bug#62734: Always fully rebuild autoloads in package-generate-autoloads

[Eli Zaretskii]

> From: Philip Kaludercic <philipk@posteo.net>
> Cc: leo.gaskin@le0.gs,  62734@debbugs.gnu.org
> Date: Sat, 29 Apr 2023 08:19:35 +0000
> 
> >>   Rebuilding an installation means scraping for new autoload
> >>   cookies, re-compiling Emacs Lisp files, building and installing
> >>   any documentation, downloading any missing dependencies.
> >
> > Thanks.  As a tangent: this is confusing terminology, so it is
> > unfortunate that it was selected for this operation.
> 
> Is there a different way you would have put this?  What we do here is
> sort of combining the steps that the ELPA server and a local
> package-install do into one, and in my mind this constitutes building
> software.

Compilation?  Setup?

"Building" is a strange term when we are talking about a Lisp package.

> > I don't follow: a file that didn't change doesn't need its autoloads
> > updated, right?
> 
> Right, but if we want to add some additional code to the autoloads (as
> we are with package.el, when injecting load-path modification), then
> loaddefs-generate does not re-use the old data, and instead just throws
> it away and creates a new file with contents of EXTRA-DATA, but without
> any autoload information.

That is definitely a bug.  But a package without autoloads is still a
valid use case, and we should support it.

> I think the central issue here is the
> 
>   (and (not defs) extra-data)
> 
> check.  Just because we did not find any new definitions to autoload
> /and/ EXTRA-DATA is non-nil, does not mean the old contents should be
> discarded.  The else-case already does the right thing, so I really do
> think that getting rid of the `if' could resolve the issue:

What happens if a package has no autoloads?  The doc string says in
that case passing EXTRA-DATA will produce OUTPUT-FILE regardless.
Does your patch handle that?  (It's hard to tell, given all the
whitespace changes in the patch.)