ViewVC logotype

Contents of /code/trunk/HACKING

Parent Directory Parent Directory | Revision Log Revision Log

Revision 114 - (hide annotations) (download)
Fri Mar 9 10:15:12 2007 UTC (8 years, 1 month ago) by ph10
Original Path: code/trunk/maint/Tech.Notes
File size: 17135 byte(s)
Rename "maintain" as "maint".

1 nigel 41 Technical Notes about PCRE
2     --------------------------
4 nigel 91 These are very rough technical notes that record potentially useful information
5     about PCRE internals.
7 nigel 75 Historical note 1
8     -----------------
10 nigel 41 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 nigel 63 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 nigel 75 subject string. In the terminology of Jeffrey Friedl's book, it was a "DFA
19 nigel 93 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.
24 nigel 41
25 nigel 75 Historical note 2
26     -----------------
28 nigel 91 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
35     Friedl's terminology.
36 nigel 41
37 nigel 75 OK, here's the real stuff
38     -------------------------
40 nigel 77 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 nigel 93 first pass through the pattern is helpful for other reasons.
46 nigel 41
47 nigel 93 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 are also present in the 7.0 release).
72     A side effect of this work is 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 is that this is not a big
76     issue.
78 nigel 77 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. Not surprising, since it is intended to be as
84     compatible with Perl as possible. This is the function most users of PCRE will
85     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 nigel 91 facilities are available, and those that are do not always work in quite the
96 nigel 77 same way. See the user documentation for details.
98 nigel 93 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 nigel 77 Format of compiled patterns
105     ---------------------------
107 nigel 41 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 nigel 75 item is either implicit in the opcode or contained in the data bytes that
110     follow it.
111 nigel 41
112 nigel 75 In many cases below "two-byte" data values are specified. This is in fact just
113 nigel 93 a default when the number is an offset within the compiled pattern. PCRE can be
114     compiled to use 3-byte or 4-byte values for these offsets (impairing the
115 nigel 75 performance). This is necessary only when patterns whose compiled length is
116 nigel 93 greater than 64K are going to be processed. In this description, we assume the
117     "normal" compilation options. "Two-byte" data values that are counts (e.g. for
118     quantifiers) are always just two bytes.
119 nigel 75
120     A list of all the opcodes follows:
122 nigel 41 Opcodes with no following data
123     ------------------------------
125     These items are all just one byte long
127     OP_END end of pattern
128     OP_ANY match any character
129 nigel 75 OP_ANYBYTE match any single byte, even in UTF-8 mode
130 nigel 41 OP_SOD match start of data: \A
131 nigel 71 OP_SOM, start of match (subject + offset): \G
132 nigel 41 OP_CIRC ^ (start of data, or after \n in multiline)
135     OP_NOT_DIGIT \D
136     OP_DIGIT \d
138     OP_WHITESPACE \s
140     OP_WORDCHAR \w
141     OP_EODN match end of data or \n at end: \Z
142     OP_EOD match end of data: \z
143     OP_DOLL $ (end of data, or before \n in multiline)
144 nigel 75 OP_EXTUNI match an extended Unicode character
145 nigel 93 OP_ANYNL match any Unicode newline sequence
146 nigel 75
147 nigel 41
148     Repeating single characters
149     ---------------------------
151 nigel 75 The common repeats (*, +, ?) when applied to a single character use the
152     following opcodes:
153 nigel 41
154     OP_STAR
155     OP_MINSTAR
156 nigel 93 OP_POSSTAR
157 nigel 41 OP_PLUS
158     OP_MINPLUS
159 nigel 93 OP_POSPLUS
160 nigel 41 OP_QUERY
162 nigel 93 OP_POSQUERY
163 nigel 41
164 nigel 75 In ASCII mode, these are two-byte items; in UTF-8 mode, the length is variable.
165 nigel 93 Those with "MIN" in their name are the minimizing versions. Those with "POS" in
166     their names are possessive versions. Each is followed by the character that is
167     to be repeated. Other repeats make use of
168 nigel 41
169     OP_UPTO
170     OP_MINUPTO
171 nigel 93 OP_POSUPTO
172 nigel 41 OP_EXACT
174     which are followed by a two-byte count (most significant first) and the
175     repeated character. OP_UPTO matches from 0 to the given number. A repeat with a
176     non-zero minimum and a fixed maximum is coded as an OP_EXACT followed by an
177 nigel 93 OP_UPTO (or OP_MINUPTO or OPT_POSUPTO).
178 nigel 41
180     Repeating character types
181     -------------------------
183     Repeats of things like \d are done exactly as for single characters, except
184     that instead of a character, the opcode for the type is stored in the data
185     byte. The opcodes are:
189 nigel 93 OP_TYPEPOSSTAR
190 nigel 41 OP_TYPEPLUS
192 nigel 93 OP_TYPEPOSPLUS
193 nigel 41 OP_TYPEQUERY
195 nigel 93 OP_TYPEPOSQUERY
