While the GNU autotools family (automake, autoconf and
libtool) is powerful and fairly straightforward to use, it's also very
slow and causes vast bloat in distributed source code.
For instance, my freedt package contains several small programs, but the
configure script that autoconf generates is three times larger
than all the source code put together.
An alternative would be to keep the Makefile.am and
configure.in formats as they are, but write a program that can
interpret them rather than compiling them to standard Makefiles and shell
scripts.
This should be much faster -- since it could optimise configure tests and avoid
the shell overhead for many operations, and avoid the need for recursive make
usage -- and reduce the need to ship the vast autotools output in source
tarballs.