Re: Removing unloaded functions from auto-mode-alist.

[Lute Kamstra]

David Kastrup <address@hidden> writes:

> Lute Kamstra <address@hidden> writes:
>
>> Consider the following situation:
>>
>> function 'a is autoloaded: (autoload "blabla" ...).
>> function 'b is autoloaded: (autoload "other" ...).
>> function 'c is defined.
>> function 'd is unbound.
>>
>> "blabla" defines 'a, 'b, 'c, and 'd as functions.
>>
>> Do (require 'blabla) and then (unload-feature 'blabla).
>>
>> Currently, all four functions will be unbound by unload-feature.  You
>> propose to let unload-feature restore both 'a and 'b to their previous
>> autoloads [1].
>
> Sure.
>
>> But what should be done with 'c?
>>
>> I think restoring 'c to its previous definition would be the right
>> thing to do.
>
> Not at all.  The purpose of unload-feature is to be able to restore a
> state, most particular to conserve memory.  So unload-feature should
> not waste memory by keeping in effect a history of load sequences
> around.  Its purpose should be confined to unloading those features
> that _can_ reasonably be unloaded.  And that means no functions
> whatsoever that _redefine_ stuff.  The main purpose of the autoload
> restoration is to restore autoloads into the package itself, not
> foreign autoloads.

For me, it is most intuitive/logical that unload-feature tries to undo
the effects of loading the feature.  It is very uncommon that a
feature redefines a function (or a variable), so keeping track of such
cases will not waste much memory.

> I think that unload-feature should in the case of c being redefined
> simply barf and refuse to unload the feature.

That's quite extreme.  And it would require you too keep track of
redefinitions.  Why not use that tracking machinery to just restore
these rare cases?

>> But that would be quite a substantial change.  It's probably best to
>> leave this alone until after the release.
>
> Lute, _any_ change in that area is best left alone.  Richard has
> already clearly stated that we are not going to fiddle with it before
> the release, and I can only agree.  The current state of brokenness
> has been there for a long time, and nobody really complained.  We
> can't hope to get a serious amount of testing for this sort of stuff
> in before the release.  And the effects might be memory leakage and
> similar hard to find things that take people months to figure out.
>
> We can't invest the time to make sure that nothing like this will
> happen.

Ok.

Lute.