fix julia_type<void>

[mofeing]

fixes JuliaInterop/CxxWrap.jl#463

the erroring tests seem to be unrelated to the changed code

[barche]

Type void is already mapped here:

libcxxwrap-julia/src/jlcxx.cpp

Line 397 in dd11c21

set_julia_type<void>((jl_datatype_t*)julia_type("Cvoid", jl_base_module),false);

So the first thing to find out is why this fails in your setup.

With regards to this PR, the error on e.g. Ubuntu is:

/home/runner/work/libcxxwrap-julia/libcxxwrap-julia/include/jlcxx/type_conversion.hpp:502:8: error: extra qualification not allowed [-fpermissive]
  502 | struct jlcxx::julia_type_factory<void>
      |        ^~~~~

So there is a syntax error. But adding a separate julia_type_factory for void is not the correct way to solve this, it just fights the symptom of a deeper problem.

[mofeing]

With regards to this PR, the error on e.g. Ubuntu is:

/home/runner/work/libcxxwrap-julia/libcxxwrap-julia/include/jlcxx/type_conversion.hpp:502:8: error: extra qualification not allowed [-fpermissive]
  502 | struct jlcxx::julia_type_factory<void>
      |        ^~~~~

So there is a syntax error. But adding a separate julia_type_factory for void is not the correct way to solve this, it just fights the symptom of a deeper problem.

ah right, that is a typo sorry

Type void is already mapped here:

libcxxwrap-julia/src/jlcxx.cpp

Line 397 in dd11c21

set_julia_type<void>((jl_datatype_t*)julia_type("Cvoid", jl_base_module),false);

So the first thing to find out is why this fails in your setup.

yeah, that's weird. i even tried calling register_core_types and register_core_cxxwrap_type by myself inside the define_module_julia but doesn't work. 🫠

one annotation: we are using bazel as build system, and due to a symbol conflict between Julia's LLVM and our LLVM, we only export some C-API symbols marked manually. AFAIK only define_module_julia is required right?

[mofeing]

okay, so i run this inside the function definining the module

std::cout << "has_julia_type<void>() = " << jlcxx::has_julia_type<void>() << std::endl;

jlcxx::register_core_types();
jlcxx::register_core_cxxwrap_types();

std::cout << "has_julia_type<void>() = " << jlcxx::has_julia_type<void>() << std::endl;

and it returns this during Julia precompilation

has_julia_type<void>() = 0
has_julia_type<void>() = 0

checking register_core_type, the only thing that i can imagine it might be happening is that for some weird problem the static bool registered might default to true?

if i take the inner code from ..., and i put it in my code...

jlcxx::set_julia_type<void>((jl_datatype_t*)jlcxx::julia_type("Cvoid", jl_base_module),false);
jlcxx::set_julia_type<void*>(jl_voidpointer_type,false);
jlcxx::set_julia_type<float>(jl_float32_type,false);
jlcxx::set_julia_type<double>(jl_float64_type,false);

jlcxx::set_julia_type<jl_datatype_t*>(jl_any_type,false);
jlcxx::set_julia_type<jl_value_t*>(jl_any_type,false);

std::cout << "has_julia_type<void>() = " << jlcxx::has_julia_type<void>() << std::endl;

i get

has_julia_type<void>() = 1

[barche]

I am not able to reproduce this, what kind of system are you running all this on, exactly?

[mofeing]

i managed to reproduce the error with a MWE and packaged everything here https://github.com/mofeing/ReproducerCxxWrapVoidBug

if you activate the package's env and call ]build you will get the error when the package is precompiling. if you then comment this line, https://github.com/mofeing/ReproducerCxxWrapVoidBug/blob/299185e0aa5b9d6655c396dfbf89617b8b9e4329/deps/CMakeLists.txt#L18, the one adding the -fvisibility=hidden (but exporting define_julia_module), the error will disappear (i believe -fvisibility is just for clang so for gcc, you need to use the equivalent flag).

AFAIK only the julia module defining function is ccalled, but it seems like the addresses of the C++ function stored there are not correct or directly skipped. not sure about how this last part happens.

tldr: the problem is that we are filtering all symbols by default except some in a passlist, and the C++ symbols are not exported.

i'm closing because my problem is solved, but might be good to record this somewhere in the docs?