A Eiffel: The Essentials

This appendix addresses people who are familiar with the object-oriented approach but do not know Eiffel very well. It introduces all the concepts needed to understand the core of this thesis. However, it is not an exhaustive presentation of the Eiffel language. The reference manual of the current version of Eiffel is the book by Meyer Eiffel: The Language. The next version of the language is defined in the third edition of this book, which is currently available as a draft.

[Meyer 1992]. [Meyer 200?b].

A.1 SETTING UP THE VOCABULARY

First, Eiffel uses vocabulary that sometimes differs from the one used in other object-oriented languages like Java or C#. This section sets up the vocabulary with references to the terminology used in these other languages you may be familiar with.

Structure of an Eiffel program

The basic unit of an Eiffel program is the class. There is no notion of module or assembly like in .NET, no notion of package like in Java (no import-like keyword). Classes are grouped into clusters, which are often associated with a file directory. Indeed, an Eiffel class is stored in a file (with the extension .e); therefore it is natural to associate a cluster with a directory. But this is not compulsory. It is a logical separation, not necessary a physical one. Clusters may contain subclusters, like a file directory may contain subdirectories. An Eiffel system is a set of classes (typically a set of clusters that contain classes) that can be assembled to produce an executable. It is close to what is usually called a program. Eiffel also introduces a notion of universe. It is a superset of the system. It corresponds to all the classes present in the clusters defined in an Eiffel system, even if these classes are not needed for the program execution. Eiffel uses a notion of root class, which is the class that is instantiated first using its creation procedure (the constructor) known as the root creation procedure. An Eiffel system corresponds to the classes needed by the root class directly or indirectly (the classes that are reachable from the root creation procedure). The universe contains all classes in all the clusters specified in the system. The definition of what an Eiffel system contains is done in an Ace file, which is a configuration file written in a language called LACE (Language for Assembly Classes in Eiffel).

374
Classes

EIFFEL: THE ESSENTIALS A.1

A class is a representation of an Abstract Data Type (ADT). Every object is an instance of a class. The object creation uses a so-called creation procedure, which is similar to the notion of constructor in languages such as Java or C#. A class is characterized by a set of features (operations), which may be either attributes or routines. (In Java/C# terminology, features tend to be called members, attributes are called fields and routines are called methods.) Eiffel further distinguishes between routines that return a result (functions) and routines that do not return a result (procedures). This is a classification by implementation: routines vs. attributes, namely computation vs. memory. There is another classification: by role. Features can be either commands (if they do not return a result) or queries (if they do return a result). Then, queries can be either functions if they involve some computation or attributes if the value is stored in memory. The following picture shows the different feature categories:
Command
No result

Procedure

No result

Routine
Returns result Computation

Feature
Returns result Computation

Function

Feature

Feature categories (by role and by implementation)

Query
Memory

Memory

Attribute

Design principles
Eiffel is not only a programming language but also an object-oriented method to build high-quality software. As a method, it brings some design principles: As mentioned above, Eiffel distinguishes between commands and queries. Even if not enforced by any compiler, the Eiffel method strongly encourages following the Command/Query Separation principle: A feature should not both change the objects state and return a result about this object. In other words, a function should be side-effect-free. As Meyer likes to present it: Asking a question should not change the answer. Another important principle, which is Information Hiding: The supplier of a module (typically a class) must select the subset of the modules properties that will be available officially to its client (the public part); the remaining properties build the secret part. The Eiffel language provides the ability to enforce this principle by allowing to define fine-grained levels of availability of a class to its clients. Another principle enforced by the Eiffel method and language is the principle of Uniform Access, which says that all features offered by a class should be available through a uniform notation, which does not betray whether features are implemented through storage (attributes) or through computation (routines). Indeed, in Eiffel, one cannot know when writing x f whether f is a routine or an attribute; the syntax is the same.

[Meyer 1997], p 751.

[Meyer 1997], p 5153.

[Meyer 1997], p 57.

A.2 THE BASICS OF EIFFEL BY EXAMPLE

375

Types
As mentioned earlier, in Eiffel, every object is an instance of a class. There is no exception; even basic types like INTEGERs or REALs are represented as a class. Besides, Eiffel is strongly typed; every program entity is declared of a certain type. A type is based on a class. In case of non-generic classes, type and class are the same. In the case of generic classes, a class is the basis for many different types. The majority of types are reference types, which means that values of a certain type are references to an object, not the object itself. There is a second category of types, called expanded types, where values are actual objects. It is the case of basic types in particular. For example, the value 5 of type INTEGER is really an object of type INTEGER with value 5, not a pointer to an object of type INTEGER with a field containing the value 5.

See Genericity, page 387.

A.2 THE BASICS OF EIFFEL BY EXAMPLE

This section shows you what an Eiffel class looks like with an example.

Structure of a class
The basic structure of an Eiffel class is the following:
class CLASS_NAME feature -- Comment ... feature -- Comment ... end

Basic structure of an Eiffel class

It starts with the keyword class and finishes with the keyword end, and in-between a set of features grouped by feature clauses introduced by the keyword feature and a comment. The comment is introduced by two consecutive dashes and is not compulsory (although recommended to improved readability and understandability).

