bug#73886: 29.4; Confusing info about void function cells in Emacs Lisp manual

[Eli Zaretskii <eliz@gnu.org>]

> From: Stefan Monnier <monnier@iro.umontreal.ca>
> Cc: ulm@gentoo.org,  73886@debbugs.gnu.org
> Date: Sun, 20 Oct 2024 12:56:03 -0400
> 
> >> Oops.  Looks like I missed this part when I changed it back around
> >> Emacs-24.4:
> >> 
> >>     ** In 'symbol-function', nil and "unbound" are indistinguishable.
> >>     'symbol-function' does not signal a 'void-function' error any more.
> >>     To determine if a symbol's function definition is void, use 'fboundp'.
> >
> > Could you explain the rationale for that change?  I tried to look for
> > relevant discussions around that date, but came up empty-handed.
> 
> I can't remember discussing it, no.  It was a kind of "executive
> decision".
> 
> Having a special "void" (`Qundefined`) non-value for the `symbol-value`
> is needed for `boundp` since variables can contain *any* value, but not
> for the `symbol-function` part where we can use any normal value (I
> chose `nil`) to play this role as long as it doesn't collide with values
> normally held in the `symbol-function` slot, like function names,
> function values, cons cells, vectors, ...
> 
> The upside was a simplification in various chunks of code which used to
> do things like `(and (fboundp SYM) (symbol-function SYM))` which can now
> be simplified to `(symbol-function SYM)`.
> 
> I remember two "motivators", i.e. places where the need to pay attention
> to the special void case annoyed me enough to look into this and make
> the change, one was `nadvice.el` and the other was `cl-letf`.
> [ So, it was no accident that the change happened in the same release as
>   the addition of `nadvice.el`.  ]
> 
> In both cases the issue is that we want to deal with "places"
> (generalized variables) and that abstraction works well for those places
> which *always* contain a value, but not as well for those special places
> that can be "unbound", so removing the "unbound" case from
> `symbol-function` resulted in a welcome simplification.
> For the same reason I dislike EIEIO's notion of `slot-boundp` and have
> already considered marking it obsolete.

Thanks, I've now updated the documentation, and I'm closing this bug.