Invalid calculation of MINfloat16/MAXfloat16 in pandas/algos.pyx

[vonosmas]

The code generated by Cython from pandas/algos.pyx contains an undefined behavior (invalid cast of floating-point number to integer):

easy_install -U Cython
git clone git://github.com/pydata/pandas.git
python setup.py install

The generated code for the line from pandas/algos.pyx

cdef np.float16_t MINfloat16 = np.NINF

in pandas/algos.c contains:

typedef npy_float16 __pyx_t_5numpy_float16_t;
//... more code
__pyx_t_5numpy_float16_t __pyx_t_10;
//... more code
__pyx_t_10 = __pyx_PyFloat_AsFloat(__pyx_t_1);

The file numpy-1.9.2-py2.7-linux-x86_64.egg/numpy/core/include/numpy/npy_common.h
defines npy_float16 as 16-bit integer type (I guess numpy implementation has a custom code that carefully represents these 16 bits as mantiss+exponent). Casting floating-point value -inf to a 16-bit integer type is an undefined behavior according to C standard.

I wasn't able to find any place in pandas code that uses MINfloat16 (and MAXfloat16). Is it safe to just delete them?

[jreback]

can u show an example with the error?

the float16 constants are prob hanging around - as it's not an explicitly supported type (though u can actually pass an array in) - it's just that all of the algos will cast

[vonosmas]

Unfortunately, I see this error only if I compile generated .c files with clang -fsanitize=float-cast-overflow (http://clang.llvm.org/docs/UsersManual.html#controlling-code-generation). Looks like Cython directives are somewhat limited... Let me know if I should craft reproduction steps.

[jreback]

dupe of #10382