NAME Aspect - Aspect-Oriented Programming (AOP) for Perl SYNOPSIS package Person; sub create { # ... } sub set_name { # ... } sub get_address { # ... } package main; use Aspect; ### USING REUSABLE ASPECTS # There can only be one. aspect Singleton => 'Person::create'; # Profile all setters to find any slow ones aspect Profiler => call qr/^Person::set_/; ### WRITING YOUR OWN ADVICE # Defines a collection of events my $pointcut = call qr/^Person::[gs]et_/; # Advice will live as long as $before is in scope my $before = before { print "g/set will soon be next"; } $pointcut; # Advice will live forever, because it is created in void context after { print "g/set has just been called"; } $pointcut; # Advice runs conditionally based on multiple factors before { print "get will be called next, and we are within Tester::run_tests"; } call qr/^Person::get_/ & cflow tester => 'Tester::run_tests'; # Complex condition hijack of a method if some condition is true around { if ( $_->self->customer_name eq 'Adam Kennedy' ) { # Ensure I always have cash $_->return_value('One meeeelion dollars'); } else { # Take a dollar off everyone else $_->proceed; $_->return_value( $_->return_value - 1 ); } } call 'Bank::Account::balance'; DESCRIPTION Aspect-Oriented Programming (AOP) is a programming method developed by Xerox PARC and others. The basic idea is that in complex class systems there are certain aspects or behaviors that cannot normally be expressed in a coherent, concise and precise way. One example of such aspects are design patterns, which combine various kinds of classes to produce a common type of behavior. Another is logging. See for more info. The Perl "Aspect" module closely follows the terminology of the AspectJ project (). However due to the dynamic nature of the Perl language, several "AspectJ" features are useless for us: exception softening, mixin support, out-of-class method declarations, and others. The Perl "Aspect" module is focused on subroutine matching and wrapping. It allows you to select collections of subroutines using a flexible pointcut language, and modify their behavior in any way you want. Terminology Join Point An event that occurs during the running of a program. Currently only calls to subroutines are recognized as join points. Pointcut An expression that selects a collection of join points. For example: all calls to the class "Person", that are in the call flow of some "Company", but *not* in the call flow of "Company::make_report". "Aspect" supports "call()", and "cflow()" pointcuts, and logical operators ("&", "|", "!") for constructing more complex pointcuts. See the Aspect::Pointcut documentation. Advice A pointcut, with code that will run when it matches. The code can be run before or after the matched sub is run. Advice Code The code that is run before or after a pointcut is matched. It can modify the way that the matched sub is run, and the value it returns. Weave The installation of advice code on subs that match a pointcut. Weaving happens when you create the advice. Unweaving happens when the advice goes out of scope. The Aspect An object that installs advice. A way to package advice and other Perl code, so that it is reusable. Features * Create and remove pointcuts, advice, and aspects. * Flexible pointcut language: select subs to match using string equality, regexp, or "CODE" ref. Match currently running sub, or a sub in the call flow. Build pointcuts composed of a logical expression of other pointcuts, using conjunction, disjunction, and negation. * In advice code, you can: modify parameter list for matched sub, modify return value, decide if to proceed to matched sub, access "CODE" ref for matched sub, and access the context of any call flow pointcuts that were matched, if they exist. * Add/remove advice and entire aspects during run-time. Scope of advice and aspect objects, is the scope of their effect. * A reusable aspect library. The Wormhole, aspect, for example. A base class makes it easy to create your own reusable aspects. The Memoize aspect is an example of how to interface with AOP-like modules from CPAN. Why create this module? Perl is a highly dynamic language, where everything this module does can be done without too much difficulty. All this module does, is make it even easier, and bring these features under one consistent interface. I have found it useful in my work in several places: * Saves me from typing an entire line of code for almost every "Test::Class" test method, because I use the TestClass aspect. * I use the Wormhole aspect, so that my methods can acquire implicit context, and so I don't need to pass too many parameters all over the place. Sure I could do it with "caller()" and "Hook::LexWrap", but this is much easier. * Using custom advice to modify class behavior: register objects when constructors are called, save object state on changes to it, etc. All this, while cleanly separating these concerns from the effected class. They exist as an independent aspect, so the class remains unpolluted. The "Aspect" module is different from "Hook::Lexwrap" (which it uses for the actual wrapping) in two respects: * Select join points using flexible pointcut language instead of the sub name. For example: select all calls to "Account" objects that are in the call flow of "Company::make_report". * More options when writing the advice code. You can, for example, run the original sub, or append parameters to it. Using Aspect.pm This package is a facade on top of the Perl AOP framework. It allows you to create pointcuts, advice, and aspects in a simple declarative fastion. You will be mostly working with this package ("Aspect"), and the advice context package. When you "use Aspect;" you will import a family of around a dozen functions. These are all factories that allow you to create pointcuts, advice, and aspects. Pointcuts Pointcuts select join points, so that an advice can run code when they happen. The most common pointcut you will probably use is "call()". For example: $p = call 'Person::get_address'; This selects the calling of "Person::get_address()" as