gcc/testsuite/gcc.dg/vect/vect-simd-11.c

   1 /* { dg-require-effective-target size32plus } */
   2 /* { dg-additional-options "-fopenmp-simd" } */
   3 /* { dg-additional-options "-mavx" { target avx_runtime } } */
   4 /* { dg-final { scan-tree-dump-times "vectorized \[1-3] loops" 2 "vect" { target i?86-*-* x86_64-*-* } } } */
   5
   6 #ifndef main
   7 #include "tree-vect.h"
   8 #endif
   9
  10 int r, a[1024], b[1024];
  11 unsigned short r2, b2[1024];
  12 unsigned char r3, b3[1024];
  13
  14 __attribute__((noipa)) void
  15 foo (int *a, int *b, unsigned short *b2, unsigned char *b3)
  16 {
  17   #pragma omp simd reduction (inscan, +:r, r2, r3)
  18   for (int i = 0; i < 1024; i++)
  19     {
  20       { r += a[i]; r2 += a[i]; r3 += a[i]; }
  21       #pragma omp scan inclusive(r, r2, r3)
  22       {
  23         b[i] = r;
  24         b2[i] = r2;
  25         b3[i] = r3;
  26       }
  27     }
  28 }
  29
  30 __attribute__((noipa)) int
  31 bar (unsigned short *s2p, unsigned char *s3p)
  32 {
  33   int s = 0;
  34   unsigned short s2 = 0;
  35   unsigned char s3 = 0;
  36   #pragma omp simd reduction (inscan, +:s, s2, s3)
  37   for (int i = 0; i < 1024; i++)
  38     {
  39       {
  40         s += 2 * a[i];
  41         s2 += 2 * a[i];
  42         s3 += 2 * a[i];
  43       }
  44       #pragma omp scan inclusive(s, s2, s3)
  45       { b[i] = s; b2[i] = s2; b3[i] = s3; }
  46     }
  47   *s2p = s2;
  48   *s3p = s3;
  49   return s;
  50 }
  51
  52 __attribute__((noipa)) void
  53 baz (int *a, int *b, unsigned short *b2, unsigned char *b3)
  54 {
  55   #pragma omp simd reduction (inscan, +:r, r2, r3) if (simd: 0)
  56   for (int i = 0; i < 1024; i++)
  57     {
  58       {
  59         r += a[i];
  60         r2 += a[i];
  61         r3 += a[i];
  62       }
  63       #pragma omp scan inclusive(r, r2, r3)
  64       {
  65         b[i] = r;
  66         b2[i] = r2;
  67         b3[i] = r3;
  68       }
  69     }
  70 }
  71
  72 __attribute__((noipa)) int
  73 qux (unsigned short *s2p, unsigned char *s3p)
  74 {
  75   int s = 0;
  76   unsigned short s2 = 0;
  77   unsigned char s3 = 0;
  78   #pragma omp simd reduction (inscan, +:s, s2, s3) simdlen (1)
  79   for (int i = 0; i < 1024; i++)
  80     {
  81       { s += 2 * a[i]; s2 += 2 * a[i]; s3 += 2 * a[i]; }
  82       #pragma omp scan inclusive(s, s2, s3)
  83       { b[i] = s; b2[i] = s2; b3[i] = s3; }
  84     }
  85   *s2p = s2;
  86   *s3p = s3;
  87   return s;
  88 }
  89
  90 int
  91 main ()
  92 {
  93   int s = 0;
  94   unsigned short s2;
  95   unsigned char s3;
  96 #ifndef main
  97   check_vect ();
  98 #endif
  99   for (int i = 0; i < 1024; ++i)
 100     {
 101       a[i] = i;
 102       b[i] = -1;
 103       b2[i] = -1;
 104       b3[i] = -1;
 105       asm ("" : "+g" (i));
 106     }
 107   foo (a, b, b2, b3);
 108   if (r != 1024 * 1023 / 2
 109       || r2 != (unsigned short) r
 110       || r3 != (unsigned char) r)
 111     abort ();
 112 #pragma GCC novector
 113   for (int i = 0; i < 1024; ++i)
 114     {
 115       s += i;
 116       if (b[i] != s
 117           || b2[i] != (unsigned short) s
 118           || b3[i] != (unsigned char) s)
 119         abort ();
 120       else
 121         {
 122           b[i] = 25;
 123           b2[i] = 24;
 124           b3[i] = 26;
 125         }
 126     }
 127   if (bar (&s2, &s3) != 1024 * 1023)
 128     abort ();
 129   if (s2 != (unsigned short) (1024 * 1023)
 130       || s3 != (unsigned char) (1024 * 1023))
 131     abort ();
 132   s = 0;
 133 #pragma GCC novector
 134   for (int i = 0; i < 1024; ++i)
 135     {
 136       s += 2 * i;
 137       if (b[i] != s
 138           || b2[i] != (unsigned short) s
 139           || b3[i] != (unsigned char) s)
 140         abort ();
 141       else
 142         {
 143           b[i] = -1;
 144           b2[i] = -1;
 145           b3[i] = -1;
 146         }
 147     }
 148   r = 0;
 149   r2 = 0;
 150   r3 = 0;
 151   baz (a, b, b2, b3);
 152   if (r != 1024 * 1023 / 2
 153       || r2 != (unsigned short) r
 154       || r3 != (unsigned char) r)
 155     abort ();
 156   s = 0;
 157 #pragma GCC novector
 158   for (int i = 0; i < 1024; ++i)
 159     {
 160       s += i;
 161       if (b[i] != s
 162           || b2[i] != (unsigned short) s
 163           || b3[i] != (unsigned char) s)
 164         abort ();
 165       else
 166         {
 167           b[i] = 25;
 168           b2[i] = 24;
 169           b3[i] = 26;
 170         }
 171     }
 172   s2 = 0;
 173   s3 = 0;
 174   if (qux (&s2, &s3) != 1024 * 1023)
 175     abort ();
 176   if (s2 != (unsigned short) (1024 * 1023)
 177       || s3 != (unsigned char) (1024 * 1023))
 178     abort ();
 179   s = 0;
 180 #pragma GCC novector
 181   for (int i = 0; i < 1024; ++i)
 182     {
 183       s += 2 * i;
 184       if (b[i] != s
 185           || b2[i] != (unsigned short) s
 186           || b3[i] != (unsigned char) s)
 187         abort ();
 188     }
 189   return 0;
 190 }