Re: [PATCH] rebase --preserve-merges keeps empty merge commits

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Martin von Zweigbergk wrote:
> I'm working on a re-roll of
> http://thread.gmane.org/gmane.comp.version-control.git/205796
>
> and finally got around to including test cases for what you fixed in
> this patch. I want to make sure I'm testing what you fixed here. See
> questions below.

Thanks for that.  I should have done this myself.

> On Sat, Jan 12, 2013 at 12:46 PM, Phil Hord <hordp@xxxxxxxxx> wrote:
>> Since 90e1818f9a  (git-rebase: add keep_empty flag, 2012-04-20)
>> 'git rebase --preserve-merges' fails to preserve empty merge commits
>> unless --keep-empty is also specified.  Merge commits should be
>> preserved in order to preserve the structure of the rebased graph,
>> even if the merge commit does not introduce changes to the parent.
>>
>> Teach rebase not to drop merge commits only because they are empty.
> Consider a history like
>
> # a---b---c
> #      \   \
> #       d---l
> #        \
> #         e
> #          \
> #           C
>
> where 'l' is tree-same with 'd' and 'C' introduces the same change as 'c'.
>
> My test case runs 'git rebase -p e l' and expects the result to look like
>
> # a---b---c
> #      \   \
> #       d   \
> #        \   \
> #         e---l
>

This is probably right, but it is not exactly the case that caused my itch.
I think my branch looked like this:

# a---b---c
#      \   
#       d---f
#        \   \
#         e---g
#              \
#               l

where g is tree-same with f.  That is, e merged with f, but all of e's
changes were dropped in the merge.

So when I ran 'git rebase -p c l', I expected to end up with this:

# a---b---c
#          \   
#           d---f
#            \   \
#             e---g
#                  \
#                   l

But instead, I got an error because git-rebase--interactive.sh decided
that g was empty, so it dropped it by commenting it out of the todo
list:

pick d
pick e
pick f
#pick g
pick l

At the end of this attempt, I got some odd error about a cherry-pick
have incorrect parameters or somesuch.  I bisected the problem to a
commit that clued me in to one of my commits being silently dropped.
And that is specifically what I fixed.

This happened only because 'is_empty_commit' checks for tree-sameness
with the first parent; it does not consider whether there are multiple
parents.  Perhaps it should.

>> A special case which is not handled by this change is for a merge commit
>> whose parents are now the same commit because all the previous different
>> parents have been dropped as a result of this rebase or some previous
>> operation.
> And for this case, the test case runs 'git rebase -p C l'. Is that
> what you meant here?
>
> Before your patch, git would just say "Nothing to do"

Huh.  That is worse than I thought.

> and after your
> patch, we get
>
> # a---b---c
> #      \   \
> #       d   \
> #        \   \
> #         e   \
> #          \   \
> #           C---l
>
> As you say, your patch doesn't try to handle this case, but at least
> the new behavior seems better. I think we would ideally want the
> recreated 'l' to have only 'C' as parent in this case. Does that make
> sense?

This is not what I meant, but it is a very interesting corner case.  I
am not sure I have a solid opinion on what the result should be here.
I feel like it should look the same as you show here, since neither
'c' nor 'C' is a candidate for collapsing during this rebase.  But I may
be missing some subtlety here.


Here is the corner case I was thinking of.  I did not test this to see
if this will happen, but I conceived that it might.  Suppose you have
this tree where

# a---b---c
#      \   
#       d---g---l
#        \ /
#         C

where 'C' introduced the same changes as 'c'.

When I execute 'git rebase -p l c', I expect that I will end up with
this:

# a---b---c---d---
#              \  \
#               ---g---l

That is, 'C' gets skipped because it introduces the same changes already
seen in 'c'.  So 'g' now has two parents: 'd' and 'C^'.  But 'C^' is 'd',
so 'g' now has two parents, both of whom are 'd'.  

I think it should collapse to this instead:

# a---b---c---d---g---l

I don't think this occurs because of my patch, and I am not sure it
occurs at all.  It is something that I considered when I was thinking of
failure scenarios for my patch.

I expect it also may happen if 'C' is an already-empty commit, or if
it is made empty after conflict resolution involving the user. I
mentioned it because I thought my patch _could_ address this if my
is_merge_commit test would also consider whether the parents are
distinct from each other or not.

I hope this is clear, but please let me know if I made it too confusing.

Phil

--
To unsubscribe from this list: send the line "unsubscribe git" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html


[Index of Archives]     [Linux Kernel Development]     [Gcc Help]     [IETF Annouce]     [DCCP]     [Netdev]     [Networking]     [Security]     [V4L]     [Bugtraq]     [Yosemite]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Linux SCSI]     [Fedora Users]