Background

Design-by-contract is a software engineering technique in which each module of a software system specifies explicitly what input (or data or arguments) it requires, and what output (or information or results) it guarantees to produce in response.

These specifications form the "clauses" of a contract between a module and the client software that uses it. If the client software abides by the input requirements, the module guarantees to produce the correct output. Hence by verifying these clauses at each interaction with a module, the overall behaviour of the system can be confidently predicted.

Design-by-contract reinforces the benefits of modular design techniques by inserting explicit compile-time or run-time checks on a contract. These checks are most often found in object-oriented languages and are typically implemented as pre-conditions and post-conditions on methods, and invariants on classes.

Note that these features differ from simple verification statements such as the C "assert" statement. Conditions and invariants are properties of a class, and are inherited by derived classes.

An additional capacity that is often provided in design-by-contract systems is the ability to selectively disable checking in production code. This allows the contractual testing to be carried out during implementation, without impinging on the performance of the final system.

Adding design-by-contract to Perl

The Class::Contract module provides a framework for specifying methods and attributes for a class (much like the existing class definition modules Class::Struct, Class::MethodMaker, and Class::Generate). Class::Contract allows both per-object and per-class methods and attributes to be defined. Attributes may be scalar-, array-, hash-, or object-based.

Class::Contract differs from other class-specification modules (except Class::Generate) in that it also provides the ability to specify invariant conditions on classes, and pre- and post-conditions on methods and attributes. All of these clauses are fully inheritable, and may be selectively disabled. It differs from all other modules in that it has a cleaner, simpler specification syntax, and -- more importantly -- it enforces encapsulation of object attributes, thereby ensuring that the class contract cannot be subverted.

Defining classes

Class::Contract provides an explicit syntax for defining the attributes, methods, and constructors of a class. The class itself is defined using the "contract" subroutine. "contract" takes a single argument -- a subroutine reference or a block. That block is executed once and the results used to construct and install the various components of the class in the current package:

        package Queue;
        contract {
          # specification of class Queue attributes and methods here
        };

Defining attributes

Attributes are defined within the "contract" block via the "attr" subroutine. Attributes must be given a name, and may also be given a type: "SCALAR", "ARRAY", "HASH", or a class name:

        contract {
                attr 'last';                   # Scalar attribute (by default)
                attr 'lest' => SCALAR;         # Scalar attribute
                attr 'list' => ARRAY;          # Array attribute
                attr 'lost' => HASH;           # Hash attribute
                attr 'lust' => MyClass;        # Object attribute
        };

For each attribute so declared, Class::Contract creates an accessor -- a method that returns a reference to the attribute in question. Code using these accessors might look like this:

        ${$obj->last}++;
        push @{$obj->list}, $newitem;
        print $obj->lost->{'marbles'};
        $obj->lust->after('technology stocks');

Attributes are normally object-specific, but it is also possible to define attributes that are shared by all objects of a class. Class objects are specified by prefixing the call to "attr" with a call to the "class" subroutine:

        class Queue;
        contract {
                class attr 'obj_count';
        };

The accessor for this shared attribute can now be called either as an object method:

        print ${$obj->obj_count};

or as a class method:

        print ${Queue->obj_count};

In order to ensure that the clauses of a class' contract (see below) are honoured, both class and object attributes are only accessible via their accessors, and those accessors may only be called within methods belonging to the same class hierarchy. Objects are implemented as "flyweight scalars" in order to ensure this strict encapsulation is preserved.

Defining methods

Methods are defined in much the same way as attributes. The "method" subroutine is used to specify the name of a method, then the "impl" subroutine is used to provide an implementation for it:

        contract {
                attr list => ARRAY;
                method 'next';
                    impl { shift @{self->list} };
                method 'enqueue';
                    impl { push @{self->list}, $_[1] };
        };

"impl" takes a block (or a reference to a subroutine), which is used as the implementation of the method named by the preceding "method" call. Within that block, the subroutine "self" may be used to return a reference to the object on which the method was called. Unlike, regular OO Perl, the object reference is not passed as the method's first argument. (Note: this change occurred in version 1.10)

Like attributes, methods normally belong to -- and are accessed via -- a specific object. To define methods that belong to the entire class, the "class" qualifier is once again used:

        contract {
                class attr 'obj_count';
                class method 'inc_count';
                        impl { ${self->obj_count}++ };
        };

Note that the "self" subroutine can still be used -- within a class method it returns the appropriate class name, rather than an object reference.

