trunk/doc/pcreperform.3

.TH PCREPERFORM 3
.SH NAME
PCRE - Perl-compatible regular expressions
.SH "PCRE PERFORMANCE"
.rs
.sp
Two aspects of performance are discussed below: memory usage and processing
time. The way you express your pattern as a regular expression can affect both
of them.
.
.SH "COMPILED PATTERN MEMORY USAGE"
.rs
.sp
Patterns are compiled by PCRE into a reasonably efficient byte code, so that
most simple patterns do not use much memory. However, there is one case where
the memory usage of a compiled pattern can be unexpectedly large. If a
parenthesized subpattern has a quantifier with a minimum greater than 1 and/or
a limited maximum, the whole subpattern is repeated in the compiled code. For
example, the pattern
.sp
  (abc|def){2,4}
.sp
is compiled as if it were
.sp
  (abc|def)(abc|def)((abc|def)(abc|def)?)?
.sp
(Technical aside: It is done this way so that backtrack points within each of
the repetitions can be independently maintained.)
.P
For regular expressions whose quantifiers use only small numbers, this is not
usually a problem. However, if the numbers are large, and particularly if such
repetitions are nested, the memory usage can become an embarrassment. For
example, the very simple pattern
.sp
  ((ab){1,1000}c){1,3}
.sp
uses 51K bytes when compiled. When PCRE is compiled with its default internal
pointer size of two bytes, the size limit on a compiled pattern is 64K, and
this is reached with the above pattern if the outer repetition is increased
from 3 to 4. PCRE can be compiled to use larger internal pointers and thus
handle larger compiled patterns, but it is better to try to rewrite your
pattern to use less memory if you can.
.P
One way of reducing the memory usage for such patterns is to make use of PCRE's
.\" HTML <a href="pcrepattern.html#subpatternsassubroutines">
.\" </a>
"subroutine"
.\"
facility. Re-writing the above pattern as
.sp
  ((ab)(?2){0,999}c)(?1){0,2}
.sp
reduces the memory requirements to 18K, and indeed it remains under 20K even
with the outer repetition increased to 100. However, this pattern is not
exactly equivalent, because the "subroutine" calls are treated as
.\" HTML <a href="pcrepattern.html#atomicgroup">
.\" </a>
atomic groups
.\"
into which there can be no backtracking if there is a subsequent matching
failure. Therefore, PCRE cannot do this kind of rewriting automatically.
Furthermore, there is a noticeable loss of speed when executing the modified
pattern. Nevertheless, if the atomic grouping is not a problem and the loss of
speed is acceptable, this kind of rewriting will allow you to process patterns
that PCRE cannot otherwise handle.
.
.
.SH "STACK USAGE AT RUN TIME"
.rs
.sp
When \fBpcre_exec()\fP is used for matching, certain kinds of pattern can cause
it to use large amounts of the process stack. In some environments the default
process stack is quite small, and if it runs out the result is often SIGSEGV.
This issue is probably the most frequently raised problem with PCRE. Rewriting
your pattern can often help. The
.\" HREF
\fBpcrestack\fP
.\"
documentation discusses this issue in detail.
.
.
.SH "PROCESSING TIME"
.rs
.sp
Certain items in regular expression patterns are processed more efficiently
than others. It is more efficient to use a character class like [aeiou] than a
set of single-character alternatives such as (a|e|i|o|u). In general, the
simplest construction that provides the required behaviour is usually the most
efficient. Jeffrey Friedl's book contains a lot of useful general discussion
about optimizing regular expressions for efficient performance. This document
contains a few observations about PCRE.
.P
Using Unicode character properties (the \ep, \eP, and \eX escapes) is slow,
because PCRE has to scan a structure that contains data for over fifteen
thousand characters whenever it needs a character's property. If you can find
an alternative pattern that does not use character properties, it will probably
be faster.
.P
By default, the escape sequences \eb, \ed, \es, and \ew, and the POSIX
character classes such as [:alpha:] do not use Unicode properties, partly for
backwards compatibility, and partly for performance reasons. However, you can
set PCRE_UCP if you want Unicode character properties to be used. This can
double the matching time for items such as \ed, when matched with
\fBpcre_exec()\fP; the performance loss is less with \fBpcre_dfa_exec()\fP, and
in both cases there is not much difference for \eb.
.P
When a pattern begins with .* not in parentheses, or in parentheses that are
not the subject of a backreference, and the PCRE_DOTALL option is set, the
pattern is implicitly anchored by PCRE, since it can match only at the start of
a subject string. However, if PCRE_DOTALL is not set, PCRE cannot make this
optimization, because the . metacharacter does not then match a newline, and if
the subject string contains newlines, the pattern may match from the character
immediately following one of them instead of from the very start. For example,
the pattern
.sp
  .*second
.sp
matches the subject "first\enand second" (where \en stands for a newline
character), with the match starting at the seventh character. In order to do
this, PCRE has to retry the match starting after every newline in the subject.
