bug#74243: [PATCH] Speed up vc-hg-state by treating ignored files as unregistered

[Dmitry Gutov <dmitry@gutov.dev>]

Hi, just to interject a little.

On 26/11/2024 09:52, Sean Whitton wrote:
> (I think there's a mistake here: an ignored file is not a file "under
> version control", so `unregistered' should say "not under version
> control and not ignored".  Would you agree?)
> 
> Thanks for pointing out the involvement of find-file-hook and
> after-save-hook.  The problem you describe is not at all Hg-specific:
> vc-state gets called in a context where speed matters, but it's also the
> primary entry point for any code that wants to know the state of a file,
> some of which might care more about accuracy than speed.
> 
> To put it another way, the code assumes throughout that finding out the
> file state will always be fast.  But it also assumes the information is
> accurate if present.  This makes me queasy about your original patch.
> It does not seem wise to return something we don't know to be true only
> on the basis that it all works out fine for now.

This FR reminds me of a similar change in vc-git-state that we ended up 
installing in bug#11757 (in 2012). Then stayed with it until 2017 when 
bug#19343 was filed and fixed (provided a recent enough Git is used) - 
see also the problem scenario described there.

So it seems both a reasonable change and ultimately not ideal. Depending 
on how many users we think might be affected by performance here.

> The 'nil' return value might provide us with a way out, however.
> Could we add an optional argument to vc-state that means "just return
> nil if finding out the state properly might be slow"?
> Could we make vc-after-save and the relevant find-file-hook entry pass
> that option through, and do something sensible with a nil return value?
> 
> If they get nil, they would clear out the saved property, and possibly
> update the mode line display to "????" or something.  Maybe we'd want a
> user option (that could go in your large repo's .dir-locals.el, so it's
> set-and-forget) to opt-in to not knowing the file state as often.

A user option seems like an easier choice.

Solutions that clear cache under some conditions or other tend to be 
more complex, and slow down at least some combined scenarios (e.g. one 
of my use cases is saving the buffer and having diff-hl-mode use its vc 
state from after-save-hook).