The source for this post is online at 2013-12-02-sharish.rkt.
Racket has the ability to construct immutable cyclic data, such as an infinite list of 1s:
In this post, I discuss how this feature could be implemented.
In our implementation of shared, called sharish, we will attempt to use the same concepts as Racket, namely placeholders. The idea is that we first create a placeholder for each identifier, then we fill it in (by calling placeholder-extract! with a function that sets the identifier), then we close all the cycles in one go.
This allows cycles because the names are bound before any of the xes are evaluated and they can mention the names wherever they need to.
The one caveat with our implementation is that we need to use mutation. shared cheats by calling make-reader-graph which is implemented in the virtual machine, so it can mutate data-structures that aren’t normally mutable at the Racket level.
Here’s an extensive test suite:
(struct posn (x y) #:mutable #:transparent) (sharish ([a (mcons 1 x)] [b (mcons #t x)] [c (mcons a x)] [d (vector a b c d)] [e (box d)] [f (posn d e)] [x (mlist a b c d e f x)]) ( #:t (eq? 1 (mcar a)) #:t (eq? x (mcdr a)) #:t (eq? #t (mcar b)) #:t (eq? x (mcdr b)) #:t (eq? a (mcar c)) #:t (eq? x (mcdr c)) #:t (eq? a (vector-ref d 0)) #:t (eq? b (vector-ref d 1)) #:t (eq? c (vector-ref d 2)) #:t (eq? d (vector-ref d 3)) #:t (eq? d (unbox e)) #:t (eq? d (posn-x f)) #:t (eq? e (posn-y f)) #:t (eq? a (mlist-ref x 0)) #:t (eq? b (mlist-ref x 1)) #:t (eq? c (mlist-ref x 2)) #:t (eq? d (mlist-ref x 3)) #:t (eq? e (mlist-ref x 4)) #:t (eq? f (mlist-ref x 5)) #:t (eq? x (mlist-ref x 6))))
The implementation revolves around the placeholder data-structure. It contains a flag to note if the extraction is complete, a queue of data-structures to fill in with its value (more on this below), and its value.
All the real work happens in placeholder-extract!. We’ll define it recursively by cases on the first position:
(define (placeholder-extract! v fill!) (define loop placeholder-extract!) (match v <value-cases> <container-cases> <ph-case> <struct-case>))
All the work happens on the placeholder case. The idea is that if the placeholder is already extracted, then just look up the value. Otherwise, we have to detect if we are a cyclic reference or the initial reference. We can do this by seeing if there are any "fillers". It is important to detect this, because otherwise we will go into an infinite loop. After the value is available, we fill in all the cyclic references and then record that we’ve already walked it.
[(? placeholder?) (cond [(placeholder-extracted? v) (fill! (placeholder-value v))] [else (define first-time? (null? (placeholder-fillers v))) (set-placeholder-fillers! v (cons fill! (placeholder-fillers v))) (when first-time? (loop (placeholder-value v) (λ (x) (for ([fill! (in-list (placeholder-fillers v))]) (fill! x)) (set-placeholder-extracted?! v #t))))])]
This makes all the other cases very simple. If we see an atomic value, we just send it to the filler:
If we see a container, then we recur on each piece with a filler appropriate to the piece:
Finally, if we reach a structure, we see if it has any reflective information available. If it does, then it’s just a variety of container:
[(app (λ (x) (call-with-values (λ () (struct-info x)) list)) (list (? struct-type? st) #f)) (define-values (name init-field-cnt auto-field-cnt accessor-proc mutator-proc immutable-k-list super-type skipped?) (struct-type-info st)) (for ([i (in-range init-field-cnt)]) (loop (accessor-proc v i) (λ (x) (mutator-proc v i x)))) (fill! v)]
The main difference between this code and the real implementation is that the real implementation can either (a) modify things that aren’t really mutable or (b) transform mutable versions of data-structures to the immutable versions after they’ve been filled in.
This code also shows you cycles have to occur in "first-order" positions. For instance, you couldn’t call append and expect there to be sharing in the lists, because append is a function that operates on lists and not a constructor.
The great thing about this code is that it’s under 100 lines... including the test suite.
But first let’s remember what we learned today!
Cyclic data in a strict language requires mutation somewhere.
Cyclic data is dangerous because you might go into an infinite loop looking at it.
If you’d like to run this exact code at home, you should put it in this order:
Or just download the raw version.