The LLVM backend that has been written generates code that directly runs on the Videocore GPU, which also handles the early boot process. Your link is concerned with code that runs on the ARM core, and interfaces with the GPU via the existing code running on the GPU