MMX Codes running slower then C++ codes

void tom::add(void* btr)
{

         __declspec(align(8))short* b =(short*)btr;

         int j;


/*
for(j = 0; j < 4; j++)
    {
    /// 1st stage add.
    int s0 = (int)(b[j]     + b[j+3]);
    int s3 = (int)(b[j]     - b[j+3]);
    int s1 = (int)(b[j+1] + b[j+2]);
    int s2 = (int)(b[j+1] - b[j+2]);

    /// 2nd stage add.

    b[j]        = (short)(s0 + s1);
    b[j+8]  = (short)(s0 - s1);
    b[j+4]  = (short)(s2 + (s3 << 1));
    b[j+12] = (short)(s3 - (s2 << 1));
    }//end for j...

   */
    __m64*b1 = (__m64*)b;
        j=0;
    __m64 f0 = _mm_set_pi16(b[j+12],b[j+8],b[j+4],b[j]);
   __m64 f1 = _mm_set_pi16(b[j+13],b[j+9],b[j+5],b[j+1]);
    __m64 f2 = _mm_set_pi16(b[j+14],b[j+10],b[j+6],b[j+2]);
    __m64 f3 = _mm_set_pi16(b[j+15],b[j+11],b[j+7],b[j+3]);
    for(j = 0; j < 4; j+=4)
       {

         // stage one add
       __m64 s0 =_mm_add_pi16(f0,f3);
    __m64 s3 =_mm_sub_pi16(f0,f3);
    __m64 s1 =_mm_add_pi16(f1,f2);
     __m64 s2 =_mm_sub_pi16(f1,f2);
         // stage two add
    *(&b1[j]) =_mm_add_pi16(s0,s1);
    *(&b1[j+2]) =_mm_sub_pi16(s0,s1);
    *(&b1[j+1]) =_mm_add_pi16(s2,_mm_slli_pi16(s3, 1));  
    *(&b1[j+3]) =_mm_sub_pi16(s3,_mm_slli_pi16(s2, 1));
    }
      _mm_empty();
}