Re: Alternative formulations of keywords Marc Feeley 12 Apr 2006 13:05 UTC

On 12-Apr-06, at 8:16 AM, John Cowan wrote:

> Marc Feeley scripsit:
>
>> If you try to formalize the cases in which it does work and the cases
>> in which it doesn't you will realize that it is very hard to specify
>> precisely.  You have to assume a particular set of powerful analyzes
>> that are performed by the compiler, and your semantics will depend on
>> the existence of these analyzes.  This places difficult constraints
>> on the Scheme implementation.
>
> It's syntactic sugar, so I'm fine with it not working in any case
> that's
> at all tricky.

Please define precisely what is a "tricky case".  If you do this
exercice you will see that it is very hard to formalize a compile
time only mechanism for named optional parameters that is useful.

>
>> For example can your proposed approach work in this case:
>>
>>    (define (f g)
>>      (g foo: 11 bar: 22))
>
> Plainly no.  The compiler should cough with "keywords in call of
> unknown
> function".
>
>> or in this case:
>>
>>    (define (f #!key (x 11) (y 22)) (+ x y))
>>    (define (g z) (f y: z))
>>    (define (h) (set! f (lambda (#!key (y 33) (z 44)) (* y z))))
>
> Again, plainly no.  I might have said "only if the compiler can prove
> that h is always called before g is", but that's precisely the sort
> of tricky analysis neither of us would want to depend on.
>
> Here the error is "keyword y: not known for function f."
>
> In short, the definition of a function must be either global or
> lexically apparent for it to be callable with keywords.

Most Scheme systems allow separate compilation (think of "load").  If
one file contains:

(define (f #!key (x 11) (y 22)) (+ x y))
(define (g z) (f y: z))

and the other contains:

(set! f (lambda (#!key (y 33) (z 44)) (* y z)))

You have the same problem.  So it doesn't suffice for the function
definition to be global.

The way I see it you are forbidding the use of higher-order functions
in combination with named optional parameters.  That would be a
serious limitation for a functional language like Scheme.

Marc