Access is simplified and abstracted through modules, which are a way of grouping related classes, methods, and variables. Rather than being tied to a single class, each method belongs to a family called a generic function. Each generic function can operate on one or more related classes, and can be extended across one or more modules. We'll talk more about generic functions, polymorphism, and modules later in this article.
Dylan has many other features that distinguish it from C++, including:
On this issue's CD, you'll find a freeware Dylan interpreter, called Marlais, that you can use to execute code written in Dylan. Simply run the application and enter your code at the prompt. Also on the CD are the code samples you'll see in this article, plus the Dylan Interim Reference Manual and other Dylan goodies.
Apple's implementation of Dylan, called Apple Dylan, is planned to ship later this year. One great feature of Apple Dylan is that it allows you to call existing C and C-compatible code libraries, such as the Macintosh Toolbox. See "Creole: Using the Toolbox and Other C Code From Within Dylan Code" for details.
Apple has designed a cross-language extension to the Dylan language. This extension, called Creole in Apple Dylan, allows you to build programs with parts written in both Dylan and C or C-compatible languages. We at Apple hope the extension will be supported by other Dylan implementations, but since the extension isn't part of the standard Dylan language, it's not required. (The Marlais interpreter on this issue's CD doesn't support it.) In the future, Apple will also support the System Object Model (SOM) extension, which is used by OpenDoc. Here we'll take a look at some features of Apple Dylan's Creole implementation.
Once you import C interfaces into Dylan, you can call C functions and refer to C structs as if they were Dylan functions and objects. There's no need to translate the C headers first; Creole reads them directly. In the following simple example, we import the interface file OSUtils.h, which contains the Toolbox function SysBeep; we can then, for instance, call SysBeep(1) from Dylan.
define interface #include "OSUtils.h", import: {"SysBeep"}; end interface;Creole provides these additional facilities:
A define interface statement imports one or mointerface files and creates Dylan classes, constants, and functions corresponding to the C types, constants, and functions in the interface files. Like any Dylan expression, a define interface statement exists in a particular module, as do the variables that it defines. You can export and rename these variables using module options just as you would for normal Dylan variables (as discussed later under "The Role of Modules").
Many options are available to override Creole's default behavior. For example, you can do any of the following:
Listing 1. SimMogul -- basic version
define class <project> (<object>) slot script; // All you need is a hot script slot star; // and a big name. end class <project>; // Last two words are optional. define class <actor> (<object>) slot name; // Actor's name slot salary; // Cost to hire slot fans; // Audience size end class; define class <script> (<object>) slot name; // Script's name slot fx-budget; // Cost of special effects end class; define class <sci-fi> (<script>) end class; define class <romance> (<script>) end class;
Figure 1. Inheritance hierarchy of SimMogul classes
The first thing you might notice about this code is that Dylan identifiers draw on a richer stock of characters than do most languages. Dylan identifiers are case-insensitive and can include characters like <, >, *, +, and -, which are traditionally reserved for operators. As a result, operators like these must be surrounded by spaces when used in formulas (as you'll see later in the definitions for profits in Listing 5).
As shown in Listing 1, each define class statement begins with the name of the class being defined, followed by its parent (or superclass) in parentheses. Dylan supports multiple inheritance; multiple superclasses would be listed in the parentheses, separated by commas. For this short example, however, we'll stick to single inheritance.
<project> is a simple class with two slots (comparable to data members in C++) named script and star. This is a pretty basic structure that doesn't include any options, but it illustrates the syntax for class definitions. There's no need to create constructor or destructor methods; that's taken care of automatically. The last two words, class <project>, are optional, but if you provide them, the name must match that of the class being defined. You can just say end or end class instead if you like, which is what we've done for the remaining classes.
Listing 2 shows a version of SimMogul that does contain some type declarations. The definition of <actor>, for example, has a slot declared as name :: <string>, which specifies the type of the name slot. The compiler will generate code that guarantees that only strings can be stored in this slot; attempts to store anything else will cause an error.
