Junio C Hamano <gitster@xxxxxxxxx> 于2021年7月28日周三 上午5:00写道：
>
> Junio C Hamano <gitster@xxxxxxxxx> writes:
>
> > Phillip Wood <phillip.wood123@xxxxxxxxx> writes:
> >
> >> This will break git-rebase--preserve-merges.sh which uses
> >> GIT_CHERRY_PICK_HELP to set the help and ensure CHERRY_PICK_HEAD is
> >> removed when picking commits.
> >
> > Ahh, I didn't realize we still had scripted rebase backends that
> > called cherry-pick as an executable. I was hoping that all rebase
> > backends by now would be calling into the cherry-pick machinery
> > directly, bypassing cmd_cherry_pick(), and that was why I suggested
> > to catch stray one the end-users set manually in the environment
> > and clear it there.
> >
> >> I'm a bit confused as to what the
> >> problem is - how is 'git cherry-pick' being run with
> >> GIT_CHERRY_PICK_HELP set in the environment outside of a rebase (your
> >> explanation in [1] does not mention how GIT_CHERRY_PICK_HELP is set)?
> >
> > I didn't press for the information too hard, but I guessed that it
> > was perhaps because somebody like stackoverflow suggested to set a
> > message in their environment to get a "better message."
>
> A good way forward may be to relieve sequencer.c::print_advice() of
> the responsibility of optinally removing CHERRY_PICK_HEAD; make it a
> separate function that bases its decision on a more direct cue, not
> on the presense of a custom message in GIT_CHERRY_PICK_HELP, make
> do_pick_commit(), which is the sole caller of print_advice(), call
> it after calling print_advice().
>
I think this function "check_need_delete_cherry_pick_head()" should be before
print_advice(), like this:
+ const char *help_msgs = NULL;
+
error(command == TODO_REVERT
? _("could not revert %s... %s")
: _("could not apply %s... %s"),
short_commit_name(commit), msg.subject);
- print_advice(r, res == 1, opts);
+ if (((opts->action == REPLAY_PICK &&
+ !opts->rebase_preserve_merges_mode) ||
+ (help_msgs = check_need_delete_cherry_pick_head(r))) &&
+ res == 1)
+ print_advice(opts, help_msgs);
> I do not offhand know what that "direct cue" should be, but we may
> already have an appropriate field in the replay_opts structure;
> "replay.action is neither REVERT nor PICK" could be a good enough
> approximation, I dunno.
>
> Otherwise we can allocate a new bit in the structure, have relevant
> callers set it, and teach cherry-pick an unadvertised command line
> option that sets the bit, and use that option only from
> git-rebase--preserve-merges when it makes a call to cherry-pick.
> When "rebase -p" is either retired or rewritten in C, we can retire
> the option from cherry-pick.
>
I think this one can be easily achieved.
> Workable?
Thanks.
--
ZheNing Hu
