Re: [PATCH v2] Test t5560: Fix test when run with dash

From: Michael Haggerty <mhagger@alum.mit.edu>
Date: Jan 21, 2010, 16:15 UTC
Message-ID: <4B587DAE.9030208@alum.mit.edu>
In-Reply-To: <7v4ommoo4p.fsf@alter.siamese.dyndns.org>

Junio C Hamano wrote:

Show 17 quoted lines

> Johannes Sixt <j6t@kdbg.org> writes:
>>> Yesterday, I saw rebase--interactive has a few codepaths where "output"
>>> shell function was used with the single-shot export; perhaps they need to
>>> also be fixed.
>> I knew these spots, and they were discussed when that code was introduced. 
>> Before I sent out the mail you were responding to, I tried various ways to 
>> show the failure in rebase--interactive, but it didn't fail...
> 
> It may be the case that the single-shot-ness of these GIT_AUTHOR_NAME
> exports do not matter at all in that program, even though the original
> versions may have been written carefully not to leak the value suitable
> for the current commit to later rounds.
> 
> I think the recent updates from Michael actually depends on the
> distinction not to matter.  For example, do_with_author() in 7756ecf
> (rebase -i: Extract function do_with_author, 2010-01-14) invokes "$@"
> that could be a shell function.

I have to say that I am a little bit over my head here. I didn't try to follow the complete data path of the GIT_AUTHOR_* shell variables, nor do I know exactly what git commands they affect. I just tried to locally refactor the code based on my mistaken assumption that shell functions are treated much like external commands WRT export of shell variables.

The use of the GIT_AUTHOR_* variables in git-rebase--interactive.sh were and are a bit peculiar anyway, since the variables are already set before do_with_author() is invoked, and the values are left to hang around afterwards. The do_with_author() function only tries to export these already-set variables.

So I suppose that the simplest solution is to export these variables explicitly in do_with_author(), something like this (similar to the third code block that was replaced by the do_with_author() function):

do_with_author() {
	export GIT_AUTHOR_NAME GIT_AUTHOR_EMAIL GIT_AUTHOR_DATE &&
	"$@"
}

But to ensure that this is a correct solution would require verification that these now-exported variables don't cause unwanted side-effects during any other external command invocations. Alternatively, I suppose that the variables could be exported within a subshell that also invokes the "$@" command; this subshell could even source the $AUTHOR_SCRIPT file if it were thought advantageous not to set the GIT_AUTHOR_* variables in the git-rebase--interactive.sh script at all.

Help would be most appreciated; I probably won't have time to work on this myself for a week or two.

Michael