Re: [PATCH 0/3] rebase: learn --keep-base

["Ævar Arnfjörð Bjarmason" <avarab@gmail.com>]

On Tue, Mar 26 2019, Denton Liu wrote:

> Hi Ævar,
>
> On Tue, Mar 26, 2019 at 03:35:34PM +0100, Ævar Arnfjörð Bjarmason wrote:
>>
>> On Sat, Mar 23 2019, Denton Liu wrote:
>>
>> > This series teaches rebase the --keep-base option.
>> >
>> > 'git rebase --keep-base <upstream>' is equivalent to
>> > 'git rebase --onto <upstream>... <upstream>' or
>> > 'git rebase --onto $(git merge-base <upstream> HEAD) <upstream>' .
>> >
>> > This seems to be a common case that people (including myself!) run into; I was
>> > able to find these StackOverflow posts about this use case:
>> >
>> > * https://stackoverflow.com/questions/53234798/can-i-rebase-on-a-branchs-fork-point-without-explicitly-specifying-the-parent
>> > * https://stackoverflow.com/questions/41529128/how-do-you-rebase-only-changes-between-two-branches-into-another-branch
>> > * https://stackoverflow.com/a/4207357
>>
>> Like with another series of yours I think this would be best squashed
>> into one patch.
>
> Will do.
>
>>
>> Maybe I've misunderstood this but isn't this like --fork-point except
>> with just plain "git merge-base" instead of "git merge-base
>> --fork-point", but then again 2/3 shows multiple base aren't supported,
>> but merge-base supports that.
>>
>
> --fork-point gets used to determine the _set of_ commits which are to be
> rebased, whereas --keep-base (and --onto) determine the base where that
> set of commits will be spliced. As a result, these two options cover
> orthogonal use-cases.

Right. After playing with this a bit more though --fork-point is mostly
there, it it does find the same fork point, as evidenced all your tests
(that aren't asserting incompatibility with other options) passing with
this:

    diff --git a/t/t3416-rebase-onto-threedots.sh b/t/t3416-rebase-onto-threedots.sh
    index 9c2548423b..ab2d50e69a 100755
    --- a/t/t3416-rebase-onto-threedots.sh
    +++ b/t/t3416-rebase-onto-threedots.sh
    @@ -116,7 +116,7 @@ test_expect_success 'rebase --keep-base master from topic' '
            git checkout topic &&
            git reset --hard G &&

    -       git rebase --keep-base master &&
    +       git rebase $(git merge-base --fork-point master HEAD) &&
            git rev-parse C >base.expect &&
            git merge-base master HEAD >base.actual &&
            test_cmp base.expect base.actual &&
    @@ -140,7 +140,7 @@ test_expect_success 'rebase -i --keep-base master from topic' '
            git reset --hard G &&

            set_fake_editor &&
    -       EXPECT_COUNT=2 git rebase -i --keep-base master &&
    +       EXPECT_COUNT=2 git rebase -i $(git merge-base --fork-point master HEAD) &&
            git rev-parse C >base.expect &&
            git merge-base master HEAD >base.actual &&
            test_cmp base.expect base.actual &&

I've poked at some of this recently in
https://public-inbox.org/git/20190221214059.9195-3-avarab@gmail.com/ as
noted in the feedback there (I haven't gotten around to v2 yet) it's
entirely possible that I haven't understood this at all :)

But it seems to me that this patch/implementation conflates two
unrelated things.

Once is that we use --fork-point to mean that we're going to find the
divergence point with "merge-base --fork-point". This gets you halfway
to where you want to be, i.e. AFAICT the --keep-base and --fork-point
will always find the same commit for "git rebase" and "git rebase
--keep-base". See the "options.restrict_revision = get_fork_point(...)"
part of the code.

The other, which you want to disable, is that --fork-point *also* says
"OK, once we've found the divergence point, let's then rebase it on the
latest upstream. Or in the example above the "master" part of "git
merge-base --fork-point master HEAD".

Shouldn't --keep-base just be implemented in terms of skipping *that*
part, i.e. we find the fork point using the upstream info, but then
don't rebase *on* upstream.

The reason the distinction matters is because with your patch these two
act differently:

    git rebase --keep-base
    git rebase $(git merge-base @{u} HEAD)

The latter will skip work ("Current branch master is up to date"), but
--keep-base will always re-rebase things. There's some cases where
--fork-point does that, which I was trying to address with my linked WIP
patch above.

Whereas the thing you actually want to work is:

    git rebase -i --keep-base
    git rebase -i $(git merge-base @{u} HEAD)

I.e. to have both of those allow you to re-arrange/fixup whatever and
still rebase on the same divergence point with @{u}, and won't run
rebase when there's no work to do unless you give it --force-rebase.

> reason that --onto already disallows multiple bases. If we have multiple
> bases, how do we determine which one is the "true base" to use? It makes
> more sense to error out and let the user manually specify it.

Ah, makes sense.

>> I'd find something like the "DISCUSSION ON FORK-POINT MODE" in
>> git-merge-base helpful with examples of what we'd pick in the various
>> scenarios, and also if whatever commit this picks was something you
>> could have "git merge-base" spew out, so you could get what rebase would
>> do here from other tooling (which maybe is possible, but I'm confused by
>> the "no multiple bases"...).
>
> If I'm understanding you correctly then yes, this could be done with
> other tooling. See the 0/3 for equivalent commands.
>
> Perhaps I should update the rebase documentation to mention that
> --fork-point and --keep-base are orthogonal because it's unclear for
> you, it's probably unclear for other users as well.
>
> Thanks,
>
> Denton