MAINT: Fix errors seen on new python 3.8

[seberg]

One of this is a small issue exposed by new warnings, the others are
simply adapting our test suit to stricter integer coercion rules
(avoiding float -> int conversions).

The last one is that we assumed pickle protocl 5 would be in 3.8, but
it seems that it did not make the cut.

Closes gh-13412

---

Marking as draft, since the pickle 5 protocol was approved for inclusion in 3.8, so I am not sure this is correct or python 3.8a3 is just lagging behind there. @pitrou do you know this quickly?

[pitrou]

PEP 574 is accepted but its implementation needs reviewing before it's integrated.

[seberg]

So we can assume that it will not be in 3.8 and change the test accordingly?

[pitrou]

For now we can assume it but I hope it will be integrated in time for 3.8 beta 1.

[seberg]

Thanks @pitrou, then I will just remove that for the moment. I do not care about the failure in the alpha release I think, I guess we probably will not officially support 3.8 in 1.17 anyway.

[mattip]

According to this mail, pickle 5 is slated for inclusion in 3.8, but it does seem a bit iffy since the first beta is supposed to be May 31.

[pitrou]

By the way, the C sources need re-cythonizing to fix them for Python 3.8:
cython/cython#2938 (comment)

[seberg]

I do not think we ship any cython generated C-sources for a while now. But we may have checks flying around to ensure that nobody uses too old cython versions?

[pitrou]

numpy/random/mtrand/mtrand.c is in the repo, so it won't be re-generated if you pip install from e.g. git master.

[seberg]

No, but the file is in .gitignore, so you may have to do a git clean -xdf or similar to make sure cython will regenerate it.

[mattip]

We ship the C code in the sdist, so that issue is relevant for our wheels/sdist creation

[pitrou]

@seberg Hmm, it is weird then. Perhaps pip installs its own Cython when building from an external source?

[mattip]

This seems like a good cleanup and is ready for merge, if there are no objections. We can handle any further changes and a possible bump of cython version in a future PR.

[seberg]

Since it got lost again here. Eric had pointed out that the commit message was wrong (it said it modified the pickle5 test, not deleted it), I fixed the commit message. I did not feel that xfail is helpful, since it is not directly related to numpy whether or not the test fails.

[pitrou]

For the record, I've just merged the PEP 574 PR (python/cpython#7076) so unless something bad happens it will be in 3.8 beta1.