Re: [PATCH 1/2] merge: fix numerus bugs around "trivial merge" area

From: Junio C Hamano <gitster@pobox.com>
Date: Aug 24, 2008, 01:58 UTC
Message-ID: <7vtzdbuqb6.fsf@gitster.siamese.dyndns.org>
In-Reply-To: <7vfxovy06u.fsf_-_@gitster.siamese.dyndns.org>

Junio C Hamano <gitster@pobox.com> writes:

Show 7 quoted lines

> The primary logic of the merge, however, had an assumption that the
> process never read the index in-core, and write_cache_as_tree() call it
> makes from write_tree_trivial() will always read from the on-disk index
> the strategies created and write it out as a tree.  This assumption is now
> broken by the above fix.  It now calls discard_cache() before calling
> write_tree_trivial() when it wants to write the on-disk index as a tree to
> fix this issue.

By the way, in the medium term, if we are serious about making an internal call to merge_recursive() from cmd_merge(), I think we may be better off making it the responsibility for try_merge_strategy() to leave an committable state in the in-core index (aka "the_index") when they return with 0 (success) status. After calling external ones via the run_command interface, it should do a read_cache() (after calling discard_cache() if needed); if it calls merge_recursive(), hopefully you already have the committable state in the in-core index.

That way, when automerge succeeds, write_tree_trivial() can write that in-core index out and create the tree object to be committed. The callchain to use merge_recursive() can avoid having to write to the on-disk index, read it again and write out the tree from it.