Book example
An Eiffel class may contain other clauses than the basic ones just shown. For example, it may start with an indexing clause (introduced by the keyword indexing), which should gives general information about the class. The following class BOOK is a typical example of what an Eiffel class looks like. (If you do not understand everything, dont panic; each new notion will be described after the class text.)
indexing description: "Representation of a book" class BOOK create make

Class representation of a book in a library

376
feature {NONE} -- Initialization

EIFFEL: THE ESSENTIALS A.2

make (a_title: like title; some_authors: like authors) is -- Set title to a_title and authors to some_authors. require a_title_not_void: a_title /= Void a_title_not_empty: not a_title is_empty do title := a_title authors := some_authors ensure title_set: title = a_title authors_set: authors = some_authors end

feature -- Access title: STRING -- Title of the book authors: STRING -- Authors of the book -- (if several authors, of the form: -- "first_author, second_author, ...") feature -- Status report is_borrowed: BOOLEAN -- Is book currently borrowed (i.e. not in the library)? feature -- Basic operation borrow is -- Borrow book. require not_borrowed: not is_borrowed do is_borrowed := True ensure is_borrowed: is_borrowed end return is -- Return book. require is_borrowed: is_borrowed do is_borrowed := False ensure not_borrowed: not is_borrowed end invariant title_not_void: title /= Void title_not_empty: not title is_empty end

A.2 THE BASICS OF EIFFEL BY EXAMPLE

Lets have a closer look at this class BOOK: The optional indexing clause was mentioned before. Each entry (there may be several) has two parts: a tag description and the text Representation of a book. The tag name definition is not a keyword; you can use any name, although it is quite traditional to use definition to describe the general purpose of the class. Other commonly used tags include author, date, note, etc. The indexing clause is interesting both for the programmers, the people who will use the class, and for tools which may use this indexed information to do different kinds of things with the class. After the class keyword and class name, BOOK, we find a clause introduced by the keyword create. It introduces the name of each creation procedure (constructor) of the class. Indeed, in Eiffel (contrary to Java or C#), creation procedures do not have a predefined name; they can have any name, although make is commonly used. Here, the class declares only one creation procedure called make. There may be several, in which case the names would be separated by commas. There may also be no create clause at all, which means that the class has the default creation procedure called default_create. (The feature default_create is defined in class ANY from which any Eiffel class inherits. This appendix will come back to this point later when mentioning inheritance.) The class name, NONE, in curly brackets between the keyword feature and the comment -- Initialization is an application of information hiding. It means that the features listed in this feature clause are exported to NONE. NONE is a virtual class that inherits from all classes (it is at the bottom of the class hierarchy). A feature exported to NONE means that no client class can access it. It can only be used within the class or one of its descendants. (It is close to protected in languages such as Java or C#.) It is possible to have any class name between these curly brackets, providing a fine-grained exportation mechanism. You may wonder why a creation procedure is exported to NONE. It does not mean that the class cannot be instantiated because clients cannot access the creation procedure. Not at all. In fact, in Eiffel, creation procedures are not special features. They are normal features that can be called as creation procedure in expressions of the form create my_book make but also as normal features in instructions of the form my_book make (where my_book must already be instantiated) to reinitialize the object for example. Exporting a creation routine to NONE means that it can only be used as a creation procedure; it cannot be called by clients later on as a normal procedure.

377

The basic structure of an Eiffel routine is the following:

routine_name (arg_1: TYPE_1; arg_2: TYPE_2): RETURN_TYPE is -- Comment do ... Implementation here (set of instructions) end

Basic structure of an Eiffel routine

It may also have a require clause after the comment to introduce preconditions and an ensure clause before the end keyword to express postconditions. It is the case of the procedure make. This section will not say more about preconditions and postconditions for the moment. They will be covered in detail in the next section about Design by Contract. A routine may also have a local clause, listing the local variables used in this routine; it is located before the do keyword (after the require clause if any).

See Design by Contract, page 378.

378

EIFFEL: THE ESSENTIALS A.2

If you compare the basic scheme of a routine shown above and the text of make on the previous page, you can deduce that the type of the first argument a_title is like title and the type of the argument some_authors is like authors. What does this mean? It is called anchored types in Eiffel. In like title, title is the anchor. It means that the argument a_title has the same type as the attribute title defined below in the class text. It avoids having to redefine several routines when the type of an attribute changes for example. It will become clearer when we talk about inheritance. But just as a glimpse: imagine BOOK has a descendant class called DICTIONARY and DICTIONARY redefines title to be of type TITLE instead of STRING, then make also needs to be redefined to take an argument of type TITLE. Anchored types avoid this redefinition avalanche by anchoring the type of a certain entity to the type of another query (function or attribute). The anchor can also be Current (a reference to the current object, like this in C# or Java). The text of feature make also shows the syntax for assignment := (not = like in Java and C#; in Eiffel = is the reference equality, like == in Java and C#). You may also encounter syntax of the form ?= which corresponds to an assignment attempt. The semantics of an assignment attempt a ?= b is to check that the type B of b conforms to the type A of a; then, if B conforms to A, b is assigned to a like a normal assignment; if not, a is set to Void. Therefore, the typical scheme of assignment attempts is as follows:
a: A b: B a ?= b if a /= Void then ... end

