Re: [PATCH 1/2] diff: Fix modified lines stats with --stat and --numstat

[Thomas Guyot-Sionnest <tguyot@gmail.com>]

On Fri, 18 Sep 2020 at 10:46, Taylor Blau <me@ttaylorr.com> wrote:
>
> On Fri, Sep 18, 2020 at 07:32:55AM -0400, Thomas Guyot-Sionnest wrote:
> > -     same_contents = oideq(&one->oid, &two->oid);
> > +     if (one->is_stdin && two->is_stdin)
> > +             same_contents = !strcmp(one->data, two->data);
>
> Hmm. A couple of thoughts here:
>
>   - strcmp seems like a slow-down here, since we'll have to go through
>     at worst the smaller of one->data and two->data to compare each of
>     them.
>
>   - strcmp is likely not the right way to do that, since we could be
>     diffing binary content, in which case we'd want to continue over
>     NULs and instead stop at a fixed length (the minimum of the length
>     of one->data and two->data, specifically). I'd have expected memcmp
>     here instead.
>

You're absolutely right - this is a bug I managed to figure out last
night - first real incursion into git C code - and I definitely didn't
think this through. TBH so far I coded mostly with tools dealing in
plaintext and C strings.

>   - Why do we have to do this at all all the way up in
>     'builtin_diffstat'? I would expect these to contain the right
>     OIDs by the time they are given back to us from the call to
>     'diff_fill_oid_info' in 'run_diffstat'.
>
> So, my last point is the most important of the three. I'd expect
> something more along the lines of:
>
>   1. diff_fill_oid_info resolve the link to the pipe, and
>   2. index_path handles the resolved fd.
>
> ...but it looks like that is already what it's doing? I'm confused why
> this doesn't work as-is.

So the idea is to checksum the data and write a valid oid. I'll see if
I can figure that out. Thanks for the hint though else I would likely
have gone with a buffer and memcmp. Your solution seems cleaner, and
there is a few other uses of oideq's that look dubious at best with
the case of null oids / buffered data so it's definitely a better
approach.

> > +     else
> > +             same_contents = oideq(&one->oid, &two->oid);
> >
> >       if (diff_filespec_is_binary(o->repo, one) ||
> >           diff_filespec_is_binary(o->repo, two)) {
> > diff --git a/t/t3206-range-diff.sh b/t/t3206-range-diff.sh
> > index e024cff65c..4715e75b68 100755
> > --- a/t/t3206-range-diff.sh
> > +++ b/t/t3206-range-diff.sh
> > @@ -258,11 +258,11 @@ test_expect_success 'changed commit with --stat diff option' '
> >            a => b | 0
> >            1 file changed, 0 insertions(+), 0 deletions(-)
> >       3:  $(test_oid t3) ! 3:  $(test_oid c3) s/11/B/
> > -          a => b | 0
> > -          1 file changed, 0 insertions(+), 0 deletions(-)
> > +          a => b | 2 +-
> > +          1 file changed, 1 insertion(+), 1 deletion(-)
> >       4:  $(test_oid t4) ! 4:  $(test_oid c4) s/12/B/
> > -          a => b | 0
> > -          1 file changed, 0 insertions(+), 0 deletions(-)
> > +          a => b | 2 +-
> > +          1 file changed, 1 insertion(+), 1 deletion(-)
>
> The tests definitely demonstrate that the old behavior was wrong,
> though...
>

For the records I verified the actual diffs (I think they're even
hardcoded in the earlier tests) and the remaining 0-add/del's are also
valid.

Regards,

Thomas