This and other RFCs are available on the web at http://dev.perl.org/rfc/ =head1 TITLE Subroutines : Pre- and post- handlers for subroutines =head1 VERSION Maintainer: Damian Conway <[EMAIL PROTECTED]> Date: 20 Sep 2000 Last Modified: 29 Sep 2000 Mailing List: [EMAIL PROTECTED] Number: 271 Version: 3 Status: Frozen Frozen since: v2 =head1 ABSTRACT In response to the desiderata set out in RFC 194 and in the Class::Contract module, this RFC proposes a generic "handler" mechanism that can install behaviours around (before or after) a subroutine invocation. =head1 DESCRIPTION =head2 Overall semantics It is proposed to provide a mechanism which allows two sequences of "handlers" to be associated with a specific subroutine or built-in function (hereafter referred to as the I<primary>). One sequence of handlers (the I<prefix sequence>) would be called whenever the primary is invoked, but I<before> the body of the primary is executed. Each handler would itself be a subroutine, and each would be called with the argument list being passed to the primary it prefixes, plus an extra argument (specifically, $_[-1]) representing a slot for the eventual return value of the primary. This last argument would normally have the value C<undef>. The second sequence of handlers would be called I<after> the body of the primary has executed, but I<before> the return value is returned to the invoking scope. Once again, each handler would itself be a subroutine, and each would be called with the same argument list as the primary it postfixes, plus the "return value(s)" argument ($_[-1]). For a postfix handlers, this extra argument would hold a reference to an array containing the value(s) actually returned from its primary. Normally prefix handlers would be prepended to the appropriate prefix sequence, whilst postfix handlers would be appended to the appropriate postfix sequence, thereby preserving the symmetry of pre- and post-fix handlers. The mechanisms for setting up such sequences are described in L<General Syntax of Handler Installers>. =head2 Prefix Handler Semantics A prefix handler may do anything that any other subroutine may do. A typical action might be to trace or log the invocation of the primary (the C<pre> syntax is used to set up a handler, and is explained in L<General Syntax of Handler Installers>). package Foo; # set up a tracing prefix handler for the subroutine &Foo::bar... pre bar { local $" = ',' my @caller = caller; print "Called foo with args (@_[0..$#_-1])\n", "from @caller[1,2]\n", "in contexts: ", join(", ",want); } Note that this (correctly) implies that a handler receives the same information from C<caller> and C<want> (RFC 21) as its primary would. However, C<caller> would be extended to return an extra value when called from a prefix or postfix handler. Specifically, that extra value would be a reference to the handler's primary. Thus a handler could detect which primary it is prefixing (or postfixing): sub handler { my $primary = (caller(0))[10]; ... } or (under RFC 259): sub handler { my $primary = caller->{primary}; ... } Another common usage would be to acquire resources that the primary will use: pre read_file { flock $_[0], LOCK_EX; } (For the obvious complement of this usage, see L<Postfix Handler Semantics>). Note that, in all cases, the return value of the handler is ignored. =head3 Special semantics: changes to arguments Each handler receives the same argument list: the list with which the primary was originally called, plus an extra element representing the return value. If a handler changes one of the original arguments -- through one of the aliases in its @_ array, or by splicing @_ itself -- those changes are passed on to subsequent handlers and to the primary itself. For example: pre tax_payable_on { $_[0] -= 20.00; # routinely underquote sales price } sub tax_payable_on { # now sees prices $20 less than # specified argument printf("Tax: %.2lf", $_[0] * 0.1); } $price = 99.95; tax_payable_on($price); # prints 8.00 $price = 29.95; tax_payable_on($price); # prints 0.99 $price = 9.95; tax_payable_on($price); # prints -1.01 (a profit!) Changes to individual elements of @_ would also be propagated back to the original argument (as usual). Note that it would still be possible to set up a handler that changes its primary's argument list but I<doesn't> propagate the changes back to the original arguments. For example, the following prefix handler allows C<CORE::open> to handle URLs, but I<changes> the second