Re: [StGit PATCH] Read several objects at once with git cat-file --batch

From: Karl Hasselström <kha@treskal.com>
Date: Aug 10, 2008, 23:58 UTC
Message-ID: <20080810235845.GA5106@diana.vm.bytemark.co.uk>
In-Reply-To: <b0943d9e0808101525w1e6f1e60h162a4b137d6dca6c@mail.gmail.com>

On 2008-08-10 23:25:08 +0100, Catalin Marinas wrote:

Show 11 quoted lines

> 2008/8/8 Karl Hasselström <kha@treskal.com>:
>
> > Instead of spawning a separate cat-file process for every blob and
> > commit we want to read. This speeds things up slightly: about 6-8%
> > when uncommitting and rebasing 1470 linux-kernel patches
> > (perftest.py rebase-newrebase-add-file-linux).
>
> Which version of Git got the --batch option to git-cat-file? It
> might be possible that default Git in Debian (testing) or Ubuntu
> doesn't have this option. Maybe we could still have the original
> behaviour as a fallback.

Hmm, I never realized it was so new. It's not in any 1.5.5 release, but it is in 1.5.6-rc0. So realistically, we'll have to require version 1.5.6 for this. (The patch should add a paragraph about the required git version to some doc file.)

A fallback is certainly conceivable. The only nontrivial thing about it would be to detect when it's necessary. And to make sure both code paths are tested ...

(I have a similar patch that uses diff-tree --stdin, but that needs a git with the patches I posted some hours ago.)

> Otherwise, the patch looks allright. It took me a bit of time to see
> why we need the new run_background() function (but in my current
> Git, 1.5.3.4.206.g58ba4, there wasn't such an option; I had to
> upgrade).

Thanks. I'm not 100% pleased with the background running stuff yet. It works, but could use some refactoring so that callers don't have to know too much.

-- 
Karl Hasselström, kha@treskal.com
      www.treskal.com/kalle