On Wed, May 15, 2019 at 10:41:06PM +0300, Janne Blomqvist wrote:
> > 2) I don't think float _Complex is
> > passed the same as std::complex<float> and similar for others;
> > std::complex<float> is in libstdc++ a C++ class with with
> > __complex__ float as its sole non-static data member and with non-trivial
> > constructors; which means it is passed/returned via a hidden reference;
> > when the argument is actually FLOAT_COMPLEX * or FLOAT_COMPLEX &,
> > you except for aliasing don't have to care that much, but if
> > that complex argument has VALUE attribute in Fortran and so the
> > C prototype would be FLOAT_COMPLEX, then std::complex<float> is
> > passed in the end as std::complex<float> &, while float _Complex
> > the same as __complex__ float; and ditto for functions returning
> > complex
>
> Ugh, I guess that's right. Any good way around it? Except print a
> warning in the header that passing std::complex<> by value doesn't
> work?
Perhaps we can use different macros for the two cases, define
__GFORTRAN_FLOAT_COMPLEX to std::complex<float> and
__GFORTRAN_FLOAT_COMPLEX_VALUE to __complex__ float and use the
former in the __GFORTRAN_FLOAT_COMPLEX * arguments and
the latter for VALUE dummy args and return value.
For the return value case, guess it will be fine, when one does
__complex__ float fortran_fn_ (void);
std::complex<float> ret = fortran_fn_ ();
it will work just fine, but for the arguments trying to
void fortran_sub_ (__complex__ float);
std::complex<float> f = 2.0f + 4.0fi;
fortran_sub_ (f);
will not work (but one will get at least errors). One can use non-standard
fortran_sub_ (f.__rep ());
Jakub