.P
If you are using such a pattern with subject strings that do not contain
newlines, the best performance is obtained by setting PCRE_DOTALL, or starting
the pattern with ^.* or ^.*? to indicate explicit anchoring. That saves PCRE
from having to scan along the subject looking for a newline to restart at.
.P
Beware of patterns that contain nested indefinite repeats. These can take a
long time to run when applied to a string that does not match. Consider the
pattern fragment
.sp
  ^(a+)*
.sp
This can match "aaaa" in 16 different ways, and this number increases very
rapidly as the string gets longer. (The * repeat can match 0, 1, 2, 3, or 4
times, and for each of those cases other than 0 or 4, the + repeats can match
different numbers of times.) When the remainder of the pattern is such that the
entire match is going to fail, PCRE has in principle to try every possible
variation, and this can take an extremely long time, even for relatively short
strings.
.P
An optimization catches some of the more simple cases such as
.sp
  (a+)*b
.sp
where a literal character follows. Before embarking on the standard matching
procedure, PCRE checks that there is a "b" later in the subject string, and if
there is not, it fails the match immediately. However, when there is no
following literal this optimization cannot be used. You can see the difference
by comparing the behaviour of
.sp
  (a+)*\ed
.sp
with the pattern above. The former gives a failure almost instantly when
applied to a whole line of "a" characters, whereas the latter takes an
appreciable time with strings longer than about 20 characters.
.P
In many cases, the solution to this kind of performance issue is to use an
atomic group or a possessive quantifier.
.
.
.SH AUTHOR
.rs
.sp
.nf
Philip Hazel
University Computing Service
Cambridge CB2 3QH, England.
.fi
.
.
.SH REVISION
.rs
.sp
.nf
Last updated: 16 May 2010
Copyright (c) 1997-2010 University of Cambridge.
.fi
1	.TH PCREPERFORM 3
2	.SH NAME
3	PCRE - Perl-compatible regular expressions
4	.SH "PCRE PERFORMANCE"
5	.rs
6	.sp
7	Two aspects of performance are discussed below: memory usage and processing
8	time. The way you express your pattern as a regular expression can affect both
9	of them.
10	.
11	.SH "COMPILED PATTERN MEMORY USAGE"
12	.rs
13	.sp
14	Patterns are compiled by PCRE into a reasonably efficient byte code, so that
15	most simple patterns do not use much memory. However, there is one case where
16	the memory usage of a compiled pattern can be unexpectedly large. If a
17	parenthesized subpattern has a quantifier with a minimum greater than 1 and/or
18	a limited maximum, the whole subpattern is repeated in the compiled code. For
19	example, the pattern
20	.sp
21	(abc\|def){2,4}
22	.sp
23	is compiled as if it were
24	.sp
25	(abc\|def)(abc\|def)((abc\|def)(abc\|def)?)?
26	.sp
27	(Technical aside: It is done this way so that backtrack points within each of
28	the repetitions can be independently maintained.)
29	.P
30	For regular expressions whose quantifiers use only small numbers, this is not
31	usually a problem. However, if the numbers are large, and particularly if such
32	repetitions are nested, the memory usage can become an embarrassment. For
33	example, the very simple pattern
34	.sp
35	((ab){1,1000}c){1,3}
36	.sp
37	uses 51K bytes when compiled. When PCRE is compiled with its default internal
38	pointer size of two bytes, the size limit on a compiled pattern is 64K, and
39	this is reached with the above pattern if the outer repetition is increased
40	from 3 to 4. PCRE can be compiled to use larger internal pointers and thus
41	handle larger compiled patterns, but it is better to try to rewrite your
42	pattern to use less memory if you can.
43	.P
44	One way of reducing the memory usage for such patterns is to make use of PCRE's
45	.\" HTML <a href="pcrepattern.html#subpatternsassubroutines">
46	.\" </a>
47	"subroutine"
48	.\"
49	facility. Re-writing the above pattern as
50	.sp
51	((ab)(?2){0,999}c)(?1){0,2}
52	.sp
53	reduces the memory requirements to 18K, and indeed it remains under 20K even
54	with the outer repetition increased to 100. However, this pattern is not
55	exactly equivalent, because the "subroutine" calls are treated as
56	.\" HTML <a href="pcrepattern.html#atomicgroup">
57	.\" </a>
58	atomic groups
59	.\"
60	into which there can be no backtracking if there is a subsequent matching
61	failure. Therefore, PCRE cannot do this kind of rewriting automatically.
62	Furthermore, there is a noticeable loss of speed when executing the modified
63	pattern. Nevertheless, if the atomic grouping is not a problem and the loss of
64	speed is acceptable, this kind of rewriting will allow you to process patterns
65	that PCRE cannot otherwise handle.
