2 X<reference> X<pointer> X<data structure> X<structure> X<struct>
4 perlref - Perl references and nested data structures
8 This is complete documentation about all aspects of references.
9 For a shorter, tutorial introduction to just the essential features,
14 Before release 5 of Perl it was difficult to represent complex data
15 structures, because all references had to be symbolic--and even then
16 it was difficult to refer to a variable instead of a symbol table entry.
17 Perl now not only makes it easier to use symbolic references to variables,
18 but also lets you have "hard" references to any piece of data or code.
19 Any scalar may hold a hard reference. Because arrays and hashes contain
20 scalars, you can now easily build arrays of arrays, arrays of hashes,
21 hashes of arrays, arrays of hashes of functions, and so on.
23 Hard references are smart--they keep track of reference counts for you,
24 automatically freeing the thing referred to when its reference count goes
25 to zero. (Reference counts for values in self-referential or
26 cyclic data structures may not go to zero without a little help; see
27 L</"Circular References"> for a detailed explanation.)
28 If that thing happens to be an object, the object is destructed. See
29 L<perlobj> for more about objects. (In a sense, everything in Perl is an
30 object, but we usually reserve the word for references to objects that
31 have been officially "blessed" into a class package.)
33 Symbolic references are names of variables or other objects, just as a
34 symbolic link in a Unix filesystem contains merely the name of a file.
35 The C<*glob> notation is something of a symbolic reference. (Symbolic
36 references are sometimes called "soft references", but please don't call
37 them that; references are confusing enough without useless synonyms.)
38 X<reference, symbolic> X<reference, soft>
39 X<symbolic reference> X<soft reference>
41 In contrast, hard references are more like hard links in a Unix file
42 system: They are used to access an underlying object without concern for
43 what its (other) name is. When the word "reference" is used without an
44 adjective, as in the following paragraph, it is usually talking about a
46 X<reference, hard> X<hard reference>
48 References are easy to use in Perl. There is just one overriding
49 principle: in general, Perl does no implicit referencing or dereferencing.
50 When a scalar is holding a reference, it always behaves as a simple scalar.
51 It doesn't magically start being an array or hash or subroutine; you have to
52 tell it explicitly to do so, by dereferencing it.
54 That said, be aware that Perl version 5.14 introduces an exception
55 to the rule, for syntactic convenience. Experimental array and hash container
56 function behavior allows array and hash references to be handled by Perl as
57 if they had been explicitly syntactically dereferenced. See
58 L<perl5140delta/"Syntactical Enhancements">
59 and L<perlfunc> for details.
61 =head2 Making References
62 X<reference, creation> X<referencing>
64 References can be created in several ways.
71 By using the backslash operator on a variable, subroutine, or value.
72 (This works much like the & (address-of) operator in C.)
73 This typically creates I<another> reference to a variable, because
74 there's already a reference to the variable in the symbol table. But
75 the symbol table reference might go away, and you'll still have the
76 reference that the backslash returned. Here are some examples:
84 It isn't possible to create a true reference to an IO handle (filehandle
85 or dirhandle) using the backslash operator. The most you can get is a
86 reference to a typeglob, which is actually a complete symbol table entry.
87 But see the explanation of the C<*foo{THING}> syntax below. However,
88 you can still use type globs and globrefs as though they were IO handles.
91 X<array, anonymous> X<[> X<[]> X<square bracket>
92 X<bracket, square> X<arrayref> X<array reference> X<reference, array>
94 A reference to an anonymous array can be created using square
97 $arrayref = [1, 2, ['a', 'b', 'c']];
99 Here we've created a reference to an anonymous array of three elements
100 whose final element is itself a reference to another anonymous array of three
101 elements. (The multidimensional syntax described later can be used to
102 access this. For example, after the above, C<< $arrayref->[2][1] >> would have
105 Taking a reference to an enumerated list is not the same
106 as using square brackets--instead it's the same as creating
107 a list of references!
109 @list = (\$a, \@b, \%c);
110 @list = \($a, @b, %c); # same thing!
112 As a special case, C<\(@foo)> returns a list of references to the contents
113 of C<@foo>, not a reference to C<@foo> itself. Likewise for C<%foo>,
114 except that the key references are to copies (since the keys are just
115 strings rather than full-fledged scalars).
