On Saturday 18 June 2011, Jim Meyering wrote:
> Stefano Lattarini wrote:
> >> ...
> >> >
> >> > In order to work with the upcoming new Automake testsuite harness, coreutils
> >> > have two possibilities:
> >> >  1. move the `shell_or_perl_' subroutine's functionality into a real acript,
> >> >     and define the LOG_COMPILER to point to it; or
> >> >  2. add a `.pl' extension to the perl test scripts, and define PL_LOG_COMPILER
> >> >     appropriately (might be a little tricky, considering the hops that the
> >> >     `shell_or_perl_' subroutine goes through in order to get the flags and
> >> >     imports right).
> >>
> >> 1) sounds preferable.
> >>
> > But it has a serious drawback: the redirection `9>&2' placed at the end
> > of TESTS_ENVIRONMENT will be rendered useless by the final exec done
> > in the new `shell_or_perl' script (at least for with shells using the
> > `cloexec' flag on fds > 2); this will bring back the problems fixed by
> > commit `v8.12-82-g6b68745' :-(
> >
Well, no, that's not true; the "problematic" shells only set the 'close-on-exec'
flag on the file descriptors they themselves create, not on the ones they inherit
from their parents:

  $ ksh -c 'exec 9>&1; sh -c "echo foo >&9"'
  sh: 9: Bad file descriptor
  $ ksh -c 'sh -c "echo foo >&9"' 9>&1
  foo
  $ (exec 9>&1; ksh -c 'sh -c "echo foo >&9"')
  foo

I should have tested better before reporting an imaginary problem.

> > So I now think we should go with solution (2).
> 
> Ok.
> 
I've gone with the less invasive solution (1) instead.  See the attached
patch.  I will soon post a follow up that tries to compensate for the
extra forks I've introduced here by removing a couple of grep calls from
the `shell-or-perl' script (this is for Cygwin, I know that the overhead
on Unix is utterly irrelevant).

Regards,
  Stefano