F2PY installation crashes sometimes

[aichhorn]

The f2py call for the fortran modules is not always stable. The BuildF2PYModule of cmake executes the command
f2py -c --f90exec= ....
This results in many cases in bad linking of the library file. Instead, the command should be
f2py -c --fcompiler= ...
where is gnu95 or intelem, for example. Problem: can not be detected easily. Maybe set gnu95 as standard, that should work always.

[parcollet]

Do you mean that in FindBuildF2pyModule.cmake line 38
... --f90exec=${CMAKE_Fortran_COMPILER} ...
CMAKE_Fortran_COMPILER is wrong ?
So cmake does not detect correctly the fortran compiler on your machine ??

What if we put gnu95 as a standard and people want to compile with ifort ?

[aichhorn]

The CMAKE_Fortran_COMPILER is correct in the module, the problem is that one should not use the --f90exec flag with specifying just the F90 executable. That leads to wrong linking in many cases.
The --fcompiler is the better flag, but you do not give the executable, but the 'vendor'. The two flags behave differently. The point is, if you put just --f90exec=ifort, then the .so is compiled with ifort, linked to some gfortran libraries.
One could introduce a variable, like CMAKE_F2PY_vendor, that specifies the vendor and is set to gnu95 by default. If wanted, can be changed by hand. Detection of the possible values of -fcompiler can be done by
f2py -c --help-fcompiler

[mferrero]

@aichhorn: is there a way to reproduce this issue? I would like to avoid the vendor flag. Are you sure the fact that there are links to libgfortran is not a consequence of linking to lapack?

[aichhorn]

Right now, I cannot reproduce this issue myself either. I tried all possible combinations of gfortran and ifort, with lapack and without, and everything worked out fine. This seems to be a well hidden problem in the f2py linking procedure that I don't understand.
(I tried it for the modules vertex.so and hubbard_I.so) We could close this issue, if nobody else is having this problem.

[mferrero]

OK, let's close the issue for now and if ever the problem appears again we can still reopen it...