118 X<hash, anonymous> X<{> X<{}> X<curly bracket>
119 X<bracket, curly> X<brace> X<hashref> X<hash reference> X<reference, hash>
121 A reference to an anonymous hash can be created using curly
129 Anonymous hash and array composers like these can be intermixed freely to
130 produce as complicated a structure as you want. The multidimensional
131 syntax described below works for these too. The values above are
132 literals, but variables and expressions would work just as well, because
133 assignment operators in Perl (even within local() or my()) are executable
134 statements, not compile-time declarations.
136 Because curly brackets (braces) are used for several other things
137 including BLOCKs, you may occasionally have to disambiguate braces at the
138 beginning of a statement by putting a C<+> or a C<return> in front so
139 that Perl realizes the opening brace isn't starting a BLOCK. The economy and
140 mnemonic value of using curlies is deemed worth this occasional extra
143 For example, if you wanted a function to make a new hash and return a
144 reference to it, you have these options:
146 sub hashem { { @_ } } # silently wrong
147 sub hashem { +{ @_ } } # ok
148 sub hashem { return { @_ } } # ok
150 On the other hand, if you want the other meaning, you can do this:
152 sub showem { { @_ } } # ambiguous (currently ok, but may change)
153 sub showem { {; @_ } } # ok
154 sub showem { { return @_ } } # ok
156 The leading C<+{> and C<{;> always serve to disambiguate
157 the expression to mean either the HASH reference, or the BLOCK.
160 X<subroutine, anonymous> X<subroutine, reference> X<reference, subroutine>
161 X<scope, lexical> X<closure> X<lexical> X<lexical scope>
163 A reference to an anonymous subroutine can be created by using
164 C<sub> without a subname:
166 $coderef = sub { print "Boink!\n" };
168 Note the semicolon. Except for the code
169 inside not being immediately executed, a C<sub {}> is not so much a
170 declaration as it is an operator, like C<do{}> or C<eval{}>. (However, no
171 matter how many times you execute that particular line (unless you're in an
172 C<eval("...")>), $coderef will still have a reference to the I<same>
173 anonymous subroutine.)
175 Anonymous subroutines act as closures with respect to my() variables,
176 that is, variables lexically visible within the current scope. Closure
177 is a notion out of the Lisp world that says if you define an anonymous
178 function in a particular lexical context, it pretends to run in that
179 context even when it's called outside the context.
181 In human terms, it's a funny way of passing arguments to a subroutine when
182 you define it as well as when you call it. It's useful for setting up
183 little bits of code to run later, such as callbacks. You can even
184 do object-oriented stuff with it, though Perl already provides a different
185 mechanism to do that--see L<perlobj>.
187 You might also think of closure as a way to write a subroutine
188 template without using eval(). Here's a small example of how
193 return sub { my $y = shift; print "$x, $y!\n"; };
195 $h = newprint("Howdy");
196 $g = newprint("Greetings");
206 Greetings, earthlings!
208 Note particularly that $x continues to refer to the value passed
209 into newprint() I<despite> "my $x" having gone out of scope by the
210 time the anonymous subroutine runs. That's what a closure is all
213 This applies only to lexical variables, by the way. Dynamic variables
214 continue to work as they have always worked. Closure is not something
215 that most Perl programmers need trouble themselves about to begin with.
218 X<constructor> X<new>
220 References are often returned by special subroutines called constructors. Perl
221 objects are just references to a special type of object that happens to know
222 which package it's associated with. Constructors are just special subroutines
223 that know how to create that association. They do so by starting with an
224 ordinary reference, and it remains an ordinary reference even while it's also
225 being an object. Constructors are often named C<new()>. You I<can> call them
228 $objref = new Doggie( Tail => 'short', Ears => 'long' );
230 But that can produce ambiguous syntax in certain cases, so it's often
231 better to use the direct method invocation approach:
233 $objref = Doggie->new(Tail => 'short', Ears => 'long');
236 $terminal = Term::Cap->Tgetent( { OSPEED => 9600 });
239 $main = MainWindow->new();
240 $menubar = $main->Frame(-relief => "raised",
246 References of the appropriate type can spring into existence if you
247 dereference them in a context that assumes they exist. Because we haven't
248 talked about dereferencing yet, we can't show you any examples yet.
