bug#73801: 31.0.50; project-try-vc sometimes set wrong cache project-vc-extra-root-markers

[Eli Zaretskii <eliz@gnu.org>]

> Date: Tue, 29 Oct 2024 22:31:04 +0200
> Cc: 73801@debbugs.gnu.org, i@fuzy.me
> From: Dmitry Gutov <dmitry@gutov.dev>
> 
> On 29/10/2024 15:33, Eli Zaretskii wrote:
> >> Date: Tue, 29 Oct 2024 04:44:13 +0200
> >> From: Dmitry Gutov <dmitry@gutov.dev>
> >> Cc: Zhengyi Fu <i@fuzy.me>
> >>
> >> Since I see some changes added to the release branch still,
> >>
> >> On 28/10/2024 06:06, Dmitry Gutov wrote:
> >>> It would be nice to get either of the patches into Emacs 30, too, but it
> >>> might be a little late given where it is in the pretest.
> >>
> >> Eli, could we install either of the fixes for this bug to emacs-30 too?
> >>
> >> The one I installed on master is longer but should result in less I/O,
> >> while the patch by Zhengyi Fu is a one-liner, which might feel a little
> >> safer.
> > 
> > I don't understand the implications of that one-line (nor, TBH, the
> > analysis of the original problem), so I'm not sure these changes are
> > safe.
> 
> The original problem was due to project-try-vc being invoked recursively 
> on a parent directory while a variable that affects its computation is 
> bound to nil. The function itself (project-try-vc) memoizes its return 
> value. As a result, any subsequent call to it with the same argument 
> outside of the said binding could return wrong result.
> 
> The first fix (one-liner) made sure that we're calling it with the same 
> argument that is passed to the current call, ensuring that the cache 
> will be rewritten after it returns.
> 
> The second fix (mine) was to extract the value computation into a helper 
> function, making the recursive call to not be memoized.
> 
> > How do we know that catering to this corner case will not screw
> > other corner cases?
> 
> Difficult to guarantee that 100%, but this specific case seems important 
> enough, while at the same time we can infer that the change won't affect 
> the majority scenario because the code is guarded by these conditions:
> 
>      (when root
>        (when (not backend)
>           ...

FWIW, I have a bad feeling about this, but if you are confident, feel
free to backport.