gistfile1.txt

#include <emmintrin.h>
#include <stdio.h>

int main(void)
{
   // Attempt to divide by 0x8000. With integer math this is right shift by 15.
   // IEEE floats are laid out like this:
   // [1        - 8        -       23]
   // [Sign bit - Exponent - Mantissa]
   // The value is calculated as: (-1)^sign * (1 + mantissa) * (2 ^ (exponent - 127)).
   // Thus, we'll need to subtract the exponent part by 15. 2 ^ (-15) = 1.0/0x8000.
   static float buf[] __attribute__((aligned(16))) = {0x4000, 0x3000, 0x2000, 0x1000};
   printf("Before: %f, %f, %f, %f\n", buf[0], buf[1], buf[2], buf[3]);

   // Load vector into 128-bit register.
   __m128i reg = _mm_load_si128((__m128i*)buf);

   // We're working on 4 values at a time, 32-bit for each.
   // exp gets the exponent part, mantissa gets sign bit and mantissa itself which we don't care about for now.
   // _mm_set1_epi32() splats a value into a vector. splat(10) = [10, 10, 10, 10].
   __m128i exp = _mm_and_si128(reg, _mm_set1_epi32(0x7f800000));

   __m128i mantissa = _mm_and_si128(reg, _mm_set1_epi32(0x807fffff));

   // Subtract 15 from the exponent (shift by 23 to get to the exponent).
   // Mask again incase we overflow our bounaries.
   exp = _mm_and_si128(_mm_sub_epi32(exp, _mm_set1_epi32(15 << 23)), _mm_set1_epi32(0x7f800000));

   // OR together fixed up exponent and old mantissa/sign.
   _mm_store_ps(buf, (__m128)_mm_or_si128(exp, mantissa));

   printf("After: %f, %f, %f, %f\n", buf[0], buf[1], buf[2], buf[3]);
}