Performance isn't (my) motivation for avoiding fork/exec (and the
equivalent on Win32) but rather avoiding the added complexity of
marshalling/IPC with fork/exec, as opposed to simply calling into a
native Haskell function and crossing process boundaries and having to
deal with the various things that can go wrong with the additional
moving parts you encounter when controlling an external process. So this
would IMO simplify code paths, and moreover I'd expect opportunities to
actually make the Haskell cpphs API richer (in case it isn't already)
and more tailored to GHC's lexer/parser pipeline and error-reporting.