Listing 2. SimMogul -- embellished version
define class <project> (<object>) slot script; // All you need is a hot script slot star; // and a big name. end class <project>; // Last two words are optional. define class <actor> (<object>) slot name :: <string>, // Actor's name required-init-keyword: name:; slot salary :: <number>, // Cost to hire init-value: 1000000, init-keyword: salary:; slot fans :: <number>, // Audience size init-value: 1000000, init-keyword: fans:; end class;
define class <script> (<object>) slot name :: <string>, // Script's name required-init-keyword: name:; slot fx-budget :: <number>, // Cost of special effects init-value: 10000, init-keyword: fx-budget:; end class; define class <sci-fi> (<script>) inherited slot fx-budget, init-value: 20000000; end class; define class <romance> (<script>) inherited slot fx-budget, init-value: 0; end class; define variable arnold = make(<actor>, name: "Arnold", fans: 10000000); define variable betty = make(<actor>, name: "Betty", fans: 5000000); define variable tender :: <script> = make(<romance>, name: "Tender Sunshine"); define variable zarx :: <script> = make(<sci-fi>, name: "Land of the Zarx-Eaters"); define constant $ticket-price = 7;Another reason to provide type declarations is that it allows the compiler to generate more efficient code. For example, if you wrote code that stores an appropriately declared value in an <actor>'s name slot, at compile time the compiler would be able to deduce the value's type. Values that are known to be strings will be stored efficiently, with no runtime type checking. Those known not to be strings will generate compile-time warnings, just as they would in a strongly typed language. If you choose to leave off declarations, the compiler will insert instructions for runtime type checking, so you'll have crash-safe code no matter what. This is an example of how Dylan always lets you compile in a way that's both maximally safe and efficient.
In general, Dylan programs should crash much less often than comparable C programs, because most errors will be detected and handled gracefully and automatically. Automatic memory management is one big source of this safety, since it eliminates the majority of bugs that usually come from manually operating on raw memory pointers. Dylan's ability to ensure safety, however, is limited when working with code imported from outside Dylan, such as the Macintosh Toolbox, which forces Dylan programmers to use raw memory pointers in some cases. Apple Dylan will insulate programmers as much as possible from these pointers with an application framework.
Since the slots in <project> don't have init-keywords, you can't provide values for them when you use make to make projects. If a project is created with make(<project>), the slots are uninitialized, and any attempt to read their values in this uninitialized state is an error that will be detected and reported.
The name slot in <actor> has a required-init-keyword: option, which is used further down to specify the name of the arnold object. Required init-keywords are commonly used for slots with no default value because this requires callers to provide a value when they make objects.
The other slots in <actor>, salary and fans, have default values as well as init-keywords. When an actor is created, the slot's value can be defaulted (for example, arnold's salary) or overridden (for example, arnold's fans). Slots can also be initialized with the init-function: option, which calls a function to compute the default value.
The declaration salary :: <number> restricts the salary slot to hold only numbers. Notice that we didn't choose a specific numeric type for the salary slot type (such as short, int, long, or double), though we easily could have. Dylan provides a rich library of numeric types, including integers of unlimited size (which are good for devalued currencies and salaries of major athletes). By using <number> instead of a more specific numeric type, your type declaration becomes a tool for documentation and error checking, even while you're in the midst of rapid prototyping. We're not obliged to make some arbitrary and premature optimization at this stage, as we would with C or C++. Using <number> captures as much of our design as we want for now; we can always come back and tune it later.
A Dylan class inherits slots from all its superclasses and can also define its own new slots, just as in C++. All slots in a given class must be unique; there cannot be two different slots with the same name. You can override some properties of an inherited slot, however, by partially respecifying the slot. Taking a look at the definition of <sci-fi> in Listing 2, we see that it overrides the default init-value for fx-budget inherited from <script> with a somewhat higher value. The keyword inherited indicates that the slot is inherited from a superclass; it's not a new slot with the coincidentally identical (and therefore illegal) name.
You can specify many other interesting options for slots, such as class allocation, which shares a singly allocated value used by all instances of that class; class allocation roughly corresponds to a static data member in C++. Dylan also lets you provide virtual allocation for slots. Rather than being stored in the slot, a virtual slot's value is computed by a function each time the slot is referenced. This feature is missing from C++ and is very different from what C++ refers to as virtual data members.
