Re: [PATCH] merge: allow using --no-ff and --ff-only at the same time

[Miklos Vajna <vmiklos@suse.cz>]

[-- Attachment #1: Type: text/plain, Size: 3046 bytes --]

Hi Michael,

On Mon, Jul 01, 2013 at 04:52:29PM +0200, Michael Haggerty <mhagger@alum.mit.edu> wrote:
> On 07/01/2013 09:01 AM, Miklos Vajna wrote:
> > 1347483 (Teach 'git merge' and 'git pull' the option --ff-only,
> > 2009-10-29) says this is not allowed, as they contradict each other.
> > 
> > However, --ff-only is about asserting the input of the merge, and
> > --no-ff is about instructing merge to always create a merge commit, i.e.
> > it makes sense to use these options together in some workflow, e.g. when
> > branches are integrated by rebasing then merging, and the maintainer
> > wants to be sure the branch is rebased.
> 
> That is one interpretation of what these options should mean, and I
> agree that it is one way of reading the manpage (which says
> 
> --ff-only::
> 	Refuse to merge and exit with a non-zero status unless the
> 	current `HEAD` is already up-to-date or the merge can be
> 	resolved as a fast-forward.
> 
> ).  However, I don't think that the manpage unambiguously demands this
> interpretation, and that (more importantly) most users would be very
> surprised if --ff-only and --no-ff were not opposites.

Yes, I agree that that's an unfortunate naming. --ff and --no-ff is the
opposite of each other, however --ff-only is independent, and I would
even rename it to something like --ff-input-only -- but I don't think
it's worth to do so, seeing the cost of it (probably these options are
used by scripts as well).

> How does it hurt?  If I have configuration value merge.ff set to "only"
> and run "git merge --no-ff" and then I merge a branch that *cannot* be
> fast forwarded, the logic of your patch would require the merge to be
> rejected, no?  But I think it is more plausible to expect that the
> command line option takes precedence.

Hmm, I did not remember that actually merge.ff even uses the same
configuration slot for these switches. :-( Yes, that would make sense to
fix, once the switches can be combined. Maybe merge.ff and
merge.ff-only?

> In my opinion, your use case shouldn't be supported by the command
> because (1) it is confusing,

I don't see why it would be confusing. I think using these two options
together is one way to try to get the benefits of both rebase (cleaner
history) and merge (keeping the history of which commits came from a
given merge).

> (2) it is not very common,

Hard to argue that argument. :-) No idea what counts as common, my
motivation is that some projects (e.g. syslog-ng) integrate *every*
feature branch this way, and doing this "manually" (as in indeed
manually or by using a helper script) seems suboptimal, when the support
for this is already mostly in merge.c, just disabled.

> easy to work around:
> 
>     if git merge-base --is-ancestor HEAD $branch
>     then
>         git merge --no-ff $branch
>     else
>         echo "fatal: Not possible to fast-forward, aborting."
>         exit 1
>     fi

Right, that's indeed a viable workaround for the problem.

Miklos

[-- Attachment #2: Digital signature --]
[-- Type: application/pgp-signature, Size: 198 bytes --]