Typical use of assignment attempts

The typical use of assignment attempts is in conjunction with persistence mechanisms because one cannot be sure of the exact type of the persistent data being retrieved. The next two feature clauses Access and Status report introduce three attributes: title and authors of type STRING, and is_borrowed of type BOOLEAN. The general scheme for an attribute is the following:
attribute_name: ATTRIBUTE_TYPE -- Comment

In the current version of Eiffel, attributes cannot have preconditions or postconditions. It will be possible in the next version of the language. The routines borrow and return follow the same scheme as the feature make described before. It is worth mentioning though the two possible values for BOOLEANs, namely True and False, which are both keywords. The last part of the class is the invariant, which has two clauses in this particular example. Contracts (preconditions, postconditions, class invariants) are explained in the next section about Design by Contract. One last comment about the class BOOK: the use of Void. Void is a feature of type NONE defined in class ANY. It is the equivalent of null in other languages like C# or Java. It corresponds to a reference attached to no object.

Structure of an Eiffel attribute

See Assertions on attributes, page 391.

Design by Contract
Design by Contract is a method of software construction, which suggests building software systems that will cooperate on the basis of precisely defined contracts.

A.2 THE BASICS OF EIFFEL BY EXAMPLE The method

Design by Contract is a method to reason about software that accompanies the programmer at any step of the software development. Even if it is called design by contract, it does not only target the design stage of an application. It is useful as a method of analysis and design, but it also helps during implementation because the software specification will have been clearly stated using assertions (boolean expressions). Design by Contract is also useful to debug and test the software against this specification. The idea of Design by Contract is to make the goal of a particular piece of software explicit. Indeed, when developers start a new project and build a new application, it is to satisfy the need of a client, match a certain specification. The Design by Contract method suggests writing this specification down to serve as a contract between the clients (the users) and the suppliers (the programmers). This idea of contract defined by some obligations and benefits is an analogy with the notion of contract in business: the supplier has some obligations to his clients and the clients also have some obligations to their supplier. What is an obligation for the supplier is a benefit for the client, and conversely.

379

Different kinds of contracts

There are three main categories of contracts: preconditions, postconditions, and class invariants: Preconditions are conditions under which a routine will execute properly; they have to be satisfied by the client when calling the routine. They are an obligation for the clients and a benefit for the supplier (which can rely on them). A precondition violation is the manifestation of a bug in the client (which fails to satisfy the precondition). Precondition clauses are introduced by the keyword require in an Eiffel routine:
routine_name (arg_1: TYPE_1; arg_2: TYPE_2): RETURN_TYPE is -- Comment require tag_1: boolean_expression_1 tag_2: boolean_expression_2 do ... Implementation here (set of instructions) end

Structure of an Eiffel routine with precondition

Each precondition clause is of the form tag: expression where the tag can be any identifier and the expression is a boolean expression (the actual assertion). The tag is optional; but it is very useful for documentation and debugging purposes. Postconditions are properties that are satisfied at the end of the routine execution. They are benefits for the clients and obligations for the supplier. A postcondition violation is the manifestation of a bug in the supplier (which fails to satisfy what it guarantees to its clients).

380

EIFFEL: THE ESSENTIALS A.2

Postcondition clauses are introduced by the keyword ensure in an Eiffel routine:
routine_name (arg_1: TYPE_1; arg_2: TYPE_2): RETURN_TYPE is -- Comment do ... Implementation here (set of instructions) ensure tag_1: boolean_expression_1 tag_2: boolean_expression_2 end

Structure of an Eiffel routine with postcondition

Of course, a routine may have both preconditions and postconditions; hence both a require and an ensure clause (like in the previous BOOK class). Class invariants capture global properties of the class. They are consistency constraints applicable to all instances of a class. They must be satisfied after the creation of a new instance and preserved by all the routines of the class. More precisely, it must be satisfied after the execution of any feature by any client. (This rule applies to qualified calls of the form x f only, namely client calls. Implementation calls unqualified calls and calls to non-exported features do not have to preserve the class invariant.) Class invariants are introduced by the keyword invariant in an Eiffel class

class CLASS_NAME feature -- Comment ... invariant tag_1: boolean_expression_1 tag_2: boolean_expression_2 end

Structure of an Eiffel class with class invariant

There are three other kinds of assertions: Check instructions: Expressions ensuring that a certain property is satisfied at a specific point of a methods execution. They help document a piece of software and make it more readable for future implementers. In Eiffel, check instructions are introduced by the keyword check as follows:
routine_name (arg_1: TYPE_1; arg_2: TYPE_2): RETURN_TYPE is -- Comment do ... Implementation here (set of instructions) check tag_1: boolean_expression_1 tag_2: boolean_expression_2 end ... Implementation here (set of instructions) end

Structure of an Eiffel routine with check instruction

Loop invariants: Conditions, which have to be satisfied at each loop iteration and when exiting the loop. They help guarantee that a loop is correct. Loop variants: Integer expressions ensuring that a loop is finite. It decreases by one at each loop iteration and has to remain positive. This appendix will show the syntax of loop variants and invariants later when introducing the syntax of loops.

See Syntax of loops, page 383.

