On Sat, Jul 25, 2015 at 2:09 AM, Jeff King <peff@xxxxxxxx> wrote:
> Yeah, I think this hunk is to blame (though I just read the code and did not
> test):
>
> @@ -658,6 +665,8 @@ fi
> if test "$(cat "$dotest/threeway")" = t
> then
> threeway=t
> +else
> + threeway=f
> fi
>
> It comes after the command-line option parsing, so it overrides our option (I
> think that running "git am -3" followed by "git am --no-3way" would have the
> same problem). It cannot just check whether $threeway is unset, though, as it
> may have come from the config.
Thanks for the detailed analysis, I completely agree. Note that the
code that handles the --message-id option somewhat handles the case
where $messageid is unset:
case "$(cat "$dotest/messageid")" in
t)
messageid=-m ;;
f)
messageid= ;;
esac
However, it still does not handle "git am --no-message-id" followed by
"git am --message-id", or "git -c am.messageid=true am" followed by
"git am --no-message-id". I think the same thing occurs for
--scissors/--no-scissors, as well as the git-apply options as well.
The real problem is that the state directory loading code comes after
the config loading and option parsing code, and thus overrides any
variables set.
> We'd need a separate variable, the way the code
> is ordered now.
If we are just fixing --3way, adding one extra variable won't be that
bad. However, I think that if we are using this approach to fix all of
the options, then it would introduce too much code complexity.
> Ideally the code would just be ordered as:
>
> - load config from git-config
>
> - override that with defaults inherited from a previous run
>
> - override that with command-line parsing
So I'm more in favor of this solution. It's feels much more natural to
me, rather than attempting to workaround the existing code structure.
> but I don't know if there are other ordering gotchas that would break.
For the C code, there won't be any problem, but yeah, fixing it in
git-am.sh might need a bit more effort.
> It does look like that is how Paul's builtin/am.c does it, which makes
> me think it might not be broken. It's also possibly I've horribly
> misdiagnosed the bug. ;)
Nah, it follows the same structure as git-am.sh and so will exhibit
the same behavior. It currently does something like this:
1. am_state_init() (config settings are loaded)
2. parse_options()
3. if (am_in_progress()) am_load(); else am_setup();
So it would be quite trivial to change the control flow such that it is:
1. am_state_init()
2. if (am_in_progress()) am_load()
3. parse_options();
4 if (!am_in_progress()) am_setup()
The next question is, should any options set on the command-line
affect subsequent invocations? If yes, then the control flow will be
like:
1. am_state_init();
2. if (am_in_progress()) am_load();
3. parse_options();
4. if (am_in_progress()) am_save_opts(); else am_setup();
where am_save_opts() will write the updated variables back to the
state directory. What do you think?
Since the builtin-am series is in 'next' already, and the fix in C is
straightforward, to save time and effort I'm wondering if we could
just do "am.threeWay patch -> builtin-am series -> bugfix patch in C".
My university term is starting soon so I may not have so much time,
but I'll see what I can do :-/
Junio, how do you want to proceed?
Thanks,
Paul
--
To unsubscribe from this list: send the line "unsubscribe git" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html