Re: [PATCH 2/3] rebase: accept -<number> as another way of saying HEAD~<number>

[Jeff King <peff@peff.net>]

On Mon, Mar 03, 2014 at 10:37:11AM +0100, Matthieu Moy wrote:

> Michael Haggerty <mhagger@alum.mit.edu> writes:
> 
> > Or perhaps "-NUM" should fail with an error message if any of the last
> > NUM commits are merges.  In that restricted scenario (which probably
> > accounts for 99% of rebases), "-NUM" is equivalent to "HEAD~NUM".
> 
> Makes sense to me. So, -NUM would actually mean "rebase the last NUM
> commits" (as well as being an alias for HEAD~NUM), but would fail when
> it does not make sense (with an error message explaining the situation
> and pointing the user to HEAD~N if this is what he wanted).

Yeah, I agree with this. I think "-NUM" is only useful for linear
string-of-pearls history. With merges, it either means:

  - linearize history (a la git-log), go back NUM commits, and treat the
    result as the upstream. This is useless, because it's difficult to
    predict where you're going to end up. Using explicit "~" and "^"
    relative-commit operators to specify the upstream is more flexible
    if you really want to do this (but I don't know why you would).

  - do not use an UNINTERESTING upstream as the cutoff, but instead
    try to rebase exactly NUM commits. In the face of non-linear
    history, I'm not even sure what this would mean, or how we would
    implement it.

Neither of those is useful, so I don't think erroring out on them is a
bad thing. And by erroring out (rather than documenting some weird and
useless behavior), we don't have to worry about backwards compatibility
if we later _do_ come up with a useful behavior (the best I can think of
is that "--first-parent -3" might have a use, but I think that should
already be covered, as the results of --first-parent are by-definition
linear).

> This would actually be a feature for me: I often want to rebase "recent
> enough" history, and when my @{upstream} isn't well positionned, I
> randomly type HEAD~N without remembering what N should be. When N is too
> small, the rebase doesn't reach the interesting commit, and when N is
> too big, it reaches a merge commit and I get a bunch of commits I'm not
> allowed to edit in my todo-list. Then I have to abort the commit
> manually. With -N failing on merge commits, the rebase would abort
> itself automatically.

I do the same thing.

-Peff