argument: use Regexp::Common; pre CORE::open { $_[1] = "lynx -source $_[1] |" if $_[1] =~ $RE{URL}; } On the other hand, the following handler does exactly the same thing I<without> propagating the changed filename back to the original second argument: pre CORE::open { splice @_, 1, 1, "lynx -source $_[1] |" if $_[1] =~ $RE{URL}; } Note that splicing @_ is also the mechanism for altering the I<number> of arguments passed to the primary. For example, the following prefix handler changes the semantics of C<CORE::open> when called with only one argument: pre CORE::open { splice @_, $#_, 0, $_ if @_ == 2; } With this handler in effect, a call to C<open SOMEFILEHANDLE;> will look for the filename in $_ instead of $SOMEFILEHANDLE. Note that, in the above example, it is important to splice the extra argument I<before> index $#_ (rather than just pushing the extra argument onto the end of @_), because the last element of @_ is always the return value. The ability to rearrange arguments in this way opens the door to some of the intriguing possibilities suggested by the Lingua::Romana::Perligata module. For example, here's how to make C<bless>'s arguments order-independent: pre CORE::bless { splice @_, 0, 2, reverse @_[0..1] if !ref($_[0]) && ref($_[1]); } # and later... $obj = bless 'ClassName', { attr => $val }; =head3 Special semantics: changes to return value slot If a handler assigns the value to the return value slot, either explicitly (C<$_[-1] = $newval>) or implicitly (C<splice @_, -1, 1, $newval>), then the remaining prefix and postfix handlers are still called, but the primary itself is I<not> called. Instead, the value in $_[-1] is used as the return value for the primary. This allows short-circuiting techniques such as memoization to be easily implemented: my %sin_cache; pre CORE::sin { $_[-1] = $sin_cache{$_[0]}; # short-circuit if exists $sin_cache{$_[0]} # if value cached } post CORE::sin { $sin_cache{$_[0]} = $_[-1]; # cache the return val } This feature might also be used to short-circuit upon failure to acquire resources: pre read_file { flock($_[0], LOCK_EX|LOCK_NB) or $_[-1] = undef; # Can't lock file so... # "No data for you!" } Note that this second example (correctly) implies that short-circuiting occurs as a result of an I<assignment> to $_[-1], even if that assignment doesn't change the value of $_[-1]. =head3 Special semantics: exceptions If a handler throws an exception, that exception is immediately propagated back to the calling scope, without having invoked any other handlers (nor the primary itself). This is useful for setting up preconditions on subroutine and methods (for Design-By-Contact programming). For example: package PriorityList; pre pop { my ($self) = @_; croak "Can't pop empty PriorityList" unless @{$self->{list}} > 0; } An exception might also be an appropriate response on resource acquisition failure: pre read_file { flock($_[0], LOCK_EX|LOCK_NB) or croak "Couldn't get immediate lock"; } =head2 Postfix Handler Semantics When called, a postfix handlers may do anything that any other subroutine may do. Just like a prefix handler, it may print out tracing information: package Foo; post bar { local $" = ',' my @caller = caller; print "Finished call to foo from @caller[1,2]\n", "returning: @{$_[-1]}"; } or release resources: post read_file { flock $_[0], LOCK_UN; } or alter the primary's argument list (for example: to effect changes to the original values after the primary has finished with them): post issue_licence { $_[0]--; # decrement licence counter carp "You just used your last licence" if $_[0] == 0; } or change the return value: post tax_payable_on { $_[-1] -= 1.00; # routinely underquote tax payable } sub tax_payable_on { printf("Tax: %.2lf", $_[0] * 0.1); } tax_payable_on(99.95); # prints 9.00 tax_payable_on(29.95); # prints 2.00 tax_payable_on( 9.95); # prints -0.01 or throw an exception (for example, to implement method post-conditions in a DBC system): package PriorityList; post pop { my ($self, $return_val) = @_; croak "pop returned undef (something odd is happening)" unless defined $return_val; } =head2 General Syntax of Handler Installers It is proposed that two built-in functions -- C<pre> and C<post> -- be introduced. These may be used to install prefix or postfix handlers for any subroutine (or for all subroutines in a single package). Each would take two arguments: =over 4 =item 1. Either a subroutine name (possibly as a bareword), or a subroutine reference, or a class name (possibly as a bareword). This argument would be mandatory. =item 2. Either a block or subroutine reference implementing the handler, or a reference to a hash whose single value is a sub reference implementing the handler, or a string or bareword specifying the logical name of a handler. This argument would be optional. =back Thus, in the notation proposed by RFC 128: pre ( [""&]identifier; [""&\%]option) post ( [""&]identifier; [""&\%]option) with all the parsing niceties that implies (e.g. optional barewords in either position, optional separating comma and optional terminating semicolon if the second argument is a block, etc.) =head2 Installing C<pre> Handlers A call to C<pre> would normally install the handler implementation specified by the second argument) as a prefix handler for the subroutine specified by the first argument. The handler would be installed by I<prepending> it to an internal list of handlers for the subroutine (or package). Note that C<pre> (and later, C<post>) are I<not> compile-time directives, but are regular built-in functions that are called as the program executes. This means that they can be used either to set up handlers during compilation (e.g. by calling them in the body of a module) or during execution (by calling them from the main package or from some other subroutine). If the first argument is a string or bareword, the prefix handler is installed for the subroutine or method of the same name in the current package. For example: package Foo; pre bar { print "about to bar()..." }; installs a prefix handler for the subroutine C<&Foo::bar>. If the specified name is qualified (i.e. contains at least one instance of "::"), then the handler is installed for the subroutine in the qualifying package. For example: package Foo; pre Coola::bar, sub { print "about to bar()..." }; installs a prefix handler for the subroutine C<&Coola::bar>. If the first argument ends in "::" and is therefore the name of a package -- rather than a subroutine -- then the handler is installed in a special I<package prefix sequence>, and is shared by every existing (and future) subroutine in the package. For example: pre Cafe::Bar::, sub { print "calling some method of class Cafe::Bar" }; installs a prefix handler that will be invoked whenever I<any> subroutine from package Cafe::Bar is called. If the first argument is a subroutine reference, the handler is installed for the referent. For example: package Foo; pre \&Zan::Zi::bar, sub { print "about to bar()..." }; installs a prefix handler for the subroutine C<&Zan::Zi::bar>, whilst: package Foo; my $anon = sub { print "I am the sub with no name\n" }; pre $anon, sub { print "about to bar()..." }; installs a prefix handler for the anonymous subroutine referred to by $anon. It should be noted that the following two calls are I<not> identical in their effects: pre "Cross::bar", sub { print "high jumping..." }; pre \&Cross::bar, sub { print "high jumping..." }; For an explanation of the difference between them see, L<Inheritance of Handlers>. =head2 Installing C<post> Handlers The semantics of the C<post> function are symmetrical with those of the C<pre> function described in the previous section. C<post> works in the same way as C<pre> -- with the same semantics for various forms of the first argument -- except: =over 4 =item * the handlers it registers are called I<after> the primary returns (instead of before) =item * postfix handlers are called in the same order as they were registered (instead of the reverse order) =item * postfix handlers and package-wide postfix handlers are appended to separate sequences: the C<postfix sequence> and C<package postfix sequence> respectively =back =head2 Logically Identified Handlers In order to later manipulate or remove prefix and postfix handlers, it is convenient to give them a logical name. To this end, the second argument to C<pre> or C<post> may be specified as reference to a one-element hash: pre Lum::bar, { PAIN => sub { print "ouch!..." } }; post Lum::bar, { PAIN => sub { print "...ahhh!" } }; These have exactly the same handler-setting effects as: pre Lum::bar, sub { print "ouch!..." }; post Lum::bar, sub { print "...aaah!" }; except that the specified handlers now have the logical name "PAIN" associated with them. This is important, because at a later point in the code it might be desirable to replace one or both of the handlers (whilst keeping them in the same sequential position): pre Lum::bar, { PAIN => sub { print "OUCH!!!!" } }; That is, when a handler is re-specified with a logical name, if a handler of that name already exists in the appropriate prefix or postfix handler sequence, the new handler replaces the existing one in situ, rather than being prepended or appended to the handler sequence. The logical name could also be used to I<remove> the handler completely: pre Lum::bar, { PAIN => undef }; which would delete the named handler from the sequence. A subsequent C<pre> specification of a handler with the logical name "PAIN" would therefore I<prepend> the new handler to the prefix sequence (since it no longer appears within the subroutine's prefix sequence). If the aim is to remove a named handler, but leave its logical name as a placeholder for future re-instatement, the following should be used instead: pre Lum::bar, { PAIN => sub{} }; =head3 Accessing named handlers It is also possible to retrieve a reference to a named handler, by passing just its logical name as the second argument. For example: $pain_handler = pre Lum::bar, 'PAIN'; $tree_handler = post \&Coola::bar, 'IS_SHADY'; $package_handler = pre Cafe::Bar::, 'CHECK_SUGAR'; Finally, it is possible to retrieve the entire prefix or postfix handler sequence for a particular subroutine, by passing just its name/reference to C<pre>: $pain_handler_array_ref = pre Lum::bar; $tree_handler_array_ref = post \&Coola::bar; $package_handler_array_ref = pre Cafe::Bar::; This returns a reference to the actual handler sequence for the subroutine, in which each element is a reference to a hash whose single key is the logical name of the handler (or "" if it has no logical name) and whose single value is a reference to the handler itself. That is, the structure of the returned value is like this: [ { 'NAME' => sub {...} }, { 'NAME' => sub {...} }, { '' => sub {...} }, { 'NAME' => sub {...} }, { '' => sub {...} }, ] Changes to this array change the actual prefix handling of the corresponding subroutine. For example, a particular module might require that any prefix handler that it installs for C<&CORE::open> (the handler mechanism applies to built-in functions too) is the very last handler that is called. Hence the module cannot use: package LastOpen; sub import { pre CORE::open, \&LastOpen::handler; } as that would I<prepend> the C<&LastOpen::handler> subroutine to the sequence of handlers for C<CORE::open>). Instead, the LastOpen module would use: package LastOpen; sub import { my $open_prefix_sequence = pre Core::open; push @{$open_prefix_sequence}, { "" => \&LastOpen::handler }; } thereby ensuring the prefix handler is called at the end of C<CORE::open>'s prefix sequence. =head2 Inheritance of Handlers L<Installing C<pre> Handlers> noted that the following two calls to C<pre> are not identical in effect: pre \&Cross::bar, sub { print "high jumping..." }; pre "Cross::bar", sub { print "high jumping..." }; The difference between them lies in the way they are treated when the class Cross is inherited by some other class. For example: package Irate; use base 'Cross'; sub bar { print "humbug!" } If a prefix (or postfix) handler for C<&Cross::bar> has been defined via a direct reference to the subroutine: pre \&Cross::bar, sub { die if $_[1] < $qualifying_height } post \&Cross::bar, sub { die unless $_[0]->still_balanced() }; then that handler is I<not> inherited by C<Irate::bar>. Handlers defined in this way are referred to as being I<non-heritable>. However, if the handler was defined using a string or bareword representing C<&Cross::bar>'s name: package Cross; pre bar => sub { die if $_[1] < $qualifying_height }; post 'Cross::bar' { die unless $_[0]->still_balanced() } then that handler I<is> inherited by C<Irate::bar>. Hence, handlers defined in this way are referred to as being I<heritable>. Heritable handlers are associated with the I<identifier> (i.e. name) of a particular subroutine, not with its I<identity> (i.e. reference). Thus it is possible to specify prefix or postfix handlers for a subroutine that exists C<in name only> within a package. Specifically, it is possible to associate a heritable handler with a subroutine that is not defined in a package but is, instead, inherited from some ancestral package. =head3 Inheritance of postfix handlers Inheritance of postfix handlers is simple. The theory of Design-By-Contact programming states that heritable postconditions on a method must be as strong or stronger in a derived class. That is, a derived-class method may make extra