Nice! Pretty much the same as #292, but I think I like this one more.

I explored if generating a custom TypeList impl (which would eliminate the recusion problem completely) would a sensible alternative.

It's definetely possible, but I think it'd require a breaking change to be nice. What I had in mind was

fn dependency_types() -> impl TypeList {
  #[derive(Copy, Clone)]
  struct Deps;
  impl TypeList for Deps {
    fn for_each(self, v: &mut impl TypeVisitor) {
      v.visit::<TypeOfField1>();
      v.visit::<TypeOfField2>();
      Something::dependency_types().for_each(v);
      SomethingElse::generics().for_each(v);
    }
  }
  Deps
}

This breaks with generic parameters though. The naive for_each impl would use them, but it can't. to make that work, Deps would need to be generic as well. Definetely something we could do right now, but it's kinda messy.

If we changed fn dependency_types() -> impl TypeList to fn dependency_types(v: &mut impl TypeVisitor), these problems should go away. That would remove one level of indirection.

So tl;dr: I think we should go ahead with this, and keep this change in mind if someone still runs into the issue or for a breaking 9.0 release some time in the future.

Note to myself: Add two tests from other branch.

So, I've added some reference counting (probably doesn't have a big performance impact, but it felt like the right thing to do), added the test from the other branch and used a HashSet for types as well.

From my side, this is ready to merge.

Thanks @kamadorueda for getting this of the ground!