1  /* { dg-do compile } */
       2  /* { dg-options "-O2 -funroll-loops -march=nocona" } */
       3  
       4  #include <mmintrin.h>
       5  
       6  int W;
       7  void f()
       8  {
       9    int j;
      10    int B, C;
      11    unsigned char* S;
      12    __m64 *T = (__m64 *) &W;
      13  
      14    for (j = 0; j < 16; j++, T++)
      15    {
      16      T[0] = T[1] = _mm_set1_pi8(*S);
      17      S += W;
      18    }
      19  
      20    C = 3 * B;
      21  
      22    __m64 E = _mm_set_pi16(3 * B, 3 * B, 3 * B, 5 * B);
      23    __m64 G = _mm_set1_pi16(3 * B);
      24  
      25    for (j = 0; j < 16; j++)
      26    {
      27      __m64 R = _mm_set1_pi16(B + j * C);
      28      R = _m_paddw(R, E);
      29      R = _m_paddw(R, G);
      30      T[0] = _mm_srai_pi16(R, 3);
      31    }
      32  }