9.3 Tagged Type Operations

You can use three options when you define the operations on a tagged type and its descendants. These categories are primitive abstract, primitive nonabstract, and class-wide operations. An abstract operation must be overridden for a nonabstract derived type. A nonabstract operation may be redefined for a subclass. A class-wide operation cannot be overridden by a subclass definition. A class-wide operation can be redefined for the derivation class rooted in the derived type; however, this practice is discouraged because of the ambiguities it introduces in the code. Through careful usage of these options, you can ensure that your abstractions preserve class-wide properties, as discussed in Guideline 9.2.1. As stated above, this principle requires that any type that is visibly derived from some parent type must fully support the semantics of the parent type.

Primitive Operations and Redispatching

guideline

Consider declaring a primitive abstract operation based on the absence of a meaningful "default" behavior.
Consider declaring a primitive nonabstract operation based on the presence of a meaningful "default" behavior.
When overriding an operation, the overriding subprogram should not raise exceptions that are not known to the users of the overridden subprogram.
If redispatching is used in the implementation of the operations of a type, with the specific intent that some of the redispatched-to operations be overridden by specializations for the derived types, then document this intent clearly in the specification as part of the "interface" of a parent type with its derived types.
When redispatching is used (for any reason) in the implementation of a primitive operation of a tagged type, then document (in some project-consistent way) this use in the body of the operation subprogram so that it can be easily found during maintenance.

example

This example (Volan 1994) is intended to show a clean derivation of a square from a rectangle. You do not want to derive Square from Rectangle because Rectangle has semantics that are inappropriate for Square. (For instance, you can make a rectangle with any arbitrary height and width, but you should not be able to make a square this way.) Instead, both Square and Rectangle should be derived from some common abstract type, such as:

Any_Rectangle:
type Figure is abstract tagged
   record
      ...
   end record;
type Any_Rectangle is abstract new Figure with private;
-- No Make function for this; it's abstract.
function Area (R: Any_Rectangle) return Float;
  -- Overrides abstract Area function inherited from Figure.
  -- Computes area as Width(R) * Height(R), which it will
  -- invoke via dispatching calls.
function Width (R: Any_Rectangle) return Float is abstract;
function Height (R: Any_Rectangle) return Float is abstract;
type Rectangle is new Any_Rectangle with private;
function Make_Rectangle (Width, Height: Float) return Rectangle;
function Width (R: Rectangle) return Float;
function Height (R: Rectangle) return Float;
-- Area for Rectangle inherited from Any_Rectangle
type Square is new Any_Rectangle with private;
function Make_Square (Side_Length: Float) return Square;
function Side_Length (S: Square) return Float;
function Width (S: Square) return Float;
function Height (S: Square) return Float;
-- Area for Square inherited from Any_Rectangle
...
-- In the body, you could just implement Width and Height for
-- Square as renamings of Side_Length:
function Width (S: Square) return Float renames Side_Length;
function Height (S: Square) return Float renames Side_Length;
function Area (R: Any_Rectangle) return Float is
begin
  return Width(Any_Rectangle'Class(R)) * Height(Any_Rectangle'Class(R));
  -- Casting [sic, i.e., converting] to the class-wide type causes the function calls to
  -- dynamically dispatch on the 'Tag of R.
  -- [sic, i.e., redispatch on the tag of R.]
end Area;