A.3 MORE ADVANCED EIFFEL MECHANISMS Benefits

The benefits of Design by Contract are both technical and managerial. Among other benefits we find: Software correctness: Contracts help build software right in the first place (as opposed to the more common approach of trying to debug software into correctness). This first use of contracts is purely methodological: the Design by Contract method advises you to think about each routines requirements and write them as part of your software text. This is only a method, some guidelines for software developers, but it is also perhaps the main benefit of contracts, because it helps you design and implement correct software right away. Documentation: Contracts serve as a basis for documentation: the documentation is automatically generated from the contracts, which means that it will always be up-to-date, correct and precise, exactly what the clients need to know about. Debugging and testing: Contracts make it much easier to detect bugs in a piece of software, since the program execution just stops at the mistaken points (faults will occur closer to the source of error). It becomes even more obvious with assertions tags (i.e. identifiers before the assertion text itself). Contracts are also of interest for testing because they can serve as a basis for black-box test case generation. Management: Contracts help understand the global purpose of a program without having to go into the code in depth, which is especially appreciable when you need to explain your work to less-technical persons. It provides a common vocabulary and facilitates communication. Besides, it provides a solid specification that facilitates reuse and component-based development, which is of interest for both managers and developers.

381

A.3 MORE ADVANCED EIFFEL MECHANISMS

Lets describe more advanced Eiffel mechanisms, typically the facilities on which the pattern library relies on.
See Pattern Library, page 26.

Book library example

This section uses an example to introduce these mechanisms. Because we talked about books in the previous section, here is the example of a library where users can borrow and return books. Here is a possible implementation of an Eiffel class LIBRARY:
indexing description: "Library where users can borrow books" class LIBRARY inherit ANY redefine default_create end

Class representation of a book library

382
feature {NONE} -- Initialization default_create is -- Create books. do create books make end

EIFFEL: THE ESSENTIALS A.3

feature -- Access books: LINKED_LIST [BOOK] -- Books available in the library feature -- Element change extend (a_book: BOOK) is -- Extend books with a_book. require a_book_not_void: a_book /= Void a_book_not_in_library: not books has (a_book) do books extend (a_book) ensure one_more: books count = old book count + 1 book_added: books last = a_book end

..

remove (a_book: BOOK) is -- Remove a_book from books. require a_book_not_void: a_book /= Void book_in_library: books has (a_book) do books start books search (a_book) books remove ensure one_less: books count = old books count 1 book_not_in_library: not books has (a_book) end

. . .

..

feature -- Output display_books is -- Display title of all books available in the library. do if books is_empty then io put_string ("No book available at the moment") else from books start until books after loop io put_string (books item title) books forth end end end

..

. ..

.. .

feature -- Basic operation

A.3 MORE ADVANCED EIFFEL MECHANISMS

borrow_all is -- Borrow all books available in the library. do from books start until books after loop books item borrow books forth end ensure end invariant books_not_void: books /= Void no_void_book: not books has (Void) end

383

.. . .

all_borrowed: books for_all (agent {BOOK} is_borrowed)

This example introduces two controls we had not encountered before: conditional structures (in feature display_books) and loops (in display_books and borrow_all). Here is the syntax scheme for conditional structures:
if some_condition_1 then do_something_1 elseif some_condition_2 then do_something_2 else do_something_else end

Syntax of conditional structures

The elseif and else branches are optional. Here is the syntax scheme for loops (there is only one kind of loops in Eiffel):
from initialization_instructions invariant loop_invariant variant loop_variant until exit_condition loop loop_instructions end

Syntax of loops

The variant and invariant clauses are optional. The from clause is compulsory but it may be empty. Lets now discover the other Eiffel techniques used in this example:

Inheritance
The class LIBRARY contains a clause we have not seen yet: an inherit clause. It introduces the classes from which class LIBRARY inherits. Here LIBRARY inherits from ANY.

384

EIFFEL: THE ESSENTIALS A.3

ANY is the class from which any Eiffel class inherits. If you remember, we saw before that NONE is the class that inherits from any Eiffel class, which means we have a closed hierarchy here:

ANY

See Business Object Notation (BON), page 394.

All Eiffel classes

Global inheritance structure

NONE
In fact, one does not need to write that a class inherits from ANY; it is the default. Here, the class LIBRARY expresses the inheritance link explicitly to be able to redefine the feature default_create inherited from ANY. Redefinition allows changing the body of a routine (the do clause) and changing the routine signature if the new signature conforms to the parent one (which basically means that the base class of the new argument types inherits from the base class of the argument types in the parent, and same thing for the result type if it is a function). The routine body can be changed, but it still has to follow the routines contract. The Design by Contract method specifies precise rules regarding contracts and inheritance: preconditions are or-ed and can only be weakened, postconditions are and-ed and can only be strengthened (not to give clients any bad surprise). Class invariants are also and-ed in descendant classes (subclasses). As suggested by the inheritance figure on the previous page, a class may inherit from one or several other classes, in which case we talk about multiple inheritance. Contrary to languages like C# or Java, Eiffel supports multiple inheritance of classes. (It is restricted to interfaces in the Java/C# worlds.) Allowing multiple inheritance means that a class may get features from two different parents (superclasses) with the same name. Eiffel provides a renaming mechanisms to handle name clashes. For example, if we have a class C that inherits from A and B, and A and B both define a feature f, it is possible to rename the feature f from A as g to solve the name clash. Here is the Eiffel syntax:
class C inherit A rename f as g end B feature -- Comment ... end

Redefinition of return type and argument types follows the rules of covariance. See Non-conforming inheritance, page 391. See Design by Contract, page 378.

