Jeff King [off-list ref] writes:
Right. There's definitely a question of whether --cherry-pick is the
most useful thing in such a situation, but the current behavior was
simply broken. It did not behave as advertised, and it treated one side
of the history different from the other. So whether we want to do
anything else to help that case, I think this at least makes
--cherry-pick sane. :)
Yes; I think "sane" does not always equal to "useful", though ;-)
Here are two related histories to think about.
...
I suspect for most operations we care less about "remove all
cherry-picks from both sides", but rather "give me one side, omitting
commits that are already on the other side" (because you want to rebase
or cherry-pick some subset).
Yes again. It means "--cherry-pick" inherently is not symmetric.
One side is the reference side (i.e. mainline), and the goal is to
remove from the other side what is in the reference side.
What we are seeing in this discussion is that "--cherry-pick" and
symmetric difference do not conceptually play well together.
But the behaviour with the patch makes the most sense when
cherry-pick is applied to a symmetric difference (provided that
doing so makes sense in the first place, that is).
On Wed, Jan 13, 2021 at 12:21:30PM -0800, Junio C Hamano wrote:
quoted
I suspect for most operations we care less about "remove all
cherry-picks from both sides", but rather "give me one side, omitting
commits that are already on the other side" (because you want to rebase
or cherry-pick some subset).
Yes again. It means "--cherry-pick" inherently is not symmetric.
One side is the reference side (i.e. mainline), and the goal is to
remove from the other side what is in the reference side.
What we are seeing in this discussion is that "--cherry-pick" and
symmetric difference do not conceptually play well together.
But the behaviour with the patch makes the most sense when
cherry-pick is applied to a symmetric difference (provided that
doing so makes sense in the first place, that is).
I didn't realize --cherry-pick would work without a symmetric
difference. The documentation says:
Omit any commit that introduces the same change as another commit on
the “other side” when the set of commits are limited with symmetric
difference.
And indeed, I think it silently does nothing with:
git rev-list --cherry-pick A..B
(because there is nothing on the "other" side to match).
So you do need some symmetric difference in order to define the two
sets, but you might only want to see one of the sides. And that is
basically what --cherry does. But having looked at the implementation of
cherry_pick_list(), it is quite happy to swap the sides internally. I
guess if we were going to make the output unsymmetric, the first thing
would be to stop doing that swap. :)
For the specifics of reverts and replays, though, I think the weirdness
has nothing to do with left/right, or showing one side. It's the
mismatched count. So if we were to make any change, it would be to keep
a count of how many times each commit appears on the other side, and
cancel them one for one.
I.e., this:
o--M--W--M--o--
\
o--M--o---
might plausibly want to show "M" once on the upper branch. But:
o--M--W--M--o--
\
o--M--W--M--o--
should never show M, I wouldn't think.
I'm not sure if that would be violating what existing callers expect
from "--cherry-pick", though (i.e., if it would need another name, or an
extra option to trigger).
-Peff