1  /* PR rtl-optimization/89445 */
       2  /* { dg-do run { target { avx512f && mmap } } } */
       3  /* { dg-options "-O2 -mavx512f" } */
       4  
       5  #include "avx512f-check.h"
       6  
       7  #include <sys/mman.h>
       8  #ifndef MAP_ANONYMOUS
       9  #define MAP_ANONYMOUS MAP_ANON
      10  #endif
      11  #ifndef MAP_ANON
      12  #define MAP_ANON 0
      13  #endif
      14  #ifndef MAP_FAILED
      15  #define MAP_FAILED ((void *)-1)
      16  #endif
      17  
      18  __attribute__ ((noipa))
      19  void daxpy (unsigned long n, double a, double const *__restrict x,
      20  	    double *__restrict y)
      21  {
      22    const __m512d v_a = _mm512_broadcastsd_pd (_mm_set_sd (a));
      23    const __mmask16 final = (1U << (n % 8u)) - 1;
      24    __mmask16 mask = 65535u;
      25    unsigned long i;
      26    for (i = 0; i < n * sizeof (double); i += 8 * sizeof (double))
      27      {
      28        if (i + 8 * sizeof (double) > n * sizeof (double))
      29  	mask = final;
      30        __m512d v_x = _mm512_maskz_loadu_pd (mask, (char const *) x + i);
      31        __m512d v_y = _mm512_maskz_loadu_pd (mask, (char const *) y + i);
      32        __m512d tmp = _mm512_fmadd_pd (v_x, v_a, v_y);
      33        _mm512_mask_storeu_pd ((char *) y + i, mask, tmp);
      34      }
      35  }
      36  
      37  static const double x[] = { 1, 2, 3, 4 };
      38  
      39  static void
      40  avx512f_test (void)
      41  {
      42    char *ptr
      43      = (char *) mmap (NULL, 2 * 4096, PROT_READ | PROT_WRITE,
      44  		     MAP_PRIVATE | MAP_ANONYMOUS, -1, 0);
      45    if (ptr == MAP_FAILED)
      46      return;
      47  
      48    munmap (ptr + 4096, 4096);
      49    double *y = (double *) (ptr + 4096 - sizeof (x));
      50    __builtin_memcpy (y, x, sizeof (x));
      51    daxpy (sizeof (x) / sizeof (x[0]), 1.0, x, y);
      52    if (y[0] != 2.0 || y[1] != 4.0 || y[2] != 6.0 || y[3] != 8.0)
      53      abort ();
      54  }