guile/devel/modules/design-notes.texi

@c devel/modules/desgin-notes.texi

@c TODO
@c - distill wishlist, index
@c - in Findings, characterize current module system wrt wishlist

@node Module System Design Notes
@chapter Module System Design Notes

This chapter documents module system design history.  At the moment
(guile-1.5.4, 2002-02-08), the module system is supposedly undergoing
redesign; the provisional implementation works but has problems, notably
making compilation difficult.

Because module systems design is (was?) an area of active research and
development in the Scheme community, many different features are possible.
This section is a historical record of the selection process used by the Guile
hackers (if one can be discerned).

@menu
* Wishlist::
* Findings::
* Selection Criteria::
* Rationale Statements::
* Specification::
@end menu

@node Wishlist
@subsection Wishlist

In the guile-related mailing lists of yore, discussion resulted in the
following desirable traits.  Note that some of these contradict each other.

@itemize @bullet

@item
support separate compilation

@item
hierarchical module names

@item
support relative references within the module name space (so that a
module within a package can use a sibling module without knowing the
prefix of the module name)

@item
support re-use of code (the same implementation can be presented to
the world through several interfaces)

@item
support individual and group renaming of bindings when using other
modules

@item
easy to import and re-export entire interfaces (so that a main
interface in a package can function as a "relay" and publish
bindings from many modules in the package)

@item
support both variable and syntactic bindings (these should be
clearly separated, though) and mesh well with hygienic macro
systems

@item
hygienic implies that we shouldn't need to care to export bindings
which an exported macro introduces at the point of its use

@item
autoloading

@item
unmemoization protocol

@item
cleanliness

A module should be able to be totally clean.  There should be no
need to have *any* extra bindings in a module (a la
%module-interface or `define-module').

Therefore, we should have at least one dedicated "command" or
"config" or "repl" module.

It would probably be a good idea to follow other good Scheme
interpreters' lead and introduce the ,<command> syntax for walking
around modules, inspecting things, entering the debugger, etc.
Such commands can be looked up in this repl module.

If we insist on not using ,<command> syntax, we are forced to let
the module use list consist of a "sticky" part and the rest, where
the "sticky" part is only available at the repl prompt and not to
the code within the module, and which follows us when we walk around
in the system.

@item
well integrated with the translator framework

We should be able to say that a module uses a different syntax or
language.

Note here the nice config language of the Scheme48 module system
where it is possible to separate code from module specification: the
module specification can use scheme syntax and rest in one file,
while the module itself can use the syntax of another language.

This config language also supports re-use of code in a neat way.

@item
examine connection with object system: how easy is it to support
Java and other class-centered object systems?

@item
easy to export the same module under several different names

@item
easily supports both compiled and interpreted modules

@item
compiled modules can by dynamically loaded or statically linked in
(libltdl might provide this automatically)

@item
convenient syntax for referencing bindings in modules that are
loaded but not used

(Assuming this distinction exists.)  But maybe group renaming is a better
solution to a similar class of problems.

@item
ability to unuse a module (and have it get collected when there are
no more references)

@item
orthoganality between source files, directories and modules. i.e.  ability to
have multiple modules in one source file and/or multiple source files in one
module

@item
backward compatibility

@item
whenever possible the module's meta-information should be stored
within the module itself (only possible for scheme files)

@item
the compiler stores the meta-information into a single file and updates it
accordingly

(FIXME: per module, per package, directory?, per project?)  This
meta-information should still be human readable (Sun's EJB use XML for their
deployment descriptors).

@item
use the file system as module repository

Since guile is a GNU project we can expect a GNU (or Unix) environment.  That
means that we should use the file system as the module repository.  (This
would help us avoid modules pointing to files which are pointing to other
files telling the user "hey, that's me (the module) over there".)

@item
every module has exactly @emph{one} owner who is responsible for the
module @emph{and} its interface (this contradicts with the "more than one
interface" concept)

@item
support module collections

Provide "packages" with a package scope for people working together on a
project.  In some sense a module is a view on the underlying package.

@item
ability to request (i.e. import or access) complete packages

@item
support module "generations" (or "versions")

Whenever a new module fails to handle a request (due to an error) it will be
replaced by the old version.

@item
help the user to handle exceptions (note: exceptions
are not errors, see above)

@item
no special configuration language (like @code{,in} etc.)

You can always press Control-D to terminate the module's repl and return to
the config module.

@item
both C and Scheme level interfaces

@item
programming interface to module system primitives

One should be able to make customized module systems from the low-level
interface, as an alternative to using the default module system.  The
default module system should use the same low-level procedures.

@item
Here are some features Keisuke Nishida desires to support his VM/compiler
[snarfed directly from post <m37l33z0cl.wl@kei.cwru.edu> dated 2001-02-06,
and requires formatting]:

 * There is no "current module".

 * Module variables are globally identified by an identifier
   like "system::core::car".

 * Bindings are solved syntactically, either by a translator
   or a compiler.  If you write just "car", it is expanded to
   "system::core::car" by a translator or compiler, depending
   on what modules you use.

 * An interpreter (repl) may memorize the "current working module".
   It tells the translator or the compiler what modules should be
   used to identify a variable.  So, during interactive sessions,
   a user may feel as if there *is* the current module.

 * But the fact is, all variables are globally identified at
   syntax level.  Therefore, the compiler can identify all
   variables at compile time.  This means the following code
   is not allowed:

     ;; I want to access `foo' in the module named some-module-name
     (let ((module (lookup-module some-module-name)))
       (set! (current-module) module)
       (foo x))
           -> ERROR: Unbound variable: current-module

     (let ((module (lookup-module some-module-name)))
       (module::foo x))
           -> ERROR: There is no variable named "module::foo"

   Instead, you should write as follows if you need a dynamic access:

     (let ((module (lookup-module some-module-name)))
       ((module-ref module 'foo) x))

     (let ((module (lookup-module some-module-name)))
       ((module 'foo) x))      ;; if module is an applicable smob

@end itemize

@c $Date: 2002-02-08 10:50:36 $
@node Findings
@subsection Findings

This section briefly describes module system truths that are one step more
detailed than "module systems are hairy".  These truths are not self-evident;
we rely on research, active experimentation and lots of discussion.  The goal
of this section is to save ourselves from rehashing that which was hashed
previously.

@itemize @bullet

@item Kent Dybvig's module system

A paper is available at
@uref{http://www.cs.indiana.edu/~dyb/papers/popl99.ps.gz,
      http://www.cs.indiana.edu/~dyb/papers/popl99.ps.gz}.

This was discussed in 2000-11 and 2000-12.

@item Distinction between Top-Level Environment and Module

These two are different beasts!  Each of the following needs to be
well-defined with respect to both of these concepts: @code{eval},
@code{define}, @code{define-public}, @code{define-module}, @code{load},
working from REPL, top-level @code{begin}, [add here].

In guile-1.4, the distinction was not clear.

@item Current module system internals

@xref{Top,Module Internals,,module-snippets}, for implemetation
details of the module system up to and including guile-1.6.x.

@item [add here]

@end itemize

@node Selection Criteria
@subsection Selection Criteria

@node Rationale Statements
@subsection Rationale Statements

@node Specification
@subsection Specification


@c devel/modules/desgin-notes.texi ends here