66	.
67	.
68	.SH "STACK USAGE AT RUN TIME"
69	.rs
70	.sp
71	When \fBpcre_exec()\fP is used for matching, certain kinds of pattern can cause
72	it to use large amounts of the process stack. In some environments the default
73	process stack is quite small, and if it runs out the result is often SIGSEGV.
74	This issue is probably the most frequently raised problem with PCRE. Rewriting
75	your pattern can often help. The
76	.\" HREF
77	\fBpcrestack\fP
78	.\"
79	documentation discusses this issue in detail.
80	.
81	.
82	.SH "PROCESSING TIME"
83	.rs
84	.sp
85	Certain items in regular expression patterns are processed more efficiently
86	than others. It is more efficient to use a character class like [aeiou] than a
87	set of single-character alternatives such as (a\|e\|i\|o\|u). In general, the
88	simplest construction that provides the required behaviour is usually the most
89	efficient. Jeffrey Friedl's book contains a lot of useful general discussion
90	about optimizing regular expressions for efficient performance. This document
91	contains a few observations about PCRE.
92	.P
93	Using Unicode character properties (the \ep, \eP, and \eX escapes) is slow,
94	because PCRE has to scan a structure that contains data for over fifteen
95	thousand characters whenever it needs a character's property. If you can find
96	an alternative pattern that does not use character properties, it will probably
97	be faster.
98	.P
99	By default, the escape sequences \eb, \ed, \es, and \ew, and the POSIX
100	character classes such as [:alpha:] do not use Unicode properties, partly for
101	backwards compatibility, and partly for performance reasons. However, you can
102	set PCRE_UCP if you want Unicode character properties to be used. This can
103	double the matching time for items such as \ed, when matched with
104	\fBpcre_exec()\fP; the performance loss is less with \fBpcre_dfa_exec()\fP, and
105	in both cases there is not much difference for \eb.
106	.P
107	When a pattern begins with .* not in parentheses, or in parentheses that are
108	not the subject of a backreference, and the PCRE_DOTALL option is set, the
109	pattern is implicitly anchored by PCRE, since it can match only at the start of
110	a subject string. However, if PCRE_DOTALL is not set, PCRE cannot make this
111	optimization, because the . metacharacter does not then match a newline, and if
112	the subject string contains newlines, the pattern may match from the character
113	immediately following one of them instead of from the very start. For example,
114	the pattern
115	.sp
116	.*second
117	.sp
118	matches the subject "first\enand second" (where \en stands for a newline
119	character), with the match starting at the seventh character. In order to do
120	this, PCRE has to retry the match starting after every newline in the subject.
121	.P
122	If you are using such a pattern with subject strings that do not contain
123	newlines, the best performance is obtained by setting PCRE_DOTALL, or starting
124	the pattern with ^.* or ^.*? to indicate explicit anchoring. That saves PCRE
125	from having to scan along the subject looking for a newline to restart at.
126	.P
127	Beware of patterns that contain nested indefinite repeats. These can take a
128	long time to run when applied to a string that does not match. Consider the
129	pattern fragment
130	.sp
131	^(a+)*
132	.sp
133	This can match "aaaa" in 16 different ways, and this number increases very
134	rapidly as the string gets longer. (The * repeat can match 0, 1, 2, 3, or 4
135	times, and for each of those cases other than 0 or 4, the + repeats can match
136	different numbers of times.) When the remainder of the pattern is such that the
137	entire match is going to fail, PCRE has in principle to try every possible
138	variation, and this can take an extremely long time, even for relatively short
139	strings.
140	.P
141	An optimization catches some of the more simple cases such as
142	.sp
143	(a+)*b
144	.sp
145	where a literal character follows. Before embarking on the standard matching
146	procedure, PCRE checks that there is a "b" later in the subject string, and if
147	there is not, it fails the match immediately. However, when there is no
148	following literal this optimization cannot be used. You can see the difference
149	by comparing the behaviour of
150	.sp
151	(a+)*\ed
152	.sp
153	with the pattern above. The former gives a failure almost instantly when
154	applied to a whole line of "a" characters, whereas the latter takes an
155	appreciable time with strings longer than about 20 characters.
156	.P
157	In many cases, the solution to this kind of performance issue is to use an
158	atomic group or a possessive quantifier.
159	.
160	.
161	.SH AUTHOR
162	.rs
163	.sp
164	.nf
165	Philip Hazel
166	University Computing Service
167	Cambridge CB2 3QH, England.
168	.fi
169	.
170	.
171	.SH REVISION
172	.rs
173	.sp
174	.nf
175	Last updated: 16 May 2010
176	Copyright (c) 1997-2010 University of Cambridge.
177	.fi
Name	Value
svn:eol-style	native
svn:keywords	"Author Date Id Revision Url"