Defining constructors

Class::Contract requires constructors to be explicitly defined using the "ctor" subroutine:

        contract {
                ctor 'new';
                    impl { @{self->list} = ( $_[0] ) }
        };

Note that the implementation section of a constructor doesn't specify code to build or bless the new object. That is taken care of automatically (in order to ensure the correct "flyweight" implementation of the object).

Instead, the constructor implementation is invoked after the object has been created and blessed into the class. Hence the implementation only needs to initialize the various attributes of the "self" object. In addition, the return value of the implementation is ignored: constructor calls always return a reference to the newly created object.

Any attribute that is not initialized by a constructor is automatically "default initialized". By default, scalar attributes remain "undef", array and hash attributes are initialized to an empty array or hash, and object attributes are initialized by having their "new" constructor called (with no arguments). This is the only reasonable default for object attributes, but it is usually advisable to initialize them explicitly in the constructor.

It is also possible to define a "class constructor", which may be used to initialize class attributes:

        contract {
                class attr 'obj_count';
                class ctor;
                        impl { ${self->obj_count} = 0 };
        };

The class constructor is invoked at the very end of the call to "contract" in which the class is defined.

Note too that the class constructor does not require a name. It may, however, be given one, so that it can be explicitly called again (as a class method) later in the program:

        class MyClass;
        contract {
                class attr 'obj_count';
                class ctor 'reset';
                        impl { ${self->obj_count} = 0 };
        };
        # and later...
        MyClass->reset;

Defining destructors

Destructors are also explicitly defined under Class::Contract, using the "dtor" subroutine:

        contract {
                dtor;
                    impl { print STDLOG "Another object died\n" }
        };

As with the constructor, the implementation section of a destructor doesn't specify code to clean up the "flyweight" implementation of the object. Class::Contract takes care of that automatically.

Instead, the implementation is invoked before the object is deallocated, and may be used to clean up any of the internal structure of the object (for example to break reference cycles).

It is also possible to define a "class destructor", which may be used to clean up class attributes:

        contract {
                class attr 'obj_count';
                class dtor;
                    impl { print STDLOG "Total was ${self->obj_count}\n" };
        };

The class destructor is invoked from an "END" block within Class::Contract (although the implementation itself is a closure, so it executes in the namespace of the original class).

Constraining class elements

As described so far, Class::Contract doesn't provide any features that differ greatly from those of any other class definition module. But Class::Contract does have one significant difference: it allows the class designer to specify "clauses" that implement and enforce a contract on the class's interface.

Contract clauses are specified as labelled blocks of code, associated with a particular class, method, or attribute definition.

Class invariants

Classes may be given invariants: clauses than must be satisfied at the end of any method call that is invoked from outside the class itself. For example, to specify that a class's object count attribute must never fall below zero:

        contract {
                invar { ${self->obj_count} >= 0 };
        };

The block following "invar" is treated as if it were a class method that is automatically invoked after every other method invocation. If the method returns false, "croak" is invoked with the error message: 'Class invariant at %s failed' (where the '%s' is replaced by the file and line number at which the invariant was defined).

This error message can be customized, using the "failmsg" subroutine:

        contract {
                invar { ${self->obj_count} >= 0 };
                    failmsg 'Anti-objects detected by invariant at %s';
        };

Once again, the '%s' is replaced by the appropriate file name and line number. A "failmsg" can be specified after other types of clause too (see below).

A class may have as many invariants as it requires, and they may be specified anywhere throughout the the body of the "contract".

Attribute and method pre- and post-conditions

Pre- and post-conditions on methods and attributes are specified using the "pre" and "post" subroutines respectively.

