Re: t7800-difftool.sh failure on pu

Hi Ramsay - thank you for running the tests and trying this out.


On Thu, Mar 29, 2012 at 2:12 PM, Ramsay Jones
[off-list ref] wrote:

With the current pu branch, I have t7800.3 (difftool ignores bad --tool values)
failing on Linux (I haven't tried cygwin or mingw yet). The failure is caused
by the test for the value of the exit code; for me the exit code is 9 not 1.

This is interesting.  On my Ubuntu box, I am able to run all of t7800
without error.  This has been my primary development platform and the
tests have consistently passed for me on that system.

I have had trouble getting the test suite to run on msysgit.  However,
I just tried the simple test sequence that you demonstrated and found
that I get the same failure on that platform (i.e. "9 Bad file
descriptor").  So it appears that Carp is broken on some platforms.

I have investigated, briefly, and found *two* alternatives for a fix. ;-)

The first option is to (effectively) revert commit 0440ed72 ("difftool: replace
system call with Git::command_noisy", 22-03-2012), like so:

Thanks again for spending the time to dig into the problem.  Based on
feedback from Junio and David in later emails, I plan to revert commit
0440ed72.

The second option is a bit of a mystery, since I don't see why it is necessary
or why it works! :-P

First take a look at the following:

$ perl -e 'print $!+0, " $!\n";'
0
$ echo $?
0

$ perl -e 'use Carp qw(croak); print $!+0, " $!\n";'
9 Bad file descriptor
$ echo $?
0

I agree that this is mysterious.  On my Ubuntu box, this Perl
one-liner prints and exits with 0 for me (i.e. no bad file descriptor
error).  However, I was able to replicate the failure on msysgit.
This makes me wonder if other users of Git.pm have had the same
problem.  The primary user appears to be 'git-svn.perl' and its
related tests.  A quick review of those files did not show any mention
of problems with Carp.  Perhaps they simply do not check the exact
exit code on failure?