1  /* { dg-do run } */
       2  /* { dg-require-effective-target sse4 } */
       3  /* { dg-options "-O2 -msse4.1" } */
       4  
       5  #ifndef CHECK_H
       6  #define CHECK_H "sse4_1-check.h"
       7  #endif
       8  
       9  #ifndef TEST
      10  #define TEST sse4_1_test
      11  #endif
      12  
      13  #include CHECK_H
      14  
      15  #include <smmintrin.h>
      16  
      17  #include <string.h>
      18  
      19  #define lmskN  0x00
      20  #define lmsk0  0x01
      21  #define lmsk1  0x02
      22  #define lmsk01 0x03
      23  
      24  #define hmskA  0x30
      25  #define hmsk0  0x10
      26  #define hmsk1  0x20
      27  #define hmsk01 0x30
      28  #define hmskN  0x00
      29  
      30  #ifndef HIMASK
      31  #define HIMASK hmskA
      32  #endif
      33  
      34  #ifndef LOMASK
      35  #define LOMASK lmsk01
      36  #endif
      37  
      38  static void
      39  TEST (void)
      40  {
      41    union
      42      {
      43        __m128d x;
      44        double d[2];
      45      } val1[4], val2[4], res[4], chk[4];
      46    int i, j;
      47    double tmp;
      48  
      49    for (i = 0; i < 4; i++)
      50      {
      51        val1[i].d [0] = 2.;
      52        val1[i].d [1] = 3.;
      53  
      54        val2[i].d [0] = 10.;
      55        val2[i].d [1] = 100.;
      56  
      57        tmp = 0.;
      58        for (j = 0; j < 2; j++)
      59  	if ((HIMASK & (0x10 << j)))
      60  	  tmp += val1[i].d [j] * val2[i].d [j];
      61  
      62        for (j = 0; j < 2; j++)
      63          if ((LOMASK & (1 << j)))
      64  	  chk[i].d[j] = tmp;
      65      }
      66  
      67    for (i = 0; i < 4; i++)
      68      {
      69        res[i].x = _mm_dp_pd (val1[i].x, val2[i].x, HIMASK | LOMASK); 
      70        if (memcmp (&res[i], &chk[i], sizeof (chk[i])))
      71  	abort ();
      72      }
      73  }