[racket-dev] `letrec' and continuations

On May 20, 2011, at 4:42 PM, Sam Tobin-Hochstadt wrote:

> On Fri, May 20, 2011 at 4:39 PM, Matthias Felleisen
> <matthias at ccs.neu.edu> wrote:
>> 
>> -- I think my preferred solution would be to wrap letrec so that continuations grabbed during the setup set up a continuation mark that labels them as 'dangerous'. When you reinvoke them, the existence of the mark tells you that the reference cells should be reinitialized (probably only the ones on the control flow from the continuation point).
>> 
>> -- An alternative could be to stick a lexical identifier into letrec declarations that gets removed from the scope once the letrec is established. It would reappear when you invoke a continuation from the RHS and thus you'd know to reini the ref cells. BUT, this requires a mechanism that is not expressible at the surface of Racket. And it's odd.
> 
> I think the key missing piece here is that Matthew wants to avoid
> having the reference cells *at all*.  If you use `let*', you don't get
> any reference cells.


He wants an optimizer that compiles away ref cells for recursive declaration constructs when possible. 

I think that this is much more easily doable with an internal define compiler that goes around the letrec intermediate steps. 

So I am proposing to leave letrec around for the cases when you need an extremely general recursion construct AND to modify it so that it is safe against call/cc. I would go as far as giving up on ref cell elimination for letrec. 

Yes, I forgot to add that this means we need to eliminate letrec from the code base in many cases.