The source for this post is online at 2013-02-04-generativity.rkt.
Racket comes with built-in records called structures, naturally. In this post, I discuss how you might have implemented them yourself and how that implementation might be different than how Racket’s work.
Let’s start with a simple example of a structure:
(jstruct posn posn? (posn-x posn-y)) (define p (posn 1 2)) (check-true (posn? p)) (check-equal? 1 (posn-x p)) (check-equal? 2 (posn-y p))
We define a posn with two fields, x and y, then construct one, test its posn-ness and the extract its fields.
One way to implement this is with lists of a fixed length:
(define-syntax-rule (jstruct struct struct? (struct-f ...)) (begin (define (struct struct-f ...) (list struct-f ...)) (define (struct? v) (and (list? v) (= (length v) (length '(struct-f ...))))) (define (struct-f l) (for/or ([v (in-list l)] [f (in-list '(struct-f ...))]) (and (eq? 'struct-f f) v))) ...))
This is pretty clearly a bad implementation for a few reasons. One problem is that the struct accessors are O(n) in the number of fields, but that’s just to simplify the macro. We can fix that with a hash-table instead. (Although the best thing would be to use a vector where we compute the offsets in the macro, but I don’t want to get into the macro machinery to do that.)
The other problem, that the hash implementation shares is that any hash table (list) that happens to have the right structure counts as a posn.
We can fix that by having a non-forgeable value:
(define jstruct-key (gensym 'dont-tase-me-bro)) (define-syntax-rule (jstruct struct struct? (struct-f ...)) (begin (define (struct struct-f ...) (make-immutable-hasheq (list (cons jstruct-key jstruct-key) (cons 'struct-f struct-f) ...))) (define (struct? v) (and (hash-eq? v) (immutable? v) (eq? (hash-ref v jstruct-key #f) jstruct-key) (hash-has-key? v 'struct-f) ...)) (define (struct-f v) (hash-ref v 'struct-f)) ...))
(We will assume that the jstruct-key identifier is not exported from the code implementing the structure macro.)
Although, this value isn’t really non-forgeable because we could extract the value from the hashes that get exposed by the instances:
(define a-posn (posn 1 2)) (define the-key (for/or ([v (in-hash-values a-posn)]) (and (not (number? v)) v))) (define not-a-posn-2 (hasheq the-key the-key 'posn-x #f 'posn-y #t)) (check-true (posn? not-a-posn-2))
The problem is that we don’t have a way to restrict access to the data-structure we use to implement the structures. Many people assume that the feature structures provide is the ability to efficiently represent records and extract their values, etc. This is not the case. Structures provide a way to implement sealing, or tying core data-structures to their creators and only their creators.
Supposing that we had sealing, then we could implement structures as follows:
(define-syntax-rule (jstruct struct struct? (struct-f ...)) (begin (define-values (seal seal? unseal) (make-seal)) (define (struct struct-f ...) (seal (make-immutable-hasheq (list (cons 'struct-f struct-f) ...)))) (define (struct? v) (seal? v)) (define (struct-f v) (hash-ref (unseal v) 'struct-f)) ...))
But how could we implement seals? Is it possible to do it inside of the language or do we have rely on extending the Racket VM?
It turns out that if your language provides first-class functions, then you already have one data-structure that is automatically sealed. (Closures are data-structures because they hold their environment.) We can use this by implementing sealed values as functions waiting for a dynamically-generated key to unlock them:
With seals in place, our previous hacks do not work:
(define not-a-posn (hasheq 'posn-x #f 'posn-y #t)) (check-false (posn? not-a-posn)) (define a-posn (posn 1 2)) (check-false (hash? a-posn))
This is roughly how Racket’s sealing (called make-struct-type) works except (1) it is much more efficient, (2) it is baked into structures, and (3) it cooperates with the inspector system to allow some "master" code to view the private values of its "slaves".
This implementation is so similar that it has a behavior that many Racket users find a little confusing. Consider this program:
(module server racket/base (require rackunit) <seals> <hash-struct3> <struct-example> (provide posn posn? posn-x posn-y)) (module a racket/base (require 'server) (define p (posn 1 2)) (provide p)) (module b racket/base (require 'server) (define f posn-x) (provide f)) (module connector racket/base (require racket/sandbox) (define a-eval (make-evaluator 'racket/base '(require a))) (define b-eval (make-evaluator 'racket/base '(require b))) (check-equal? ((b-eval 'f) (a-eval 'p)) 1))
It may seem like this test should pass because both a and b use the same server module. Unfortunately, because they are evaluated separately, the server module runs twice, generating a different seal each time. This is called generativity.
The same thing happens in Racket and often shows up when people use sandboxes or code that uses sandboxes. For example, DrRacket keeps user programs separate from itself, so most structures do not inter-operate. The Web server runs each servlet in a separate place, so they are all separate from each other and the server. In the case of the Web server, #:servlet-namespace argument to functions like serve/servlet exists to explicitly allow servlets to share some (but not all) modules.
By the way, if you use this code at home, make sure you put the code in this order: