gcc/testsuite/gcc.target/i386/ssse3-pmaddubsw.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_pmaddubsw (int *i1, int *i2, int *r)
  27 {
  28   __m64 t1 = *(__m64 *) i1;
  29   __m64 t2 = *(__m64 *) i2;
  30   *(__m64 *) r = _mm_maddubs_pi16 (t1, t2);
  31   _mm_empty ();
  32 }
  33
  34 /* Test the 128-bit form */
  35 static void
  36 ssse3_test_pmaddubsw128 (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_maddubs_epi16 (t1, t2);
  42 }
  43
  44 static short
  45 signed_saturate_to_word(int x)
  46 {
  47   if (x > (int) 0x7fff)
  48     return 0x7fff;
  49
  50   if (x < (int) 0xffff8000)
  51     return 0x8000;
  52
  53   return (short) x;
  54 }
  55
  56 /* Routine to manually compute the results */
  57 static void
  58 compute_correct_result (int *i1, int *i2, int *r)
  59 {
  60   unsigned char *ub1 = (unsigned char *) i1;
  61   char *sb2 = (char *) i2;
  62   short *sout = (short *) r;
  63   int t0;
  64   int i;
  65
  66   for (i = 0; i < 8; i++)
  67     {
  68       t0 = ((int) ub1[2 * i] * (int) sb2[2 * i] +
  69             (int) ub1[2 * i + 1] * (int) sb2[2 * i + 1]);
  70       sout[i] = signed_saturate_to_word (t0);
  71     }
  72 }
  73
  74 static void
  75 ssse3_test (void)
  76 {
  77   int i;
  78   int r [4] __attribute__ ((aligned(16)));
  79   int ck [4];
  80   int fail = 0;
  81
  82   for (i = 0; i < 256; i += 8)
  83     {
  84       /* Manually compute the result */
  85       compute_correct_result (&vals[i + 0], &vals[i + 4], ck);
  86
  87       /* Run the 64-bit tests */
  88       ssse3_test_pmaddubsw (&vals[i + 0], &vals[i + 4], &r[0]);
  89       ssse3_test_pmaddubsw (&vals[i + 2], &vals[i + 6], &r[2]);
  90       fail += chk_128 (ck, r);
  91
  92       /* Run the 128-bit tests */
  93       ssse3_test_pmaddubsw128 (&vals[i + 0], &vals[i + 4], r);
  94       fail += chk_128 (ck, r);
  95     }
  96
  97   if (fail != 0)
  98     abort ();
  99 }