Re: [PATCH v3 13/15] replay: add --advance or 'cherry-pick' mode

From: Christian Couder <christian.couder@gmail.com>
Date: Sep 7, 2023, 08:35 UTC
Message-ID: <CAP8UFD0T64UcDLUrbjrR0DvD4DQLuZ==ATk9ewSMkMn+9VhMAg@mail.gmail.com>
In-Reply-To: <87h6qzst8u.fsf@iotcl.com>

On Thu, Jun 22, 2023 at 12:09 PM Toon Claes <toon@iotcl.com> wrote:

Show 13 quoted lines

>
>
> Christian Couder <christian.couder@gmail.com> writes:
>
> > +     /*
> > +      * When the user specifies e.g.
> > +      *   git replay origin/main..mybranch
> > +      *   git replay ^origin/next mybranch1 mybranch2
>
> When I'm trying these, I'm getting the error:
>     error: option --onto or --advance is mandatory
>
> In what situation can I omit both --onto and --advance?

It was possible with version 1 of this series as one of the patches allowed the command to guess the base:

https://lore.kernel.org/git/20230407072415.1360068-13-christian.couder@gmail.com/

so --onto wasn't needed to specify it.

Comments on that patch said that it might be better to focus on a plumbing command first and for that the patch wasn't needed, so I removed it in version 2.

Show 6 quoted lines

> > +static void determine_replay_mode(struct rev_cmdline_info *cmd_info,
> > +                               const char *onto_name,
> > +                               const char **advance_name,
> > +                               struct commit **onto,
>
> Would it make sense to call this target?

This is more the new base we are rebasing target branches onto. So if we want to change the name, `--base` or `--newbase` would make more sense. `git rebase` already has `--onto` though, so, if we want to be consistent with it, we should keep `--onto` for this.

Show 14 quoted lines

> > +                               struct strset **update_refs)
> > +{
> > +     struct ref_info rinfo;
> > +
> > +     get_ref_information(cmd_info, &rinfo);
> > +     if (!rinfo.positive_refexprs)
> > +             die(_("need some commits to replay"));
> > +     if (onto_name && *advance_name)
> > +             die(_("--onto and --advance are incompatible"));
>
> Do we actually need to disallow this? I mean from git-replay's point of
> view, there's no technical limitation why can cannot support both modes
> at once. The update-ref commands in the output will update both target
> and source branches, but it's not up to us whether that's desired.

I am not sure what you call "target" and "source" branches. Anyway here is in simple terms the way the command works:

  1) it takes either `--onto <newbase>` or `--advance <branch>` and
then one or more <revision-range>,
  2) it replays all the commits in the <revision-range> onto either
<newbase> or <branch>,
  3) in case of `--advance`, it outputs a single command for `git
update-ref --stdin` to advance <branch> to the last commit that was
just replayed,
  4) in case of `--onto`, it outputs a number of commands for `git
update-ref --stdin` to update the branches in <revision-range> to
where the tip commits of these branches have been replayed.

So `--advance` is like a cherry-pick, and `--onto` is like a rebase.

It would be possible to do both a rebase onto a branch and a cherry-pick of the rebased commits onto that branch, but this is not very common and you can achieve the same result by just rebasing and then using `git reset` or `git update-ref` to make the branch point to the result of the rebase. So I don't see the point of complicating the command at this point.

Show 15 quoted lines

> > +     else if (onto_name) {
>
> No need to 'else' here IMHO.
>
> > +             *onto = peel_committish(onto_name);
> > +             if (rinfo.positive_refexprs <
> > +                 strset_get_size(&rinfo.positive_refs))
> > +                     die(_("all positive revisions given must be references"));
>
> I tested this locally with the following command:
>
> $ git replay --onto main OID..OID
>
> This command didn't give any errors, neither did it return any
> update-ref lines. I would have expected to hit this die().

Yeah, this might be unexpected.

I tested it too and 'rinfo.positive_refexprs' is 1 while 'strset_get_size(&rinfo.positive_refs)' is 0 with such a command.

The result does not look wrong though. Above that code there is:

    if (!rinfo.positive_refexprs)
        die(_("need some commits to replay"));

so it looks like there is at least a check that the revision range passed to the command contains positive commits.

It might be possible that users prefer a command that outputs nothing when there is nothing to replay instead of erroring out.

Show 17 quoted lines

> > +     } else if (*advance_name) {
> > +             struct object_id oid;
> > +             char *fullname = NULL;
> > +
> > +             *onto = peel_committish(*advance_name);
> > +             if (repo_dwim_ref(the_repository, *advance_name, strlen(*advance_name),
> > +                          &oid, &fullname, 0) == 1) {
> > +                     *advance_name = fullname;
> > +             } else {
> > +                     die(_("argument to --advance must be a reference"));
> > +             }
> > +             if (rinfo.positive_refexprs > 1)
> > +                     die(_("cannot advance target with multiple source branches because ordering would be ill-defined"));
>
> The sources aren't always branches, so I suggest something like:
>
> +                       die(_("cannot advance target with multiple sources because ordering would be ill-defined"));

Yeah, that looks reasonable. I have made this change in version 4 I will send very soon.

Show 14 quoted lines

> > +     determine_replay_mode(&revs.cmdline, onto_name, &advance_name,
> > +                           &onto, &update_refs);
> > +
> > +     if (!onto) /* FIXME: Should handle replaying down to root commit */
> > +             die("Replaying down to root commit is not supported yet!");
>
> When I was testing locally I tried the following:
>
> $ git replay --onto main feature
>
> I was expecting this command to find the common ancestor automatically,
> but instead I got this error. I'm fine if for now the command does not
> determine the common ancestor yet, but I think we should provide a
> better error for this scenario.

I agree that it isn't very user friendly. We could indeed try to find if there is a common ancestor, and, if that's the case, suggest another way to call the command. This is a plumbing command in its early stage though for now. So I guess it's Ok to postpone working on nicer error messages.

Show 9 quoted lines

> > +test_expect_success 'using replay on bare repo to perform basic cherry-pick' '
> > +     git -C bare replay --advance main topic1..topic2 >result-bare &&
> > +     test_cmp expect result-bare
> > +'
> > +
> >  test_done
>
> Shall we add a test case when providing both --onto and --advance? And
> one that omits both?

Ok, I have made this change in version 4.