LLVM optimizes IR before turning it into BPF instructions. Sometimes the
optimization might cut off important code if the IR was not built
correctly. -d is not enough to find these issues, therefore this patch
introduces more debug levels to also print the raw IR before it gets
optimized.