Line data    Source code

       1             : /*
       2             :   return 64 bit x / 1000
       3             :   faster than the normal gcc implementation using by about 3x
       4             :   With thanks to https://0x414b.com/2021/04/16/arm-division.html
       5             :   and https://stackoverflow.com/questions/74765410/multiply-two-uint64-ts-and-store-result-to-uint64-t-doesnt-seem-to-work
       6             : */
       7 13358072142 : static inline uint64_t uint64_div1000(uint64_t x)
       8             : {
       9 13358072142 :     x >>= 3U;
      10 13358072142 :     uint64_t a_lo = (uint32_t)x;
      11 13358072142 :     uint64_t a_hi = x >> 32;
      12 13358072142 :     const uint64_t b_lo = 0xe353f7cfU;
      13 13358072142 :     const uint64_t b_hi = 0x20c49ba5U;
      14             : 
      15 13358072142 :     uint64_t a_x_b_hi = a_hi * b_hi;
      16 13358072142 :     uint64_t a_x_b_mid = a_hi * b_lo;
      17 13358072142 :     uint64_t b_x_a_mid = b_hi * a_lo;
      18 13358072142 :     uint32_t a_x_b_lo = (a_lo * b_lo)>>32;
      19             : 
      20             :     // 64-bit product + two 32-bit values
      21 13358072142 :     uint64_t middle = a_x_b_mid + a_x_b_lo + (uint32_t)b_x_a_mid;
      22             : 
      23             :     // 64-bit product + two 32-bit values
      24 13358072142 :     uint64_t r = a_x_b_hi + (middle >> 32) + (b_x_a_mid >> 32);
      25 13358072142 :     return r >> 4U;
      26             : }