gcc/testsuite/gcc.target/i386/avx512f-gather-3.c

   1 /* { dg-do run } */
   2 /* { dg-require-effective-target avx512f } */
   3 /* { dg-options "-O3 -mavx512f -ffast-math" } */
   4
   5 #include "avx512f-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 q[N];
  15 long long **o[N];
  16 long long t[N];
  17 long long *r[N];
  18 int *s[N];
  19
  20 __attribute__((noinline, noclone)) float
  21 f1 (void)
  22 {
  23   int i;
  24   float g = 0.0;
  25   for (i = 0; i < N / 2; i++)
  26     g += f[k[i]];
  27   return g;
  28 }
  29
  30 __attribute__((noinline, noclone)) float
  31 f2 (float *p)
  32 {
  33   int i;
  34   float g = 0.0;
  35   for (i = 0; i < N / 2; i++)
  36     g += p[k[i]];
  37   return g;
  38 }
  39
  40 __attribute__((noinline, noclone)) float
  41 f3 (void)
  42 {
  43   int i;
  44   float g = 0.0;
  45   for (i = 0; i < N / 2; i++)
  46     g += *l[i];
  47   return g;
  48 }
  49
  50 __attribute__((noinline, noclone)) int
  51 f4 (void)
  52 {
  53   int i;
  54   int g = 0;
  55   for (i = 0; i < N / 2; i++)
  56     g += **m[i];
  57   return g;
  58 }
  59
  60 __attribute__((noinline, noclone)) double
  61 f5 (void)
  62 {
  63   int i;
  64   double g = 0.0;
  65   for (i = 0; i < N / 2; i++)
  66     g += d[k[i]];
  67   return g;
  68 }
  69
  70 __attribute__((noinline, noclone)) double
  71 f6 (double *p)
  72 {
  73   int i;
  74   double g = 0.0;
  75   for (i = 0; i < N / 2; i++)
  76     g += p[k[i]];
  77   return g;
  78 }
  79
  80 __attribute__((noinline, noclone)) double
  81 f7 (void)
  82 {
  83   int i;
  84   double g = 0.0;
  85   for (i = 0; i < N / 2; i++)
  86     g += *n[i];
  87   return g;
  88 }
  89
  90 __attribute__((noinline, noclone)) int
  91 f8 (void)
  92 {
  93   int i;
  94   int g = 0;
  95   for (i = 0; i < N / 2; i++)
  96     g += **o[i];
  97   return g;
  98 }
  99
 100 __attribute__((noinline, noclone)) float
 101 f9 (void)
 102 {
 103   int i;
 104   float g = 0.0;
 105   for (i = 0; i < N / 2; i++)
 106     g += f[q[i]];
 107   return g;
 108 }
 109
 110 __attribute__((noinline, noclone)) float
 111 f10 (float *p)
 112 {
 113   int i;
 114   float g = 0.0;
 115   for (i = 0; i < N / 2; i++)
 116     g += p[q[i]];
 117   return g;
 118 }
 119
 120 __attribute__((noinline, noclone)) double
 121 f11 (void)
 122 {
 123   int i;
 124   double g = 0.0;
 125   for (i = 0; i < N / 2; i++)
 126     g += d[q[i]];
 127   return g;
 128 }
 129
 130 __attribute__((noinline, noclone)) double
 131 f12 (double *p)
 132 {
 133   int i;
 134   double g = 0.0;
 135   for (i = 0; i < N / 2; i++)
 136     g += p[q[i]];
 137   return g;
 138 }
 139
 140 static void
 141 avx512f_test (void)
 142 {
 143   int i;
 144
 145   for (i = 0; i < N; i++)
 146     {
 147       asm ("");
 148       f[i] = -256.0f + i;
 149       d[i] = -258.0 + i;
 150       k[i] = (i * 731) & (N - 1);
 151       q[i] = (i * 657) & (N - 1);
 152       t[i] = (i * 657) & (N - 1);
 153       l[i] = &f[(i * 239) & (N - 1)];
 154       n[i] = &d[(i * 271) & (N - 1)];
 155       r[i] = &t[(i * 323) & (N - 1)];
 156       s[i] = &k[(i * 565) & (N - 1)];
 157       m[i] = &s[(i * 13) & (N - 1)];
 158       o[i] = &r[(i * 19) & (N - 1)];
 159     }
 160
 161   if (f1 () != 136448.0f || f2 (f) != 136448.0f || f3 () != 130304.0)
 162     abort ();
 163   if (f4 () != 261376 || f5 () != 135424.0 || f6 (d) != 135424.0)
 164     abort ();
 165   if (f7 () != 129280.0 || f8 () != 259840L || f9 () != 130816.0f)
 166     abort ();
 167   if (f10 (f) != 130816.0f || f11 () != 129792.0 || f12 (d) != 129792.0)
 168     abort ();
 169 }