Re: autoconf fails if env var U set

[Eric Blake]

On 04/13/2011 11:31 AM, Tim Wallace wrote:
> This is a bug discovered when trying to build postgresql 9.0.3 on Redhat
> Enterprise Linux 5.  If env var U is set, it fails, as pointed out by
> the Postgres maintainer.  I guess it's autoconf 2.59, and I'm guessing
> you're past that now, so possibly it has already been fixed.  Baffling
> bug, though.

It stems from the old days of ansi2knr, where automake would set $U when
translating source files into knr versions; but is relatively unused
these days.

> for ac_i in : $LIBOBJS; do test "x$ac_i" = x: && continue
>   # 1. Remove the extension, and $U if already installed.
>   ac_script='s/\$U\././;s/\.o$//;s/\.obj$//'
>   ac_i=`$as_echo "$ac_i" | sed "$ac_script"`
>   # 2. Prepend LIBOBJDIR.  When used with automake>=1.10 LIBOBJDIR
>   #    will be set to the directory where LIBOBJS objects are built.
>   ac_libobjs="$ac_libobjs \${LIBOBJDIR}$ac_i\$U.$ac_objext"
>   ac_ltlibobjs="$ac_ltlibobjs \${LIBOBJDIR}$ac_i"'$U.lo'

The issue still exists in autoconf 2.68 (and autoconf.git); but the
rules are not using $U in the configure script (read those lines
closely, and you'll note that what is really happening is that they are
generating a literal '$U' for inclusion in the AC_SUBST of LIBOBJS and
LTLIBOBJS).  So the real problem may be that automake is not prepared
for the case when $U is defined at make time, rather than on autoconf
for passing literal $U into the makefile in the first place during
AC_LIBOBJ.

Or maybe we do need to consider patching lib/autoconf/general.m4 to make
_AC_LIBOBJS_NORMALIZE only output $U into the makefile conditional on
whether ansi2knr is in use, rather than its current usage of doing it
for every package that used AC_LIBOBJ.

-- 
Eric Blake   address@hidden    +1-801-349-2682
Libvirt virtualization library http://libvirt.org

signature.asc
Description: OpenPGP digital signature