This is a live mirror of the Perl 5 development currently hosted at https://github.com/perl/perl5
Duh.
[perl5.git] / lib / overload.pm
... / ...
CommitLineData
1package overload;
2
3our $VERSION = '1.00';
4
5$overload::hint_bits = 0x20000;
6
7sub nil {}
8
9sub OVERLOAD {
10 $package = shift;
11 my %arg = @_;
12 my ($sub, $fb);
13 $ {$package . "::OVERLOAD"}{dummy}++; # Register with magic by touching.
14 *{$package . "::()"} = \&nil; # Make it findable via fetchmethod.
15 for (keys %arg) {
16 if ($_ eq 'fallback') {
17 $fb = $arg{$_};
18 } else {
19 $sub = $arg{$_};
20 if (not ref $sub and $sub !~ /::/) {
21 $ {$package . "::(" . $_} = $sub;
22 $sub = \&nil;
23 }
24 #print STDERR "Setting `$ {'package'}::\cO$_' to \\&`$sub'.\n";
25 *{$package . "::(" . $_} = \&{ $sub };
26 }
27 }
28 ${$package . "::()"} = $fb; # Make it findable too (fallback only).
29}
30
31sub import {
32 $package = (caller())[0];
33 # *{$package . "::OVERLOAD"} = \&OVERLOAD;
34 shift;
35 $package->overload::OVERLOAD(@_);
36}
37
38sub unimport {
39 $package = (caller())[0];
40 ${$package . "::OVERLOAD"}{dummy}++; # Upgrade the table
41 shift;
42 for (@_) {
43 if ($_ eq 'fallback') {
44 undef $ {$package . "::()"};
45 } else {
46 delete $ {$package . "::"}{"(" . $_};
47 }
48 }
49}
50
51sub Overloaded {
52 my $package = shift;
53 $package = ref $package if ref $package;
54 $package->can('()');
55}
56
57sub ov_method {
58 my $globref = shift;
59 return undef unless $globref;
60 my $sub = \&{*$globref};
61 return $sub if $sub ne \&nil;
62 return shift->can($ {*$globref});
63}
64
65sub OverloadedStringify {
66 my $package = shift;
67 $package = ref $package if ref $package;
68 #$package->can('(""')
69 ov_method mycan($package, '(""'), $package
70 or ov_method mycan($package, '(0+'), $package
71 or ov_method mycan($package, '(bool'), $package
72 or ov_method mycan($package, '(nomethod'), $package;
73}
74
75sub Method {
76 my $package = shift;
77 $package = ref $package if ref $package;
78 #my $meth = $package->can('(' . shift);
79 ov_method mycan($package, '(' . shift), $package;
80 #return $meth if $meth ne \&nil;
81 #return $ {*{$meth}};
82}
83
84sub AddrRef {
85 my $package = ref $_[0];
86 return "$_[0]" unless $package;
87 bless $_[0], overload::Fake; # Non-overloaded package
88 my $str = "$_[0]";
89 bless $_[0], $package; # Back
90 $package . substr $str, index $str, '=';
91}
92
93sub StrVal {
94 (OverloadedStringify($_[0]) or ref($_[0]) eq 'Regexp') ?
95 (AddrRef(shift)) :
96 "$_[0]";
97}
98
99sub mycan { # Real can would leave stubs.
100 my ($package, $meth) = @_;
101 return \*{$package . "::$meth"} if defined &{$package . "::$meth"};
102 my $p;
103 foreach $p (@{$package . "::ISA"}) {
104 my $out = mycan($p, $meth);
105 return $out if $out;
106 }
107 return undef;
108}
109
110%constants = (
111 'integer' => 0x1000,
112 'float' => 0x2000,
113 'binary' => 0x4000,
114 'q' => 0x8000,
115 'qr' => 0x10000,
116 );
117
118%ops = ( with_assign => "+ - * / % ** << >> x .",
119 assign => "+= -= *= /= %= **= <<= >>= x= .=",
120 num_comparison => "< <= > >= == !=",
121 '3way_comparison'=> "<=> cmp",
122 str_comparison => "lt le gt ge eq ne",
123 binary => "& | ^",
124 unary => "neg ! ~",
125 mutators => '++ --',
126 func => "atan2 cos sin exp abs log sqrt int",
127 conversion => 'bool "" 0+',
128 iterators => '<>',
129 dereferencing => '${} @{} %{} &{} *{}',
130 special => 'nomethod fallback =');
131
132use warnings::register;
133sub constant {
134 # Arguments: what, sub
135 while (@_) {
136 if (@_ == 1) {
137 warnings::warnif ("Odd number of arguments for overload::constant");
138 last;
139 }
140 elsif (!exists $constants {$_ [0]}) {
141 warnings::warnif ("`$_[0]' is not an overloadable type");
142 }
143 elsif (!ref $_ [1] || "$_[1]" !~ /CODE\(0x[\da-f]+\)$/) {
144 # Can't use C<ref $_[1] eq "CODE"> above as code references can be
145 # blessed, and C<ref> would return the package the ref is blessed into.
146 if (warnings::enabled) {
147 $_ [1] = "undef" unless defined $_ [1];
148 warnings::warn ("`$_[1]' is not a code reference");
149 }
150 }
151 else {
152 $^H{$_[0]} = $_[1];
153 $^H |= $constants{$_[0]} | $overload::hint_bits;
154 }
155 shift, shift;
156 }
157}
158
159sub remove_constant {
160 # Arguments: what, sub
161 while (@_) {
162 delete $^H{$_[0]};
163 $^H &= ~ $constants{$_[0]};
164 shift, shift;
165 }
166}
167
1681;
169
170__END__
171
172=head1 NAME
173
174overload - Package for overloading perl operations
175
176=head1 SYNOPSIS
177
178 package SomeThing;
179
180 use overload
181 '+' => \&myadd,
182 '-' => \&mysub;
183 # etc
184 ...
185
186 package main;
187 $a = new SomeThing 57;
188 $b=5+$a;
189 ...
190 if (overload::Overloaded $b) {...}
191 ...
192 $strval = overload::StrVal $b;
193
194=head1 DESCRIPTION
195
196=head2 Declaration of overloaded functions
197
198The compilation directive
199
200 package Number;
201 use overload
202 "+" => \&add,
203 "*=" => "muas";
204
205declares function Number::add() for addition, and method muas() in
206the "class" C<Number> (or one of its base classes)
207for the assignment form C<*=> of multiplication.
