PERLFAQ6(1)
NNAAMMEE
perlfaq6 - Regexps ($Revision: 1.17 $, $Date: 1997/04/24
22:44:10 $)
DDEESSCCRRIIPPTTIIOONN
This section is surprisingly small because the rest of the
FAQ is littered with answers involving regular
expressions. For example, decoding a URL and checking
whether something is a number are handled with regular
expressions, but those answers are found elsewhere in this
document (in the section on Data and the Networking one on
networking, to be precise).
HHooww ccaann II hhooppee ttoo uussee rreegguullaarr eexxpprreessssiioonnss wwiitthhoouutt ccrreeaattiinngg
iilllleeggiibbllee aanndd uunnmmaaiinnttaaiinnaabbllee ccooddee??
Three techniques can make regular expressions maintainable
and understandable.
Comments Outside the Regexp
Describe what you're doing and how you're doing it,
using normal Perl comments.
# turn the line into the first word, a colon, and the
# number of characters on the rest of the line
s/^(\w+)(.*)/ lc($1) . ":" . length($2) /ge;
Comments Inside the Regexp
The /x modifier causes whitespace to be ignored in a
regexp pattern (except in a character class), and also
allows you to use normal comments there, too. As you
can imagine, whitespace and comments help a lot.
/x lets you turn this:
s{<(?:[^>'"]*|".*?"|'.*?')+>}{}gs;
into this:
s{ < # opening angle bracket
(?: # Non-backreffing grouping paren
[^>'"] * # 0 or more things that are neither > nor ' nor "
| # or else
".*?" # a section between double quotes (stingy match)
| # or else
'.*?' # a section between single quotes (stingy match)
) + # all occurring one or more times
> # closing angle bracket
}{}gsx; # replace with nothing, i.e. delete
It's still not quite so clear as prose, but it is very
useful for describing the meaning of each part of the
pattern.
Different Delimiters
While we normally think of patterns as being delimited
with / characters, they can be delimited by almost any
character. the perlre manpage describes this. For
example, the s/// above uses braces as delimiters.
Selecting another delimiter can avoid quoting the
delimiter within the pattern:
s/\/usr\/local/\/usr\/share/g; # bad delimiter choice
s#/usr/local#/usr/share#g; # better
II''mm hhaavviinngg ttrroouubbllee mmaattcchhiinngg oovveerr mmoorree tthhaann oonnee lliinnee..
WWhhaatt''ss wwrroonngg??
Either you don't have newlines in your string, or you
aren't using the correct modifier(s) on your pattern.
There are many ways to get multiline data into a string.
If you want it to happen automatically while reading
input, you'll want to set $/ (probably to '' for
paragraphs or undef for the whole file) to allow you to
read more than one line at a time.
Read the perlre manpage to help you decide which of /s and
/m (or both) you might want to use: /s allows dot to
include newline, and /m allows caret and dollar to match
next to a newline, not just at the end of the string. You
do need to make sure that you've actually got a multiline
string in there.
For example, this program detects duplicate words, even
when they span line breaks (but not paragraph ones). For
this example, we don't need /s because we aren't using dot
in a regular expression that we want to cross line
boundaries. Neither do we need /m because we aren't
wanting caret or dollar to match at any point inside the
record next to newlines. But it's imperative that $/ be
set to something other than the default, or else we won't
actually ever have a multiline record read in.
$/ = ''; # read in more whole paragraph, not just one line
while ( <> ) {
while ( /\b(\w\S+)(\s+\1)+\b/gi ) {
print "Duplicate $1 at paragraph $.\n";
}
}
Here's code that finds sentences that begin with "From "
(which would be mangled by many mailers):
$/ = ''; # read in more whole paragraph, not just one line
while ( <> ) {
while ( /^From /gm ) { # /m makes ^ match next to \n
print "leading from in paragraph $.\n";
}
}
Here's code that finds everything between START and END in
a paragraph:
undef $/; # read in whole file, not just one line or paragraph
while ( <> ) {
while ( /START(.*?)END/sm ) { # /s makes . cross line boundaries
print "$1\n";
}
}
HHooww ccaann II ppuullll oouutt lliinneess bbeettwweeeenn ttwwoo ppaatttteerrnnss tthhaatt aarree
tthheemmsseellvveess oonn ddiiffffeerreenntt lliinneess??
