xref: /openbsd-src/gnu/usr.bin/perl/pod/perlre.pod (revision 3d61058aa5c692477b6d18acfbbdb653a9930ff9)
1=head1 NAME
2X<regular expression> X<regex> X<regexp>
3
4perlre - Perl regular expressions
5
6=head1 DESCRIPTION
7
8This page describes the syntax of regular expressions in Perl.
9
10If you haven't used regular expressions before, a tutorial introduction
11is available in L<perlretut>.  If you know just a little about them,
12a quick-start introduction is available in L<perlrequick>.
13
14Except for L</The Basics> section, this page assumes you are familiar
15with regular expression basics, like what is a "pattern", what does it
16look like, and how it is basically used.  For a reference on how they
17are used, plus various examples of the same, see discussions of C<m//>,
18C<s///>, C<qr//> and C<"??"> in L<perlop/"Regexp Quote-Like Operators">.
19
20New in v5.22, L<C<use re 'strict'>|re/'strict' mode> applies stricter
21rules than otherwise when compiling regular expression patterns.  It can
22find things that, while legal, may not be what you intended.
23
24=head2 The Basics
25X<regular expression, version 8> X<regex, version 8> X<regexp, version 8>
26
27Regular expressions are strings with the very particular syntax and
28meaning described in this document and auxiliary documents referred to
29by this one.  The strings are called "patterns".  Patterns are used to
30determine if some other string, called the "target", has (or doesn't
31have) the characteristics specified by the pattern.  We call this
32"matching" the target string against the pattern.  Usually the match is
33done by having the target be the first operand, and the pattern be the
34second operand, of one of the two binary operators C<=~> and C<!~>,
35listed in L<perlop/Binding Operators>; and the pattern will have been
36converted from an ordinary string by one of the operators in
37L<perlop/"Regexp Quote-Like Operators">, like so:
38
39 $foo =~ m/abc/
40
41This evaluates to true if and only if the string in the variable C<$foo>
42contains somewhere in it, the sequence of characters "a", "b", then "c".
43(The C<=~ m>, or match operator, is described in
44L<perlop/m/PATTERN/msixpodualngc>.)
45
46Patterns that aren't already stored in some variable must be delimited,
47at both ends, by delimiter characters.  These are often, as in the
48example above, forward slashes, and the typical way a pattern is written
49in documentation is with those slashes.  In most cases, the delimiter
50is the same character, fore and aft, but there are a few cases where a
51character looks like it has a mirror-image mate, where the opening
52version is the beginning delimiter, and the closing one is the ending
53delimiter, like
54
55 $foo =~ m<abc>
56
57Most times, the pattern is evaluated in double-quotish context, but it
58is possible to choose delimiters to force single-quotish, like
59
60 $foo =~ m'abc'
61
62If the pattern contains its delimiter within it, that delimiter must be
63escaped.  Prefixing it with a backslash (I<e.g.>, C<"/foo\/bar/">)
64serves this purpose.
65
66Any single character in a pattern matches that same character in the
67target string, unless the character is a I<metacharacter> with a special
68meaning described in this document.  A sequence of non-metacharacters
69matches the same sequence in the target string, as we saw above with
70C<m/abc/>.
71
72Only a few characters (all of them being ASCII punctuation characters)
73are metacharacters.  The most commonly used one is a dot C<".">, which
74normally matches almost any character (including a dot itself).
75
76You can cause characters that normally function as metacharacters to be
77interpreted literally by prefixing them with a C<"\">, just like the
78pattern's delimiter must be escaped if it also occurs within the
79pattern.  Thus, C<"\."> matches just a literal dot, C<"."> instead of
80its normal meaning.  This means that the backslash is also a
81metacharacter, so C<"\\"> matches a single C<"\">.  And a sequence that
82contains an escaped metacharacter matches the same sequence (but without
83the escape) in the target string.  So, the pattern C</blur\\fl/> would
84match any target string that contains the sequence C<"blur\fl">.
85
86The metacharacter C<"|"> is used to match one thing or another.  Thus
87
88 $foo =~ m/this|that/
89
90is TRUE if and only if C<$foo> contains either the sequence C<"this"> or
91the sequence C<"that">.  Like all metacharacters, prefixing the C<"|">
92with a backslash makes it match the plain punctuation character; in its
93case, the VERTICAL LINE.
94
95 $foo =~ m/this\|that/
96
97is TRUE if and only if C<$foo> contains the sequence C<"this|that">.
98
99You aren't limited to just a single C<"|">.
100
101 $foo =~ m/fee|fie|foe|fum/
102
103is TRUE if and only if C<$foo> contains any of those 4 sequences from
104the children's story "Jack and the Beanstalk".
105
106As you can see, the C<"|"> binds less tightly than a sequence of
107ordinary characters.  We can override this by using the grouping
108metacharacters, the parentheses C<"("> and C<")">.
109
110 $foo =~ m/th(is|at) thing/
111
112is TRUE if and only if C<$foo> contains either the sequence S<C<"this
113thing">> or the sequence S<C<"that thing">>.  The portions of the string
114that match the portions of the pattern enclosed in parentheses are
115normally made available separately for use later in the pattern,
116substitution, or program.  This is called "capturing", and it can get
117complicated.  See L</Capture groups>.
118
119The first alternative includes everything from the last pattern
120delimiter (C<"(">, C<"(?:"> (described later), I<etc>. or the beginning
121of the pattern) up to the first C<"|">, and the last alternative
122contains everything from the last C<"|"> to the next closing pattern
123delimiter.  That's why it's common practice to include alternatives in
124parentheses: to minimize confusion about where they start and end.
125
126Alternatives are tried from left to right, so the first
127alternative found for which the entire expression matches, is the one that
128is chosen. This means that alternatives are not necessarily greedy. For
129example: when matching C<foo|foot> against C<"barefoot">, only the C<"foo">
130part will match, as that is the first alternative tried, and it successfully
131matches the target string. (This might not seem important, but it is
132important when you are capturing matched text using parentheses.)
133
134Besides taking away the special meaning of a metacharacter, a prefixed
135backslash changes some letter and digit characters away from matching
136just themselves to instead have special meaning.  These are called
137"escape sequences", and all such are described in L<perlrebackslash>.  A
138backslash sequence (of a letter or digit) that doesn't currently have
139special meaning to Perl will raise a warning if warnings are enabled,
140as those are reserved for potential future use.
141
142One such sequence is C<\b>, which matches a boundary of some sort.
143C<\b{wb}> and a few others give specialized types of boundaries.
144(They are all described in detail starting at
145L<perlrebackslash/\b{}, \b, \B{}, \B>.)  Note that these don't match
146characters, but the zero-width spaces between characters.  They are an
147example of a L<zero-width assertion|/Assertions>.  Consider again,
148
149 $foo =~ m/fee|fie|foe|fum/
150
151It evaluates to TRUE if, besides those 4 words, any of the sequences
152"feed", "field", "Defoe", "fume", and many others are in C<$foo>.  By
153judicious use of C<\b> (or better (because it is designed to handle
154natural language) C<\b{wb}>), we can make sure that only the Giant's
155words are matched:
156
157 $foo =~ m/\b(fee|fie|foe|fum)\b/
158 $foo =~ m/\b{wb}(fee|fie|foe|fum)\b{wb}/
159
160The final example shows that the characters C<"{"> and C<"}"> are
161metacharacters.
162
163Another use for escape sequences is to specify characters that cannot
164(or which you prefer not to) be written literally.  These are described
165in detail in L<perlrebackslash/Character Escapes>, but the next three
166paragraphs briefly describe some of them.
167
168Various control characters can be written in C language style: C<"\n">
169matches a newline, C<"\t"> a tab, C<"\r"> a carriage return, C<"\f"> a
170form feed, I<etc>.
171
172More generally, C<\I<nnn>>, where I<nnn> is a string of three octal
173digits, matches the character whose native code point is I<nnn>.  You
174can easily run into trouble if you don't have exactly three digits.  So
175always use three, or since Perl 5.14, you can use C<\o{...}> to specify
176any number of octal digits.
177
178Similarly, C<\xI<nn>>, where I<nn> are hexadecimal digits, matches the
179character whose native ordinal is I<nn>.  Again, not using exactly two
180digits is a recipe for disaster, but you can use C<\x{...}> to specify
181any number of hex digits.
182
183Besides being a metacharacter, the C<"."> is an example of a "character
184class", something that can match any single character of a given set of
185them.  In its case, the set is just about all possible characters.  Perl
186predefines several character classes besides the C<".">; there is a
187separate reference page about just these, L<perlrecharclass>.
188
189You can define your own custom character classes, by putting into your
190pattern in the appropriate place(s), a list of all the characters you
191want in the set.  You do this by enclosing the list within C<[]> bracket
192characters.  These are called "bracketed character classes" when we are
193being precise, but often the word "bracketed" is dropped.  (Dropping it
194usually doesn't cause confusion.)  This means that the C<"["> character
195is another metacharacter.  It doesn't match anything just by itself; it
196is used only to tell Perl that what follows it is a bracketed character
197class.  If you want to match a literal left square bracket, you must
198escape it, like C<"\[">.  The matching C<"]"> is also a metacharacter;
199again it doesn't match anything by itself, but just marks the end of
200your custom class to Perl.  It is an example of a "sometimes
201metacharacter".  It isn't a metacharacter if there is no corresponding
202C<"[">, and matches its literal self:
203
204 print "]" =~ /]/;  # prints 1
205
206The list of characters within the character class gives the set of
207characters matched by the class.  C<"[abc]"> matches a single "a" or "b"
208or "c".  But if the first character after the C<"["> is C<"^">, the
209class instead matches any character not in the list.  Within a list, the
210C<"-"> character specifies a range of characters, so that C<a-z>
211represents all characters between "a" and "z", inclusive.  If you want
212either C<"-"> or C<"]"> itself to be a member of a class, put it at the
213start of the list (possibly after a C<"^">), or escape it with a
214backslash.  C<"-"> is also taken literally when it is at the end of the
215list, just before the closing C<"]">.  (The following all specify the
216same class of three characters: C<[-az]>, C<[az-]>, and C<[a\-z]>.  All
217are different from C<[a-z]>, which specifies a class containing
218twenty-six characters, even on EBCDIC-based character sets.)
219
220There is lots more to bracketed character classes; full details are in
221L<perlrecharclass/Bracketed Character Classes>.
222
223=head3 Metacharacters
224X<metacharacter>
225X<\> X<^> X<.> X<$> X<|> X<(> X<()> X<[> X<[]>
226
227L</The Basics> introduced some of the metacharacters.  This section
228gives them all.  Most of them have the same meaning as in the I<egrep>
229command.
230
231Only the C<"\"> is always a metacharacter.  The others are metacharacters
232just sometimes.  The following tables lists all of them, summarizes
233their use, and gives the contexts where they are metacharacters.
234Outside those contexts or if prefixed by a C<"\">, they match their
235corresponding punctuation character.  In some cases, their meaning
236varies depending on various pattern modifiers that alter the default
237behaviors.  See L</Modifiers>.
238
239
240            PURPOSE                                  WHERE
241 \   Escape the next character                    Always, except when
242                                                  escaped by another \
243 ^   Match the beginning of the string            Not in []
244       (or line, if /m is used)
245 ^   Complement the [] class                      At the beginning of []
246 .   Match any single character except newline    Not in []
247       (under /s, includes newline)
248 $   Match the end of the string                  Not in [], but can
249       (or before newline at the end of the       mean interpolate a
250       string; or before any newline if /m is     scalar
251       used)
252 |   Alternation                                  Not in []
253 ()  Grouping                                     Not in []
254 [   Start Bracketed Character class              Not in []
255 ]   End Bracketed Character class                Only in [], and
256                                                    not first
257 *   Matches the preceding element 0 or more      Not in []
258       times
259 +   Matches the preceding element 1 or more      Not in []
260       times
261 ?   Matches the preceding element 0 or 1         Not in []
262       times
263 {   Starts a sequence that gives number(s)       Not in []
264       of times the preceding element can be
265       matched
266 {   when following certain escape sequences
267       starts a modifier to the meaning of the
268       sequence
269 }   End sequence started by {
270 -   Indicates a range                            Only in [] interior
271 #   Beginning of comment, extends to line end    Only with /x modifier
272
273Notice that most of the metacharacters lose their special meaning when
274they occur in a bracketed character class, except C<"^"> has a different
275meaning when it is at the beginning of such a class.  And C<"-"> and C<"]">
276are metacharacters only at restricted positions within bracketed
277character classes; while C<"}"> is a metacharacter only when closing a
278special construct started by C<"{">.
279
280In double-quotish context, as is usually the case,  you need to be
281careful about C<"$"> and the non-metacharacter C<"@">.  Those could
282interpolate variables, which may or may not be what you intended.
283
284These rules were designed for compactness of expression, rather than
285legibility and maintainability.  The L</E<sol>x and E<sol>xx> pattern
286modifiers allow you to insert white space to improve readability.  And
287use of S<C<L<re 'strict'|re/'strict' mode>>> adds extra checking to
288catch some typos that might silently compile into something unintended.
289
290By default, the C<"^"> character is guaranteed to match only the
291beginning of the string, the C<"$"> character only the end (or before the
292newline at the end), and Perl does certain optimizations with the
293assumption that the string contains only one line.  Embedded newlines
294will not be matched by C<"^"> or C<"$">.  You may, however, wish to treat a
295string as a multi-line buffer, such that the C<"^"> will match after any
296newline within the string (except if the newline is the last character in
297the string), and C<"$"> will match before any newline.  At the
298cost of a little more overhead, you can do this by using the
299C<L</E<sol>m>> modifier on the pattern match operator.  (Older programs
300did this by setting C<$*>, but this option was removed in perl 5.10.)
301X<^> X<$> X</m>
302
303To simplify multi-line substitutions, the C<"."> character never matches a
304newline unless you use the L<C<E<sol>s>|/s> modifier, which in effect tells
305Perl to pretend the string is a single line--even if it isn't.
306X<.> X</s>
307
308=head2 Modifiers
309
310=head3 Overview
311
312The default behavior for matching can be changed, using various
313modifiers.  Modifiers that relate to the interpretation of the pattern
314are listed just below.  Modifiers that alter the way a pattern is used
315by Perl are detailed in L<perlop/"Regexp Quote-Like Operators"> and
316L<perlop/"Gory details of parsing quoted constructs">.  Modifiers can be added
317dynamically; see L</Extended Patterns> below.
318
319=over 4
320
321=item B<C<m>>
322X</m> X<regex, multiline> X<regexp, multiline> X<regular expression, multiline>
323
324Treat the string being matched against as multiple lines.  That is, change C<"^"> and C<"$"> from matching
325the start of the string's first line and the end of its last line to
326matching the start and end of each line within the string.
327
328=item B<C<s>>
329X</s> X<regex, single-line> X<regexp, single-line>
330X<regular expression, single-line>
331
332Treat the string as single line.  That is, change C<"."> to match any character
333whatsoever, even a newline, which normally it would not match.
334
335Used together, as C</ms>, they let the C<"."> match any character whatsoever,
336while still allowing C<"^"> and C<"$"> to match, respectively, just after
337and just before newlines within the string.
338
339=item B<C<i>>
340X</i> X<regex, case-insensitive> X<regexp, case-insensitive>
341X<regular expression, case-insensitive>
342
343Do case-insensitive pattern matching.  For example, "A" will match "a"
344under C</i>.
345
346If locale matching rules are in effect, the case map is taken from the
347current
348locale for code points less than 255, and from Unicode rules for larger
349code points.  However, matches that would cross the Unicode
350rules/non-Unicode rules boundary (ords 255/256) will not succeed, unless
351the locale is a UTF-8 one.  See L<perllocale>.
352
353There are a number of Unicode characters that match a sequence of
354multiple characters under C</i>.  For example,
355C<LATIN SMALL LIGATURE FI> should match the sequence C<fi>.  Perl is not
356currently able to do this when the multiple characters are in the pattern and
357are split between groupings, or when one or more are quantified.  Thus
358
359 "\N{LATIN SMALL LIGATURE FI}" =~ /fi/i;          # Matches
360 "\N{LATIN SMALL LIGATURE FI}" =~ /[fi][fi]/i;    # Doesn't match!
361 "\N{LATIN SMALL LIGATURE FI}" =~ /fi*/i;         # Doesn't match!
362
363 # The below doesn't match, and it isn't clear what $1 and $2 would
364 # be even if it did!!
365 "\N{LATIN SMALL LIGATURE FI}" =~ /(f)(i)/i;      # Doesn't match!
366
367Perl doesn't match multiple characters in a bracketed
368character class unless the character that maps to them is explicitly
369mentioned, and it doesn't match them at all if the character class is
370inverted, which otherwise could be highly confusing.  See
371L<perlrecharclass/Bracketed Character Classes>, and
372L<perlrecharclass/Negation>.
373
374=item B<C<x>> and B<C<xx>>
375X</x>
376
377Extend your pattern's legibility by permitting whitespace and comments.
378Details in L</E<sol>x and  E<sol>xx>
379
380=item B<C<p>>
381X</p> X<regex, preserve> X<regexp, preserve>
382
383Preserve the string matched such that C<${^PREMATCH}>, C<${^MATCH}>, and
384C<${^POSTMATCH}> are available for use after matching.
385
386In Perl 5.20 and higher this is ignored. Due to a new copy-on-write
387mechanism, C<${^PREMATCH}>, C<${^MATCH}>, and C<${^POSTMATCH}> will be available
388after the match regardless of the modifier.
389
390=item B<C<a>>, B<C<d>>, B<C<l>>, and B<C<u>>
391X</a> X</d> X</l> X</u>
392
393These modifiers, all new in 5.14, affect which character-set rules
394(Unicode, I<etc>.) are used, as described below in
395L</Character set modifiers>.
396
397=item B<C<n>>
398X</n> X<regex, non-capture> X<regexp, non-capture>
399X<regular expression, non-capture>
400
401Prevent the grouping metacharacters C<()> from capturing. This modifier,
402new in 5.22, will stop C<$1>, C<$2>, I<etc>... from being filled in.