208
209Arguments of this directive come in (key, value) pairs. Legal values
210are values legal inside a C<&{ ... }> call, so the name of a
211subroutine, a reference to a subroutine, or an anonymous subroutine
212will all work. Note that values specified as strings are
213interpreted as methods, not subroutines. Legal keys are listed below.
214
215The subroutine C<add> will be called to execute C<$a+$b> if $a
216is a reference to an object blessed into the package C<Number>, or if $a is
217not an object from a package with defined mathemagic addition, but $b is a
218reference to a C<Number>. It can also be called in other situations, like
219C<$a+=7>, or C<$a++>. See L<MAGIC AUTOGENERATION>. (Mathemagical
220methods refer to methods triggered by an overloaded mathematical
221operator.)
222
223Since overloading respects inheritance via the @ISA hierarchy, the
224above declaration would also trigger overloading of C<+> and C<*=> in
225all the packages which inherit from C<Number>.
226
227=head2 Calling Conventions for Binary Operations
228
229The functions specified in the C<use overload ...> directive are called
230with three (in one particular case with four, see L<Last Resort>)
231arguments. If the corresponding operation is binary, then the first
232two arguments are the two arguments of the operation. However, due to
233general object calling conventions, the first argument should always be
234an object in the package, so in the situation of C<7+$a>, the
235order of the arguments is interchanged. It probably does not matter
236when implementing the addition method, but whether the arguments
237are reversed is vital to the subtraction method. The method can
238query this information by examining the third argument, which can take
239three different values:
240
241=over 7
242
243=item FALSE
244
245the order of arguments is as in the current operation.
246
247=item TRUE
248
249the arguments are reversed.
250
251=item C<undef>
252
253the current operation is an assignment variant (as in
254C<$a+=7>), but the usual function is called instead. This additional
255information can be used to generate some optimizations. Compare
256L<Calling Conventions for Mutators>.
257
258=back
259
260=head2 Calling Conventions for Unary Operations
261
262Unary operation are considered binary operations with the second
263argument being C<undef>. Thus the functions that overloads C<{"++"}>
264is called with arguments C<($a,undef,'')> when $a++ is executed.
265
266=head2 Calling Conventions for Mutators
267
268Two types of mutators have different calling conventions:
269
270=over
271
272=item C<++> and C<-->
273
274The routines which implement these operators are expected to actually
275I<mutate> their arguments. So, assuming that $obj is a reference to a
276number,
277
278 sub incr { my $n = $ {$_[0]}; ++$n; $_[0] = bless \$n}
279
280is an appropriate implementation of overloaded C<++>. Note that
281
282 sub incr { ++$ {$_[0]} ; shift }
283
284is OK if used with preincrement and with postincrement. (In the case
285of postincrement a copying will be performed, see L<Copy Constructor>.)
286
287=item C<x=> and other assignment versions
288
289There is nothing special about these methods. They may change the
290value of their arguments, and may leave it as is. The result is going
291to be assigned to the value in the left-hand-side if different from
292this value.
293
294This allows for the same method to be used as overloaded C<+=> and
295C<+>. Note that this is I<allowed>, but not recommended, since by the
296semantic of L<"Fallback"> Perl will call the method for C<+> anyway,
297if C<+=> is not overloaded.
298
299=back
300
301B<Warning.> Due to the presence of assignment versions of operations,
302routines which may be called in assignment context may create
303self-referential structures. Currently Perl will not free self-referential
304structures until cycles are C<explicitly> broken. You may get problems
305when traversing your structures too.
306
307Say,
308
309 use overload '+' => sub { bless [ \$_[0], \$_[1] ] };
310
311is asking for trouble, since for code C<$obj += $foo> the subroutine
312is called as C<$obj = add($obj, $foo, undef)>, or C<$obj = [\$obj,
313\$foo]>. If using such a subroutine is an important optimization, one
314can overload C<+=> explicitly by a non-"optimized" version, or switch
315to non-optimized version if C<not defined $_[2]> (see
316L<Calling Conventions for Binary Operations>).
317
318Even if no I<explicit> assignment-variants of operators are present in
319the script, they may be generated by the optimizer. Say, C<",$obj,"> or
320C<',' . $obj . ','> may be both optimized to
321
322 my $tmp = ',' . $obj; $tmp .= ',';
323
324=head2 Overloadable Operations
325
326The following symbols can be specified in C<use overload> directive:
327
328=over 5
329
330=item * I<Arithmetic operations>
331
332 "+", "+=", "-", "-=", "*", "*=", "/", "/=", "%", "%=",
333 "**", "**=", "<<", "<<=", ">>", ">>=", "x", "x=", ".", ".=",
334
335For these operations a substituted non-assignment variant can be called if
336the assignment variant is not available. Methods for operations "C<+>",
337"C<->", "C<+=>", and "C<-=>" can be called to automatically generate
338increment and decrement methods. The operation "C<->" can be used to
339autogenerate missing methods for unary minus or C<abs>.
340
341See L<"MAGIC AUTOGENERATION">, L<"Calling Conventions for Mutators"> and
342L<"Calling Conventions for Binary Operations">) for details of these
343substitutions.
344
345=item * I<Comparison operations>
346
347 "<", "<=", ">", ">=", "==", "!=", "<=>",
348 "lt", "le", "gt", "ge", "eq", "ne", "cmp",
349
350If the corresponding "spaceship" variant is available, it can be
351used to substitute for the missing operation. During C<sort>ing
352arrays, C<cmp> is used to compare values subject to C<use overload>.
353
354=item * I<Bit operations>
355
356 "&", "^", "|", "neg", "!", "~",
357
358"C<neg>" stands for unary minus. If the method for C<neg> is not
359specified, it can be autogenerated using the method for
360subtraction. If the method for "C<!>" is not specified, it can be
361autogenerated using the methods for "C<bool>", or "C<\"\">", or "C<0+>".
362
363=item * I<Increment and decrement>
364
365 "++", "--",
366
367If undefined, addition and subtraction methods can be
368used instead. These operations are called both in prefix and
369postfix form.
