Cannot build ADIOS2 with clang/gfortran

[eschnett]

I am building ADIOS2 2.8.1 via Spack using clang @14.0.5 as C/C++ compiler and gfortran @11.1 as Fortran compiler. This fails when compiling the Fortran interface. The problem is that the flag -fallow-argument-mismatch is passed to clang++, which fails.

cd /tmp/eschnetter/spack-stage/spack-stage-adios2-2.8.1-bgwwyui542uytlc2b6eump6k2k3d6toj/spack-build-bgwwyui/bindings/Fortran && /home/eschnetter/src/spack/lib/spack/env/clang/clang++ -DADIOS2_USE_MPI -DMPICH_SKIP_MPICXX -DOMPI_SKIP_MPICXX -Dadios2_fortran_mpi_EXPORTS -I/tmp/eschnetter/spack-stage/spack-stage-adios2-2.8.1-bgwwyui542uytlc2b6eump6k2k3d6toj/spack-build-bgwwyui/bindings/Fortran -I/tmp/eschnetter/spack-stage/spack-stage-adios2-2.8.1-bgwwyui542uytlc2b6eump6k2k3d6toj/spack-src/bindings/C -I/tmp/eschnetter/spack-stage/spack-stage-adios2-2.8.1-bgwwyui542uytlc2b6eump6k2k3d6toj/spack-src/source -I/tmp/eschnetter/spack-stage/spack-stage-adios2-2.8.1-bgwwyui542uytlc2b6eump6k2k3d6toj/spack-build-bgwwyui/source -isystem /home/eschnetter/src/spack/opt/spack/linux-ubuntu18.04-skylake_avx512/clang-14.0.5/mpitrampoline-4.0.2-qxq4cp72man4yt75u6qmanoja2oj2aiy/include -O3 -DNDEBUG -fPIC -fallow-argument-mismatch -fcray-pointer -std=gnu++11 -MD -MT bindings/Fortran/CMakeFiles/adios2_fortran_mpi.dir/f2c/adios2_f2c_io_mpi.cpp.o -MF CMakeFiles/adios2_fortran_mpi.dir/f2c/adios2_f2c_io_mpi.cpp.o.d -o CMakeFiles/adios2_fortran_mpi.dir/f2c/adios2_f2c_io_mpi.cpp.o -c /tmp/eschnetter/spack-stage/spack-stage-adios2-2.8.1-bgwwyui542uytlc2b6eump6k2k3d6toj/spack-src/bindings/Fortran/f2c/adios2_f2c_io_mpi.cpp
clang-14: error: unknown argument: '-fallow-argument-mismatch'
bindings/Fortran/CMakeFiles/adios2_fortran_mpi.dir/build.make:104: recipe for target 'bindings/Fortran/CMakeFiles/adios2_fortran_mpi.dir/f2c/adios2_f2c_adios_mpi.cpp.o' failed

Note that I have configured ADIOS2 with MPItrampoline as MPI implementation. This library adds the flags -fallow-argument-mismatch and -fcray-pointers to the Fortran flags. I do not think that this is relevant because many of ADIOS2 C/C++ source files have been successfully compiled at this time, without using these flags.

I notice that bindings/Fortran/CMakeLists.txt defines adios2_fortran as a list of Fortran source codes, although some of these files are instead C++ files.

[eisenhauer]

Probably we need more information here. Maybe run "make VERBOSE=1" so we can see what is being compiled when this happens?

[eschnett]

I attach the complete configure and build output from Spack:

• spack-build-env.txt
• spack-configure-args.txt
• spack-build-01-cmake-out.txt
• spack-build-02-build-out.txt

[eisenhauer]

@chuckatkins I don't see anything wrong in the cmake spec, but the build seems to be applying CMake_Fortran_FLAGS to C++ compilations when building the adios2_fortran bindings library (which is a mix of C++ and Fortran source). Possible cmake problem?

[eschnett]

I have since encountered a very similar problem in superlu_dist. This also points to a cmake problem.

But what about this line in your CMakeLists.txt:

  target_compile_definitions(adios2_fortran PRIVATE "$<$<COMPILE_LANGUAGE:Fortran>:ADIOS2_HAVE_MPI_F>")

Doesn't this declare adios2_fortran to be Fortran code?

[eisenhauer]

I can't speak as authoritatively as others might, but i think the answer is that that line just adds -DADIOS2_HAVE_MPI_F to the command line for fortran compiles done in building adios2_fortran dependencies. I don't think it declares all the dependencies of adios2_fortran to be fortran source. CMake should infer what is fortran and what is C++ from the filename, unless you override that with something like: "set_source_files_properties(hello-world.f PROPERTIES LANGUAGE Fortran)"