unless I'm mistaken, only x86 needs -ffloat-store.  x86_64 certainly
doesn't need it, because it uses SSE2 with the correct-sized floating
point registers and doesn't store temporary results with more
precision than results in memory.