253 A reference can be created by using a special syntax, lovingly known as
254 the *foo{THING} syntax. *foo{THING} returns a reference to the THING
255 slot in *foo (which is the symbol table entry which holds everything
258 $scalarref = *foo{SCALAR};
259 $arrayref = *ARGV{ARRAY};
260 $hashref = *ENV{HASH};
261 $coderef = *handler{CODE};
263 $globref = *foo{GLOB};
264 $formatref = *foo{FORMAT};
265 $globname = *foo{NAME}; # "foo"
266 $pkgname = *foo{PACKAGE}; # "main"
268 Most of these are self-explanatory, but C<*foo{IO}>
269 deserves special attention. It returns
270 the IO handle, used for file handles (L<perlfunc/open>), sockets
271 (L<perlfunc/socket> and L<perlfunc/socketpair>), and directory
272 handles (L<perlfunc/opendir>). For compatibility with previous
273 versions of Perl, C<*foo{FILEHANDLE}> is a synonym for C<*foo{IO}>, though it
274 is deprecated as of 5.8.0. If deprecation warnings are in effect, it will warn
277 C<*foo{THING}> returns undef if that particular THING hasn't been used yet,
278 except in the case of scalars. C<*foo{SCALAR}> returns a reference to an
279 anonymous scalar if $foo hasn't been used yet. This might change in a
282 C<*foo{NAME}> and C<*foo{PACKAGE}> are the exception, in that they return
283 strings, rather than references. These return the package and name of the
284 typeglob itself, rather than one that has been assigned to it. So, after
285 C<*foo=*Foo::bar>, C<*foo> will become "*Foo::bar" when used as a string,
286 but C<*foo{PACKAGE}> and C<*foo{NAME}> will continue to produce "main" and
289 C<*foo{IO}> is an alternative to the C<*HANDLE> mechanism given in
290 L<perldata/"Typeglobs and Filehandles"> for passing filehandles
291 into or out of subroutines, or storing into larger data structures.
292 Its disadvantage is that it won't create a new filehandle for you.
293 Its advantage is that you have less risk of clobbering more than
294 you want to with a typeglob assignment. (It still conflates file
295 and directory handles, though.) However, if you assign the incoming
296 value to a scalar instead of a typeglob as we do in the examples
297 below, there's no risk of that happening.
299 splutter(*STDOUT); # pass the whole glob
300 splutter(*STDOUT{IO}); # pass both file and dir handles
304 print $fh "her um well a hmmm\n";
307 $rec = get_rec(*STDIN); # pass the whole glob
308 $rec = get_rec(*STDIN{IO}); # pass both file and dir handles
317 =head2 Using References
318 X<reference, use> X<dereferencing> X<dereference>
320 That's it for creating references. By now you're probably dying to
321 know how to use references to get back to your long-lost data. There
322 are several basic methods.
328 Anywhere you'd put an identifier (or chain of identifiers) as part
329 of a variable or subroutine name, you can replace the identifier with
330 a simple scalar variable containing a reference of the correct type:
333 push(@$arrayref, $filename);
334 $$arrayref[0] = "January";
335 $$hashref{"KEY"} = "VALUE";
337 print $globref "output\n";
339 It's important to understand that we are specifically I<not> dereferencing
340 C<$arrayref[0]> or C<$hashref{"KEY"}> there. The dereference of the
341 scalar variable happens I<before> it does any key lookups. Anything more
342 complicated than a simple scalar variable must use methods 2 or 3 below.
343 However, a "simple scalar" includes an identifier that itself uses method
344 1 recursively. Therefore, the following prints "howdy".
346 $refrefref = \\\"howdy";
351 Anywhere you'd put an identifier (or chain of identifiers) as part of a
352 variable or subroutine name, you can replace the identifier with a
353 BLOCK returning a reference of the correct type. In other words, the
354 previous examples could be written like this:
356 $bar = ${$scalarref};
357 push(@{$arrayref}, $filename);
358 ${$arrayref}[0] = "January";
359 ${$hashref}{"KEY"} = "VALUE";
361 $globref->print("output\n"); # iff IO::Handle is loaded
363 Admittedly, it's a little silly to use the curlies in this case, but
364 the BLOCK can contain any arbitrary expression, in particular,
365 subscripted expressions:
367 &{ $dispatch{$index} }(1,2,3); # call correct routine
369 Because of being able to omit the curlies for the simple case of C<$$x>,
370 people often make the mistake of viewing the dereferencing symbols as
371 proper operators, and wonder about their precedence. If they were,
372 though, you could use parentheses instead of braces. That's not the case.
