provide hooks to customize counsel-git-grep behavior on large repos

[articuluxe]

I hope these (non-breaking) changes are OK. The current solution (du -s) can take from 7 - 24 sec. on a medium-size repo, which already takes so long that it doesn't matter that we limit the minimum input to 3 characters. These changes just provide a customization point for users to change this behavior.

[basil-conto]

It's about twice as fast calling

(read (car (process-lines "du" "-s")))

plus you avoid the shell, unless that was intentional for some reason.

[articuluxe]

Thanks, seems to work. I've pushed this change.

[abo-abo]

What makes this more efficient?

[basil-conto]

@abo-abo Calling du directly instead of via the shell cuts out the middle man (the shell).

I just noticed that shell-command-to-string calls process-file (which inspects file handlers) instead of call-process.

It is possible to keep this behaviour without calling the shell, thus still halving subprocess overhead:

(read (with-output-to-string
        (process-file "du" nil standard-output nil "-s")))

Obviously this overhead reduction is completely negligible compared to the execution time of du. My comment was more about avoiding the shell where it's not needed than making any noticeable optimisation.

[abo-abo]

I agree that not piping it all through the shell is a bit more efficient.

On the other hand, using the shell is a lot more flexible:

• You can copy paste the command into an actual shell to test it
• You can add/remove flags more easily, since it's just in the shell format

[basil-conto]

@abo-abo Those aren't the arguments in favour of the shell's flexibility that I would have made (think piping, logical operators...), but agreed, the shell is of course more flexible (and thus problematic when users use anything other than compliant bourne/bash shells with corresponding Unix utils).

[abo-abo]

Thanks, I merged the first commit. I think not merging the second one keeps things more simple.

[articuluxe]

Thanks for merging.

In case anyone is curious, my local change to require 3 chars for counsel-git-grep, regardless of repo size, is to execute the following after 'counsel is loaded (e.g. in use-package's :config stanza):

   (setq counsel--git-grep-count-func
         (lambda() (1+ counsel--git-grep-count-threshold)))

[icarus-sparry]

Please see issue #1502 where a suggested fix uses the flexibility of the shell to at least stop counsel-git-grep from failing.