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

   1 /* { dg-do run } */
   2 /* { dg-require-effective-target ssse3 } */
   3 /* { dg-options "-O2 -fno-strict-aliasing -mssse3" } */
   4
   5 #include "ssse3-check.h"
   6 #include "ssse3-vals.h"
   7
   8 #include <tmmintrin.h>
   9
  10 /* Test the 64-bit form */
  11 static void
  12 ssse3_test_phaddd (int *i1, int *i2, int *r)
  13 {
  14   __m64 t1 = *(__m64 *) i1;
  15   __m64 t2 = *(__m64 *) i2;
  16   *(__m64 *) r = _mm_hadd_pi32 (t1, t2);
  17   _mm_empty();
  18 }
  19
  20 /* Test the 128-bit form */
  21 static void
  22 ssse3_test_phaddd128 (int *i1, int *i2, int *r)
  23 {
  24   /* Assumes incoming pointers are 16-byte aligned */
  25   __m128i t1 = *(__m128i *) i1;
  26   __m128i t2 = *(__m128i *) i2;
  27   *(__m128i *) r = _mm_hadd_epi32 (t1, t2);
  28 }
  29
  30 /* Routine to manually compute the results */
  31 static void
  32 compute_correct_result(int *i1, int *i2, int *r)
  33 {
  34   int i;
  35
  36   for (i = 0; i < 2; i++)
  37     r[i] = i1[2 * i] + i1[2 * i + 1];
  38   for (i = 0; i < 2; i++)
  39     r[i + 2] = i2[2 * i] + i2[2 * i + 1];
  40 }
  41
  42 static void
  43 ssse3_test (void)
  44 {
  45   int i;
  46   int r [4] __attribute__ ((aligned(16)));
  47   int ck [4];
  48   int fail = 0;
  49
  50
  51   for (i = 0; i < 256; i += 8)
  52     {
  53       /* Manually compute the result */
  54       compute_correct_result (&vals[i + 0], &vals[i + 4], ck);
  55
  56       /* Run the 64-bit tests */
  57       ssse3_test_phaddd (&vals[i + 0], &vals[i + 2], &r[0]);
  58       ssse3_test_phaddd (&vals[i + 4], &vals[i + 6], &r[2]);
  59       fail += chk_128 (ck, r);
  60
  61       /* Run the 128-bit tests */
  62       ssse3_test_phaddd128 (&vals[i + 0], &vals[i + 4], r);
  63       fail += chk_128 (ck, r);
  64     }
  65
  66   if (fail != 0)
  67     abort ();
  68 }