Very Interesting! I guess I would still expect the optimization to use the library implementation of memcpy to be invalid when interacting with memory mapped IO. In this case the issue was the size of writes being issued (byte vs word), but you can easily imagine other issues that would not matter something like memcpy. I would have thought volatile might have signaled this, even if the issue here is not related to values being cached in registers. Does zig have a version of volatile?