Re: "git am" and then "git am -3" regression?

From: Paul Tan <pyokagan@gmail.com>
Date: Jul 26, 2015, 05:03 UTC
Message-ID: <CACRoPnR=DSETucY78Xo0RNxHKkqDnTCYFvHsSzWAG7X7z3_DKQ@mail.gmail.com>
In-Reply-To: <20150724180921.GA17730@peff.net>

On Sat, Jul 25, 2015 at 2:09 AM, Jeff King <peff@peff.net> wrote:

Show 15 quoted lines

> 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.

Show 7 quoted lines

> 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