[racket] Splicing `values' in-place

On Thu, Jul 11, 2013 at 09:18:30AM -0600, Matthew Flatt wrote:
> To elaborate on "currently not possible" (because this idea shows up
> from time to time), allowing splicing of results in function-call
> subexpressions would break equivalences that are currently exploited by
> macros and the compiler.
> 
> For example, many macros assume that
> 
>  (rator rand1 rand2 ... randn)
> 
> can be rewritten as
> 
>  (let ([r rator]
>        [a1 rand1]
>        [a2 rand2]
>        ...
>        [an randn])
>    (r a1 a2 ... an))
> 
> That would not be the case if the `rand's can produce multiple values.
> (I assume that you don't want to allow binding multiple values to a
> variable in `let'.) I think that disallowing this kind of
> transformation would make many macros more difficult to implement ---
> and maybe impossible, in some cases.
> 
> The Racket compiler takes advantage of transformations like the one
> above to speed up your code. Although the compiler could still perform
> transformations when the relevant subexpressions are known to be
> single-valued, I think the transformations would apply much less often
> than now.
> 
> Along similar lines, various tools can tell statically you that
> 
>  (cons e1 e2 e3)
> 
> will be an arity error (assuming that `cons' is the usual binding).
> That kind of support would become much weaker, since `e2' might return
> zero values while `e1' and `e3' return a single value.

All this might be possible in typed racket, if the specific type(s) of 
a multiple returned value were known.

> 
> In short, the kind of splicing that you suggest is a significant sense
> more "dynamic" than Racket.

And suitable typing would make it less dynamic again.

-- hendrik