gcc/testsuite/gcc.target/i386/ssse3-palignr.c

   1 /* { dg-do run { target i?86-*-* x86_64-*-* } } */
   2 /* { dg-options "-O2 -mssse3" } */
   3 #include <tmmintrin.h>
   4 #include <string.h>
   5 #include <stdlib.h>
   6 #include "../../gcc.dg/i386-cpuid.h"
   7 #include "ssse3-vals.h"
   8
   9 static void ssse3_test (void);
  10
  11 int
  12 main ()
  13 {
  14   unsigned long cpu_facilities;
  15
  16   cpu_facilities = i386_cpuid_ecx ();
  17
  18   /* Run SSSE3 test only if host has SSSE3 support.  */
  19   if ((cpu_facilities & bit_SSSE3))
  20     ssse3_test ();
  21
  22   exit (0);
  23 }
  24
  25 /* Test the 64-bit form */
  26 static void
  27 ssse3_test_palignr (int *i1, int *i2, unsigned int imm, int *r)
  28 {
  29   __m64 t1 = *(__m64 *) i1;
  30   __m64 t2 = *(__m64 *) i2;
  31
  32   switch (imm)
  33     {
  34     case 0:
  35       *(__m64 *) r = _mm_alignr_pi8 (t1, t2, 0);
  36       break;
  37     case 1:
  38       *(__m64 *) r = _mm_alignr_pi8 (t1, t2, 1);
  39       break;
  40     case 2:
  41       *(__m64 *) r = _mm_alignr_pi8 (t1, t2, 2);
  42       break;
  43     case 3:
  44       *(__m64 *) r = _mm_alignr_pi8 (t1, t2, 3);
  45       break;
  46     case 4:
  47       *(__m64 *) r = _mm_alignr_pi8 (t1, t2, 4);
  48       break;
  49     case 5:
  50       *(__m64 *) r = _mm_alignr_pi8 (t1, t2, 5);
  51       break;
  52     case 6:
  53       *(__m64 *) r = _mm_alignr_pi8 (t1, t2, 6);
  54       break;
  55     case 7:
  56       *(__m64 *) r = _mm_alignr_pi8 (t1, t2, 7);
  57       break;
  58     case 8:
  59       *(__m64 *) r = _mm_alignr_pi8 (t1, t2, 8);
  60       break;
  61     case 9:
  62       *(__m64 *) r = _mm_alignr_pi8 (t1, t2, 9);
  63       break;
  64     case 10:
  65       *(__m64 *) r = _mm_alignr_pi8 (t1, t2, 10);
  66       break;
  67     case 11:
  68       *(__m64 *) r = _mm_alignr_pi8 (t1, t2, 11);
  69       break;
  70     case 12:
  71       *(__m64 *) r = _mm_alignr_pi8 (t1, t2, 12);
  72       break;
  73     case 13:
  74       *(__m64 *) r = _mm_alignr_pi8 (t1, t2, 13);
  75       break;
  76     case 14:
  77       *(__m64 *) r = _mm_alignr_pi8 (t1, t2, 14);
  78       break;
  79     case 15:
  80       *(__m64 *) r = _mm_alignr_pi8 (t1, t2, 15);
  81       break;
  82     default:
  83       *(__m64 *) r = _mm_alignr_pi8 (t1, t2, 16);
  84       break;
  85     }
  86
  87    _mm_empty();
  88 }
  89
  90 /* Test the 128-bit form */
  91 static void
  92 ssse3_test_palignr128 (int *i1, int *i2, unsigned int imm, int *r)
  93 {
  94   /* Assumes incoming pointers are 16-byte aligned */
  95   __m128i t1 = *(__m128i *) i1;
  96   __m128i t2 = *(__m128i *) i2;
  97
  98   switch (imm)
  99     {
 100     case 0:
 101       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 0);
 102       break;
 103     case 1:
 104       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 1);
 105       break;
 106     case 2:
 107       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 2);
 108       break;
 109     case 3:
 110       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 3);
 111       break;
 112     case 4:
 113       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 4);
 114       break;
 115     case 5:
 116       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 5);
 117       break;
 118     case 6:
 119       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 6);
 120       break;
 121     case 7:
 122       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 7);
 123       break;
 124     case 8:
 125       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 8);
 126       break;
 127     case 9:
 128       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 9);
 129       break;
 130     case 10:
 131       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 10);
 132       break;
 133     case 11:
 134       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 11);
 135       break;
 136     case 12:
 137       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 12);
 138       break;
 139     case 13:
 140       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 13);
 141       break;
