GHC doesn't compile

[Sven Panne]

I wrote:
> [...] A quick look in the CVS log reveals that the test in question was
> added as a pure performance improvement for Digital UNIX, so removing it
> would probably be the best and most defensive way to go.

I've just completely rewritten the ar-related autoconf magic, making it much
more modular and hopefully more correct. :-) I decided to keep the support
for -input, but made the test more robust. Could you try the HEAD, please,
and tell me if things improved for you.

Cheers,
    S.