guarantees about its behaviour, but may never renege on those guarantees already made by its base-class counterpart. This implies that any postfix handlers inherited by a derived-class method act as if they are appended to its heritable postfix sequence, and thus are called I<after> the derived-class postfix handlers, at which time they have the opportunity to impose additional constraints on the method. For example: package Cat; sub new { my ($class, $name, $weight) = @_; bless { name => $name, weight => $weight }, $class; } post new { die "Anti-matter cat detected" if $_[-1]->{weight} <= 0; } package Tiger; use base 'Cat'; post new { die "Tiger died of shame" if $_[-1]->{name} eq 'Fluffy'; } Given the above definitions, a Cat object need only have a positive weight, but a Tiger object must have both a positive weight I<and> an unembarrassing name. Note that these same inheritance semantics also apply to package-wide postcondition handlers (since they are always specified by name, not by reference). This makes it possible to set up inheritable class invariants. For example: package Cat; pre Cat::, sub { $_[0]->is_dry() or die "Wet cat :-(" } =head3 Inheritance of prefix handlers The inheritance behaviour of prefix handlers is less straightforward. DBC theory states that the heritable preconditions defined on a derived-class method must be no stronger (and may well be weaker) than those for its base-class counterpart. In other words, a derived-class method is allowed to be more permissive in its requirements, but not more demanding. This implies that any heritable prefix handlers inherited from a base class are executed I<disjunctively> with those heritable handlers specified in the derived class. In practice this means that the set of heritable prefix handlers from the base class(es) is invoked within an C<eval> and if none of them throws an exception, that success suffices to satisfy the precondition constraint on the method, in which case the heritable prefix handlers of the derived class are not called (although the non-heritable handlers are still invoked). If the eval terminates because some inherited handler threw an exception, then the full set of prefix handlers (heritable and non-heritable) for the derived-class are tried instead. For example, given: package Cat; sub new { my ($class, $claw_len, $caged) = @_; bless { claw_len => $claw_len, caged => $caged }, $class; } pre new { my ($class, $claw_len) = @_; die "Too dangerous" if $claw_len > 0.5; } package Tiger; use base 'Cat'; post new { my ($class, $claw_len, $caged) = @_; die "Too dangerous" unless $caged; } then a Cat object must have properly trimmed claws, whereas a Tiger may either have properly trimmed claws (i.e. the heritable precondition succeeds) I<or> be safely caged (i.e. the derived precondition succeeds). =head1 MIGRATION ISSUES Minimal. Code with: sub pre { ... } and sub post { ... } might be affected. Surreptitious name mangling by p52p6 (C<the_subroutine_formerly_known_as_pre>???) would probably solve this in most cases. Various modules which provide or make use of mechanisms like this (e.g. the Hook:: hierarchy, or Memoize) might like to take advantage of the standardized facility. =head1 IMPLEMENTATION A prototype (pure Perl) implementation featuring everything except package-wide handlers will be available (as the Hook::SubHandlers module) some time real soon now. It will be archived as: http://www.csse.monash.edu.au/~damian/CPAN/Hook-SubHandlers.tar.gz and subsequently added to the CPAN. As Jarrko points out in RFC 194, the implementation details of true built-in C<pre> and C<post> functions will depend very heavily on what becomes of typeglobs. =head1 ACKNOWLEDGEMENTS Thanks to Uri Guttman and Mark-Jason Dominus for their valuable feedback and suggestions. =head1 REFERENCES RFC 21: Subroutines: Replace C<wantarray> with a generic C<want> function RFC 23: Higher order functions RFC 128: Subroutines: Extend subroutine contexts to include name parameters and lazy arguments RFC 137: Overview: Perl OO should I<not> be fundamentally changed. RFC 194: Standardise Function Pre- and Post-Handling RFC 259: Builtins : Make use of hashref context for garrulous builtins Meyer, B., "Object Oriented Software Construction, 2nd Edition", Prentice Hall. Conway, D., "Lingua::Romana::Perligata -- Perl for the XXI-imus Century", Proc. Perl Conference 4.0, O'Reilly, 2000, pp. 1-16. Also available as: http://www.csse.monash.edu.au/~damian/papers/HTML/Perligata.html