Also included is a define constant statement, which looks just like define variable, except that once you give it a value, the running program isn't allowed to change it. The $ in the name $ticket-price is something of a coincidence. By convention, all constants in Dylan are given names beginning with a dollar character, as in define constant $pi = 3.14159.
It's worth noting that define constant doesn't restrict mutable objects from being mutated. Some collections, such as vectors, are mutable in that the value of an element can change, and class instances are mutable in that a slot can change (unless you declare the slot as a constant in the class definition, of course). Since define constant describes the identifier, not the object, what it really means is that the identifier will always refer to that particular object, and to no other object. This is the same as a const pointer in C++, where the pointer is not allowed to change but the object pointed to may be mutated.
$ticket-price is a real constant after all, because its value of 7 (like all numbers) cannot be mutated; for example, you cannot change the 7 to an 8 without changing the object itself.
When we say everything's an object in Dylan, we mean everything. A variable is just a way of naming an object so that you can refer to it in your program. Since you can refer to functions or classes just as easily as you can refer to numbers, we think of them as "variables." So don't be shocked when you see documentation referring to something like print as a variable. It's just a variable whose value happens to be a function.
name(a) // Gets the name of actor a name-setter(new, a) // Sets the name of actor a to new salary(a) // Gets the salary of actor a salary-setter(new, a) // Sets the salary of actor a to new fans(a) // Gets the audience size of actor a fans-setter(new, a) // Sets the audience size of actor a to newSlot access in Dylan looks exactly like a function call, even though the compiler may implement slot access much more efficiently. Alternatively, you can use the more traditional dot notation for slot access. Therefore, the syntax object.property is exactly equivalent to property(object). You can use whichever syntax best fits the situation.
This functional interface is a great feature, because it allows a class's implementation details to remain an abstraction for the users of a class. The fans property, which indicates the box office drawing power, might be stored as a slot in some classes or it might be computed on the fly by a function for other classes. Users will always see a functional interface, and never need to know about the internal implementation.
Whenever a slot reference appears on the left side of an assignment statement, the reference is translated into a call to the appropriate setter function. For example, these are all equivalent ways of changing the name slot of the arnold object:
arnold.name := "Arnie"; name(arnold) := "Arnie"; name-setter("Arnie", arnold);Slots can also take a setter: option, which lets you provide the name of the setter function. The default is to give it a name like name-setter, but you can use a different name, or specify that no setter at all should be created. If there is no setter function, you effectively make the slot's value read-only. As you'll see later in the section on modules, you can also control read and write access to slots by selectively exporting getter and setter functions to other modules.
So, when Dylan sees a call to name(x), depending on what type of object is, one of several methods is selected and executed. In Dylan, name is called a generic function, consisting of a family of name methods that implement the functionality of name for various classes (see Figure 2). Each of these methods "belongs to" its generic function (in this case, name) rather than to a class. This is a key point; it's the core difference between C++'s object model and Dylan's.
Figure 2. Generic function containing several methods
C++'s virtual methods are polymorphic only to the extent that they share a common ancestor. In C++, if you wanted name to work on both actors and scripts, you'd have to create a class (for example, nameableObject) just to contain the name method, and then modify both actor and script classes to inherit from it. This scenario creates quite a few unwanted complications. First, it clutters up your object hierarchies with unnatural "glue" classes that have little to do with the problem domain being represented. Second, it requires you to add inheritance links to bring together classes that otherwise have no reason to be connected, which reduces modularity. Multiple inheritance is extremely awkward in C++ (much less so in Dylan), so you usually want to avoid it wherever possible when using C++.
You also may not have the desire or the ability to change classes near the root of a C++ class hierarchy, either because you don't have access to the affected classes' source code, or because the recompilation time would be very costly. The latter is usually not a problem in Dylan, because most commercial Dylan implementations (including Apple Dylan) provide incremental compilation, which means you can edit, recompile, and relink classes in a matter of seconds.
Listing 3. Method specificity
define method double (x) // No type declaration, default handler pair(x, x); // Works on any type of object end method double; define method double (x :: <number>) // Works on all numbers 2 * x; // Returns 2 times x end method double; define method double (x :: <string>) // Works on all strings concatenate(x, x); // Returns "stringstring" end method double; define method double (x == cup:) // Works on the symbol cup: pint:; // Returns the symbol pint: end method double;The first method in Listing 3 has no specialization at all, so it's equivalent to a default specialization of x :: <object>, which means it will work on anything. It returns a new structure (an instance of the built-in class <pair>), containing two pointers to the argument x.
