1  /* { dg-do run } */
       2  /* { dg-require-effective-target sse3 } */
       3  /* { dg-options "-O2 -msse3 -mfpmath=sse" } */
       4  
       5  #ifndef CHECK_H
       6  #define CHECK_H "sse3-check.h"
       7  #endif
       8  
       9  #ifndef TEST
      10  #define TEST sse3_test
      11  #endif
      12  
      13  #include CHECK_H
      14  
      15  #include <pmmintrin.h>
      16  
      17  static void
      18  sse3_test_movddup_mem (double *i1, double *r)
      19  {
      20    __m128d t1 = _mm_loaddup_pd (i1);
      21  
      22    _mm_storeu_pd (r, t1);
      23  }
      24  
      25  static double cnst1 [2] = {1.0, 1.0};
      26  
      27  static void
      28  sse3_test_movddup_reg (double *i1, double *r)
      29  {
      30    __m128d t1 = _mm_loadu_pd (i1);
      31    __m128d t2 = _mm_loadu_pd (&cnst1[0]);
      32  
      33    t1  = _mm_mul_pd (t1, t2);
      34    t2  = _mm_movedup_pd (t1);
      35  
      36    _mm_storeu_pd (r, t2);
      37  }
      38  
      39  static void
      40  sse3_test_movddup_reg_subsume_unaligned (double *i1, double *r)
      41  {
      42    __m128d t1 = _mm_loadu_pd (i1);
      43    __m128d t2 = _mm_movedup_pd (t1);
      44  
      45    _mm_storeu_pd (r, t2);
      46  }
      47  
      48  static void
      49  sse3_test_movddup_reg_subsume_ldsd (double *i1, double *r)
      50  {
      51    __m128d t1 = _mm_load_sd (i1);
      52    __m128d t2 = _mm_movedup_pd (t1);
      53  
      54    _mm_storeu_pd (r, t2);
      55  }
      56  
      57  static void
      58  sse3_test_movddup_reg_subsume (double *i1, double *r)
      59  {
      60    __m128d t1 = _mm_load_pd (i1);
      61    __m128d t2 = _mm_movedup_pd (t1);
      62  
      63    _mm_storeu_pd (r, t2);
      64  }
      65  
      66  static int
      67  chk_pd (double *v1, double *v2)
      68  {
      69    int i;
      70    int n_fails = 0;
      71  
      72    for (i = 0; i < 2; i++)
      73      if (v1[i] != v2[i])
      74        n_fails += 1;
      75  
      76    return n_fails;
      77  }
      78  
      79  static double p1[2] __attribute__ ((aligned(16)));
      80  static double p2[2];
      81  static double ck[2];
      82  
      83  static double vals[80] =
      84    {
      85      100.0,  200.0, 300.0, 400.0, 5.0, -1.0, .345, -21.5,
      86      1100.0, 0.235, 321.3, 53.40, 0.3, 10.0, 42.0, 32.52,
      87      32.6,   123.3, 1.234, 2.156, 0.1, 3.25, 4.75, 32.44,
      88      12.16,  52.34, 64.12, 71.13, -.1, 2.30, 5.12, 3.785,
      89      541.3,  321.4, 231.4, 531.4, 71., 321., 231., -531.,
      90      23.45,  23.45, 23.45, 23.45, 23.45, 23.45, 23.45, 23.45,
      91      23.45,  -1.43, -6.74, 6.345, -20.1, -20.1, -40.1, -40.1,
      92      1.234,  2.345, 3.456, 4.567, 5.678, 6.789, 7.891, 8.912,
      93      -9.32,  -8.41, -7.50, -6.59, -5.68, -4.77, -3.86, -2.95,
      94      9.32,  8.41, 7.50, 6.59, -5.68, -4.77, -3.86, -2.95
      95    };
      96  
      97  static void
      98  TEST (void)
      99  {
     100    int i;
     101    int fail = 0;
     102  
     103    for (i = 0; i < 80; i += 1)
     104      {
     105        p1[0] = vals[i+0];
     106  
     107        ck[0] = p1[0];
     108        ck[1] = p1[0];
     109  
     110        sse3_test_movddup_mem (p1, p2);
     111  
     112        fail += chk_pd (ck, p2);
     113  
     114        sse3_test_movddup_reg (p1, p2);
     115  
     116        fail += chk_pd (ck, p2);
     117  
     118        sse3_test_movddup_reg_subsume (p1, p2);
     119  
     120        fail += chk_pd (ck, p2);
     121  
     122        sse3_test_movddup_reg_subsume_unaligned (p1, p2);
     123  
     124        fail += chk_pd (ck, p2);
     125  
     126        sse3_test_movddup_reg_subsume_ldsd (p1, p2);
     127  
     128        fail += chk_pd (ck, p2);
     129      }
     130  
     131    if (fail != 0)
     132      abort ();
     133  }