Re: [PATCH] Teach git log --check to return an appropriate error code

[Junio C Hamano <gitster@pobox.com>]

Johannes Schindelin <Johannes.Schindelin@gmx.de> writes:

> On Sat, 9 Aug 2008, Peter Valdemar Mørch wrote:
>
>> 	Whether or not a check fails is stored in the
>> 	DIFF_OPT_CHECK_FAILED field of flags in struct diff_options.
>> 	This flag-field is only set (diff.c:1644), never cleared.
>
> That is a side effect.  How wise is it to rely on that?

Hmm, good point.

The bit will never be cleared during a single diff run by design, because
it needs to be cumulative in order to check a patch that describes changes
to multiple paths --- iow, the API sequence is (1) the caller to the diff
machinery resets the bit to zero and then (2) the caller exercises the
diff machinery and expects the machinery to set the bit if even a single
failure is detected, or leaves it unset if there is none.

So unless you (log_tree_diff(), the caller of diff machinery), decide to
explicitly reset the bit (or decide to use a freshly allocated and
initialized diff_options for each commit it feeds diff_tree_sha1()), the
assumption would hold.  We need to see how plausible it would be for us to
break that assumption in the future.

Future versions of log_tree_diff() may want to tweak opt->diffopt per
commit, when we have options for "use larger -U<lines> value after hitting
this commit", or "use this pathspec to limit the diff output after hitting
this commit", for example.  But even in these cases, I think it is
implausible to start from a freshly initialized diff_options structure.
The code most likely would start from the copy of what was in use and
update only the necessary fields, without disturbing the state variables.

So I think you are worried a bit too much in this case, even though it is
a valid concern in principle.  It might warrant a comment somewhere inside
log_tree_diff() to tell people not to re-initialize opt->diffopt per
commit without thinking, though.

One interesting option that might be interesting to add to the log family
would be to show only commits that fail the checkdiff tests.  I suspect
necessary change for doing so would go to log_tree_diff() codepath.