gcc/testsuite/gcc.target/i386/avx2-gather-3.c

   1 /* { dg-do run } */
   2 /* { dg-require-effective-target avx2 } */
   3 /* { dg-options "-O3 -mavx2 -ffast-math -mtune=skylake" } */
   4
   5 #include "avx2-check.h"
   6
   7 #define N 1024
   8 float f[N];
   9 double d[N];
  10 int k[N];
  11 float *l[N];
  12 double *n[N];
  13 int **m[N];
  14 long **o[N];
  15 long q[N];
  16 long *r[N];
  17 int *s[N];
  18
  19 __attribute__((noinline, noclone)) float
  20 f1 (void)
  21 {
  22   int i;
  23   float g = 0.0;
  24   for (i = 0; i < N / 2; i++)
  25     g += f[k[i]];
  26   return g;
  27 }
  28
  29 __attribute__((noinline, noclone)) float
  30 f2 (float *p)
  31 {
  32   int i;
  33   float g = 0.0;
  34   for (i = 0; i < N / 2; i++)
  35     g += p[k[i]];
  36   return g;
  37 }
  38
  39 __attribute__((noinline, noclone)) float
  40 f3 (void)
  41 {
  42   int i;
  43   float g = 0.0;
  44   for (i = 0; i < N / 2; i++)
  45     g += *l[i];
  46   return g;
  47 }
  48
  49 __attribute__((noinline, noclone)) int
  50 f4 (void)
  51 {
  52   int i;
  53   int g = 0;
  54   for (i = 0; i < N / 2; i++)
  55     g += **m[i];
  56   return g;
  57 }
  58
  59 __attribute__((noinline, noclone)) double
  60 f5 (void)
  61 {
  62   int i;
  63   double g = 0.0;
  64   for (i = 0; i < N / 2; i++)
  65     g += d[k[i]];
  66   return g;
  67 }
  68
  69 __attribute__((noinline, noclone)) double
  70 f6 (double *p)
  71 {
  72   int i;
  73   double g = 0.0;
  74   for (i = 0; i < N / 2; i++)
  75     g += p[k[i]];
  76   return g;
  77 }
  78
  79 __attribute__((noinline, noclone)) double
  80 f7 (void)
  81 {
  82   int i;
  83   double g = 0.0;
  84   for (i = 0; i < N / 2; i++)
  85     g += *n[i];
  86   return g;
  87 }
  88
  89 __attribute__((noinline, noclone)) int
  90 f8 (void)
  91 {
  92   int i;
  93   int g = 0;
  94   for (i = 0; i < N / 2; i++)
  95     g += **o[i];
  96   return g;
  97 }
  98
  99 __attribute__((noinline, noclone)) float
 100 f9 (void)
 101 {
 102   int i;
 103   float g = 0.0;
 104   for (i = 0; i < N / 2; i++)
 105     g += f[q[i]];
 106   return g;
 107 }
 108
 109 __attribute__((noinline, noclone)) float
 110 f10 (float *p)
 111 {
 112   int i;
 113   float g = 0.0;
 114   for (i = 0; i < N / 2; i++)
 115     g += p[q[i]];
 116   return g;
 117 }
 118
 119 __attribute__((noinline, noclone)) double
 120 f11 (void)
 121 {
 122   int i;
 123   double g = 0.0;
 124   for (i = 0; i < N / 2; i++)
 125     g += d[q[i]];
 126   return g;
 127 }
 128
 129 __attribute__((noinline, noclone)) double
 130 f12 (double *p)
 131 {
 132   int i;
 133   double g = 0.0;
 134   for (i = 0; i < N / 2; i++)
 135     g += p[q[i]];
 136   return g;
 137 }
 138
 139 static void
 140 avx2_test (void)
 141 {
 142   int i;
 143
 144   for (i = 0; i < N; i++)
 145     {
 146       asm ("");
 147       f[i] = -256.0f + i;
 148       d[i] = -258.0 + i;
 149       k[i] = (i * 731) & (N - 1);
 150       q[i] = (i * 657) & (N - 1);
 151       l[i] = &f[(i * 239) & (N - 1)];
 152       n[i] = &d[(i * 271) & (N - 1)];
 153       r[i] = &q[(i * 323) & (N - 1)];
 154       s[i] = &k[(i * 565) & (N - 1)];
 155       m[i] = &s[(i * 13) & (N - 1)];
 156       o[i] = &r[(i * 19) & (N - 1)];
 157     }
 158
 159   if (f1 () != 136448.0f || f2 (f) != 136448.0f || f3 () != 130304.0)
 160     abort ();
 161   if (f4 () != 261376 || f5 () != 135424.0 || f6 (d) != 135424.0)
 162     abort ();
 163   if (f7 () != 129280.0 || f8 () != 259840L || f9 () != 130816.0f)
 164     abort ();
 165   if (f10 (f) != 130816.0f || f11 () != 129792.0 || f12 (d) != 129792.0)
 166     abort ();
 167 }