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
15#include "ssse3-vals.h"
16
17#include <tmmintrin.h>
18
19#ifndef __AVX__
20/* Test the 64-bit form */
21static void
22ssse3_test_pshufb (int *i1, int *i2, int *r)
23{
24  __m64 t1 = *(__m64 *) i1;
25  __m64 t2 = *(__m64 *) i2;
26  *(__m64 *)r = _mm_shuffle_pi8 (t1, t2);
27  _mm_empty ();
28}
29#endif
30
31/* Test the 128-bit form */
32static void
33ssse3_test_pshufb128 (int *i1, int *i2, int *r)
34{
35  /* Assumes incoming pointers are 16-byte aligned */
36  __m128i t1 = *(__m128i *) i1;
37  __m128i t2 = *(__m128i *) i2;
38  *(__m128i *)r = _mm_shuffle_epi8 (t1, t2);
39}
40
41#ifndef __AVX__
42/* Routine to manually compute the results */
43static void
44compute_correct_result_64 (int *i1, int *i2, int *r)
45{
46  char *b1 = (char *) i1;
47  char *b2 = (char *) i2;
48  char *bout = (char *) r;
49  int i;
50  char select;
51
52  for (i = 0; i < 16; i++)
53    {
54      select = b2[i];
55      if (select & 0x80)
56	bout[i] = 0;
57      else if (i < 8)
58	bout[i] = b1[select & 0x7];
59      else
60	bout[i] = b1[8 + (select & 0x7)];
61    }
62}
63#endif
64
65static void
66compute_correct_result_128 (int *i1, int *i2, int *r)
67{
68  char *b1 = (char *) i1;
69  char *b2 = (char *) i2;
70  char *bout = (char *) r;
71  int i;
72  char select;
73
74  for (i = 0; i < 16; i++)
75    {
76      select = b2[i];
77      if (select & 0x80)
78	bout[i] = 0;
79      else
80	bout[i] = b1[select & 0xf];
81    }
82}
83
84static void
85TEST (void)
86{
87  int i;
88  int r [4] __attribute__ ((aligned(16)));
89  int ck [4];
90  int fail = 0;
91
92  for (i = 0; i < 256; i += 8)
93    {
94#ifndef __AVX__
95      /* Manually compute the result */
96      compute_correct_result_64 (&vals[i + 0], &vals[i + 4], ck);
97
98      /* Run the 64-bit tests */
99      ssse3_test_pshufb (&vals[i + 0], &vals[i + 4], &r[0]);
100      ssse3_test_pshufb (&vals[i + 2], &vals[i + 6], &r[2]);
101      fail += chk_128 (ck, r);
102#endif
103
104      /* Recompute the result for 128-bits */
105      compute_correct_result_128 (&vals[i + 0], &vals[i + 4], ck);
106
107      /* Run the 128-bit tests */
108      ssse3_test_pshufb128 (&vals[i + 0], &vals[i + 4], r);
109      fail += chk_128 (ck, r);
110    }
111
112  if (fail != 0)
113    abort ();
114}
115