111
|
1 /* { dg-require-effective-target vect_int } */
|
|
2
|
|
3 #include <stdarg.h>
|
|
4 #include "tree-vect.h"
|
|
5
|
|
6 #define N 96
|
|
7
|
|
8 unsigned short in[N*8];
|
|
9
|
|
10 int
|
|
11 main1 ()
|
|
12 {
|
|
13 int i;
|
|
14 unsigned short out[N*8];
|
|
15
|
|
16 for (i = 0; i < N*8; i++)
|
|
17 {
|
|
18 in[i] = i&63;
|
131
|
19 asm volatile ("" ::: "memory");
|
111
|
20 }
|
|
21
|
|
22 for (i = 0; i < N; i++)
|
|
23 {
|
|
24 out[i*8] = in[i*8];
|
|
25 out[i*8 + 1] = in[i*8 + 1];
|
|
26 out[i*8 + 2] = in[i*8 + 2];
|
|
27 out[i*8 + 3] = in[i*8 + 3];
|
|
28 out[i*8 + 4] = in[i*8 + 4];
|
|
29 out[i*8 + 5] = in[i*8 + 5];
|
|
30 out[i*8 + 6] = in[i*8 + 6];
|
|
31 out[i*8 + 7] = in[i*8 + 7];
|
|
32 }
|
|
33
|
|
34 /* check results: */
|
|
35 for (i = 0; i < N; i++)
|
|
36 {
|
|
37 if (out[i*8] != in[i*8]
|
|
38 || out[i*8 + 1] != in[i*8 + 1]
|
|
39 || out[i*8 + 2] != in[i*8 + 2]
|
|
40 || out[i*8 + 3] != in[i*8 + 3]
|
|
41 || out[i*8 + 4] != in[i*8 + 4]
|
|
42 || out[i*8 + 5] != in[i*8 + 5]
|
|
43 || out[i*8 + 6] != in[i*8 + 6]
|
|
44 || out[i*8 + 7] != in[i*8 + 7])
|
|
45 abort ();
|
|
46 }
|
|
47
|
|
48 for (i = 0; i < N*2; i++)
|
|
49 {
|
|
50 out[i*4] = in[i*4];
|
|
51 out[i*4 + 1] = in[i*4 + 1];
|
|
52 out[i*4 + 2] = in[i*4 + 2];
|
|
53 out[i*4 + 3] = in[i*4 + 3];
|
|
54 }
|
|
55
|
|
56 /* check results: */
|
|
57 for (i = 0; i < N*2; i++)
|
|
58 {
|
|
59 if (out[i*4] != in[i*4]
|
|
60 || out[i*4 + 1] != in[i*4 + 1]
|
|
61 || out[i*4 + 2] != in[i*4 + 2]
|
|
62 || out[i*4 + 3] != in[i*4 + 3])
|
|
63 abort ();
|
|
64 }
|
|
65
|
|
66 for (i = 0; i < N/2; i++)
|
|
67 {
|
|
68 out[i*16] = in[i*16];
|
|
69 out[i*16 + 1] = in[i*16 + 1];
|
|
70 out[i*16 + 2] = in[i*16 + 2];
|
|
71 out[i*16 + 3] = in[i*16 + 3];
|
|
72 out[i*16 + 4] = in[i*16 + 4];
|
|
73 out[i*16 + 5] = in[i*16 + 5];
|
|
74 out[i*16 + 6] = in[i*16 + 6];
|
|
75 out[i*16 + 7] = in[i*16 + 7];
|
|
76 out[i*16 + 8] = in[i*16 + 8];
|
|
77 out[i*16 + 9] = in[i*16 + 9];
|
|
78 out[i*16 + 10] = in[i*16 + 10];
|
|
79 out[i*16 + 11] = in[i*16 + 11];
|
|
80 out[i*16 + 12] = in[i*16 + 12];
|
|
81 out[i*16 + 13] = in[i*16 + 13];
|
|
82 out[i*16 + 14] = in[i*16 + 14];
|
|
83 out[i*16 + 15] = in[i*16 + 15];
|
|
84 }
|
|
85
|
|
86 /* check results: */
|
|
87 for (i = 0; i < N/2; i++)
|
|
88 {
|
|
89 if (out[i*16] != in[i*16]
|
|
90 || out[i*16 + 1] != in[i*16 + 1]
|
|
91 || out[i*16 + 2] != in[i*16 + 2]
|
|
92 || out[i*16 + 3] != in[i*16 + 3]
|
|
93 || out[i*16 + 4] != in[i*16 + 4]
|
|
94 || out[i*16 + 5] != in[i*16 + 5]
|
|
95 || out[i*16 + 6] != in[i*16 + 6]
|
|
96 || out[i*16 + 7] != in[i*16 + 7]
|
|
97 || out[i*16 + 8] != in[i*16 + 8]
|
|
98 || out[i*16 + 9] != in[i*16 + 9]
|
|
99 || out[i*16 + 10] != in[i*16 + 10]
|
|
100 || out[i*16 + 11] != in[i*16 + 11]
|
|
101 || out[i*16 + 12] != in[i*16 + 12]
|
|
102 || out[i*16 + 13] != in[i*16 + 13]
|
|
103 || out[i*16 + 14] != in[i*16 + 14]
|
|
104 || out[i*16 + 15] != in[i*16 + 15])
|
|
105 abort ();
|
|
106 }
|
|
107
|
|
108 /* SLP with unrolling by 8. */
|
|
109 for (i = 0; i < N/4; i++)
|
|
110 {
|
|
111 out[i*9] = in[i*9];
|
|
112 out[i*9 + 1] = in[i*9 + 1];
|
|
113 out[i*9 + 2] = in[i*9 + 2];
|
|
114 out[i*9 + 3] = in[i*9 + 3];
|
|
115 out[i*9 + 4] = in[i*9 + 4];
|
|
116 out[i*9 + 5] = in[i*9 + 5];
|
|
117 out[i*9 + 6] = in[i*9 + 6];
|
|
118 out[i*9 + 7] = in[i*9 + 7];
|
|
119 out[i*9 + 8] = in[i*9 + 8];
|
|
120 }
|
|
121
|
|
122 /* check results: */
|
|
123 for (i = 0; i < N/4; i++)
|
|
124 {
|
|
125 if (out[i*9] != in[i*9]
|
|
126 || out[i*9 + 1] != in[i*9 + 1]
|
|
127 || out[i*9 + 2] != in[i*9 + 2]
|
|
128 || out[i*9 + 3] != in[i*9 + 3]
|
|
129 || out[i*9 + 4] != in[i*9 + 4]
|
|
130 || out[i*9 + 5] != in[i*9 + 5]
|
|
131 || out[i*9 + 6] != in[i*9 + 6]
|
|
132 || out[i*9 + 7] != in[i*9 + 7]
|
|
133 || out[i*9 + 8] != in[i*9 + 8])
|
|
134 abort ();
|
|
135 }
|
|
136
|
|
137
|
|
138 return 0;
|
|
139 }
|
|
140
|
|
141 int main (void)
|
|
142 {
|
|
143 check_vect ();
|
|
144
|
|
145 main1 ();
|
|
146
|
|
147 return 0;
|
|
148 }
|
|
149
|
|
150 /* { dg-final { scan-tree-dump-times "vectorized 4 loops" 1 "vect" } } */
|
|
151 /* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 4 "vect" } } */
|
|
152
|