1  /* { dg-do run } */
       2  /* { dg-options "-O2 -mavx2 -ftree-vectorize -fvect-cost-model=unlimited -mprefer-vector-width=256" } */
       3  /* { dg-require-effective-target avx2 } */
       4  
       5  #include "avx2-check.h"
       6  #include <string.h>
       7  #include "pr106010-3a.c"
       8  
       9  void
      10  avx2_test (void)
      11  {
      12    _Complex double* pd_src = (_Complex double*) malloc (32);
      13    _Complex double* pd_dst = (_Complex double*) malloc (32);
      14    _Complex double* pd_exp = (_Complex double*) malloc (32);
      15    _Complex float* ps_src = (_Complex float*) malloc (32);
      16    _Complex float* ps_dst = (_Complex float*) malloc (32);
      17    _Complex float* ps_exp = (_Complex float*) malloc (32);
      18    _Complex long long* epi64_src = (_Complex long long*) malloc (32);
      19    _Complex long long* epi64_dst = (_Complex long long*) malloc (32);
      20    _Complex long long* epi64_exp = (_Complex long long*) malloc (32);
      21    _Complex int* epi32_src = (_Complex int*) malloc (32);
      22    _Complex int* epi32_dst = (_Complex int*) malloc (32);
      23    _Complex int* epi32_exp = (_Complex int*) malloc (32);
      24    _Complex short* epi16_src = (_Complex short*) malloc (32);
      25    _Complex short* epi16_dst = (_Complex short*) malloc (32);
      26    _Complex short* epi16_exp = (_Complex short*) malloc (32);
      27    _Complex char* epi8_src = (_Complex char*) malloc (32);
      28    _Complex char* epi8_dst = (_Complex char*) malloc (32);
      29    _Complex char* epi8_exp = (_Complex char*) malloc (32);
      30    char* p = (char* ) malloc (32);
      31    char* q = (char* ) malloc (32);
      32  
      33    __builtin_memset (pd_dst, 0, 32);
      34    __builtin_memset (ps_dst, 0, 32);
      35    __builtin_memset (epi64_dst, 0, 32);
      36    __builtin_memset (epi32_dst, 0, 32);
      37    __builtin_memset (epi16_dst, 0, 32);
      38    __builtin_memset (epi8_dst, 0, 32);
      39  
      40    for (int i = 0; i != 32; i++)
      41      p[i] = i;
      42    __builtin_memcpy (pd_src, p, 32);
      43    __builtin_memcpy (ps_src, p, 32);
      44    __builtin_memcpy (epi64_src, p, 32);
      45    __builtin_memcpy (epi32_src, p, 32);
      46    __builtin_memcpy (epi16_src, p, 32);
      47    __builtin_memcpy (epi8_src, p, 32);
      48  
      49    for (int i = 0; i != 16; i++)
      50      {
      51        p[i] = i + 16;
      52        p[i + 16] = i;
      53      }
      54    __builtin_memcpy (pd_exp, p, 32);
      55    __builtin_memcpy (epi64_exp, p, 32);
      56  
      57    for (int i = 0; i != 8; i++)
      58      {
      59        p[i] = i + 8;
      60        p[i + 8] = i;
      61        p[i + 16] = i + 24;
      62        p[i + 24] = i + 16;
      63        q[i] = i + 24;
      64        q[i + 8] = i + 16;
      65        q[i + 16] = i + 8;
      66        q[i + 24] = i;
      67      }
      68    __builtin_memcpy (ps_exp, p, 32);
      69    __builtin_memcpy (epi32_exp, q, 32);
      70  
      71  
      72    for (int i = 0; i != 4; i++)
      73      {
      74        q[i] = i + 28;
      75        q[i + 4] = i + 24;
      76        q[i + 8] = i + 20;
      77        q[i + 12] = i + 16;
      78        q[i + 16] = i + 12;
      79        q[i + 20] = i + 8;
      80        q[i + 24] = i + 4;
      81        q[i + 28] = i;
      82      }
      83    __builtin_memcpy (epi16_exp, q, 32);
      84  
      85    for (int i = 0; i != 2; i++)
      86      {
      87        q[i] = i + 14;
      88        q[i + 2] = i + 12;
      89        q[i + 4] = i + 10;
      90        q[i + 6] = i + 8;
      91        q[i + 8] = i + 6;
      92        q[i + 10] = i + 4;
      93        q[i + 12] = i + 2;
      94        q[i + 14] = i;
      95        q[i + 16] = i + 30;
      96        q[i + 18] = i + 28;
      97        q[i + 20] = i + 26;
      98        q[i + 22] = i + 24;
      99        q[i + 24] = i + 22;
     100        q[i + 26] = i + 20;
     101        q[i + 28] = i + 18;
     102        q[i + 30] = i + 16;
     103      }
     104    __builtin_memcpy (epi8_exp, q, 32);
     105  
     106    foo_pd (pd_dst, pd_src);
     107    foo_ps (ps_dst, ps_src);
     108    foo_epi64 (epi64_dst, epi64_src);
     109    foo_epi32 (epi32_dst, epi32_src);
     110    foo_epi16 (epi16_dst, epi16_src);
     111    foo_epi8 (epi8_dst, epi8_src);
     112    if (__builtin_memcmp (pd_dst, pd_exp, 32) != 0)
     113      __builtin_abort ();
     114    if (__builtin_memcmp (ps_dst, ps_exp, 32) != 0)
     115      __builtin_abort ();
     116    if (__builtin_memcmp (epi64_dst, epi64_exp, 32) != 0)
     117      __builtin_abort ();
     118    if (__builtin_memcmp (epi32_dst, epi32_exp, 32) != 0)
     119      __builtin_abort ();
     120    if (__builtin_memcmp (epi16_dst, epi16_exp, 32) != 0)
     121      __builtin_abort ();
     122    if (__builtin_memcmp (epi8_dst, epi8_exp, 32) != 0)
     123      __builtin_abort ();
     124  
     125    return;
     126  }