optimizing_all_64_bit.cpp

#if defined(__x86_64__)
inline uint64_t Divide128Div64To64(uint64_t high, uint64_t low,
                                   uint64_t divisor, uint64_t* remainder) {
  uint64_t result;
  __asm__("divq %[v]"
          : "=a"(result), "=d"(*remainder) // Ouput parametrs, =a for rax, =d for rdx, [v] is an
                                           // alias for divisor, input paramters "a" and "d" for low and high.
          : [v] "r"(divisor), "a"(low), "d"(high));
  return result;
}
#endif

tu_int __udivmodti4(tu_int dividend, tu_int divisor, tu_int* remainder) {
…

#if defined(__x86_64__)
  if (divisor.high == 0) {
    remainder->high = 0;
    if (dividend.high < divisor) {
      uint64_t quotient =
          Divide128Div64To64(dividend.high, dividend.low,
                             divisor.low, &remainder->low);
        return quotient;
    } else {
      tu_int quotient;
      quotient.high = Divide128Div64To64(0, dividend.high, divisor.low,
                                         &dividend.high);
      quotient.low = Divide128Div64To64(dividend.high, dividend.low,
                                        divisor.low, &remainder->low);
      return quotient;
    }
  }
#endif
...
}