Renaming mechanism

A.3 MORE ADVANCED EIFFEL MECHANISMS

As mentioned above, Eiffel also provides the ability to redefine features inherited from a parent using a redefine clause. In the redefined version, it is possible to call the version from the parent by using the Precursor. Here is an example:
class C inherit A redefine f end feature -- Comment f (args: SOME_TYPE) is -- Comment do Precursor {A} (args) ... end ... end

385

Redefinition with call to Precursor

Eiffel also enables to undefine a routine, that is to say making it deferred (abstract). A deferred feature is a feature that is not implemented yet. It has no do clause but a deferred clause. Yet it can have routine preconditions and postconditions. Here is the general structure of a deferred routine:
routine_name (arg_1: TYPE_1; arg_2: TYPE_2): RETURN_TYPE is -- Comment require ... Some precondition clauses deferred ensure ... Some postcondition clauses end

Structure of a deferred routine

A class that has at least one deferred feature must be declared as deferred. In the current version of Eiffel, it is also true that a deferred class must have at least one deferred feature. In the next version of the language, it will be possible to declare a class deferred even if all its features are effective (implemented). Besides, contrary to other languages like Java or C#, a deferred class can declare attributes. It can also express a class invariant. To come back to inheritance and undefinition, here is the syntax that allows to make a routine deferred when inheriting it from a parent class:
deferred class C inherit A undefine f end feature -- Comment ... end

Undefinition mechanism

386

EIFFEL: THE ESSENTIALS A.3

This example supposes that class A declares a feature f, which is effective in A. The class C inherits f from A and makes it deferred by listing it in the clause undefine. Therefore, class C must be declared as deferred (it has at least one deferred feature). Another problem that arises with multiple inheritance is the case of repeated inheritance (also known as the diamond structure). For example, we have a class D that inherits from B and C, which themselves inherit from A. Class A has a feature f. B renames it as g and redefines it. C leaves f unchanged. Here is the corresponding class diagram:

See Business Object Notation (BON), page 394.

B f g g++ D

Repeated inheritance (diamond structure)

The class D inherits two different features from the same feature f in A. The problem occurs when talking about dynamic binding: which feature should be applied? There is another inheritance adaptation clause called select, which provides the ability to say: in case there is a conflict, use the version coming from this parent. For example, if we want to retain the feature g coming from B, we would write:
class D inherit B select g end C feature -- Comment ... end

Selection mechanism

The last inheritance adaptation clause is the export clause. It gives the possibility to restrict or enlarge the exportation status of an inherited routine. Here is the general scheme:
class B inherit A export {NONE} f -- Makes f secret in B (it may have been exported in A) {ANY} g -- Makes g exported in B (it may have been secret in A) {D, E} x, z -- Makes x and z exported to only classes D and E

Export mechanism

end feature -- Comment ... end

A.3 MORE ADVANCED EIFFEL MECHANISMS

Eiffel offers a keyword all to designate all the features of a class (defined in the class itself and inherited). It is often used to make all features inherited from a parent class secret (in case of implementation inheritance for example):
class B inherit A export end feature -- Comment ... end {NONE} all

387

Making all inherited features secret

The last point we need to see about inheritance is the order of adaptation clauses. Here it is:
class D inherit B rename e as k export {NONE} all {D} r undefine m redefine b select g end C feature -- Comment ... end

Order of the inheritance adaptation clauses

Genericity
Lets come back to our class LIBRARY. The next mechanism we had not encountered before is genericity. Indeed, the class LIBRARY declares a list of books:
books: LINKED_LIST [BOOK] -- Books available in the library
See Class representation of a book library, page 381.

Attribute of a generic type

The attribute books is of type LINKED_LIST [BOOK], which is derived from the generic class LINKED_LIST [G], where G is the formal generic parameter. A generic class describes a type template. One must provide a type, called actual generic parameter (for example here BOOK), to derive a directly usable type like LINKED_LIST [BOOK] . Genericity is crucial for software reusability and extendibility. Besides, most componentized versions of design patterns rely on genericity. The example of LINKED_LIST [G] is a case of unconstrained genericity. There are cases where it is needed to impose a constraint on the actual generic parameters. Lets take an example. Say we want to represent vectors as a generic class VECTOR [G], which has a feature plus to be able to add two vectors, and we also want to be able to have vectors of vectors like VECTOR [VECTOR [INTEGER]].

388

EIFFEL: THE ESSENTIALS A.3

Lets try to write the feature plus of class VECTOR [G]. In fact, it is unlikely to be called plus; it would rather be an infix feature +, which is a special notation to allow writing vector_a + vector_b instead of vector_a plus (vector_b). There also exists a prefix notation to be able to write - my_integer for example.

