Read this section for an explanation of the history and purpose of generic programming and the precursor language, Ada. Ada is an Object-Oriented programming language used predominantly in military applications. It shares many of its features – including
its structure and typing, with C++. While both languages were developed in the same year, Ada's rapid adoption by certain circles within the Computer Science world led computer scientists to modify C++ by adopting some of Ada's features in subsequent
C++ releases. These changes are seen as an important evolutionary step in Object-Oriented programming. These materials cover generic programming, precursor Ada, and ANSI/ISO.
The idea of code reuse arises from the necessity for constructing large software systems combining well-established building blocks. The reusability of code improves the productivity and the quality of software. The generic units are one of the ways in which the Ada language supports this characteristic. A generic unit is a subprogram or package that defines algorithms in terms of types and operations that are not defined until the user instantiates them.
Note to C++ programmers: generic units are similar to C++ templates.
For example, to define a procedure for swapping variables of any (non-limited) type:
generic type Element_T is private; -- Generic formal type parameter procedure Swap (X, Y : in out Element_T);
procedure Swap (X, Y : in out Element_T) is Temporary : constant Element_T := X; begin X := Y; Y := Temporary; end Swap;
The Swap subprogram is said to be generic. The subprogram specification is preceded by the generic formal part consisting of the reserved word generic followed by a list of generic formal parameters which may be empty. The entities declared as generic are not directly usable, it is necessary to instantiate them.
To be able to use Swap, it is necessary to create an instance for the wanted type. For example:
procedure Swap_Integers is new Swap (Integer);
Now the Swap_Integers procedure can be used for variables of type Integer.
The generic procedure can be instantiated for all the needed types. It can be instantiated with different names or, if the same identifier is used in the instantiation, each declaration overloads the procedure:
procedure Instance_Swap is new Swap (Float); procedure Instance_Swap is new Swap (Day_T); procedure Instance_Swap is new Swap (Element_T => Stack_T);
Similarly, generic packages can be used, for example, to implement a stack of any kind of elements:
generic Max: Positive; type Element_T is private; package Generic_Stack is procedure Push (E: Element_T); function Pop return Element_T; end Generic_Stack;
package body Generic_Stack is Stack: array (1 .. Max) of Element_T; Top : Integer range 0 .. Max := 0; -- initialise to empty -- ... end Generic_Stack;
A stack of a given size and type could be defined in this way:
declare package Float_100_Stack is new Generic_Stack (100, Float); use Float_100_Stack; begin Push (45.8); -- ... end;
The generic unit declares generic formal parameters, which can be:
When instantiating the generic, the programmer passes one actual parameter for each formal. Formal values and subprograms can have defaults, so passing an actual for them is optional.
Formal parameters of mode in accept any value, constant, or variable of the designated type. The actual is copied into the generic instance, and behaves as a constant inside the generic; this implies that the designated type cannot be limited. It is possible to specify a default value, like this:
generic Object : in Natural := 0;
For mode in out, the actual must be a variable.
One limitation with generic formal objects is that they are never considered static, even if the actual happens to be static. If the object is a number, it cannot be used to create a new type. It can however be used to create a new derived type, or a subtype:
generic Size : in Natural := 0; package P is type T1 is mod Size; -- illegal! type T2 is range 1 .. Size; -- illegal! type T3 is new Integer range 1 .. Size; -- OK subtype T4 is Integer range 1 .. Size; -- OK end P;
The reason why formal objects are nonstatic is to allow the compiler to emit the object code for the generic only once, and to have all instances share it, passing it the address of their actual object as a parameter. This bit of compiler technology is called shared generics. If formal objects were static, the compiler would have to emit one copy of the object code, with the object embedded in it, for each instance, potentially leading to an explosion in object code size (code bloat).
(Note to C++ programmers: in C++, since formal objects can be static, the compiler cannot implement shared generics in the general case; it would have to examine the entire body of the generic before deciding whether or not to share its object code. In contrast, Ada generics are designed so that the compiler can instantiate a generic without looking at its body.)
The syntax allows the programmer to specify which type categories are acceptable as actuals. As a rule of thumb: The syntax expresses how the generic sees the type, i.e. it assumes the worst, not how the creator of the instance sees the type.
This is the syntax of RM 12.5:
formal_type_declaration ::= type defining_identifier[discriminant_part] is formal_type_definition; formal_type_definition ::= formal_private_type_definition | formal_derived_type_definition | formal_discrete_type_definition | formal_signed_integer_type_definition | formal_modular_type_definition | formal_floating_point_definition | formal_ordinary_fixed_point_definition | formal_decimal_fixed_point_definition | formal_array_type_definition | formal_access_type_definiton | formal_interface_type_definition
This is quite complex, so some examples are given below. A type declared with the syntax
type T (<>)denotes a type with unknown discriminants. This is the Ada vernacular for indefinite types, i.e. types for which objects cannot be declared without giving an initial expression. An example of such a type is one with a discriminant without default, another example is an unconstrained array type.
