bug#48155: [dejagnu-1.6.3-rc4] make check depends on SHELL

[Jacob Bachmeyer]

Kiyoshi KANAZAWA wrote:
> Hello,
> Tried dejagnu-1.6.3-rc4 on Solaris x86/x64 with many SHELLs & found
> make check fails with SHELL=/bin/csh,
> which does not occur with dejagnu-1.6.3-rc3.
> I know csh is not recommend, but still have to say something changed worse.
>   

The root cause of this issue is a program called "shellcheck" and the 
decision by the maintainers of config.guess to take its advice and break 
backwards compatibility with pre-POSIX shells, while Solaris 10 /bin/sh 
is still a pre-POSIX shell, leading to bug#47382 when I imported the 
newer config.guess in preparation for the 1.6.3 release.  Now DejaGnu 
must find some way to choose a shell that can run config.guess and use 
it for that purpose if config.guess must be run.

DejaGnu is installed in an architecture-independent tree, so 
substitutions made by configure are not acceptable here:  the installed 
DejaGnu can be shared with other machines after installation.

The solution adopted in 1.6.3-rc4 is to check the environment for 
variables named CONFIG_SHELL and SHELL (in that order) and use the first 
one found to run config.guess, falling back to simply running 
config.guess directly (which will use /bin/sh) if neither is set.  
Obviously, this causes problems if CONFIG_SHELL is not set and $SHELL is 
not a POSIX Bourne shell.  In an effort to limit those problems, runtest 
now also checks that the output of config.guess looks plausible and 
aborts very early if it does not.

I am very interested in the "Native configuration is ..." line from 
"make check" in the failing case; runtest should detect that 
config.guess did not produce a valid result and abort instead of running 
any tests at all.


I see a few options:

1.  Check $SHELL (apparently [exec $::env(SHELL) -c {A=$(echo X); B=Y; 
echo $A${B}Z}] should be enough to exclude csh and non-POSIX Bourne 
shells, producing "XYZ" if $SHELL is acceptable) and either:
1a.  Unset SHELL if the check fails.  This would mean that testsuites 
could depend on SHELL naming a POSIX Bourne shell if it is set during a 
testing run.
1b.  Immediately bail out if the check fails, informing the user that 
CONFIG_SHELL must be set to the executable name of a POSIX Bourne shell 
if SHELL is set to something else.

2.  Improve the error message produced when config.guess fails to run to 
explain that CONFIG_SHELL must be set to the executable name of a POSIX 
Bourne shell if SHELL is set to something else.

The 1a option does lead to possible future support for /finding/ a POSIX 
Bourne shell and thus allowing testsuites to expect that SHELL will be a 
"standard" shell.


-- Jacob