libgomp/testsuite/libgomp.c/scan-17.c

   1 /* { dg-require-effective-target size32plus } */
   2 /* { dg-additional-options "-O2 -fopenmp -fdump-tree-vect-details" } */
   3 /* { dg-additional-options "-msse2" { target sse2_runtime } } */
   4 /* { dg-additional-options "-mavx" { target avx_runtime } } */
   5 /* { dg-final { scan-tree-dump-times "vectorized \[2-6] loops" 2 "vect" { target sse2_runtime } } } */
   6
   7 extern void abort (void);
   8 float r = 1.0f, a[1024], b[1024];
   9
  10 __attribute__((noipa)) void
  11 foo (float *a, float *b)
  12 {
  13   #pragma omp for simd reduction (inscan, *:r)
  14   for (int i = 0; i < 1024; i++)
  15     {
  16       b[i] = r;
  17       #pragma omp scan exclusive(r)
  18       r *= a[i];
  19     }
  20 }
  21
  22 __attribute__((noipa)) float
  23 bar (void)
  24 {
  25   float s = -__builtin_inff ();
  26   #pragma omp parallel for simd reduction (inscan, max:s)
  27   for (int i = 0; i < 1024; i++)
  28     {
  29       b[i] = s;
  30       #pragma omp scan exclusive(s)
  31       s = s > a[i] ? s : a[i];
  32     }
  33   return s;
  34 }
  35
  36 int
  37 main ()
  38 {
  39   float s = 1.0f;
  40   for (int i = 0; i < 1024; ++i)
  41     {
  42       if (i < 80)
  43         a[i] = (i & 1) ? 0.25f : 0.5f;
  44       else if (i < 200)
  45         a[i] = (i % 3) == 0 ? 2.0f : (i % 3) == 1 ? 4.0f : 1.0f;
  46       else if (i < 280)
  47         a[i] = (i & 1) ? 0.25f : 0.5f;
  48       else if (i < 380)
  49         a[i] = (i % 3) == 0 ? 2.0f : (i % 3) == 1 ? 4.0f : 1.0f;
  50       else
  51         switch (i % 6)
  52           {
  53           case 0: a[i] = 0.25f; break;
  54           case 1: a[i] = 2.0f; break;
  55           case 2: a[i] = -1.0f; break;
  56           case 3: a[i] = -4.0f; break;
  57           case 4: a[i] = 0.5f; break;
  58           case 5: a[i] = 1.0f; break;
  59           default: a[i] = 0.0f; break;
  60           }
  61       b[i] = -19.0f;
  62       asm ("" : "+g" (i));
  63     }
  64   #pragma omp parallel
  65   foo (a, b);
  66   if (r * 16384.0f != 0.125f)
  67     abort ();
  68   float m = -175.25f;
  69   for (int i = 0; i < 1024; ++i)
  70     {
  71       if (b[i] != s)
  72         abort ();
  73       else
  74         b[i] = -231.75f;
  75       s *= a[i];
  76       a[i] = m - ((i % 3) == 1 ? 2.0f : (i % 3) == 2 ? 4.0f : 0.0f);
  77       m += 0.75f;
  78     }
  79   if (bar () != 592.0f)
  80     abort ();
  81   s = -__builtin_inff ();
  82   for (int i = 0; i < 1024; ++i)
  83     {
  84       if (b[i] != s)
  85         abort ();
  86       if (s < a[i])
  87         s = a[i];
  88     }
  89   return 0;
  90 }