1  /* PR target/49411 */
       2  /* { dg-do assemble } */
       3  /* { dg-options "-O0 -mf16c -maes -mpclmul" } */
       4  /* { dg-require-effective-target f16c } */
       5  /* { dg-require-effective-target vaes } */
       6  /* { dg-require-effective-target vpclmul } */
       7  
       8  #include <x86intrin.h>
       9  
      10  __m128i i1, i2, i3, i4;
      11  __m128 a1, a2, a3, a4;
      12  __m128d d1, d2, d3, d4;
      13  __m256i l1, l2, l3, l4;
      14  __m256 b1, b2, b3, b4;
      15  __m256d e1, e2, e3, e4;
      16  __m64 m1, m2, m3, m4;
      17  int k1, k2, k3, k4;
      18  float f1, f2, f3, f4;
      19  
      20  void
      21  test8bit (void)
      22  {
      23    i1 = _mm_cmpistrm (i2, i3, 255);
      24    k1 = _mm_cmpistri (i2, i3, 255);
      25    k1 = _mm_cmpistra (i2, i3, 255);
      26    k1 = _mm_cmpistrc (i2, i3, 255);
      27    k1 = _mm_cmpistro (i2, i3, 255);
      28    k1 = _mm_cmpistrs (i2, i3, 255);
      29    k1 = _mm_cmpistrz (i2, i3, 255);
      30    i1 = _mm_cmpestrm (i2, k2, i3, k3, 255);
      31    k1 = _mm_cmpestri (i2, k2, i3, k3, 255);
      32    k1 = _mm_cmpestra (i2, k2, i3, k3, 255);
      33    k1 = _mm_cmpestrc (i2, k2, i3, k3, 255);
      34    k1 = _mm_cmpestro (i2, k2, i3, k3, 255);
      35    k1 = _mm_cmpestrs (i2, k2, i3, k3, 255);
      36    k1 = _mm_cmpestrz (i2, k2, i3, k3, 255);
      37    b1 = _mm256_blend_ps (b2, b3, 255);
      38    k1 = _cvtss_sh (f1, 255);
      39    i1 = _mm256_cvtps_ph (b2, 255);
      40    b1 = _mm256_dp_ps (b2, b3, 255);
      41    e1 = _mm256_permute2f128_pd (e2, e3, 255);
      42    b1 = _mm256_permute2f128_ps (b2, b3, 255);
      43    l1 = _mm256_permute2f128_si256 (l2, l3, 255);
      44    b1 = _mm256_permute_ps (b2, 255);
      45    i1 = _mm_aeskeygenassist_si128 (i2, 255);
      46    i1 = _mm_blend_epi16 (i2, i3, 255);
      47    i1 = _mm_clmulepi64_si128 (i2, i3, 255);
      48    i1 = _mm_cvtps_ph (a1, 255);
      49    d1 = _mm_dp_pd (d2, d3, 255);
      50    a1 = _mm_dp_ps (a2, a3, 255);
      51    a1 = _mm_insert_ps (a2, a3, 255);
      52    i1 = _mm_mpsadbw_epu8 (i2, i3, 255);
      53    a1 = _mm_permute_ps (a2, 255);
      54    i1 = _mm_slli_si128 (i2, 255);
      55    i1 = _mm_srli_si128 (i2, 255);
      56  }
      57  
      58  void
      59  test5bit (void)
      60  {
      61    d1 = _mm_cmp_sd (d2, d3, 31);
      62    a1 = _mm_cmp_ss (a2, a3, 31);
      63    d1 = _mm_cmp_pd (d2, d3, 31);
      64    a1 = _mm_cmp_ps (a2, a3, 31);
      65    e1 = _mm256_cmp_pd (e2, e3, 31);
      66    b1 = _mm256_cmp_ps (b2, b3, 31);
      67  }
      68  
      69  void
      70  test4bit (void)
      71  {
      72    d1 = _mm_round_pd (d2, 15);
      73    d1 = _mm_round_sd (d2, d3, 15);
      74    a1 = _mm_round_ps (a2, 15);
      75    a1 = _mm_round_ss (a2, a2, 15);
      76    a1 = _mm_blend_ps (a2, a3, 15);
      77    e1 = _mm256_blend_pd (e2, e3, 15);
      78    e1 = _mm256_round_pd (e2, 15);
      79    b1 = _mm256_round_ps (b2, 15);
      80  }
      81  
      82  void
      83  test2bit (void)
      84  {
      85    d1 = _mm_blend_pd (d2, d3, 3);
      86  }
      87  
      88  void
      89  test1bit (void)
      90  {
      91    d1 = _mm256_extractf128_pd (e2, 1);
      92    a1 = _mm256_extractf128_ps (b2, 1);
      93    i1 = _mm256_extractf128_si256 (l2, 1);
      94    e1 = _mm256_insertf128_pd (e2, d1, 1);
      95    b1 = _mm256_insertf128_ps (b2, a1, 1);
      96    l1 = _mm256_insertf128_si256 (l2, i1, 1);
      97  }