GPU compilers prefer to inline everything, and they try to reuse partial results if they can, so it’s easy to get out of order dependencies in places you might not expect.