Re: Git Evolve

[Taylor Blau <me@ttaylorr.com>]

On Tue, Oct 02, 2018 at 11:11:11AM +0200, Ævar Arnfjörð Bjarmason wrote:

You timed this email quite well ;-).

> On Tue, Oct 02 2018, Taylor Blau wrote:
>
> > Hi Stefan,
> >
> > On Sat, Sep 29, 2018 at 04:00:04PM -0700, Stefan Xenos wrote:
> >> Hello, List!
> >>
> >> I'm interested in porting something like Mercurial's evolve command to
> >> Git.
> >
> > Welcome to Git :-). I think that the discussion in this thread is good,
> > but it's not why I'm replying. I have also wanted a Mercurial feature in
> > Git, but a different one than yours.
> >
> > Specifically, I've wanted the 'hg absorb' command. My understanding of
> > the commands functionality is that it builds a sort of flamegraph-esque
> > view of the blame, and then cascades downwards parts of a change. I am
> > sure that I'm not doing the command justice, so I'll defer to [1] where
> > it is explained in more detail.
> >
> > The benefit of this command is that it gives you a way to--without
> > ambiguity--absorb changes into earlier commits, and in fact, the
> > earliest commit that they make sense to belong to.
> >
> > This would simplify my workflow greatly when re-rolling patches, as I
> > often want to rewrite a part of an earlier commit. This is certainly
> > possible by a number of different `git rebase` invocations (e.g., (1)
> > create fixup commits, and then re-order them, or (2) mark points in your
> > history as 'edit', and rewrite them in a detached state, and I'm sure
> > many more).
> >
> > I'm curious if you or anyone else has thought about how this might work
> > in Git.
>
> I've wanted a "git absorb" for a while, but have done no actual work on
> it, I just found out about it.
>
> I think a combination of these two heuristics would probably do the
> trick:
>
>  1. If a change in your "git diff" output has a hunk whose lines overlap
>     with an earlier commit in the @{u}.. range, we do the equivalent of
>     "git add -p", select that hunk, and "git commit --fixup <that
>     commit>". We fixup the most recent commit that matches (otherwise
>     commit>we'd conflict).

I had imagined this working slightly differently. I think about it in
terms of a flamegraph-shape, where each line is affected by gravity.
Consider this:

       [---------------]
  L0   L1   L2   L3   L4   L5   L6   L7

Here's a line in a diff that affects L1-L4. Were we to create a
``fixup'' to L3-L4, it would look like this:

                 [-----] --|
       [---------|-----] <-|
  L0   L1   L2   L3   L4   L5   L6   L7

The commit owning the adjacent edit hunk is the one that gets the
changes applied to it.

Consider instead the case where we have two overlapping parts of a
change:

                 [-----|----]
       [---------|-----]
  L0   L1   L2   L3   L4   L5   L6   L7

The left-hand side of the top-most hunk belongs to the hunk below it,
but the right-hand side is ``affected by gravity'' down to the base:

                 [-----|
       [---------|-----]|---]
  L0   L1   L2   L3   L4   L5   L6   L7

And thus could be reapplied anywhere. I have not spent time proving this,
but I believe that this resolves appropriate bases without ambiguity
(or, at least can detect when finding a base introduces ambiguity).

>  2. Have some mode where we fall back from #1 and consider changes to
>     entire files, if that's unambiguous.
>
> The neat thing about this would be that you could tweak how promiscuous
> #1 would be via the -U option to git-diff, and #2 would just be a
> special case of -U9999999999999 (we should really add a -Uinf...).
>
> Then once you ran this you could run "git rebase -i --autosquash" to see
> how the TODO list would look, and optionally have some "git absorb
> --now" or whatever to do the "git add -p", "git commit --fixup" and "git
> rebase --autosquash" all in one go.

That's cool. I hadn't known about '--autosquash' before, but it now
makes sense to me why I have often seen patches that begin with
"fixup!".

Another thought that I am reminded of was an off-list discussion with
Peff, where we thought that a particularly Git-like way to implement
something like this would be to generate a set of commits that would be
immediately '--autosquash'-able.

> > [1]: http://files.lihdd.net/hgabsorb-note.pdf

Thanks,
Taylor