libgomp/testsuite/libgomp.c/scan-16.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 int r, a[1024], b[1024];
   9
  10 #pragma omp declare reduction (foo: int: omp_out += omp_in) initializer (omp_priv = 0)
  11
  12 __attribute__((noipa)) void
  13 foo (int *a, int *b)
  14 {
  15   #pragma omp for simd reduction (inscan, foo:r)
  16   for (int i = 0; i < 1024; i++)
  17     {
  18       b[i] = r;
  19       #pragma omp scan exclusive(r)
  20       r += a[i];
  21     }
  22 }
  23
  24 __attribute__((noipa)) int
  25 bar (void)
  26 {
  27   int s = 0;
  28   #pragma omp parallel
  29   #pragma omp for simd simdlen (1) reduction (inscan, foo:s)
  30   for (int i = 0; i < 1024; i++)
  31     {
  32       b[i] = s;
  33       #pragma omp scan exclusive(s)
  34       s += 2 * a[i];
  35     }
  36   return s;
  37 }
  38
  39 __attribute__((noipa)) void
  40 baz (int *a, int *b)
  41 {
  42   #pragma omp parallel for simd if (simd: 0) reduction (inscan, foo:r)
  43   for (int i = 0; i < 1024; i++)
  44     {
  45       b[i] = r;
  46       #pragma omp scan exclusive(r)
  47       r += a[i];
  48     }
  49 }
  50
  51 __attribute__((noipa)) int
  52 qux (void)
  53 {
  54   int s = 0;
  55   #pragma omp parallel for simd reduction (inscan, foo:s)
  56   for (int i = 0; i < 1024; i++)
  57     {
  58       b[i] = s;
  59       #pragma omp scan exclusive(s)
  60       s += 2 * a[i];
  61     }
  62   return s;
  63 }
  64
  65 int
  66 main ()
  67 {
  68   int s = 0;
  69   for (int i = 0; i < 1024; ++i)
  70     {
  71       a[i] = i;
  72       b[i] = -1;
  73       asm ("" : "+g" (i));
  74     }
  75   #pragma omp parallel
  76   foo (a, b);
  77   if (r != 1024 * 1023 / 2)
  78     abort ();
  79   for (int i = 0; i < 1024; ++i)
  80     {
  81       if (b[i] != s)
  82         abort ();
  83       else
  84         b[i] = 25;
  85       s += i;
  86     }
  87   if (bar () != 1024 * 1023)
  88     abort ();
  89   s = 0;
  90   for (int i = 0; i < 1024; ++i)
  91     {
  92       if (b[i] != s)
  93         abort ();
  94       else
  95         b[i] = -1;
  96       s += 2 * i;
  97     }
  98   r = 0;
  99   baz (a, b);
 100   if (r != 1024 * 1023 / 2)
 101     abort ();
 102   s = 0;
 103   for (int i = 0; i < 1024; ++i)
 104     {
 105       if (b[i] != s)
 106         abort ();
 107       else
 108         b[i] = -25;
 109       s += i;
 110     }
 111   if (qux () != 1024 * 1023)
 112     abort ();
 113   s = 0;
 114   for (int i = 0; i < 1024; ++i)
 115     {
 116       if (b[i] != s)
 117         abort ();
 118       s += 2 * i;
 119     }
 120   return 0;
 121 }