I just meant that the compiler in your example is computing x / 1000000 as follows: *q = (x * 1125899907) >> 50; This works out OK because there's a 32x32-

I just meant that the compiler in your example is computing x / 1000000 as follows: *q = (x * 1125899907) >> 50; This works out OK because there's a 32x32->64 instruction, arm64's umull in this case. If it didn't have one, and it had a 32/32->32 division instruction, it might have emitted that instead of hand-rolling the same scheme using 32x32->32 multiplications. Note that it's technically an arithmetic approximation that just happens to work for all possible values of x, since obviously 1<<50 is only divisible by 1<<(0 to 50).