libgomp/testsuite/libgomp.c/scan-18.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 unsigned short r2, b2[1024];
  10 unsigned char r3, b3[1024];
  11
  12 __attribute__((noipa)) void
  13 foo (int *a, int *b, unsigned short *b2, unsigned char *b3)
  14 {
  15   #pragma omp for simd reduction (inscan, +:r, r2, r3)
  16   for (int i = 0; i < 1024; i++)
  17     {
  18       {
  19         b[i] = r;
  20         b2[i] = r2;
  21         b3[i] = r3;
  22       }
  23       #pragma omp scan exclusive(r, r2, r3)
  24       { r += a[i]; r2 += a[i]; r3 += a[i]; }
  25     }
  26 }
  27
  28 __attribute__((noipa)) int
  29 bar (unsigned short *s2p, unsigned char *s3p)
  30 {
  31   int s = 0;
  32   unsigned short s2 = 0;
  33   unsigned char s3 = 0;
  34   #pragma omp parallel
  35   #pragma omp for simd reduction (inscan, +:s, s2, s3)
  36   for (int i = 0; i < 1024; i++)
  37     {
  38       { b[i] = s; b2[i] = s2; b3[i] = s3; }
  39       #pragma omp scan exclusive(s, s2, s3)
  40       {
  41         s += 2 * a[i];
  42         s2 += 2 * a[i];
  43         s3 += 2 * a[i];
  44       }
  45     }
  46   *s2p = s2;
  47   *s3p = s3;
  48   return s;
  49 }
  50
  51 __attribute__((noipa)) void
  52 baz (int *a, int *b, unsigned short *b2, unsigned char *b3)
  53 {
  54   #pragma omp parallel for simd reduction (inscan, +:r, r2, r3) if (simd: 0)
  55   for (int i = 0; i < 1024; i++)
  56     {
  57       {
  58         b[i] = r;
  59         b2[i] = r2;
  60         b3[i] = r3;
  61       }
  62       #pragma omp scan exclusive(r, r2, r3)
  63       {
  64         r += a[i];
  65         r2 += a[i];
  66         r3 += a[i];
  67       }
  68     }
  69 }
  70
  71 __attribute__((noipa)) int
  72 qux (unsigned short *s2p, unsigned char *s3p)
  73 {
  74   int s = 0;
  75   unsigned short s2 = 0;
  76   unsigned char s3 = 0;
  77   #pragma omp parallel for simd simdlen (1) reduction (inscan, +:s, s2, s3)
  78   for (int i = 0; i < 1024; i++)
  79     {
  80       { b[i] = s; b2[i] = s2; b3[i] = s3; }
  81       #pragma omp scan exclusive(s, s2, s3)
  82       { s += 2 * a[i]; s2 += 2 * a[i]; s3 += 2 * a[i]; }
  83     }
  84   *s2p = s2;
  85   *s3p = s3;
  86   return s;
  87 }
  88
  89 int
  90 main ()
  91 {
  92   int s = 0;
  93   unsigned short s2;
  94   unsigned char s3;
  95   for (int i = 0; i < 1024; ++i)
  96     {
  97       a[i] = i;
  98       b[i] = -1;
  99       b2[i] = -1;
 100       b3[i] = -1;
 101       asm ("" : "+g" (i));
 102     }
 103   #pragma omp parallel
 104   foo (a, b, b2, b3);
 105   if (r != 1024 * 1023 / 2
 106       || r2 != (unsigned short) r
 107       || r3 != (unsigned char) r)
 108     abort ();
 109   for (int i = 0; i < 1024; ++i)
 110     {
 111       if (b[i] != s
 112           || b2[i] != (unsigned short) s
 113           || b3[i] != (unsigned char) s)
 114         abort ();
 115       else
 116         {
 117           b[i] = 25;
 118           b2[i] = 24;
 119           b3[i] = 26;
 120         }
 121       s += i;
 122     }
 123   if (bar (&s2, &s3) != 1024 * 1023)
 124     abort ();
 125   if (s2 != (unsigned short) (1024 * 1023)
 126       || s3 != (unsigned char) (1024 * 1023))
 127     abort ();
 128   s = 0;
 129   for (int i = 0; i < 1024; ++i)
 130     {
 131       if (b[i] != s
 132           || b2[i] != (unsigned short) s
 133           || b3[i] != (unsigned char) s)
 134         abort ();
 135       else
 136         {
 137           b[i] = -1;
 138           b2[i] = -1;
 139           b3[i] = -1;
 140         }
 141       s += 2 * i;
 142     }
 143   r = 0;
 144   r2 = 0;
 145   r3 = 0;
 146   baz (a, b, b2, b3);
 147   if (r != 1024 * 1023 / 2
 148       || r2 != (unsigned short) r
 149       || r3 != (unsigned char) r)
 150     abort ();
 151   s = 0;
 152   for (int i = 0; i < 1024; ++i)
 153     {
 154       if (b[i] != s
 155           || b2[i] != (unsigned short) s
 156           || b3[i] != (unsigned char) s)
 157         abort ();
 158       else
 159         {
 160           b[i] = 25;
 161           b2[i] = 24;
 162           b3[i] = 26;
 163         }
 164       s += i;
 165     }
 166   s2 = 0;
 167   s3 = 0;
 168   if (qux (&s2, &s3) != 1024 * 1023)
 169     abort ();
 170   if (s2 != (unsigned short) (1024 * 1023)
 171       || s3 != (unsigned char) (1024 * 1023))
 172     abort ();
 173   s = 0;
 174   for (int i = 0; i < 1024; ++i)
 175     {
 176       if (b[i] != s
 177           || b2[i] != (unsigned short) s
 178           || b3[i] != (unsigned char) s)
 179         abort ();
 180       s += 2 * i;
 181     }
 182   return 0;
 183 }