This chapter describes Scheme 48's module system. The module system is unique in the extent to which it supports both static linking and rapid turnaround during program development. The design was influenced by Standard ML modules[7] and by the module system for Scheme Xerox[4]. It has also been shaped by the needs of Scheme 48, which is designed to run both on workstations and on relatively small (less than 1 Mbyte) embedded controllers.
Except where noted, everything described here is implemented in Scheme 48, and exercised by the Scheme 48 implementation and some application programs.
Unlike the Common Lisp package system, the module system described here controls the mapping of names to denotations, not the mapping of strings to symbols.
The module system supports the structured division of a corpus of Scheme software into a set of modules. Each module has its own isolated namespace, with visibility of bindings controlled by module descriptions written in a special configuration language.
A module may be instantiated multiple times, producing several packages, just as a lambda-expression can be instantiated multiple times to produce several different procedures. Since single instantiation is the normal case, we will defer discussion of multiple instantiation until a later section. For now you can think of a package as simply a module's internal environment mapping names to denotations.
A module exports bindings by providing views onto the underlying package. Such a view is called a structure (terminology from Standard ML). One module may provide several different views. A structure is just a subset of the package's bindings. The particular set of names whose bindings are exported is the structure's interface.
A module imports bindings from other modules by either opening or accessing some structures that are built on other packages. When a structure is opened, all of its exported bindings are visible in the client package.
For example:
(define-structure foo (export a c cons)
(open scheme)
(begin (define a 1)
(define (b x) (+ a x))
(define (c y) (* (b a) y))))
(define-structure bar (export d)
(open scheme foo)
(begin (define (d w) (+ a (c w)))))
This configuration defines two structures, foo and bar. foo is a view on a package in which the scheme structure's bindings (including define and +) are visible, together with bindings for a, b, and c. foo's interface is (export a c cons), so of the bindings in its underlying package, foo only exports those three. Similarly, structure bar consists of the binding of d from a package in which both scheme's and foo's bindings are visible. foo's binding of cons is imported from the Scheme structure and then re-exported.
A module's body, the part following begin in the above example, is evaluated in an isolated lexical scope completely specified by the package definition's open and access clauses. In particular, the binding of the syntactic operator define-structure is not visible unless it comes from some opened structure. Similarly, bindings from the scheme structure aren't visible unless they become so by scheme (or an equivalent structure) being opened.
The configuration language consists of top-level defining forms for modules and interfaces. Its syntax is given in figure 1.
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| Figure 1: The configuration language. | ||||||||||||||||||||||||||||||||||||||||||||||||||||||||
A define-structure form introduces a binding of a name to a structure. A structure is a view on an underlying package which is created according to the clauses of the define-structure form. Each structure has an interface that specifies which bindings in the structure's underlying package can be seen via that structure in other packages.
An open clause specifies which structures will be opened up for use inside the new package. At least one structure must be specified or else it will be impossible to write any useful programs inside the package, since define, lambda, cons, etc. will be unavailable. Packages typically include scheme, which exports all bindings appropriate to Revised5 Scheme, in an open clause. For building structures that export structures, there is a defpackage package that exports the operators of the configuration language. Many other structures, such as record and hash table facilities, are also available in the Scheme 48 implementation.
The modify, subset, and prefix forms produce new views on existing structures by renaming or hiding exported names. Subset returns a new structure that exports only the listed names from its
⟨structure⟩ argument. With-prefix returns a new structure that adds
⟨prefix⟩ to each of the names exported by the
⟨structure⟩ argument. For example, if structure s exports a and b, then
(subset s (a))
exports only a and
(with-prefix s p/)
exports a as p/a and b as p/b.
Both subset and with-prefix are simple macros that expand into uses of modify, a more general renaming form. In a modify structure specification the
⟨command⟩s are applied to the names exported by
⟨structure⟩ to produce a new set of names for the
⟨structure⟩'s bindings. Expose makes only the listed names visible. Hide makes all but the listed names visible. Rename makes each
⟨name⟩0 visible as
⟨name⟩1 name and not visible as
⟨name⟩0 , while alias makes each
⟨name⟩0 visible as both
⟨name⟩0 and
⟨name⟩1. Prefix adds
⟨name⟩ to the beginning of each exported name. The modifiers are applied from right to left. Thus
(modify scheme (prefix foo/) (rename (car bus))))
makes car available as foo/bus.
