bug#39634: All keyowrds hash to the same value

[lloda <lloda@sarc.name>]

> On 25 Feb 2020, at 21:56, Andy Wingo <wingo@igalia.com> wrote:
> 
> On Thu 20 Feb 2020 17:19, Ludovic Courtès <ludo@gnu.org> writes:
> 
>> Of all the scm_tc7_ values listed in ‘scm.h’, the following are not
>> explicitly listed (so they go to the default case that hashes the first
>> word):
> 
> Reformatting your list so I can check one by one :)
> 
>>  variable,
>>  hashtable,
>>  fluid,
>>  dynamic_state,
>>  frame,
>>  atomic_box,
>>  program,
>>  vm_cont,
>>  weak_set,
>>  weak_table,
>>  port
> 
> No equal? implementation, so should hashq() instead.
> 
>>  bytevector,
>>  array,
>>  bitvector,
> 
> These have equal? implementations, and what's more, a bitvector can
> equal? an array... I think we have another bug!
> 
>  ;; Project 2d array as 1d array (scm_tc7_array)
>  (define x
>    (make-shared-array #2b((#t #t #t)) (lambda (i) (list 0 i)) '(0 2)))
>  ;; scm_tc7_bitvector
>  (define y #*111)
> 
>  (equal? x y) ;; => #t
>  (equal? (hash x #xffffffff) (hash y #xffffffff)) ;; => #f
> 
> Similarly for 1-d scm_tc7_array versus regular vectors, bytevectors,
> etc.
> 
> Fixing this will not be straightforward...  I think basically 1d arrays
> need some special hashing logic so that e.g. vectors and 1d arrays hash
> to the same thing.

I cannot check at the moment but I think that use of make-shared-array is special cased to return a bitvector because the shared array and the root happen to be equivalent. So your x isn't a scm_tc7_array but a scm_tc7_bitvector. The same is true for the other vector types. You can see that if you make a shared array with bounds '(0 1) instead of '(0 2) for example, or non-unit step or non-zero lower bound or anything that cannot be represented as a root vector.

Now it is true that functionally a root vector and a 1d array with the same bounds and the same elements are equivalent even if the array has non-unit stride and so on, but we had that logic before were you could use the root vector functions on arrays and it was an absolute mess. I think there should be a logic to hash n-d arrays that extends to 1-d arrays so there's no need to make special cases. All vectors can be treated as 1-d arrays so that should work fine for those too.

Partially related, I have a series of patches on wip-vector-cleanup to make sure that the various vector implementations don't depend on arrays (as they still do on some cases) but rather the other way around, strictly. I haven't posted about it b/c it changes a few interfaces and I haven't figured out the deprecation route.

regards





> 
>>  stringbuf,
>>  values,
> 
> These are never exposed to Scheme, and never compared using equal?
> AFAIU.  No need for special cases.
> 
> Basically I think the tc7 case should default to hashq, and include
> special cases for the ones that have equal? implementations or which
> have read syntax.
> 
> Sound right to you?
> 
> Andy
> 
> 
>