using long double c++ math functions result in a segfault with mingwpy's gcc 4.9.2

[EdwinBennink]

I noticed that using c++ math functions (like std::fmod) with long doubles result in a segfault using the current mingwpy version. This is gcc version 4.9.2 (x86_64-win32-seh-rev201506, mingwpy build).

Nathanial smith suggested:

Maybe the problem is that libstdc++ is getting compiled with -mlong-double-80, and that has ABI effects that make it incompatible with code compiled with -mlong-double-64? Is mingwpy's tweak of the spec file defaults happening before or after libstdc++ gets built?

To reproduce the segfault, compile foo.cxx below without any compiler flags:

#include <cstdio>
#include <cmath>

int main()
{       
    printf("fmod -> %f\n", fmod((long double)1, 16));
    printf("std::fmod -> %f\n", std::fmod((long double)1, 16));

    return 0;
}

Forcing 80 bit long doubles by adding the -mlong-double-80 compiler flag and including CRT_fp10.o does not fix the issue, probably since the mlong-double-64 flag is hardcoded in the 4.9.2 compiler specs.

[EdwinBennink]

The following issue with numpy is probably connected. I built numpy 1.11.0 for Python 3.4.4 (Win64) with mingwpy's gcc 4.9.2 (x86_64-win32-seh-rev201506). The following python code resulted in a segfault when calling fmodl:

import numpy
import operator
a = numpy.array([1, 2, 3], dtype='g')
[operator.floordiv(a_, b_) for a_, b_ in zip(a,a)]

Backtrace:

#0  0x000000006c08cad0 in fmodl () from /d/develop/release/lib/site-packages/numpy-1.11.0-py3.4-win-amd64.egg/numpy/core/umath.pyd
#1  0x000000006c07fccc in npy_fmodl (y=1, x=1) at numpy/core/src/npymath/npy_math.c.src:482
#2  npy_divmodl (a=1, b=1, modulus=0x2ef218) at numpy/core/src/npymath/npy_math.c.src:621
#3  0x000000006c067cf7 in longdouble_ctype_floor_divide (out=0x2ef210, b=<optimized out>, a=<optimized out>) at numpy/core/src/umath/scalarmath.c.src:288
#4  longdouble_floor_divide (a=<optimized out>, b=0x38013a8) at numpy/core/src/umath/scalarmath.c.src:862
#5  0x0000000062d1732f in python34!_PyFloat_Unpack8 () from /d/develop/release/python34.dll

[carlkl]

as mingwpy handles long doubles identical to doubles numpy (hint: MSVC compatibility) has to take care of it during build.

That was the case with a recently refused numpy PR. This PR has to be rebuild (be my) and published again. When this is done, is should be possible to build numpy with mingwpy again.