370
371=item * I<Transcendental functions>
372
373 "atan2", "cos", "sin", "exp", "abs", "log", "sqrt", "int"
374
375If C<abs> is unavailable, it can be autogenerated using methods
376for "E<lt>" or "E<lt>=E<gt>" combined with either unary minus or subtraction.
377
378Note that traditionally the Perl function L<int> rounds to 0, thus for
379floating-point-like types one should follow the same semantic. If
380C<int> is unavailable, it can be autogenerated using the overloading of
381C<0+>.
382
383=item * I<Boolean, string and numeric conversion>
384
385 "bool", "\"\"", "0+",
386
387If one or two of these operations are not overloaded, the remaining ones can
388be used instead. C<bool> is used in the flow control operators
389(like C<while>) and for the ternary "C<?:>" operation. These functions can
390return any arbitrary Perl value. If the corresponding operation for this value
391is overloaded too, that operation will be called again with this value.
392
393As a special case if the overload returns the object itself then it will
394be used directly. An overloaded conversion returning the object is
395probably a bug, because you're likely to get something that looks like
396C<YourPackage=HASH(0x8172b34)>.
397
398=item * I<Iteration>
399
400 "<>"
401
402If not overloaded, the argument will be converted to a filehandle or
403glob (which may require a stringification). The same overloading
404happens both for the I<read-filehandle> syntax C<E<lt>$varE<gt>> and
405I<globbing> syntax C<E<lt>${var}E<gt>>.
406
407=item * I<Dereferencing>
408
409 '${}', '@{}', '%{}', '&{}', '*{}'.
410
411If not overloaded, the argument will be dereferenced I<as is>, thus
412should be of correct type. These functions should return a reference
413of correct type, or another object with overloaded dereferencing.
414
415As a special case if the overload returns the object itself then it
416will be used directly (provided it is the correct type).
417
418The dereference operators must be specified explicitly they will not be passed to
419"nomethod".
420
421=item * I<Special>
422
423 "nomethod", "fallback", "=",
424
425see L<SPECIAL SYMBOLS FOR C<use overload>>.
426
427=back
428
429See L<"Fallback"> for an explanation of when a missing method can be
430autogenerated.
431
432A computer-readable form of the above table is available in the hash
433%overload::ops, with values being space-separated lists of names:
434
435 with_assign => '+ - * / % ** << >> x .',
436 assign => '+= -= *= /= %= **= <<= >>= x= .=',
437 num_comparison => '< <= > >= == !=',
438 '3way_comparison'=> '<=> cmp',
439 str_comparison => 'lt le gt ge eq ne',
440 binary => '& | ^',
441 unary => 'neg ! ~',
442 mutators => '++ --',
443 func => 'atan2 cos sin exp abs log sqrt',
444 conversion => 'bool "" 0+',
445 iterators => '<>',
446 dereferencing => '${} @{} %{} &{} *{}',
447 special => 'nomethod fallback ='
448
449=head2 Inheritance and overloading
450
451Inheritance interacts with overloading in two ways.
452
453=over
454
455=item Strings as values of C<use overload> directive
456
457If C<value> in
458
459 use overload key => value;
460
461is a string, it is interpreted as a method name.
462
463=item Overloading of an operation is inherited by derived classes
464
465Any class derived from an overloaded class is also overloaded. The
466set of overloaded methods is the union of overloaded methods of all
467the ancestors. If some method is overloaded in several ancestor, then
468which description will be used is decided by the usual inheritance
469rules:
470
471If C<A> inherits from C<B> and C<C> (in this order), C<B> overloads
472C<+> with C<\&D::plus_sub>, and C<C> overloads C<+> by C<"plus_meth">,
473then the subroutine C<D::plus_sub> will be called to implement
474operation C<+> for an object in package C<A>.
475
476=back
477
478Note that since the value of the C<fallback> key is not a subroutine,
479its inheritance is not governed by the above rules. In the current
480implementation, the value of C<fallback> in the first overloaded
481ancestor is used, but this is accidental and subject to change.
482
483=head1 SPECIAL SYMBOLS FOR C<use overload>
484
485Three keys are recognized by Perl that are not covered by the above
486description.
487
488=head2 Last Resort
489
490C<"nomethod"> should be followed by a reference to a function of four
491parameters. If defined, it is called when the overloading mechanism
492cannot find a method for some operation. The first three arguments of
493this function coincide with the arguments for the corresponding method if
494it were found, the fourth argument is the symbol
495corresponding to the missing method. If several methods are tried,
496the last one is used. Say, C<1-$a> can be equivalent to
497
498 &nomethodMethod($a,1,1,"-")
499
500if the pair C<"nomethod" =E<gt> "nomethodMethod"> was specified in the
501C<use overload> directive.
502
503The C<"nomethod"> mechanism is I<not> used for the dereference operators
504( ${} @{} %{} &{} *{} ).
505
506
507If some operation cannot be resolved, and there is no function
508assigned to C<"nomethod">, then an exception will be raised via die()--
509unless C<"fallback"> was specified as a key in C<use overload> directive.
510
511
512=head2 Fallback
513
514The key C<"fallback"> governs what to do if a method for a particular
515operation is not found. Three different cases are possible depending on
516the value of C<"fallback">:
517
518=over 16
519
520=item * C<undef>
521
522Perl tries to use a
523substituted method (see L<MAGIC AUTOGENERATION>). If this fails, it
524then tries to calls C<"nomethod"> value; if missing, an exception
525will be raised.
526
527=item * TRUE
528
529The same as for the C<undef> value, but no exception is raised. Instead,
530it silently reverts to what it would have done were there no C<use overload>
531present.
532
533=item * defined, but FALSE
534
535No autogeneration is tried. Perl tries to call
536C<"nomethod"> value, and if this is missing, raises an exception.
537
538=back
539
540B<Note.> C<"fallback"> inheritance via @ISA is not carved in stone
541yet, see L<"Inheritance and overloading">.
542
543=head2 Copy Constructor
544
545The value for C<"="> is a reference to a function with three
546arguments, i.e., it looks like the other values in C<use
547overload>. However, it does not overload the Perl assignment
548operator. This would go against Camel hair.