defined in the symbol table during weave-time. The string is a pointcut spec, and can be expressed in three ways: "string" Select only the sub whose name is equal to the spec string. "regexp" Select only the subs whose name matches the regexp. The following will match all the subs defined on the "Person" class, but not on the "Person::Address" class. $p = call qr/^Person::\w+$/; "CODE" ref Select only subs, where the supplied code, when run with the sub name as only parameter, returns true. The following will match all calls to subs whose name isa key in the hash %subs_to_match: $p = call sub { exists $subs_to_match{shift()} } Pointcuts can be combined to form logical expressions, because they overload "&", "|", and "!", with factories that create composite pointcut objects. Be careful not to use the non-overloadable "&&", and "||" operators, because you will get no error message. Select all calls to "Person", which are not calls to the constructor: $p = call qr/^Person::\w+$/ & ! call 'Person::create'; The second pointcut you can use, is "cflow()". It selects only the subs that are in call flow of its spec. Here we select all calls to "Person", only if they are in the call flow of some method in "Company": $p = call qr/^Person::\w+$/ & cflow company => qr/^Company::\w+$/; The "cflow()" pointcut takes two parameters: a context key, and a pointcut spec. The context key is used in advice code to access the context (params, sub name, etc.) of the sub found in the call flow. In the example above, the key can be used to access the name of the specific sub on "Company" that was found in the call flow of the "Person" method.The second parameter is a pointcut spec, that should match the sub required from the call flow. See the Aspect::Pointcut docs for more info. Advice An advice definition is just some code that will run on a match of some pointcut. The "advice" can run around the entire call to allow lexical variables to capture custom information on the way into the function that will be needed when it exists, or it can be more specific and only run before the sub, after the sub runs and returns, after the sub throws an exception, or after the sub runs regardless of the result. Using a more specific advice type will allow the optimiser to generate smaller and faster hooks into your code. You create advice using "around", "before", "after_returning", "after_throwing" or "after()". These take a "CODE" ref, and a pointcut, and install the code on the subs that match the pointcut. For example: after { print "Person::get_address has returned!\n"; } call 'Person::get_address'; The advice code is run with one parameter: the advice context. You use it to learn how the matched sub was run, modify parameters, return value, and if it is run at all. When the advice is created in void context, it remains enabled until the interpreter dies, or the symbol table reloaded. However, advice code can also be applied to matching pointcuts (i.e. the advice is enabled) for only a specific scope by declare it in scalar context and storing the returned guard object. This allows you to neatly control enabling and disabling of advice: SCOPE: { my $advice = before { print "called!\n" } $pointcut; # Do something while the device is enabled } # The advice is now disabled Please note that due to the internal mechanism used to achieve this lexical scoping, you may see a slight loss of memory and a slight slow down of the function, even after the advice has gone out of scope. Lexically creating and removing advice many times is recommended against, and doing so hundreds or thousands of times may result in significant memory consumption of performance loss for the functions matched by your pointcut. Aspects Aspects are just plain old Perl objects, that install advice, and do other AOP-like things, like install methods on other classes, or mess around with the inheritance hierarchy of other classes. A good base class for them is Aspect::Modular, but you can use any Perl object as long as the class inherits from Aspect::Library. If the aspect class exists immediately below the namespace "Aspect::Library", then it can be easily created with the following shortcut. aspect Singleton => 'Company::create'; This will create an Aspect::Library::Singleton object. This reusable aspect is included in the "Aspect" distribution, and forces singleton behavior on some constructor, in this case, "Company::create()". Such aspects share a similar behaviour to advice. If enabled in void context they will be installed permanently, but if called in scalar context they will return a guard object that allows the aspect to be enabled only until the end of the current scope. Internals Due to the dynamic nature of Perl, there is no need for processing of source or byte code, as required in the Java and .NET worlds. The implementation is very simple: when you create advice, its pointcut is matched using "match_define()" to find every sub defined in the symbol table that might match against the pointcut (potentially subject to further runtime conditions). Those that match, will get a special wrapper installed. The wrapper only executes if, during run-time, a compiled context test for the pointcut returns true. The wrapper code creates an advice context, and gives it to the advice code. Some pointcuts like "call()" are static, so the compiled run-time function always returns true, and "match_define()" returns true if the sub name matches the pointcut spec. Some pointcuts like "cflow()" are dynamic, so "match_define()" always returns true, but the compiled run-time function returns true only if some condition within the point is true. To make this process faster, when the advice is installed, the pointcut will not use itself directly for the compiled run-time function but will additionally generate a "curried" (optimised) version of itself. This curried version uses the fact that the run-time check will only be called if it matches the "call()" pointcut pattern, and so no "call()" pointcuts needed to be tested at run-time unless they are in deep and