Alternatively, you could just wait until defining types Rectangle and Square to provide actual Area functions:

type Any_Rectangle is abstract new Figure with private;
-- Inherits abstract Area function from Figure,
-- but that's okay, Any_Rectangle is abstract too.
function Width (R: Any_Rectangle) return Float is abstract;
function Height (R: Any_Rectangle) return Float is abstract;
type Rectangle is new Any_Rectangle with private;
function Make_Rectangle (Width, Height: Float) return Rectangle;
function Width (R: Rectangle) return Float;
function Height (R: Rectangle) return Float;
function Area (R: Rectangle) return Float; -- Overrides Area from Figure
type Square is new Any_Rectangle with private;
function Make_Square (Side_Length: Float) return Square;
function Side_Length (S: Square) return Float;
function Width (S: Square) return Float;
function Height (S: Square) return Float;
function Area (S: Square) return Float;  -- Overrides Area from Figure
...
function Area (R: Rectangle) return Float is
begin
  return Width(R) * Height(R); -- Non-dispatching calls
end Area;
function Area (S: Square) return Float is
begin
  return Side_Length(S) ** 2;
end Area;

rationale

The behavior of a nonabstract operation can be interpreted as the expected behavior for all members of the class; therefore, the behavior must be a meaningful default for all descendants. If the operation must be tailored based on the descendant abstraction (e.g., computing the area of a geometric shape depends on the specific shape), then the operation should be primitive and possibly abstract. The effect of making the operation abstract is that it guarantees that each descendant must define its own version of the operation. Thus, when there is no acceptable basic behavior, an abstract operation is appropriate because a new version of the operation must be provided with each derivation.

All operations declared in the same package as the tagged type and following the tagged type's declaration but before the next type declaration are considered its primitive operations. Therefore, when a new type is derived from the tagged type, it inherits the primitive operations. If there are any operations that you do not want to be inherited, you must choose whether to declare them as class-wide operations (see Guideline 9.3.2) or to declare them in a separate package (e.g., a child package).

Exceptions are part of the semantics of the class. By modifying the exceptions, you are violating the semantic properties of the class-wide type (see Guideline 9.2.1).

There are (at least) two distinct users of a tagged type and its primitives. The "ordinary" user uses the type and its primitives without enhancement. The "extending" user extends the type by deriving a type based on the existing (tagged) type. Extending users and maintainers must determine the ramifications of a possibly incorrect extension. The guidelines here try to strike a balance between too much documentation (that can then easily get out of synch with the actual code) and an appropriate level of documentation to enhance the maintainability of the code.

One of the major maintenance headaches associated with inheritance and dynamic binding relates to undocumented interdependencies among primitive (dispatching) operations of tagged types (the equivalent of "methods" in typical object-oriented terminology). If a derived type inherits some and overrides other primitive operations, there is the question of what indirect effects on the inherited primitives are produced. If no redispatching is used, the primitives may be inherited as "black boxes." If redispatching is used internally, then when inherited, the externally visible behavior of an operation may change, depending on what other primitives are overridden. Maintenance problems (here, finding and fixing bugs) occur when someone overrides incorrectly (on purpose or by accident) an operation used in redispatching. Because this overriding can invalidate the functioning of another operation defined perhaps several levels of inheritance up from the incorrect operation, it can be extremely difficult to track down.

In the object-oriented paradigm, redispatching is often used to parameterize abstractions. In other words, certain primitives are intended to be overridden precisely because they are redispatching. These primitives may even be declared as abstract, requiring that they be overridden. Because they are redispatching, they act as "parameters" for the other operations. Although in Ada much of this parameterization can be done using generics, there are cases where the redispatching approach leads to a clearer object-oriented design. When you document the redispatching connection between the operations that are to be overridden and the operations that use them, you make the intended use of the type much clearer.

Hence, any use of redispatching within a primitive should be considered part of the "interface" of the primitive, at least as far as any inheritor, and requires documentation at the specification level. The alternative (i.e., not providing such documentation in the specification) is to have to delve deep into the code of all the classes in the derivation hierarchy in order to map out the redispatching calls. Such detective work compromises the black-box nature of object-oriented class definitions. Note that if you follow Guideline 9.2.1 on preserving the semantics of the class-wide dispatching operations in the extensions of derived types, you will minimize or avoid the problems discussed here about redispatching.