You can use Perl's somewhat exotic .. operator (documented
in the perlop manpage):
perl -ne 'print if /START/ .. /END/' file1 file2 ...
If you wanted text and not lines, you would use
perl -0777 -pe 'print "$1\n" while /START(.*?)END/gs' file1 file2 ...
But if you want nested occurrences of START through END,
you'll run up against the problem described in the
question in this section on matching balanced text.
II ppuutt aa rreegguullaarr eexxpprreessssiioonn iinnttoo $$// bbuutt iitt ddiiddnn''tt wwoorrkk..
WWhhaatt''ss wwrroonngg??
$/ must be a string, not a regular expression. Awk has to
be better for something. :-)
Actually, you could do this if you don't mind reading the
whole file into memory:
undef $/;
@records = split /your_pattern/, <FH>;
The Net::Telnet module (available from CPAN) has the
capability to wait for a pattern in the input stream, or
timeout if it doesn't appear within a certain time.
## Create a file with three lines.
open FH, ">file";
print FH "The first line\nThe second line\nThe third line\n";
close FH;
## Get a read/write filehandle to it.
$fh = new FileHandle "+<file";
## Attach it to a "stream" object.
use Net::Telnet;
$file = new Net::Telnet (-fhopen => $fh);
## Search for the second line and print out the third.
$file->waitfor('/second line\n/');
print $file->getline;
HHooww ddoo II ssuubbssttiittuuttee ccaassee iinnsseennssiittiivveellyy oonn tthhee LLHHSS,, bbuutt
pprreesseerrvviinngg ccaassee oonn tthhee RRHHSS??
It depends on what you mean by "preserving case". The
following script makes the substitution have the same
case, letter by letter, as the original. If the
substitution has more characters than the string being
substituted, the case of the last character is used for
the rest of the substitution.
# Original by Nathan Torkington, massaged by Jeffrey Friedl
#
sub preserve_case($$)
{
my ($old, $new) = @_;
my ($state) = 0; # 0 = no change; 1 = lc; 2 = uc
my ($i, $oldlen, $newlen, $c) = (0, length($old), length($new));
my ($len) = $oldlen < $newlen ? $oldlen : $newlen;
for ($i = 0; $i < $len; $i++) {
if ($c = substr($old, $i, 1), $c =~ /[\W\d_]/) {
$state = 0;
} elsif (lc $c eq $c) {
substr($new, $i, 1) = lc(substr($new, $i, 1));
$state = 1;
} else {
substr($new, $i, 1) = uc(substr($new, $i, 1));
$state = 2;
}
}
# finish up with any remaining new (for when new is longer than old)
if ($newlen > $oldlen) {
if ($state == 1) {
substr($new, $oldlen) = lc(substr($new, $oldlen));
} elsif ($state == 2) {
substr($new, $oldlen) = uc(substr($new, $oldlen));
}
}
return $new;
}
$a = "this is a TEsT case";
$a =~ s/(test)/preserve_case($1, "success")/gie;
print "$a\n";
This prints:
this is a SUcCESS case
HHooww ccaann II mmaakkee \\ww match accented characters?
See the perllocale manpage.
HHooww ccaann II mmaattcchh aa llooccaallee--ssmmaarrtt vveerrssiioonn ooff //[[aa--zzAA--ZZ]]//?
One alphabetic character would be /[^\W\d_]/, no matter
what locale you're in. Non-alphabetics would be /[\W\d_]/
(assuming you don't consider an underscore a letter).
HHooww ccaann II qquuoottee aa vvaarriiaabbllee ttoo uussee iinn aa rreeggeexxpp??
The Perl parser will expand $variable and @variable
references in regular expressions unless the delimiter is
a single quote. Remember, too, that the right-hand side
of a s/// substitution is considered a double-quoted
string (see the perlop manpage for more details).
Remember also that any regexp special characters will be
acted on unless you precede the substitution with \Q.
Here's an example:
$string = "to die?";
$lhs = "die?";
$rhs = "sleep no more";
$string =~ s/\Q$lhs/$rhs/;
# $string is now "to sleep no more"
Without the \Q, the regexp would also spuriously match
"di".
WWhhaatt iiss //oo really for?
Using a variable in a regular expression match forces a
re-evaluation (and perhaps recompilation) each time
through. The /o modifier locks in the regexp the first
time it's used. This always happens in a constant regular
expression, and in fact, the pattern was compiled into the
internal format at the same time your entire program was.