403
404  "hello" =~ /(hi|hello)/;   # $1 is "hello"
405  "hello" =~ /(hi|hello)/n;  # $1 is undef
406
407This is equivalent to putting C<?:> at the beginning of every capturing group:
408
409  "hello" =~ /(?:hi|hello)/; # $1 is undef
410
411C</n> can be negated on a per-group basis. Alternatively, named captures
412may still be used.
413
414  "hello" =~ /(?-n:(hi|hello))/n;   # $1 is "hello"
415  "hello" =~ /(?<greet>hi|hello)/n; # $1 is "hello", $+{greet} is
416                                    # "hello"
417
418=item Other Modifiers
419
420There are a number of flags that can be found at the end of regular
421expression constructs that are I<not> generic regular expression flags, but
422apply to the operation being performed, like matching or substitution (C<m//>
423or C<s///> respectively).
424
425Flags described further in
426L<perlretut/"Using regular expressions in Perl"> are:
427
428  c  - keep the current position during repeated matching
429  g  - globally match the pattern repeatedly in the string
430
431Substitution-specific modifiers described in
432L<perlop/"s/PATTERN/REPLACEMENT/msixpodualngcer"> are:
433
434  e  - evaluate the right-hand side as an expression
435  ee - evaluate the right side as a string then eval the result
436  o  - pretend to optimize your code, but actually introduce bugs
437  r  - perform non-destructive substitution and return the new value
438
439=back
440
441Regular expression modifiers are usually written in documentation
442as I<e.g.>, "the C</x> modifier", even though the delimiter
443in question might not really be a slash.  The modifiers C</imnsxadlup>
444may also be embedded within the regular expression itself using
445the C<(?...)> construct, see L</Extended Patterns> below.
446
447=head3 Details on some modifiers
448
449Some of the modifiers require more explanation than given in the
450L</Overview> above.
451
452=head4 C</x> and  C</xx>
453
454A single C</x> tells
455the regular expression parser to ignore most whitespace that is neither
456backslashed nor within a bracketed character class, nor within the characters
457of a multi-character metapattern like C<(?i: ... )>.  You can use this to
458break up your regular expression into more readable parts.
459Also, the C<"#"> character is treated as a metacharacter introducing a
460comment that runs up to the pattern's closing delimiter, or to the end
461of the current line if the pattern extends onto the next line.  Hence,
462this is very much like an ordinary Perl code comment.  (You can include
463the closing delimiter within the comment only if you precede it with a
464backslash, so be careful!)
465
466Use of C</x> means that if you want real
467whitespace or C<"#"> characters in the pattern (outside a bracketed character
468class, which is unaffected by C</x>), then you'll either have to
469escape them (using backslashes or C<\Q...\E>) or encode them using octal,
470hex, or C<\N{}> or C<\p{name=...}> escapes.
471It is ineffective to try to continue a comment onto the next line by
472escaping the C<\n> with a backslash or C<\Q>.
473
474You can use L</(?#text)> to create a comment that ends earlier than the
475end of the current line, but C<text> also can't contain the closing
476delimiter unless escaped with a backslash.
477
478A common pitfall is to forget that C<"#"> characters (outside a
479bracketed character class) begin a comment under C</x> and are not
480matched literally.  Just keep that in mind when trying to puzzle out why
481a particular C</x> pattern isn't working as expected.
482Inside a bracketed character class, C<"#"> retains its non-special,
483literal meaning.
484
485Starting in Perl v5.26, if the modifier has a second C<"x"> within it,
486the effect of a single C</x> is increased.  The only difference is that
487inside bracketed character classes, non-escaped (by a backslash) SPACE
488and TAB characters are not added to the class, and hence can be inserted
489to make the classes more readable:
490
491    / [d-e g-i 3-7]/xx
492    /[ ! @ " # $ % ^ & * () = ? <> ' ]/xx
493
494may be easier to grasp than the squashed equivalents
495
496    /[d-eg-i3-7]/
497    /[!@"#$%^&*()=?<>']/
498
499Note that this unfortunately doesn't mean that your bracketed classes
500can contain comments or extend over multiple lines.  A C<#> inside a
501character class is still just a literal C<#>, and doesn't introduce a
502comment.  And, unless the closing bracket is on the same line as the
503opening one, the newline character (and everything on the next line(s)
504until terminated by a C<]> will be part of the class, just as if you'd
505written C<\n>.
506
507Taken together, these features go a long way towards
508making Perl's regular expressions more readable.  Here's an example:
509
510    # Delete (most) C comments.
511    $program =~ s {
512	/\*	# Match the opening delimiter.
513	.*?	# Match a minimal number of characters.
514	\*/	# Match the closing delimiter.
515    } []gsx;
516
517Note that anything inside
518a C<\Q...\E> stays unaffected by C</x>.  And note that C</x> doesn't affect
519space interpretation within a single multi-character construct.  For
520example C<(?:...)> can't have a space between the C<"(">,
521C<"?">, and C<":">.  Within any delimiters for such a construct, allowed
522spaces are not affected by C</x>, and depend on the construct.  For
523example, all constructs using curly braces as delimiters, such as
524C<\x{...}> can have blanks within but adjacent to the braces, but not
525elsewhere, and no non-blank space characters.  An exception are Unicode
526properties which follow Unicode rules, for which see
527L<perluniprops/Properties accessible through \p{} and \P{}>.
528X</x>
529
530The set of characters that are deemed whitespace are those that Unicode
531calls "Pattern White Space", namely:
532
533 U+0009 CHARACTER TABULATION
534 U+000A LINE FEED
535 U+000B LINE TABULATION
536 U+000C FORM FEED
537 U+000D CARRIAGE RETURN
538 U+0020 SPACE
539 U+0085 NEXT LINE
540 U+200E LEFT-TO-RIGHT MARK
541 U+200F RIGHT-TO-LEFT MARK
542 U+2028 LINE SEPARATOR
543 U+2029 PARAGRAPH SEPARATOR
544
545=head4 Character set modifiers
546
547C</d>, C</u>, C</a>, and C</l>, available starting in 5.14, are called
548the character set modifiers; they affect the character set rules
549used for the regular expression.
550
551The C</d>, C</u>, and C</l> modifiers are not likely to be of much use
552to you, and so you need not worry about them very much.  They exist for
553Perl's internal use, so that complex regular expression data structures
554can be automatically serialized and later exactly reconstituted,
555including all their nuances.  But, since Perl can't keep a secret, and
556there may be rare instances where they are useful, they are documented
557here.
558
559The C</a> modifier, on the other hand, may be useful.  Its purpose is to
560allow code that is to work mostly on ASCII data to not have to concern
561itself with Unicode.
562
563Briefly, C</l> sets the character set to that of whatever B<L>ocale is in
564effect at the time of the execution of the pattern match.
565
566C</u> sets the character set to B<U>nicode.
567
568C</a> also sets the character set to Unicode, BUT adds several
569restrictions for B<A>SCII-safe matching.
570
571C</d> is the old, problematic, pre-5.14 B<D>efault character set
572behavior.  Its only use is to force that old behavior.
573
574At any given time, exactly one of these modifiers is in effect.  Their
575existence allows Perl to keep the originally compiled behavior of a
576regular expression, regardless of what rules are in effect when it is
577actually executed.  And if it is interpolated into a larger regex, the
578original's rules continue to apply to it, and don't affect the other
579parts.
580
581The C</l> and C</u> modifiers are automatically selected for
582regular expressions compiled within the scope of various pragmas,
583and we recommend that in general, you use those pragmas instead of
584specifying these modifiers explicitly.  For one thing, the modifiers
585affect only pattern matching, and do not extend to even any replacement
586done, whereas using the pragmas gives consistent results for all
587appropriate operations within their scopes.  For example,
588
589 s/foo/\Ubar/il
590
591will match "foo" using the locale's rules for case-insensitive matching,
592but the C</l> does not affect how the C<\U> operates.  Most likely you
593want both of them to use locale rules.  To do this, instead compile the
594regular expression within the scope of C<use locale>.  This both
595implicitly adds the C</l>, and applies locale rules to the C<\U>.   The
596lesson is to C<use locale>, and not C</l> explicitly.
597
598Similarly, it would be better to use C<use feature 'unicode_strings'>
599instead of,
600
601 s/foo/\Lbar/iu
602
603to get Unicode rules, as the C<\L> in the former (but not necessarily
604the latter) would also use Unicode rules.
605
606More detail on each of the modifiers follows.  Most likely you don't
607need to know this detail for C</l>, C</u>, and C</d>, and can skip ahead
608to L<E<sol>a|/E<sol>a (and E<sol>aa)>.
609
610=head4 /l
611
612means to use the current locale's rules (see L<perllocale>) when pattern
613matching.  For example, C<\w> will match the "word" characters of that
614locale, and C<"/i"> case-insensitive matching will match according to
615the locale's case folding rules.  The locale used will be the one in
616effect at the time of execution of the pattern match.  This may not be
617the same as the compilation-time locale, and can differ from one match
618to another if there is an intervening call of the
619L<setlocale() function|perllocale/The setlocale function>.
620
621Prior to v5.20, Perl did not support multi-byte locales.  Starting then,
622UTF-8 locales are supported.  No other multi byte locales are ever
623likely to be supported.  However, in all locales, one can have code
624points above 255 and these will always be treated as Unicode no matter
625what locale is in effect.
626
627Under Unicode rules, there are a few case-insensitive matches that cross
628the 255/256 boundary.  Except for UTF-8 locales in Perls v5.20 and
629later, these are disallowed under C</l>.  For example, 0xFF (on ASCII
630platforms) does not caselessly match the character at 0x178, C<LATIN
631CAPITAL LETTER Y WITH DIAERESIS>, because 0xFF may not be C<LATIN SMALL
632LETTER Y WITH DIAERESIS> in the current locale, and Perl has no way of
633knowing if that character even exists in the locale, much less what code
634point it is.
635
636In a UTF-8 locale in v5.20 and later, the only visible difference
637between locale and non-locale in regular expressions should be tainting,
638if your perl supports taint checking (see L<perlsec>).
639
640This modifier may be specified to be the default by C<use locale>, but
641see L</Which character set modifier is in effect?>.
642X</l>
643
644=head4 /u
645
646means to use Unicode rules when pattern matching.  On ASCII platforms,
647this means that the code points between 128 and 255 take on their
648Latin-1 (ISO-8859-1) meanings (which are the same as Unicode's).
649(Otherwise Perl considers their meanings to be undefined.)  Thus,
650under this modifier, the ASCII platform effectively becomes a Unicode
651platform; and hence, for example, C<\w> will match any of the more than
652100_000 word characters in Unicode.
653
654Unlike most locales, which are specific to a language and country pair,
655Unicode classifies all the characters that are letters I<somewhere> in
656the world as
657C<\w>.  For example, your locale might not think that C<LATIN SMALL
658LETTER ETH> is a letter (unless you happen to speak Icelandic), but
659Unicode does.  Similarly, all the characters that are decimal digits
660somewhere in the world will match C<\d>; this is hundreds, not 10,
661possible matches.  And some of those digits look like some of the 10
662ASCII digits, but mean a different number, so a human could easily think
663a number is a different quantity than it really is.  For example,
664C<BENGALI DIGIT FOUR> (U+09EA) looks very much like an
665C<ASCII DIGIT EIGHT> (U+0038), and C<LEPCHA DIGIT SIX> (U+1C46) looks
666very much like an C<ASCII DIGIT FIVE> (U+0035).  And, C<\d+>, may match
667strings of digits that are a mixture from different writing systems,
668creating a security issue.  A fraudulent website, for example, could
669display the price of something using U+1C46, and it would appear to the
670user that something cost 500 units, but it really costs 600.  A browser
671that enforced script runs (L</Script Runs>) would prevent that
672fraudulent display.  L<Unicode::UCD/num()> can also be used to sort this
673out.  Or the C</a> modifier can be used to force C<\d> to match just the
674ASCII 0 through 9.
675
676Also, under this modifier, case-insensitive matching works on the full
677set of Unicode
678characters.  The C<KELVIN SIGN>, for example matches the letters "k" and
679"K"; and C<LATIN SMALL LIGATURE FF> matches the sequence "ff", which,
680if you're not prepared, might make it look like a hexadecimal constant,
681presenting another potential security issue.  See
682L<https://unicode.org/reports/tr36> for a detailed discussion of Unicode
683security issues.
684
685This modifier may be specified to be the default by C<use feature
686'unicode_strings>, C<use locale ':not_characters'>, or
687C<L<use v5.12|perlfunc/use VERSION>> (or higher),
688but see L</Which character set modifier is in effect?>.
689X</u>
690
691=head4 /d
692
693B<IMPORTANT:> Because of the unpredictable behaviors this
694modifier causes, only use it to maintain weird backward compatibilities.
695Use the
696L<< C<unicode_strings>|feature/"The 'unicode_strings' feature" >>
697feature
698in new code to avoid inadvertently enabling this modifier by default.
699
700What does this modifier do? It "Depends"!
701
702This modifier means to use platform-native matching rules
703except when there is cause to use Unicode rules instead, as follows:
704
705=over 4
706
707=item 1
708
709the target string's L<UTF8 flag|perlunifaq/What is "the UTF8 flag"?>
710(see below) is set; or
711
712=item 2
713
714the pattern's L<UTF8 flag|perlunifaq/What is "the UTF8 flag"?>
715(see below) is set; or
716
717=item 3
718
719the pattern explicitly mentions a code point that is above 255 (say by
720C<\x{100}>); or
721
722=item 4
723
724the pattern uses a Unicode name (C<\N{...}>);  or
725
726=item 5
727
728the pattern uses a Unicode property (C<\p{...}> or C<\P{...}>); or
729
730=item 6
731
732the pattern uses a Unicode break (C<\b{...}> or C<\B{...}>); or
733
734=item 7
735
736the pattern uses C<L</(?[ ])>>
737
738=item 8
739
740the pattern uses L<C<(*script_run: ...)>|/Script Runs>
741
742=back
743
744Regarding the "UTF8 flag" references above: normally Perl applications
745shouldn't think about that flag. It's part of Perl's internals,
746so it can change whenever Perl wants. C</d> may thus cause unpredictable
747results. See L<perlunicode/The "Unicode Bug">. This bug
748has become rather infamous, leading to yet other (without swearing) names
749for this modifier like "Dicey" and "Dodgy".
750
751Here are some examples of how that works on an ASCII platform:
752
753 $str =  "\xDF";        #
754 utf8::downgrade($str); # $str is not UTF8-flagged.
755 $str =~ /^\w/;         # No match, since no UTF8 flag.
756
757 $str .= "\x{0e0b}";    # Now $str is UTF8-flagged.
758 $str =~ /^\w/;         # Match! $str is now UTF8-flagged.
759 chop $str;
760 $str =~ /^\w/;         # Still a match! $str retains its UTF8 flag.
761
762Under Perl's default configuration this modifier is automatically
763selected by default when none of the others are, so yet another name
764for it (unfortunately) is "Default".
765
766Whenever you can, use the
767L<< C<unicode_strings>|feature/"The 'unicode_strings' feature" >>
768to cause X</u> to be the default instead.
769
770=head4 /a (and /aa)
771
772This modifier stands for ASCII-restrict (or ASCII-safe).  This modifier
773may be doubled-up to increase its effect.
774
775When it appears singly, it causes the sequences C<\d>, C<\s>, C<\w>, and
776the Posix character classes to match only in the ASCII range.  They thus
777revert to their pre-5.6, pre-Unicode meanings.  Under C</a>,  C<\d>
778always means precisely the digits C<"0"> to C<"9">; C<\s> means the five
779characters C<[ \f\n\r\t]>, and starting in Perl v5.18, the vertical tab;
780C<\w> means the 63 characters
781C<[A-Za-z0-9_]>; and likewise, all the Posix classes such as
782C<[[:print:]]> match only the appropriate ASCII-range characters.
783
784This modifier is useful for people who only incidentally use Unicode,
785and who do not wish to be burdened with its complexities and security
786concerns.
787
788With C</a>, one can write C<\d> with confidence that it will only match
789ASCII characters, and should the need arise to match beyond ASCII, you
790can instead use C<\p{Digit}> (or C<\p{Word}> for C<\w>).  There are
791similar C<\p{...}> constructs that can match beyond ASCII both white
792space (see L<perlrecharclass/Whitespace>), and Posix classes (see
793L<perlrecharclass/POSIX Character Classes>).  Thus, this modifier
794doesn't mean you can't use Unicode, it means that to get Unicode
795matching you must explicitly use a construct (C<\p{}>, C<\P{}>) that
796signals Unicode.
797
798As you would expect, this modifier causes, for example, C<\D> to mean
799the same thing as C<[^0-9]>; in fact, all non-ASCII characters match
800C<\D>, C<\S>, and C<\W>.  C<\b> still means to match at the boundary
801between C<\w> and C<\W>, using the C</a> definitions of them (similarly
802for C<\B>).
803
804Otherwise, C</a> behaves like the C</u> modifier, in that
805case-insensitive matching uses Unicode rules; for example, "k" will
806match the Unicode C<\N{KELVIN SIGN}> under C</i> matching, and code
807points in the Latin1 range, above ASCII will have Unicode rules when it
808comes to case-insensitive matching.
809
810To forbid ASCII/non-ASCII matches (like "k" with C<\N{KELVIN SIGN}>),
811specify the C<"a"> twice, for example C</aai> or C</aia>.  (The first
812occurrence of C<"a"> restricts the C<\d>, I<etc>., and the second occurrence
813adds the C</i> restrictions.)  But, note that code points outside the
814ASCII range will use Unicode rules for C</i> matching, so the modifier
815doesn't really restrict things to just ASCII; it just forbids the
816intermixing of ASCII and non-ASCII.
817
818To summarize, this modifier provides protection for applications that
819don't wish to be exposed to all of Unicode.  Specifying it twice
820gives added protection.