To come back to class VECTOR [G], a first sketch may look as follows:
class VECTOR [G] create make feature {NONE} -- Initialization make (max_index: INTEGER) is -- Initialize vector as an array with indexes from 1 to max_index. require ... do ... end feature -- Access count: INTEGER -- Number of elements in vector item (i: INTEGER): G is -- Vector element at index i require ... do ... end infix + (other: like Current): like Current is -- Sum with other require other_not_void: other /= Void consistent: other count = count local i: INTEGER do create Result make (1, count) from i := 1 until i > count loop Result put (item (i) + other item (i), i) -- Requires an operation + on elements of type G. i := i + 1 end ensure sum_not_void: Result /= Void end

Addable vectors

. .

... invariant ... end

A.3 MORE ADVANCED EIFFEL MECHANISMS

Our implementation of the + operation requires a + operation on elements of type G, which means that we cannot accept any kind of actual generic parameters. We need them to have such a feature +. Here is when constrained genericity comes into play: it allows to say that actual generic parameters must conform to a certain type, for example here NUMERIC. (It basically means that the base class of the actual generic parameter must inherit from class NUMERIC.) Here is the corresponding syntax:
class VECTOR [G > NUMERIC]

389

See Non-conforming inheritance, page 391.

It is not allowed to have multiple constraints, say class C [G > {A, B}]. It may be allowed in the next version of the language. Another kind of constraint is to impose that actual generic parameters must have certain creation procedures. For example, the notation:
class MY_CLASS [G > ANY create default_create end]

[Meyer 200?b].

means that any actual generic parameter of MY_CLASS must conform to ANY and expose default_create in its list of creation procedures (introduced by the keyword create in an Eiffel class text).

Agents
There is still one mysterious point in the class LIBRARY, the postcondition of borrow_ all:
ensure all_borrowed: books for_all (agent {BOOK} is_borrowed)
See Class representation of a book library, page 381.

Use of agents in contracts

What the postcondition of borrow_all does is to test for all items of type BOOK in the list books whether it is_borrowed. The postcondition will evaluate to True if all books are borrowed. But what does this agent mean? An agent is an encapsulation of a routine ready to be called. (To make things simple, you may consider an agent as a typed function pointer.) It is used to handle event-driven development. For example, if you want to associate an action with the event button is selected, you will write:
my_button select_actions extend (agent my_routine)

Events with agents

where my_routine is a routine of the class where this line appears. A typical agent expression is of the form:
agent my_function (?, a, b)

where a and b are closed arguments (they are set at the time of the agents definition) whereas ? is an open argument (it will be set at the time of any call to the agent). It is also possible to construct an agent with a routine that is not declared in the class itself. The syntax becomes:
agent some_object some_routine (?, a, b)

Open and closed arguments

where some_object is the target of the call. It is a closed target. The agent expression used in the postcondition of borrow_all had an open target of type BOOK:
agent {BOOK} is_borrowed

Open and closed arguments Open target

390
To call an agent, you simply have to write:
my_agent call ([maybe_some_arguments])

EIFFEL: THE ESSENTIALS A.4

where call is a feature of class ROUTINE. Here is the class diagram of the three agent types:

Calling an agent
See Business Object Notation (BON), page 394.

ROUTINE [BASE, ARGS -> TUPLE] call

Agent types
PROCEDURE [BASE, ARGS -> TUPLE] FUNCTION [BASE, ARGS -> TUPLE, RES]

An important property of the Eiffel agent mechanism: it is completely type-safe. Agents are a recent addition to the Eiffel language. They were introduced in 1999. Therefore, they are not described in the reference manual Eiffel: The Language. However, there is an entire chapter of the next revision of the book devoted to agents.

[Dubois 1999]. See chapter 25 of [Meyer 200?b].

Performance
Using agents has a performance overhead. To measure that overhead, I performed one million direct calls to a routine that does nothing and one million agent calls to the same routine. Without agents, the duration was two seconds (2s per call); with agents, it was fourteen seconds (14s per call), thus seven times as slow. But in practice, one calls routines that do something. Therefore I added an implementation to the routine (a loop that executes do_nothing, twenty times) and did the same tests. The results were the following: 33s (33s per call) without agents; 46s (46s per call) with agents; thus 1.4 times as slow. In a real application, the number of agent calls in the whole code will be less significant. Typically, no more than 5% of the feature calls will be calls to agents. Therefore the execution of an application using agents will be about 0.07 times as slow, which is a acceptable performance overhead in most cases.

do_nothing is a procedure defined in ANY which does nothing.

A.4 TOWARDS AN EIFFEL STANDARD

Eiffel is currently being standardized through the ECMA international organization. The working group in charge of the Eiffel standardization examines some issues of the Eiffel language and discusses possible extensions. I am an active member of the group; I am responsible for preparing the meetings agenda and for writing the meeting minutes.
[ECMA-Web].

ECMA standardization
The process started in March 2002 when ECMA accepted the proposal to standardize Eiffel. It resulted in the creation of a new working group TG4, part of the Technical Committee 39 (originally scripting languages, although this is just for historical reasons). The first group meeting took place at ETH Zurich in Switzerland in June 2002. Jan van den Beld, secretary general of ECMA, took part in the meeting and explained how the standardization work should proceed.