The package's body is specified by begin and/or files clauses. begin and files have the same semantics, except that for begin the text is given directly in the package definition, while for files the text is stored somewhere in the file system. The body consists of a Scheme program, that is, a sequence of definitions and expressions to be evaluated in order. In practice, we always use files in preference to begin; begin exists mainly for expository purposes.
A name's imported binding may be lexically overridden or shadowed by defining the name using a defining form such as define or define-syntax. This will create a new binding without having any effect on the binding in the opened package. For example, one can do (define car 'chevy) without affecting the binding of the name car in the scheme package.
Assignments (using set!) to imported and undefined variables are not allowed. In order to set! a top-level variable, the package body must contain a define form defining that variable. Applied to bindings from the scheme structure, this restriction is compatible with the requirements of the Revised5 Scheme report.
It is an error for two of a package's opened structures to export two different bindings for the same name. However, the current implementation does not check for this situation; a name's binding is always taken from the structure that is listed first within the open clause. This may be fixed in the future.
File names in a files clause can be symbols, strings, or lists (Maclisp-style “namelists”). A “.scm” file type suffix is assumed. Symbols are converted to file names by converting to upper or lower case as appropriate for the host operating system. A namelist is an operating-system-independent way to specify a file obtained from a subdirectory. For example, the namelist (rts record) specifies the file record.scm in the rts subdirectory.
If the define-structure form was itself obtained from a file, then file names in files clauses are interpreted relative to the directory in which the file containing the define-structure form was found. You can't at present put an absolute path name in the files list.
An interface can be thought of as the type of a structure. In its basic form it is just a list of variable names, written (export name ...). However, in place of a name one may write (name type), indicating the type of name's binding. The type field is optional, except that exported macros must be indicated with type :syntax.
Interfaces may be either anonymous, as in the example in the introduction, or they may be given names by a define-interface form, for example
(define-interface foo-interface (export a c cons)) (define-structure foo foo-interface ...)
In principle, interfaces needn't ever be named. If an interface had to be given at the point of a structure's use as well as at the point of its definition, it would be important to name interfaces in order to avoid having to write them out twice, with risk of mismatch should the interface ever change. But they don't.
Still, there are several reasons to use define-interface:
It is important to separate the interface definition from the package definitions when there are multiple distinct structures that have the same interface — that is, multiple implementations of the same abstraction.
It is conceptually cleaner, and often useful for documentation purposes, to separate a module's specification (interface) from its implementation (package).
Our experience is that configurations that are separated into interface definitions and package definitions are easier to read; the long lists of exported bindings just get in the way most of the time.
The compound-interface operator forms an interface that is the union of two or more component interfaces. For example,
(define-interface bar-interface (compound-interface foo-interface (export mumble)))
defines bar-interface to be foo-interface with the name mumble added.
Hygienic macros, as described in [2, 3], are implemented. Structures may export macros; auxiliary names introduced into the expansion are resolved in the environment of the macro's definition.
For example, the scheme structure's delay macro is defined by the rewrite rule
(delay exp) % % —>% (make-promise (lambda () exp)).
The variable make-promise is defined in the scheme structure's underlying package, but is not exported. A use of the delay macro, however, always accesses the correct definition of make-promise. Similarly, the case macro expands into uses of cond, eqv?, and so on. These names are exported by scheme, but their correct bindings will be found even if they are shadowed by definitions in the client package.
There are define-module and define forms for defining modules that are intended to be instantiated multiple times. But these are pretty kludgey — for example, compiled code isn't shared between the instantiations — so we won't describe them yet. If you must know, figure it out from the following grammar.
|
Scheme 48 has a static linker that produces stand-alone heap images from module descriptions. The programmer specifies a particular procedure in a particular structure to be the image's startup procedure (entry point), and the linker traces dependency links as given by open and access clauses to determine the composition of the heap image.
There is not currently any provision for separate compilation; the only input to the static linker is source code. However, it will not be difficult to implement separate compilation. The unit of compilation is one module (not one file). Any opened or accessed structures from which macros are obtained must be processed to the extent of extracting its macro definitions. The compiler knows from the interface of an opened or accessed structure which of its exports are macros. Except for macros, a module may be compiled without any knowledge of the implementation of its opened and accessed structures. However, inter-module optimization may be available as an option.
