1 /* { dg-require-effective-target vect_int } */
2
3 #include <stdarg.h>
4 #include "tree-vect.h"
5
6 #define N 64
7
8 int
9 main1 ()
10 {
11 int i;
12 unsigned short out[N*8];
13 unsigned short in[N*8];
14 unsigned int in2[N*8];
15 unsigned int out2[N*8];
16
17 for (i = 0; i < N*8; i++)
18 {
19 in[i] = in2[i] = i;
20 asm volatile ("" ::: "memory");
21 }
22
23 /* Induction is SLPable. */
24 for (i = 0; i < N; i++)
25 {
26 out[i*8] = in[i*8] + i;
27 out[i*8 + 1] = in[i*8 + 1] + i;
28 out[i*8 + 2] = in[i*8 + 2] + i;
29 out[i*8 + 3] = in[i*8 + 3] + i;
30 out[i*8 + 4] = in[i*8 + 4] + i;
31 out[i*8 + 5] = in[i*8 + 5] + i;
32 out[i*8 + 6] = in[i*8 + 6] + i;
33 out[i*8 + 7] = in[i*8 + 7] + i;
34 }
35
36 /* check results: */
37 for (i = 0; i < N; i++)
38 {
39 if (out[i*8] != in[i*8] + i
40 || out[i*8 + 1] != in[i*8 + 1] + i
41 || out[i*8 + 2] != in[i*8 + 2] + i
42 || out[i*8 + 3] != in[i*8 + 3] + i
43 || out[i*8 + 4] != in[i*8 + 4] + i
44 || out[i*8 + 5] != in[i*8 + 5] + i
45 || out[i*8 + 6] != in[i*8 + 6] + i
46 || out[i*8 + 7] != in[i*8 + 7] + i)
47 abort ();
48 }
49
50 /* Induction is SLPable. */
51 for (i = 0; i < N/2; i++)
52 {
53 out2[i*12] = in2[i*12] + i;
54 out2[i*12 + 1] = in2[i*12 + 1] + i;
55 out2[i*12 + 2] = in2[i*12 + 2] + i;
56 out2[i*12 + 3] = in2[i*12 + 3] + i;
57 out2[i*12 + 4] = in2[i*12 + 4] + i;
58 out2[i*12 + 5] = in2[i*12 + 5] + i;
59 out2[i*12 + 6] = in2[i*12 + 6] + i;
60 out2[i*12 + 7] = in2[i*12 + 7] + i;
61 out2[i*12 + 8] = in2[i*12 + 8] + i;
62 out2[i*12 + 9] = in2[i*12 + 9] + i;
63 out2[i*12 + 10] = in2[i*12 + 10] + i;
64 out2[i*12 + 11] = in2[i*12 + 11] + i;
65 }
66
67 /* check results: */
68 for (i = 0; i < N/2; i++)
69 {
70 if (out2[i*12] != in2[i*12] + i
71 || out2[i*12 + 1] != in2[i*12 + 1] + i
72 || out2[i*12 + 2] != in2[i*12 + 2] + i
73 || out2[i*12 + 3] != in2[i*12 + 3] + i
74 || out2[i*12 + 4] != in2[i*12 + 4] + i
75 || out2[i*12 + 5] != in2[i*12 + 5] + i
76 || out2[i*12 + 6] != in2[i*12 + 6] + i
77 || out2[i*12 + 7] != in2[i*12 + 7] + i
78 || out2[i*12 + 8] != in2[i*12 + 8] + i
79 || out2[i*12 + 9] != in2[i*12 + 9] + i
80 || out2[i*12 + 10] != in2[i*12 + 10] + i
81 || out2[i*12 + 11] != in2[i*12 + 11] + i)
82 abort ();
83 }
84
85 /* Not power of 2 but SLPable. */
86 for (i = 0; i < N/2; i++)
87 {
88 out2[i*12] = in2[i*12] + 1;
89 out2[i*12 + 1] = in2[i*12 + 1] + 2;
90 out2[i*12 + 2] = in2[i*12 + 2] + 3;
91 out2[i*12 + 3] = in2[i*12 + 3] + 4;
92 out2[i*12 + 4] = in2[i*12 + 4] + 5;
93 out2[i*12 + 5] = in2[i*12 + 5] + 6;
94 out2[i*12 + 6] = in2[i*12 + 6] + 7;
95 out2[i*12 + 7] = in2[i*12 + 7] + 8;
96 out2[i*12 + 8] = in2[i*12 + 8] + 9;
97 out2[i*12 + 9] = in2[i*12 + 9] + 10;
98 out2[i*12 + 10] = in2[i*12 + 10] + 11;
99 out2[i*12 + 11] = in2[i*12 + 11] + 12;
100 }
101
102 /* check results: */
103 for (i = 0; i < N/2; i++)
104 {
105 if (out2[i*12] != in2[i*12] + 1
106 || out2[i*12 + 1] != in2[i*12 + 1] + 2
107 || out2[i*12 + 2] != in2[i*12 + 2] + 3
108 || out2[i*12 + 3] != in2[i*12 + 3] + 4
109 || out2[i*12 + 4] != in2[i*12 + 4] + 5
110 || out2[i*12 + 5] != in2[i*12 + 5] + 6
111 || out2[i*12 + 6] != in2[i*12 + 6] + 7
112 || out2[i*12 + 7] != in2[i*12 + 7] + 8
113 || out2[i*12 + 8] != in2[i*12 + 8] + 9
114 || out2[i*12 + 9] != in2[i*12 + 9] + 10
115 || out2[i*12 + 10] != in2[i*12 + 10] + 11
116 || out2[i*12 + 11] != in2[i*12 + 11] + 12)
117 abort ();
118 }
119
120
121
122 return 0;
123 }
124
125 int main (void)
126 {
127 check_vect ();
128
129 main1 ();
130
131 return 0;
132 }
133
134 /* { dg-final { scan-tree-dump-times "vectorized 2 loops" 1 "vect" { target { { vect_interleave && vect_extract_even_odd } && { ! vect_pack_trunc } } } } } */
135 /* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 2 "vect" { target { ! vect_pack_trunc } } } } */
136 /* { dg-final { scan-tree-dump-times "vectorized 3 loops" 1 "vect" { target { { vect_interleave && vect_extract_even_odd } && vect_pack_trunc } } } } */
137 /* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 3 "vect" { target vect_pack_trunc xfail vect_variable_length } } } */