A.4 TOWARDS AN EIFFEL STANDARD

The goal is to have a first draft of the standard ready in 2004. To achieve this goal, the group meets at least four times a year and keeps interacting by email in between. The group members are all Eiffel experts with several years experience and coming from both academia and industry. For the moment, members include ETH Zurich with Bertrand Meyer and me, LORIA in France with Dominique Colnet, Monash University, in Australia, represented by Christine Mingins. This is for the academia. Industry members are AXA Rosenberg with Mark Howard and ric Bezault, Eiffel Software with Emmanuel Stapf and Bertrand Meyer, and Enea Data with Kim Waldn and Paul Cohen.

391

New mechanisms
This section presents some extensions to the Eiffel language that have been preapproved by the committee. (To be finally approved, a mechanism needs to be implemented in at least one Eiffel compiler. For the four extensions presented here, there is no doubt about the final acceptation. Besides, three of these extensions are already implemented at least in part.)

Assertions on attributes
As mentioned before, attributes cannot have contracts in the current version of Eiffel. Only routines can have contracts. This discrimination between routines and attributes goes against the Uniform Access principle presented at the beginning of this appendix. Indeed, clients should not have to know whether a feature is implemented by computation or by storage; they just need to know that the class offers this service, no matter what its implementation is. Therefore the team agreed to introduce a new syntax for attributes, with a new keyword attribute, that allows putting preconditions and postcondition:
attribute_name: ATTRIBUTE_TYPE is -- Comment require ... Some precondition clauses attribute ensure ... Some postcondition clauses end
See Structure of an Eiffel attribute, page 378. See Design principles, page 374.

Assertions on attributes

The current notation:

attribute_name: ATTRIBUTE_TYPE -- Comment

becomes a shortcut for:

attribute_name: ATTRIBUTE_TYPE is -- Comment attribute end

Non-conforming inheritance
In the current version of Eiffel, inheritance always brings conformance. For example, if a class B inherits from a class A, then type B conforms to type A. As a consequence, an assignment like a1 := b1 where b1 is of type B and a1 is declared of type A is allowed. It is also possible to pass an instance of type B as argument of a feature expecting an A (for example, f (b1) with f declared as f (arg: A)).

392

EIFFEL: THE ESSENTIALS A.4

However, sometimes, it may be useful to have inheritance without conformance, namely having subclassing without polymorphism. Non-conforming gives descendant classes access to the parents features but forbids such assignments and arguments passing as those described above (between entities of the descendant type and entities of the parent type). This facility is useful only in specific cases like implementation inheritance. For example, we could have a class TEXTBOOK that inherits from BOOK. This inheritance relation should be conformant; we want both subclassing and polymorphism. But this class TEXTBOOK may need access to some helper features defined in a class TEXTBOOK_SUPPORT. One way to get access to these features is to declare an attribute of type TEXTBOOK_SUPPORT and have a client relationship. Another way to do it is to use what is usually called implementation inheritance, that is to say inheriting from TEXTBOOK_SUPPORT just to be able to use these features. In that case, we just need subclassing (to get the features), not conformance (we dont want to assign an entity of type TEXTBOOK to an entity of type TEXTBOOK_SUPPORT). Hence the use of non-conforming inheritance. Another example taken from EiffelBase is the class ARRAYED_LIST, which inherits from LIST and ARRAY. For the moment, both inheritance links are conformant (there is no other choice!). But what we really want is a class ARRAYED_ LIST that inherits from LIST in a conformant way and a non-conformant link between ARRAYED_LIST and ARRAY (just to get the features of ARRAY, which are useful for the implementation of ARRAYED_LIST). The syntax is not completely fixed yet. For the moment, the idea is to use the keyword expanded in front of the parent class name in the inherit clause to specify that the inheritance link is non-conformant (hence the name expanded inheritance, which is sometimes used):
class B inherit C expanded A feature -- Comment ... end
What is the relation between non-conforming inheritance and expanded types? If a class B inherits from a class A, which is declared as expanded, then there is no conformance between B and A. Hence the use of the keyword expanded.
[EiffelBase-Web].

Possible syntax of nonconforming inheritance

Non-conforming inheritance is currently being implemented into the SmartEiffel compiler.

Automatic type conversion

The third mechanism is automatic type conversion. The goal is to be able to add, for example a complex number and an integer, or a real and an integer. To do this, the group decided to include a two-side conversion mechanism into the Eiffel language. It is possible to convert from and to convert to a particular type thanks to a new clause convert.

A.4 TOWARDS AN EIFFEL STANDARD

Here is the syntax:
class MY_CLASS create from_type_1 convert from_type_1 ({TYPE_1}), to_type_2: {TYPE_2} feature -- Conversion from_type_1 (arg: TYPE_1) is -- Convert from arg. do ... end to_type_2: TYPE_2 is -- New object of type TYPE_2 do ... end end

393

Type conversion mechanism

If you have an instance of type TYPE_1 and you want to add it to an instance of MY_ CLASS, the instance of TYPE_1 will be automatically converted to an instance of MY_ CLASS by calling the conversion procedure from_type_1, which needs to be declared as a creation procedure. On the other hand, if you have an instance of type MY_CLASS and you want to add it to an instance of type TYPE_2 , your instance of MY_CLASS will be automatically converted to TYPE_2 by calling the function to_type_2. Here are typical cases that this new mechanism permits to write:
my_attribute: MY_TYPE attribute_1: TYPE_1 attribute_2: TYPE_2 my_attribute + attribute_1 -- Equivalent to: -- my_attribute + create {MY_TYPE} from_type_1 (attribute_1)

attribute_2 + my_attribute -- Equivalent to: -- attribute_2 + my_attribute to_type_2

Examples of automatic type conversions

Thus, it becomes possible to add complex numbers and integers cleanly and completely transparently to the clients. Part of the automatic type conversion mechanism is already implemented in the ISE Eiffel compiler.

