git commit file completion recently broke

git commit file completion recently broke

[Jacob Keller]

Hi,

I'm still investigating, but thought I'd send an email. I recently
updated to jch branch, and found that completion for git commit does
not work as expected.

If I have a git repository with a modified file in a subdirectiory,
then git commit <TAB> produces the name of the subdirectory instead of
the file names.

This occurs regardless of where I run the git commit command.

Thanks,
Jake

Re: git commit file completion recently broke

[Jacob Keller]

On Wed, Dec 6, 2017 at 3:53 PM, Jacob Keller <jacob.keller@gmail.com> wrote:
> Hi,
>
> I'm still investigating, but thought I'd send an email. I recently
> updated to jch branch, and found that completion for git commit does
> not work as expected.
>
> If I have a git repository with a modified file in a subdirectiory,
> then git commit <TAB> produces the name of the subdirectory instead of
> the file names.
>
> This occurs regardless of where I run the git commit command.
>
> Thanks,
> Jake

I think I narrowed this down to "git diff-index --name-only --relative
HEAD" producing a list of files *not* relative to the current
directory.

Thanks,
Jake

Re: git commit file completion recently broke

[Jacob Keller]

On Wed, Dec 6, 2017 at 4:01 PM, Jacob Keller <jacob.keller@gmail.com> wrote:
> On Wed, Dec 6, 2017 at 3:53 PM, Jacob Keller <jacob.keller@gmail.com> wrote:
>> Hi,
>>
>> I'm still investigating, but thought I'd send an email. I recently
>> updated to jch branch, and found that completion for git commit does
>> not work as expected.
>>
>> If I have a git repository with a modified file in a subdirectiory,
>> then git commit <TAB> produces the name of the subdirectory instead of
>> the file names.
>>
>> This occurs regardless of where I run the git commit command.
>>
>> Thanks,
>> Jake
>
> I think I narrowed this down to "git diff-index --name-only --relative
> HEAD" producing a list of files *not* relative to the current
> directory.
>
> Thanks,
> Jake

I've started a git bisect to see if i can find the source of the problem.

Thanks,
Jake

Re: git commit file completion recently broke

[Jeff King]

On Wed, Dec 06, 2017 at 04:01:51PM -0800, Jacob Keller wrote:

> I think I narrowed this down to "git diff-index --name-only --relative
> HEAD" producing a list of files *not* relative to the current
> directory.

Hmm, my guess would have been something funny in the setup code
forgetting our original prefix.

But nope, it looks like the culprit is f7923a5ece (diff: use
skip_to_optional_val(), 2017-12-04), which switched over parsing of
"--relative".

-Peff

Re: git commit file completion recently broke

[Jeff King]

On Wed, Dec 06, 2017 at 07:22:35PM -0500, Jeff King wrote:

> On Wed, Dec 06, 2017 at 04:01:51PM -0800, Jacob Keller wrote:
> 
> > I think I narrowed this down to "git diff-index --name-only --relative
> > HEAD" producing a list of files *not* relative to the current
> > directory.
> 
> Hmm, my guess would have been something funny in the setup code
> forgetting our original prefix.
> 
> 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

Re: git commit file completion recently broke

[Jacob Keller]

On Wed, Dec 6, 2017 at 4:24 PM, Jeff King <peff@peff.net> wrote:
> On Wed, Dec 06, 2017 at 07:22:35PM -0500, Jeff King wrote:
>
>> On Wed, Dec 06, 2017 at 04:01:51PM -0800, Jacob Keller wrote:
>>
>> > I think I narrowed this down to "git diff-index --name-only --relative
>> > HEAD" producing a list of files *not* relative to the current
>> > directory.
>>
>> Hmm, my guess would have been something funny in the setup code
>> forgetting our original prefix.
>>
>> 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.

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'll look into that.

