[racket-dev] PLAI - Manual management of closures

From: Robby Findler (robby at eecs.northwestern.edu)
Date: Wed Jan 4 22:05:48 EST 2012

I'm glad to see you doing this!

FWIW, I would have gone with 1. I would have explained it to the
students by saying that an earlier stage of the compiler must do this
and now primitives are a different kind of things than other

But not really have tried it, I agree 3 is the next best thing.


On Wed, Jan 4, 2012 at 8:55 PM, Jay McCarthy <jay.mccarthy at gmail.com> wrote:
> I've just pushed a new version of the GC language for PLAI that
> requires students to manually manage their closures, rather than
> treating them as "flat" values.
> The code is tested (the entire normal GC test suite is ported) but I
> have not yet pushed the documentation (to discourage early use and
> confusion from students before we've figured out the best way to
> present the issues.)
> *** Why ***
> In the first language, closures are treated as "flat" values, except
> that they aren't really flat so students had to use the
> 'procedure-roots' function to discover any locations that the closure
> held---i.e. its free identifiers. I have always found this really
> cheap and I find that students are very confused by when they need to
> use this function because it comes out of left field.
> In this language, there are a few new functions the student must
> implement:
> gc:closure : code-ptr (vectorof location) -> location
> gc:closure-code-ptr : location -> code-ptr
> gc:closure-env-ref : location integer -> location
> gc:closure? : location -> boolean
> *** Implications ***
> Closures represent a totally different kind of data structure for
> students. Flats are fixed size with no pointers inside. Conses are
> fixed sized with pointers inside. Closures are variable size with
> pointers inside. A representation on the heap is obvious to me:
> 'closure code-ptr n free-var_0 ... free-var_n
> but I have no idea how this will affect how students perform on the
> assignment.
> One benefit is that lots of lame M&S allocation routines simply won't
> work and students will have to have something more robust.
> *** Issues ***
> There is a bit of trouble in the implementation with regards to
> primitive functions like add1, sub1. Here are the approaches I thought
> to try:
> 1. Syntactically restrict them to first-order uses.
>   (map add1 l) must become (map (lambda (x) (add1 x)) l)
>   Pros: No influence on collector/mutator
>   Cons: Hacky?, Weird language.
> 2. Allocate all of them with 'native' code pointers and no free-vars
>   at startup
>   Pros: Not hacky, each primitive only counts once
>   Cons: Every mutator has a very large minimum heap size
> 3. Allocate them at each use with 'native' code pointers ...
>   Pros: A little hacky, Small minimum heap sizes, Pay for what you need
>   Cons: Multiple uses count multiple times, Hard to predict minimum
>   heap sizes
> 4. Allocate each a single time but only if they are used
>   Pros: Same as 3
>   Cons: None of 3, but the implementation is not obvious
> Currently I have gone with 3 after doing 1 and finding it
> distasteful. Are there options I haven't thought of? What should I
> choose? Anyone know how to do 4?
> Jay
> --
> Jay McCarthy <jay.mccarthy at gmail.com>
> Assistant Professor / Brigham Young University
> http://faculty.cs.byu.edu/~jay
> "The glory of God is Intelligence" - D&C 93

Posted on the dev mailing list.