gcc/testsuite/gcc.dg/vect/vect-outer-simd-2.c

   1 /* { dg-require-effective-target vect_simd_clones } */
   2 /* { dg-additional-options "-fopenmp-simd -ffast-math" } */
   3 #include <stdlib.h>
   4 #include "tree-vect.h"
   5 #define N 64
   6
   7 float *px, *py;
   8 float *tx, *ty;
   9 float *x1, *z1, *t1, *t2;
  10
  11 static void inline bar (const float cx, float cy,
  12                          float *vx, float *vy)
  13 {
  14   int j;
  15     for (j = 0; j < N; ++j)
  16     {
  17         const float dx  = cx - px[j];
  18         const float dy  = cy - py[j];
  19         *vx               -= dx * tx[j];
  20         *vy               -= dy * ty[j];
  21     }
  22 }
  23
  24 __attribute__((noinline, noclone)) void foo1 (int n)
  25 {
  26   int i;
  27 #pragma omp simd
  28   for (i=0; i<n; i++)
  29     bar (px[i], py[i], x1+i, z1+i);
  30 }
  31
  32 __attribute__((noinline, noclone)) void foo2 (int n)
  33 {
  34   volatile int i;
  35   for (i=0; i<n; i++)
  36     bar (px[i], py[i], x1+i, z1+i);
  37 }
  38
  39
  40 int main ()
  41 {
  42   float *X = (float*)malloc (N * 8 * sizeof (float));
  43   int i;
  44   int n = N - 1;
  45   check_vect ();
  46   px = &X[0];
  47   py = &X[N * 1];
  48   tx = &X[N * 2];
  49   ty = &X[N * 3];
  50   x1 = &X[N * 4];
  51   z1 = &X[N * 5];
  52   t1 = &X[N * 6];
  53   t2 = &X[N * 7];
  54
  55   for (i=0; i<N; i++)
  56     {
  57       px[i] = (float) (i+2);
  58       tx[i] = (float) (i+1);
  59       py[i] = (float) (i+4);
  60       ty[i] = (float) (i+3);
  61       x1[i] = z1[i] = 1.0f;
  62     }
  63   foo1 (n);  /* vector variant.  */
  64   for (i=0; i<N;i++)
  65     {
  66       t1[i] = x1[i]; x1[i] = 1.0f;
  67       t2[i] = z1[i]; z1[i] = 1.0f;
  68     }
  69   foo2 (n);  /* scalar variant.  */
  70 #pragma GCC novector
  71   for (i=0; i<N; i++)
  72     if (x1[i] != t1[i] || z1[i] != t2[i])
  73       abort ();
  74   return 0;
  75 }
  76 /* { dg-final { scan-tree-dump "OUTER LOOP VECTORIZED" "vect" } } */