549
550This operation is called in the situations when a mutator is applied
551to a reference that shares its object with some other reference, such
552as
553
554 $a=$b;
555 ++$a;
556
557To make this change $a and not change $b, a copy of C<$$a> is made,
558and $a is assigned a reference to this new object. This operation is
559done during execution of the C<++$a>, and not during the assignment,
560(so before the increment C<$$a> coincides with C<$$b>). This is only
561done if C<++> is expressed via a method for C<'++'> or C<'+='> (or
562C<nomethod>). Note that if this operation is expressed via C<'+'>
563a nonmutator, i.e., as in
564
565 $a=$b;
566 $a=$a+1;
567
568then C<$a> does not reference a new copy of C<$$a>, since $$a does not
569appear as lvalue when the above code is executed.
570
571If the copy constructor is required during the execution of some mutator,
572but a method for C<'='> was not specified, it can be autogenerated as a
573string copy if the object is a plain scalar.
574
575=over 5
576
577=item B<Example>
578
579The actually executed code for
580
581 $a=$b;
582 Something else which does not modify $a or $b....
583 ++$a;
584
585may be
586
587 $a=$b;
588 Something else which does not modify $a or $b....
589 $a = $a->clone(undef,"");
590 $a->incr(undef,"");
591
592if $b was mathemagical, and C<'++'> was overloaded with C<\&incr>,
593C<'='> was overloaded with C<\&clone>.
594
595=back
596
597Same behaviour is triggered by C<$b = $a++>, which is consider a synonym for
598C<$b = $a; ++$a>.
599
600=head1 MAGIC AUTOGENERATION
601
602If a method for an operation is not found, and the value for C<"fallback"> is
603TRUE or undefined, Perl tries to autogenerate a substitute method for
604the missing operation based on the defined operations. Autogenerated method
605substitutions are possible for the following operations:
606
607=over 16
608
609=item I<Assignment forms of arithmetic operations>
610
611C<$a+=$b> can use the method for C<"+"> if the method for C<"+=">
612is not defined.
613
614=item I<Conversion operations>
615
616String, numeric, and boolean conversion are calculated in terms of one
617another if not all of them are defined.
618
619=item I<Increment and decrement>
620
621The C<++$a> operation can be expressed in terms of C<$a+=1> or C<$a+1>,
622and C<$a--> in terms of C<$a-=1> and C<$a-1>.
623
624=item C<abs($a)>
625
626can be expressed in terms of C<$aE<lt>0> and C<-$a> (or C<0-$a>).
627
628=item I<Unary minus>
629
630can be expressed in terms of subtraction.
631
632=item I<Negation>
633
634C<!> and C<not> can be expressed in terms of boolean conversion, or
635string or numerical conversion.
636
637=item I<Concatenation>
638
639can be expressed in terms of string conversion.
640
641=item I<Comparison operations>
642
643can be expressed in terms of its "spaceship" counterpart: either
644C<E<lt>=E<gt>> or C<cmp>:
645
646 <, >, <=, >=, ==, != in terms of <=>
647 lt, gt, le, ge, eq, ne in terms of cmp
648
649=item I<Iterator>
650
651 <> in terms of builtin operations
652
653=item I<Dereferencing>
654
655 ${} @{} %{} &{} *{} in terms of builtin operations
656
657=item I<Copy operator>
658
659can be expressed in terms of an assignment to the dereferenced value, if this
660value is a scalar and not a reference.
661
662=back
663
664=head1 Losing overloading
665
666The restriction for the comparison operation is that even if, for example,
667`C<cmp>' should return a blessed reference, the autogenerated `C<lt>'
668function will produce only a standard logical value based on the
669numerical value of the result of `C<cmp>'. In particular, a working
670numeric conversion is needed in this case (possibly expressed in terms of
671other conversions).
672
673Similarly, C<.=> and C<x=> operators lose their mathemagical properties
674if the string conversion substitution is applied.
675
676When you chop() a mathemagical object it is promoted to a string and its
677mathemagical properties are lost. The same can happen with other
678operations as well.
679
680=head1 Run-time Overloading
681
682Since all C<use> directives are executed at compile-time, the only way to
683change overloading during run-time is to
684
685 eval 'use overload "+" => \&addmethod';
686
687You can also use
688
689 eval 'no overload "+", "--", "<="';
690
691though the use of these constructs during run-time is questionable.
692
693=head1 Public functions
694
695Package C<overload.pm> provides the following public functions:
696
697=over 5
698
699=item overload::StrVal(arg)
700
701Gives string value of C<arg> as in absence of stringify overloading.
702
703=item overload::Overloaded(arg)
704
705Returns true if C<arg> is subject to overloading of some operations.
706
707=item overload::Method(obj,op)
708
709Returns C<undef> or a reference to the method that implements C<op>.
710
711=back
712
713=head1 Overloading constants
714
715For some application Perl parser mangles constants too much. It is possible
716to hook into this process via overload::constant() and overload::remove_constant()
717functions.
718
719These functions take a hash as an argument. The recognized keys of this hash
720are
721
722=over 8
723
724=item integer
725
726to overload integer constants,
727
728=item float
729
730to overload floating point constants,
731
732=item binary
733
734to overload octal and hexadecimal constants,
735
736=item q
737
738to overload C<q>-quoted strings, constant pieces of C<qq>- and C<qx>-quoted
739strings and here-documents,
740
741=item qr
742
743to overload constant pieces of regular expressions.
744
745=back
746
747The corresponding values are references to functions which take three arguments:
748the first one is the I<initial> string form of the constant, the second one
749is how Perl interprets this constant, the third one is how the constant is used.
750Note that the initial string form does not
751contain string delimiters, and has backslashes in backslash-delimiter
752combinations stripped (thus the value of delimiter is not relevant for
753processing of this string). The return value of this function is how this
754constant is going to be interpreted by Perl. The third argument is undefined
755unless for overloaded C<q>- and C<qr>- constants, it is C<q> in single-quote
756context (comes from strings, regular expressions, and single-quote HERE
757documents), it is C<tr> for arguments of C<tr>/C<y> operators,
758it is C<s> for right-hand side of C<s>-operator, and it is C<qq> otherwise.
759
760Since an expression C<"ab$cd,,"> is just a shortcut for C<'ab' . $cd . ',,'>,
761it is expected that overloaded constant strings are equipped with reasonable
762overloaded catenation operator, otherwise absurd results will result.
763Similarly, negative numbers are considered as negations of positive constants.
764
765Note that it is probably meaningless to call the functions overload::constant()
766and overload::remove_constant() from anywhere but import() and unimport() methods.
