use CMPLX[FL]? macros instead of I

[fdopen]

The usage of I provided by "complex.h" is error-prone, e.g double complex z = Double_val(x) + I*(Double_val(y)) might give NaN as creal(z) for x = 0.0 and y = infinity.

The macros were introduced by C11, therefore I didn't make any assumptions, if they are available or not.

[yallop]

Let me check that I understand the issue here. C11 (or, at least this draft) says:

NOTE
These macros act as if the implementation supported imaginary types and the definitions were:

#define CMPLX(x, y) ((double complex)((double)(x) + \
_Imaginary_I * (double)(y)))

and also

The macro I expands to either _Imaginary_I or _Complex_I. If _Imaginary_I is not defined, I shall expand to _Complex_I.

So the CMPLX macros are preferable to using I because they are guaranteed to use imaginary arithmetic, whereas I may use complex arithmetic on some implementations, which can result in more NaN propagation.

[fdopen]

Yes, and glibc and many other stdlibs define I as _Complex_I. Therefore the surprising result is currently easy to demonstrate on the toplevel:

ComplexL.of_complex Complex.({re = 3.; im = nan});;
- : ComplexL.t = <complexld nan + nani>
ComplexL.of_complex Complex.({re = 7.; im = infinity});;
- : ComplexL.t = <complexld nan + infi>
ComplexL.of_complex Complex.({re = infinity; im = infinity});;
- : ComplexL.t = <complexld nan + infi>

_Imaginary_I is already mentioned in C99, but it is an optional feature. CMPLX was introduced by C11, but already seems to be available on more platforms.