[BUG] Should cython object buffers check for NULL?

[seberg]

Pandas seems to semi-regularly run into issues with NumPy's logic which currently says that object arrays may be filled with NULL at initialization so that NULL is accepted everywhere to have the same meaning as None.
(See also pandas-dev/pandas#47097)

This is a bit weird, since NumPy also fills the array with None almost always, so in the few places where it doesn't it is unexpected!

I have opened numpy/numpy#21817 to solve this in NumPy. The intention would be that NumPy for now should accept NULL, but defines it as incorrect and will never produce it on its own (with some "internal" exceptions).

Now, I am not sure what the best solution is here and if you think that Cython should fix this (or we should do both), I can look into it.
In some sense, a fix in cython might be best... then pandas can just upgrade its Cython dependency and stop worrying about these oddities.

[seberg]

I think this would be the right changes (probably, I did not test it thoroughly):

diff --git a/Cython/Compiler/ExprNodes.py b/Cython/Compiler/ExprNodes.py
index c20a76bd4..19aa787e0 100644
--- a/Cython/Compiler/ExprNodes.py
+++ b/Cython/Compiler/ExprNodes.py
@@ -4578,17 +4578,17 @@ class BufferIndexNode(_IndexingBaseNode):
         buffer_entry, ptrexpr = self.buffer_lookup_code(code)
 
         if self.buffer_type.dtype.is_pyobject:
-            # Must manage refcounts. Decref what is already there
-            # and incref what we put in.
+            # Must manage refcounts. XDecref what is already there
+            # and incref what we put in (NumPy allows there to be NULL)
             ptr = code.funcstate.allocate_temp(buffer_entry.buf_ptr_type,
                                                manage_ref=False)
             rhs_code = rhs.result()
             code.putln("%s = %s;" % (ptr, ptrexpr))
-            code.put_gotref("*%s" % ptr, self.buffer_type.dtype)
-            code.putln("__Pyx_INCREF(%s); __Pyx_DECREF(*%s);" % (
+            code.put_xgotref("*%s" % ptr, self.buffer_type.dtype)
+            code.putln("__Pyx_INCREF(%s); __Pyx_XDECREF(*%s);" % (
                 rhs_code, ptr))
             code.putln("*%s %s= %s;" % (ptr, op, rhs_code))
-            code.put_giveref("*%s" % ptr, self.buffer_type.dtype)
+            code.put_xgiveref("*%s" % ptr, self.buffer_type.dtype)
             code.funcstate.release_temp(ptr)
         else:
             # Simple case
@@ -4609,8 +4609,11 @@ class BufferIndexNode(_IndexingBaseNode):
             # is_temp is True, so must pull out value and incref it.
             # NOTE: object temporary results for nodes are declared
             #       as PyObject *, so we need a cast
-            code.putln("%s = (PyObject *) *%s;" % (self.result(), self.buffer_ptr_code))
-            code.putln("__Pyx_INCREF((PyObject*)%s);" % self.result())
+            res = self.result()
+            code.putln("%s = (PyObject *) *%s;" % (res, self.buffer_ptr_code))
+            # NumPy does (occasionally) allow NULL to denote None.
+            code.putln("if (%s == NULL) %s = Py_None;" % (res, res))
+            code.putln("__Pyx_INCREF((PyObject*)%s);" % res)
 
     def free_subexpr_temps(self, code):
         for temp in self.index_temps:

I had a brief look at the test, but would have to look more to figure out how to get the NULLs in there best.

[da-woods]

I guess the follow-up question is: is there an equivalent issue with object memoryviews?

[seberg]

Oh, this only fixes np.ndaray and not object[:, :]? Because in that case... yes, and I should look into it.

[da-woods]

I believe so (although I haven't actually tested what typed memoryviews do). I'll re-open this issue and so it covers them too

[da-woods]

So it looks like this did affect memoryviews too (which I'm slightly surprised at but I guess is good). So false alarm I think

[da-woods]

I'll just create a copy of the tests for memoryviews and then it can be closed properly