bug#32503: 26.1; Byte-compiled functions don't hash consistently

bug#32503: 26.1; Byte-compiled functions don't hash consistently

[Adam Porter]

I noticed that byte-compiled functions don't hash consistently.  Here's
an ECM from Noam Postavsky
<https://github.com/skeeto/emacs-memoize/pull/10#issuecomment-415066682>:

    (let ((obj1 (byte-compile (lambda (x) x)))
          (obj2 (byte-compile (lambda (x) x))))
      (list (equal obj1 obj2)
            (eq obj1 obj2)
            (= (sxhash obj1)
               (sxhash obj2))))
    ;=> (t nil nil)

According to "(elisp) Defining Hash":

    If two objects OBJ1 and OBJ2 are equal, then ‘(sxhash OBJ1)’ and
    ‘(sxhash OBJ2)’ are the same integer.

This makes it impossible to use byte-compiled functions as keys or parts
of keys in hash tables.  A workaround is to `prin1-to-string` the
byte-compiled function and use that as the key.

Thanks.

bug#32503: 26.1; Byte-compiled functions don't hash consistently

[Tom Tromey]

Adam> I noticed that byte-compiled functions don't hash consistently.

The Lisp_Vectorlike case in Fequal has a number of special cases which
don't seem to be reflected in sxhash:

	if (OVERLAYP (o1))
[...]
	if (MARKERP (o1))
[...]
	if (WINDOW_CONFIGURATIONP (o1))
[...]
	/* Aside from them, only true vectors, char-tables, compiled
	   functions, and fonts (font-spec, font-entity, font-object)
	   are sensible to compare, so eliminate the others now.  */

(Bignums and bool vectors are special-cased as well but are handled in
sxhash.)  This is followed by code to loop over the pseudovector's
elements.

However, the Lisp_Vectorlike case in sxhash doesn't do any of this.

Probably there should be a rule that adding a new type should
necessitate adding a new hasher (much like how print syntax is needed);
or certainly if the type has non-trivial equal behavior.

Tom

bug#32503: 26.1; Byte-compiled functions don't hash consistently

[Lars Ingebrigtsen]

Eli Zaretskii <eliz@gnu.org> writes:

>> Sorry, haven't had a chance to look at it yet. I've been treating it as
>> low-ish priority because pdumping outside loadup isn't supported yet. Is
>> there some reason to expedite this work?
>
> Not that I can see, no.  I just wanted to be sure this isn't
> forgotten.
>
> Thanks.

This was a long thread, and I only skimmed it lightly.  But I noticed
that it was merged with this bug report:

Adam Porter <adam@alphapapa.net> writes:

> I noticed that byte-compiled functions don't hash consistently.  Here's
> an ECM from Noam Postavsky
> <https://github.com/skeeto/emacs-memoize/pull/10#issuecomment-415066682>:
>
>     (let ((obj1 (byte-compile (lambda (x) x)))
>           (obj2 (byte-compile (lambda (x) x))))
>       (list (equal obj1 obj2)
>             (eq obj1 obj2)
>             (= (sxhash obj1)
>                (sxhash obj2))))
>     ;=> (t nil nil)

And this test case no longer fails in Emacs 28 (but it fails in Emacs
27).  So is there more to be done in these merged bug reports?

-- 
(domestic pets only, the antidote for overdose, milk.)
   bloggy blog: http://lars.ingebrigtsen.no