|Generic formal type||Acceptable actual types|
||Any type at all. The actual type can be limited or not, indefinite or definite, but the generic treats it as limited and indefinite, i.e. does not assume that assignment is available for the type.|
||Any nonlimited type: the generic knows that it is possible to assign to variables of this type, but it is not possible to declare objects of this type without initial value.|
||Any nonlimited definite type: the generic knows that it is possible to assign to variables of this type and to declare objects without initial value.|
||Any tagged type, abstract or concrete, limited or not.|
||Any concrete tagged type, limited or not.|
||Any nonlimited tagged type, abstract or concrete.|
||Any nonlimited, concrete tagged type.|
||Any type derived from
||Any type, abstract or concrete, derived from
||Any concrete type, derived from the tagged type
||Any discrete type: integer, modular, or enumeration.|
||Any signed integer type|
||Any modular type|
||Any (non-decimal) fixed point type|
||Any decimal fixed point type|
||Any floating point type|
||Any array type with index of type
||Any access type pointing to objects of type
In the body we can only use the operations predefined for the type category of the formal parameter. That is, the generic specification is a contract between the generic implementor and the client instantiating the generic unit. This is different to the parametric features of other languages, such as C++.
It is possible to further restrict the set of acceptable actual types like so:
|Generic formal type||Acceptable actual types|
||Definite or indefinite types (loosely speaking: types with or without discriminants, but other forms of indefiniteness exist)|
||Types with a discriminant of type DT (it is possible to specify several discriminants, too)|
||Definite types (loosely speaking types without a discriminant or with a discriminant with default value)|
It is possible to pass a subprogram as a parameter to a generic. The generic specifies a generic formal subprogram, complete with parameter list and return type (if the subprogram is a function). The actual must match this parameter profile. It is not necessary that the names of parameters match, though.
Here is the specification of a generic subprogram that takes another subprogram as its parameter:
generic type Element_T is private; with function "*" (X, Y: Element_T) return Element_T; function Square (X : Element_T) return Element_T;
And here is the body of the generic subprogram; it calls parameter as it would any other subprogram.
function Square (X: Element_T) return Element_T is begin return X * X; -- The formal operator "*". end Square;
This generic function could be used, for example, with matrices, having defined the matrix product.
with Square; with Matrices; procedure Matrix_Example is function Square_Matrix is new Square (Element_T => Matrices.Matrix_T, "*" => Matrices.Product); A : Matrices.Matrix_T := Matrices.Identity; begin A := Square_Matrix (A); end Matrix_Example;
It is possible to specify a default with "the box" (
is<>), like this:
generic type Element_T is private; with function "*" (X, Y: Element_T) return Element_T is <>;
This means that if, at the point of instantiation, a function "*" exists for the actual type, and if it is directly visible, then it will be used by default as the actual subprogram.
One of the main uses is passing needed operators. The following example shows this:
generic type Element_Type is private; ... with function "<" (Left : in Element_Type; Right : in Element_Type) return Boolean is <>; procedure Search (Elements : in Array_Type; Search : in Element_Type; Found : out Boolean; Index : out Index_Type'Base) ...
A generic formal can be a package; it must be an instance of a generic package, so that the generic knows the interface exported by the package:
generic with package P is new Q (<>);
This means that the actual must be an instance of the generic package Q. The box after Q means that we do not care which actual generic parameters were used to create the actual for P. It is possible to specify the exact parameters, or to specify that the defaults must be used, like this:
generic -- P1 must be an instance of Q with the specified actual parameters: with package P1 is new Q (Param1 => X, Param2 => Y); -- P2 must be an instance of Q where the actuals are the defaults: with package P2 is new Q;
It is all or nothing: if you specify the generic parameters, you must specify all of them. Similarly, if you specify no parameters and no box, then all the generic formal parameters of Q must have defaults. The actual package must, of course, match these constraints.
The generic sees both the public part and the generic parameters of the actual package (Param1 and Param2 in the above example).
This feature allows the programmer to pass arbitrarily complex types as parameters to a generic unit, while retaining complete type safety and encapsulation. (example needed)
It is not possible for a package to list itself as a generic formal, so no generic recursion is possible. The following is illegal:
with A; generic with package P is new A (<>); package A; -- illegal: A references itself
In fact, this is only a particular case of:
with A; -- illegal: A does not exist yet at this point! package A;
which is also illegal, despite the fact that A is no longer generic.
To instantiate a generic unit, use the keyword new:
function Square_Matrix is new Square (Element_T => Matrices.Matrix_T, "*" => Matrices.Product);
Notes of special interest to C++ programmers:
As a consequence of the above, Ada does not permit template metaprogramming. However, this design has significant advantages:
A generic unit can be nested inside another unit, which itself may be generic. Even though no special rules apply (just the normal rules about generics and the rules about nested units), novices may be confused. It is important to understand the difference between a generic unit and instances of a generic unit.