Use of /o is irrelevant unless variable interpolation is
used in the pattern, and if so, the regexp engine will
neither know nor care whether the variables change after
the pattern is evaluated the very first time.
/o is often used to gain an extra measure of efficiency by
not performing subsequent evaluations when you know it
won't matter (because you know the variables won't
change), or more rarely, when you don't want the regexp to
notice if they do.
For example, here's a "paragrep" program:
$/ = ''; # paragraph mode
$pat = shift;
while (<>) {
print if /$pat/o;
}
HHooww ddoo II uussee aa rreegguullaarr eexxpprreessssiioonn ttoo ssttrriipp CC ssttyyllee
ccoommmmeennttss ffrroomm aa ffiillee??
While this actually can be done, it's much harder than
you'd think. For example, this one-liner
perl -0777 -pe 's{/\*.*?\*/}{}gs' foo.c
will work in many but not all cases. You see, it's too
simple-minded for certain kinds of C programs, in
particular, those with what appear to be comments in
quoted strings. For that, you'd need something like this,
created by Jeffrey Friedl:
$/ = undef;
$_ = <>;
s#/\*[^*]*\*+([^/*][^*]*\*+)*/|("(\\.|[^"\\])*"|'(\\.|[^'\\])*'|\n+|.[^/"'\\]*)#$2#g;
print;
This could, of course, be more legibly written with the /x
modifier, adding whitespace and comments.
CCaann II uussee PPeerrll rreegguullaarr eexxpprreessssiioonnss ttoo mmaattcchh bbaallaanncceedd tteexxtt??
Although Perl regular expressions are more powerful than
"mathematical" regular expressions, because they feature
conveniences like backreferences (\1 and its ilk), they
still aren't powerful enough. You still need to use non-
regexp techniques to parse balanced text, such as the text
enclosed between matching parentheses or braces, for
example.
An elaborate subroutine (for 7-bit ASCII only) to pull out
balanced and possibly nested single chars, like ` and ', {
and }, or ( and ) can be found in
http://www.perl.com/CPAN/authors/id/TOMC/scripts/pull_quotes.gz
.
The C::Scan module from CPAN contains such subs for
internal usage, but they are undocumented.
WWhhaatt ddooeess iitt mmeeaann tthhaatt rreeggeexxppss aarree ggrreeeeddyy?? HHooww ccaann II ggeett
aarroouunndd iitt??
Most people mean that greedy regexps match as much as they
can. Technically speaking, it's actually the quantifiers
(?, *, +, {}) that are greedy rather than the whole
pattern; Perl prefers local greed and immediate
gratification to overall greed. To get non-greedy
versions of the same quantifiers, use (??, *?, +?, {}?).
An example:
$s1 = $s2 = "I am very very cold";
$s1 =~ s/ve.*y //; # I am cold
$s2 =~ s/ve.*?y //; # I am very cold
Notice how the second substitution stopped matching as
soon as it encountered "y ". The *? quantifier
effectively tells the regular expression engine to find a
match as quickly as possible and pass control on to
whatever is next in line, like you would if you were
playing hot potato.
HHooww ddoo II pprroocceessss eeaacchh wwoorrdd oonn eeaacchh lliinnee??
Use the split function:
while (<>) {
foreach $word ( split ) {
# do something with $word here
}
}
Note that this isn't really a word in the English sense;
it's just chunks of consecutive non-whitespace characters.
To work with only alphanumeric sequences, you might
consider
while (<>) {
foreach $word (m/(\w+)/g) {
# do something with $word here
}
}
HHooww ccaann II pprriinntt oouutt aa wwoorrdd--ffrreeqquueennccyy oorr lliinnee--ffrreeqquueennccyy
ssuummmmaarryy??
To do this, you have to parse out each word in the input
stream. We'll pretend that by word you mean chunk of
alphabetics, hyphens, or apostrophes, rather than the non-
whitespace chunk idea of a word given in the previous
question:
while (<>) {
while ( /(\b[^\W_\d][\w'-]+\b)/g ) { # misses "`sheep'"
$seen{$1}++;
}
}
while ( ($word, $count) = each %seen ) {
print "$count $word\n";
}
If you wanted to do the same thing for lines, you wouldn't
need a regular expression:
while (<>) {
$seen{$_}++;
}
while ( ($line, $count) = each %seen ) {
print "$count $line";
}
If you want these output in a sorted order, see the
section on Hashes.
