name-rev --stdin is slow

name-rev --stdin has some really high start-up time. The example below is on my git.git branch. Is this unavoidable to make name-rev fast for naming high numbers of revisions -- i.e. is it about amortization? Still, 0.7 seconds seems pretty excessive, and on linux-2.6 it even takes 6 seconds.

Anyone care to look into it?

$ time echo HEAD | git name-rev --stdin
HEAD

real 0m0.748s
user 0m0.588s
sys 0m0.080s

$ time git name-rev HEAD
HEAD master

real 0m0.041s
user 0m0.016s
sys 0m0.028s

Actually, I think it's the reverse: without --stdin, name-rev can do a first pass that lets it eliminate a lot of data from consideration; with --stdin, it doesn't know if an item it hasn't seen is going to need some data, and so it's conservative and doesn't eliminate anything, which means that each request is slower, at least until it's cached everything that you turn out to care about.

-Daniel
*This .sig left intentionally blank*

Are you sure that's the case? time git-name-rev --stdin < /dev/null gives the same high startup time without looking up anything.

Anyways, it would be great if someone could try to fix that, since a 6-second startup time on a repository like linux-2.6 makes the --stdin option unusable for applications like gitweb (for which it would actually be quite useful to reduce the number of forks).

-- Lea

Ah, okay. There's a startup pass that goes through all of the refs (tags in particular), and does stuff with them, but only if they aren't filtered out.

It might be possible to collect all of the input, set up the filter, and do the (now cheaper) initial pass, but I don't think that would offer any real advantage over just collecting them in the caller and calling name-rev with a long command line.

-Daniel
*This .sig left intentionally blank*