Hello,
I sent this bug report almost a year ago but never received a response, nor
has the problem been addressed, so I thought I would try again.
In Autoconf 2.57 (all releases) and in the CVS version, there is a problem
with AC_CONFIG_COMMANDS when it is used to create a file in a directory which
does not already exist at the time config.status is run. Consider the
following configure.ac:
AC_PREREQ([2.57])
AC_INIT([myprog], [0])
AC_CONFIG_COMMANDS([foo/bar], [echo "test" > foo/bar])
AC_OUTPUT
This results in a config.status which produces these error messages:
configure: creating ./config.status
./config.status: line 1: cd: foo: No such file or directory
./config.status: line 1: cd: foo: No such file or directory
./config.status: line 1: cd: foo: No such file or directory
./config.status: line 1: cd: foo: No such file or directory
config.status: executing foo/bar commands
These errors are caused by the following lines in config.status which are
executed right before the "foo/bar" target is invoked:
ac_abs_builddir=`cd "$ac_dir" && cd $ac_builddir && pwd`
ac_abs_top_builddir=`cd "$ac_dir" && cd ${ac_top_builddir}. && pwd`
ac_abs_srcdir=`cd "$ac_dir" && cd $ac_srcdir && pwd`
ac_abs_top_srcdir=`cd "$ac_dir" && cd $ac_top_srcdir && pwd`
The problem is that these commands are trying to change to the non-existent
"foo" directory.
This raises the question of whether AC_CONFIG_COMMANDS should be be
responsible for creating the "foo" directory itself before running the
target's commands, much like AC_CONFIG_FILES will create missing directories.
However, considering that AC_CONFIG_COMMANDS is intended to be generic
(that is, the target does not necessarily have to represent a file), then one
can successfully argue that AC_CONFIG_COMMANDS should _not_ create the "foo"
directory. This leads to the obvious answer that the target should itself
be responsible for creating its own directory if necessary. Thus, a
re-written configure.ac which creates its own output directory would appear
as follows:
AC_PREREQ([2.57])
AC_INIT([myprog], [0])
AC_CONFIG_COMMANDS([foo/bar],
[AS_MKDIR_P([foo])
echo "test" > foo/bar])
AC_OUTPUT
Unfortunately, this version _also_ fails with the exact same error messages.
The reason it fails is because config.status still attempts to compute
ac_abs_builddir, ac_abs_top_builddir, ac_abs_srcdir, and ac_abs_top_srcdir
_before_ the target's commands have been invoked (before `mkdir' is issued),
thus the problem is not solved.
One very ugly hack to work around this problem is to move the `mkdir' to
AC_CONFIG_COMMANDS's "init-cmds", as follows:
AC_PREREQ([2.57])
AC_INIT([myprog], [0])
AC_CONFIG_COMMANDS([foo/bar],
[echo "test" > foo/bar],
[AS_ESCAPE(AS_MKDIR_P([foo]), [$`\])])
AC_OUTPUT
Unfortunately, this solution, though it "works", is incorrect because it
causes "foo" to be created when _any_ config.status target is invoked, even
though "foo" should only be created when the "foo/bar" target is invoked.
This means that even the invocation "config.status --help" will incorrectly
create the "foo" directory.
Note that the ugly AS_ESCAPE invocation, for this case, is required to
guarantee that AS_MKDIR_P works correcly on platforms which do not have a
working "mkdir -p" command. AS_ESCAPE ensures that the variables and
subshell environments (such as $ac_dir, $ac_dirs, `dirname`, etc.) which
result from the expansion of AS_MKDIR_P are properly protected. Thus, these
variables and subshells will be evaluated correctly in the context of
config.status, rather than in the context of configure.
I would be interested in seeing a fix for this problem. This is a
real-world issue which I discovered when converting the Crystal Space project
(http://crystal.sourceforge.net/) to use Autoconf.
-- ES