821
822This modifier may be specified to be the default by C<use re '/a'>
823or C<use re '/aa'>.  If you do so, you may actually have occasion to use
824the C</u> modifier explicitly if there are a few regular expressions
825where you do want full Unicode rules (but even here, it's best if
826everything were under feature C<"unicode_strings">, along with the
827C<use re '/aa'>).  Also see L</Which character set modifier is in
828effect?>.
829X</a>
830X</aa>
831
832=head4 Which character set modifier is in effect?
833
834Which of these modifiers is in effect at any given point in a regular
835expression depends on a fairly complex set of interactions.  These have
836been designed so that in general you don't have to worry about it, but
837this section gives the gory details.  As
838explained below in L</Extended Patterns> it is possible to explicitly
839specify modifiers that apply only to portions of a regular expression.
840The innermost always has priority over any outer ones, and one applying
841to the whole expression has priority over any of the default settings that are
842described in the remainder of this section.
843
844The C<L<use re 'E<sol>foo'|re/"'/flags' mode">> pragma can be used to set
845default modifiers (including these) for regular expressions compiled
846within its scope.  This pragma has precedence over the other pragmas
847listed below that also change the defaults. Note that the /x modifier does
848NOT affect C<split STR> patterns.
849
850Otherwise, C<L<use locale|perllocale>> sets the default modifier to C</l>;
851and C<L<use feature 'unicode_strings|feature>>, or
852C<L<use v5.12|perlfunc/use VERSION>> (or higher) set the default to
853C</u> when not in the same scope as either C<L<use locale|perllocale>>
854or C<L<use bytes|bytes>>.
855(C<L<use locale ':not_characters'|perllocale/Unicode and UTF-8>> also
856sets the default to C</u>, overriding any plain C<use locale>.)
857Unlike the mechanisms mentioned above, these
858affect operations besides regular expressions pattern matching, and so
859give more consistent results with other operators, including using
860C<\U>, C<\l>, I<etc>. in substitution replacements.
861
862If none of the above apply, for backwards compatibility reasons, the
863C</d> modifier is the one in effect by default.  As this can lead to
864unexpected results, it is best to specify which other rule set should be
865used.
866
867=head4 Character set modifier behavior prior to Perl 5.14
868
869Prior to 5.14, there were no explicit modifiers, but C</l> was implied
870for regexes compiled within the scope of C<use locale>, and C</d> was
871implied otherwise.  However, interpolating a regex into a larger regex
872would ignore the original compilation in favor of whatever was in effect
873at the time of the second compilation.  There were a number of
874inconsistencies (bugs) with the C</d> modifier, where Unicode rules
875would be used when inappropriate, and vice versa.  C<\p{}> did not imply
876Unicode rules, and neither did all occurrences of C<\N{}>, until 5.12.
877
878=head2 Regular Expressions
879
880=head3 Quantifiers
881
882Quantifiers are used when a particular portion of a pattern needs to
883match a certain number (or numbers) of times.  If there isn't a
884quantifier the number of times to match is exactly one.  The following
885standard quantifiers are recognized:
886X<metacharacter> X<quantifier> X<*> X<+> X<?> X<{n}> X<{n,}> X<{n,m}>
887
888    *           Match 0 or more times
889    +           Match 1 or more times
890    ?           Match 1 or 0 times
891    {n}         Match exactly n times
892    {n,}        Match at least n times
893    {,n}        Match at most n times
894    {n,m}       Match at least n but not more than m times
895
896(If a non-escaped curly bracket occurs in a context other than one of
897the quantifiers listed above, where it does not form part of a
898backslashed sequence like C<\x{...}>, it is either a fatal syntax error,
899or treated as a regular character, generally with a deprecation warning
900raised.  To escape it, you can precede it with a backslash (C<"\{">) or
901enclose it within square brackets  (C<"[{]">).
902This change will allow for future syntax extensions (like making the
903lower bound of a quantifier optional), and better error checking of
904quantifiers).
905
906The C<"*"> quantifier is equivalent to C<{0,}>, the C<"+">
907quantifier to C<{1,}>, and the C<"?"> quantifier to C<{0,1}>.  I<n> and I<m> are limited
908to non-negative integral values less than a preset limit defined when perl is built.
909This is usually 65534 on the most common platforms.  The actual limit can
910be seen in the error message generated by code such as this:
911
912    $_ **= $_ , / {$_} / for 2 .. 42;
913
914By default, a quantified subpattern is "greedy", that is, it will match as
915many times as possible (given a particular starting location) while still
916allowing the rest of the pattern to match.  If you want it to match the
917minimum number of times possible, follow the quantifier with a C<"?">.  Note
918that the meanings don't change, just the "greediness":
919X<metacharacter> X<greedy> X<greediness>
920X<?> X<*?> X<+?> X<??> X<{n}?> X<{n,}?> X<{,n}?> X<{n,m}?>
921
922    *?        Match 0 or more times, not greedily
923    +?        Match 1 or more times, not greedily
924    ??        Match 0 or 1 time, not greedily
925    {n}?      Match exactly n times, not greedily (redundant)
926    {n,}?     Match at least n times, not greedily
927    {,n}?     Match at most n times, not greedily
928    {n,m}?    Match at least n but not more than m times, not greedily
929
930Normally when a quantified subpattern does not allow the rest of the
931overall pattern to match, Perl will backtrack. However, this behaviour is
932sometimes undesirable. Thus Perl provides the "possessive" quantifier form
933as well.
934
935 *+     Match 0 or more times and give nothing back
936 ++     Match 1 or more times and give nothing back
937 ?+     Match 0 or 1 time and give nothing back
938 {n}+   Match exactly n times and give nothing back (redundant)
939 {n,}+  Match at least n times and give nothing back
940 {,n}+  Match at most n times and give nothing back
941 {n,m}+ Match at least n but not more than m times and give nothing back
942
943For instance,
944
945   'aaaa' =~ /a++a/
946
947will never match, as the C<a++> will gobble up all the C<"a">'s in the
948string and won't leave any for the remaining part of the pattern. This
949feature can be extremely useful to give perl hints about where it
950shouldn't backtrack. For instance, the typical "match a double-quoted
951string" problem can be most efficiently performed when written as:
952
953   /"(?:[^"\\]++|\\.)*+"/
954
955as we know that if the final quote does not match, backtracking will not
956help. See the independent subexpression
957C<L</(?E<gt>I<pattern>)>> for more details;
958possessive quantifiers are just syntactic sugar for that construct. For
959instance the above example could also be written as follows:
960
961   /"(?>(?:(?>[^"\\]+)|\\.)*)"/
962
963Note that the possessive quantifier modifier can not be combined
964with the non-greedy modifier. This is because it would make no sense.
965Consider the follow equivalency table:
966
967    Illegal         Legal
968    ------------    ------
969    X??+            X{0}
970    X+?+            X{1}
971    X{min,max}?+    X{min}
972
973=head3 Escape sequences
974
975Because patterns are processed as double-quoted strings, the following
976also work:
977
978 \t          tab                   (HT, TAB)
979 \n          newline               (LF, NL)
980 \r          return                (CR)
981 \f          form feed             (FF)
982 \a          alarm (bell)          (BEL)
983 \e          escape (think troff)  (ESC)
984 \cK         control char          (example: VT)
985 \x{}, \x00  character whose ordinal is the given hexadecimal number
986 \N{name}    named Unicode character or character sequence
987 \N{U+263D}  Unicode character     (example: FIRST QUARTER MOON)
988 \o{}, \000  character whose ordinal is the given octal number
989 \l          lowercase next char (think vi)
990 \u          uppercase next char (think vi)
991 \L          lowercase until \E (think vi)
992 \U          uppercase until \E (think vi)
993 \Q          quote (disable) pattern metacharacters until \E
994 \E          end either case modification or quoted section, think vi
995
996Details are in L<perlop/Quote and Quote-like Operators>.
997
998=head3 Character Classes and other Special Escapes
999
1000In addition, Perl defines the following:
1001X<\g> X<\k> X<\K> X<backreference>
1002
1003 Sequence   Note    Description
1004  [...]     [1]  Match a character according to the rules of the
1005                   bracketed character class defined by the "...".
1006                   Example: [a-z] matches "a" or "b" or "c" ... or "z"
1007  [[:...:]] [2]  Match a character according to the rules of the POSIX
1008                   character class "..." within the outer bracketed
1009                   character class.  Example: [[:upper:]] matches any
1010                   uppercase character.
1011  (?[...])  [8]  Extended bracketed character class
1012  \w        [3]  Match a "word" character (alphanumeric plus "_", plus
1013                   other connector punctuation chars plus Unicode
1014                   marks)
1015  \W        [3]  Match a non-"word" character
1016  \s        [3]  Match a whitespace character
1017  \S        [3]  Match a non-whitespace character
1018  \d        [3]  Match a decimal digit character
1019  \D        [3]  Match a non-digit character
1020  \pP       [3]  Match P, named property.  Use \p{Prop} for longer names
1021  \PP       [3]  Match non-P
1022  \X        [4]  Match Unicode "eXtended grapheme cluster"
1023  \1        [5]  Backreference to a specific capture group or buffer.
1024                   '1' may actually be any positive integer.
1025  \g1       [5]  Backreference to a specific or previous group,
1026  \g{-1}    [5]  The number may be negative indicating a relative
1027                   previous group and may optionally be wrapped in
1028                   curly brackets for safer parsing.
1029  \g{name}  [5]  Named backreference
1030  \k<name>  [5]  Named backreference
1031  \k'name'  [5]  Named backreference
1032  \k{name}  [5]  Named backreference
1033  \K        [6]  Keep the stuff left of the \K, don't include it in $&
1034  \N        [7]  Any character but \n.  Not affected by /s modifier
1035  \v        [3]  Vertical whitespace
1036  \V        [3]  Not vertical whitespace
1037  \h        [3]  Horizontal whitespace
1038  \H        [3]  Not horizontal whitespace
1039  \R        [4]  Linebreak
1040
1041=over 4
1042
1043=item [1]
1044
1045See L<perlrecharclass/Bracketed Character Classes> for details.
1046
1047=item [2]
1048
1049See L<perlrecharclass/POSIX Character Classes> for details.
1050
1051=item [3]
1052
1053See L<perlunicode/Unicode Character Properties> for details
1054
1055=item [4]
1056
1057See L<perlrebackslash/Misc> for details.
1058
1059=item [5]
1060
1061See L</Capture groups> below for details.
1062
1063=item [6]
1064
1065See L</Extended Patterns> below for details.
1066
1067=item [7]
1068
1069Note that C<\N> has two meanings.  When of the form C<\N{I<NAME>}>, it
1070matches the character or character sequence whose name is I<NAME>; and
1071similarly
1072when of the form C<\N{U+I<hex>}>, it matches the character whose Unicode
1073code point is I<hex>.  Otherwise it matches any character but C<\n>.
1074
1075=item [8]
1076
1077See L<perlrecharclass/Extended Bracketed Character Classes> for details.
1078
1079=back
1080
1081=head3 Assertions
1082
1083Besides L<C<"^"> and C<"$">|/Metacharacters>, Perl defines the following
1084zero-width assertions:
1085X<zero-width assertion> X<assertion> X<regex, zero-width assertion>
1086X<regexp, zero-width assertion>
1087X<regular expression, zero-width assertion>
1088X<\b> X<\B> X<\A> X<\Z> X<\z> X<\G>
1089
1090 \b{}   Match at Unicode boundary of specified type
1091 \B{}   Match where corresponding \b{} doesn't match
1092 \b     Match a \w\W or \W\w boundary
1093 \B     Match except at a \w\W or \W\w boundary
1094 \A     Match only at beginning of string
1095 \Z     Match only at end of string, or before newline at the end
1096 \z     Match only at end of string
1097 \G     Match only at pos() (e.g. at the end-of-match position
1098        of prior m//g)
1099
1100A Unicode boundary (C<\b{}>), available starting in v5.22, is a spot
1101between two characters, or before the first character in the string, or
1102after the final character in the string where certain criteria defined
1103by Unicode are met.  See L<perlrebackslash/\b{}, \b, \B{}, \B> for
1104details.
1105
1106A word boundary (C<\b>) is a spot between two characters
1107that has a C<\w> on one side of it and a C<\W> on the other side
1108of it (in either order), counting the imaginary characters off the
1109beginning and end of the string as matching a C<\W>.  (Within
1110character classes C<\b> represents backspace rather than a word
1111boundary, just as it normally does in any double-quoted string.)
1112The C<\A> and C<\Z> are just like C<"^"> and C<"$">, except that they
1113won't match multiple times when the C</m> modifier is used, while
1114C<"^"> and C<"$"> will match at every internal line boundary.  To match
1115the actual end of the string and not ignore an optional trailing
1116newline, use C<\z>.
1117X<\b> X<\A> X<\Z> X<\z> X</m>
1118
1119The C<\G> assertion can be used to chain global matches (using
1120C<m//g>), as described in L<perlop/"Regexp Quote-Like Operators">.
1121It is also useful when writing C<lex>-like scanners, when you have
1122several patterns that you want to match against consequent substrings
1123of your string; see the previous reference.  The actual location
1124where C<\G> will match can also be influenced by using C<pos()> as
1125an lvalue: see L<perlfunc/pos>. Note that the rule for zero-length
1126matches (see L</"Repeated Patterns Matching a Zero-length Substring">)
1127is modified somewhat, in that contents to the left of C<\G> are
1128not counted when determining the length of the match. Thus the following
1129will not match forever:
1130X<\G>
1131
1132     my $string = 'ABC';
1133     pos($string) = 1;
1134     while ($string =~ /(.\G)/g) {
1135         print $1;
1136     }
1137
1138It will print 'A' and then terminate, as it considers the match to
1139be zero-width, and thus will not match at the same position twice in a
1140row.
1141
1142It is worth noting that C<\G> improperly used can result in an infinite
1143loop. Take care when using patterns that include C<\G> in an alternation.
1144
1145Note also that C<s///> will refuse to overwrite part of a substitution
1146that has already been replaced; so for example this will stop after the
1147first iteration, rather than iterating its way backwards through the
1148string:
1149
1150    $_ = "123456789";
1151    pos = 6;
1152    s/.(?=.\G)/X/g;
1153    print; 	# prints 1234X6789, not XXXXX6789
1154
1155
1156=head3 Capture groups
1157
1158The grouping construct C<( ... )> creates capture groups (also referred to as
1159capture buffers). To refer to the current contents of a group later on, within
1160the same pattern, use C<\g1> (or C<\g{1}>) for the first, C<\g2> (or C<\g{2}>)
1161for the second, and so on.
1162This is called a I<backreference>.
1163X<regex, capture buffer> X<regexp, capture buffer>
1164X<regex, capture group> X<regexp, capture group>
1165X<regular expression, capture buffer> X<backreference>
1166X<regular expression, capture group> X<backreference>
1167X<\g{1}> X<\g{-1}> X<\g{name}> X<relative backreference> X<named backreference>
1168X<named capture buffer> X<regular expression, named capture buffer>
1169X<named capture group> X<regular expression, named capture group>
1170X<%+> X<$+{name}> X<< \k<name> >>
1171There is no limit to the number of captured substrings that you may use.
1172Groups are numbered with the leftmost open parenthesis being number 1, I<etc>.  If
1173a group did not match, the associated backreference won't match either. (This
1174can happen if the group is optional, or in a different branch of an
1175alternation.)
1176You can omit the C<"g">, and write C<"\1">, I<etc>, but there are some issues with
1177this form, described below.
1178
1179You can also refer to capture groups relatively, by using a negative number, so
1180that C<\g-1> and C<\g{-1}> both refer to the immediately preceding capture
1181group, and C<\g-2> and C<\g{-2}> both refer to the group before it.  For
1182example:
1183
1184        /
1185         (Y)            # group 1
1186         (              # group 2
1187            (X)         # group 3
1188            \g{-1}      # backref to group 3
1189            \g{-3}      # backref to group 1
1190         )
1191        /x
1192
1193would match the same as C</(Y) ( (X) \g3 \g1 )/x>.  This allows you to
1194interpolate regexes into larger regexes and not have to worry about the
1195capture groups being renumbered.
1196
1197You can dispense with numbers altogether and create named capture groups.
1198The notation is C<(?E<lt>I<name>E<gt>...)> to declare and C<\g{I<name>}> to
1199reference.  (To be compatible with .Net regular expressions, C<\g{I<name>}> may
1200also be written as C<\k{I<name>}>, C<\kE<lt>I<name>E<gt>> or C<\k'I<name>'>.)
1201I<name> must not begin with a number, nor contain hyphens.
1202When different groups within the same pattern have the same name, any reference
1203to that name assumes the leftmost defined group.  Named groups count in
1204absolute and relative numbering, and so can also be referred to by those
1205numbers.
1206(It's possible to do things with named capture groups that would otherwise
1207require C<(??{})>.)
1208
1209Capture group contents are dynamically scoped and available to you outside the
1210pattern until the end of the enclosing block or until the next successful
1211match in the same scope, whichever comes first.
1212See L<perlsyn/"Compound Statements"> and
1213L<perlvar/"Scoping Rules of Regex Variables"> for more details.
1214
1215You can access the contents of a capture group by absolute number (using
1216C<"$1"> instead of C<"\g1">, I<etc>); or by name via the C<%+> hash,
1217using C<"$+{I<name>}">.
1218
1219Braces are required in referring to named capture groups, but are optional for
1220absolute or relative numbered ones.  Braces are safer when creating a regex by
1221concatenating smaller strings.  For example if you have C<qr/$x$y/>, and C<$x>
1222contained C<"\g1">, and C<$y> contained C<"37">, you would get C</\g137/> which
1223is probably not what you intended.
1224
1225If you use braces, you may also optionally add any number of blank
1226(space or tab) characters within but adjacent to the braces, like
1227S<C<\g{ -1 }>>, or S<C<\k{ I<name> }>>.