complex nested coolean logic. It also handles collapsing any boolean logic impacted by the safe removal of the "call()" pointcuts. If you use only "call()" pointcuts (alone or in boolean combinations) the currying results in a null test (the pointcut is optimised away entirely) and so the need to make a run-time point test will be removed altogether from the generated advice hooks, reducing call overheads significantly. If your pointcut does not have any static conditions (i.e. "call") then the wrapper code will need to be installed into every function on the symbol table. This is highly discouraged and liable to result in hooks on unusual functions and unwanted side effects. FUNCTIONS TO BE COMPLETED LIMITATIONS Inheritance Support Support for inheritance is lacking. Consider the following two classes: package Automobile; ... sub compute_mileage { ... } package Van; use base 'Automobile'; And the following two advice: before { print "Automobile!\n" } call 'Automobile::compute_mileage'; before { print "Van!\n" } call 'Van::compute_mileage'; Some join points one would expect to be matched by the call pointcuts above, do not: $automobile = Automobile->new; $van = Van->new; $automobile->compute_mileage; # Automobile! $van->compute_mileage; # Automobile!, should also print Van! "Van!" will never be printed. This happens because "Aspect" installs advice code on symbol table entries. "Van::compute_mileage" does not have one, so nothing happens. Until this is solved, you have to do the thinking about inheritance yourself. Performance You may find it very easy to shoot yourself in the foot with this module. Consider this advice: # Do not do this! before { print $_->sub_name; } cflow company => 'MyApp::Company::make_report'; The advice code will be installed on every sub loaded. The advice code will only run when in the specified call flow, which is the correct behavior, but it will be *installed* on every sub in the system. This can be slow. It happens because the "cflow()" pointcut matches *all* subs during weave-time. It matches the correct sub during run-time. The solution is to narrow the pointcut: # Much better before { print $_->sub_name; } call qr/^MyApp::/ & cflow company => 'MyApp::Company::make_report'; TO DO There are a number of things that could be added, if people have an interest in contributing to the project. Documentation * cookbook * tutorial * example of refactoring a useful CPAN module using aspects Pointcuts * new pointcuts: execution, cflowbelow, within, advice, calledby. Sure you can implement them today with Perl treachery, but it is too much work. * need a way to match subs with an attribute, attributes::get() will not work for some reason * isa() support for method pointcuts as Gaal Yahas suggested: match methods on class hierarchies without callbacks * Perl join points: phasic- BEGIN/INIT/CHECK/END * The previous items indicate a need for a real join point specification language Weaving * look into byte code manipulation with B:: modules- could be faster, no need to mess with caller, and could add many more pointcut types. All we need to do for sub pointcuts is add 2 gotos to selected subs. * a debug flag to print out subs that were matched on match_define * warnings when over 1000 methods wrapped * support more pulling (vs. pushing) of aspects into packages: attributes, package specific join points * add whatever constructs required for mocking packages, objects, builtins * allow finer control of advice execution order Reusable Aspects * need better example for wormhole- something less tedius * use Scalar-Footnote for adding aspect state to objects, e.g. in Listenable. Problem is it is still in developer release state * Listenable: when listeners go out of scope, they should be removed from listenables, so you don't have to remember to remove them manually * Listenable: should overload some operator on listenables so that it is easier to add/remove listeners, e.g.: $button += (click => sub { print 'click!' }); * design aspects: DBC, threading, more GOF patterns * middleware aspects: security, load balancing, timeout/retry, distribution * Perl aspects: add use strict/warning/Carp to all matched packages. Actually, Spiffy, Toolkit, and Toolset do this already very nicely. * interface with existing Perl modules for logging, tracing, param checking, generally all things that are AOPish on CPAN. One should be able to use it all through one consistent interface. If I have a good set of pointcuts, I should be able to do all kinds of cross- cutting things with them. * UnderscoreContext aspect: subs that match will, if called with no parameters, get $_, and if in void context, return value will set $_. Allows you to use your subs like builtins, that fall back on $_. So if we have a sub: sub replace_foo { my $in = shift; $in =~ s/foo/bar; $in } Then both calls would be equivalent: $_ = replace_foo($_); replace_foo; * a generic FriendParamAppender aspect, that adds to a param list for affected methods, any object the method requires. Heuristics are applied to find the friend: maybe it is available in the call flow? Perhaps someone in the call flow has an accessor that can get it? Maybe a lexical in some sub in the call flow has it? The point is to cover all cases where we pass objects around, so that we don't have to. A generalization of the wormhole aspect. SUPPORT Please report any bugs or feature requests through the web interface at . INSTALLATION See perlmodinstall for information and options on installing Perl modules. AVAILABILITY The latest version of this module is available from the Comprehensive Perl Archive Network (CPAN). Visit to find a CPAN site near you. Or see . AUTHORS Adam Kennedy Marcel Grünauer Ran Eilam SEE ALSO You can find AOP examples in the "examples/" directory of the distribution. Aspect::Library::Memoize Aspect::Library::Profiler Aspect::Library::Trace COPYRIGHT Copyright 2001 by Marcel Grünauer Some parts copyright 2009 - 2011 Adam Kennedy. This library is free software; you can redistribute it and/or modify it under the same terms as Perl itself.