Re: [PATCH v2 2/7] add tests for rebasing with patch-equivalence present

On Thu, May 30, 2013 at 5:54 AM, Johannes Sixt [off-list ref] wrote:

Am 30.05.2013 07:30, schrieb Martin von Zweigbergk:

quoted

On Wed, May 29, 2013 at 12:09 AM, Johannes Sixt [off-list ref] wrote:

quoted

Am 5/29/2013 8:39, schrieb Martin von Zweigbergk:

quoted

+#       f
+#      /
+# a---b---c---g---h
+#      \
+#       d---G---i

...

quoted

+test_run_rebase () {
+     result=$1
+     shift
+     test_expect_$result "rebase $* --onto drops patches in onto" "
+             reset_rebase &&
+             git rebase $* --onto h f i &&
+             test_cmp_rev h HEAD~2 &&
+             test_linear_range 'd i' h..

Isn't this expectation wrong? The upstream of the rebased branch is f, and
it does not contain G. Hence, G should be replayed. Since h is the
reversal of g, the state at h is the same as at c, and applying G should
succeed (it is the same change as g). Therefore, I think the correct
expectation is:

                test_linear_range 'd G i' h..

Good question! It is really not obvious what the right answer is. Some
arguments in favor of dropping 'G':

1. Let's say origin/master points to 'b' when you start the 'd G i'
branch. You then send the 'G' patch to Junio who applies it as 'g'
(cherry-picking direction is reversed compared to figure, but same
effect). You then "git pull --rebase" when master on origin points to
'h'. Because of the cleverness in 'git pull --rebase', it issues 'git
rebase --onto h b i'.

The reason for this git pull cleverness is to be prepared for rewritten
history:

   b'--c'--g'--h'
  /
 a---b
      \
       d---G---i

to avoid that b is rebased.

Right. It doesn't currently drop 'G' and maybe it shouldn't, so let's
leave it as is for now, I would say.

quoted

2. In the test a little before the above one, we instead do 'git
rebase --onto f h i' and make sure that the 'G' is _not_ lost. In that
case it doesn't matter what's in $branch..$upstream. Do we agree that
$branch..$upstream should never matter (instead, $upstream is only
used to find merge base with $branch)?

No, we do not agree. $branch..$upstream should be the set of patches
that should be omitted. $branch..$onto should not matter. $onto is only
used to specify the destination of the rebased commits.

Ok. As I said to Felipe, I'm not sure why I was so convinced that it's
bad to lose the patch in 'git rebase --onto f h i'. It can result in
lost work, but it seems rare enough that no one has reported it,
AFAIK.

I'll change those tests in a re-roll, and perhaps I'll drop a few of
them. Let me know if you (anyone) disagree.


Martin

PS. Thanks for a meticulous review!