767From these methods they may be called as
768
769 sub import {
770 shift;
771 return unless @_;
772 die "unknown import: @_" unless @_ == 1 and $_[0] eq ':constant';
773 overload::constant integer => sub {Math::BigInt->new(shift)};
774 }
775
776B<BUGS> Currently overloaded-ness of constants does not propagate
777into C<eval '...'>.
778
779=head1 IMPLEMENTATION
780
781What follows is subject to change RSN.
782
783The table of methods for all operations is cached in magic for the
784symbol table hash for the package. The cache is invalidated during
785processing of C<use overload>, C<no overload>, new function
786definitions, and changes in @ISA. However, this invalidation remains
787unprocessed until the next C<bless>ing into the package. Hence if you
788want to change overloading structure dynamically, you'll need an
789additional (fake) C<bless>ing to update the table.
790
791(Every SVish thing has a magic queue, and magic is an entry in that
792queue. This is how a single variable may participate in multiple
793forms of magic simultaneously. For instance, environment variables
794regularly have two forms at once: their %ENV magic and their taint
795magic. However, the magic which implements overloading is applied to
796the stashes, which are rarely used directly, thus should not slow down
797Perl.)
798
799If an object belongs to a package using overload, it carries a special
800flag. Thus the only speed penalty during arithmetic operations without
801overloading is the checking of this flag.
802
803In fact, if C<use overload> is not present, there is almost no overhead
804for overloadable operations, so most programs should not suffer
805measurable performance penalties. A considerable effort was made to
806minimize the overhead when overload is used in some package, but the
807arguments in question do not belong to packages using overload. When
808in doubt, test your speed with C<use overload> and without it. So far
809there have been no reports of substantial speed degradation if Perl is
810compiled with optimization turned on.
811
812There is no size penalty for data if overload is not used. The only
813size penalty if overload is used in some package is that I<all> the
814packages acquire a magic during the next C<bless>ing into the
815package. This magic is three-words-long for packages without
816overloading, and carries the cache table if the package is overloaded.
817
818Copying (C<$a=$b>) is shallow; however, a one-level-deep copying is
819carried out before any operation that can imply an assignment to the
820object $a (or $b) refers to, like C<$a++>. You can override this
821behavior by defining your own copy constructor (see L<"Copy Constructor">).
822
823It is expected that arguments to methods that are not explicitly supposed
824to be changed are constant (but this is not enforced).
825
826=head1 Metaphor clash
827
828One may wonder why the semantic of overloaded C<=> is so counter intuitive.
829If it I<looks> counter intuitive to you, you are subject to a metaphor
830clash.
831
832Here is a Perl object metaphor:
833
834I< object is a reference to blessed data>
835
836and an arithmetic metaphor:
837
838I< object is a thing by itself>.
839
840The I<main> problem of overloading C<=> is the fact that these metaphors
841imply different actions on the assignment C<$a = $b> if $a and $b are
842objects. Perl-think implies that $a becomes a reference to whatever
843$b was referencing. Arithmetic-think implies that the value of "object"
844$a is changed to become the value of the object $b, preserving the fact
845that $a and $b are separate entities.
846
847The difference is not relevant in the absence of mutators. After
848a Perl-way assignment an operation which mutates the data referenced by $a
849would change the data referenced by $b too. Effectively, after
850C<$a = $b> values of $a and $b become I<indistinguishable>.
851
852On the other hand, anyone who has used algebraic notation knows the
853expressive power of the arithmetic metaphor. Overloading works hard
854to enable this metaphor while preserving the Perlian way as far as
855possible. Since it is not possible to freely mix two contradicting
856metaphors, overloading allows the arithmetic way to write things I<as
857far as all the mutators are called via overloaded access only>. The
858way it is done is described in L<Copy Constructor>.
859
860If some mutator methods are directly applied to the overloaded values,
861one may need to I<explicitly unlink> other values which references the
862same value:
863
864 $a = new Data 23;
865 ...
866 $b = $a; # $b is "linked" to $a
867 ...
868 $a = $a->clone; # Unlink $b from $a
869 $a->increment_by(4);
870
871Note that overloaded access makes this transparent:
872
873 $a = new Data 23;
874 $b = $a; # $b is "linked" to $a
875 $a += 4; # would unlink $b automagically
876
877However, it would not make
878
879 $a = new Data 23;
880 $a = 4; # Now $a is a plain 4, not 'Data'
881
882preserve "objectness" of $a. But Perl I<has> a way to make assignments
883to an object do whatever you want. It is just not the overload, but
884tie()ing interface (see L<perlfunc/tie>). Adding a FETCH() method
885which returns the object itself, and STORE() method which changes the
886value of the object, one can reproduce the arithmetic metaphor in its
887completeness, at least for variables which were tie()d from the start.
888
889(Note that a workaround for a bug may be needed, see L<"BUGS">.)
890
891=head1 Cookbook
892
893Please add examples to what follows!
894
895=head2 Two-face scalars
896
897Put this in F<two_face.pm> in your Perl library directory:
898
899 package two_face; # Scalars with separate string and
900 # numeric values.
901 sub new { my $p = shift; bless [@_], $p }
902 use overload '""' => \&str, '0+' => \&num, fallback => 1;
903 sub num {shift->[1]}
904 sub str {shift->[0]}
905
906Use it as follows:
907
908 require two_face;
909 my $seven = new two_face ("vii", 7);
910 printf "seven=$seven, seven=%d, eight=%d\n", $seven, $seven+1;
911 print "seven contains `i'\n" if $seven =~ /i/;
912
913(The second line creates a scalar which has both a string value, and a
914numeric value.) This prints:
915
916 seven=vii, seven=7, eight=8
917 seven contains `i'
918
919=head2 Two-face references
920
921Suppose you want to create an object which is accessible as both an
922array reference and a hash reference, similar to the
923L<pseudo-hash|perlref/"Pseudo-hashes: Using an array as a hash">
924builtin Perl type. Let's make it better than a pseudo-hash by
925allowing index 0 to be treated as a normal element.