Thanks,
Jake

Re: git commit file completion recently broke

[Jeff King]

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

Re: git commit file completion recently broke

[Jacob Keller]

On Wed, Dec 6, 2017 at 4:56 PM, Jeff King <peff@peff.net> wrote:
> 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:
>

That obviously won't work for any case which sues
skip_to_optional_val_default() (since these provide a default value to
give in case none is provided.

>   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))
>

It obviously wouldn't. The only sensible solution is to have
"skip_to_optional_val_something()" which does this new behavior.

Or, change skip_to_optional_val() behave this new way, but
skip_to_optional_val_default() behave in the current way.

> 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

What you outlined above is probably the best we can do. We could even
add some extra helper which does that for us if we want.

I sent a patch that merely reverts the change to --relative and adds a
test for now though.

Thanks,
Jake

Re: git commit file completion recently broke

[Jeff King]

On Wed, Dec 06, 2017 at 05:04:02PM -0800, Jacob Keller wrote:

> What you outlined above is probably the best we can do. We could even
> add some extra helper which does that for us if we want.
> 
> I sent a patch that merely reverts the change to --relative and adds a
> test for now though.

Thanks. I'm fine with reverting --relative to its original form, or
having it do the NULL thing I mentioned.  Christian can decide, since
it's his series.

I do think we may want to take the test improvement as a separate patch.

-Peff

Re: git commit file completion recently broke

[Christian Couder]

On Thu, Dec 7, 2017 at 1:56 AM, Jeff King <peff@peff.net> wrote:

> 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;
>   }

Yeah, that is a possible fix.

> IOW, the problem isn't in the design of the skip function, but just how
> it was used in this particular case.

I agree.

> 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=").

I discussed that with Junio and yeah there are many callers that want
"--foo" to be the same as "--foo=".

By the way I wonder if "--relative=" makes any sense, and if we should
emit a warning or just die in this case.

Re: git commit file completion recently broke

[Jeff King]

On Thu, Dec 07, 2017 at 09:14:31AM +0100, Christian Couder wrote:

> > 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=").

Oof, I lost all ability to type in that last sentence. :)

It looks like you deciphered my meaning, though.

> I discussed that with Junio and yeah there are many callers that want
> "--foo" to be the same as "--foo=".

Yeah, if this is the only case, then just calling the "_default" variant
with NULL for this instance makes sense to me.

> By the way I wonder if "--relative=" makes any sense, and if we should
> emit a warning or just die in this case.

I also wondered about that. It does function as "relative to the root of
the tree". But of course if you want that you can just omit
"--relative". Still, it could be a minor convenience for somebody who is
filling in "--relative" in a script. That's reaching, I think, but I
don't see any particular reason to _forbid_ it (especially since it has
worked for many years).

-Peff

Re: git commit file completion recently broke

[Junio C Hamano]

Christian Couder <christian.couder@gmail.com> writes:

>> 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=").
>
> I discussed that with Junio and yeah there are many callers that want
> "--foo" to be the same as "--foo=".

Yup, the original thread has details and me saying that assuming all
of them want --foo and --foo= the same is questionable.  The likely
fix would be to use the _default variant with NULL, which was added
exactly for cases like this.

Re: git commit file completion recently broke

[Jacob Keller]

On Thu, Dec 7, 2017 at 7:24 AM, Junio C Hamano <gitster@pobox.com> wrote:
> Christian Couder <christian.couder@gmail.com> writes:
>
>>> 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=").
>>
>> I discussed that with Junio and yeah there are many callers that want
>> "--foo" to be the same as "--foo=".
>
> Yup, the original thread has details and me saying that assuming all
> of them want --foo and --foo= the same is questionable.  The likely
> fix would be to use the _default variant with NULL, which was added
> exactly for cases like this.
>

Slightly more complex. You have to use the _default variant, pass in
arg instead of options->prefix, and then make sure arg was set before
overwriting options->prefix. If you just use _default with NULL, it
will not quite fix the problem.

Thanks,
Jake