1228
1229The C<\g> and C<\k> notations were introduced in Perl 5.10.0.  Prior to that
1230there were no named nor relative numbered capture groups.  Absolute numbered
1231groups were referred to using C<\1>,
1232C<\2>, I<etc>., and this notation is still
1233accepted (and likely always will be).  But it leads to some ambiguities if
1234there are more than 9 capture groups, as C<\10> could mean either the tenth
1235capture group, or the character whose ordinal in octal is 010 (a backspace in
1236ASCII).  Perl resolves this ambiguity by interpreting C<\10> as a backreference
1237only if at least 10 left parentheses have opened before it.  Likewise C<\11> is
1238a backreference only if at least 11 left parentheses have opened before it.
1239And so on.  C<\1> through C<\9> are always interpreted as backreferences.
1240There are several examples below that illustrate these perils.  You can avoid
1241the ambiguity by always using C<\g{}> or C<\g> if you mean capturing groups;
1242and for octal constants always using C<\o{}>, or for C<\077> and below, using 3
1243digits padded with leading zeros, since a leading zero implies an octal
1244constant.
1245
1246The C<\I<digit>> notation also works in certain circumstances outside
1247the pattern.  See L</Warning on \1 Instead of $1> below for details.
1248
1249Examples:
1250
1251    s/^([^ ]*) *([^ ]*)/$2 $1/;     # swap first two words
1252
1253    /(.)\g1/                        # find first doubled char
1254         and print "'$1' is the first doubled character\n";
1255
1256    /(?<char>.)\k<char>/            # ... a different way
1257         and print "'$+{char}' is the first doubled character\n";
1258
1259    /(?'char'.)\g1/                 # ... mix and match
1260         and print "'$1' is the first doubled character\n";
1261
1262    if (/Time: (..):(..):(..)/) {   # parse out values
1263        $hours = $1;
1264        $minutes = $2;
1265        $seconds = $3;
1266    }
1267
1268    /(.)(.)(.)(.)(.)(.)(.)(.)(.)\g10/   # \g10 is a backreference
1269    /(.)(.)(.)(.)(.)(.)(.)(.)(.)\10/    # \10 is octal
1270    /((.)(.)(.)(.)(.)(.)(.)(.)(.))\10/  # \10 is a backreference
1271    /((.)(.)(.)(.)(.)(.)(.)(.)(.))\010/ # \010 is octal
1272
1273    $x = '(.)\1';        # Creates problems when concatenated.
1274    $y = '(.)\g{1}';     # Avoids the problems.
1275    "aa" =~ /${x}/;      # True
1276    "aa" =~ /${y}/;      # True
1277    "aa0" =~ /${x}0/;    # False!
1278    "aa0" =~ /${y}0/;    # True
1279    "aa\x08" =~ /${x}0/;  # True!
1280    "aa\x08" =~ /${y}0/;  # False
1281
1282Several special variables also refer back to portions of the previous
1283match.  C<$+> returns whatever the last bracket match matched.
1284C<$&> returns the entire matched string.  (At one point C<$0> did
1285also, but now it returns the name of the program.)  C<$`> returns
1286everything before the matched string.  C<$'> returns everything
1287after the matched string. And C<$^N> contains whatever was matched by
1288the most-recently closed group (submatch). C<$^N> can be used in
1289extended patterns (see below), for example to assign a submatch to a
1290variable.
1291X<$+> X<$^N> X<$&> X<$`> X<$'>
1292
1293These special variables, like the C<%+> hash and the numbered match variables
1294(C<$1>, C<$2>, C<$3>, I<etc>.) are dynamically scoped
1295until the end of the enclosing block or until the next successful
1296match, whichever comes first.  (See L<perlsyn/"Compound Statements">.)
1297X<$+> X<$^N> X<$&> X<$`> X<$'>
1298X<$1> X<$2> X<$3> X<$4> X<$5> X<$6> X<$7> X<$8> X<$9>
1299X<@{^CAPTURE}>
1300
1301The C<@{^CAPTURE}> array may be used to access ALL of the capture buffers
1302as an array without needing to know how many there are. For instance
1303
1304    $string=~/$pattern/ and @captured = @{^CAPTURE};
1305
1306will place a copy of each capture variable, C<$1>, C<$2> etc, into the
1307C<@captured> array.
1308
1309Be aware that when interpolating a subscript of the C<@{^CAPTURE}>
1310array you must use demarcated curly brace notation:
1311
1312    print "${^CAPTURE[0]}";
1313
1314See L<perldata/"Demarcated variable names using braces"> for more on
1315this notation.
1316
1317B<NOTE>: Failed matches in Perl do not reset the match variables,
1318which makes it easier to write code that tests for a series of more
1319specific cases and remembers the best match.
1320
1321B<WARNING>: If your code is to run on Perl 5.16 or earlier,
1322beware that once Perl sees that you need one of C<$&>, C<$`>, or
1323C<$'> anywhere in the program, it has to provide them for every
1324pattern match.  This may substantially slow your program.
1325
1326Perl uses the same mechanism to produce C<$1>, C<$2>, I<etc>, so you also
1327pay a price for each pattern that contains capturing parentheses.
1328(To avoid this cost while retaining the grouping behaviour, use the
1329extended regular expression C<(?: ... )> instead.)  But if you never
1330use C<$&>, C<$`> or C<$'>, then patterns I<without> capturing
1331parentheses will not be penalized.  So avoid C<$&>, C<$'>, and C<$`>
1332if you can, but if you can't (and some algorithms really appreciate
1333them), once you've used them once, use them at will, because you've
1334already paid the price.
1335X<$&> X<$`> X<$'>
1336
1337Perl 5.16 introduced a slightly more efficient mechanism that notes
1338separately whether each of C<$`>, C<$&>, and C<$'> have been seen, and
1339thus may only need to copy part of the string.  Perl 5.20 introduced a
1340much more efficient copy-on-write mechanism which eliminates any slowdown.
1341
1342As another workaround for this problem, Perl 5.10.0 introduced C<${^PREMATCH}>,
1343C<${^MATCH}> and C<${^POSTMATCH}>, which are equivalent to C<$`>, C<$&>
1344and C<$'>, B<except> that they are only guaranteed to be defined after a
1345successful match that was executed with the C</p> (preserve) modifier.
1346The use of these variables incurs no global performance penalty, unlike
1347their punctuation character equivalents, however at the trade-off that you
1348have to tell perl when you want to use them.
1349X</p> X<p modifier>
1350
1351=head2 Quoting metacharacters
1352
1353Backslashed metacharacters in Perl are alphanumeric, such as C<\b>,
1354C<\w>, C<\n>.  Unlike some other regular expression languages, there
1355are no backslashed symbols that aren't alphanumeric.  So anything
1356that looks like C<\\>, C<\(>, C<\)>, C<\[>, C<\]>, C<\{>, or C<\}> is
1357always
1358interpreted as a literal character, not a metacharacter.  This was
1359once used in a common idiom to disable or quote the special meanings
1360of regular expression metacharacters in a string that you want to
1361use for a pattern. Simply quote all non-"word" characters:
1362
1363    $pattern =~ s/(\W)/\\$1/g;
1364
1365(If C<use locale> is set, then this depends on the current locale.)
1366Today it is more common to use the C<L<quotemeta()|perlfunc/quotemeta>>
1367function or the C<\Q> metaquoting escape sequence to disable all
1368metacharacters' special meanings like this:
1369
1370    /$unquoted\Q$quoted\E$unquoted/
1371
1372Beware that if you put literal backslashes (those not inside
1373interpolated variables) between C<\Q> and C<\E>, double-quotish
1374backslash interpolation may lead to confusing results.  If you
1375I<need> to use literal backslashes within C<\Q...\E>,
1376consult L<perlop/"Gory details of parsing quoted constructs">.
1377
1378C<quotemeta()> and C<\Q> are fully described in L<perlfunc/quotemeta>.
1379
1380=head2 Extended Patterns
1381
1382Perl also defines a consistent extension syntax for features not
1383found in standard tools like B<awk> and
1384B<lex>.  The syntax for most of these is a
1385pair of parentheses with a question mark as the first thing within
1386the parentheses.  The character after the question mark indicates
1387the extension.
1388
1389A question mark was chosen for this and for the minimal-matching
1390construct because 1) question marks are rare in older regular
1391expressions, and 2) whenever you see one, you should stop and
1392"question" exactly what is going on.  That's psychology....
1393
1394=over 4
1395
1396=item C<(?#I<text>)>
1397X<(?#)>
1398
1399A comment.  The I<text> is ignored.
1400Note that Perl closes
1401the comment as soon as it sees a C<")">, so there is no way to put a literal
1402C<")"> in the comment.  The pattern's closing delimiter must be escaped by
1403a backslash if it appears in the comment.
1404
1405See L</E<sol>x> for another way to have comments in patterns.
1406
1407Note that a comment can go just about anywhere, except in the middle of
1408an escape sequence.   Examples:
1409
1410 qr/foo(?#comment)bar/'  # Matches 'foobar'
1411
1412 # The pattern below matches 'abcd', 'abccd', or 'abcccd'
1413 qr/abc(?#comment between literal and its quantifier){1,3}d/
1414
1415 # The pattern below generates a syntax error, because the '\p' must
1416 # be followed immediately by a '{'.
1417 qr/\p(?#comment between \p and its property name){Any}/
1418
1419 # The pattern below generates a syntax error, because the initial
1420 # '\(' is a literal opening parenthesis, and so there is nothing
1421 # for the  closing ')' to match
1422 qr/\(?#the backslash means this isn't a comment)p{Any}/
1423
1424 # Comments can be used to fold long patterns into multiple lines
1425 qr/First part of a long regex(?#
1426   )remaining part/
1427
1428=item C<(?adlupimnsx-imnsx)>
1429
1430=item C<(?^alupimnsx)>
1431X<(?)> X<(?^)>
1432
1433Zero or more embedded pattern-match modifiers, to be turned on (or
1434turned off if preceded by C<"-">) for the remainder of the pattern or
1435the remainder of the enclosing pattern group (if any).
1436
1437This is particularly useful for dynamically-generated patterns,
1438such as those read in from a
1439configuration file, taken from an argument, or specified in a table
1440somewhere.  Consider the case where some patterns want to be
1441case-sensitive and some do not:  The case-insensitive ones merely need to
1442include C<(?i)> at the front of the pattern.  For example:
1443
1444    $pattern = "foobar";
1445    if ( /$pattern/i ) { }
1446
1447    # more flexible:
1448
1449    $pattern = "(?i)foobar";
1450    if ( /$pattern/ ) { }
1451
1452These modifiers are restored at the end of the enclosing group. For example,
1453
1454    ( (?i) blah ) \s+ \g1
1455
1456will match C<blah> in any case, some spaces, and an exact (I<including the case>!)
1457repetition of the previous word, assuming the C</x> modifier, and no C</i>
1458modifier outside this group.
1459
1460These modifiers do not carry over into named subpatterns called in the
1461enclosing group. In other words, a pattern such as C<((?i)(?&I<NAME>))> does not
1462change the case-sensitivity of the I<NAME> pattern.
1463
1464A modifier is overridden by later occurrences of this construct in the
1465same scope containing the same modifier, so that
1466
1467    /((?im)foo(?-m)bar)/
1468
1469matches all of C<foobar> case insensitively, but uses C</m> rules for
1470only the C<foo> portion.  The C<"a"> flag overrides C<aa> as well;
1471likewise C<aa> overrides C<"a">.  The same goes for C<"x"> and C<xx>.
1472Hence, in
1473
1474    /(?-x)foo/xx
1475
1476both C</x> and C</xx> are turned off during matching C<foo>.  And in
1477
1478    /(?x)foo/x
1479
1480C</x> but NOT C</xx> is turned on for matching C<foo>.  (One might
1481mistakenly think that since the inner C<(?x)> is already in the scope of
1482C</x>, that the result would effectively be the sum of them, yielding
1483C</xx>.  It doesn't work that way.)  Similarly, doing something like
1484C<(?xx-x)foo> turns off all C<"x"> behavior for matching C<foo>, it is not
1485that you subtract 1 C<"x"> from 2 to get 1 C<"x"> remaining.
1486
1487Any of these modifiers can be set to apply globally to all regular
1488expressions compiled within the scope of a C<use re>.  See
1489L<re/"'/flags' mode">.
1490
1491Starting in Perl 5.14, a C<"^"> (caret or circumflex accent) immediately
1492after the C<"?"> is a shorthand equivalent to C<d-imnsx>.  Flags (except
1493C<"d">) may follow the caret to override it.
1494But a minus sign is not legal with it.
1495
1496Note that the C<"a">, C<"d">, C<"l">, C<"p">, and C<"u"> modifiers are special in
1497that they can only be enabled, not disabled, and the C<"a">, C<"d">, C<"l">, and
1498C<"u"> modifiers are mutually exclusive: specifying one de-specifies the
1499others, and a maximum of one (or two C<"a">'s) may appear in the
1500construct.  Thus, for
1501example, C<(?-p)> will warn when compiled under C<use warnings>;
1502C<(?-d:...)> and C<(?dl:...)> are fatal errors.
1503
1504Note also that the C<"p"> modifier is special in that its presence
1505anywhere in a pattern has a global effect.
1506
1507Having zero modifiers makes this a no-op (so why did you specify it,
1508unless it's generated code), and starting in v5.30, warns under L<C<use
1509re 'strict'>|re/'strict' mode>.
1510
1511=item C<(?:I<pattern>)>
1512X<(?:)>
1513
1514=item C<(?adluimnsx-imnsx:I<pattern>)>
1515
1516=item C<(?^aluimnsx:I<pattern>)>
1517X<(?^:)>
1518
1519This is for clustering, not capturing; it groups subexpressions like
1520C<"()">, but doesn't make backreferences as C<"()"> does.  So
1521
1522    @fields = split(/\b(?:a|b|c)\b/)
1523
1524matches the same field delimiters as
1525
1526    @fields = split(/\b(a|b|c)\b/)
1527
1528but doesn't spit out the delimiters themselves as extra fields (even though
1529that's the behaviour of L<perlfunc/split> when its pattern contains capturing
1530groups).  It's also cheaper not to capture
1531characters if you don't need to.
1532
1533Any letters between C<"?"> and C<":"> act as flags modifiers as with
1534C<(?adluimnsx-imnsx)>.  For example,
1535
1536    /(?s-i:more.*than).*million/i
1537
1538is equivalent to the more verbose
1539
1540    /(?:(?s-i)more.*than).*million/i
1541
1542Note that any C<()> constructs enclosed within this one will still
1543capture unless the C</n> modifier is in effect.
1544
1545Like the L</(?adlupimnsx-imnsx)> construct, C<aa> and C<"a"> override each
1546other, as do C<xx> and C<"x">.  They are not additive.  So, doing
1547something like C<(?xx-x:foo)> turns off all C<"x"> behavior for matching
1548C<foo>.
1549
1550Starting in Perl 5.14, a C<"^"> (caret or circumflex accent) immediately
1551after the C<"?"> is a shorthand equivalent to C<d-imnsx>.  Any positive
1552flags (except C<"d">) may follow the caret, so
1553
1554    (?^x:foo)
1555
1556is equivalent to
1557
1558    (?x-imns:foo)
1559
1560The caret tells Perl that this cluster doesn't inherit the flags of any
1561surrounding pattern, but uses the system defaults (C<d-imnsx>),
1562modified by any flags specified.
1563
1564The caret allows for simpler stringification of compiled regular
1565expressions.  These look like
1566
1567    (?^:pattern)
1568
1569with any non-default flags appearing between the caret and the colon.
1570A test that looks at such stringification thus doesn't need to have the
1571system default flags hard-coded in it, just the caret.  If new flags are
1572added to Perl, the meaning of the caret's expansion will change to include
1573the default for those flags, so the test will still work, unchanged.
1574
1575Specifying a negative flag after the caret is an error, as the flag is
1576redundant.
1577
1578Mnemonic for C<(?^...)>:  A fresh beginning since the usual use of a caret is
1579to match at the beginning.
1580
1581=item C<(?|I<pattern>)>
1582X<(?|)> X<Branch reset>
1583
1584This is the "branch reset" pattern, which has the special property
1585that the capture groups are numbered from the same starting point
1586in each alternation branch. It is available starting from perl 5.10.0.
1587
1588Capture groups are numbered from left to right, but inside this
1589construct the numbering is restarted for each branch.
1590
1591The numbering within each branch will be as normal, and any groups
1592following this construct will be numbered as though the construct
1593contained only one branch, that being the one with the most capture
1594groups in it.
1595
1596This construct is useful when you want to capture one of a
1597number of alternative matches.
1598
1599Consider the following pattern.  The numbers underneath show in
1600which group the captured content will be stored.
1601
1602
1603    # before  ---------------branch-reset----------- after
1604    / ( a )  (?| x ( y ) z | (p (q) r) | (t) u (v) ) ( z ) /x
1605    # 1            2         2  3        2     3     4
1606
1607Be careful when using the branch reset pattern in combination with
1608named captures. Named captures are implemented as being aliases to
1609numbered groups holding the captures, and that interferes with the
1610implementation of the branch reset pattern. If you are using named
1611captures in a branch reset pattern, it's best to use the same names,
1612in the same order, in each of the alternations:
1613
1614   /(?|  (?<a> x ) (?<b> y )
1615      |  (?<a> z ) (?<b> w )) /x
1616
1617Not doing so may lead to surprises:
1618
1619  "12" =~ /(?| (?<a> \d+ ) | (?<b> \D+))/x;
1620  say $+{a};    # Prints '12'
1621  say $+{b};    # *Also* prints '12'.
1622
1623The problem here is that both the group named C<< a >> and the group
1624named C<< b >> are aliases for the group belonging to C<< $1 >>.
1625
1626=item Lookaround Assertions
1627X<look-around assertion> X<lookaround assertion> X<look-around> X<lookaround>
1628
1629Lookaround assertions are zero-width patterns which match a specific
1630pattern without including it in C<$&>. Positive assertions match when
1631their subpattern matches, negative assertions match when their subpattern
1632fails. Lookbehind matches text up to the current match position,
1633lookahead matches text following the current match position.
