gcc/testsuite/gcc.target/i386/ssse3-pabsb.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_pabsb (int *i1, int *r)
  27 {
  28   __m64 t1 = *(__m64 *) i1;
  29   *(__m64 *) r = _mm_abs_pi8 (t1);
  30   _mm_empty ();
  31 }
  32
  33 /* Test the 128-bit form */
  34 static void
  35 ssse3_test_pabsb128 (int *i1, int *r)
  36 {
  37   /* Assumes incoming pointers are 16-byte aligned */
  38   __m128i t1 = *(__m128i *) i1;
  39   *(__m128i *) r = _mm_abs_epi8 (t1);
  40 }
  41
  42 /* Routine to manually compute the results */
  43 static void
  44 compute_correct_result (int *i1, int *r)
  45 {
  46   char *b1 = (char *) i1;
  47   char *bout = (char *) r;
  48   int i;
  49
  50   for (i = 0; i < 16; i++)
  51     if (b1[i] < 0)
  52       bout[i] = -b1[i];
  53     else
  54       bout[i] = b1[i];
  55 }
  56
  57 static void
  58 ssse3_test (void)
  59 {
  60   int i;
  61   int r [4] __attribute__ ((aligned(16)));
  62   int ck [4];
  63   int fail = 0;
  64
  65   for (i = 0; i < 256; i += 4)
  66     {
  67       /* Manually compute the result */
  68       compute_correct_result(&vals[i + 0], ck);
  69
  70       /* Run the 64-bit tests */
  71       ssse3_test_pabsb (&vals[i + 0], &r[0]);
  72       ssse3_test_pabsb (&vals[i + 2], &r[2]);
  73       fail += chk_128 (ck, r);
  74
  75       /* Run the 128-bit tests */
  76       ssse3_test_pabsb128 (&vals[i + 0], r);
  77       fail += chk_128 (ck, r);
  78     }
  79
  80   if (fail != 0)
  81     abort ();
  82 }