arithmetic-with-high-word-output.c

#define LO32(x) (x&((UINT64_C(1)<<32)-1))
#define HI32(x) (x >> 32)

// Returns the high 64-bits of the 128-bit result.
// the `result` variable specifies the address of the low 64-bits
static uint64_t add_u64(uint64_t *result, uint64_t a, uint64_t b)
{
  uint64_t al = LO32(a);
  uint64_t ah = HI32(a);
  uint64_t bl = LO32(b);
  uint64_t bh = HI32(b);

  uint64_t r0 = LO32(al+bl);
  uint64_t r1 = HI32(al+bl) + LO32(ah+bh);
  *result = r0 + (r1 << 32);

  return HI32(ah+bh);
}

// Returns the high 64-bits of the 128-bit result
// the `result` is the address of the variable that receives the low 64 bits
static uint64_t mul_u64(uint64_t *result, uint64_t a, uint64_t b)
{
  uint64_t al = LO32(a);
  uint64_t bl = LO32(b);
  
  uint64_t tmp = (al * bl);
  uint64_t r3 = LO32(tmp);
  uint64_t k = HI32(tmp);
  
  uint64_t ah = HI32(a);
  uint64_t bh = HI32(b);

  tmp = (ah * bl) + k;
  k = LO32(tmp);
  uint64_t r1 = HI32(tmp);
  
  tmp = (al * bh) + k;
  k = HI32(tmp);
  
  *result = (tmp << 32) + r3;
  return (ah * bh) + r1 + k;
}


#undef LO32
#undef HI32