1634
1635=over 4
1636
1637=item C<(?=I<pattern>)>
1638
1639=item C<(*pla:I<pattern>)>
1640
1641=item C<(*positive_lookahead:I<pattern>)>
1642X<(?=)>
1643X<(*pla>
1644X<(*positive_lookahead>
1645X<look-ahead, positive> X<lookahead, positive>
1646
1647A zero-width positive lookahead assertion.  For example, C</\w+(?=\t)/>
1648matches a word followed by a tab, without including the tab in C<$&>.
1649
1650=item C<(?!I<pattern>)>
1651
1652=item C<(*nla:I<pattern>)>
1653
1654=item C<(*negative_lookahead:I<pattern>)>
1655X<(?!)>
1656X<(*nla>
1657X<(*negative_lookahead>
1658X<look-ahead, negative> X<lookahead, negative>
1659
1660A zero-width negative lookahead assertion.  For example C</foo(?!bar)/>
1661matches any occurrence of "foo" that isn't followed by "bar".  Note
1662however that lookahead and lookbehind are NOT the same thing.  You cannot
1663use this for lookbehind.
1664
1665If you are looking for a "bar" that isn't preceded by a "foo", C</(?!foo)bar/>
1666will not do what you want.  That's because the C<(?!foo)> is just saying that
1667the next thing cannot be "foo"--and it's not, it's a "bar", so "foobar" will
1668match.  Use lookbehind instead (see below).
1669
1670=item C<(?<=I<pattern>)>
1671
1672=item C<\K>
1673
1674=item C<(*plb:I<pattern>)>
1675
1676=item C<(*positive_lookbehind:I<pattern>)>
1677X<(?<=)>
1678X<(*plb>
1679X<(*positive_lookbehind>
1680X<look-behind, positive> X<lookbehind, positive> X<\K>
1681
1682A zero-width positive lookbehind assertion.  For example, C</(?<=\t)\w+/>
1683matches a word that follows a tab, without including the tab in C<$&>.
1684
1685Prior to Perl 5.30, it worked only for fixed-width lookbehind, but
1686starting in that release, it can handle variable lengths from 1 to 255
1687characters as an experimental feature.  The feature is enabled
1688automatically if you use a variable length positive lookbehind assertion.
1689
1690In Perl 5.35.10 the scope of the experimental nature of this construct
1691has been reduced, and experimental warnings will only be produced when
1692the construct contains capturing parenthesis. The warnings will be
1693raised at pattern compilation time, unless turned off, in the
1694C<experimental::vlb> category.  This is to warn you that the exact
1695contents of capturing buffers in a variable length positive lookbehind
1696is not well defined and is subject to change in a future release of perl.
1697
1698Currently if you use capture buffers inside of a positive variable length
1699lookbehind the result will be the longest and thus leftmost match possible.
1700This means that
1701
1702    "aax" =~ /(?=x)(?<=(a|aa))/
1703    "aax" =~ /(?=x)(?<=(aa|a))/
1704    "aax" =~ /(?=x)(?<=(a{1,2}?)/
1705    "aax" =~ /(?=x)(?<=(a{1,2})/
1706
1707will all result in C<$1> containing C<"aa">. It is possible in a future
1708release of perl we will change this behavior.
1709
1710There is a special form of this construct, called C<\K>
1711(available since Perl 5.10.0), which causes the
1712regex engine to "keep" everything it had matched prior to the C<\K> and
1713not include it in C<$&>. This effectively provides non-experimental
1714variable-length lookbehind of any length.
1715
1716And, there is a technique that can be used to handle variable length
1717lookbehinds on earlier releases, and longer than 255 characters.  It is
1718described in
1719L<http://www.drregex.com/2019/02/variable-length-lookbehinds-actually.html>.
1720
1721Note that under C</i>, a few single characters match two or three other
1722characters.  This makes them variable length, and the 255 length applies
1723to the maximum number of characters in the match.  For
1724example C<qr/\N{LATIN SMALL LETTER SHARP S}/i> matches the sequence
1725C<"ss">.  Your lookbehind assertion could contain 127 Sharp S
1726characters under C</i>, but adding a 128th would generate a compilation
1727error, as that could match 256 C<"s"> characters in a row.
1728
1729The use of C<\K> inside of another lookaround assertion
1730is allowed, but the behaviour is currently not well defined.
1731
1732For various reasons C<\K> may be significantly more efficient than the
1733equivalent C<< (?<=...) >> construct, and it is especially useful in
1734situations where you want to efficiently remove something following
1735something else in a string. For instance
1736
1737  s/(foo)bar/$1/g;
1738
1739can be rewritten as the much more efficient
1740
1741  s/foo\Kbar//g;
1742
1743Use of the non-greedy modifier C<"?"> may not give you the expected
1744results if it is within a capturing group within the construct.
1745
1746=item C<(?<!I<pattern>)>
1747
1748=item C<(*nlb:I<pattern>)>
1749
1750=item C<(*negative_lookbehind:I<pattern>)>
1751X<(?<!)>
1752X<(*nlb>
1753X<(*negative_lookbehind>
1754X<look-behind, negative> X<lookbehind, negative>
1755
1756A zero-width negative lookbehind assertion.  For example C</(?<!bar)foo/>
1757matches any occurrence of "foo" that does not follow "bar".
1758
1759Prior to Perl 5.30, it worked only for fixed-width lookbehind, but
1760starting in that release, it can handle variable lengths from 1 to 255
1761characters as an experimental feature.  The feature is enabled
1762automatically if you use a variable length negative lookbehind assertion.
1763
1764In Perl 5.35.10 the scope of the experimental nature of this construct
1765has been reduced, and experimental warnings will only be produced when
1766the construct contains capturing parentheses. The warnings will be
1767raised at pattern compilation time, unless turned off, in the
1768C<experimental::vlb> category.  This is to warn you that the exact
1769contents of capturing buffers in a variable length negative lookbehind
1770is not well defined and is subject to change in a future release of perl.
1771
1772Currently if you use capture buffers inside of a negative variable length
1773lookbehind the result may not be what you expect, for instance:
1774
1775    say "axfoo"=~/(?=foo)(?<!(a|ax)(?{ say $1 }))/ ? "y" : "n";
1776
1777will output the following:
1778
1779    a
1780    no
1781
1782which does not make sense as this should print out "ax" as the "a" does
1783not line up at the correct place. Another example would be:
1784
1785    say "yes: '$1-$2'" if "aayfoo"=~/(?=foo)(?<!(a|aa)(a|aa)x)/;
1786
1787will output the following:
1788
1789    yes: 'aa-a'
1790
1791It is possible in a future release of perl we will change this behavior
1792so both of these examples produced more reasonable output.
1793
1794Note that we are confident that the construct will match and reject
1795patterns appropriately, the undefined behavior strictly relates to the
1796value of the capture buffer during or after matching.
1797
1798There is a technique that can be used to handle variable length
1799lookbehind on earlier releases, and longer than 255 characters.  It is
1800described in
1801L<http://www.drregex.com/2019/02/variable-length-lookbehinds-actually.html>.
1802
1803Note that under C</i>, a few single characters match two or three other
1804characters.  This makes them variable length, and the 255 length applies
1805to the maximum number of characters in the match.  For
1806example C<qr/\N{LATIN SMALL LETTER SHARP S}/i> matches the sequence
1807C<"ss">.  Your lookbehind assertion could contain 127 Sharp S
1808characters under C</i>, but adding a 128th would generate a compilation
1809error, as that could match 256 C<"s"> characters in a row.
1810
1811Use of the non-greedy modifier C<"?"> may not give you the expected
1812results if it is within a capturing group within the construct.
1813
1814=back
1815
1816=item C<< (?<I<NAME>>I<pattern>) >>
1817
1818=item C<(?'I<NAME>'I<pattern>)>
1819X<< (?<NAME>) >> X<(?'NAME')> X<named capture> X<capture>
1820
1821A named capture group. Identical in every respect to normal capturing
1822parentheses C<()> but for the additional fact that the group
1823can be referred to by name in various regular expression
1824constructs (like C<\g{I<NAME>}>) and can be accessed by name
1825after a successful match via C<%+> or C<%->. See L<perlvar>
1826for more details on the C<%+> and C<%-> hashes.
1827
1828If multiple distinct capture groups have the same name, then
1829C<$+{I<NAME>}> will refer to the leftmost defined group in the match.
1830
1831The forms C<(?'I<NAME>'I<pattern>)> and C<< (?<I<NAME>>I<pattern>) >>
1832are equivalent.
1833
1834B<NOTE:> While the notation of this construct is the same as the similar
1835function in .NET regexes, the behavior is not. In Perl the groups are
1836numbered sequentially regardless of being named or not. Thus in the
1837pattern
1838
1839  /(x)(?<foo>y)(z)/
1840
1841C<$+{foo}> will be the same as C<$2>, and C<$3> will contain 'z' instead of
1842the opposite which is what a .NET regex hacker might expect.
1843
1844Currently I<NAME> is restricted to simple identifiers only.
1845In other words, it must match C</^[_A-Za-z][_A-Za-z0-9]*\z/> or
1846its Unicode extension (see L<utf8>),
1847though it isn't extended by the locale (see L<perllocale>).
1848
1849B<NOTE:> In order to make things easier for programmers with experience
1850with the Python or PCRE regex engines, the pattern C<<
1851(?PE<lt>I<NAME>E<gt>I<pattern>) >>
1852may be used instead of C<< (?<I<NAME>>I<pattern>) >>; however this form does not
1853support the use of single quotes as a delimiter for the name.
1854
1855=item C<< \k<I<NAME>> >>
1856
1857=item C<< \k'I<NAME>' >>
1858
1859=item C<< \k{I<NAME>} >>
1860
1861Named backreference. Similar to numeric backreferences, except that
1862the group is designated by name and not number. If multiple groups
1863have the same name then it refers to the leftmost defined group in
1864the current match.
1865
1866It is an error to refer to a name not defined by a C<< (?<I<NAME>>) >>
1867earlier in the pattern.
1868
1869All three forms are equivalent, although with C<< \k{ I<NAME> } >>,
1870you may optionally have blanks within but adjacent to the braces, as
1871shown.
1872
1873B<NOTE:> In order to make things easier for programmers with experience
1874with the Python or PCRE regex engines, the pattern C<< (?P=I<NAME>) >>
1875may be used instead of C<< \k<I<NAME>> >>.
1876
1877=item C<(?{ I<code> })>
1878X<(?{})> X<regex, code in> X<regexp, code in> X<regular expression, code in>
1879
1880B<WARNING>: Using this feature safely requires that you understand its
1881limitations.  Code executed that has side effects may not perform identically
1882from version to version due to the effect of future optimisations in the regex
1883engine.  For more information on this, see L</Embedded Code Execution
1884Frequency>.
1885
1886This zero-width assertion executes any embedded Perl code.  It always
1887succeeds, and its return value is set as C<$^R>.
1888
1889In literal patterns, the code is parsed at the same time as the
1890surrounding code. While within the pattern, control is passed temporarily
1891back to the perl parser, until the logically-balancing closing brace is
1892encountered. This is similar to the way that an array index expression in
1893a literal string is handled, for example
1894
1895    "abc$array[ 1 + f('[') + g()]def"
1896
1897In particular, braces do not need to be balanced:
1898
1899    s/abc(?{ f('{'); })/def/
1900
1901Even in a pattern that is interpolated and compiled at run-time, literal
1902code blocks will be compiled once, at perl compile time; the following
1903prints "ABCD":
1904
1905    print "D";
1906    my $qr = qr/(?{ BEGIN { print "A" } })/;
1907    my $foo = "foo";
1908    /$foo$qr(?{ BEGIN { print "B" } })/;
1909    BEGIN { print "C" }
1910
1911In patterns where the text of the code is derived from run-time
1912information rather than appearing literally in a source code /pattern/,
1913the code is compiled at the same time that the pattern is compiled, and
1914for reasons of security, C<use re 'eval'> must be in scope. This is to
1915stop user-supplied patterns containing code snippets from being
1916executable.
1917
1918In situations where you need to enable this with C<use re 'eval'>, you should
1919also have taint checking enabled, if your perl supports it.
1920Better yet, use the carefully constrained evaluation within a Safe compartment.
1921See L<perlsec> for details about both these mechanisms.
1922
1923From the viewpoint of parsing, lexical variable scope and closures,
1924
1925    /AAA(?{ BBB })CCC/
1926
1927behaves approximately like
1928
1929    /AAA/ && do { BBB } && /CCC/
1930
1931Similarly,
1932
1933    qr/AAA(?{ BBB })CCC/
1934
1935behaves approximately like
1936
1937    sub { /AAA/ && do { BBB } && /CCC/ }
1938
1939In particular:
1940
1941    { my $i = 1; $r = qr/(?{ print $i })/ }
1942    my $i = 2;
1943    /$r/; # prints "1"
1944
1945Inside a C<(?{...})> block, C<$_> refers to the string the regular
1946expression is matching against. You can also use C<pos()> to know what is
1947the current position of matching within this string.
1948
1949The code block introduces a new scope from the perspective of lexical
1950variable declarations, but B<not> from the perspective of C<local> and
1951similar localizing behaviours. So later code blocks within the same
1952pattern will still see the values which were localized in earlier blocks.
1953These accumulated localizations are undone either at the end of a
1954successful match, or if the assertion is backtracked (compare
1955L</"Backtracking">). For example,
1956
1957  $_ = 'a' x 8;
1958  m<
1959     (?{ $cnt = 0 })               # Initialize $cnt.
1960     (
1961       a
1962       (?{
1963           local $cnt = $cnt + 1;  # Update $cnt,
1964                                   # backtracking-safe.
1965       })
1966     )*
1967     aaaa
1968     (?{ $res = $cnt })            # On success copy to
1969                                   # non-localized location.
1970   >x;
1971
1972will initially increment C<$cnt> up to 8; then during backtracking, its
1973value will be unwound back to 4, which is the value assigned to C<$res>.
1974At the end of the regex execution, C<$cnt> will be wound back to its initial
1975value of 0.
1976
1977This assertion may be used as the condition in a
1978
1979    (?(condition)yes-pattern|no-pattern)
1980
1981switch.  If I<not> used in this way, the result of evaluation of I<code>
1982is put into the special variable C<$^R>.  This happens immediately, so
1983C<$^R> can be used from other C<(?{ I<code> })> assertions inside the same
1984regular expression.
1985
1986The assignment to C<$^R> above is properly localized, so the old
1987value of C<$^R> is restored if the assertion is backtracked; compare
1988L</"Backtracking">.
1989
1990Note that the special variable C<$^N>  is particularly useful with code
1991blocks to capture the results of submatches in variables without having to
1992keep track of the number of nested parentheses. For example:
1993
1994  $_ = "The brown fox jumps over the lazy dog";
1995  /the (\S+)(?{ $color = $^N }) (\S+)(?{ $animal = $^N })/i;
1996  print "color = $color, animal = $animal\n";
1997
1998The use of this construct disables some optimisations globally in the
1999pattern, and the pattern may execute much slower as a consequence.
2000Use a C<*> instead of the C<?> block to create an optimistic form of
2001this construct. C<(*{ ... })> should not disable any optimisations.
2002
2003=item C<(*{ I<code> })>
2004X<(*{})> X<regex, optimistic code>
2005
2006This is *exactly* the same as C<(?{ I<code> })> with the exception
2007that it does not disable B<any> optimisations at all in the regex engine.
2008How often it is executed may vary from perl release to perl release.
2009In a failing match it may not even be executed at all.
2010
2011=item C<(??{ I<code> })>
2012X<(??{})>
2013X<regex, postponed> X<regexp, postponed> X<regular expression, postponed>
2014
2015B<WARNING>: Using this feature safely requires that you understand its
2016limitations.  Code executed that has side effects may not perform
2017identically from version to version due to the effect of future
2018optimisations in the regex engine.  For more information on this, see
2019L</Embedded Code Execution Frequency>.
2020
2021This is a "postponed" regular subexpression.  It behaves in I<exactly> the
2022same way as a C<(?{ I<code> })> code block as described above, except that
2023its return value, rather than being assigned to C<$^R>, is treated as a
2024pattern, compiled if it's a string (or used as-is if its a qr// object),
2025then matched as if it were inserted instead of this construct.
2026
2027During the matching of this sub-pattern, it has its own set of
2028captures which are valid during the sub-match, but are discarded once
2029control returns to the main pattern. For example, the following matches,
2030with the inner pattern capturing "B" and matching "BB", while the outer
2031pattern captures "A";
2032
2033    my $inner = '(.)\1';
2034    "ABBA" =~ /^(.)(??{ $inner })\1/;
2035    print $1; # prints "A";
2036
2037Note that this means that  there is no way for the inner pattern to refer
2038to a capture group defined outside.  (The code block itself can use C<$1>,
2039I<etc>., to refer to the enclosing pattern's capture groups.)  Thus, although
2040
2041    ('a' x 100)=~/(??{'(.)' x 100})/
2042
2043I<will> match, it will I<not> set C<$1> on exit.
2044
2045The following pattern matches a parenthesized group:
2046
2047 $re = qr{
2048            \(
2049            (?:
2050               (?> [^()]+ )  # Non-parens without backtracking
2051             |
2052               (??{ $re })   # Group with matching parens
2053            )*
2054            \)
2055         }x;
2056
2057See also
2058L<C<(?I<PARNO>)>|/(?I<PARNO>) (?-I<PARNO>) (?+I<PARNO>) (?R) (?0)>
2059for a different, more efficient way to accomplish
2060the same task.
2061
2062Executing a postponed regular expression too many times without
2063consuming any input string will also result in a fatal error.  The depth
2064at which that happens is compiled into perl, so it can be changed with a
2065custom build.
2066
2067The use of this construct disables some optimisations globally in the pattern,
2068and the pattern may execute much slower as a consequence.
2069
2070=item C<(?I<PARNO>)> C<(?-I<PARNO>)> C<(?+I<PARNO>)> C<(?R)> C<(?0)>
2071X<(?PARNO)> X<(?1)> X<(?R)> X<(?0)> X<(?-1)> X<(?+1)> X<(?-PARNO)> X<(?+PARNO)>
2072X<regex, recursive> X<regexp, recursive> X<regular expression, recursive>
2073X<regex, relative recursion> X<GOSUB> X<GOSTART>
2074
2075Recursive subpattern. Treat the contents of a given capture buffer in the
2076current pattern as an independent subpattern and attempt to match it at
2077the current position in the string. Information about capture state from
2078the caller for things like backreferences is available to the subpattern,
2079but capture buffers set by the subpattern are not visible to the caller.
