Re: [RFD/PATCH 3/5] checkout: Use remote refspecs when DWIMming tracking branches

From: Junio C Hamano <gitster@pobox.com>
Date: Apr 19, 2013, 19:44 UTC
Message-ID: <7vzjwuz4ov.fsf@alter.siamese.dyndns.org>
In-Reply-To: <1366352442-501-4-git-send-email-johan@herland.net>

Johan Herland <johan@herland.net> writes:

Show 22 quoted lines

> The DWIM mode of checkout allows you to run "git checkout foo" when there is
> no existing local ref or path called "foo", and there is exactly one remote
> with a remote-tracking branch called "foo". Git will then automatically
> create a new local branch called "foo" using the remote-tracking "foo" as
> its starting point and configured upstream.
>
> However, the current code hardcodes the assumption that all remote-tracking
> branches are located at refs/remotes/$remote/*, and that "git checkout foo"
> must find exactly one ref matching "refs/remotes/*/foo" to succeed.
> This approach fails if a user has customized the refspec of a given remote to
> place remote-tracking branches elsewhere.
>
> The better way to find a tracking branch is to use the fetch refspecs for the
> configured remotes to deduce the available candidate remote-tracking branches
> corresponding to a remote branch of the requested name, and if exactly one of
> these exists (and points to a valid SHA1), then that is the remote-tracking
> branch we will use.
>
> For example, in the case of "git checkout foo", we map "refs/heads/foo"
> through each remote's refspec to find the available candidate remote-tracking
> branches, and if exactly one of these candidates exist in our local repo, then
> we have found the upstream for the new local branch "foo".

Once you introduce a concrete "foo" as a name in the example, it would be far easier to understand if you spelled all the other assumptions in the example out.

I am _guessing_ that you mean a case like this:

	[remote "origin"]
		fetch = refs/heads/*:refs/remotes/origin/*
	[remote "xyzzy"]
		fetch = refs/heads/*:refs/remotes/xyzzy/nitfol/*
	[remote "frotz"]
		fetch = refs/heads/*:refs/remotes/frotz/nitfol/*

and refs/remotes/origin/foo or refs/remotes/frotz/nitfol/foo do not exist but refs/remotes/xyzzy/nitfol/foo does. And the user says "git checkout foo". Instead of finding a remote ref that matches "refs/remotes/*/foo" pattern (and assuming the part that matched * is the name of the remote), you can iterate the RHS of the refspecs to see if there is a unique match.

Then the new branch can unambiguously find that its upstream is xyzzy's foo.

I think it makes sense to update the semantics like that.

Wouldn't the traditional case (i.e. without "nitfol/" in the xyzzy/frotz remotes above) be a mere special case for this new logic? You mentioned there is a regression caught by existing tests if you go this route, but I do not see how that happens.