[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
loss of abstraction
Michael Sperber wrote:
> Just out of curiosity---from this statement, and more like it on
> this list, I'm getting the vague impression that people regard using
> lists to represent compound syntax as something new. (Of course,
> Scheme macro systems have been doing that forever.) Am I
> misunderstanding the vibes here?
Well, I'm most familiar with the portable syntax case expander from
Chez Scheme [http://scheme.com/csug/syntax.html] that SISC uses.
After I was using the portable syntax case expander for some time, I
became aware that compound syntax objects were represented by lists.
However, as far as I knew this was simply an internal representation.
It was never clear to me that using CAR, CDR, and CONS as a procedural
interface with that macro system was at all supported. I might luck
out, and have a macro work when I wrote it, but how was I to know that
it would work tomorrow or in some other implementation of syntax-case?
Of course, perhaps the procedural interface was always supported and
in fact so obvious that the manual didn't need to mention it :-) As
you asked, I'm reporting on my vibes here...
Then I learned more about PLT Scheme's macro system. I noticed that
PLT Scheme did not support CAR, CDR, and CONS with syntax objects
directly. (Though of course I can use SYNTAX->LIST etc.) I also saw
that PLT Scheme had useful source code location tracking capabilities
that greatly enhanced the macro error reporting.
So my *impression* as I saw the progression towards more more powerful
macro systems with better error reporting was that we were going from
a procedural interface being available but not supported to not being
available. A necessary loss in order to support more sophisticated
Michael Sperber wrote:
> No, the problem is the loss of abstraction.
A line of argument occurred to me this morning. Let me try it out...
In Java, java.util.List is an interface. This is a powerful
abstraction, because any methods that operate on lists can be used
directly with any object that can usefully have a list abstraction.
java.util.List is probably not a useful abstraction for a Java parse
tree. However, it would be a useful abstraction for Scheme syntax.
Suppose we consider CAR, CDR, and CONS as an abstract interface to an
underlying representation. Being able to traverse Scheme syntax as if
it were a list is useful abstraction and allows us to use directly use
powerful list procedures such as SRFI-1.
But what about CONS? CAR and CDR could in theory be, for example,
generic procedures that have the ability operate on different types.
But CONS returns a pair.
I don't know if any language has a good solution to the issue of what
to do when you want to construct new objects of your preferred type.
In Java, you might have a procedure that you wished to be able to
return a list of the type you desired. (Think of a Java
implementation of FOLD, for example). You could pass to the procedure
an instance of a factory class to create a new list of your desired
type. However, as a practical matter, I find (in Java) is once I get
to the point of creating factory classes and so on, I'm often creating
more new code in the framework with the factory classes etc. then I'm
able to take out by removing code duplication. The cure turns out to
be worse than the disease.
As an alternative, I may be able to return a particular implementation
of java.util.List, which user code then simply uses as a list without
being concerned with what class is implementing it.
java.util.Collections.singletonList(Object) works in this way, for
By analogy, CAR and CDR might be generic procedures that operate on
any kind of pair, or object oriented procedures that can operate on
pair subtypes, while CONS always returns a plain pair.
As an interface, CAR, CDR, and CONS have some advantages and
disadvantages compared to Java's java.util.List interface. In Java,
for example, you can insert a new item at the beginning of a list.
However, the list procedures available in Scheme through SRFI-1 and
simple tail recursion are richer (and far easier to use) than are the
procedures available in Java's java.util.Collections or in the Apache
I'm not sure that this is just a matter of available libraries. What
I find in programming in Scheme compared to other languages is that
there is tremendous synergy in the recurring pattern of the list
concept: procedure calls as a list of arguments, data as lists,
programs as list, macros that operate on lists.
Java's collections library is a modern, sophisticated specification.
It directly supports powerful abstractions such as lists and sets as
interfaces. And yet, as a practical matter, I find that for much of
the work I do Scheme's synergy around the list concept turns out to be
more powerful (I can write code with more features in less time) then
Java's sophistication in this area.
I look at the SRFI proposal, and I say, "wow, I get to use the
procedural interface of CAR, CDR, and CONS with syntax objects and
thus use libraries such as SRFI-1 directly without conversions? *And*
I get source code location information? Neat!"
When you say the issue is the loss of abstraction, correct me if I'm
wrong, what I imagine is an argument that goes like this: most current
Scheme implementations do not support CAR and CDR as generic
procedures or object oriented methods. In these Scheme
implementations, the only way to get an object that can be operated on
by CAR and CDR is to return a pair. We are thus constraining the
implementation to use lists made up of standard pairs.
I think this is a valid argument.
But let me ask. Suppose you are using a typical Scheme implementation
in which CAR and CDR operate only on plain pairs. You are then
constrained to implement compound syntax objects as lists. You have
lost the ability to use some other abstraction. As a practical
matter, what impact does that have on you? What would you like to be
able to do with a syntax object abstraction that you'd not be able to
do if you've lost that abstraction?