Frozen classes
Another mechanism pre-approved at ECMA is to allow frozen classes in Eiffel, meaning classes from which one cannot inherit. The syntax is simple: the header of a frozen class is extended with the keyword frozen as shown next:
frozen class MY_CLASS

Header of a frozen class

Section 18.3 describes the syntax and semantics of frozen classes in detail.

394

EIFFEL: THE ESSENTIALS A.5

This extension is directly useful to the work presented in this thesis: it enables writing correct singletons in Eiffel, which is not possible with the current version of the language. Frozen classes are not supported by classic Eiffel compilers yet but they are already accepted by the ISE Eiffel for .NET compiler.

Singleton pattern, 18.1, page 289.

A.5 BUSINESS OBJECT NOTATION (BON)

This last section describes the BON method, with a focus on notation. It only introduces a small subset of BON what you need to know to understand the class diagrams appearing in this thesis.
[Waldn-Web].

The method
The Business Object Notation (BON) is a method for analysis and design of objectoriented systems, which emphasizes seamlessness, reversibility and Design by Contract. It was developed between 1989 and 1993 by Jean-Marc Nerson and Kim Waldn to provide the Eiffel programming language and method with a notation for analysis and design. BON stresses simplicity and well-defined semantics. In that respect, it is almost at the opposite of widely-used design notations such as the Unified Modeling Language (UML) or the Rationale Uniform Process (RUP). As mentioned above, one priority of BON is to bring seamlessness into software development, to narrow the gap between analysis, design, and implementation by using the same concepts and the same semantics for the notation on the tree stages. Therefore BON does not have state-charts or entity-relationship diagrams like in UML because they are not compatible with what is available at the implementation level and would prevent reversibility. Instead, BON relies on pure object-oriented concepts like classes, client and inheritance relationships.

Notation
Here is a catalog of the notations used throughout this thesis: In BON, classes are represented as ellipses, sometimes referred to as bubbles:

NAME_OF_ NON_GENERIC_CLASS

NAME_OF_ GENERIC_CLASS [G]

Notation for a class

The ellipse may contain information about the class properties, for example: Deferred class: Class that is declared as deferred.

* CLASS_NAME
Effective class: Class that is not declared as deferred but has at least one deferred parent, or redefines at least one feature.

Notation for a deferred class

+ CLASS_NAME

Notation for an effective class

A.5 BUSINESS OBJECT NOTATION (BON)

Persistent class: Class that inherits from STORABLE or has an indexing tag persistent.

395

CLASS_NAME
Interfaced class: Class that interfaces with other programming languages (for example C) through external features.

Notation for a persistent class

CLASS_NAME
Reused class: Class that comes from a precompiled library. (ISE Eiffel compiler provides the ability to compile a library once and for all and then reuse it in so-called precompiled form.)

Notation for an interfaced class

CLASS_NAME
Root class: Class that is the programs entry point.

Notation for a reused class

CLASS_NAME
It is also possible to specify the features of a class by writing the feature names next to the ellipse representing the class:

Notation for a root class

CLASS_NAME name_of_feature_1 name_of_feature_2

BON also permits to express the different categories of features: Deferred feature: Non-implemented feature.

Notation for features

* CLASS_NAME feature_name*
Effective feature: Feature that was deferred in the parent and is implemented in the current class.

Notation for a deferred feature

+ CLASS_NAME feature_name+
Undefined feature: Feature that was effective in the parent and is made deferred in the current class through the undefinition mechanism.

Notation for an effective feature

CLASS_NAME feature_name-

Notation for an undefined feature

396

EIFFEL: THE ESSENTIALS A.5

Redefined feature: Feature that was effective in the parent and whose signature and/or implementation is changed in the current class through the redefinition mechanism.

CLASS_NAME feature_name++
BON also provides a notation for feature renaming:

Notation for a redefined feature

CLASS_NAME f g

Notation for feature renaming

Classes may be grouped into clusters, which are represented as red stippled-rounded rectangles:

cluster name

CLASS_NAME

Notation for a cluster

Client relationship between two classes is represented as a blue double-arrow:

CLIENT_CLASS_NAME

feature_name

SUPPLIER_CLASS_NAME

Notation for client/supplier relationship

Inheritance relationship is represented as a red single-arrow:

PARENT_CLASS_NAME

Notation for (conforming) inheritance

HEIR_CLASS_NAME

Here is the notation for non-conforming inheritance:

PARENT_CLASS_NAME

Notation for (non-conforming) inheritance

HEIR_CLASS_NAME