We can't handle 128bit long double in softfloat yet, but most libraries
are compiled like this. It's better to have the compiler fail than to
produce code with a mismatching ABI in the default setting.