BLD: Fix npy_isnan of integer variables in cephes

[fweimer]

Commit ede5590 (BUG: special: fix
ufunc results for nan arguments) started to apply npy_isnan to integer
arguments by mistake. GCC's __builtin_isnan issues errors when
applied to integers, so depending on which implementation is chosen by
numpy, this can result in build errors.

[pv]

Thanks, lgtm. Just noticed this myself yesterday on a different machine...

[person142]

My bad, thanks for the fix @fweimer.

[person142]

This means we should add a nan check to _legacy_cast_check:

https://github.com/scipy/scipy/blob/master/scipy/special/_legacy.pxd#L40

Otherwise if you pass in a nan it gets cast to an int, which could then in theory give you some garbage non-nan result.

[pv]

Correct, but I think the comparison is a floating point comparison so whatever value nan gets cast to, is not equal to nan.

[pv]

Of course, if it's undefined behavior, the compiler may choose to emit code running "rm -rf /". (see e.g. https://kristerw.blogspot.it/2017/09/why-undefined-behavior-may-call-never.html :)

[person142]

Of course, if it's undefined behavior, the compiler may choose to emit code running "rm -rf /".

Thankfully on my box I don't get anything that drastic, but the cast to int gives you -2147483648. Now combining that with

>>> sc.bdtrc(-2147483648, 1, 0.0)
1.0

makes one sad. :-(

So on master:

>>> import scipy.special as sc
>>> import numpy as np
>>> sc.bdtrc(np.nan, 1, 0.5)
__main__:1: RuntimeWarning: floating point number truncated to an integer
1.0