2080
2081Similar to C<(??{ I<code> })> except that it does not involve executing any
2082code or potentially compiling a returned pattern string; instead it treats
2083the part of the current pattern contained within a specified capture group
2084as an independent pattern that must match at the current position. Also
2085different is the treatment of capture buffers, unlike C<(??{ I<code> })>
2086recursive patterns have access to their caller's match state, so one can
2087use backreferences safely.
2088
2089I<PARNO> is a sequence of digits (not starting with 0) whose value reflects
2090the paren-number of the capture group to recurse to. C<(?R)> recurses to
2091the beginning of the whole pattern. C<(?0)> is an alternate syntax for
2092C<(?R)>. If I<PARNO> is preceded by a plus or minus sign then it is assumed
2093to be relative, with negative numbers indicating preceding capture groups
2094and positive ones following. Thus C<(?-1)> refers to the most recently
2095declared group, and C<(?+1)> indicates the next group to be declared.
2096Note that the counting for relative recursion differs from that of
2097relative backreferences, in that with recursion unclosed groups B<are>
2098included.
2099
2100The following pattern matches a function C<foo()> which may contain
2101balanced parentheses as the argument.
2102
2103  $re = qr{ (                   # paren group 1 (full function)
2104              foo
2105              (                 # paren group 2 (parens)
2106                \(
2107                  (             # paren group 3 (contents of parens)
2108                  (?:
2109                   (?> [^()]+ ) # Non-parens without backtracking
2110                  |
2111                   (?2)         # Recurse to start of paren group 2
2112                  )*
2113                  )
2114                \)
2115              )
2116            )
2117          }x;
2118
2119If the pattern was used as follows
2120
2121    'foo(bar(baz)+baz(bop))'=~/$re/
2122        and print "\$1 = $1\n",
2123                  "\$2 = $2\n",
2124                  "\$3 = $3\n";
2125
2126the output produced should be the following:
2127
2128    $1 = foo(bar(baz)+baz(bop))
2129    $2 = (bar(baz)+baz(bop))
2130    $3 = bar(baz)+baz(bop)
2131
2132If there is no corresponding capture group defined, then it is a
2133fatal error.  Recursing deeply without consuming any input string will
2134also result in a fatal error.  The depth at which that happens is
2135compiled into perl, so it can be changed with a custom build.
2136
2137The following shows how using negative indexing can make it
2138easier to embed recursive patterns inside of a C<qr//> construct
2139for later use:
2140
2141    my $parens = qr/(\((?:[^()]++|(?-1))*+\))/;
2142    if (/foo $parens \s+ \+ \s+ bar $parens/x) {
2143       # do something here...
2144    }
2145
2146B<Note> that this pattern does not behave the same way as the equivalent
2147PCRE or Python construct of the same form. In Perl you can backtrack into
2148a recursed group, in PCRE and Python the recursed into group is treated
2149as atomic. Also, modifiers are resolved at compile time, so constructs
2150like C<(?i:(?1))> or C<(?:(?i)(?1))> do not affect how the sub-pattern will
2151be processed.
2152
2153=item C<(?&I<NAME>)>
2154X<(?&NAME)>
2155
2156Recurse to a named subpattern. Identical to C<(?I<PARNO>)> except that the
2157parenthesis to recurse to is determined by name. If multiple parentheses have
2158the same name, then it recurses to the leftmost.
2159
2160It is an error to refer to a name that is not declared somewhere in the
2161pattern.
2162
2163B<NOTE:> In order to make things easier for programmers with experience
2164with the Python or PCRE regex engines the pattern C<< (?P>I<NAME>) >>
2165may be used instead of C<< (?&I<NAME>) >>.
2166
2167=item C<(?(I<condition>)I<yes-pattern>|I<no-pattern>)>
2168X<(?()>
2169
2170=item C<(?(I<condition>)I<yes-pattern>)>
2171
2172Conditional expression. Matches I<yes-pattern> if I<condition> yields
2173a true value, matches I<no-pattern> otherwise. A missing pattern always
2174matches.
2175
2176C<(I<condition>)> should be one of:
2177
2178=over 4
2179
2180=item an integer in parentheses
2181
2182(which is valid if the corresponding pair of parentheses
2183matched);
2184
2185=item a lookahead/lookbehind/evaluate zero-width assertion;
2186
2187=item a name in angle brackets or single quotes
2188
2189(which is valid if a group with the given name matched);
2190
2191=item the special symbol C<(R)>
2192
2193(true when evaluated inside of recursion or eval).  Additionally the
2194C<"R"> may be
2195followed by a number, (which will be true when evaluated when recursing
2196inside of the appropriate group), or by C<&I<NAME>>, in which case it will
2197be true only when evaluated during recursion in the named group.
2198
2199=back
2200
2201Here's a summary of the possible predicates:
2202
2203=over 4
2204
2205=item C<(1)> C<(2)> ...
2206
2207Checks if the numbered capturing group has matched something.
2208Full syntax: C<< (?(1)then|else) >>
2209
2210=item C<(E<lt>I<NAME>E<gt>)> C<('I<NAME>')>
2211
2212Checks if a group with the given name has matched something.
2213Full syntax: C<< (?(<name>)then|else) >>
2214
2215=item C<(?=...)> C<(?!...)> C<(?<=...)> C<(?<!...)>
2216
2217Checks whether the pattern matches (or does not match, for the C<"!">
2218variants).
2219Full syntax: C<< (?(?=I<lookahead>)I<then>|I<else>) >>
2220
2221=item C<(?{ I<CODE> })>
2222
2223Treats the return value of the code block as the condition.
2224Full syntax: C<< (?(?{ I<CODE> })I<then>|I<else>) >>
2225
2226Note use of this construct may globally affect the performance
2227of the pattern. Consider using C<(*{ I<CODE> })>
2228
2229=item C<(*{ I<CODE> })>
2230
2231Treats the return value of the code block as the condition.
2232Full syntax: C<< (?(*{ I<CODE> })I<then>|I<else>) >>
2233
2234=item C<(R)>
2235
2236Checks if the expression has been evaluated inside of recursion.
2237Full syntax: C<< (?(R)I<then>|I<else>) >>
2238
2239=item C<(R1)> C<(R2)> ...
2240
2241Checks if the expression has been evaluated while executing directly
2242inside of the n-th capture group. This check is the regex equivalent of
2243
2244  if ((caller(0))[3] eq 'subname') { ... }
2245
2246In other words, it does not check the full recursion stack.
2247
2248Full syntax: C<< (?(R1)I<then>|I<else>) >>
2249
2250=item C<(R&I<NAME>)>
2251
2252Similar to C<(R1)>, this predicate checks to see if we're executing
2253directly inside of the leftmost group with a given name (this is the same
2254logic used by C<(?&I<NAME>)> to disambiguate). It does not check the full
2255stack, but only the name of the innermost active recursion.
2256Full syntax: C<< (?(R&I<name>)I<then>|I<else>) >>
2257
2258=item C<(DEFINE)>
2259
2260In this case, the yes-pattern is never directly executed, and no
2261no-pattern is allowed. Similar in spirit to C<(?{0})> but more efficient.
2262See below for details.
2263Full syntax: C<< (?(DEFINE)I<definitions>...) >>
2264
2265=back
2266
2267For example:
2268
2269    m{ ( \( )?
2270       [^()]+
2271       (?(1) \) )
2272     }x
2273
2274matches a chunk of non-parentheses, possibly included in parentheses
2275themselves.
2276
2277A special form is the C<(DEFINE)> predicate, which never executes its
2278yes-pattern directly, and does not allow a no-pattern. This allows one to
2279define subpatterns which will be executed only by the recursion mechanism.
2280This way, you can define a set of regular expression rules that can be
2281bundled into any pattern you choose.
2282
2283It is recommended that for this usage you put the DEFINE block at the
2284end of the pattern, and that you name any subpatterns defined within it.
2285
2286Also, it's worth noting that patterns defined this way probably will
2287not be as efficient, as the optimizer is not very clever about
2288handling them.
2289
2290An example of how this might be used is as follows:
2291
2292  /(?<NAME>(?&NAME_PAT))(?<ADDR>(?&ADDRESS_PAT))
2293   (?(DEFINE)
2294     (?<NAME_PAT>....)
2295     (?<ADDRESS_PAT>....)
2296   )/x
2297
2298Note that capture groups matched inside of recursion are not accessible
2299after the recursion returns, so the extra layer of capturing groups is
2300necessary. Thus C<$+{NAME_PAT}> would not be defined even though
2301C<$+{NAME}> would be.
2302
2303Finally, keep in mind that subpatterns created inside a DEFINE block
2304count towards the absolute and relative number of captures, so this:
2305
2306    my @captures = "a" =~ /(.)                  # First capture
2307                           (?(DEFINE)
2308                               (?<EXAMPLE> 1 )  # Second capture
2309                           )/x;
2310    say scalar @captures;
2311
2312Will output 2, not 1. This is particularly important if you intend to
2313compile the definitions with the C<qr//> operator, and later
2314interpolate them in another pattern.
2315
2316=item C<< (?>I<pattern>) >>
2317
2318=item C<< (*atomic:I<pattern>) >>
2319X<(?E<gt>pattern)>
2320X<(*atomic>
2321X<backtrack> X<backtracking> X<atomic> X<possessive>
2322
2323An "independent" subexpression, one which matches the substring
2324that a standalone I<pattern> would match if anchored at the given
2325position, and it matches I<nothing other than this substring>.  This
2326construct is useful for optimizations of what would otherwise be
2327"eternal" matches, because it will not backtrack (see L</"Backtracking">).
2328It may also be useful in places where the "grab all you can, and do not
2329give anything back" semantic is desirable.
2330
2331For example: C<< ^(?>a*)ab >> will never match, since C<< (?>a*) >>
2332(anchored at the beginning of string, as above) will match I<all>
2333characters C<"a"> at the beginning of string, leaving no C<"a"> for
2334C<ab> to match.  In contrast, C<a*ab> will match the same as C<a+b>,
2335since the match of the subgroup C<a*> is influenced by the following
2336group C<ab> (see L</"Backtracking">).  In particular, C<a*> inside
2337C<a*ab> will match fewer characters than a standalone C<a*>, since
2338this makes the tail match.
2339
2340C<< (?>I<pattern>) >> does not disable backtracking altogether once it has
2341matched. It is still possible to backtrack past the construct, but not
2342into it. So C<< ((?>a*)|(?>b*))ar >> will still match "bar".
2343
2344An effect similar to C<< (?>I<pattern>) >> may be achieved by writing
2345C<(?=(I<pattern>))\g{-1}>.  This matches the same substring as a standalone
2346C<a+>, and the following C<\g{-1}> eats the matched string; it therefore
2347makes a zero-length assertion into an analogue of C<< (?>...) >>.
2348(The difference between these two constructs is that the second one
2349uses a capturing group, thus shifting ordinals of backreferences
2350in the rest of a regular expression.)
2351
2352Consider this pattern:
2353
2354    m{ \(
2355          (
2356            [^()]+           # x+
2357          |
2358            \( [^()]* \)
2359          )+
2360       \)
2361     }x
2362
2363That will efficiently match a nonempty group with matching parentheses
2364two levels deep or less.  However, if there is no such group, it
2365will take virtually forever on a long string.  That's because there
2366are so many different ways to split a long string into several
2367substrings.  This is what C<(.+)+> is doing, and C<(.+)+> is similar
2368to a subpattern of the above pattern.  Consider how the pattern
2369above detects no-match on C<((()aaaaaaaaaaaaaaaaaa> in several
2370seconds, but that each extra letter doubles this time.  This
2371exponential performance will make it appear that your program has
2372hung.  However, a tiny change to this pattern
2373
2374    m{ \(
2375          (
2376            (?> [^()]+ )        # change x+ above to (?> x+ )
2377          |
2378            \( [^()]* \)
2379          )+
2380       \)
2381     }x
2382
2383which uses C<< (?>...) >> matches exactly when the one above does (verifying
2384this yourself would be a productive exercise), but finishes in a fourth
2385the time when used on a similar string with 1000000 C<"a">s.  Be aware,
2386however, that, when this construct is followed by a
2387quantifier, it currently triggers a warning message under
2388the C<use warnings> pragma or B<-w> switch saying it
2389C<"matches null string many times in regex">.
2390
2391On simple groups, such as the pattern C<< (?> [^()]+ ) >>, a comparable
2392effect may be achieved by negative lookahead, as in C<[^()]+ (?! [^()] )>.
2393This was only 4 times slower on a string with 1000000 C<"a">s.
2394
2395The "grab all you can, and do not give anything back" semantic is desirable
2396in many situations where on the first sight a simple C<()*> looks like
2397the correct solution.  Suppose we parse text with comments being delimited
2398by C<"#"> followed by some optional (horizontal) whitespace.  Contrary to
2399its appearance, C<#[ \t]*> I<is not> the correct subexpression to match
2400the comment delimiter, because it may "give up" some whitespace if
2401the remainder of the pattern can be made to match that way.  The correct
2402answer is either one of these:
2403
2404    (?>#[ \t]*)
2405    #[ \t]*(?![ \t])
2406
2407For example, to grab non-empty comments into C<$1>, one should use either
2408one of these:
2409
2410    / (?> \# [ \t]* ) (        .+ ) /x;
2411    /     \# [ \t]*   ( [^ \t] .* ) /x;
2412
2413Which one you pick depends on which of these expressions better reflects
2414the above specification of comments.
2415
2416In some literature this construct is called "atomic matching" or
2417"possessive matching".
2418
2419Possessive quantifiers are equivalent to putting the item they are applied
2420to inside of one of these constructs. The following equivalences apply:
2421
2422    Quantifier Form     Bracketing Form
2423    ---------------     ---------------
2424    PAT*+               (?>PAT*)
2425    PAT++               (?>PAT+)
2426    PAT?+               (?>PAT?)
2427    PAT{min,max}+       (?>PAT{min,max})
2428
2429Nested C<(?E<gt>...)> constructs are not no-ops, even if at first glance
2430they might seem to be.  This is because the nested C<(?E<gt>...)> can
2431restrict internal backtracking that otherwise might occur.  For example,
2432
2433 "abc" =~ /(?>a[bc]*c)/
2434
2435matches, but
2436
2437 "abc" =~ /(?>a(?>[bc]*)c)/
2438
2439does not.
2440
2441=item C<(?[ ])>
2442
2443See L<perlrecharclass/Extended Bracketed Character Classes>.
2444
2445=back
2446
2447=head2 Backtracking
2448X<backtrack> X<backtracking>
2449
2450NOTE: This section presents an abstract approximation of regular
2451expression behavior.  For a more rigorous (and complicated) view of
2452the rules involved in selecting a match among possible alternatives,
2453see L</Combining RE Pieces>.
2454
2455A fundamental feature of regular expression matching involves the
2456notion called I<backtracking>, which is currently used (when needed)
2457by all regular non-possessive expression quantifiers, namely C<"*">,
2458C<*?>, C<"+">, C<+?>, C<{n,m}>, and C<{n,m}?>.  Backtracking is often
2459optimized internally, but the general principle outlined here is valid.
2460
2461For a regular expression to match, the I<entire> regular expression must
2462match, not just part of it.  So if the beginning of a pattern containing a
2463quantifier succeeds in a way that causes later parts in the pattern to
2464fail, the matching engine backs up and recalculates the beginning
2465part--that's why it's called backtracking.
2466
2467Here is an example of backtracking:  Let's say you want to find the
2468word following "foo" in the string "Food is on the foo table.":
2469
2470    $_ = "Food is on the foo table.";
2471    if ( /\b(foo)\s+(\w+)/i ) {
2472        print "$2 follows $1.\n";
2473    }
2474
2475When the match runs, the first part of the regular expression (C<\b(foo)>)
2476finds a possible match right at the beginning of the string, and loads up
2477C<$1> with "Foo".  However, as soon as the matching engine sees that there's
2478no whitespace following the "Foo" that it had saved in C<$1>, it realizes its
2479mistake and starts over again one character after where it had the
2480tentative match.  This time it goes all the way until the next occurrence
2481of "foo". The complete regular expression matches this time, and you get
2482the expected output of "table follows foo."
2483
2484Sometimes minimal matching can help a lot.  Imagine you'd like to match
2485everything between "foo" and "bar".  Initially, you write something
2486like this:
2487
2488    $_ =  "The food is under the bar in the barn.";
2489    if ( /foo(.*)bar/ ) {
2490        print "got <$1>\n";
2491    }
2492
2493Which perhaps unexpectedly yields:
2494
2495  got <d is under the bar in the >
2496
2497That's because C<.*> was greedy, so you get everything between the
2498I<first> "foo" and the I<last> "bar".  Here it's more effective
2499to use minimal matching to make sure you get the text between a "foo"
2500and the first "bar" thereafter.
2501
2502    if ( /foo(.*?)bar/ ) { print "got <$1>\n" }
2503  got <d is under the >
2504
2505Here's another example. Let's say you'd like to match a number at the end
2506of a string, and you also want to keep the preceding part of the match.
2507So you write this:
2508
2509    $_ = "I have 2 numbers: 53147";
2510    if ( /(.*)(\d*)/ ) {                                # Wrong!
2511        print "Beginning is <$1>, number is <$2>.\n";
2512    }
2513
2514That won't work at all, because C<.*> was greedy and gobbled up the
2515whole string. As C<\d*> can match on an empty string the complete
2516regular expression matched successfully.
