> By leaf, you mean the derived subtype? That sounds a bit quirky (you
> need the aforementioned "if"), how about the base heap type instead?

I think this fails (currently) because a subtype defined in Python using
a "class" statement will automatically get subtype_dealloc as its dealloc
function, which will in turn unconditionally decrefs the type (after
calling the tp_dealloc function from its nearest ancestor which doesn't
use subtype_dealloc).

> Speaking of which, what does this refactoring bring exactly? The type
> declarations are slightly less verbose, but other than that, is there a
> point?

It's really about the PEP 3121 changes: modules shouldn't have any
global variables shared across interpreters, so that module cleanup
can work properly. Ultimately, this can lead to gc-based shutdown,
module unloading, and better separation of interpreters.

In addition, it further reduces the differences between "extension
types" and "classes" (which supposedly were removed by dropping
old-style classes, yet some inconsistencies remain where the
interpreter offers some features only to heaptypes).