The PRECALL/CALL bytecode pair for calls is new (still in review at time of writing) and is not as efficient as it could be.

Some possible improvements are:

Transfer refcount of func when making a frame.

NULL call_shape.kwnames after use instead of in PRECALL, as many specialized instructions already assert that it is NULL.

Specialize the PRECALL instructions. Either for simple cases like `type(arg)` so we can skip the CALL, or for complex cases like `PythonClass(args)` where PRECALL can create object, and set up the frame so that `CALL` calls the __init__ function.