Hi Jim,

On 29/10/2023 07:36, Jim Porter wrote:
> By default, this is just a theoretical problem, but if you customize 
> 'xref-search-program-alist' and 'xref-search-program' to include some 
> other program, this can cause real issues. For example, I tried to add 
> "ag" to this list[1], and unfortunately, it just doesn't work in this 
> case. The results for submodules get duplicated, and there's no way I 
> can see with ag to search only the specified *files*, ignoring any 
> directories. (Looking at the definition for ripgrep, I'm guessing the 
> "-g '!*/'" is the trick for that program, but nothing similar works for 
> ag.)
> 
> Attached are two possible patches for this: a minimal version that just 
> fixes 'project-find-regexp', and a maximal version that fixes this in 
> general, and should theoretically speed up 'project-search' and 
> 'project-query-replace-regexp' since they no longer need to call 
> 'file-regular-p' on every file.

I kept this unfortunate situation around because every obvious fix 
brought non-negligible performance regressions: the version with 
file-regular-p slowed one of my examples (Mozilla's repo) by 370%. Your 
cl-set-difference version slowed it down by 10-14% -- better, but still 
something that seemed worse on balance when I tried this before, so I 
preferred to work around it in command implementations: the "-s" or 
"--no-messages" flags xref-search-program-alist.

And that's not to mention usage over Tramp (which would be affected by 
the +1 process call that you mentioned as well, but that seems unavoidable).

Anyway, after recent experience micro-optimizing list operations, I came 
up with this version where the impact seems minimal.

WDYT?