For attributes, pre-conditions are called before the attribute's accessor is invoked, and post-conditions are called after the reference returned by the accessor is no longer accessible. This is achieved by having the accessor return a tied scalar whose "DESTROY" method invokes the post-condition.

Method pre-conditions are tested before their method's implementation is invoked; post-conditions are tested after the implementation finishes (but before the method's result is returned). Constructors are (by definition) class methods and may have pre- and post-conditions, just like any other method.

Both types of condition clause receive the same argument list as the accessor or method implementation that they constrain. Both are expected to return a false value if they fail:

        contract {
                class attr 'obj_count';
                    post { ${&value} > 0 };
                      failmsg 'Anti-objects detected by %s';
                method 'inc_count';
                    post { ${self->obj_count} < 1000000 };
                      failmsg 'Too many objects!';
                    impl { ${self->obj_count}++ };
        };

Note that within the pre- and post-conditions of an attribute, the special "value" subroutine returns a reference to the attribute itself, so that conditions can check properties of the attribute they guard.

Methods and attributes may have as many distinct pre- and post-conditions as they require, specified in any convenient order.

Checking state changes.

Post-conditions and invariants can access the previous state of an object or the class, via the "old" subroutine. Within any post-condition or invariant, this subroutine returns a reference to a copy of the object or class state, as it was just before the current method or accessor was called.

For example, an "append" method might use "old" to verify the appropriate change in size of an object:

        contract {
            method 'append';
                post { @{self->queue} == @{old->queue} + @_ }
                impl { push @{self->queue}, @_ };
        };

Note that the implementation's return value is also available in the method's post-condition(s) and the class's invariants, through the subroutine "value". In the above example, the implementation of "append" returns the new size of the queue (i.e. what "push" returns), so the post-condition could also be written:

        contract {
            method 'append';
                post { ${&value} == @{old->queue} + @_ }
                impl { push @{self->queue}, @_ };
        };

Note that "value" will return a reference to a scalar or to an array, depending on the context in which the method was originally called.

Clause control

Any type of clause may be declared optional:

        contract {
                optional invar { @{self->list} > 0 };
                failmsg 'Empty queue detected at %s after call';
        };

By default, optional clauses are still checked every time a method or accessor is invoked, but they may also be switched off (and back on) at run-time, using the "check" method:

        local $_ = 'Queue';         # Specify in $_ which class to disable
        check my %contract => 0;    # Disable optional checks for class Queue

This (de)activation is restricted to the scope of the hash that is passed as the first argument to "check". In addition, the change only affects the class whose name is held in the variable $_ at the time "check" is called. This makes it easy to (de)activate checks for a series of classes:

        check %contract => 0 for qw(Queue PriorityQueue DEQueue);  # Turn off
        check %contract => 1 for qw(Stack PriorityStack Heap);     # Turn on

The special value '__ALL__' may also be used as a (pseudo-)class name:

        check %contract => 0 for __ALL__;

This enables or disables checking on every class defined using Class::Contract. But note that only clauses that were originally declared "optional" are affected by calls to "check". Non-optional clauses are always checked.

Optional clauses are typically universally disabled in production code, so Class::Contract provides a short-cut for this. If the module is imported with the single argument 'production', optional clauses are universally and irrevocably deactivated. In fact, the "optional" subroutine is replaced by:

        sub Class::Contract::optional {}

so that optional clauses impose no run-time overhead at all.

In production code, contract checking ought to be disabled completely, and the requisite code optimized away. To do that, simply change:

  use Class::Contract;

to

  use Class::Contract::Production;

Inheritance

The semantics of class inheritance for Class::Contract classes differ in several respects from those of normal object-oriented Perl.

To begin with, classes defined using Class::Contract have a static inheritance hierarchy. The inheritance relationships of contracted classes are defined using the "inherits" subroutine within the class's "contract" block:

        package PriorityQueue;
        contract {
                inherits qw( Queue OrderedContainer );
        };

That means that ancestor classes are fixed at compile-time (rather than being determined at run-time by the @ISA array). Note that multiple inheritance is supported.

