patchGit.pm: Don't return 'undef' in vector context.

Previously, the Git->repository()->config('non-existent.key') evaluated to as true in a vector context. Return an empty list instead. --- I don't know whether this breaks anything, because I don't use most of the git perl scripts. I can't imagine that there is a script that relies on the fact that config('non-existent.key') actually returns (''), in an array context. Is this a reasonable change?

  perl/Git.pm |    2 +-
  1 files changed, 1 insertions(+), 1 deletions(-)

Dan Zwell <dzwell@gmail.com> writes:

I did not examine the callers but my gut feeling is that it would be simpler and cleaner to always return () without checking the context. In scalar context:

	sub null {
        	...
                return ();
	}
	my $scalar = null();

would assign undef to $scalar anyway.

I generally try to stay away from functions that changes their return values depending on the context, because they tend to make reading the callers to find bugs more difficult. An exception is a function that tries to mimic a built-in operator, because reading the callers to such a function, as long as it is clear which built-in the function is imitating, can apply the same knowledge on how the callee would behave you already have by knowing Perl itself.

The same thing can be said about functions with prototypes to force certain context on the caller's side. Avoid it unless there is a good reason.

Maybe it is just me, but that's my reaction.

Junio C Hamano wrote:

That's reasonable. I'll resend this as part of the git-add--interactive color patches. This can be cherry-picked out, but some of the other stuff I want to do depends on it (a helper function that I wrote, config_with_default($repo, $key, $default)).

Dan

Dan Zwell <dzwell@gmail.com> writes:

Sure; as long as the "Don't return 'undef'" remains a separate patch early in the series, that is easy to manage.

Thanks.

Hi,

On Thu, Nov 15, 2007 at 10:39:26PM -0800, Junio C Hamano wrote:

[...]

In fact, if you simply return without any value, it will evaluate to "false" no matter which context it has been called in ("()" in list context, "undef" in scalar context, etc.). So, it's generally a good idea to use "return;" to indicate an error.

Cheers, Sebastian

[Cc: Dan Zwell <dzwell@gmail.com>, Petr Baudis <pasky@suse.cz>,
     git@vger.kernel.org]

Dan Zwell wrote:

Nice.

By the way, what do you think about changing Git.pm config handling to the 'eager' one used currently by gitweb, namely reading all the config to hash, and later getting config values from hash instead of calling git-config? Or at least make it an option?

Jakub Narebski wrote:

That seems appropriate, though it may be a slight trade-off between complexity and efficiency. I don't think it's strictly necessary, at least for git-add--interactive. My experience is that the nine calls to config() (that I am adding) do not slow down the program from a user perspective (though I haven't tested on a slower computer).

The big reason to do it would be if you wanted to convert gitweb to use the standard config() call from Git.pm. Because right now, config() isn't efficient, but it probably doesn't need to be.

Dan

On Saturday, 17 November 2007, Dan Zwell wrote:

On one hand it adds a bit of complexity, as you have to check if
config was parsed (and either parse, as it is done now in gitweb,
or perhaps fallback on calling git-config per variable), and you
have to do type checking / conversion to bool and int (size suffixes!)
in Perl, and deal with single-value and multi-value variables.
 
On the other hand I think that error handling will be simplier.

The problem is not so much with a slower computer, as operating systems with inefficient, slow fork implementation, like MS Windows and (if I remember) correctly MacOS X.

But it is true that the config() performance is needed less for desktop (commands like git-add--interactive) than for web (gitweb for example).

There are two reasons against convering gitweb to use Git.pm, at least for now.

First, it makes installing gitweb harder, as one would have to install (and perhaps compile) Git.pm to use gitweb. It would add additional dependency. But perhaps this is not as much a problem as I think...

Second, gitweb code contain in few places pipelines (e.g. tar.gz and tar.bz2 snapshot pipelines); even if some of them can be eliminated, some will be added (syntax highlighting in blob view). Git.pm doesn't as of yet support this...

It would be nice if Git.pm could take advantage of libification project, and git have fast bindings not only for Python but also for Perl. If I remember correctly the early attempts to use "git library" directly were abandoned becaue they were not sufficiently portable, and the fallback-to-Perl idea was thought too complex to implement...