The default specialization might not be satisfactory for all objects, of course, so the second method specializes the behavior for the case where x is a <number>. In this case, double returns the argument multiplied by 2. For a <string>, the third method returns a new string created by concatenating the argument to itself.
A #rest parameter collects an arbitrary number of arguments as a sequence. For example, the following function takes one required argument, view, and any number of additional arguments. A for loop is used inside the function to iterate over the arguments.
define method polygon (view :: <view>, #rest points) for (p in points) ... end for; end method;Here's an example of using this function:
polygon(myWindow, p1, p2, p3, p4, p5); // Typical usageKeyword parameters specified with #key are quite handy, especially for functions with many parameters, which often take default values. As we saw earlier, make takes keyword parameters in order to create objects. These can be provided in any order by the caller, or omitted entirely if default values are specified. The keywords themselves provide an extra degree of clarity to the calling code, since they serve to document the arguments they introduce. For example:
define method rent-car (customer :: <person>, // Two required parameters location :: <city>, // and up to 4 keywords #key color = white:, // Default color is white sunroof? = #f, // Default no sunroof automatic? = #t, // Default automatic shift days = 3) // Default 3-day rental ... end method;Notice the usage of #t and #f. These are the Dylan values for Boolean true and false, respectively.
Some examples of using this function are as follows:
rent-car(arnold, dallas, days: 7, sunroof?: #t); rent-car(betty, dallas, days: 8, color: #"red"); rent-car(colin, vegas); // Everything defaultedYou also have the option of specifying the return parameters for Dylan functions, as illustrated in Listing 4. This provides more information to the compiler to assist in optimization, as well as documents your code for other users. Dylan functions can return multiple values, which means the caller can get zero, one, or more than one value from the callee. This lets you program in a cleaner, more functional style than in C. In Dylan, you don't need to mix your input and output parameters and bash inputs to make them outputs, or clutter your code with definitions for funny data structures that do nothing more than carry the results of one function back to another.
Listing 4. Example of return declarations and multiple return values
define class <brick> (<object>) slot vert; slot horiz; slot depth; slot density; end class; define method calculate-weight (b :: <brick>) => weight :: <number>; // Declares return parameter let (x, y, z) = bounding-box(b); // Binds multiple values x * y * z * b.density; // Returns one value end method; define method bounding-box (b :: <brick>) => (height :: <number>, width :: <number>, length :: <number>); values(b.vert, b.horiz, b.depth); // Returns three values end method;All methods in a generic function must be congruent. Basically, this means they must all take the same number of required parameters, and they must agree on taking keyword and rest values. There are a few more options you can specify for a generic function using the define generic statement, which can also constrain method congruency.
There are two methods in the profits generic function defined in Listing 5. The second of these methods is more specialized than the first one, because its script parameter (<sci-fi>) is more specific than that in the first (<script>). It just happens that the script parameter is in the second position. When selecting the method to handle a call like profits(betty, tender), Dylan determines that the first method is the only one that's applicable, so that's the one that's used (see Figure 3). It turns out that both methods are applicable in a call like profits(arnold, zarx). The second method is more specific, so that's what gets invoked.
Listing 5.
define method profits (star :: <actor>, script :: <script>) (star.fans * $ticket-price) // Money from ticket sales - (script.fx-budget + star.salary); // minus expenses end method; define method profits (star :: <actor>, script :: <sci-fi>) next-method() / 2; // Sci-fi is out of fashion these days end method;
Figure 3. Method selection based on all arguments (not just the first one)
The body of the second profits method uses a special trick to inherit functionality provided in the base method. It calls next-method, a Dylan function that calls the next appropriate method in the generic function, in decreasing order of specificity. In the example, next-method gets a numeric value calculated by the first method, divides it by 2, and returns that to the caller. As a result, you don't have to write the basic equation twice; new methods have the option of calling up the specificity chain and doing what they want with the results. You can also add code to perform tasks before or after calling next-method.
