libgomp/testsuite/libgomp.c/scan-14.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) if (simd:0)
  16   for (int i = 0; i < 1024; i++)
  17     {
  18       { r += a[i]; r2 += a[i]; r3 += a[i]; }
  19       #pragma omp scan inclusive(r, r2, r3)
  20       {
  21         b[i] = r;
  22         b2[i] = r2;
  23         b3[i] = r3;
  24       }
  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) simdlen (1)
  36   for (int i = 0; i < 1024; i++)
  37     {
  38       {
  39         s += 2 * a[i];
  40         s2 += 2 * a[i];
  41         s3 += 2 * a[i];
  42       }
  43       #pragma omp scan inclusive(s, s2, s3)
  44       { b[i] = s; b2[i] = s2; b3[i] = s3; }
  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)
  55   for (int i = 0; i < 1024; i++)
  56     {
  57       {
  58         r += a[i];
  59         r2 += a[i];
  60         r3 += a[i];
  61       }
  62       #pragma omp scan inclusive(r, r2, r3)
  63       {
  64         b[i] = r;
  65         b2[i] = r2;
  66         b3[i] = r3;
  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 reduction (inscan, +:s, s2, s3)
  78   for (int i = 0; i < 1024; i++)
  79     {
  80       { s += 2 * a[i]; s2 += 2 * a[i]; s3 += 2 * a[i]; }
  81       #pragma omp scan inclusive(s, s2, s3)
  82       { b[i] = s; b2[i] = s2; b3[i] = s3; }
  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       s += i;
 112       if (b[i] != s
 113           || b2[i] != (unsigned short) s
 114           || b3[i] != (unsigned char) s)
 115         abort ();
 116       else
 117         {
 118           b[i] = 25;
 119           b2[i] = 24;
 120           b3[i] = 26;
 121         }
 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       s += 2 * i;
 132       if (b[i] != s
 133           || b2[i] != (unsigned short) s
 134           || b3[i] != (unsigned char) s)
 135         abort ();
 136       else
 137         {
 138           b[i] = -1;
 139           b2[i] = -1;
 140           b3[i] = -1;
 141         }
 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       s += i;
 155       if (b[i] != s
 156           || b2[i] != (unsigned short) s
 157           || b3[i] != (unsigned char) s)
 158         abort ();
 159       else
 160         {
 161           b[i] = 25;
 162           b2[i] = 24;
 163           b3[i] = 26;
 164         }
 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       s += 2 * i;
 177       if (b[i] != s
 178           || b2[i] != (unsigned short) s
 179           || b3[i] != (unsigned char) s)
 180         abort ();
 181     }
 182   return 0;
 183 }