1  /* { dg-do run } */
       2  /* { dg-options "-O3 -mpower8-vector -Wno-psabi" } */
       3  /* { dg-require-effective-target p8vector_hw } */
       4  
       5  #ifndef CHECK_H
       6  #define CHECK_H "ssse3-check.h"
       7  #endif
       8  
       9  #ifndef TEST
      10  #define TEST ssse3_test
      11  #endif
      12  
      13  #include CHECK_H
      14  
      15  #include "ssse3-vals.h"
      16  
      17  #include <tmmintrin.h>
      18  
      19  #ifndef __AVX__
      20  /* Test the 64-bit form */
      21  static void
      22  ssse3_test_phsubd (__m64 *i1, __m64 *i2, __m64 *r)
      23  {
      24    *r = _mm_hsub_pi32 (*i1, *i2);
      25    _mm_empty ();
      26  }
      27  #endif
      28  
      29  /* Test the 128-bit form */
      30  static void
      31  ssse3_test_phsubd128 (__m128i *i1, __m128i *i2, __m128i *r)
      32  {
      33    /* Assumes incoming pointers are 16-byte aligned */
      34    *(__m128i *) r = _mm_hsub_epi32 (*i1, *i2);
      35  }
      36  
      37  /* Routine to manually compute the results */
      38  static void
      39  compute_correct_result (int *i1, int *i2, int *r)
      40  {
      41    int i;
      42  
      43    for (i = 0; i < 2; i++)
      44      r[i] = i1[2 * i] - i1[2 * i + 1];
      45    for (i = 0; i < 2; i++)
      46      r[i + 2] = i2[2 * i] - i2[2 * i + 1];
      47  }
      48  
      49  static void
      50  TEST (void)
      51  {
      52    int i;
      53    union data r __attribute__ ((aligned(16)));
      54    union data ck;
      55    int fail = 0;
      56  
      57    for (i = 0; i < ARRAY_SIZE (vals) - 1; i++)
      58      {
      59        /* Manually compute the result */
      60        compute_correct_result (&vals[i + 0].w[0], &vals[i + 1].w[0], &ck.w[0]);
      61  
      62  #ifndef __AVX__
      63        /* Run the 64-bit tests */
      64        ssse3_test_phsubd (&vals[i + 0].ll[0], &vals[i + 0].ll[1], &r.ll[0]);
      65        ssse3_test_phsubd (&vals[i + 1].ll[0], &vals[i + 1].ll[1], &r.ll[1]);
      66        fail += chk_128 (ck.m[0], r.m[0]);
      67  #endif
      68  
      69        /* Run the 128-bit tests */
      70        ssse3_test_phsubd128 (&vals[i + 0].m[0], &vals[i + 1].m[0], &r.m[0]);
      71        fail += chk_128 (ck.m[0], r.m[0]);
      72      }
      73  
      74    if (fail != 0)
      75      abort ();
      76  }