bug#42966: 28.0.50; vc-dir: wrong backend

[Dmitry Gutov]

On 26.10.2020 23:54, Glenn Morris wrote:
> Dmitry Gutov wrote:
>
> > > That makes sense, but it's just a performance hack, isn't it?  The
> > > result should be the same as the less invasive "loop over all the
> > > backends and collect the most specific one".
> >
> > Pretty much. Except it should naturally limit the traversal up the
> > directory tree, so it feels like a good architecture, not just a
> > "hack".
>
> Indeed, it just seems like the Right Thing to Do, not a hack.
> Not having been paying attention, I was surprised to see the adopted solution
> goes for "loop over every VC backend, and every directory up the tree,
> then filter the results", rather than "walk up the directory tree,
> stopping when a backend claims responsibility".

I didn't want to insist on it because upon some thinking it seemed to me 
that the remote case might be the only problematic one. And 
one-traversal-per-backend might be more optimizable by Tramp (e.g. via a 
mini-program) than one-check-per-directory-level. But perhaps the latter 
could be optimized using a file handler or an advice just as well.

> I would think efficiency matters in such a frequent operation.
> As a (completely unscientific) data point, my first single core
> bootstrap build after this change took about 5% longer than before
> (22m5s v 21m4s). But of course a single measurement means nothing.

Was that with a rotating disk?

A few more experiments should help, to establish whether that was a 
fluke or not.