> How come that for the same code, Clang shows totally different results than GCC and shows that the basic loop is thirty thousand times faster with -O2 and -O3?

Because clang turns it into `return data+loop_size;`.

In reality of course if you're considering Duff's device, your operations should be much heavier than any of the operations the author tested. And yes, those cases still exist, but people usually consider it cheaper to perpetually throw more compute resources at the problem.