HHooww ccaann II ddoo aapppprrooxxiimmaattee mmaattcchhiinngg??
See the module String::Approx available from CPAN.
HHooww ddoo II eeffffiicciieennttllyy mmaattcchh mmaannyy rreegguullaarr eexxpprreessssiioonnss aatt
oonnccee??
The following is super-inefficient:
while (<FH>) {
foreach $pat (@patterns) {
if ( /$pat/ ) {
# do something
}
}
}
Instead, you either need to use one of the experimental
Regexp extension modules from CPAN (which might well be
overkill for your purposes), or else put together
something like this, inspired from a routine in Jeffrey
Friedl's book:
sub _bm_build {
my $condition = shift;
my @regexp = @_; # this MUST not be local(); need my()
my $expr = join $condition => map { "m/\$regexp[$_]/o" } (0..$#regexp);
my $match_func = eval "sub { $expr }";
die if $@; # propagate $@; this shouldn't happen!
return $match_func;
}
sub bm_and { _bm_build('&&', @_) }
sub bm_or { _bm_build('||', @_) }
$f1 = bm_and qw{
xterm
(?i)window
};
$f2 = bm_or qw{
\b[Ff]ree\b
\bBSD\B
(?i)sys(tem)?\s*[V5]\b
};
# feed me /etc/termcap, prolly
while ( <> ) {
print "1: $_" if &$f1;
print "2: $_" if &$f2;
}
WWhhyy ddoonn''tt wwoorrdd--bboouunnddaarryy sseeaarrcchheess wwiitthh \\bb work for me?
Two common misconceptions are that \b is a synonym for
\s+, and that it's the edge between whitespace characters
and non-whitespace characters. Neither is correct. \b is
the place between a \w character and a \W character (that
is, \b is the edge of a "word"). It's a zero-width
assertion, just like ^, $, and all the other anchors, so
it doesn't consume any characters. the perlre manpage
describes the behaviour of all the regexp metacharacters.
Here are examples of the incorrect application of \b, with
fixes:
"two words" =~ /(\w+)\b(\w+)/; # WRONG
"two words" =~ /(\w+)\s+(\w+)/; # right
" =matchless= text" =~ /\b=(\w+)=\b/; # WRONG
" =matchless= text" =~ /=(\w+)=/; # right
Although they may not do what you thought they did, \b and
\B can still be quite useful. For an example of the
correct use of \b, see the example of matching duplicate
words over multiple lines.
An example of using \B is the pattern \Bis\B. This will
find occurrences of "is" on the insides of words only, as
in "thistle", but not "this" or "island".
WWhhyy ddooeess uussiinngg $$&&,, $$``,, oorr $$'' ssllooww mmyy pprrooggrraamm ddoowwnn??
Because once Perl sees that you need one of these
variables anywhere in the program, it has to provide them
on each and every pattern match. The same mechanism that
handles these provides for the use of $1, $2, etc., so you
pay the same price for each regexp that contains capturing
parentheses. But if you never use $&, etc., in your
script, then regexps without capturing parentheses won't
be penalized. So avoid $&, $', and $` if you can, but if
you can't (and some algorithms really appreciate them),
once you've used them once, use them at will, because
you've already paid the price.
WWhhaatt ggoooodd iiss \\GG in a regular expression?
The notation \G is used in a match or substitution in
conjunction the /g modifier (and ignored if there's no /g)
to anchor the regular expression to the point just past
where the last match occurred, i.e. the pos() point.
For example, suppose you had a line of text quoted in
standard mail and Usenet notation, (that is, with leading
> characters), and you want change each leading > into a
corresponding :. You could do so in this way:
s/^(>+)/':' x length($1)/gem;
Or, using \G, the much simpler (and faster):
s/\G>/:/g;
A more sophisticated use might involve a tokenizer. The
following lex-like example is courtesy of Jeffrey Friedl.
