I'd name the new bytecode(s) explicitly here.

I have a question about gi_iframe. Below (L31) it is defined as an array of length 1 of object pointers. But everywhere it's used, it is cast to InterpreterFrame *. That's not an object or object pointer AFAICT. So what's going on here? Is the declaration wrong? An intentional lie?

It is a lie.

We don't want to expose InterpreterFrame in public headers, so we can't do the sensible thing and declare gi_frame as:

    InterpreterFrame gi_frame;

as C won't allow incomplete types in structs, even as the last member.

We could improve this by breaking up the header, so the public API sees a different definition. Still a lie, but a more elegant one.
Public header:

typedef struct {
    _PyGenObject_HEAD(gi)
} PyGenObject;

Private header:

typedef struct {
    _PyGenObject_HEAD(gi)
    InterpreterFrame gi_frame;
} PyGenObject;

Probably best done in a different PR, though.

Gotcha. Any type other than PyObject* would be better though :-). And if you replace the casts with a macro it’s easier to fix later. I have an idea for the macro but it’s too painful to type on a phone.

The first field of InterpreterFrame is PyFunctionObject * and the first few fields are all PyObject *, so declaring gi_frame as PyObject * gi_frame[1] seemed safe. Do let me know what your macro is, once you at a PC.

Do let me know what your macro is, once you at a PC.

See https://bugs.python.org/issue40120#msg365465