Re: [BUG] git cherry-pick does not complain about unknown options
On Mon, Jul 09, 2018 at 04:16:16PM +0200, Andrei Rybak wrote:
> I was trying to cherry pick commits, while simultaneously changing the
> author. Unfortunately, cherry-pick doesn't have the same --author
> option as git-commit. However, instead of complaining about unknown
> option:
Yeah, its "--author" option is interpreted by revision.c similar to "log --author" (because we just pass all unknown options to the traversal machinery). So you could say:
git cherry-pick --author=peff v1.0..v2.0
to pick all of my commits between the two versions. I'm not sure if that would ever be _useful_, but that's how it has behaved since 7e2bfd3f99 (revert: allow cherry-picking more than one commit, 2010-06-02), I think.
I agree that having something similar to commit's "--author" (or even just "--reset-author") would be useful. When I've had to do this before, I usually just cherry-pick and then follow-up with "commit --amend --author" (or use "rebase -i" if there are several commits). You can also do "cherry-pick -n $commit" followed by "commit -c $commit".
Show 7 quoted lines
> All commits in tests existed in repository:
>
> $ git cherry-pick --author='TEST' <one-commit> # case 1
> error: BUG: expected exactly one commit from walk
> fatal: cherry-pick failed
> $ echo $?
> 128
I think it's reasonable for this to issue an error. But a BUG is definitely wrong. The error is more like "empty commit set passed", which is what we get for something like:
$ git cherry-pick HEAD..HEAD
error: empty commit set passed
fatal: cherry-pick failed
that triggers the revision walker but has no positive commits. But we don't trigger that code because the single-rev case has some magic to allow a cherry-pick during another sequencer operation.
So I think that BUG message needs to be softened.
> $ git cherry-pick --author='TEST' <commit1> <commit2> # case 2
> $ echo $?
> 0
And here we come up with an empty set, but we don't flag it as an error at all. I could even accept that philosophically as "empty input, nothing to do", but we clearly do flag it as an error in the case I showed above.
I think the problem is that we catch the error in sequencer.c:prepare_revs(), after prepare_revision_walk() returns no commits. But we _do_ have a starting commit for this case, and it's not culled until we actually ask get_revisions() to start walking.
I think we'd want something like this:
diff --git a/sequencer.c b/sequencer.c
index 4034c0461b..b978157399 100644
--- a/sequencer.c
+++ b/sequencer.c
@@ -1860,8 +1860,6 @@ static int prepare_revs(struct replay_opts *opts)
if (prepare_revision_walk(opts->revs))
return error(_("revision walk setup failed"));
- if (!opts->revs->commits)
- return error(_("empty commit set passed"));
return 0;
}
@@ -2314,6 +2312,10 @@ static int walk_revs_populate_todo(struct todo_list *todo_list,
short_commit_name(commit), subject_len, subject);
unuse_commit_buffer(commit, commit_buffer);
}
+
+ if (!todo_list->nr)
+ return error(_("empty commit set passed"));
+
return 0;
}
-Peff
Andrei Rybak