926
927 package two_refs;
928 use overload '%{}' => \&gethash, '@{}' => sub { $ {shift()} };
929 sub new {
930 my $p = shift;
931 bless \ [@_], $p;
932 }
933 sub gethash {
934 my %h;
935 my $self = shift;
936 tie %h, ref $self, $self;
937 \%h;
938 }
939
940 sub TIEHASH { my $p = shift; bless \ shift, $p }
941 my %fields;
942 my $i = 0;
943 $fields{$_} = $i++ foreach qw{zero one two three};
944 sub STORE {
945 my $self = ${shift()};
946 my $key = $fields{shift()};
947 defined $key or die "Out of band access";
948 $$self->[$key] = shift;
949 }
950 sub FETCH {
951 my $self = ${shift()};
952 my $key = $fields{shift()};
953 defined $key or die "Out of band access";
954 $$self->[$key];
955 }
956
957Now one can access an object using both the array and hash syntax:
958
959 my $bar = new two_refs 3,4,5,6;
960 $bar->[2] = 11;
961 $bar->{two} == 11 or die 'bad hash fetch';
962
963Note several important features of this example. First of all, the
964I<actual> type of $bar is a scalar reference, and we do not overload
965the scalar dereference. Thus we can get the I<actual> non-overloaded
966contents of $bar by just using C<$$bar> (what we do in functions which
967overload dereference). Similarly, the object returned by the
968TIEHASH() method is a scalar reference.
969
970Second, we create a new tied hash each time the hash syntax is used.
971This allows us not to worry about a possibility of a reference loop,
972which would lead to a memory leak.
973
974Both these problems can be cured. Say, if we want to overload hash
975dereference on a reference to an object which is I<implemented> as a
976hash itself, the only problem one has to circumvent is how to access
977this I<actual> hash (as opposed to the I<virtual> hash exhibited by the
978overloaded dereference operator). Here is one possible fetching routine:
979
980 sub access_hash {
981 my ($self, $key) = (shift, shift);
982 my $class = ref $self;
983 bless $self, 'overload::dummy'; # Disable overloading of %{}
984 my $out = $self->{$key};
985 bless $self, $class; # Restore overloading
986 $out;
987 }
988
989To remove creation of the tied hash on each access, one may an extra
990level of indirection which allows a non-circular structure of references:
991
992 package two_refs1;
993 use overload '%{}' => sub { ${shift()}->[1] },
994 '@{}' => sub { ${shift()}->[0] };
995 sub new {
996 my $p = shift;
997 my $a = [@_];
998 my %h;
999 tie %h, $p, $a;
1000 bless \ [$a, \%h], $p;
1001 }
1002 sub gethash {
1003 my %h;
1004 my $self = shift;
1005 tie %h, ref $self, $self;
1006 \%h;
1007 }
1008
1009 sub TIEHASH { my $p = shift; bless \ shift, $p }
1010 my %fields;
1011 my $i = 0;
1012 $fields{$_} = $i++ foreach qw{zero one two three};
1013 sub STORE {
1014 my $a = ${shift()};
1015 my $key = $fields{shift()};
1016 defined $key or die "Out of band access";
1017 $a->[$key] = shift;
1018 }
1019 sub FETCH {
1020 my $a = ${shift()};
1021 my $key = $fields{shift()};
1022 defined $key or die "Out of band access";
1023 $a->[$key];
1024 }
1025
1026Now if $baz is overloaded like this, then C<$baz> is a reference to a
1027reference to the intermediate array, which keeps a reference to an
1028actual array, and the access hash. The tie()ing object for the access
1029hash is a reference to a reference to the actual array, so
1030
1031=over
1032
1033=item *
1034
1035There are no loops of references.
1036
1037=item *
1038
1039Both "objects" which are blessed into the class C<two_refs1> are
1040references to a reference to an array, thus references to a I<scalar>.
1041Thus the accessor expression C<$$foo-E<gt>[$ind]> involves no
1042overloaded operations.
1043
1044=back
1045
1046=head2 Symbolic calculator
1047
1048Put this in F<symbolic.pm> in your Perl library directory:
1049
1050 package symbolic; # Primitive symbolic calculator
1051 use overload nomethod => \&wrap;
1052
1053 sub new { shift; bless ['n', @_] }
1054 sub wrap {
1055 my ($obj, $other, $inv, $meth) = @_;
1056 ($obj, $other) = ($other, $obj) if $inv;
1057 bless [$meth, $obj, $other];
1058 }
1059
1060This module is very unusual as overloaded modules go: it does not
1061provide any usual overloaded operators, instead it provides the L<Last
1062Resort> operator C<nomethod>. In this example the corresponding
1063subroutine returns an object which encapsulates operations done over
1064the objects: C<new symbolic 3> contains C<['n', 3]>, C<2 + new
1065symbolic 3> contains C<['+', 2, ['n', 3]]>.
1066
1067Here is an example of the script which "calculates" the side of
1068circumscribed octagon using the above package:
1069
1070 require symbolic;
1071 my $iter = 1; # 2**($iter+2) = 8
1072 my $side = new symbolic 1;
1073 my $cnt = $iter;
1074
1075 while ($cnt--) {
1076 $side = (sqrt(1 + $side**2) - 1)/$side;
1077 }
1078 print "OK\n";
1079
1080The value of $side is
1081
1082 ['/', ['-', ['sqrt', ['+', 1, ['**', ['n', 1], 2]],
1083 undef], 1], ['n', 1]]
1084
1085Note that while we obtained this value using a nice little script,
1086there is no simple way to I<use> this value. In fact this value may
1087be inspected in debugger (see L<perldebug>), but ony if
1088C<bareStringify> B<O>ption is set, and not via C<p> command.
1089
1090If one attempts to print this value, then the overloaded operator
1091C<""> will be called, which will call C<nomethod> operator. The
1092result of this operator will be stringified again, but this result is
1093again of type C<symbolic>, which will lead to an infinite loop.