 142     case 14:
 143       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 14);
 144       break;
 145     case 15:
 146       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 15);
 147       break;
 148     case 16:
 149       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 16);
 150       break;
 151     case 17:
 152       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 17);
 153       break;
 154     case 18:
 155       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 18);
 156       break;
 157     case 19:
 158       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 19);
 159       break;
 160     case 20:
 161       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 20);
 162       break;
 163     case 21:
 164       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 21);
 165       break;
 166     case 22:
 167       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 22);
 168       break;
 169     case 23:
 170       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 23);
 171       break;
 172     case 24:
 173       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 24);
 174       break;
 175     case 25:
 176       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 25);
 177       break;
 178     case 26:
 179       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 26);
 180       break;
 181     case 27:
 182       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 27);
 183       break;
 184     case 28:
 185       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 28);
 186       break;
 187     case 29:
 188       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 29);
 189       break;
 190     case 30:
 191       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 30);
 192       break;
 193     case 31:
 194       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 31);
 195       break;
 196     default:
 197       *(__m128i *) r = _mm_alignr_epi8 (t1, t2, 32);
 198       break;
 199     }
 200 }
 201
 202 /* Routine to manually compute the results */
 203 static void
 204 compute_correct_result_128 (int *i1, int *i2, unsigned int imm, int *r)
 205 {
 206   char buf [32];
 207   char *bout = (char *) r;
 208   int i;
 209
 210   memcpy (&buf[0], i2, 16);
 211   memcpy (&buf[16], i1, 16);
 212
 213   for (i = 0; i < 16; i++)
 214     if (imm >= 32 || imm + i >= 32)
 215       bout[i] = 0;
 216     else
 217       bout[i] = buf[imm + i];
 218 }
 219
 220 static void
 221 compute_correct_result_64 (int *i1, int *i2, unsigned int imm, int *r)
 222 {
 223   char buf [16];
 224   char *bout = (char *)r;
 225   int i;
 226
 227   /* Handle the first half */
 228   memcpy (&buf[0], i2, 8);
 229   memcpy (&buf[8], i1, 8);
 230
 231   for (i = 0; i < 8; i++)
 232     if (imm >= 16 || imm + i >= 16)
 233       bout[i] = 0;
 234     else
 235       bout[i] = buf[imm + i];
 236
 237   /* Handle the second half */
 238   memcpy (&buf[0], &i2[2], 8);
 239   memcpy (&buf[8], &i1[2], 8);
 240
 241   for (i = 0; i < 8; i++)
 242     if (imm >= 16 || imm + i >= 16)
 243       bout[i + 8] = 0;
 244     else
 245       bout[i + 8] = buf[imm + i];
 246 }
 247
 248 static void
 249 ssse3_test (void)
 250 {
 251   int i;
 252   int r [4] __attribute__ ((aligned(16)));
 253   int ck [4];
 254   unsigned int imm;
 255   int fail = 0;
 256
 257   for (i = 0; i < 256; i += 8)
 258     for (imm = 0; imm < 100; imm++)
 259       {
 260         /* Manually compute the result */
 261         compute_correct_result_64 (&vals[i + 0], &vals[i + 4], imm, ck);
 262
 263         /* Run the 64-bit tests */
 264         ssse3_test_palignr (&vals[i + 0], &vals[i + 4], imm, &r[0]);
 265         ssse3_test_palignr (&vals[i + 2], &vals[i + 6], imm, &r[2]);
 266         fail += chk_128 (ck, r);
 267
 268         /* Recompute the results for 128-bits */
 269         compute_correct_result_128 (&vals[i + 0], &vals[i + 4], imm, ck);
 270
 271         /* Run the 128-bit tests */
 272         ssse3_test_palignr128 (&vals[i + 0], &vals[i + 4], imm, r);
 273         fail += chk_128 (ck, r);
 274       }
 275
 276   if (fail != 0)
 277     abort ();
 278 }