ENH: Avoid re-encapsulation of the default handler - #20345

@mattip @seberg What do you think?

Seems reasonable to me. Do we expect users to test identity? Because if not, we could just not make this a public symbol, but I don't really care either way.

Do we expect users to test identity?

Define users :). My points in favor of exporting it publicly are:

• There is already a way to get (a borrowed reference to) it, but can be considered an overkill:

  PyObject *handler = PyDataMem_SetHandler(NULL);
  PyObject *default_handler = PyDataMem_SetHandler(handler);
  Py_XDECREF(handler);
  Py_XDECREF(default_handler);

• Maybe in the future NumPy could offer more than one native allocator (e.g. pooled/unpooled versions). PyDataMem_SetHandler(NULL) will always set the default one but still all the available options should be accessible by the users. For example:

  Py_XDECREF(PyDataMem_SetHandler(NULL)); // sets the default allocator
  Py_XDECREF(PyDataMem_SetHandler(PyDataMem_ArenaHandler)); // sets the arena allocator
  Py_XDECREF(PyDataMem_SetHandler(PyDataMem_DefaultHandler)); // sets the default allocator (for API consistency)

• It could enable a feature in the NumPy Allocator project, roughly equivalent to this:

  import numpy
  import numpy_allocator
  
  class default_allocator(metaclass=numpy_allocator.type):
      _handler_ = numpy_allocator.default_handler()
  
  with another_allocator:
      # do stuff
      with default_allocator:
          test = numpy.ndarray(())
  assert default_allocator.handles(test)

Good, any reason is enough of a reason for me in this case ;).

@mattip, if you like this, lets put it in for 1.22?

LGTM, thanks. It simplifies things a bit, at the cost of another exported API value