Class-Wide Operations

guideline

Consider using a class-wide operation (i.e., an operation with parameter[s] of a class-wide type) when an operation can be written, compiled, and tested without knowing all the possible descendants of a given tagged type (Barnes 1996).
Consider using a class-wide operation when you do not want an operation to be inherited and/or overridden.

example

The following example is adapted from Barnes (1996) using the geometric objects from the example of Guideline 9.2.1 and declaring the following functions as primitives in the package specification:

function Area (O : in Object) return Float;

function Area (C : in Circle) return Float;

function Area (S : in Shape) return Float;

A function for computing the moment of a force about a fulcrum can now be created using a class-wide type as follows:

function Moment (OC : Object'Class) return Float is
begin
   return OC.X_Coord*Area(OC);
end Moment;

Because Moment accepts the class-wide formal parameter of Object'Class, it can be called with an actual parameter that is any derivation of type Object. Assuming that all derivations of type object have defined a function for Area, Moment will dispatch to the appropriate function when called. For example:

C : Circle;
M : Float;

...

-- Moment will dispatch to the Area function for the Circle type.
M := Moment(C);

rationale

The use of class-wide operations avoids unnecessary duplication of code. Run-time dispatching may be used where necessary to invoke appropriate type-specific operations based on an operand's tag.

See also Guideline 8.4.3 for a discussion of class-wide pointers in an object-oriented programming framework registry.

Constructors

Ada does not define a unique syntax for constructors. In Ada a constructor for a type is defined as an operation that produces as a result a constructed object, i.e., an initialized instance of the type.

guideline

Avoid declaring a constructor as a primitive abstract operation.
Use a primitive abstract operation to declare an initialization function or constructor only when objects of the inheriting derived types will not require additional parameters for initialization.
Consider using access discriminants to provide parameters to default initialization.
Use constructors for explicit initialization.
Consider splitting the initialization and construction of an object.
Consider declaring a constructor operation in a child package.
Consider declaring a constructor operation to return an access value to the constructed object (Dewar 1995).

example

The following example illustrates the declaration of a constructor in a child package:

--------------------------------------------------------------------------
package Game is
   type Game_Piece is tagged ...
   ...

end Game;
--------------------------------------------------------------------------
package Game.Constructors is
   function Make_Piece return Game_Piece;
   ...
end Game.Constructors;
--------------------------------------------------------------------------

The following example shows how to split the initialization and construction of an object:

type Vehicle is tagged ...

procedure Initialize (Self : in out Vehicle;
                      Make : in     String);

...

type Car is new Vehicle with ... ;
type Car_Ptr is access all Car'Class;

...

procedure Initialize (Self  : in out Car_Ptr;
                      Make  : in     String;
                      Model : in     String) is
begin -- Initialize
   Initialize (Vehicle (Self.all), Make);
   ...
   -- initialization of Car
end Initialize;

function Create (Make  : in String;
                 Model : in String) return Car_Ptr is
   Temp_Ptr : Car_Ptr;
begin -- Create
   Temp_Ptr := new Car;
   Initialize (Temp_Ptr, Make, Model);
   return Temp_Ptr;
end Create;

rationale

Constructor operations for the types in a type hierarchy (assuming tagged types and their derivatives) usually differ in their parameter profiles. The constructor will typically need more parameters because of the added components in the descendant types. You run into a problem when you let constructor operations be inherited because you now have operations for which there is no meaningful implementation (default or overridden). Effectively, you violate the class-wide properties (see Guideline 9.2.1) because the root constructor will not successfully construct a descendant object. Inherited operations cannot add parameters to their parameter profile, so these are inappropriate to use as constructors.

You cannot initialize a limited type at its declaration, so you may need to use an access discriminant and rely on default initialization. For a tagged type, however, you should not assume that any default initialization is sufficient, and you should declare constructors. For limited types, the constructors must be separate procedures or functions that return an access to the limited type.

