ONCE. The check for matching an empty string in an unbounded repeat is
366 handled entirely at runtime, so there is just this one opcode.
369 Conditional subpatterns
370 -----------------------
372 These are like other subpatterns, but they start with the opcode OP_COND, or
373 OP_SCOND for one that might match an empty string in an unbounded repeat. If
374 the condition is a back reference, this is stored at the start of the
375 subpattern using the opcode OP_CREF followed by two bytes containing the
376 reference number. OP_NCREF is used instead if the reference was generated by
377 name (so that the runtime code knows to check for duplicate names).
379 If the condition is "in recursion" (coded as "(?(R)"), or "in recursion of
380 group x" (coded as "(?(Rx)"), the group number is stored at the start of the
381 subpattern using the opcode OP_RREF or OP_NRREF (cf OP_NCREF), and a value of
382 zero for "the whole pattern". For a DEFINE condition, just the single byte
383 OP_DEF is used (it has no associated data). Otherwise, a conditional subpattern
384 always starts with one of the assertions.
390 Recursion either matches the current regex, or some subexpression. The opcode
391 OP_RECURSE is followed by an value which is the offset to the starting bracket
392 from the start of the whole pattern. From release 6.5, OP_RECURSE is
393 automatically wrapped inside OP_ONCE brackets (because otherwise some patterns
394 broke it). OP_RECURSE is also used for "subroutine" calls, even though they
395 are not strictly a recursion.
401 OP_CALLOUT is followed by one byte of data that holds a callout number in the
402 range 0 to 254 for manual callouts, or 255 for an automatic callout. In both
403 cases there follows a two-byte value giving the offset in the pattern to the
404 start of the following item, and another two-byte item giving the length of the
411 If any of the /i, /m, or /s options are changed within a pattern, an OP_OPT
412 opcode is compiled, followed by one byte containing the new settings of these
413 flags. If there are several alternatives, there is an occurrence of OP_OPT at
414 the start of all those following the first options change, to set appropriate
415 options for the start of the alternative. Immediately after the end of the
416 group there is another such item to reset the flags to their previous values. A
417 change of flag right at the very start of the pattern can be handled entirely
418 at compile time, and so does not cause anything to be put into the compiled
OSDN Git Service
