Hi Junio,
On Tue, 9 Jan 2018, Junio C Hamano wrote:
> Johannes Schindelin <Johannes.Schindelin@xxxxxx> writes:
>
> > diff --git a/t/t0021/rot13-filter.pl b/t/t0021/rot13-filter.pl
> > index f1678851de9..470107248eb 100644
> > --- a/t/t0021/rot13-filter.pl
> > +++ b/t/t0021/rot13-filter.pl
> > @@ -31,7 +31,22 @@
> > #
> >
> > use 5.008;
> > -use lib (split(/:/, $ENV{GITPERLLIB}));
> > +sub gitperllib {
> > +...
> > + if ($ENV{GITPERLLIB} =~ /;/) {
> > + return split(/;/, $ENV{GITPERLLIB});
> > + }
> > + return split(/:/, $ENV{GITPERLLIB});
> > +}
>
> This cannot be the whole story for a few reasons.
>
> - In t/test-lib.sh we see this:
>
> GITPERLLIB="$GIT_BUILD_DIR"/perl/blib/lib:"$GIT_BUILD_DIR"/perl/blib/arch/auto/Git
> export GITPERLLIB
>
> If this part wants to split with ';', then the joining needs to
> be done with ';' to match, no?
No.
It is a lot more complicated than that. As you know, on Linux there is
this implicit assumption that path lists are colon-separated. As a
consequence, Cygwin does the same (because it would be too hard to port
all those Linux/Unix projects to stop assuming colon-separated path lists,
right?)
This is what Cygwin's Bash does (and hence the MSYS2 Bash used by Git for
Windows, too).
Then the MSYS2 Bash calls git.exe, which is *not* an MSYS2 program, hence
the MSYS2 runtime knows that it has to convert the path lists to Windows
paths separated by semicolons.
The next thing happening in our case is that the Perl script is called
from git.exe. Now, the MSYS2 runtime (implicitly spun up by the MSYS2 Perl
interpreter) does *not* convert those path lists back to Unix-like paths
separated by colons.
And that's why the Unix shell script can happily construct the
colon-separated list, and the Perl script will *still* receive the
semicolon-separated version of it.
> - In addition to t0021, there are similar split with colon in 0202,
> 9000 and 9700, yet I am getting the feeling that you observed the
> issue only in0021, to which I do not think of a good explanation
> why.
Here is the good explanation: t0021 relies on a Perl package that is not
yet installed. t0202 relies on Git::I18N, of which there is a version
installed in Git for Windows' SDK. (I do not bother to slow down the test
runs by the Subversion tests, I always skip all of them, that's why t9*
does not matter to me.)
Granted, that is a bug, and an old one at that: the test suite should test
what is in the current revision, not what happens to be installed into the
system locations.
But this late in the game, I am really uncomfortable with the idea to rush
through an equivalent fix to all Perl scripts: who knows what it breaks?
Ciao,
Dscho
