PyLong to std::size_t conversion overflows

[Jan Blechta]

Conversion from PyLong to std::size_t fails for values >= 2**63 on 64-bit machines. The key of the problem is in dolfin/swig/typemaps/primitives.i:Py_convert_std_size_t. Solution of the problem should probably wait until Py3 transition is finished.

MWE:

from dolfin import *
m = UnitIntervalMesh(1)
f = CellFunction('size_t', m)
f[0] = 2**63
# TypeError: (size_t) expected positive 'int' for argument 3

[Martin Sandve Alnæs]

There are no outstanding py3 branches, feel free to fix away.

[Jan Blechta]

What about testing with py3? Is it handled by buildbots, @johannes_ring?

Is somewhere written what python versions do we support?

[Johannes Ring]

@blechta - Yes, we have a buildbot running Python 3.

[Martin Sandve Alnæs]

FEniCS dev requires 2.7, and the py3 buildbot passes all but a few tests. Some of the failures are a bit weird and may or may not be hard to debug. Feel free to look at them.

[Jan Blechta]

And which is the lowest py3 version required?

[Martin Sandve Alnæs]

I don't know. I think Aslak used 3.2 while I've used 3.4. @johannes_ring what 3.x does the buildbot use?

[Johannes Ring]

@martinal - 3.4, the one in Ubuntu 14.04.

[Martin Sandve Alnæs]

@blechta we haven't discussed which 3.x version to require, but I think there are minor issues that are more compatible between 2.7 and 3.2 than with 3.0 and possibly 3.1, so requiring 3.2 makes sense. For all I care it's just as well to just jump straight to 3.4. Personally I'll stick with 2.7 for now, as 3.x seems a lot slower.

[Jan Blechta]

Fix issue 377 - conversion of large PyLong to size_t.

→ <<cset 7530d4f31f98>>

[Jan Blechta]

Fix in pull request #162. Nevertheless some strange issues persist

from dolfin import *
import numpy as np
m = UnitIntervalMesh(1)
f = CellFunction('size_t', m)

f[0] = np.uintp(2**63 - 1)
print f[0] == 2**63 - 1 # True
print f[0] == np.uintp(2**63 -    1) # True
print f[0] == np.uintp(2**63 -  512) # True!!!
print f[0] == np.uintp(2**63 + 1024) # True!!!

f[0] = 2** 64 - 1 # now working
print f[0] == 2**64 - 1 # True
f[0] = np.uintp(2** 64 - 1) # not yet working

[Jan Blechta]

Fix in pull request #162.

[Johan Hake]

I have now fixed the last conversion. We needed to treat numpy integers different to Python ints, at least for Python 2. Will push and merge as soon as it passes local tests.

[Prof Garth Wells]

@johanhake Is this now resolved?

[Jan Blechta]

It is not. Third and fourth print statement from the example still print True.

[Anders Logg (Chalmers)]

Removing milestone: 1.7 (automated comment)

[Prof Garth Wells]

This will be dealt with by #897.