gcc/testsuite/gcc.target/i386/ssse3-phsubsw.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_phsubsw (int *i1, int *i2, int *r)
  13 {
  14   __m64 t1 = *(__m64 *) i1;
  15   __m64 t2 = *(__m64 *) i2;
  16
  17   *(__m64 *) r = _mm_hsubs_pi16 (t1, t2);
  18
  19   _mm_empty ();
  20 }
  21
  22 /* Test the 128-bit form */
  23 static void
  24 ssse3_test_phsubsw128 (int *i1, int *i2, int *r)
  25 {
  26   /* Assumes incoming pointers are 16-byte aligned */
  27   __m128i t1 = *(__m128i *) i1;
  28   __m128i t2 = *(__m128i *) i2;
  29   *(__m128i *) r = _mm_hsubs_epi16 (t1, t2);
  30 }
  31
  32 static short
  33 signed_saturate_to_word (int x)
  34 {
  35   if (x > (int )0x7fff)
  36     return 0x7fff;
  37
  38   if (x < (int) 0xffff8000)
  39     return 0x8000;
  40
  41   return (short)x;
  42 }
  43
  44 /* Routine to manually compute the results */
  45 static void
  46 compute_correct_result (int *i1, int *i2, int *r)
  47 {
  48   short *s1 = (short *) i1;
  49   short *s2 = (short *) i2;
  50   short *sout = (short *) r;
  51   int i;
  52
  53   for (i = 0; i < 4; i++)
  54     sout[i] = signed_saturate_to_word (s1[2 * i] - s1[2 * i + 1]);
  55
  56   for (i = 0; i < 4; i++)
  57     sout[i + 4] = signed_saturate_to_word (s2[2 * i] - s2[2 * i + 1]);
  58 }
  59
  60 static void
  61 ssse3_test (void)
  62 {
  63   int i;
  64   int r [4] __attribute__ ((aligned(16)));
  65   int ck [4];
  66   int fail = 0;
  67
  68   for (i = 0; i < 256; i += 8)
  69     {
  70       /* Manually compute the result */
  71       compute_correct_result (&vals[i + 0], &vals[i + 4], ck);
  72
  73       /* Run the 64-bit tests */
  74       ssse3_test_phsubsw (&vals[i + 0], &vals[i + 2], &r[0]);
  75       ssse3_test_phsubsw (&vals[i + 4], &vals[i + 6], &r[2]);
  76       fail += chk_128 (ck, r);
  77
  78       /* Run the 128-bit tests */
  79       ssse3_test_phsubsw128 (&vals[i + 0], &vals[i + 4], r);
  80       fail += chk_128 (ck, r);
  81     }
  82
  83   if (fail != 0)
  84     abort ();
  85 }