Re: [PATCH v2] sequencer: do not squash 'reword' commits when we hit conflicts

[Johannes Schindelin <Johannes.Schindelin@gmx.de>]

Hi Phillip,

On Mon, 18 Jun 2018, Phillip Wood wrote:

> On 17/06/18 20:28, Johannes Schindelin wrote:
> > 
> > On Sun, 17 Jun 2018, Phillip Wood wrote:
> > 
> >> On 17/06/18 06:37, Elijah Newren wrote:
> >>> Ever since commit 18633e1a22 ("rebase -i: use the rebase--helper
> >>> builtin", 2017-02-09), when a commit marked as 'reword' in an
> >>> interactive rebase has conflicts and fails to apply, when the rebase
> >>> is resumed that commit will be squashed into its parent with its
> >>> commit message taken.
> >>>
> >>> The issue can be understood better by looking at commit 56dc3ab04b
> >>> ("sequencer (rebase -i): implement the 'edit' command", 2017-01-02),
> >>> which introduced error_with_patch() for the edit command.  For the
> >>> edit command, it needs to stop the rebase whether or not the patch
> >>> applies cleanly.  If the patch does apply cleanly, then when it
> >>> resumes it knows it needs to amend all changes into the previous
> >>> commit.  If it does not apply cleanly, then the changes should not
> >>> be amended.  Thus, it passes !res (success of applying the 'edit'
> >>> commit) to error_with_patch() for the to_amend flag.
> >>>
> >>> The problematic line of code actually came from commit 04efc8b57c
> >>> ("sequencer (rebase -i): implement the 'reword' command", 2017-01-02).
> >>> Note that to get to this point in the code:
> >>>    * !!res (i.e. patch application failed)
> >>>    * item->command < TODO_SQUASH
> >>>    * item->command != TODO_EDIT
> >>>    * !is_fixup(item->command) [i.e. not squash or fixup]
> >>> So that means this can only be a failed patch application that is
> >>> either a pick, revert, or reword.  For any of those cases we want a
> >>> new commit, so we should not set the to_amend flag.
> >>
> >> Unfortunately I'm not sure it's that simple. Looking and do_pick()
> >> sometimes reword amends HEAD and sometimes it does not. In the
> >> "normal" case then the commit is picked and committed with '--edit'.
> >> However when fast-forwarding the code fast forwards to the commit to
> >> be reworded and then amends it. If the root commit is being reworded
> >> it takes the same code path. While these cases cannot fail with
> >> conflicts, it is possible for the user to cancel the commit or for
> >> them to fail due to collisions with untracked files.
> >>
> >> If I remember correctly the shell version always picks the commit and
> >> then calls 'git commit --amend' afterwards which is less efficient
> >> but consistent.
> >>
> >> I'm afraid I don't have a simple solution for fixing this, as
> >> currently pick_commits() does not know if the commit was called with
> >> AMEND_MSG, I guess that means adding some kind of flag for do_pick()
> >> to set.
> > 
> > Oh, you're right, the fast-forwarding path would pose a problem. I
> > think there is an easy way to resolve this, though: in the case that
> > we do want to amend the to-be-reworded commit, we simply have to see
> > whether HEAD points to the very same commit mentioned in the `reword`
> > command:
> 
> That's clever, I think to get it to work for rewording the root commit,
> it will need to do something like comparing HEAD to squash-onto as well.

... because squash-onto is a fresh, empty root commit (to be "amended"
when a non-root commit is to be picked as a new root commit). Good point.

> > -- snip --
> > diff --git a/sequencer.c b/sequencer.c
> > index 2dad7041960..99d33d4e063 100644
> > --- a/sequencer.c
> > +++ b/sequencer.c
> > @@ -3691,10 +3691,22 @@ static int pick_commits(struct todo_list
> > *todo_list, struct replay_opts *opts)
> >                                         intend_to_amend();
> >                                 return error_failed_squash(item->commit, opts,
> >                                         item->arg_len, item->arg);
> > -                       } else if (res && is_rebase_i(opts) && item->commit)
> > +                       } else if (res && is_rebase_i(opts) && item->commit) {
> > +                               int to_amend = 0;
> > +
> > +                               if (item->command == TODO_REWORD) {
> > +                                       struct object_id head;
> > +
> > +                                       if (!get_oid("HEAD", &head) &&
> > +					    !oidcmp(&item->commit->object.oid,
> > +                                                   &head))
> > +                                               to_amend = 1;

This would now become

					if (!get_oid("HEAD", &head) &&
					    (!oidcmp(&item->commit->object.oid,
						     &head) ||
					     (opts->have_squash_onto &&
					      !oidcmp(&opts->squash_onto,
						      &head))))
						to_amend = 1;

This is awfully indented, so a better idea would probably be to avoid the
extra block just to declare `head`:


-                       } else if (res && is_rebase_i(opts) && item->commit)
+                       } else if (res && is_rebase_i(opts) && item->commit) {
+                               int to_amend = 0;
+                               struct object_id oid;
+
+				/*
+				 * If we fast-forwarded already, or if we
+				 * are about to create a new root commit,
+				 * we definitely want to amend, otherwise
+				 * we do not.
+				 */
+                               if (item->command == TODO_REWORD &&
+				    !get_oid("HEAD", &oid) &&
+				    (!oidcmp(&item->commit->object.oid, &oid) ||
+				     (opts->have_squash_onto &&
+				      !oidcmp(&opts->squash_onto, &head))))
+					to_amend = 1;
+
                                return res | error_with_patch(item->commit,
                                        item->arg, item->arg_len, opts, res,
-                                       item->command == TODO_REWORD);
+                                       to_amend);
+                       }
                } else if (item->command == TODO_EXEC) {


What do you think?

And: could you perchance add a regression test for a failing pick onto
squash-onto (I am desperately in need of sleeping, otherwise I would do it
myself)? Something that executes `reset [new root]` and then `pick abcdef`
where abcdef would overwrite an untracked file should trigger this
relatively easily.

Thanks,
Dscho