In particular, we are testing the output of deriving
Generic instances.

We were getting the RTS linked twice, which caused odd things to happen.

We also use it internally, which saves us running "ghc --print-libdir"
on Windows.

It was failing for me on OS X as canonicalizePath was altering the
capitalisation of the path (on a case-insensitive filesystem).

It now also prints diagnostic info if it fails.

I think ideally we should give the info from both the old and new
messages; i.e. "You wrote X, which I reduced to Y, which isn't valid".
That way if Y is an abstract newtype then it will be clearer why
the type wasn't valid. We used to only give X, and now we only give Y.
Finding X is now a little messier, especially in cases like ccfail005.

You can only use newtypes in the FFI if the constructor is visible

cf Trac #5509

It now knows about all the differences between the extensions that
GHC knows, and the extensions that Cabal knows.

This fixes the testsuite driver with python 2.5. Patch from Florian Weimer.

It was failing due to quoting differences between Windows and other
platforms.