The example shows using a constructor in a child package. By declaring constructor operations in either a child package or a nested package, you avoid the problems associated with making them primitive operations. Because they are no longer primitive operations, they cannot be inherited. By declaring them in a child package (see also Guidelines 4.1.6 and 4.2.2 on using child packages versus nested packages), you gain the ability to change them without affecting the clients of the parent package (Taft 1995b).

You should put the construction logic and initialization logic in distinct subprograms so that you are able to call the initialization routine for the parent tagged type.

notes

When you extend a tagged type (regardless whether it is an abstract type), you can choose to declare as abstract some of the additional operations. Doing so, however, means that the derived type must also be declared as abstract. If this newly derived type has inherited any functions that name it as the return type, these inherited functions now also become abstract (Barnes 1996). If one of these primitive functions served as the constructor function, you have now violated the first guideline in that the constructor has become a primitive abstract operation.

Equality

guideline

When you redefine the "=" operator on a tagged type, make sure that it has the expected behavior in extensions of this type and override it if necessary.

example

The following example is adapted from the discussion of equality and inheritance in Barnes (1996):

----------------------------------------------------------------------------
package Object_Package is

   Epsilon : constant Float := 0.01;

   type Object is tagged
      record
         X_Coordinate : Float;
         Y_Coordinate : Float;
      end record;

   function "=" (A, B : Object) return Boolean;

end Object_Package;
----------------------------------------------------------------------------
package body Object_Package is

   -- redefine equality to be when two objects are located within a delta
   -- of the same point
   function "=" (A, B : Object) return Boolean is
   begin
      return (A.X_Coordinate - B.X_Coordinate) ** 2
           + (A.Y_Coordinate - B.Y_Coordinate) ** 2 < Epsilon**2;
   end "=";

end Object_Package;

----------------------------------------------------------------------------
with Object_Package;  use Object_Package;
package Circle_Package_1 is
   type Circle is new Object with
      record
         Radius : Float;
      end record;
  function "=" (A, B : Circle) return Boolean;
end Circle_Package_1;
----------------------------------------------------------------------------
package body Circle_Package_1 is

   -- Equality is overridden, otherwise two circles must have exactly
   -- equal radii to be considered equal.
   function "=" (A, B : Circle) return Boolean is
   begin
      return (Object(A) = Object(B)) and
             (abs (A.Radius - B.Radius) < Epsilon);
   end "=";

end Circle_Package_1;
----------------------------------------------------------------------------
with Object_Package;  use Object_Package;
package Circle_Package_2 is

   type Circle is new Object with
      record
         Radius : Float;
      end record;

   -- don't override equality in this package

end Circle_Package_2;
----------------------------------------------------------------------------
with Object_Package;
with Circle_Package_1;
with Circle_Package_2;
with Ada.Text_IO;
procedure Equality_Test is
   use type Object_Package.Object;
   use type Circle_Package_1.Circle;
   use type Circle_Package_2.Circle;
   Object_1 : Object_Package.Object;
   Object_2 : Object_Package.Object;
   Circle_1 : Circle_Package_1.Circle;
   Circle_2 : Circle_Package_1.Circle;
   Circle_3 : Circle_Package_2.Circle;
   Circle_4 : Circle_Package_2.Circle;
begin
   Object_1 := (X_Coordinate => 1.000, Y_Coordinate => 2.000);
   Object_2 := (X_Coordinate => 1.005, Y_Coordinate => 2.000);
   -- These Objects are considered equal.  Equality has been redefined to be
   -- when two objects are located within a delta of the same point.
   if Object_1 = Object_2 then
      Ada.Text_IO.Put_Line ("Objects equal.");
   else
      Ada.Text_IO.Put_Line ("Objects not equal.");
   end if;
   Circle_1 := (X_Coordinate => 1.000, Y_Coordinate => 2.000, Radius => 5.000);
   Circle_2 := (X_Coordinate => 1.005, Y_Coordinate => 2.000, Radius => 5.005);
   -- These Circles are considered equal.  Equality has been redefined to be
   -- when the X-Y locations of the circles and their radii are both within
   -- the delta.
   if Circle_1 = Circle_2 then
      Ada.Text_IO.Put_Line ("Circles equal.");
   else
      Ada.Text_IO.Put_Line ("Circles not equal.");
   end if;
   Circle_3 := (X_Coordinate => 1.000, Y_Coordinate => 2.000, Radius => 5.000);
   Circle_4 := (X_Coordinate => 1.005, Y_Coordinate => 2.000, Radius => 5.005);
   -- These Circles are not considered equal because predefined equality of
   -- the extension component Radius will evaluate to False.
   if Circle_3 = Circle_4 then
      Ada.Text_IO.Put_Line ("Circles equal.");
   else
      Ada.Text_IO.Put_Line ("Circles not equal.");
   end if;
end Equality_Test;