Method implementations are only inherited if they are not explicitly provided. As with normal OO Perl, a method's implementation is inherited from the left-most ancestral class that provides a method of the same name (though with Class::Contract, this is determined at compile-time).

Constructors are a special case, however. Their "constructive" behaviour is always specific to the current class, and hence involves no inheritance under any circumstances. However, the "initialising" behaviour specified by a constructor's "impl" block is inherited. In fact, the implementations of all base class constructors are called automatically by the derived class constructor (in left-most, depth-first order), and passed the same argument list as the invoked constructor. This behaviour is much more like that of other OO programming languages (for example, Eiffel or C++).

Methods in a base class can also be declared as being abstract:

        contract {
            abstract method 'remove';
                post { ${self->count} == ${old->count}-1 };
        };

Abstract methods act like placeholders in an inheritance hierarchy. Specifically, they have no implementation, existing only to reserve the name of a method and to associate pre- and post-conditions with it.

An abstract method cannot be directly called (although its associated conditions may be). If such a method is ever invoked, it immediately calls "croak". Therefore, the presence of an abstract method in a base class requires the derived class to redefine that method, if the derived class is to be usable. To ensure this, any constructor built by Class::Contract will refuse to create objects belonging to classes with abstract methods.

Methods in a base class can also be declared as being private:

        contract {
            private method 'remove';
                impl { pop @{self->queue} };
        };

Private methods may only be invoked by the class or one of its descendants.

Inheritance and condition checking

Attribute accessors and object methods inherit all post-conditions of every ancestral accessor or method of the same name. Objects and classes also inherit all invariants from any ancestor classes. That is, methods accumulate all the post- and invariant checks that their ancestors performed, as well as any new ones they define for themselves, and must satisfy all of them in order to execute successfully.

Pre-conditions are handled slightly differently. The principles of design-by-contract programming state that pre-conditions in derived classes can be no stronger than those in base classes (and may well be weaker). In other words, a derived class must handle every case that its base class handled, but may choose to handle other cases as well, by being less demanding regarding its pre-conditions.

Meyers suggests an efficient way to achieve this relaxation of constraints without the need for detailed logical analysis of pre-conditions. His solution is to allow a derived class method or accessor to run if either the pre-conditions it inherits are satisfied or its own pre-conditions are satisfied. This is precisely the semantics that Class::Contract uses when checking pre-conditions in derived classes.

A complete example

The following code implements a PriorityStack class, in which elements pushed onto the stack "sink" until they encounter an element with lower priority. Note the use of "old" to check that object state has changed correctly, and the use of explicit dispatch (e.g. "self->Stack::pop") to invoke inherited methods from the derived-class methods that redefine them.

        package PriorityStack;
        use Class::Contract;
        contract {
            # Reuse existing implementation...
            inherits 'Stack';
            # Name the constructor (nothing special to do, so no implementation)
            ctor 'new';
            method 'push';
                # Check that data to be added is okay...
                pre  { defined $_[0] };
                    failmsg 'Cannot push an undefined value';
                pre  { $_[1] > 0 };
                    failmsg 'Priority must be greater than zero';
                # Check that push increases stack depth appropriately...
                post { self->count == old->count+1 };
                # Check that the right thing was left on top...
                post { old->top->{'priority'} <= self->top->{'priority'} };
                # Implementation reuses inherited methods: pop any higher
                # priority entries, push the new entry, then re-bury it...
                impl {
                    my ($newval, $priority) = @_[0,1];
                    my @betters;
                    unshift @betters, self->Stack::pop 
                        while self->count
                           && self->Stack::top->{'priority'} > $priority;
                    self->Stack::push( {'val'=>$newval, priority=>$priority} );
                    self->Stack::push( $_ )  foreach @betters;
                };
            method 'pop';
                # Check that pop decreases stack depth appropriately...
                post { self->count == old->count-1 };
                # Reuse inherited method...
                impl {
                    return  unless self->count;
                    return self->Stack::pop->{'val'};
                };
            method 'top';
                post { old->count == self->count }
                impl {
                    return  unless self->count;
                    return self->Stack::top->{'val'};
                };
        };