196 nigel 41 OP_TYPEUPTO
198 nigel 93 OP_TYPEPOSUPTO
199 nigel 41 OP_TYPEEXACT
202 nigel 75 Match by Unicode property
203     -------------------------
205     OP_PROP and OP_NOTPROP are used for positive and negative matches of a
206     character by testing its Unicode property (the \p and \P escape sequences).
207 nigel 91 Each is followed by two bytes that encode the desired property as a type and a
208     value.
209 nigel 75
210 nigel 91 Repeats of these items use the OP_TYPESTAR etc. set of opcodes, followed by
211     three bytes: OP_PROP or OP_NOTPROP and then the desired property type and
212     value.
213 nigel 75
215     Matching literal characters
216 nigel 41 ---------------------------
218 nigel 75 The OP_CHAR opcode is followed by a single character that is to be matched
219     casefully. For caseless matching, OP_CHARNC is used. In UTF-8 mode, the
220     character may be more than one byte long. (Earlier versions of PCRE used
221     multi-character strings, but this was changed to allow some new features to be
222     added.)
223 nigel 41
225     Character classes
226     -----------------
228 nigel 75 If there is only one character, OP_CHAR or OP_CHARNC is used for a positive
229     class, and OP_NOT for a negative one (that is, for something like [^a]).
230     However, in UTF-8 mode, the use of OP_NOT applies only to characters with
231     values < 128, because OP_NOT is confined to single bytes.
232 nigel 41
233 nigel 63 Another set of repeating opcodes (OP_NOTSTAR etc.) are used for a repeated,
234     negated, single-character class. The normal ones (OP_STAR etc.) are used for a
235     repeated positive single-character class.
237 nigel 71 When there's more than one character in a class and all the characters are less
238 nigel 75 than 256, OP_CLASS is used for a positive class, and OP_NCLASS for a negative
239 nigel 71 one. In either case, the opcode is followed by a 32-byte bit map containing a 1
240     bit for every character that is acceptable. The bits are counted from the least
241     significant end of each byte.
242 nigel 41
243 nigel 75 The reason for having both OP_CLASS and OP_NCLASS is so that, in UTF-8 mode,
244     subject characters with values greater than 256 can be handled correctly. For
245 nigel 71 OP_CLASS they don't match, whereas for OP_NCLASS they do.
247 nigel 63 For classes containing characters with values > 255, OP_XCLASS is used. It
248     optionally uses a bit map (if any characters lie within it), followed by a list
249 nigel 75 of pairs and single characters. There is a flag character than indicates
250 nigel 63 whether it's a positive or a negative class.
251 nigel 41
252 nigel 63
253 nigel 41 Back references
254     ---------------
256 nigel 53 OP_REF is followed by two bytes containing the reference number.
257 nigel 41
259     Repeating character classes and back references
260     -----------------------------------------------
262 nigel 93 Single-character classes are handled specially (see above). This section
263     applies to OP_CLASS and OP_REF. In both cases, the repeat information follows
264     the base item. The matching code looks at the following opcode to see if it is
265     one of
266 nigel 41
267     OP_CRSTAR
269     OP_CRPLUS
271     OP_CRQUERY
273     OP_CRRANGE
276     All but the last two are just single-byte items. The others are followed by
277 nigel 93 four bytes of data, comprising the minimum and maximum repeat counts. There are
278     no special possessive opcodes for these repeats; a possessive repeat is
279     compiled into an atomic group.
280 nigel 41
282     Brackets and alternation
283     ------------------------
285 nigel 43 A pair of non-capturing (round) brackets is wrapped round each expression at
286 nigel 41 compile time, so alternation always happens in the context of brackets.
287 nigel 53
288 nigel 93 [Note for North Americans: "bracket" to some English speakers, including
289     myself, can be round, square, curly, or pointy. Hence this usage.]
290 nigel 41
291 nigel 93 Non-capturing brackets use the opcode OP_BRA. Originally PCRE was limited to 99
292     capturing brackets and it used a different opcode for each one. From release
293     3.5, the limit was removed by putting the bracket number into the data for
294     higher-numbered brackets. From release 7.0 all capturing brackets are handled
295     this way, using the single opcode OP_CBRA.
296 nigel 53
297 nigel 77 A bracket opcode is followed by LINK_SIZE bytes which give the offset to the
298     next alternative OP_ALT or, if there aren't any branches, to the matching
299     OP_KET opcode. Each OP_ALT is followed by LINK_SIZE bytes giving the offset to
300 nigel 93 the next one, or to the OP_KET opcode. For capturing brackets, the bracket
301     number immediately follows the offset, always as a 2-byte item.
302 nigel 41
303     OP_KET is used for subpatterns that do not repeat indefinitely, while
304     OP_KETRMIN and OP_KETRMAX are used for indefinite repetitions, minimally or
305 nigel 77 maximally respectively. All three are followed by LINK_SIZE bytes giving (as a
