choice of file system build-dir can change how coreutils works

[Jim Meyering]

This is probably not big news, but it does bear repeating.

Today I compared two config.cache files, each generated by coreutils'
configure on the same Solaris 10 system.
The only difference is that one ./configure -C run was using /tmp
(a tmpfs file system), and the other was using an NFS-mounted
file system:

-gl_cv_header_working_fcntl_h=${gl_cv_header_working_fcntl_h=yes}
+gl_cv_header_working_fcntl_h=${gl_cv_header_working_fcntl_h='no (bad 
O_NOATIME)'

This doesn't change much in practice (dd's new noatime flag
is the only use), but it demonstrates why we should try hard
not to rely on configure-time "RUN" tests that may depend
on environmental factors like the build file system type.
However, sometimes there is no alternative.

In this case, perhaps it's ok to let dd support the notaime flag
as long as the system defines O_NOATIME.

However, note that for O_NOFOLLOW, it'd be a whole different kettle
of fish, (since it is used in security-related areas) if configure
detected support for it on the build file system, but an application
used it on a different file system, without support.