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