jhuber6 wrote: At a minimum we'd want symbol address and size, then you can add stuff to that like whether or not it's a kernel or a global, you can probably refer to what CUDA / HSA provides as well. https://github.com/llvm/llvm-project/pull/142632