Implement H5Type for complex

[mulimoen]

Closes #208

[aldanor]

This is inherently unsafe, how do we know that offsets are right? (e.g. should we check actual offsets here?)

[mulimoen]

I am not sure how we would check the offsets

[aldanor]

@mulimoen Like this, I think? https://doc.rust-lang.org/stable/std/ptr/macro.addr_of.html

[mulimoen]

That would require us to make something to point at. We could make an instance of MaybeUninit, but field access is only a future feature

[aldanor]

Hm... I guess we could do something like this, does that make sense? https://rust.godbolt.org/z/TWbavj98G

[mulimoen]

We could do

unsafe impl<T: H5Type> H5Type for Complex<T> {
    fn type_descriptor() -> TypeDescriptor {
        use std::mem::MaybeUninit;
        // MaybeUninit has the same size and is transparent, but we can
        // do some pointer arithmetics on a real instance
        let dummystruct: Complex<MaybeUninit<T>> =
            Complex::new(MaybeUninit::uninit(), MaybeUninit::uninit());

        assert_eq!(size_of_val(&dummystruct), size_of::<Complex<T>>());

        let base = addr_of!(dummystruct).cast::<u8>();
        let ptr_r = addr_of!(dummystruct.re).cast::<u8>();
        let ptr_i = addr_of!(dummystruct.im).cast::<u8>();

        TypeDescriptor::Compound(CompoundType {
            fields: vec![
                // Compatible with h5py definition of complex
                CompoundField::typed::<T>("r", unsafe { ptr_r.offset_from(base) } as usize, 0),
                CompoundField::typed::<T>("i", unsafe { ptr_i.offset_from(base) } as usize, 1),
            ],
            size: size_of::<Complex<T>>(),
        })
    }
}

[mulimoen]

Might be enough to link to the documentation on representation? https://docs.rs/num-complex/0.4.2/num_complex/struct.Complex.html#representation-and-foreign-function-interface-compatibility

[aldanor]

Maybe add this link in a comment in code? E.g. "Complex<T> should be FFI-equivalent to [T; 2]: ".

[mulimoen]

Thanks, added this as a comment under type_descriptor()

[aldanor]

Minor nits aside, needs to be rebased on master (because of the changelog).