From: Junio C Hamano <gitster@pobox.com>
Date: Wed, 06 Jan 2010 16:58:42 GMT
Subject: Re: [PATCH 9/9] rerere forget path: forget recorded resolution
Message-ID: <7vr5q31871.fsf@alter.siamese.dyndns.org>
In-Reply-To: <4B444FF9.1000004@kdbg.org>

Johannes Sixt <j6t@kdbg.org> writes:

>> Doesn't "checkout --conflict=merge" use the same ll_merge() machinery?
>
> It does, without setting up .gitattributes, either (IIUC), which is a bug IMO.

I don't think there is any bug there; git_checkattr() that is used by
git_path_check_merge() autoinitializes the attribute machinery.

However, I think you need this fix for "checkout -m" to work as intended,
if your custom merge driver has recursive attribute set to 'binary' or
something:

diff --git a/builtin-checkout.c b/builtin-checkout.c
index 64f3a11..50f5079 100644
--- a/builtin-checkout.c
+++ b/builtin-checkout.c
@@ -167,7 +167,7 @@ static int checkout_merged(int pos, struct checkout *state)
 	fill_mm(active_cache[pos+2]->sha1, &theirs);
 
 	status = ll_merge(&result_buf, path, &ancestor,
-			  &ours, "ours", &theirs, "theirs", 1);
+			  &ours, "ours", &theirs, "theirs", 0);
 	free(ancestor.ptr);
 	free(ours.ptr);
 	free(theirs.ptr);

I don't know why I decided to pass "1" to ll_merge() there; it doesn't
make any sense and looks like an untested bug to me.

The ll_merge() call jc/cache-unmerge topic adds for "rerere forget"
doesn't share this problem.