[patch] fix recent breakage of darwin, mingw and others

[Tobias Burnus]

Daniel Franke wrote:
> Tested on i686-pc-linux-gnu by myself, on x86_64-pc-mingw32 by Kai Tietz
> and on Dominique Dhumieres currently tests it on i686-darwin.
>
> Assuming no further problems, ok to commit?

OK.

Can you open a PR for getting the TARGET_* macros back after the
problem(s) is (are) solved?

Tobias

PS: With that patch, we loose the following #defines on x86-64:
  __amd64, __amd64__, __ELF__, __gnu_linux__,
  __k8, __k8__, __MMX__,  __SSE__, __SSE2__,
  __SSE2_MATH__, __SSE_MATH__, __x86_64, __x86_64__
but bootstapping is more importrant :-)

PPS: With gfortran-4.3, when we called cc1, there were even more
defines: e.g. __GCC_HAVE_SYNC_COMPARE_AND_SWAP_1, __linux, __unix
__DBL_HAS_QUIET_NAN__, __DEC64_MIN__, __FLT_EPSILON__,
__SIZEOF_POINTER__. Though I'm not sure whether
we need have those; though especially  __linux and __unix could be
useful, some of the others might be useful for C interoperabilty,
though ISO_BIND_C should handle most of this already.