From: Jeff King Date: Thu, 11 Jun 2026 08:41:32 GMT Subject: Re: [PATCH v2] ls-files: filter pathspec before lstat Message-ID: <20260611084132.GK2191159@coredump.intra.peff.net> In-Reply-To: On Tue, Jun 09, 2026 at 04:15:41PM -0700, Tamir Duberstein wrote: > On Tue, Jun 9, 2026 at 3:41 AM Jeff King wrote: > > > > On Mon, Jun 08, 2026 at 07:37:15PM -0700, Tamir Duberstein wrote: > > > > > + /* > > > + * match_pathspec() is linear in pathspec.nr, so prefilter only > > > + * the single-pathspec case. Only entries shown by show_ce() > > > + * satisfy --error-unmatch. > > > + */ > > > + if (pathspec.nr == 1 && > > > + !match_pathspec(repo->index, &pathspec, fullname.buf, > > > + fullname.len, max_prefix_len, NULL, > > > + S_ISDIR(ce->ce_mode) || > > > + S_ISGITLINK(ce->ce_mode))) > > > + continue; > > > > This feels...kind of arbitrary, no? Surely it's also faster with > > pathspec.nr == 2, and so on up to some nr closer to the size of the > > total index. It feels weird to be making an arbitrary cutoff based on > > pathspec performance in calling code like this. > > > > It is not wrong, per se, as you are optimizing your case without trying > > to hurt any others. But what do we do when somebody profiles it and > > comes along trying to bump the number to 2, or 10? > > > > I dunno. > > Yeah, absolutely it's arbitrary. The simplest answer is that others > are welcome to bump this, provided they make the case for it. OK. I can live with, I suppose, but I am tempted to say that it should just kick in always (i.e., removing the pathspec.nr check). Though I did show a case where the performance regresses, it was pretty made-up and not something I'd expect in the real world. And you'd see that same crappy performance with "git ls-files -- $(git ls-files)", without the "-m". The real solution is making the pathspec code less crappy. -Peff