373 Consider the difference below; case 0 is a short-hand version of case 1,
376 $$hashref{"KEY"} = "VALUE"; # CASE 0
377 ${$hashref}{"KEY"} = "VALUE"; # CASE 1
378 ${$hashref{"KEY"}} = "VALUE"; # CASE 2
379 ${$hashref->{"KEY"}} = "VALUE"; # CASE 3
381 Case 2 is also deceptive in that you're accessing a variable
382 called %hashref, not dereferencing through $hashref to the hash
383 it's presumably referencing. That would be case 3.
387 Subroutine calls and lookups of individual array elements arise often
388 enough that it gets cumbersome to use method 2. As a form of
389 syntactic sugar, the examples for method 2 may be written:
391 $arrayref->[0] = "January"; # Array element
392 $hashref->{"KEY"} = "VALUE"; # Hash element
393 $coderef->(1,2,3); # Subroutine call
395 The left side of the arrow can be any expression returning a reference,
396 including a previous dereference. Note that C<$array[$x]> is I<not> the
397 same thing as C<< $array->[$x] >> here:
399 $array[$x]->{"foo"}->[0] = "January";
401 This is one of the cases we mentioned earlier in which references could
402 spring into existence when in an lvalue context. Before this
403 statement, C<$array[$x]> may have been undefined. If so, it's
404 automatically defined with a hash reference so that we can look up
405 C<{"foo"}> in it. Likewise C<< $array[$x]->{"foo"} >> will automatically get
406 defined with an array reference so that we can look up C<[0]> in it.
407 This process is called I<autovivification>.
409 One more thing here. The arrow is optional I<between> brackets
410 subscripts, so you can shrink the above down to
412 $array[$x]{"foo"}[0] = "January";
414 Which, in the degenerate case of using only ordinary arrays, gives you
415 multidimensional arrays just like C's:
417 $score[$x][$y][$z] += 42;
419 Well, okay, not entirely like C's arrays, actually. C doesn't know how
420 to grow its arrays on demand. Perl does.
424 If a reference happens to be a reference to an object, then there are
425 probably methods to access the things referred to, and you should probably
426 stick to those methods unless you're in the class package that defines the
427 object's methods. In other words, be nice, and don't violate the object's
428 encapsulation without a very good reason. Perl does not enforce
429 encapsulation. We are not totalitarians here. We do expect some basic
434 Using a string or number as a reference produces a symbolic reference,
435 as explained above. Using a reference as a number produces an
436 integer representing its storage location in memory. The only
437 useful thing to be done with this is to compare two references
438 numerically to see whether they refer to the same location.
439 X<reference, numeric context>
441 if ($ref1 == $ref2) { # cheap numeric compare of references
442 print "refs 1 and 2 refer to the same thing\n";
445 Using a reference as a string produces both its referent's type,
446 including any package blessing as described in L<perlobj>, as well
447 as the numeric address expressed in hex. The ref() operator returns
448 just the type of thing the reference is pointing to, without the
449 address. See L<perlfunc/ref> for details and examples of its use.
450 X<reference, string context>
452 The bless() operator may be used to associate the object a reference
453 points to with a package functioning as an object class. See L<perlobj>.
455 A typeglob may be dereferenced the same way a reference can, because
456 the dereference syntax always indicates the type of reference desired.
