From: Junio C Hamano <gitster@pobox.com>
Date: Wed, 29 Jul 2009 08:33:29 GMT
Subject: Re: [PATCH 1/2] Demonstrate bugs when a directory is replaced with a symlink.
Message-ID: <7v4osvyjl2.fsf@alter.siamese.dyndns.org>
In-Reply-To: <4A70062A.4040008@drmicha.warpmail.net>

Michael J Gruber <git@drmicha.warpmail.net> writes:

>> +test_expect_failure 'checkout should not delete a/b-2/c/d' '
>> +	git checkout -b temp HEAD^ &&
>> +	test -f a/b-2/c/d
>> +'
>> +
>> +test_expect_failure 'merge should not delete a/b-2/c/d' '
>> +	echo x > a/x &&
>> +	git add a/x &&
>> +	git commit -m x &&
>> +	git merge master &&
>> +	test -f a/b-2/c/d
>> +'
>> +
>> +test_done
>
> Isn't the failure of the second test caused by that of the first one?
> a/b-2/c/d is gone from the worktree, and master does not touch it, so
> the merge leaves the worktree version (non-existent) as is.

To avoid that impression the second test should probably have been written
to start from a clean slate, using "reset --hard" or something.

Kjetil's patch actually fixes the first one, but the second one will still
show breakage.

I wonder if the breakage is in recursive merge or in the generic read-tree
three-way merge code.  I highly suspect that using "git merge -s resolve"
would make the test pass.  Historically recursive merge is known to be
careless in many corner cases.