rationale

Equality is applied to all components of a record. When you extend a tagged type and compare two objects of the derived type for equality, the parent components as well as the new extension components will be compared. Therefore, when you redefine equality on a tagged type and define extensions on this type, the parent components are compared using the redefined equality. The extension components are also compared, using either predefined equality or some other redefined equality if appropriate. The behavior of inherited equality differs from the behavior of other inherited operations. When other primitives are inherited, if you do not override the inherited primitive, it can only operate on the parent components of the object of the extended type. Equality, on the other hand, generally does the right thing.

Polymorphism

guideline

Consider using class-wide programming to provide run-time, dynamic polymorphism when constructing larger, reusable, extensible frameworks.
When possible, use class-wide programming rather than variant records.
Use class-wide programming to provide a consistent interface across the set of types in the tagged type hierarchy (i.e., class).
Consider using generics to define a new type in terms of an existing type, either as an extension or as a container, collection, or composite data structure.
Avoid using type extensions for parameterized abstractions when generics provide a more appropriate mechanism.

example

  generic
     type Element is private;
  package Stack is
     ...
  end Stack;

is preferable to:

  package Stack is
     type Element is tagged null record;
     -- Elements to be put on the stack must be of a descendant type
     -- of this type.
     ...
  end Stack;

rationale

Both generics and class-wide types allow a single algorithm to be applicable to multiple, specific types. With generics, you achieve polymorphism across unrelated types because the type used in the instantiation must match the generic formal part. You specify required operations using generic formal subprograms, constructing them as needed for a given instantiation. Generics are ideal for capturing relatively small, reusable algorithms and programming idioms, for example, sorting algorithms, maps, bags, and iterators. As generics become large, however, they become unwieldy, and each instantiation may involve additional generated code. Class-wide programming, including class-wide types and type extension, is more appropriate for building a large subsystem because you avoid the additional generated code and unwieldy properties of generics.

Class-wide programming enables you to take a set of heterogeneous data structures and provide a homogeneous-looking interface across the whole set. See also Guideline 9.2.1 on using tagged types to describe heterogeneous polymorphic data.

In object-oriented programming languages without generic capabilities, it was common to use inheritance to achieve much the same effect. However, this technique is generally less clear and more cumbersome to use than the equivalent explicit generic definition. The nongeneric, inheritance approach can always be recovered using a specific instantiation of the generic. Also see Guidelines 5.3.2 and 5.4.7 for a discussion of self-referential data structures.

Primitive Operations and Redispatching​

guideline​

example​

rationale​

Class-Wide Operations​

guideline​

example​

rationale​

Constructors​

guideline​

example​

rationale​

notes​

Equality​

guideline​

example​

rationale​

Polymorphism​

guideline​

example​

rationale​

Primitive Operations and Redispatching

guideline

example

rationale

Class-Wide Operations

guideline

example

rationale

Constructors

guideline

example

rationale

notes

Equality

guideline

example

rationale

Polymorphism

guideline

example

rationale