The main difficulty with separate compilation is resolution of auxiliary bindings introduced into macro expansions. The module compiler must transmit to the loader or linker the search path by which such bindings are to be resolved. In the case of the delay macro's auxiliary make-promise (see example above), the loader or linker needs to know that the desired binding of make-promise is the one apparent in delay's defining package, not in the package being loaded or linked.
During program development it is often desirable to make changes to packages and interfaces. In static languages it may be necessary to recompile and re-link a program in order for such changes to be reflected in a running system. Even in interactive Common Lisp implementations, a change to a package's exports often requires reloading clients that have already mentioned names whose bindings change. Once read resolves a use of a name to a symbol, that resolution is fixed, so a change in the way that a name resolves to a symbol can only be reflected by re-reading all such references.
The Scheme 48 development environment supports rapid turnaround in modular program development by allowing mutations to a program's configuration, and giving a clear semantics to such mutations. The rule is that variable bindings in a running program are always resolved according to current structure and interface bindings, even when these bindings change as a result of edits to the configuration. For example, consider the following:
(define-interface foo-interface (export a c))
(define-structure foo foo-interface
(open scheme)
(begin (define a 1)
(define (b x) (+ a x))
(define (c y) (* (b a) y))))
(define-structure bar (export d)
(open scheme foo)
(begin (define (d w) (+ (b w) a))))
This program has a bug. The variable b, which is free in the definition of d, has no binding in bar's package. Suppose that b was supposed to be exported by foo, but was omitted from foo-interface by mistake. It is not necessary to re-process bar or any of foo's other clients at this point. One need only change foo-interface and inform the development system of that change (using, say, an appropriate Emacs command), and foo's binding of b will be found when procedure d is called.
Similarly, it is also possible to replace a structure; clients of the old structure will be modified so that they see bindings from the new one. Shadowing is also supported in the same way. Suppose that a client package C opens a structure foo that exports a name x, and foo's implementation obtains the binding of x as an import from some other structure bar. Then C will see the binding from bar. If one then alters foo so that it shadows bar's binding of x with a definition of its own, then procedures in C that reference x will automatically see foo's definition instead of the one from bar that they saw earlier.
This semantics might appear to require a large amount of computation on every variable reference: The specified behavior requires scanning the package's list of opened structures, examining their interfaces, on every variable reference, not just at compile time. However, the development environment uses caching with cache invalidation to make variable references fast.
While it is possible to use the Scheme 48 static linker for program development, it is far more convenient to use the development environment, which supports rapid turnaround for program changes. The programmer interacts with the development environment through a command processor. The command processor is like the usual Lisp read-eval-print loop in that it accepts Scheme forms to evaluate. However, all meta-level operations, such as exiting the Scheme system or requests for trace output, are handled by commands, which are lexically distinguished from Scheme forms. This arrangement is borrowed from the Symbolics Lisp Machine system, and is reminiscent of non-Lisp debuggers. Commands are a little easier to type than Scheme forms (no parentheses, so you don't have to shift), but more importantly, making them distinct from Scheme forms ensures that programs' namespaces aren't cluttered with inappropriate bindings. Equivalently, the command set is available for use regardless of what bindings happen to be visible in the current program. This is especially important in conjunction with the module system, which puts strict controls on visibility of bindings.
The Scheme 48 command processor supports the module system with a variety of special commands. For commands that require structure names, these names are resolved in a designated configuration package that is distinct from the current package for evaluating Scheme forms given to the command processor. The command processor interprets Scheme forms in a particular current package, and there are commands that move the command processor between different packages.
Commands are introduced by a comma (,) and end at the end of line. The command processor's prompt consists of the name of the current package followed by a greater-than (>).
,open (subset foo (bar baz))
which only makes the bar and baz bindings from structure foo visible.
,config ,load foo.scm
interprets configuration language forms from the file foo.scm in the current configuration package.
user> ,config
config> (define-structure foo (export a)
(open scheme))
config> ,in foo
foo> (define a 13)
foo> a
13
In this example the command processor starts in a package called user, but the ,config command moves it into the configuration package, which has the name config. The define-structure form binds, in config, the name foo to a structure that exports a. Finally, the command ,in foo moves the command processor into structure foo's underlying package.
A package's body isn't executed (evaluated) until the package is loaded, which is accomplished by the ,load-package command.