A module's definition specifies which names are exported. This gives you control over which variables, functions, classes, and slots are private to that module and which are public. For example, suppose the code in Listing 2 lived in a module called the studio module. We could define this module with the statement below, which exports three classes and three functions. Since arnold and betty are not exported, they're private to the studio module, and are inaccessible to any code outside it.
define module studio use dylan; export <project>, <actor>, <script>, name, name-setter, profits; end moduleModules can selectively import some or all of another module's exports. Once imported, these can be used internally, extended, or reexported. We can define a new hollywood module that uses (imports everything exported from) the studio module. Notice that both modules also use the dylan module. Since the dylan module defines all the basic language primitives (like addition), it's a good idea for user-defined modules to always use it.
define module hollywood use dylan; use studio, export: name, profits; export <movie>, <tv-show>, <videogame>, do-oscars; end moduleThis definition assumes that the hollywood module defines three new classes, plus one new function for computing the Oscar winners. It may define others for internal purposes, but those are the only internal classes and functions that it exports. The module also exports two functions imported from the studio module, name and profits. Even though the hollywood module imports the <actor> class from the studio module, there's no way to access the salary slot because salary wasn't exported, and hence cannot be imported into the hollywood module (see Figure 4).
Figure 4. Selectively exporting names from modules to other modules
You can selectively export just the getter but not the setter function for a slot, which has the effect of making the slot read-only to all other modules. This is what hollywood does with name. Code in the hollywood module can change an object's name because name-setter is imported from studio, but clients outside hollywood can only read, but not set, an object's name.
You could go ahead and define a new function in the hollywood module called fans, but it would have nothing to do with the fans slot in <actor>. This new fans function would be totally unrelated, and could have a different number of parameters than the fans function in the studio module. It's like two different cities each having a street called Main Street; the references are not valid across city borders. This is another key advantage of namespaces -- they reduce the pressure to keep names unique at the expense of legibility or clarity.
You can even rename what you import, which is useful to prevent name conflicts, or to emphasize the origin of a name. For example, the following version of the hollywood module imports the <project> class from studio, but renames it. Within this hollywood module, the class is known only as <production>, not <project>. Modules have many more fancy renaming and import/export features, but we'll skip them for now.
define module hollywood use dylan; use studio, export: name, profits, rename: {<project> => <production>}; export <movie>, <tv-show>, <videogame>, do-oscars; end moduleModules let you control the interface to a portion of code by specifying exactly what you want to make public. You can even use several modules to provide high- and low-level interfaces to the same internal code -- a capability not available in C++. For example, a hollywood-tourist module would import, rename, and export a subset of documented high-level calls to one set of users, whereas a separate hollywood-insider module might import, rename, and export more detailed calls to a different audience. This helps keep the implementation and interface nicely separated.
C++ has many notions of scope, including lexical (block scope inside functions), class, file, and name space. Some people even rely on the selective inclusion of header files or verbose name prefixes ("typographic scoping") to prevent name collisions. Dylan's simpler scheme -- just lexical scope and modules -- provides precise control over the importing, exporting, and naming of classes, functions, and variables in a clean and consistent way.
Apple Dylan isn't planned to ship until later this year, but that doesn't mean you can't play with Dylan before then. If you like what you've seen here, or want to see more, check out the goodies on the CD or those available from on-line services (see "Where to Get Dylan Software and Information").
Just like the Macintosh, Dylan was carefully designed from scratch to make your life a lot more fun and productive. Enjoy, and happy hacking!
Other sources of Dylan software and documentation include the following on-line services:
Dylan discussions can be found on the Internet newsgroup called comp.lang.dylan. You can also access Dylan discussions through e-mail. Internet users can ask to be included in discussions by sending a request to info-dylan-request@cambridge.apple.com (AppleLink users can use the address info-dylan-req@cambridge.apple.com@internet#).
If you'd like to become a beta tester of Apple Dylan, please send a message, including your name, address, telephone number, and a brief statement of what you'd like to do with Apple Dylan, to AppleLink DYLAN.
Thanks to our technical reviewers Stoney Ballard, Jeff Barbose, Ken Dickey, Phil Kania, Ken Karakotsios, David Moon, Carl Nelson, and Kent Sandvik.*