Maybe you can use PyModule_AddType in here too?

@shihai1991 yes and I guess I should convert these to heap types too

@shihai1991 @vstinner I am trying to convert the NDArray_Type to a heap type but it has binary functions in its as_mapping that need access to the module state, e.g.:

nd = (NDArrayObject *)ndarray_new(&NDArray_Type, NULL, NULL);

The signature of this method is:

There doesn't seem to be a way to get the module in this case - so I believe converting to heap types might block until a future PEP. Please correct me if I am wrong but I will proceed with using PyModule_AddType without heap types.

Maybe you can take a look of victor's PR: 1c2fa78

@shihai1991 That PR does not have any types with an as_mapping method defined

If it is not possible to subclass the type, you can use PyType_GetModule().

@vstinner I think I'm missing some information. How can I use PyType_GetModule() on a method signature like:

There is no type here for me to access.

Py_TYPE(self) gives you a type.

@vstinner oh, right. Every object struct starts with PyObject_HEAD which has the type. I totally forgot about that.

Hum, would it be possible to move these static variables to be able to clear them when the extension module is unloaded? It would require to pass the module to NDArray_Type creation, so be able to get the module state from a ndarray instance.

@vstinner This seems tricky. How do you recommend passing the module to a static type like NDArray_Type. I can't put the module instance on the type or it will be overridden each time we do _testbuffer_exec. Is there a way to do this in tp_new?

@vstinner What if I just call PyUnicode_FromString every time it is used, since as you say this is just a test module. Struct and calcsize should be handled properly in any case.

Usually to convert an extension module to multiphase init, you must convert static types to heap types first.

@vstinner see my comment in response to @shihai1991 about this - I was not able to convert these to heap types

Converting _testbuffer to multiphase init allows to create two instances of the module, but the C implementation uses global variables which is incompatible with that:

You must first avoid these global variables. I suggest you to attempt to pass the module to functions using Struct and calcsize, and then get Struct and calcsize from the module. Either use PyObject_GetAttrString() which can slow down the module (which is ok, the module is only used for testing purpose), or use a module state (see my other comment about that).

@vstinner I am more comfortable with module state, anyways eventually it will need it if we migrate to heap types.