Re: git commit file completion recently broke

[Jeff King <peff@peff.net>]

On Wed, Dec 06, 2017 at 04:38:29PM -0800, Jacob Keller wrote:

> >> But nope, it looks like the culprit is f7923a5ece (diff: use
> >> skip_to_optional_val(), 2017-12-04), which switched over parsing of
> >> "--relative".
> >
> > Oh, actually, I guess I was half-right. It feeds &options->prefix as the
> > "default", meaning that we overwrite it with the empty string. I don't
> > think "--relative" works for the semantics of skip_to_optional_value,
> > since it needs:
> >
> >   --relative=foo: set prefix to "foo"
> >
> >   --relative: leave prefix untouched
> >
> > -Peff
> 
> Yep, and apparently our test suite completely lacked any tests of
> --relative on its own.
> 
> I've sent a patch to add some tests.

Great. I was also saddened by our lack of tests.

> I don't know the exact best way to fix this, I guess we could just
> revert it the changes to relative... but maybe we could add or modify
> the semantics of skip_to_optional_val()?? What if it was changed so
> that it left the value alone if no value was provided? This would
> require callers to pre-set the value they want as default, but that
> would solve relative's problem.

I think that would work for this case. But just looking at others from
the same series, I think they'd get pretty awkward. For instance we now
have:

  else if (!strcmp(arg, "--color))
	options->use_color = 1;
  else if (skip_prefix(arg, "--color=", &arg))
	/* parse "arg" as colorbool */

which became:

  else if (skip_to_optional_val_default(arg, "--color", &arg, "always"))
	/* parse "arg" as colorbool */

How would that look with the "leave it alone instead of assigning a
default" semantics? It gets pretty clumsy, because you have to
pre-assign "always" to some pointer. But then we can't reuse "arg", so
we end up with something more like:

  const char *color_val = "always";
  ...
  else if (skip_to_optional_val(arg, "--color", &color_val))

But we need one such "color_val" for every option we test for, and we
have to set all of them up before any matches (because we don't know
which one we'll actually match). Yuck.

I think we'd do better to just assign NULL when there's "=", so we can
tell the difference between "--relative", "--relative=", and
"--relative=foo" (all of which are distinct).

I think that's possible with the current scheme by doing:

  else if (skip_to_optional_val_default(arg, "--relative", &arg, NULL)) {
	options->flags.relative_name = 1;
	if (arg)
		options->prefix = arg;
  }

IOW, the problem isn't in the design of the skip function, but just how
it was used in this particular case. I do think it may make sense for
the "short" one to use NULL, like:

  skip_to_optional_val(arg, "--relative, &arg)

but maybe some other callers would be more inconvenienced (they may have
to current NULL back into the empty string if they want to string
"--foo" the same as "--foo=").

-Peff