Re: Portability of git shell scripts?

From: Armin Kunaschik <megabreit@googlemail.com>
Date: May 6, 2016, 16:57 UTC
Message-ID: <CALR6jEiWQUO0WsOVah2O+9-j2Qx7juYjP7cb1v+P=Ey0oqQszg@mail.gmail.com>
In-Reply-To: <20160504212028.GG21259@sigill.intra.peff.net>

On Wed, May 4, 2016 at 11:20 PM, Jeff King <peff@peff.net> wrote:

Show 16 quoted lines

> On Wed, May 04, 2016 at 08:17:38PM +0200, Armin Kunaschik wrote:
>
>> I'm trying to compile/test/use git 2.8.2 on AIX 6.1 with no bash available.
>> /bin/sh is a hard link to /bin/ksh which is a ksh88, a posix shell.
>> Is this supposed to work?
>
> We aim for a practical subset of Bourne shells, including bash, dash,
> ash, ksh, etc. There's at least one Bourne-ish shell known not to work,
> which is Solaris /bin/sh[1]. POSIX is usually a good guide, but we aim
> for practical portability more than adhering strictly to the standards
> document.
>
> I've tested with mksh in the past (though it's possible that we've
> introduced a regression since then). But I think we've run into problems
> with ksh93[2]. I don't know about ksh88, or what construct it doesn't
> like.  It may or may not be easy to work around.

In general ksh (88 or 93) is posix compliant... and bash is moving away from posix. :-) But I know what you mean.

Show 8 quoted lines

>> As an example: make test fails on nearly every t34* test and on tests
>> which contain rebase.
>> The installation of bash (and manually changing the shebang to
>> /bin/bash) "fixes" all rebase test failures. So obviously git-rebase
>> is not portable at some point.
>
> Right. Any modern-ish Bourne shell will do, so moving to bash is one way
> to fix it.

My last compile of git 2.2.2 did far better than the current 2.8.2. So it looks like there were more recent changes that broke portability.

Show 6 quoted lines

>> Does it make any sense to put work into making these scripts portable,
>> that is, work with posix shells?
>
> Maybe. :) If you can find what it is that ksh88 is unhappy with, we can
> see how painful it is to adapt to. But given my looking into ksh93 in
> [2], I suspect it will be easier to just use a more modern shell.

Regarding [2] this was a bug which was fixed quite fast. To me this is no real showstopper. Modernity of ksh93 depends on the letter after the 93 :-)

>> And, as last resort, is it possible to configure git use bash in some
>> or all shell scripts?
>
> You can set SHELL_PATH in your config.mak file.

I tried a build with SHELL_PATH=/bin/bash. Many problems "went away". Others appeared. I'll give it a few more days to look into it.

First finding: make test does not make it through t3513-revert-submodule.sh anymore. The test is not portable since it uses the z-flags of GNU-tar. When -z is removed, (and extension is changed back to tar) everything runs and tests smoothly.

Is this report enough to start the magic to change things?

Regards, Armin