[SYCL] Linking in static library with SYCL code should not require -Wl,--whole-archive

[dvrogozh]

With 848eba7, Ubuntu 24.04, PVC. Consider the following trivial example:

• https://github.com/dvrogozh/cmake-with-sycl/tree/fsycl-link (mind fsycl-link branch)

This project builds main-static application from main.cpp and lib.cpp in the following way:

1. main.cpp is regular C++ source and built with g++
2. lib.cpp is SYCL source and built with dpclang++ into liblib-static.a
3. main-static application is linked by g++

Review build-no-whole-archive.sh and build-whole-archive.sh scripts for the details.

main-static built without -Wl,--whole-archive will fail with:

$ ./main-static
main-static: /home/dvrogozh/git/intel-llvm/sycl/source/detail/program_manager/program_manager.cpp:1620: sycl::_V1::detail::DeviceKernelInfo& sycl::_V1::detail::ProgramManager::getDeviceKernelInfo(const sycl::_V1::detail::compile_time_kernel_info_v1::CompileTimeKernelInfoTy&): Assertion `It != m_DeviceKernelInfoMap.end()' failed.
Aborted (core dumped)

Can compiler be changed to not require -Wl,--whole-archive in this case? Note that NVidia nvcc does not require that in a similar scenario.

CC: @tahonermann, @sarnex

[sarnex]

I'm taking a look

[sarnex]

It looks like that since the prelinked object file doesn't provide any symbols needed by other object files, it never gets linked in, so the global constructor registering the device image never runs, so runtime fails. CUDA seems to handle this through a __cudaRegisterLinkedBinary function, I'll try to implement something similar for SYCL.