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 */
21 static void
ssse3_test_phaddw(int * i1,int * i2,int * r)22 ssse3_test_phaddw (int *i1, int *i2, int *r)
23 {
24 __m64 t1 = *(__m64 *) i1;
25 __m64 t2 = *(__m64 *) i2;
26 *(__m64 *) r = _mm_hadd_pi16 (t1, t2);
27 _mm_empty ();
28 }
29 #endif
30
31 /* Test the 128-bit form */
32 static void
ssse3_test_phaddw128(int * i1,int * i2,int * r)33 ssse3_test_phaddw128 (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_hadd_epi16 (t1, t2);
39 }
40
41 /* Routine to manually compute the results */
42 static void
compute_correct_result(int * i1,int * i2,int * r)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
53 for (i = 0; i < 4; i++)
54 sout[i + 4] = s2[2 * i] + s2[2 * i + 1];
55 }
56
57 static void
TEST(void)58 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 += 8)
66 {
67 /* Manually compute the result */
68 compute_correct_result (&vals[i + 0], &vals[i + 4], ck);
69
70 #ifndef __AVX__
71 /* Run the 64-bit tests */
72 ssse3_test_phaddw (&vals[i + 0], &vals[i + 2], &r[0]);
73 ssse3_test_phaddw (&vals[i + 4], &vals[i + 6], &r[2]);
74 fail += chk_128 (ck, r);
75 #endif
76
77 /* Run the 128-bit tests */
78 ssse3_test_phaddw128 (&vals[i + 0], &vals[i + 4], r);
79 fail += chk_128 (ck, r);
80 }
81
82 if (fail != 0)
83 abort ();
84 }
85