It did not work in 5.003 due to bugs in that release, but
does work in 5.004 or better. (Note the use of /c, which
prevents a failed match with /g from resetting the search
position back to the beginning of the string.)
while (<>) {
chomp;
PARSER: {
m/ \G( \d+\b )/gcx && do { print "number: $1\n"; redo; };
m/ \G( \w+ )/gcx && do { print "word: $1\n"; redo; };
m/ \G( \s+ )/gcx && do { print "space: $1\n"; redo; };
m/ \G( [^\w\d]+ )/gcx && do { print "other: $1\n"; redo; };
}
}
Of course, that could have been written as
while (<>) {
chomp;
PARSER: {
if ( /\G( \d+\b )/gcx {
print "number: $1\n";
redo PARSER;
}
if ( /\G( \w+ )/gcx {
print "word: $1\n";
redo PARSER;
}
if ( /\G( \s+ )/gcx {
print "space: $1\n";
redo PARSER;
}
if ( /\G( [^\w\d]+ )/gcx {
print "other: $1\n";
redo PARSER;
}
}
}
But then you lose the vertical alignment of the regular
expressions.
AArree PPeerrll rreeggeexxppss DDFFAAss oorr NNFFAAss?? AArree tthheeyy PPOOSSIIXX ccoommpplliiaanntt??
While it's true that Perl's regular expressions resemble
the DFAs (deterministic finite automata) of the egrep(1)
program, they are in fact implemented as NFAs (non-
deterministic finite automata) to allow backtracking and
backreferencing. And they aren't POSIX-style either,
because those guarantee worst-case behavior for all cases.
(It seems that some people prefer guarantees of
consistency, even when what's guaranteed is slowness.)
See the book "Mastering Regular Expressions" (from
O'Reilly) by Jeffrey Friedl for all the details you could
ever hope to know on these matters (a full citation
appears in the perlfaq2 manpage).
WWhhaatt''ss wwrroonngg wwiitthh uussiinngg ggrreepp oorr mmaapp iinn aa vvooiidd ccoonntteexxtt??
Strictly speaking, nothing. Stylistically speaking, it's
not a good way to write maintainable code. That's because
you're using these constructs not for their return values
but rather for their side-effects, and side-effects can be
mystifying. There's no void grep() that's not better
written as a for (well, foreach, technically) loop.
HHooww ccaann II mmaattcchh ssttrriinnggss wwiitthh mmuullttiibbyyttee cchhaarraacctteerrss??
This is hard, and there's no good way. Perl does not
directly support wide characters. It pretends that a byte
and a character are synonymous. The following set of
approaches was offered by Jeffrey Friedl, whose article in
issue #5 of The Perl Journal talks about this very matter.
Let's suppose you have some weird Martian encoding where
pairs of ASCII uppercase letters encode single Martian
letters (i.e. the two bytes "CV" make a single Martian
letter, as do the two bytes "SG", "VS", "XX", etc.). Other
bytes represent single characters, just like ASCII.
So, the string of Martian "I am CVSGXX!" uses 12 bytes to
encode the nine characters 'I', ' ', 'a', 'm', ' ', 'CV',
'SG', 'XX', '!'.
Now, say you want to search for the single character /GX/.
Perl doesn't know about Martian, so it'll find the two
bytes "GX" in the "I am CVSGXX!" string, even though that
character isn't there: it just looks like it is because
"SG" is next to "XX", but there's no real "GX". This is a
big problem.
Here are a few ways, all painful, to deal with it:
$martian =~ s/([A-Z][A-Z])/ $1 /g; # Make sure adjacent ``martian'' bytes
# are no longer adjacent.
print "found GX!\n" if $martian =~ /GX/;
Or like this:
@chars = $martian =~ m/([A-Z][A-Z]|[^A-Z])/g;
# above is conceptually similar to: @chars = $text =~ m/(.)/g;
#
foreach $char (@chars) {
print "found GX!\n", last if $char eq 'GX';
}
Or like this:
while ($martian =~ m/\G([A-Z][A-Z]|.)/gs) { # \G probably unneeded
print "found GX!\n", last if $1 eq 'GX';
}
Or like this:
die "sorry, Perl doesn't (yet) have Martian support )-:\n";
In addition, a sample program which converts half-width to
full-width katakana (in Shift-JIS or EUC encoding) is
available from CPAN as
There are many double- (and multi-) byte encodings
commonly used these days. Some versions of these have 1-,
2-, 3-, and 4-byte characters, all mixed.
AAUUTTHHOORR AANNDD CCOOPPYYRRIIGGHHTT
Copyright (c) 1997 Tom Christiansen and Nathan Torkington.
All rights reserved. See the perlfaq manpage for
distribution information.