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

   1 /* { dg-do run } */
   2 /* { dg-require-effective-target ssse3 } */
   3 /* { dg-options "-O2 -fno-strict-aliasing -mssse3" } */
   4
   5 #ifndef CHECK_H
   6 #define CHECK_H "ssse3-check.h"
   7 #endif
   8
   9 #ifndef TEST
  10 #define TEST ssse3_test
  11 #endif
  12
  13 #include CHECK_H
  14 #include "ssse3-vals.h"
  15
  16 #include <tmmintrin.h>
  17
  18 #ifndef __AVX__
  19 /* Test the 64-bit form */
  20 static void
  21 ssse3_test_phsubw (int *i1, int *i2, int *r)
  22 {
  23   __m64 t1 = *(__m64 *) i1;
  24   __m64 t2 = *(__m64 *) i2;
  25   *(__m64 *) r = _mm_hsub_pi16 (t1, t2);
  26   _mm_empty ();
  27 }
  28 #endif
  29
  30 /* Test the 128-bit form */
  31 static void
  32 ssse3_test_phsubw128 (int *i1, int *i2, int *r)
  33 {
  34   /* Assumes incoming pointers are 16-byte aligned */
  35   __m128i t1 = *(__m128i *) i1;
  36   __m128i t2 = *(__m128i *) i2;
  37
  38   *(__m128i *) r = _mm_hsub_epi16 (t1, t2);
  39 }
  40
  41 /* Routine to manually compute the results */
  42 static void
  43 compute_correct_result (int *i1, int *i2, int *r)
  44 {
  45   short *s1 = (short *) i1;
  46   short *s2 = (short *) i2;
  47   short *sout = (short *) r;
  48   int i;
  49
  50   for (i = 0; i < 4; i++)
  51     sout[i] = s1[2 * i] - s1[2 * i + 1];
  52   for (i = 0; i < 4; i++)
  53     sout[i + 4] = s2[2 * i] - s2[2 * i + 1];
  54 }
  55
  56 static void
  57 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 #ifndef __AVX__
  70       /* Run the 64-bit tests */
  71       ssse3_test_phsubw (&vals[i + 0], &vals[i + 2], &r[0]);
  72       ssse3_test_phsubw (&vals[i + 4], &vals[i + 6], &r[2]);
  73       fail += chk_128 (ck, r);
  74 #endif
  75
  76       /* Run the 128-bit tests */
  77       ssse3_test_phsubw128 (&vals[i + 0], &vals[i + 4], r);
  78       fail += chk_128 (ck, r);
  79     }
  80
  81   if (fail != 0)
  82     abort ();
  83 }