Allocating complex data (btVector3) on the stack in LLVM IR

I want to use LLVM to JIT compile some code in my game using the LLVM C/C++ API. The JIT compiled functions should be able to effect the physics of the world, which is handled by bullet3.
To do so I need/want to allocate some space on the stack for a btVector3. This class is not simply a four element array of floats, but fulfills many alignment and SIMD properties, so I think manually creating a struct type would be hard and difficult to maintain.

Is there a way to load the definitions from the btVector3.h file into a LLVM module?

I already tried a few things:

• I tried to compile the btVector3.h file into a shared object file using g++. This does not work as neither nm nor readelf recognize the resulting file as a valid object.
• I tried compiling the btVector3.h file into a .bc file using clang++ -c -emit-llvm. This also faces the same issue when using llvm-nm to check the resulting file.
• I also tried to create a proxy file that uses all necessary functions and then compile the proxy. This works for accessing functions like btVector3::x() in the resulting shared object file, but does not allow me to access the definition of the structure itself.

I can work around the issue by simplifying the btVector3 object to a 16bit align float array, and then every time I want to pass a vector to an extern function void foo(btVector3&) I could call a function like:

void _foo(float* p)
{
    btVector3 v(p[0], p[1], p[2]);

    foo(v);

    p[0] = v.x();
    p[1] = v.y();
    p[2] = v.z();
}

which I compile into a shared object file beforehand. This feels bad, but would fix the problem and I could even imagine that the JIT optimizer would inline this function, which could fix the problem.

I would be happy about either a way to directly incorporate the definitions as I have stated in the beginning. I would also be happy about any other work around ideas.