,in mumble (cons 1 2) ,in mumble ,trace foo
It is possible to set up multiple configuration packages. The default configuration package opens the following structures:
module-system, which exports define-structure and the other configuration language keywords, as well as standard types and type constructors (:syntax, :value, proc, etc.).
built-in-structures, which exports structures that are built into the initial Scheme 48 image; these include scheme, threads, tables, and records.
more-structures, which exports additional structures that are available in the development environment. A complete listing can be found in the definition of more-structures-interface at the end of the file scheme/packages.scm.
Note that it does not open scheme.
You can define additional configuration packages by making a package that opens module-system and, optionally, built-in-structures, more-structures, or other structures that export structures and interfaces.
For example:
> ,config (define-structure foo (export)
(open module-system
built-in-structures
more-structures))
> ,in foo
foo> (define-structure x (export a b)
(open scheme)
(files x))
foo>
Unfortunately, the above example does not work. The problem is that every environment in which define-structure is used must also have a way to create “syntactic towers”. A new syntactic tower is required whenever a new environment is created for compiling the source code in the package associated with a new structure. The environment's tower is used at compile time for evaluating the macro-source in
(define-syntax name macro-source) (let-syntax ((name macro-source) ...) body)
and so forth. It is a “tower” because that environment, in turn, has to say what environment to use if macro-source itself contains a use of let-syntax.
The simplest way to provide a tower maker is to pass on the one used by an existing configuration package. The special form export-syntactic-tower-maker creates an interface that exports a configuration package's tower. The following example uses export-syntactic-tower-maker and the ,structure command to obtain a tower maker and create a new configuration environment.
> ,config ,structure t (export-syntactic-tower-maker)
> ,config (define-structure foo (export)
(open module-system
t
built-in-structures
more-structures))
Before Scheme 48 1.9, export-syntactic-tower-maker was named export-reflective-tower-maker; this name is still supported for backwards compatibility.
This module system was not designed as the be-all and end-all of Scheme module systems; it was only intended to help us organize the Scheme 48 system. Not only does the module system help avoid name clashes by keeping different subsystems in different namespaces, it has also helped us to tighten up and generalize Scheme 48's internal interfaces. Scheme 48 is unusual among Lisp implementations in admitting many different possible modes of operation. Examples of such multiple modes include the following:
Linking can be either static or dynamic.
The development environment (compiler, debugger, and command processor) can run either in the same address space as the program being developed or in a different address space. The environment and user program may even run on different processors under different operating systems[8].
The virtual machine can be supported by either of two implementations of its implementation language, Prescheme.
The module system has been helpful in organizing these multiple modes. By forcing us to write down interfaces and module dependencies, the module system helps us to keep the system clean, or at least to keep us honest about how clean or not it is.
The need to make structures and interfaces second-class instead of first-class results from the requirements of static program analysis: it must be possible for the compiler and linker to expand macros and resolve variable bindings before the program is executed. Structures could be made first-class (as in FX[10]) if a type system were added to Scheme and the definitions of exported macros were defined in interfaces instead of in module bodies, but even in that case types and interfaces would remain second-class.
The prohibition on assignment to imported bindings makes substitution a valid optimization when a module is compiled as a block. The block compiler first scans the entire module body, noting which variables are assigned. Those that aren't assigned (only defined) may be assumed never assigned, even if they are exported. The optimizer can then perform a very simple-minded analysis to determine automatically that some procedures can and should have their calls compiled in line.
The programming style encouraged by the module system is consistent with the unextended Scheme language. Because module system features do not generally show up within module bodies, an individual module may be understood by someone who is not familiar with the module system. This is a great aid to code presentation and portability. If a few simple conditions are met (no name conflicts between packages, and use of files in preference to begin), then a multi-module program can be loaded into a Scheme implementation that does not support the module system. The Scheme 48 static linker satisfies these conditions, and can therefore run in other Scheme implementations. Scheme 48's bootstrap process, which is based on the static linker, is therefore nonincestuous. This contrasts with most other integrated programming environments, such as Smalltalk-80, where the system can only be built using an existing version of the system itself.
Like ML modules, but unlike Scheme Xerox modules, this module system is compositional. That is, structures are constructed by single syntactic units that compose existing structures with a body of code. In Scheme Xerox, the set of modules that can contribute to an interface is open-ended — any module can contribute bindings to any interface whose name is in scope. The module system implementation is a cross-bar that channels definitions from modules to interfaces. The module system described here has simpler semantics and makes dependencies easier to trace. It also allows for higher-order modules, which Scheme Xerox considers unimportant.