Re: new `obarray` type

[Stefan Monnier]

> It's not more palatable to this user.  It sounds more like "dumbing
> down".  There are few things more frustrating whilst debugging than
> having Emacs obfuscating information "for my own good".

The good thing about having the `obarray` type is that e get to choose
how to print them.

I can easily make it so it prints all the symbols in it, would you
prefer that?

> Indeed, why not just print _all_ vectors by printing only their size?

The reason why the patch I sent only prints the number of entries is
that currently the way obarrays are printed you only get to see some of
the symbols but not all.  But I guess you're right: it would be more
useful to list all the symbols in it.
[ Time passes... Done!  ]

> Not rarely, particularly in CC Mode, I will be dealing with obarrays
> with relatively small numbers of symbols.

Regardless of what we decide to do with obarrays, I strongly recommend
you change cc-mode to use hash-tables instead.  My experience with EIEIO
(where I "recently" moved from obarrays to hash-tables) is that it's
measurably faster and the code tends to be clearer (tho that's clearly
subjective).

> Of course I want to see these symbols' names when I ask for that
> obarray to be printed.

With hash-tables, you'll them see all, properly printed and even `read`able!

> I'm also not in favour of introducing another vector-like type without a
> very good reason.

Obarrays are very weird, currently, because they combine "plain vectors"
and "plain symbols" in a tricky way.
- Have you ever tried to do `aref` on an obarray?
- The printout lists some of the symbols, but not all.  Which ones appear
  is arbitrary, unpredictable.
- Have you ever tried to put something else than 0 in an obarray slot?
- An obarray can lead to unexpected space behavior:

    (let* ((o (obarray-make))
           (s1 (intern "s1" o))
           (s2 (intern "s2" o)))
      (set s1 (make-list 100000 t))
      s2)

  might leave you with a 10000-element list preserved as "reachable" as
  long as `s2` is reachable, even tho it's clearly not reachable any more.

BTW, my patch doesn't address this GC problem yet.

> positions.  Would a new obarray type prevent any vector operations being
> carried out on it, should any package do such things?  If so, that would
> be a Bad Thing.

Currently, I haven't changed `aref` to work on obarrays, no.  I've never
ever seen code try to do that (I guess in theory it could be potentially
useful, tho I can't think of any operation you could implement reliably
using `aref` on obarrays would be `obarray-empty-p`).

I was planning on allowing `aset` in case some package uses it to do the
equivalent of `clrhash`, but I haven't yet found any package doing that,
so I haven't bothered either.

> This change would create hassle in general for many packages, all of
> which create obarrays with (make-vector LENGTH 0), and would need
> changing to use `make-obarray'.

Of course, (make-vector LENGTH 0) still works.  And there's
obarray-make, introduced in Emacs-25, IIRC.  But even if we deprecate
(make-vector LENGTH 0) you won't get any byte-compilation warning for it
since we can't detect whether a (make-vector LENGTH 0) is meant as an
obarray or as a normal array that happens to be filled with zeroes.

> It would mean having to write yet more compatibility macros (for the
> inevitable day when old style obarrays get removed from Emacs).

To the extent that we can't detect when make-vector is used for an
obarray, I expect that it'll take *many* years until we can drop support
for "old-style" obarrays, so I wouldn't worry about it.
I suspect that even cc-mode will have switch to hash-tables before
support for old-style obarrays is dropped.


        Stefan