1 /* { dg-do run } */
2 /* { dg-require-effective-target xop } */
3 /* { dg-options "-O2 -mxop" } */
4
5 #include "xop-check.h"
6
7 #include <x86intrin.h>
8 #include <string.h>
9
10 #define NUM 10
11
12 union
13 {
14 __m128i x[NUM];
15 unsigned char ssi[NUM * 16];
16 unsigned short si[NUM * 8];
17 unsigned int li[NUM * 4];
18 unsigned long long lli[NUM * 2];
19 } dst, res, src1;
20
21 static void
init_byte()22 init_byte ()
23 {
24 int i;
25 for (i=0; i < NUM * 16; i++)
26 src1.ssi[i] = i;
27 }
28
29 static void
init_word()30 init_word ()
31 {
32 int i;
33 for (i=0; i < NUM * 8; i++)
34 src1.si[i] = i;
35 }
36
37
38 static void
init_dword()39 init_dword ()
40 {
41 int i;
42 for (i=0; i < NUM * 4; i++)
43 src1.li[i] = i;
44 }
45
46 static int
check_byte2word()47 check_byte2word ()
48 {
49 int i, j, s, t, check_fails = 0;
50 for (i = 0; i < NUM * 16; i = i + 16)
51 {
52 for (j = 0; j < 8; j++)
53 {
54 t = i + (2 * j);
55 s = (i / 2) + j;
56 res.si[s] = src1.ssi[t] + src1.ssi[t + 1] ;
57 if (res.si[s] != dst.si[s])
58 check_fails++;
59 }
60 }
61 }
62
63 static int
check_byte2dword()64 check_byte2dword ()
65 {
66 int i, j, s, t, check_fails = 0;
67 for (i = 0; i < NUM * 16; i = i + 16)
68 {
69 for (j = 0; j < 4; j++)
70 {
71 t = i + (4 * j);
72 s = (i / 4) + j;
73 res.li[s] = (src1.ssi[t] + src1.ssi[t + 1]) + (src1.ssi[t + 2]
74 + src1.ssi[t + 3]);
75 if (res.li[s] != dst.li[s])
76 check_fails++;
77 }
78 }
79 return check_fails++;
80 }
81
82 static int
check_byte2qword()83 check_byte2qword ()
84 {
85 int i, j, s, t, check_fails = 0;
86 for (i = 0; i < NUM * 16; i = i + 16)
87 {
88 for (j = 0; j < 2; j++)
89 {
90 t = i + (8 * j);
91 s = (i / 8) + j;
92 res.lli[s] = ((src1.ssi[t] + src1.ssi[t + 1]) + (src1.ssi[t + 2]
93 + src1.ssi[t + 3])) + ((src1.ssi[t + 4] + src1.ssi[t +5])
94 + (src1.ssi[t + 6] + src1.ssi[t + 7]));
95 if (res.lli[s] != dst.lli[s])
96 check_fails++;
97 }
98 }
99 return check_fails++;
100 }
101
102 static int
check_word2dword()103 check_word2dword ()
104 {
105 int i, j, s, t, check_fails = 0;
106 for (i = 0; i < (NUM * 8); i = i + 8)
107 {
108 for (j = 0; j < 4; j++)
109 {
110 t = i + (2 * j);
111 s = (i / 2) + j;
112 res.li[s] = src1.si[t] + src1.si[t + 1] ;
113 if (res.li[s] != dst.li[s])
114 check_fails++;
115 }
116 }
117 }
118
119 static int
check_word2qword()120 check_word2qword ()
121 {
122 int i, j, s, t, check_fails = 0;
123 for (i = 0; i < NUM * 8; i = i + 8)
124 {
125 for (j = 0; j < 2; j++)
126 {
127 t = i + (4 * j);
128 s = (i / 4) + j;
129 res.lli[s] = (src1.si[t] + src1.si[t + 1]) + (src1.si[t + 2]
130 + src1.si[t + 3]);
131 if (res.lli[s] != dst.lli[s])
132 check_fails++;
133 }
134 }
135 return check_fails++;
136 }
137
138 static int
check_dword2qword()139 check_dword2qword ()
140 {
141 int i, j, s, t, check_fails = 0;
142 for (i = 0; i < (NUM * 4); i = i + 4)
143 {
144 for (j = 0; j < 2; j++)
145 {
146 t = i + (2 * j);
147 s = (i / 2) + j;
148 res.lli[s] = src1.li[t] + src1.li[t + 1] ;
149 if (res.lli[s] != dst.lli[s])
150 check_fails++;
151 }
152 }
153 }
154
155 static void
xop_test(void)156 xop_test (void)
157 {
158 int i;
159
160 /* Check haddubw */
161 init_byte ();
162
163 for (i = 0; i < NUM; i++)
164 dst.x[i] = _mm_haddw_epu8 (src1.x[i]);
165
166 if (check_byte2word())
167 abort ();
168
169 /* Check haddubd */
170 for (i = 0; i < (NUM ); i++)
171 dst.x[i] = _mm_haddd_epu8 (src1.x[i]);
172
173 if (check_byte2dword())
174 abort ();
175
176 /* Check haddubq */
177 for (i = 0; i < NUM; i++)
178 dst.x[i] = _mm_haddq_epu8 (src1.x[i]);
179
180 if (check_byte2qword())
181 abort ();
182
183 /* Check hadduwd */
184 init_word ();
185
186 for (i = 0; i < (NUM ); i++)
187 dst.x[i] = _mm_haddd_epu16 (src1.x[i]);
188
189 if (check_word2dword())
190 abort ();
191
192 /* Check haddbuwq */
193
194 for (i = 0; i < NUM; i++)
195 dst.x[i] = _mm_haddq_epu16 (src1.x[i]);
196
197 if (check_word2qword())
198 abort ();
199
200 /* Check hadudq */
201 init_dword ();
202 for (i = 0; i < NUM; i++)
203 dst.x[i] = _mm_haddq_epu32 (src1.x[i]);
204
205 if (check_dword2qword())
206 abort ();
207 }
208