457 So C<${*foo}> and C<${\$foo}> both indicate the same scalar variable.
459 Here's a trick for interpolating a subroutine call into a string:
461 print "My sub returned @{[mysub(1,2,3)]} that time.\n";
463 The way it works is that when the C<@{...}> is seen in the double-quoted
464 string, it's evaluated as a block. The block creates a reference to an
465 anonymous array containing the results of the call to C<mysub(1,2,3)>. So
466 the whole block returns a reference to an array, which is then
467 dereferenced by C<@{...}> and stuck into the double-quoted string. This
468 chicanery is also useful for arbitrary expressions:
470 print "That yields @{[$n + 5]} widgets\n";
472 Similarly, an expression that returns a reference to a scalar can be
473 dereferenced via C<${...}>. Thus, the above expression may be written
476 print "That yields ${\($n + 5)} widgets\n";
478 =head2 Circular References
479 X<circular reference> X<reference, circular>
481 It is possible to create a "circular reference" in Perl, which can lead
482 to memory leaks. A circular reference occurs when two references
483 contain a reference to each other, like this:
486 my $bar = { foo => $foo };
489 You can also create a circular reference with a single variable:
494 In this case, the reference count for the variables will never reach 0,
495 and the references will never be garbage-collected. This can lead to
498 Because objects in Perl are implemented as references, it's possible to
499 have circular references with objects as well. Imagine a TreeNode class
500 where each node references its parent and child nodes. Any node with a
501 parent will be part of a circular reference.
503 You can break circular references by creating a "weak reference". A
504 weak reference does not increment the reference count for a variable,
505 which means that the object can go out of scope and be destroyed. You
506 can weaken a reference with the C<weaken> function exported by the
507 L<Scalar::Util> module.
509 Here's how we can make the first example safer:
511 use Scalar::Util 'weaken';
514 my $bar = { foo => $foo };
519 The reference from C<$foo> to C<$bar> has been weakened. When the
520 C<$bar> variable goes out of scope, it will be garbage-collected. The
521 next time you look at the value of the C<< $foo->{bar} >> key, it will
524 This action at a distance can be confusing, so you should be careful
525 with your use of weaken. You should weaken the reference in the
526 variable that will go out of scope I<first>. That way, the longer-lived
527 variable will contain the expected reference until it goes out of
530 =head2 Symbolic references
531 X<reference, symbolic> X<reference, soft>
532 X<symbolic reference> X<soft reference>
534 We said that references spring into existence as necessary if they are
535 undefined, but we didn't say what happens if a value used as a
536 reference is already defined, but I<isn't> a hard reference. If you
537 use it as a reference, it'll be treated as a symbolic
538 reference. That is, the value of the scalar is taken to be the I<name>
539 of a variable, rather than a direct link to a (possibly) anonymous
542 People frequently expect it to work like this. So it does.
545 $$name = 1; # Sets $foo
546 ${$name} = 2; # Sets $foo
547 ${$name x 2} = 3; # Sets $foofoo
548 $name->[0] = 4; # Sets $foo[0]
549 @$name = (); # Clears @foo
550 &$name(); # Calls &foo()
552 ${"${pack}::$name"} = 5; # Sets $THAT::foo without eval
554 This is powerful, and slightly dangerous, in that it's possible
555 to intend (with the utmost sincerity) to use a hard reference, and
556 accidentally use a symbolic reference instead. To protect against
561 and then only hard references will be allowed for the rest of the enclosing
562 block. An inner block may countermand that with
566 Only package variables (globals, even if localized) are visible to
567 symbolic references. Lexical variables (declared with my()) aren't in
568 a symbol table, and thus are invisible to this mechanism. For example:
577 This will still print 10, not 20. Remember that local() affects package
578 variables, which are all "global" to the package.
580 =head2 Not-so-symbolic references
582 Brackets around a symbolic reference can simply
583 serve to isolate an identifier or variable name from the rest of an
584 expression, just as they always have within a string. For example,
589 has always meant to print "pop on over", even though push is
590 a reserved word. This is generalized to work the same
591 without the enclosing double quotes, so that
593 print ${push} . "over";
597 print ${ push } . "over";
599 will have the same effect. This
600 construct is I<not> considered to be a symbolic reference when you're
604 ${ bareword }; # Okay, means $bareword.
605 ${ "bareword" }; # Error, symbolic reference.
607 Similarly, because of all the subscripting that is done using single words,
608 the same rule applies to any bareword that is used for subscripting a hash.
609 So now, instead of writing
611 $array{ "aaa" }{ "bbb" }{ "ccc" }
615 $array{ aaa }{ bbb }{ ccc }
617 and not worry about whether the subscripts are reserved words. In the
618 rare event that you do wish to do something like
622 you can force interpretation as a reserved word by adding anything that
623 makes it more than a bareword:
629 The C<use warnings> pragma or the B<-w> switch will warn you if it
630 interprets a reserved word as a string.
