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

   1 /* { dg-do run { target i?86-*-* x86_64-*-* } } */
   2 /* { dg-options "-O2 -mssse3" } */
   3 #include <tmmintrin.h>
   4 #include <stdlib.h>
   5 #include "../../gcc.dg/i386-cpuid.h"
   6 #include "ssse3-vals.h"
   7
   8 static void ssse3_test (void);
   9
  10 int
  11 main ()
  12 {
  13   unsigned long cpu_facilities;
  14
  15   cpu_facilities = i386_cpuid_ecx ();
  16
  17   /* Run SSSE3 test only if host has SSSE3 support.  */
  18   if ((cpu_facilities & bit_SSSE3))
  19     ssse3_test ();
  20
  21   exit (0);
  22 }
  23
  24 /* Test the 64-bit form */
  25 static void
  26 ssse3_test_phsubd (int *i1, int *i2, int *r)
  27 {
  28   __m64 t1 = *(__m64 *) i1;
  29   __m64 t2 = *(__m64 *) i2;
  30   *(__m64 *) r = _mm_hsub_pi32(t1, t2);
  31   _mm_empty ();
  32 }
  33
  34 /* Test the 128-bit form */
  35 static void
  36 ssse3_test_phsubd128 (int *i1, int *i2, int *r)
  37 {
  38   /* Assumes incoming pointers are 16-byte aligned */
  39   __m128i t1 = *(__m128i *) i1;
  40   __m128i t2 = *(__m128i *) i2;
  41   *(__m128i *) r = _mm_hsub_epi32 (t1, t2);
  42 }
  43
  44 /* Routine to manually compute the results */
  45 static void
  46 compute_correct_result (int *i1, int *i2, int *r)
  47 {
  48   int i;
  49
  50   for (i = 0; i < 2; i++)
  51     r[i] = i1[2 * i] - i1[2 * i + 1];
  52   for (i = 0; i < 2; i++)
  53     r[i + 2] = i2[2 * i] - i2[2 * i + 1];
  54 }
  55
  56 static void
  57 ssse3_test (void)
  58 {
  59   int i;
  60   int r [4] __attribute__ ((aligned(16)));
  61   int ck [4];
  62   int fail = 0;
  63
  64   for (i = 0; i < 256; i += 8)
  65     {
  66       /* Manually compute the result */
  67       compute_correct_result (&vals[i + 0], &vals[i + 4], ck);
  68
  69       /* Run the 64-bit tests */
  70       ssse3_test_phsubd (&vals[i + 0], &vals[i + 2], &r[0]);
  71       ssse3_test_phsubd (&vals[i + 4], &vals[i + 6], &r[2]);
  72       fail += chk_128 (ck, r);
  73
  74       /* Run the 128-bit tests */
  75       ssse3_test_phsubd128 (&vals[i + 0], &vals[i + 4], r);
  76       fail += chk_128 (ck, r);
  77     }
  78
  79   if (fail != 0)
  80     abort ();
  81 }