2517
2518    Beginning is <I have 2 numbers: 53147>, number is <>.
2519
2520Here are some variants, most of which don't work:
2521
2522    $_ = "I have 2 numbers: 53147";
2523    @pats = qw{
2524        (.*)(\d*)
2525        (.*)(\d+)
2526        (.*?)(\d*)
2527        (.*?)(\d+)
2528        (.*)(\d+)$
2529        (.*?)(\d+)$
2530        (.*)\b(\d+)$
2531        (.*\D)(\d+)$
2532    };
2533
2534    for $pat (@pats) {
2535        printf "%-12s ", $pat;
2536        if ( /$pat/ ) {
2537            print "<$1> <$2>\n";
2538        } else {
2539            print "FAIL\n";
2540        }
2541    }
2542
2543That will print out:
2544
2545    (.*)(\d*)    <I have 2 numbers: 53147> <>
2546    (.*)(\d+)    <I have 2 numbers: 5314> <7>
2547    (.*?)(\d*)   <> <>
2548    (.*?)(\d+)   <I have > <2>
2549    (.*)(\d+)$   <I have 2 numbers: 5314> <7>
2550    (.*?)(\d+)$  <I have 2 numbers: > <53147>
2551    (.*)\b(\d+)$ <I have 2 numbers: > <53147>
2552    (.*\D)(\d+)$ <I have 2 numbers: > <53147>
2553
2554As you see, this can be a bit tricky.  It's important to realize that a
2555regular expression is merely a set of assertions that gives a definition
2556of success.  There may be 0, 1, or several different ways that the
2557definition might succeed against a particular string.  And if there are
2558multiple ways it might succeed, you need to understand backtracking to
2559know which variety of success you will achieve.
2560
2561When using lookahead assertions and negations, this can all get even
2562trickier.  Imagine you'd like to find a sequence of non-digits not
2563followed by "123".  You might try to write that as
2564
2565    $_ = "ABC123";
2566    if ( /^\D*(?!123)/ ) {                # Wrong!
2567        print "Yup, no 123 in $_\n";
2568    }
2569
2570But that isn't going to match; at least, not the way you're hoping.  It
2571claims that there is no 123 in the string.  Here's a clearer picture of
2572why that pattern matches, contrary to popular expectations:
2573
2574    $x = 'ABC123';
2575    $y = 'ABC445';
2576
2577    print "1: got $1\n" if $x =~ /^(ABC)(?!123)/;
2578    print "2: got $1\n" if $y =~ /^(ABC)(?!123)/;
2579
2580    print "3: got $1\n" if $x =~ /^(\D*)(?!123)/;
2581    print "4: got $1\n" if $y =~ /^(\D*)(?!123)/;
2582
2583This prints
2584
2585    2: got ABC
2586    3: got AB
2587    4: got ABC
2588
2589You might have expected test 3 to fail because it seems to a more
2590general purpose version of test 1.  The important difference between
2591them is that test 3 contains a quantifier (C<\D*>) and so can use
2592backtracking, whereas test 1 will not.  What's happening is
2593that you've asked "Is it true that at the start of C<$x>, following 0 or more
2594non-digits, you have something that's not 123?"  If the pattern matcher had
2595let C<\D*> expand to "ABC", this would have caused the whole pattern to
2596fail.
2597
2598The search engine will initially match C<\D*> with "ABC".  Then it will
2599try to match C<(?!123)> with "123", which fails.  But because
2600a quantifier (C<\D*>) has been used in the regular expression, the
2601search engine can backtrack and retry the match differently
2602in the hope of matching the complete regular expression.
2603
2604The pattern really, I<really> wants to succeed, so it uses the
2605standard pattern back-off-and-retry and lets C<\D*> expand to just "AB" this
2606time.  Now there's indeed something following "AB" that is not
2607"123".  It's "C123", which suffices.
2608
2609We can deal with this by using both an assertion and a negation.
2610We'll say that the first part in C<$1> must be followed both by a digit
2611and by something that's not "123".  Remember that the lookaheads
2612are zero-width expressions--they only look, but don't consume any
2613of the string in their match.  So rewriting this way produces what
2614you'd expect; that is, case 5 will fail, but case 6 succeeds:
2615
2616    print "5: got $1\n" if $x =~ /^(\D*)(?=\d)(?!123)/;
2617    print "6: got $1\n" if $y =~ /^(\D*)(?=\d)(?!123)/;
2618
2619    6: got ABC
2620
2621In other words, the two zero-width assertions next to each other work as though
2622they're ANDed together, just as you'd use any built-in assertions:  C</^$/>
2623matches only if you're at the beginning of the line AND the end of the
2624line simultaneously.  The deeper underlying truth is that juxtaposition in
2625regular expressions always means AND, except when you write an explicit OR
2626using the vertical bar.  C</ab/> means match "a" AND (then) match "b",
2627although the attempted matches are made at different positions because "a"
2628is not a zero-width assertion, but a one-width assertion.
2629
2630B<WARNING>: Particularly complicated regular expressions can take
2631exponential time to solve because of the immense number of possible
2632ways they can use backtracking to try for a match.  For example, without
2633internal optimizations done by the regular expression engine, this will
2634take a painfully long time to run:
2635
2636    'aaaaaaaaaaaa' =~ /((a{0,5}){0,5})*[c]/
2637
2638And if you used C<"*">'s in the internal groups instead of limiting them
2639to 0 through 5 matches, then it would take forever--or until you ran
2640out of stack space.  Moreover, these internal optimizations are not
2641always applicable.  For example, if you put C<{0,5}> instead of C<"*">
2642on the external group, no current optimization is applicable, and the
2643match takes a long time to finish.
2644
2645A powerful tool for optimizing such beasts is what is known as an
2646"independent group",
2647which does not backtrack (see C<L</(?E<gt>pattern)>>).  Note also that
2648zero-length lookahead/lookbehind assertions will not backtrack to make
2649the tail match, since they are in "logical" context: only
2650whether they match is considered relevant.  For an example
2651where side-effects of lookahead I<might> have influenced the
2652following match, see C<L</(?E<gt>pattern)>>.
2653
2654=head2 Script Runs
2655X<(*script_run:...)> X<(sr:...)>
2656X<(*atomic_script_run:...)> X<(asr:...)>
2657
2658A script run is basically a sequence of characters, all from the same
2659Unicode script (see L<perlunicode/Scripts>), such as Latin or Greek.  In
2660most places a single word would never be written in multiple scripts,
2661unless it is a spoofing attack.  An infamous example, is
2662
2663 paypal.com
2664
2665Those letters could all be Latin (as in the example just above), or they
2666could be all Cyrillic (except for the dot), or they could be a mixture
2667of the two.  In the case of an internet address the C<.com> would be in
2668Latin, And any Cyrillic ones would cause it to be a mixture, not a
2669script run.  Someone clicking on such a link would not be directed to
2670the real Paypal website, but an attacker would craft a look-alike one to
2671attempt to gather sensitive information from the person.
2672
2673Starting in Perl 5.28, it is now easy to detect strings that aren't
2674script runs.  Simply enclose just about any pattern like either of
2675these:
2676
2677 (*script_run:pattern)
2678 (*sr:pattern)
2679
2680What happens is that after I<pattern> succeeds in matching, it is
2681subjected to the additional criterion that every character in it must be
2682from the same script (see exceptions below).  If this isn't true,
2683backtracking occurs until something all in the same script is found that
2684matches, or all possibilities are exhausted.  This can cause a lot of
2685backtracking, but generally, only malicious input will result in this,
2686though the slow down could cause a denial of service attack.  If your
2687needs permit, it is best to make the pattern atomic to cut down on the
2688amount of backtracking.  This is so likely to be what you want, that
2689instead of writing this:
2690
2691 (*script_run:(?>pattern))
2692
2693you can write either of these:
2694
2695 (*atomic_script_run:pattern)
2696 (*asr:pattern)
2697
2698(See C<L</(?E<gt>I<pattern>)>>.)
2699
2700In Taiwan, Japan, and Korea, it is common for text to have a mixture of
2701characters from their native scripts and base Chinese.  Perl follows
2702Unicode's UTS 39 (L<https://unicode.org/reports/tr39/>) Unicode Security
2703Mechanisms in allowing such mixtures.  For example, the Japanese scripts
2704Katakana and Hiragana are commonly mixed together in practice, along
2705with some Chinese characters, and hence are treated as being in a single
2706script run by Perl.
2707
2708The rules used for matching decimal digits are slightly stricter.  Many
2709scripts have their own sets of digits equivalent to the Western C<0>
2710through C<9> ones.  A few, such as Arabic, have more than one set.  For
2711a string to be considered a script run, all digits in it must come from
2712the same set of ten, as determined by the first digit encountered.
2713As an example,
2714
2715 qr/(*script_run: \d+ \b )/x
2716
2717guarantees that the digits matched will all be from the same set of 10.
2718You won't get a look-alike digit from a different script that has a
2719different value than what it appears to be.
2720
2721Unicode has three pseudo scripts that are handled specially.
2722
2723"Unknown" is applied to code points whose meaning has yet to be
2724determined.  Perl currently will match as a script run, any single
2725character string consisting of one of these code points.  But any string
2726longer than one code point containing one of these will not be
2727considered a script run.
2728
2729"Inherited" is applied to characters that modify another, such as an
2730accent of some type.  These are considered to be in the script of the
2731master character, and so never cause a script run to not match.
2732
2733The other one is "Common".  This consists of mostly punctuation, emoji,
2734characters used in mathematics and music, the ASCII digits C<0>
2735through C<9>, and full-width forms of these digits.  These characters
2736can appear intermixed in text in many of the world's scripts.  These
2737also don't cause a script run to not match.  But like other scripts, all
2738digits in a run must come from the same set of 10.
2739
2740This construct is non-capturing.  You can add parentheses to I<pattern>
2741to capture, if desired.  You will have to do this if you plan to use
2742L</(*ACCEPT) (*ACCEPT:arg)> and not have it bypass the script run
2743checking.
2744
2745The C<Script_Extensions> property as modified by UTS 39
2746(L<https://unicode.org/reports/tr39/>) is used as the basis for this
2747feature.
2748
2749To summarize,
2750
2751=over 4
2752
2753=item *
2754
2755All length 0 or length 1 sequences are script runs.
2756
2757=item *
2758
2759A longer sequence is a script run if and only if B<all> of the following
2760conditions are met:
2761
2762Z<>
2763
2764=over
2765
2766=item 1
2767
2768No code point in the sequence has the C<Script_Extension> property of
2769C<Unknown>.
2770
2771This currently means that all code points in the sequence have been
2772assigned by Unicode to be characters that aren't private use nor
2773surrogate code points.
2774
2775=item 2
2776
2777All characters in the sequence come from the Common script and/or the
2778Inherited script and/or a single other script.
2779
2780The script of a character is determined by the C<Script_Extensions>
2781property as modified by UTS 39 (L<https://unicode.org/reports/tr39/>), as
2782described above.
2783
2784=item 3
2785
2786All decimal digits in the sequence come from the same block of 10
2787consecutive digits.
2788
2789=back
2790
2791=back
2792
2793=head2 Special Backtracking Control Verbs
2794
2795These special patterns are generally of the form C<(*I<VERB>:I<arg>)>. Unless
2796otherwise stated the I<arg> argument is optional; in some cases, it is
2797mandatory.
2798
2799Any pattern containing a special backtracking verb that allows an argument
2800has the special behaviour that when executed it sets the current package's
2801C<$REGERROR> and C<$REGMARK> variables. When doing so the following
2802rules apply:
2803
2804On failure, the C<$REGERROR> variable will be set to the I<arg> value of the
2805verb pattern, if the verb was involved in the failure of the match. If the
2806I<arg> part of the pattern was omitted, then C<$REGERROR> will be set to the
2807name of the last C<(*MARK:I<NAME>)> pattern executed, or to TRUE if there was
2808none. Also, the C<$REGMARK> variable will be set to FALSE.
2809
2810On a successful match, the C<$REGERROR> variable will be set to FALSE, and
2811the C<$REGMARK> variable will be set to the name of the last
2812C<(*MARK:I<NAME>)> pattern executed.  See the explanation for the
2813C<(*MARK:I<NAME>)> verb below for more details.
2814
2815B<NOTE:> C<$REGERROR> and C<$REGMARK> are not magic variables like C<$1>
2816and most other regex-related variables. They are not local to a scope, nor
2817readonly, but instead are volatile package variables similar to C<$AUTOLOAD>.
2818They are set in the package containing the code that I<executed> the regex
2819(rather than the one that compiled it, where those differ).  If necessary, you
2820can use C<local> to localize changes to these variables to a specific scope
2821before executing a regex.
2822
2823If a pattern does not contain a special backtracking verb that allows an
2824argument, then C<$REGERROR> and C<$REGMARK> are not touched at all.
2825
2826=over 3
2827
2828=item Verbs
2829
2830=over 4
2831
2832=item C<(*PRUNE)> C<(*PRUNE:I<NAME>)>
2833X<(*PRUNE)> X<(*PRUNE:NAME)>
2834
2835This zero-width pattern prunes the backtracking tree at the current point
2836when backtracked into on failure. Consider the pattern C</I<A> (*PRUNE) I<B>/>,
2837where I<A> and I<B> are complex patterns. Until the C<(*PRUNE)> verb is reached,
2838I<A> may backtrack as necessary to match. Once it is reached, matching
2839continues in I<B>, which may also backtrack as necessary; however, should B
2840not match, then no further backtracking will take place, and the pattern
2841will fail outright at the current starting position.
2842
2843The following example counts all the possible matching strings in a
2844pattern (without actually matching any of them).
2845
2846    'aaab' =~ /a+b?(?{print "$&\n"; $count++})(*FAIL)/;
2847    print "Count=$count\n";
2848
2849which produces:
2850
2851    aaab
2852    aaa
2853    aa
2854    a
2855    aab
2856    aa
2857    a
2858    ab
2859    a
2860    Count=9
2861
2862If we add a C<(*PRUNE)> before the count like the following
2863
2864    'aaab' =~ /a+b?(*PRUNE)(?{print "$&\n"; $count++})(*FAIL)/;
2865    print "Count=$count\n";
2866
2867we prevent backtracking and find the count of the longest matching string
2868at each matching starting point like so:
2869
2870    aaab
2871    aab
2872    ab
2873    Count=3
2874
2875Any number of C<(*PRUNE)> assertions may be used in a pattern.
2876
2877See also C<<< L<< /(?>I<pattern>) >> >>> and possessive quantifiers for
2878other ways to
2879control backtracking. In some cases, the use of C<(*PRUNE)> can be
2880replaced with a C<< (?>pattern) >> with no functional difference; however,
2881C<(*PRUNE)> can be used to handle cases that cannot be expressed using a
2882C<< (?>pattern) >> alone.
2883
2884=item C<(*SKIP)> C<(*SKIP:I<NAME>)>
2885X<(*SKIP)>
2886
2887This zero-width pattern is similar to C<(*PRUNE)>, except that on
2888failure it also signifies that whatever text that was matched leading up
2889to the C<(*SKIP)> pattern being executed cannot be part of I<any> match
2890of this pattern. This effectively means that the regex engine "skips" forward
2891to this position on failure and tries to match again, (assuming that
2892there is sufficient room to match).
2893
2894The name of the C<(*SKIP:I<NAME>)> pattern has special significance. If a
2895C<(*MARK:I<NAME>)> was encountered while matching, then it is that position
2896which is used as the "skip point". If no C<(*MARK)> of that name was
2897encountered, then the C<(*SKIP)> operator has no effect. When used
2898without a name the "skip point" is where the match point was when
2899executing the C<(*SKIP)> pattern.
2900
2901Compare the following to the examples in C<(*PRUNE)>; note the string
2902is twice as long:
2903
2904 'aaabaaab' =~ /a+b?(*SKIP)(?{print "$&\n"; $count++})(*FAIL)/;
2905 print "Count=$count\n";
2906
2907outputs
2908
2909    aaab
2910    aaab
2911    Count=2
2912
2913Once the 'aaab' at the start of the string has matched, and the C<(*SKIP)>
2914executed, the next starting point will be where the cursor was when the
2915C<(*SKIP)> was executed.
2916
2917=item C<(*MARK:I<NAME>)> C<(*:I<NAME>)>
2918X<(*MARK)> X<(*MARK:NAME)> X<(*:NAME)>
2919
2920This zero-width pattern can be used to mark the point reached in a string
2921when a certain part of the pattern has been successfully matched. This
2922mark may be given a name. A later C<(*SKIP)> pattern will then skip
2923forward to that point if backtracked into on failure. Any number of
2924C<(*MARK)> patterns are allowed, and the I<NAME> portion may be duplicated.
2925
2926In addition to interacting with the C<(*SKIP)> pattern, C<(*MARK:I<NAME>)>
2927can be used to "label" a pattern branch, so that after matching, the
2928program can determine which branches of the pattern were involved in the
2929match.
2930
2931When a match is successful, the C<$REGMARK> variable will be set to the
2932name of the most recently executed C<(*MARK:I<NAME>)> that was involved
2933in the match.
2934
2935This can be used to determine which branch of a pattern was matched
2936without using a separate capture group for each branch, which in turn
2937can result in a performance improvement, as perl cannot optimize
2938C</(?:(x)|(y)|(z))/> as efficiently as something like
2939C</(?:x(*MARK:x)|y(*MARK:y)|z(*MARK:z))/>.
2940
2941When a match has failed, and unless another verb has been involved in
2942failing the match and has provided its own name to use, the C<$REGERROR>
2943variable will be set to the name of the most recently executed
2944C<(*MARK:I<NAME>)>.
2945
2946See L</(*SKIP)> for more details.
2947
2948As a shortcut C<(*MARK:I<NAME>)> can be written C<(*:I<NAME>)>.
2949
2950=item C<(*THEN)> C<(*THEN:I<NAME>)>
2951
2952This is similar to the "cut group" operator C<::> from Raku.  Like
2953C<(*PRUNE)>, this verb always matches, and when backtracked into on
2954failure, it causes the regex engine to try the next alternation in the
2955innermost enclosing group (capturing or otherwise) that has alternations.
2956The two branches of a C<(?(I<condition>)I<yes-pattern>|I<no-pattern>)> do not
2957count as an alternation, as far as C<(*THEN)> is concerned.