1094
1095Add a pretty-printer method to the module F<symbolic.pm>:
1096
1097 sub pretty {
1098 my ($meth, $a, $b) = @{+shift};
1099 $a = 'u' unless defined $a;
1100 $b = 'u' unless defined $b;
1101 $a = $a->pretty if ref $a;
1102 $b = $b->pretty if ref $b;
1103 "[$meth $a $b]";
1104 }
1105
1106Now one can finish the script by
1107
1108 print "side = ", $side->pretty, "\n";
1109
1110The method C<pretty> is doing object-to-string conversion, so it
1111is natural to overload the operator C<""> using this method. However,
1112inside such a method it is not necessary to pretty-print the
1113I<components> $a and $b of an object. In the above subroutine
1114C<"[$meth $a $b]"> is a catenation of some strings and components $a
1115and $b. If these components use overloading, the catenation operator
1116will look for an overloaded operator C<.>; if not present, it will
1117look for an overloaded operator C<"">. Thus it is enough to use
1118
1119 use overload nomethod => \&wrap, '""' => \&str;
1120 sub str {
1121 my ($meth, $a, $b) = @{+shift};
1122 $a = 'u' unless defined $a;
1123 $b = 'u' unless defined $b;
1124 "[$meth $a $b]";
1125 }
1126
1127Now one can change the last line of the script to
1128
1129 print "side = $side\n";
1130
1131which outputs
1132
1133 side = [/ [- [sqrt [+ 1 [** [n 1 u] 2]] u] 1] [n 1 u]]
1134
1135and one can inspect the value in debugger using all the possible
1136methods.
1137
1138Something is still amiss: consider the loop variable $cnt of the
1139script. It was a number, not an object. We cannot make this value of
1140type C<symbolic>, since then the loop will not terminate.
1141
1142Indeed, to terminate the cycle, the $cnt should become false.
1143However, the operator C<bool> for checking falsity is overloaded (this
1144time via overloaded C<"">), and returns a long string, thus any object
1145of type C<symbolic> is true. To overcome this, we need a way to
1146compare an object to 0. In fact, it is easier to write a numeric
1147conversion routine.
1148
1149Here is the text of F<symbolic.pm> with such a routine added (and
1150slightly modified str()):
1151
1152 package symbolic; # Primitive symbolic calculator
1153 use overload
1154 nomethod => \&wrap, '""' => \&str, '0+' => \&num;
1155
1156 sub new { shift; bless ['n', @_] }
1157 sub wrap {
1158 my ($obj, $other, $inv, $meth) = @_;
1159 ($obj, $other) = ($other, $obj) if $inv;
1160 bless [$meth, $obj, $other];
1161 }
1162 sub str {
1163 my ($meth, $a, $b) = @{+shift};
1164 $a = 'u' unless defined $a;
1165 if (defined $b) {
1166 "[$meth $a $b]";
1167 } else {
1168 "[$meth $a]";
1169 }
1170 }
1171 my %subr = ( n => sub {$_[0]},
1172 sqrt => sub {sqrt $_[0]},
1173 '-' => sub {shift() - shift()},
1174 '+' => sub {shift() + shift()},
1175 '/' => sub {shift() / shift()},
1176 '*' => sub {shift() * shift()},
1177 '**' => sub {shift() ** shift()},
1178 );
1179 sub num {
1180 my ($meth, $a, $b) = @{+shift};
1181 my $subr = $subr{$meth}
1182 or die "Do not know how to ($meth) in symbolic";
1183 $a = $a->num if ref $a eq __PACKAGE__;
1184 $b = $b->num if ref $b eq __PACKAGE__;
1185 $subr->($a,$b);
1186 }
1187
1188All the work of numeric conversion is done in %subr and num(). Of
1189course, %subr is not complete, it contains only operators used in the
1190example below. Here is the extra-credit question: why do we need an
1191explicit recursion in num()? (Answer is at the end of this section.)
1192
1193Use this module like this:
1194
1195 require symbolic;
1196 my $iter = new symbolic 2; # 16-gon
1197 my $side = new symbolic 1;
1198 my $cnt = $iter;
1199
1200 while ($cnt) {
1201 $cnt = $cnt - 1; # Mutator `--' not implemented
1202 $side = (sqrt(1 + $side**2) - 1)/$side;
1203 }
1204 printf "%s=%f\n", $side, $side;
1205 printf "pi=%f\n", $side*(2**($iter+2));
1206
1207It prints (without so many line breaks)
1208
1209 [/ [- [sqrt [+ 1 [** [/ [- [sqrt [+ 1 [** [n 1] 2]]] 1]
1210 [n 1]] 2]]] 1]
1211 [/ [- [sqrt [+ 1 [** [n 1] 2]]] 1] [n 1]]]=0.198912
1212 pi=3.182598
1213
1214The above module is very primitive. It does not implement
1215mutator methods (C<++>, C<-=> and so on), does not do deep copying
1216(not required without mutators!), and implements only those arithmetic
1217operations which are used in the example.
1218
1219To implement most arithmetic operations is easy; one should just use
1220the tables of operations, and change the code which fills %subr to
1221
1222 my %subr = ( 'n' => sub {$_[0]} );
1223 foreach my $op (split " ", $overload::ops{with_assign}) {
1224 $subr{$op} = $subr{"$op="} = eval "sub {shift() $op shift()}";
1225 }
1226 my @bins = qw(binary 3way_comparison num_comparison str_comparison);
1227 foreach my $op (split " ", "@overload::ops{ @bins }") {
1228 $subr{$op} = eval "sub {shift() $op shift()}";
1229 }
1230 foreach my $op (split " ", "@overload::ops{qw(unary func)}") {
1231 print "defining `$op'\n";
1232 $subr{$op} = eval "sub {$op shift()}";
1233 }
1234
1235Due to L<Calling Conventions for Mutators>, we do not need anything
1236special to make C<+=> and friends work, except filling C<+=> entry of
1237%subr, and defining a copy constructor (needed since Perl has no
1238way to know that the implementation of C<'+='> does not mutate
1239the argument, compare L<Copy Constructor>).
1240
1241To implement a copy constructor, add C<< '=' => \&cpy >> to C<use overload>
1242line, and code (this code assumes that mutators change things one level
1243deep only, so recursive copying is not needed):
1244
1245 sub cpy {
1246 my $self = shift;
1247 bless [@$self], ref $self;
1248 }
1249
1250To make C<++> and C<--> work, we need to implement actual mutators,
1251either directly, or in C<nomethod>. We continue to do things inside
1252C<nomethod>, thus add
1253
1254 if ($meth eq '++' or $meth eq '--') {
1255 @$obj = ($meth, (bless [@$obj]), 1); # Avoid circular reference
1256 return $obj;
1257 }
1258
1259after the first line of wrap(). This is not a most effective
1260implementation, one may consider
1261
1262 sub inc { $_[0] = bless ['++', shift, 1]; }
1263
1264instead.
