I've been using f2py to "pythonize" some Fortran codes for use on a 64-bit Windows 7 machine. I'm working with 64-bit Anaconda 4.2.0 and using MinGW installed via conda install mingw (the associated folders use the name "x86_64-w64-mingw32"). It was working well until I tried a program with quadruple precision (REAL*16) numbers. I found that the program crashed (libquadmath-0.dll was cited in the ProblemSignature of the Windows-generated crash report).
I tried to identify the simplest case that causes a problem. I ended up coming up with a fairly basic Fortran program that gives a "fatal error" under certain conditions:
PROGRAM MAIN
REAL*8 DL
REAL*16 QL
DL=2.71828
QL=3.14159
C when compiled with -ffpe-trap=overflow with MinGW distribution for 64-bit Anaconda, uncommenting either of the lines below breaks the program ("fatal error" at runtime)
C however, the second line is acceptable if DL is changed to quad precision (REAL*16)
C both lines are acceptable if both DL and QL are double precision (REAL*8)
OPEN (79)
print*, "DL=",DL
print*, "Ready to print QL!"
print*, "QL=",QL
end
As indicated in the comments above, when I compile with the Anaconda MinGW package's gfortran with the compiler option -ffpe-trap=overflow, I get a fatal error. For example, with both lines uncommented (as above), I get:
DL= 2.7182800769805908 Ready to print QL! A fatal error occurred! Backtrace for this error: #0 ffffffffffffffff
(See comments in the code for further details about what happens with different variations.)
I have tried with a standard MinGW32 distribution that I have set up on the same computer (compiling and running within msys command line), and there are no issues.
I believe I have confirmed that the libquadmath-0.dll included with the anaconda mingw package is a 64-bit dll. I have not been able to identify any issues with environment variables, and I think I have ruled out the possibility that it is using my standard MinGW32 32-bit dll by temporarily renaming the MinGW32 directory.
Aside from the 32-bit vs. 64-bit difference, I have noticed that my MinGW32 gfortran is based on GCC 4.8.1, whereas the gfortran with the 64-bit Anaconda MinGW package is listed as using GCC "4.7.0 20111220 (experimental)". I'm not sure if that is related to the issue, but I couldn't find anything obvious in the GCC changelog that would account for the issues I'm seeing.
I'd very much appreciate any tips / solutions anyone can provide.
I'd be especially interested if anyone can reproduce this error with another 64-bit MinGW distribution. At the moment, I don't know if it is an issue with GCC/gfortran, MinGW, Anaconda, or something else peculiar to my own setup.
