`msgpack` failure

[bennn]

cc @samth

Description

The msgpack package fails to build with the current TR release candidate. The source compiles, but test programs written in #lang racket do not. For example:

#lang racket/base
(require msgpack)
unpack

Error message:

foo.rkt:3:0: Type Checker: could not convert type to a contract;
 required a flat contract but generated a chaperone contract
  identifier: unpack
  type: (-> Input-Port Packable)
  in: unpack

The type for unpack is:

(: unpack (-> Input-Port
              (U Void Boolean Integer Real String Bytes (Vectorof Packable) (Listof Packable) (HashTable Packable Packable) Ext)))

where Packable is:

(define-type Packable
  (U Void
     Boolean
     Integer
     Real
     String
     Bytes
     (Vectorof Packable)
     (Listof Packable)
     (HashTable Packable Packable)
     Ext))

Minimal Example

The TR error message seems to be due to the HashTable type in Packable; here's a tiny program the raises a similar exception:

#lang racket/base

(module t typed/racket/base
  (provide f)

  (define-type T
    (HashTable (Vectorof Integer) Symbol))

  (: f (-> Symbol T))
  (define (f x)
    (hash '#(1) x)))

(require 't)

f

Error message (on Racket v6.10.1 or earlier):

bar.rkt:15:0: Type Checker: could not convert type to a contract;
 required a flat contract but generated a chaperone contract
  identifier: f
  type: (-> Symbol T)
  in: f

The program runs if the type for f is (-> Symbol (U Integer T)) because that type gets converted to a contract like (-> Symbol any/c)

Diagnosis

Since (hash (vector 1) 'a) is valid, the problem is likely the constraint in static-contracts/combinators/structural.rkt that hashtable keys must be flat:

  ((hash/sc (#:invariant #:flat) (#:invariant)) hash/c #:chaperone)

[pnwamk]

Was the motivation for requiring flat key contracts perhaps to not change the behavior of eq? based hash tables?

[samth]

I think we can fix that by distinguishing equal? based tables. Will that fix msgpack?

[bennn]

After removing #:flat, msgpack builds and all raco test -c msgpack tests pass

[pnwamk]

@samth what do you mean by distinguish in this case? As in at the type level? (I assume not since that would be non-trivial and your talking about fixing for release) Or something else?

[samth]

Ok great. We need to improve the contract generated for hash/sc if the keys aren't flat so that it doesn't have a weird contract-system error when given an eq? or eqv? based hash, and then we should be good to go.

[bennn]

@samth I think I understand that, will send a PR soon

[pnwamk]

Maybe I'll understand once the PR is ready XD

[samth]

See the docs for hash/c

[pnwamk]

I've read them -- I'll just wait for the PR and ask if needed their 👍

[bennn]

closed by #626