remove refcount semantics, now a.resize() almost always requires refc… #8050
Conversation
| @@ -4080,7 +4081,10 @@ def test___array__(self): | |||
| class TestResize(TestCase): | |||
| def test_basic(self): | |||
| x = np.array([[1, 0, 0], [0, 1, 0], [0, 0, 1]]) | |||
| x.resize((5, 5)) | |||
| if IS_PYPY: | |||
There was a problem hiding this comment.
Out of curiousity, I noticed the import also defines HAS_REFCOUNT, is this reflected in the C-api so that this could be not PyPy specific but in principle good for all non-refcounted pythons?
There was a problem hiding this comment.
HAS_REFCOUNT reflects sys.getrefcount() where this pull request deals with pyobj->ob_refcnt not being incremented at the same places in CPython and PyPy. I'm not sure the two concepts are equivalent, I could imagine a garbage collection strategy that cannot count overall refcounts but can track single-item assignments, or visa-versa
|
@juliantaylor @mattip I'm wondering if #7997 is compatible with PyPy. |
|
Looks like you picked up a stray commit somewhere, probably merged master into your branch. |
|
@charris thanks, should be fixed now |
| y = np.array([123456789e199], dtype=np.float64) | ||
| y.resize((0, n)) | ||
| if IS_PYPY: | ||
| y.resize((0, n),refcheck=False) |
There was a problem hiding this comment.
Space missing, but will merge without a fix ;).
|
Unless I missed something big or someone else wants to look over (don't know the PyPy stuff much), this all looks good to me and I will merge it tomorrow. |
|
@seberg thanks, I will fix up that space in an inevitable upcoming pull request, there are still around 30 failing tests so chances are there will be another opportunity |
|
Thanks a lot, putting this in. Hopefully, your PyPy support gets better. Just thought of another Nitpick (its a sport ;)). Not actually sure, but I think we often make it |
The refcount semantics on PyPy do not provide a reliable way to determine if other references to the buffer are alive. So I propose to more stringently require refcheck=False for a.resize(). I also undefined the refcount-related entries in the NumPy headers, so downstream issues will be detected earlier.
Note that with this commit, NumPy now fails only 33 tests on a nightly PyPy.