2958
2959Its name comes from the observation that this operation combined with the
2960alternation operator (C<"|">) can be used to create what is essentially a
2961pattern-based if/then/else block:
2962
2963  ( COND (*THEN) FOO | COND2 (*THEN) BAR | COND3 (*THEN) BAZ )
2964
2965Note that if this operator is used and NOT inside of an alternation then
2966it acts exactly like the C<(*PRUNE)> operator.
2967
2968  / A (*PRUNE) B /
2969
2970is the same as
2971
2972  / A (*THEN) B /
2973
2974but
2975
2976  / ( A (*THEN) B | C ) /
2977
2978is not the same as
2979
2980  / ( A (*PRUNE) B | C ) /
2981
2982as after matching the I<A> but failing on the I<B> the C<(*THEN)> verb will
2983backtrack and try I<C>; but the C<(*PRUNE)> verb will simply fail.
2984
2985=item C<(*COMMIT)> C<(*COMMIT:I<arg>)>
2986X<(*COMMIT)>
2987
2988This is the Raku "commit pattern" C<< <commit> >> or C<:::>. It's a
2989zero-width pattern similar to C<(*SKIP)>, except that when backtracked
2990into on failure it causes the match to fail outright. No further attempts
2991to find a valid match by advancing the start pointer will occur again.
2992For example,
2993
2994 'aaabaaab' =~ /a+b?(*COMMIT)(?{print "$&\n"; $count++})(*FAIL)/;
2995 print "Count=$count\n";
2996
2997outputs
2998
2999    aaab
3000    Count=1
3001
3002In other words, once the C<(*COMMIT)> has been entered, and if the pattern
3003does not match, the regex engine will not try any further matching on the
3004rest of the string.
3005
3006=item C<(*FAIL)> C<(*F)> C<(*FAIL:I<arg>)>
3007X<(*FAIL)> X<(*F)>
3008
3009This pattern matches nothing and always fails. It can be used to force the
3010engine to backtrack. It is equivalent to C<(?!)>, but easier to read. In
3011fact, C<(?!)> gets optimised into C<(*FAIL)> internally. You can provide
3012an argument so that if the match fails because of this C<FAIL> directive
3013the argument can be obtained from C<$REGERROR>.
3014
3015It is probably useful only when combined with C<(?{})> or C<(??{})>.
3016
3017=item C<(*ACCEPT)> C<(*ACCEPT:I<arg>)>
3018X<(*ACCEPT)>
3019
3020This pattern matches nothing and causes the end of successful matching at
3021the point at which the C<(*ACCEPT)> pattern was encountered, regardless of
3022whether there is actually more to match in the string. When inside of a
3023nested pattern, such as recursion, or in a subpattern dynamically generated
3024via C<(??{})>, only the innermost pattern is ended immediately.
3025
3026If the C<(*ACCEPT)> is inside of capturing groups then the groups are
3027marked as ended at the point at which the C<(*ACCEPT)> was encountered.
3028For instance:
3029
3030  'AB' =~ /(A (A|B(*ACCEPT)|C) D)(E)/x;
3031
3032will match, and C<$1> will be C<AB> and C<$2> will be C<"B">, C<$3> will not
3033be set. If another branch in the inner parentheses was matched, such as in the
3034string 'ACDE', then the C<"D"> and C<"E"> would have to be matched as well.
3035
3036You can provide an argument, which will be available in the var
3037C<$REGMARK> after the match completes.
3038
3039=back
3040
3041=back
3042
3043=head2 Warning on C<\1> Instead of C<$1>
3044
3045Some people get too used to writing things like:
3046
3047    $pattern =~ s/(\W)/\\\1/g;
3048
3049This is grandfathered (for \1 to \9) for the RHS of a substitute to avoid
3050shocking the
3051B<sed> addicts, but it's a dirty habit to get into.  That's because in
3052PerlThink, the righthand side of an C<s///> is a double-quoted string.  C<\1> in
3053the usual double-quoted string means a control-A.  The customary Unix
3054meaning of C<\1> is kludged in for C<s///>.  However, if you get into the habit
3055of doing that, you get yourself into trouble if you then add an C</e>
3056modifier.
3057
3058    s/(\d+)/ \1 + 1 /eg;            # causes warning under -w
3059
3060Or if you try to do
3061
3062    s/(\d+)/\1000/;
3063
3064You can't disambiguate that by saying C<\{1}000>, whereas you can fix it with
3065C<${1}000>.  The operation of interpolation should not be confused
3066with the operation of matching a backreference.  Certainly they mean two
3067different things on the I<left> side of the C<s///>.
3068
3069=head2 Repeated Patterns Matching a Zero-length Substring
3070
3071B<WARNING>: Difficult material (and prose) ahead.  This section needs a rewrite.
3072
3073Regular expressions provide a terse and powerful programming language.  As
3074with most other power tools, power comes together with the ability
3075to wreak havoc.
3076
3077A common abuse of this power stems from the ability to make infinite
3078loops using regular expressions, with something as innocuous as:
3079
3080    'foo' =~ m{ ( o? )* }x;
3081
3082The C<o?> matches at the beginning of "C<foo>", and since the position
3083in the string is not moved by the match, C<o?> would match again and again
3084because of the C<"*"> quantifier.  Another common way to create a similar cycle
3085is with the looping modifier C</g>:
3086
3087    @matches = ( 'foo' =~ m{ o? }xg );
3088
3089or
3090
3091    print "match: <$&>\n" while 'foo' =~ m{ o? }xg;
3092
3093or the loop implied by C<split()>.
3094
3095However, long experience has shown that many programming tasks may
3096be significantly simplified by using repeated subexpressions that
3097may match zero-length substrings.  Here's a simple example being:
3098
3099    @chars = split //, $string;           # // is not magic in split
3100    ($whitewashed = $string) =~ s/()/ /g; # parens avoid magic s// /
3101
3102Thus Perl allows such constructs, by I<forcefully breaking
3103the infinite loop>.  The rules for this are different for lower-level
3104loops given by the greedy quantifiers C<*+{}>, and for higher-level
3105ones like the C</g> modifier or C<split()> operator.
3106
3107The lower-level loops are I<interrupted> (that is, the loop is
3108broken) when Perl detects that a repeated expression matched a
3109zero-length substring.   Thus
3110
3111   m{ (?: NON_ZERO_LENGTH | ZERO_LENGTH )* }x;
3112
3113is made equivalent to
3114
3115   m{ (?: NON_ZERO_LENGTH )* (?: ZERO_LENGTH )? }x;
3116
3117For example, this program
3118
3119   #!perl -l
3120   "aaaaab" =~ /
3121     (?:
3122        a                 # non-zero
3123        |                 # or
3124       (?{print "hello"}) # print hello whenever this
3125                          #    branch is tried
3126       (?=(b))            # zero-width assertion
3127     )*  # any number of times
3128    /x;
3129   print $&;
3130   print $1;
3131
3132prints
3133
3134   hello
3135   aaaaa
3136   b
3137
3138Notice that "hello" is only printed once, as when Perl sees that the sixth
3139iteration of the outermost C<(?:)*> matches a zero-length string, it stops
3140the C<"*">.
3141
3142The higher-level loops preserve an additional state between iterations:
3143whether the last match was zero-length.  To break the loop, the following
3144match after a zero-length match is prohibited to have a length of zero.
3145This prohibition interacts with backtracking (see L</"Backtracking">),
3146and so the I<second best> match is chosen if the I<best> match is of
3147zero length.
3148
3149For example:
3150
3151    $_ = 'bar';
3152    s/\w??/<$&>/g;
3153
3154results in C<< <><b><><a><><r><> >>.  At each position of the string the best
3155match given by non-greedy C<??> is the zero-length match, and the I<second
3156best> match is what is matched by C<\w>.  Thus zero-length matches
3157alternate with one-character-long matches.
3158
3159Similarly, for repeated C<m/()/g> the second-best match is the match at the
3160position one notch further in the string.
3161
3162The additional state of being I<matched with zero-length> is associated with
3163the matched string, and is reset by each assignment to C<pos()>.
3164Zero-length matches at the end of the previous match are ignored
3165during C<split>.
3166
3167=head2 Combining RE Pieces
3168
3169Each of the elementary pieces of regular expressions which were described
3170before (such as C<ab> or C<\Z>) could match at most one substring
3171at the given position of the input string.  However, in a typical regular
3172expression these elementary pieces are combined into more complicated
3173patterns using combining operators C<ST>, C<S|T>, C<S*> I<etc>.
3174(in these examples C<"S"> and C<"T"> are regular subexpressions).
3175
3176Such combinations can include alternatives, leading to a problem of choice:
3177if we match a regular expression C<a|ab> against C<"abc">, will it match
3178substring C<"a"> or C<"ab">?  One way to describe which substring is
3179actually matched is the concept of backtracking (see L</"Backtracking">).
3180However, this description is too low-level and makes you think
3181in terms of a particular implementation.
3182
3183Another description starts with notions of "better"/"worse".  All the
3184substrings which may be matched by the given regular expression can be
3185sorted from the "best" match to the "worst" match, and it is the "best"
3186match which is chosen.  This substitutes the question of "what is chosen?"
3187by the question of "which matches are better, and which are worse?".
3188
3189Again, for elementary pieces there is no such question, since at most
3190one match at a given position is possible.  This section describes the
3191notion of better/worse for combining operators.  In the description
3192below C<"S"> and C<"T"> are regular subexpressions.
3193
3194=over 4
3195
3196=item C<ST>
3197
3198Consider two possible matches, C<AB> and C<A'B'>, C<"A"> and C<A'> are
3199substrings which can be matched by C<"S">, C<"B"> and C<B'> are substrings
3200which can be matched by C<"T">.
3201
3202If C<"A"> is a better match for C<"S"> than C<A'>, C<AB> is a better
3203match than C<A'B'>.
3204
3205If C<"A"> and C<A'> coincide: C<AB> is a better match than C<AB'> if
3206C<"B"> is a better match for C<"T"> than C<B'>.
3207
3208=item C<S|T>
3209
3210When C<"S"> can match, it is a better match than when only C<"T"> can match.
3211
3212Ordering of two matches for C<"S"> is the same as for C<"S">.  Similar for
3213two matches for C<"T">.
3214
3215=item C<S{REPEAT_COUNT}>
3216
3217Matches as C<SSS...S> (repeated as many times as necessary).
3218
3219=item C<S{min,max}>
3220
3221Matches as C<S{max}|S{max-1}|...|S{min+1}|S{min}>.
3222
3223=item C<S{min,max}?>
3224
3225Matches as C<S{min}|S{min+1}|...|S{max-1}|S{max}>.
3226
3227=item C<S?>, C<S*>, C<S+>
3228
3229Same as C<S{0,1}>, C<S{0,BIG_NUMBER}>, C<S{1,BIG_NUMBER}> respectively.
3230
3231=item C<S??>, C<S*?>, C<S+?>
3232
3233Same as C<S{0,1}?>, C<S{0,BIG_NUMBER}?>, C<S{1,BIG_NUMBER}?> respectively.
3234
3235=item C<< (?>S) >>
3236
3237Matches the best match for C<"S"> and only that.
3238
3239=item C<(?=S)>, C<(?<=S)>
3240
3241Only the best match for C<"S"> is considered.  (This is important only if
3242C<"S"> has capturing parentheses, and backreferences are used somewhere
3243else in the whole regular expression.)
3244
3245=item C<(?!S)>, C<(?<!S)>
3246
3247For this grouping operator there is no need to describe the ordering, since
3248only whether or not C<"S"> can match is important.
3249
3250=item C<(??{ I<EXPR> })>, C<(?I<PARNO>)>
3251
3252The ordering is the same as for the regular expression which is
3253the result of I<EXPR>, or the pattern contained by capture group I<PARNO>.
3254
3255=item C<(?(I<condition>)I<yes-pattern>|I<no-pattern>)>
3256
3257Recall that which of I<yes-pattern> or I<no-pattern> actually matches is
3258already determined.  The ordering of the matches is the same as for the
3259chosen subexpression.
3260
3261=back
3262
3263The above recipes describe the ordering of matches I<at a given position>.
3264One more rule is needed to understand how a match is determined for the
3265whole regular expression: a match at an earlier position is always better
3266than a match at a later position.
3267
3268=head2 Creating Custom RE Engines
3269
3270As of Perl 5.10.0, one can create custom regular expression engines.  This
3271is not for the faint of heart, as they have to plug in at the C level.  See
3272L<perlreapi> for more details.
3273
3274As an alternative, overloaded constants (see L<overload>) provide a simple
3275way to extend the functionality of the RE engine, by substituting one
3276pattern for another.
3277
3278Suppose that we want to enable a new RE escape-sequence C<\Y|> which
3279matches at a boundary between whitespace characters and non-whitespace
3280characters.  Note that C<(?=\S)(?<!\S)|(?!\S)(?<=\S)> matches exactly
3281at these positions, so we want to have each C<\Y|> in the place of the
3282more complicated version.  We can create a module C<customre> to do
3283this:
3284
3285    package customre;
3286    use overload;
3287
3288    sub import {
3289      shift;
3290      die "No argument to customre::import allowed" if @_;
3291      overload::constant 'qr' => \&convert;
3292    }
3293
3294    sub invalid { die "/$_[0]/: invalid escape '\\$_[1]'"}
3295
3296    # We must also take care of not escaping the legitimate \\Y|
3297    # sequence, hence the presence of '\\' in the conversion rules.
3298    my %rules = ( '\\' => '\\\\',
3299                  'Y|' => qr/(?=\S)(?<!\S)|(?!\S)(?<=\S)/ );
3300    sub convert {
3301      my $re = shift;
3302      $re =~ s{
3303                \\ ( \\ | Y . )
3304              }
3305              { $rules{$1} or invalid($re,$1) }sgex;
3306      return $re;
3307    }
3308
3309Now C<use customre> enables the new escape in constant regular
3310expressions, I<i.e.>, those without any runtime variable interpolations.
3311As documented in L<overload>, this conversion will work only over
3312literal parts of regular expressions.  For C<\Y|$re\Y|> the variable
3313part of this regular expression needs to be converted explicitly
3314(but only if the special meaning of C<\Y|> should be enabled inside C<$re>):
3315
3316    use customre;
3317    $re = <>;
3318    chomp $re;
3319    $re = customre::convert $re;
3320    /\Y|$re\Y|/;
3321
3322=head2 Embedded Code Execution Frequency
3323
3324The exact rules for how often C<(?{})> and C<(??{})> are executed in a pattern
3325are unspecified, and this is even more true of C<(*{})>.
3326In the case of a successful match you can assume that they DWIM and
3327will be executed in left to right order the appropriate number of times in the
3328accepting path of the pattern as would any other meta-pattern. How non-
3329accepting pathways and match failures affect the number of times a pattern is
3330executed is specifically unspecified and may vary depending on what
3331optimizations can be applied to the pattern and is likely to change from
3332version to version.
3333
3334For instance in
3335
3336  "aaabcdeeeee"=~/a(?{print "a"})b(?{print "b"})cde/;
3337
3338the exact number of times "a" or "b" are printed out is unspecified for
3339failure, but you may assume they will be printed at least once during
3340a successful match, additionally you may assume that if "b" is printed,
3341it will be preceded by at least one "a".
3342
3343In the case of branching constructs like the following:
3344
3345  /a(b|(?{ print "a" }))c(?{ print "c" })/;
3346
3347you can assume that the input "ac" will output "ac", and that "abc"
3348will output only "c".
3349
3350When embedded code is quantified, successful matches will call the
3351code once for each matched iteration of the quantifier.  For
3352example:
3353
3354  "good" =~ /g(?:o(?{print "o"}))*d/;
3355
3356will output "o" twice.
3357
3358For historical and consistency reasons the use of normal code blocks
3359anywhere in a pattern will disable certain optimisations. As of 5.37.7
3360you can use an "optimistic" codeblock, C<(*{ ... })> as a replacement
3361for C<(?{ ... })>, if you do *not* wish to disable these optimisations.
3362This may result in the code block being called less often than it might
3363have been had they not been optimistic.
3364
3365=head2 PCRE/Python Support
3366
3367As of Perl 5.10.0, Perl supports several Python/PCRE-specific extensions
3368to the regex syntax. While Perl programmers are encouraged to use the
3369Perl-specific syntax, the following are also accepted:
3370
3371=over 4
3372
3373=item C<< (?PE<lt>I<NAME>E<gt>I<pattern>) >>
3374
3375Define a named capture group. Equivalent to C<< (?<I<NAME>>I<pattern>) >>.
3376
3377=item C<< (?P=I<NAME>) >>
3378
3379Backreference to a named capture group. Equivalent to C<< \g{I<NAME>} >>.
3380
3381=item C<< (?P>I<NAME>) >>
3382
3383Subroutine call to a named capture group. Equivalent to C<< (?&I<NAME>) >>.
3384
3385=back
3386
3387=head1 BUGS
3388
3389There are a number of issues with regard to case-insensitive matching
3390in Unicode rules.  See C<"i"> under L</Modifiers> above.
3391
3392This document varies from difficult to understand to completely
3393and utterly opaque.  The wandering prose riddled with jargon is
3394hard to fathom in several places.
3395
3396This document needs a rewrite that separates the tutorial content
3397from the reference content.
3398
3399=head1 SEE ALSO
3400
3401The syntax of patterns used in Perl pattern matching evolved from those
3402supplied in the Bell Labs Research Unix 8th Edition (Version 8) regex
3403routines.  (The code is actually derived (distantly) from Henry
3404Spencer's freely redistributable reimplementation of those V8 routines.)
3405
3406L<perlrequick>.
3407
3408L<perlretut>.
3409
3410L<perlop/"Regexp Quote-Like Operators">.
3411
3412L<perlop/"Gory details of parsing quoted constructs">.
3413
3414L<perlfaq6>.
3415
3416L<perlfunc/pos>.
3417
3418L<perllocale>.
3419
3420L<perlebcdic>.
3421
3422I<Mastering Regular Expressions> by Jeffrey Friedl, published
3423by O'Reilly and Associates.
3424