Le mardi 08 août 2023 à 21:38 +0200, Maxime Devos a écrit :
> As such, this not working on the top-level seems a bug to me -- after 
> all, a module definition is conceptually just a big let:
> 
> <enable extra reader syntax> (if applicable)
> (let ()
>    <magic to make imports work>
>    (define ...)
>    (define-syntax-rule ...) ...
>    ;; use a new macro using syntax-local-binding
>    ;; to extract the syntax transformer (*).
>    <insert stuff in hash tables>)
> 
> (*) not sure if that precise approach actually works in this context



This is very tempting to believe, and I wish it were true, but it's not true.

At least in Guile, the <insert stuff in hash tables> part doesn't happen
at the end of evaluating the module. Each module variable is created and
inserted while evaluating the define form. Otherwise this would give
an error:

(define a 5)
(define b (module-ref (current-module) 'a))
(display b)

The consequences are not innocent.

A variable is associated to a binding which is a "place" in the terminology
of some other languages, i.e., something you can set!. For toplevel variables,
because you can use module-set!, a module variable is used as the place.

And because module variables are bound to names which are just symbols,
duplicate definitions stomp on each other. Consider this:

(define a 5)
(define (b) a)
(define a 6)
(display (b))

This is valid in Guile and prints 6. The second definition for `a` reuses
the same variable as the first one, effectively acting like a set!.
Contrast this with

(let ()
  (define a 5)
  (define (b) a)
  (define a 6)
  (display (b)))

which raises an error due to the duplicate binding.

As https://okmij.org/ftp/Scheme/macros.html#syntax-rule-dark-corner
puts it, "the top level of Scheme is indeed under-specified and treacherous".

Best,

Jean