Example 1. A generic subprogram nested in a nongeneric package.
package Bag_Of_Strings is type Bag is private; generic with procedure Operator (S : in out String); procedure Apply_To_All (B : in out Bag); private -- omitted end Bag_Of_Strings;
To use Apply_To_All, you first define the procedure to be applied to each String in the Bag. Then, you instantiate Apply_To_All, and finally you call the instance.
with Bag_Of_Strings; procedure Example_1 is procedure Capitalize (S : in out String) is separate; -- omitted procedure Capitalize_All is new Bag_Of_Strings.Apply_To_All (Operator => Capitalize); B : Bag_Of_Strings.Bag; begin Capitalize_All (B); end Example_1;
Example 2. A generic subprogram nested in a generic package
This is the same as above, except that now the Bag itself is generic:
generic type Element_Type (<>) is private; package Generic_Bag is type Bag is private; generic with procedure Operator (S : in out Element_Type); procedure Apply_To_All (B : in out Bag); private -- omitted end Generic_Bag;
As you can see, the generic formal subprogram Operator takes a parameter of the generic formal type Element_Type. This is okay: the nested generic sees everything that is in its enclosing unit.
You cannot instantiate Generic_Bag.Apply_To_All directly, so you must first create an instance of Generic_Bag, say Bag_Of_Strings, and then instantiate Bag_Of_Strings.Apply_To_All.
with Generic_Bag; procedure Example_2 is procedure Capitalize (S : in out String) is separate; -- omitted package Bag_Of_Strings is new Generic_Bag (Element_Type => String); procedure Capitalize_All is new Bag_Of_Strings.Apply_To_All (Operator => Capitalize); B : Bag_Of_Strings.Bag; begin Capitalize_All (B); end Example_2;
Example 3. A generic unit that is a child of a nongeneric unit.
Each instance of the generic child is a child of the parent unit, and so it can see the parent's public and private parts.
package Bag_Of_Strings is type Bag is private; private -- omitted end Bag_Of_Strings; generic with procedure Operator (S : in out String); procedure Bag_Of_Strings.Apply_To_All (B : in out Bag);
The differences between this and Example 1 are:
with Bag_Of_Strings.Apply_To_All; -- implicitly withs Bag_Of_Strings, too procedure Example_3 is procedure Capitalize (S : in out String) is separate; -- omitted procedure Capitalize_All is new Bag_Of_Strings.Apply_To_All (Operator => Capitalize); B : Bag_Of_Strings.Bag; begin Capitalize_All (B); end Example_3;
Example 4. A generic unit that is a child of a generic unit
This is the same as Example 3, except that now the Bag is generic, too.
generic type Element_Type (<>) is private; package Generic_Bag is type Bag is private; private -- omitted end Generic_Bag; generic with procedure Operator (S : in out Element_Type); procedure Generic_Bag.Apply_To_All (B : in out Bag); with Generic_Bag.Apply_To_All; procedure Example_4 is procedure Capitalize (S : in out String) is separate; -- omitted package Bag_Of_Strings is new Generic_Bag (Element_Type => String); procedure Capitalize_All is new Bag_Of_Strings.Apply_To_All (Operator => Capitalize); B : Bag_Of_Strings.Bag; begin Capitalize_All (B); end Example_4;
Example 5. A parameterless generic child unit
Children of a generic unit must be generic, no matter what. If you think about it, it is quite logical: a child unit sees the public and private parts of its parent, including the variables declared in the parent. If the parent is generic, which instance should the child see? The answer is that the child must be the child of only one instance of the parent, therefore the child must also be generic.
generic type Element_Type (<>) is private; type Hash_Type is (<>); with function Hash_Function (E : Element_Type) return Hash_Type; package Generic_Hash_Map is type Map is private; private -- omitted end Generic_Hash_Map;
Suppose we want a child of a Generic_Hash_Map that can serialise the map to disk; for this it needs to sort the map by hash value. This is easy to do, because we know that Hash_Type is a discrete type, and so has a less-than operator. The child unit that does the serialisation does not need any additional generic parameters, but it must be generic nevertheless, so it can see its parent's generic parameters, public and private part.
generic package Generic_Hash_Map.Serializer is procedure Dump (Item : in Map; To_File : in String); procedure Restore (Item : out Map; From_File : in String); end Generic_Hash_Map.Serializer;
To read and write a map to disk, you first create an instance of Generic_Hash_Map, for example Map_Of_Unbounded_Strings, and then an instance of Map_Of_Unbounded_Strings.Serializer:
with Ada.Strings.Unbounded; with Generic_Hash_Map.Serializer; procedure Example_5 is use Ada.Strings.Unbounded; function Hash (S : in Unbounded_String) return Integer is separate; -- omitted package Map_Of_Unbounded_Strings is new Generic_Hash_Map (Element_Type => Unbounded_String, Hash_Type => Integer, Hash_Function => Hash); package Serializer is new Map_Of_Unbounded_Strings.Serializer; M : Map_Of_Unbounded_Strings.Map; begin Serializer.Restore (Item => M, From_File => "map.dat"); end Example_5;
This work is licensed under a Creative Commons Attribution-ShareAlike 3.0 License.