1265
1266As a final remark, note that one can fill %subr by
1267
1268 my %subr = ( 'n' => sub {$_[0]} );
1269 foreach my $op (split " ", $overload::ops{with_assign}) {
1270 $subr{$op} = $subr{"$op="} = eval "sub {shift() $op shift()}";
1271 }
1272 my @bins = qw(binary 3way_comparison num_comparison str_comparison);
1273 foreach my $op (split " ", "@overload::ops{ @bins }") {
1274 $subr{$op} = eval "sub {shift() $op shift()}";
1275 }
1276 foreach my $op (split " ", "@overload::ops{qw(unary func)}") {
1277 $subr{$op} = eval "sub {$op shift()}";
1278 }
1279 $subr{'++'} = $subr{'+'};
1280 $subr{'--'} = $subr{'-'};
1281
1282This finishes implementation of a primitive symbolic calculator in
128350 lines of Perl code. Since the numeric values of subexpressions
1284are not cached, the calculator is very slow.
1285
1286Here is the answer for the exercise: In the case of str(), we need no
1287explicit recursion since the overloaded C<.>-operator will fall back
1288to an existing overloaded operator C<"">. Overloaded arithmetic
1289operators I<do not> fall back to numeric conversion if C<fallback> is
1290not explicitly requested. Thus without an explicit recursion num()
1291would convert C<['+', $a, $b]> to C<$a + $b>, which would just rebuild
1292the argument of num().
1293
1294If you wonder why defaults for conversion are different for str() and
1295num(), note how easy it was to write the symbolic calculator. This
1296simplicity is due to an appropriate choice of defaults. One extra
1297note: due to the explicit recursion num() is more fragile than sym():
1298we need to explicitly check for the type of $a and $b. If components
1299$a and $b happen to be of some related type, this may lead to problems.
1300
1301=head2 I<Really> symbolic calculator
1302
1303One may wonder why we call the above calculator symbolic. The reason
1304is that the actual calculation of the value of expression is postponed
1305until the value is I<used>.
1306
1307To see it in action, add a method
1308
1309 sub STORE {
1310 my $obj = shift;
1311 $#$obj = 1;
1312 @$obj->[0,1] = ('=', shift);
1313 }
1314
1315to the package C<symbolic>. After this change one can do
1316
1317 my $a = new symbolic 3;
1318 my $b = new symbolic 4;
1319 my $c = sqrt($a**2 + $b**2);
1320
1321and the numeric value of $c becomes 5. However, after calling
1322
1323 $a->STORE(12); $b->STORE(5);
1324
1325the numeric value of $c becomes 13. There is no doubt now that the module
1326symbolic provides a I<symbolic> calculator indeed.
1327
1328To hide the rough edges under the hood, provide a tie()d interface to the
1329package C<symbolic> (compare with L<Metaphor clash>). Add methods
1330
1331 sub TIESCALAR { my $pack = shift; $pack->new(@_) }
1332 sub FETCH { shift }
1333 sub nop { } # Around a bug
1334
1335(the bug is described in L<"BUGS">). One can use this new interface as
1336
1337 tie $a, 'symbolic', 3;
1338 tie $b, 'symbolic', 4;
1339 $a->nop; $b->nop; # Around a bug
1340
1341 my $c = sqrt($a**2 + $b**2);
1342
1343Now numeric value of $c is 5. After C<$a = 12; $b = 5> the numeric value
1344of $c becomes 13. To insulate the user of the module add a method
1345
1346 sub vars { my $p = shift; tie($_, $p), $_->nop foreach @_; }
1347
1348Now
1349
1350 my ($a, $b);
1351 symbolic->vars($a, $b);
1352 my $c = sqrt($a**2 + $b**2);
1353
1354 $a = 3; $b = 4;
1355 printf "c5 %s=%f\n", $c, $c;
1356
1357 $a = 12; $b = 5;
1358 printf "c13 %s=%f\n", $c, $c;
1359
1360shows that the numeric value of $c follows changes to the values of $a
1361and $b.
1362
1363=head1 AUTHOR
1364
1365Ilya Zakharevich E<lt>F<ilya@math.mps.ohio-state.edu>E<gt>.
1366
1367=head1 DIAGNOSTICS
1368
1369When Perl is run with the B<-Do> switch or its equivalent, overloading
1370induces diagnostic messages.
1371
1372Using the C<m> command of Perl debugger (see L<perldebug>) one can
1373deduce which operations are overloaded (and which ancestor triggers
1374this overloading). Say, if C<eq> is overloaded, then the method C<(eq>
1375is shown by debugger. The method C<()> corresponds to the C<fallback>
1376key (in fact a presence of this method shows that this package has
1377overloading enabled, and it is what is used by the C<Overloaded>
1378function of module C<overload>).
1379
1380The module might issue the following warnings:
1381
1382=over 4
1383
1384=item Odd number of arguments for overload::constant
1385
1386(W) The call to overload::constant contained an odd number of arguments.
1387The arguments should come in pairs.
1388
1389=item `%s' is not an overloadable type
1390
1391(W) You tried to overload a constant type the overload package is unaware of.
1392
1393=item `%s' is not a code reference
1394
1395(W) The second (fourth, sixth, ...) argument of overload::constant needs
1396to be a code reference. Either an anonymous subroutine, or a reference
1397to a subroutine.
1398
1399=back
1400
1401=head1 BUGS
1402
1403Because it is used for overloading, the per-package hash %OVERLOAD now
1404has a special meaning in Perl. The symbol table is filled with names
1405looking like line-noise.
1406
1407For the purpose of inheritance every overloaded package behaves as if
1408C<fallback> is present (possibly undefined). This may create
1409interesting effects if some package is not overloaded, but inherits
1410from two overloaded packages.
1411
1412Relation between overloading and tie()ing is broken. Overloading is
1413triggered or not basing on the I<previous> class of tie()d value.
1414
1415This happens because the presence of overloading is checked too early,
1416before any tie()d access is attempted. If the FETCH()ed class of the
1417tie()d value does not change, a simple workaround is to access the value
1418immediately after tie()ing, so that after this call the I<previous> class
1419coincides with the current one.
1420
1421B<Needed:> a way to fix this without a speed penalty.
1422
1423Barewords are not covered by overloaded string constants.
1424
1425This document is confusing. There are grammos and misleading language
1426used in places. It would seem a total rewrite is needed.
1427
1428=cut
1429