Re: [PATCH 7/7] xdiff: make diff3 the default conflictStyle

From: Elijah Newren <newren@gmail.com>
Date: Jun 11, 2021, 17:57 UTC
Message-ID: <CABPp-BGtYnVPijg2OWoDBM915-PPFXk1O3H=BMf_itc4dNjAxQ@mail.gmail.com>
In-Reply-To: <60c39e1c52c04_8d0f208b2@natae.notmuch>

On Fri, Jun 11, 2021 at 10:32 AM Felipe Contreras <felipe.contreras@gmail.com> wrote:

Show 25 quoted lines

>
> Elijah Newren wrote:
> > On Fri, Jun 11, 2021 at 7:25 AM Felipe Contreras <felipe.contreras@gmail.com>
> > wrote:
>
> > > Personally I have never experienced what you posted, so maybe there's
> > > something else happening behind the scenes.
> > >
> > > Maybe merge-ort changed something.
> >
> > merge-ort made no changes relative to content merges or choice of merge
> > bases.  (In fact, merge ort doesn't even handle content merges; that's the
> > xdiff layer.)  Even if merge-ort had made changes in this area, merge-ort
> > is not the default and I didn't see the necessary config tweaks in your
> > list of config options.  (I would have recommended against people using
> > merge ort until 7bec8e7fa6 ("Merge branch 'en/ort-readiness'", 2021-04-16),
> > which only made it into a release last week with 2.32.  I probably won't be
> > recommending it as the default at least until the optimization work is
> > merged and it's hard to predict how many more months that will take.)
>
> Indeed, I tested on v2.25 and found the same output.
>
> I thought of merge-ort because 1) I've never seen such kind of output
> before, and 2) grepping the code I thought I saw merge-ort being the
> default of something, but now I seem to be unable to find where.

We have briefly discussed multiple times what things might be needed to eventually make merge-ort the default (though it's not even complete yet; I'm five months into upstreaming the optimization patches with an unknown number of months left). There were also a couple patches to make the _tests_ default to using merge-ort on most platforms, while still keeping one suite that tests merge-recursive to ensure we don't add breakage. Perhaps one of those is what you're thinking of?

Show 7 quoted lines

> > It's more likely that the codebases you work with just don't have
> > criss-cross merges.
>
> Yes, that's it.
>
> I don't see why people in these kinds of codebases would like diff3
> doing that by default.

I suspect they don't[1]. What's the alternative, though? Not using diff3? Picking a different base to avoid the occasional nested conflict in the inner merge region, but which overall has much worse other side effects? I think Junio was addressing this when he recently said elsewhere in this thread that "Rejecting diff3 style output because of the way a conflicted part in the inner merge appears as a common ancestor version may be throwing the baby out with the bathwater"[2]. Sure, it's an annoyance, but diff3 is still a good option and there is no current solution to the annoying nested merge display.

Also, not sure whether it matters or not, but just for completeness I should point out that you can get nested conflict markers without criss-cross merges and without merge.conflictStyle=diff3. It's just much more rare.

[1] https://lore.kernel.org/git/CAPc5daUC+6cHpexXTO24p4mG_5eL1JmxrYm8h3UfdTh_FMka=w@mail.gmail.com/ [2] https://lore.kernel.org/git/xmqqh7i5ci3t.fsf@gitster.g/