1  /* { dg-do run } */
       2  /* { dg-require-effective-target avx } */
       3  /* { dg-options "-O2 -mavx" } */
       4  
       5  #include "avx-check.h"
       6  
       7  static void
       8  avx_test (void)
       9  {
      10    __m256i src;
      11  #ifdef __x86_64__
      12    char reg_save[16][32];
      13    char d[16][32];
      14  #else
      15    char reg_save[8][32];
      16    char d[8][32];
      17  #endif  
      18  
      19    int s[8] = {1, 2, 3, 4, 5, 6, 7, 8};
      20  
      21    __builtin_memset (d, 0, sizeof d);
      22    __builtin_memset (reg_save, -1, sizeof reg_save);
      23  
      24    src = _mm256_loadu_si256 ((__m256i*) s);
      25  
      26    _mm256_zeroall (); 
      27  
      28    __asm__ __volatile__ ("vmovdqu %%ymm0,%0":"=m"(reg_save[0]));
      29    __asm__ __volatile__ ("vmovdqu %%ymm1,%0":"=m"(reg_save[1]));
      30    __asm__ __volatile__ ("vmovdqu %%ymm2,%0":"=m"(reg_save[2]));
      31    __asm__ __volatile__ ("vmovdqu %%ymm3,%0":"=m"(reg_save[3]));
      32    __asm__ __volatile__ ("vmovdqu %%ymm4,%0":"=m"(reg_save[4]));
      33    __asm__ __volatile__ ("vmovdqu %%ymm5,%0":"=m"(reg_save[5]));
      34    __asm__ __volatile__ ("vmovdqu %%ymm6,%0":"=m"(reg_save[6]));
      35    __asm__ __volatile__ ("vmovdqu %%ymm7,%0":"=m"(reg_save[7]));
      36  #ifdef __x86_64__
      37    __asm__ __volatile__ ("vmovdqu %%ymm8,%0":"=m"(reg_save[8]));
      38    __asm__ __volatile__ ("vmovdqu %%ymm9,%0":"=m"(reg_save[9]));
      39    __asm__ __volatile__ ("vmovdqu %%ymm10,%0":"=m"(reg_save[10]));
      40    __asm__ __volatile__ ("vmovdqu %%ymm11,%0":"=m"(reg_save[11]));
      41    __asm__ __volatile__ ("vmovdqu %%ymm12,%0":"=m"(reg_save[12]));
      42    __asm__ __volatile__ ("vmovdqu %%ymm13,%0":"=m"(reg_save[13]));
      43    __asm__ __volatile__ ("vmovdqu %%ymm14,%0":"=m"(reg_save[14]));
      44    __asm__ __volatile__ ("vmovdqu %%ymm15,%0":"=m"(reg_save[15]));
      45  #endif
      46  
      47    if (__builtin_memcmp (reg_save, d, sizeof d))
      48      abort ();
      49  
      50    _mm256_storeu_si256 ((__m256i*) d, src);
      51  
      52  }