On 06.07.2012 17:44, Michael Albinus wrote:
> Dmitry Gutov <dgutov@yandex.ru> writes:
>
>>> If we assume that there are no dangerous vc commands outside Emacs, we
>>> wouldn't have a problem.
>>
>> In this case, the behavior of the first patch I posted here should be
>> acceptable, right? It's simpler, has pretty much the same effect, and
>> should be a tiny bit faster.
>
> That I don't know. Both patches do almost what we expect, and having a
> cached value for `vc-registered' sounds more performant when applied
> often enough.

With the attached patch, `vc-git-registered' is usually called only once 
during the lifetime of a buffer.

For caching to be an improvement, 'git-registered value has to be 
invalidated separately from 'vc-backend, yet less often than 
`vc-git-state' is called.

>>> Yes. I don't know, whether we will be able to handle any surprise when
>>> using caches. There will always be a scenario which lets fail a given
>>> algorithm. I fear.
>>
>> Sure, but I'm just asking for one scenario that works better with
>> explicitly caching 'git-registered, instead of not calling it in
>> vc-git-state'.
>> If `vc-git-state' doesn't call `vc-git-registered' (just assumes it's
>> t), then `vc-registered' is the latter's only client, and so its
>> return value is implicitly cached in 'vc-backend property.
>
> Maybe. Could you show a patch? (Please with ChangeLog entry, I would
> commit if it looks good).

* vc-git.el (vc-git-state): Don't call `vc-git-registered', assume it's 
always t.
(vc-git-registered): Remove caching, the function is only called once.

Thank you,

--Dmitry