[racket] Delimited continuations and parameters

On 05/14/2012 04:26 PM, Asumu Takikawa wrote:
> Hi all,
>
> Here is a code snippet that uses both delimited continuations and
> parameters (translated from the paper "Delimited Dynamic Binding"[1]):
>
>    #lang racket
>
>    (require racket/control)
>
>    (define p (make-parameter 0))
>    (define r (make-parameter 0))
>
>    ((λ (f)
>       (parameterize ([p 2])
>         (parameterize ([r 20])
>           (f 0))))
>     (parameterize ([p 1])
>       (reset
>         (parameterize ([r 10])
>           ((λ (x) (+ (p) (r)))
>            (shift f f))))))
>
> If you run this, it produces 11. The authors of [1] argue that it should
> return 12, namely because it follows the principle that "the dynamic
> bindings in scope are those in the context".
>
> That makes some sense, considering that when you eventually get to the
> call (f 0) in the first lambda, your context looks like
>
>    (parameterize ([p 2])
>      (parameterize ([r 20])
>        (f 0)))
>
>    where f = (λ (y)
>                (parameterize ([r 10])
>                  ((λ (x) (+ (p) (r)))
>                   y)))
>
> according to shift/reset semantics. From this context, p = 2, r = 10
> could make sense.
>
> That said, I don't really have an intuition for what semantics for
> dynamic binding&  delimited control is useful. Is there a pragmatic
> reason for why Racket's parameters and delimited control interact in
> this fashion?

Racket contexts don't store individual parameter values, they store 
entire *parameterizations* (which effectively contain the values of all 
parameters). So f captures the parameterization {p=1, r=10}. When f is 
called, it restores the entire parameterization, overriding the existing 
mapping {p=2, r=20} at the call site.

Jay's "mark parameter" library (at unstable/markparam) provides 
semantics that are probably closer to what you expect. (Except, IIRC, 
not inherited across threads as parameters are.)

Ryan