306 nigel 93 positive number) the offset back to the matching bracket opcode.
307 nigel 41
308     If a subpattern is quantified such that it is permitted to match zero times, it
309     is preceded by one of OP_BRAZERO or OP_BRAMINZERO. These are single-byte
310     opcodes which tell the matcher that skipping this subpattern entirely is a
311     valid branch.
313     A subpattern with an indefinite maximum repetition is replicated in the
314 nigel 75 compiled data its minimum number of times (or once with OP_BRAZERO if the
315     minimum is zero), with the final copy terminating with OP_KETRMIN or OP_KETRMAX
316     as appropriate.
317 nigel 41
318     A subpattern with a bounded maximum repetition is replicated in a nested
319 nigel 75 fashion up to the maximum number of times, with OP_BRAZERO or OP_BRAMINZERO
320     before each replication after the minimum, so that, for example, (abc){2,5} is
321 nigel 93 compiled as (abc)(abc)((abc)((abc)(abc)?)?)?, except that each bracketed group
322     has the same number.
323 nigel 41
324 nigel 93 When a repeated subpattern has an unbounded upper limit, it is checked to see
325     whether it could match an empty string. If this is the case, the opcode in the
326     final replication is changed to OP_SBRA or OP_SCBRA. This tells the matcher
327     that it needs to check for matching an empty string when it hits OP_KETRMIN or
328     OP_KETRMAX, and if so, to break the loop.
329 nigel 41
330 nigel 93
331 nigel 41 Assertions
332     ----------
334     Forward assertions are just like other subpatterns, but starting with one of
335     the opcodes OP_ASSERT or OP_ASSERT_NOT. Backward assertions use the opcodes
336     OP_ASSERTBACK and OP_ASSERTBACK_NOT, and the first opcode inside the assertion
337     is OP_REVERSE, followed by a two byte count of the number of characters to move
338 nigel 49 back the pointer in the subject string. When operating in UTF-8 mode, the count
339     is a character count rather than a byte count. A separate count is present in
340     each alternative of a lookbehind assertion, allowing them to have different
341     fixed lengths.
342 nigel 41
344 nigel 93 Once-only (atomic) subpatterns
345     ------------------------------
346 nigel 41
347     These are also just like other subpatterns, but they start with the opcode
348 nigel 93 OP_ONCE. The check for matching an empty string in an unbounded repeat is
349     handled entirely at runtime, so there is just this one opcode.
350 nigel 41
352     Conditional subpatterns
353     -----------------------
355 nigel 93 These are like other subpatterns, but they start with the opcode OP_COND, or
356     OP_SCOND for one that might match an empty string in an unbounded repeat. If
357 nigel 41 the condition is a back reference, this is stored at the start of the
358 nigel 53 subpattern using the opcode OP_CREF followed by two bytes containing the
359 nigel 93 reference number. If the condition is "in recursion" (coded as "(?(R)"), or "in
360     recursion of group x" (coded as "(?(Rx)"), the group number is stored at the
361     start of the subpattern using the opcode OP_RREF, and a value of zero for "the
362     whole pattern". For a DEFINE condition, just the single byte OP_DEF is used (it
363     has no associated data). Otherwise, a conditional subpattern always starts with
364     one of the assertions.
365 nigel 41
367 nigel 71 Recursion
368     ---------
370     Recursion either matches the current regex, or some subexpression. The opcode
371     OP_RECURSE is followed by an value which is the offset to the starting bracket
372 nigel 87 from the start of the whole pattern. From release 6.5, OP_RECURSE is
373     automatically wrapped inside OP_ONCE brackets (because otherwise some patterns
374     broke it). OP_RECURSE is also used for "subroutine" calls, even though they
375     are not strictly a recursion.
376 nigel 71
378     Callout
379     -------
381 nigel 75 OP_CALLOUT is followed by one byte of data that holds a callout number in the
382     range 0 to 254 for manual callouts, or 255 for an automatic callout. In both
383     cases there follows a two-byte value giving the offset in the pattern to the
384     start of the following item, and another two-byte item giving the length of the
385     next item.
386 nigel 71
388 nigel 41 Changing options
389     ----------------
391 nigel 63 If any of the /i, /m, or /s options are changed within a pattern, an OP_OPT
392     opcode is compiled, followed by one byte containing the new settings of these
393     flags. If there are several alternatives, there is an occurrence of OP_OPT at
394     the start of all those following the first options change, to set appropriate
395     options for the start of the alternative. Immediately after the end of the
396     group there is another such item to reset the flags to their previous values. A
397     change of flag right at the very start of the pattern can be handled entirely
398     at compile time, and so does not cause anything to be put into the compiled
399     data.
400 nigel 41
401     Philip Hazel
402 nigel 93 November 2006

ViewVC Help
Powered by ViewVC 1.1.12