gcc/testsuite/gcc.dg/vect/slp-21.c

   1 /* { dg-require-effective-target vect_int } */
   2
   3 #include <stdarg.h>
   4 #include "tree-vect.h"
   5
   6 #define N 128
   7
   8 int
   9 main1 ()
  10 {
  11   unsigned short i;
  12   unsigned short out[N*8], out2[N*8], b0, b1, b2, b3, b4, a0, a1, a2, a3, b5;
  13   unsigned short in[N*8];
  14
  15 #pragma GCC novector
  16   for (i = 0; i < N*8; i++)
  17     {
  18       in[i] = i;
  19     }
  20
  21   /* Different operations in both cases - vectorization with interleaving.  */
  22   for (i = 0; i < N; i++)
  23     {
  24       a0 = in[i*4];
  25       a1 = in[i*4 + 1];
  26       a2 = in[i*4 + 2];
  27       a3 = in[i*4 + 3];
  28
  29       b0 = a0 * 8;
  30       b1 = a1 + 7;
  31       b2 = a2 + 6;
  32       b3 = a3 * 5;
  33
  34       b4 = a2 + 4;
  35       b5 = a3 + 3;
  36
  37       out[i*4] = b0;
  38       out[i*4 + 1] = b1;
  39       out[i*4 + 2] = b2;
  40       out[i*4 + 3] = b3;
  41
  42       out2[i*4] = b0;
  43       out2[i*4 + 1] = b1;
  44       out2[i*4 + 2] = b4;
  45       out2[i*4 + 3] = b5;
  46     }
  47
  48   /* check results:  */
  49 #pragma GCC novector
  50   for (i = 0; i < N; i++)
  51     {
  52       a0 = in[i*4];
  53       a1 = in[i*4 + 1];
  54       a2 = in[i*4 + 2];
  55       a3 = in[i*4 + 3];
  56
  57       b0 = a0 * 8;
  58       b1 = a1 + 7;
  59       b2 = a2 + 6;
  60       b3 = a3 * 5;
  61
  62       b4 = a2 + 4;
  63       b5 = a3 + 3;
  64
  65       if (out[i*4] != b0
  66          || out[i*4 + 1] != b1
  67          || out[i*4 + 2] != b2
  68          || out[i*4 + 3] != b3)
  69         abort ();
  70
  71       if (out2[i*4] != b0
  72          || out2[i*4 + 1] != b1
  73          || out2[i*4 + 2] != b4
  74          || out2[i*4 + 3] != b5)
  75         abort ();
  76     }
  77
  78   /* Different operations in the first case - vectorization with interleaving.  */
  79   for (i = 0; i < N; i++)
  80     {
  81       a0 = in[i*4];
  82       a1 = in[i*4 + 1];
  83       a2 = in[i*4 + 2];
  84       a3 = in[i*4 + 3];
  85
  86       b0 = a0 + 8;
  87       b1 = a1 + 7;
  88       b2 = a2 + 6;
  89       b3 = a3 * 5;
  90
  91       b4 = a2 + 4;
  92       b5 = a3 + 3;
  93
  94       out[i*4] = b0;
  95       out[i*4 + 1] = b1;
  96       out[i*4 + 2] = b2;
  97       out[i*4 + 3] = b3;
  98
  99       out2[i*4] = b0;
 100       out2[i*4 + 1] = b1;
 101       out2[i*4 + 2] = b4;
 102       out2[i*4 + 3] = b5;
 103     }
 104
 105   /* check results:  */
 106 #pragma GCC novector
 107   for (i = 0; i < N; i++)
 108     {
 109       a0 = in[i*4];
 110       a1 = in[i*4 + 1];
 111       a2 = in[i*4 + 2];
 112       a3 = in[i*4 + 3];
 113
 114       b0 = a0 + 8;
 115       b1 = a1 + 7;
 116       b2 = a2 + 6;
 117       b3 = a3 * 5;
 118
 119       b4 = a2 + 4;
 120       b5 = a3 + 3;
 121
 122       if (out[i*4] != b0
 123          || out[i*4 + 1] != b1
 124          || out[i*4 + 2] != b2
 125          || out[i*4 + 3] != b3)
 126         abort ();
 127
 128       if (out2[i*4] != b0
 129          || out2[i*4 + 1] != b1
 130          || out2[i*4 + 2] != b4
 131          || out2[i*4 + 3] != b5)
 132         abort ();
 133     }
 134
 135
 136   /* Different operations in the second case - vectorization with interleaving.  */
 137   for (i = 0; i < N; i++)
 138     {
 139       a0 = in[i*4];
 140       a1 = in[i*4 + 1];
 141       a2 = in[i*4 + 2];
 142       a3 = in[i*4 + 3];
 143
 144       b0 = a0 + 8;
 145       b1 = a1 + 7;
 146       b2 = a2 + 6;
 147       b3 = a3 + 5;
 148
 149       b4 = a2 * 4;
 150       b5 = a3 + 3;
 151
 152       out[i*4] = b0;
 153       out[i*4 + 1] = b1;
 154       out[i*4 + 2] = b2;
 155       out[i*4 + 3] = b3;
 156
 157       out2[i*4] = b0;
 158       out2[i*4 + 1] = b1;
 159       out2[i*4 + 2] = b4;
 160       out2[i*4 + 3] = b5;
 161     }
 162
 163   /* check results:  */
 164 #pragma GCC novector
 165   for (i = 0; i < N; i++)
 166     {
 167       a0 = in[i*4];
 168       a1 = in[i*4 + 1];
 169       a2 = in[i*4 + 2];
 170       a3 = in[i*4 + 3];
 171
 172       b0 = a0 + 8;
 173       b1 = a1 + 7;
 174       b2 = a2 + 6;
 175       b3 = a3 + 5;
 176
 177       b4 = a2 * 4;
 178       b5 = a3 + 3;
 179
 180       if (out[i*4] != b0
 181          || out[i*4 + 1] != b1
 182          || out[i*4 + 2] != b2
 183          || out[i*4 + 3] != b3)
 184         abort ();
 185
 186       if (out2[i*4] != b0
 187          || out2[i*4 + 1] != b1
 188          || out2[i*4 + 2] != b4
 189          || out2[i*4 + 3] != b5)
 190         abort ();
 191     }
 192
 193
 194   return 0;
 195 }
 196
 197 int main (void)
 198 {
 199   check_vect ();
 200
 201   main1 ();
 202
 203   return 0;
 204 }
 205
 206 /* { dg-final { scan-tree-dump-times "vectorized 3 loops" 1 "vect"  { target { vect_strided4 || vect_extract_even_odd } } } } */
 207 /* { dg-final { scan-tree-dump-times "vectorized 0 loops" 1 "vect"  { target  { ! { vect_strided4 || vect_extract_even_odd } } } } } */