631 But it will no longer warn you about using lowercase words, because the
632 string is effectively quoted.
634 =head2 Pseudo-hashes: Using an array as a hash
635 X<pseudo-hash> X<pseudo hash> X<pseudohash>
637 Pseudo-hashes have been removed from Perl. The 'fields' pragma
640 =head2 Function Templates
641 X<scope, lexical> X<closure> X<lexical> X<lexical scope>
642 X<subroutine, nested> X<sub, nested> X<subroutine, local> X<sub, local>
644 As explained above, an anonymous function with access to the lexical
645 variables visible when that function was compiled, creates a closure. It
646 retains access to those variables even though it doesn't get run until
647 later, such as in a signal handler or a Tk callback.
649 Using a closure as a function template allows us to generate many functions
650 that act similarly. Suppose you wanted functions named after the colors
651 that generated HTML font changes for the various colors:
653 print "Be ", red("careful"), "with that ", green("light");
655 The red() and green() functions would be similar. To create these,
656 we'll assign a closure to a typeglob of the name of the function we're
659 @colors = qw(red blue green yellow orange purple violet);
660 for my $name (@colors) {
661 no strict 'refs'; # allow symbol table manipulation
662 *$name = *{uc $name} = sub { "<FONT COLOR='$name'>@_</FONT>" };
665 Now all those different functions appear to exist independently. You can
666 call red(), RED(), blue(), BLUE(), green(), etc. This technique saves on
667 both compile time and memory use, and is less error-prone as well, since
668 syntax checks happen at compile time. It's critical that any variables in
669 the anonymous subroutine be lexicals in order to create a proper closure.
670 That's the reasons for the C<my> on the loop iteration variable.
672 This is one of the only places where giving a prototype to a closure makes
673 much sense. If you wanted to impose scalar context on the arguments of
674 these functions (probably not a wise idea for this particular example),
675 you could have written it this way instead:
677 *$name = sub ($) { "<FONT COLOR='$name'>$_[0]</FONT>" };
679 However, since prototype checking happens at compile time, the assignment
680 above happens too late to be of much use. You could address this by
681 putting the whole loop of assignments within a BEGIN block, forcing it
682 to occur during compilation.
684 Access to lexicals that change over time--like those in the C<for> loop
685 above, basically aliases to elements from the surrounding lexical scopes--
686 only works with anonymous subs, not with named subroutines. Generally
687 said, named subroutines do not nest properly and should only be declared
688 in the main package scope.
690 This is because named subroutines are created at compile time so their
691 lexical variables get assigned to the parent lexicals from the first
692 execution of the parent block. If a parent scope is entered a second
693 time, its lexicals are created again, while the nested subs still
694 reference the old ones.
696 Anonymous subroutines get to capture each time you execute the C<sub>
697 operator, as they are created on the fly. If you are accustomed to using
698 nested subroutines in other programming languages with their own private
699 variables, you'll have to work at it a bit in Perl. The intuitive coding
700 of this type of thing incurs mysterious warnings about "will not stay
701 shared" due to the reasons explained above.
702 For example, this won't work:
706 sub inner { return $x * 19 } # WRONG
710 A work-around is the following:
714 local *inner = sub { return $x * 19 };
718 Now inner() can only be called from within outer(), because of the
719 temporary assignments of the anonymous subroutine. But when it does,
720 it has normal access to the lexical variable $x from the scope of
721 outer() at the time outer is invoked.
723 This has the interesting effect of creating a function local to another
724 function, something not normally supported in Perl.
727 X<reference, string context> X<reference, use as hash key>
729 You may not (usefully) use a reference as the key to a hash. It will be
730 converted into a string:
734 If you try to dereference the key, it won't do a hard dereference, and
735 you won't accomplish what you're attempting. You might want to do something
741 And then at least you can use the values(), which will be
742 real refs, instead of the keys(), which won't.
744 The standard Tie::RefHash module provides a convenient workaround to this.
748 Besides the obvious documents, source code can be instructive.
749 Some pathological examples of the use of references can be found
750 in the F<t/op/ref.t> regression test in the Perl source directory.
752 See also L<perldsc> and L<perllol> for how to use references to create
753 complex data structures, and L<perlootut> and L<perlobj>
754 for how to use them to create objects.