bug#50906: xref-find-references blocks Emacs: asynchronous operation?

[Dmitry Gutov]

On 05.10.2021 09:29, Helmut Eller wrote:
> On Tue, Oct 05 2021, Dmitry Gutov wrote:
>
> > What can be done here:
> >
> > - Design an "asynchronous" format for xref-show-xrefs-function to
> >    consume. FETCHER of a different shape. Not sure how it's going to
> >    work in the end -- maybe a simple-ish iterator (call a function
> >    again for more results), but ideally it would look synchronous
> >    somehow, and the concurrency would be achieved through the use of
> >    threads. Not sure if that's realistic.
> >
> > - The new kind of fetcher would need to provide a way to abort the
> >    search, since 'C-g' would not be available anymore.
> >
> > - Implement it for the common searches of course.
>
> I think promises, as used in the Javascript world, would be a good fit
> for this kind of problem.  Something like this:
> https://github.com/chuntaro/emacs-promise.

A promise is something that resolves once. We could build on top of this 
concept, but what's really needed is some sort of a lazy sequence 
(Clojure-style), or a sequence of chunks.

> > Downsides:
> >
> > - No way to quickly 'C-g' out of a search, supposedly one would have
> >    to switch to the results buffer (maybe it will be selected right
> >    away) and type 'C-c C-c'. And then kill the buffer, I guess?
>
> Maybe we could have some "promise framework" that solves this problem
> more generally, e.g., a list-promises command that works like
> list-processes and offers a command to cancel promises.

It would need be accessible by the code handling the "abort" command, 
not just by some special UI accessible to the user separately.

But some Promise/Future implementations include the "abort" 
functionality, so it can work together.

> > - The size threshold of a project where the improvement will be
> >    significant is pretty big -- for instance, searching across the
> >    Emacs checkout takes about 100-200ms (just the time the external
> >    process uses). If the search results in many matches (1000s or
> >    10000s) the results will take a while to display, but most of the
> >    time is taken up by processing of results which is implemented in
> >    Lisp. We might have Emacs which shows the first results soon, but
> >    then remains sluggish until all search results are processed. This
> >    problem could be worked around, however, by limiting the displayed
> >    number of results and having buttons like the ones at the bottom of
> >   vc-print-root-log output buffer.
> >
> > - Search results come in unsorted, and, in the case of ripgrep, sorted
> >    randomly every time the search is performed (the files, at
> >    least). We sort them now at the bottom of xref-matches-in-files, but
> >    asynchronous search results would make that infeasible.
>
> This is a good point and probably quite difficult to solve.  I'm
> wondering if it would be possible to build some kind of index, like
> search engines do.  So instead of grepping, we'd use the index and maybe
> invest more effort in ranking the results?

For xref-find-references in particular, you can build an index using 'ID 
Utils' already, and the search will be fast. The downside is you will 
need to update this index manually when the project changes. E.g. when 
you switch to a different repository branch.

And the ripgrep devs are working on something similar: 
https://github.com/BurntSushi/ripgrep/issues/1497

Not sure how far off in the future that is, though.

A really fast searcher solves the biggest part of the problem, but we'd 
still be left with very imprecise searches (many matches) locking up 
Emacs for seconds, since the Lisp overhead processing a match is 
unavoidably larger than the time it takes for a search program to print 
it. Using lazy sequences could allow us some leeway as well -- namely, 
processing only the first N hits initially, and then processing the rest 
only if the user requests that.

If we only target this kind of improvement, the "abort" functionality 
could wait. We'd still need to choose between sorting the results and 
saving on parsing the output buffer eagerly, though.