111
|
1 #include "tree-vect.h"
|
|
2
|
|
3 extern float copysignf (float, float);
|
|
4 extern float sqrtf (float);
|
|
5 extern float fabsf (float);
|
|
6 extern void abort (void);
|
|
7 float a[64], b[64], c[64], d[64];
|
|
8
|
|
9 __attribute__((noinline, noclone)) void
|
|
10 f1 (int n)
|
|
11 {
|
|
12 int i;
|
|
13 for (i = 0; i < n; i++)
|
|
14 {
|
|
15 a[4 * i + 0] = copysignf (b[4 * i + 0], c[4 * i + 0]) + 1.0f + sqrtf (d[4 * i + 0]);
|
|
16 a[4 * i + 1] = copysignf (b[4 * i + 1], c[4 * i + 1]) + 2.0f + sqrtf (d[4 * i + 1]);
|
|
17 a[4 * i + 2] = copysignf (b[4 * i + 2], c[4 * i + 2]) + 3.0f + sqrtf (d[4 * i + 2]);
|
|
18 a[4 * i + 3] = copysignf (b[4 * i + 3], c[4 * i + 3]) + 4.0f + sqrtf (d[4 * i + 3]);
|
|
19 }
|
|
20 }
|
|
21
|
|
22 __attribute__((noinline, noclone)) void
|
|
23 f2 (int n)
|
|
24 {
|
|
25 int i;
|
|
26 for (i = 0; i < 2 * n; i++)
|
|
27 {
|
|
28 a[2 * i + 0] = copysignf (b[2 * i + 0], c[2 * i + 0]) + 1.0f + sqrtf (d[2 * i + 0]);
|
|
29 a[2 * i + 1] = copysignf (b[2 * i + 1], c[2 * i + 1]) + 2.0f + sqrtf (d[2 * i + 1]);
|
|
30 }
|
|
31 }
|
|
32
|
|
33 __attribute__((noinline, noclone)) void
|
|
34 f3 (void)
|
|
35 {
|
|
36 int i;
|
|
37 for (i = 0; i < 64; i++)
|
|
38 a[i] = copysignf (b[i], c[i]) + 1.0f + sqrtf (d[i]);
|
|
39 }
|
|
40
|
|
41 __attribute__((noinline, noclone)) void
|
|
42 f4 (int n)
|
|
43 {
|
|
44 int i;
|
|
45 for (i = 0; i < 2 * n; i++)
|
|
46 {
|
|
47 a[3 * i + 0] = copysignf (b[3 * i + 0], c[3 * i + 0]) + 1.0f + sqrtf (d[3 * i + 0]);
|
|
48 a[3 * i + 1] = copysignf (b[3 * i + 1], c[3 * i + 1]) + 2.0f + sqrtf (d[3 * i + 1]);
|
|
49 a[3 * i + 2] = copysignf (b[3 * i + 2], c[3 * i + 2]) + 3.0f + sqrtf (d[3 * i + 2]);
|
|
50 }
|
|
51 }
|
|
52
|
|
53 __attribute__((noinline, noclone)) int
|
|
54 main1 ()
|
|
55 {
|
|
56 int i;
|
|
57
|
|
58 for (i = 0; i < 64; i++)
|
|
59 {
|
|
60 asm ("");
|
|
61 b[i] = (i & 1) ? -4 * i : 4 * i;
|
|
62 c[i] = (i & 2) ? -8 * i : 8 * i;
|
|
63 d[i] = i * i;
|
|
64 }
|
|
65 f1 (16);
|
|
66 for (i = 0; i < 64; i++)
|
|
67 if (fabsf (((i & 2) ? -4 * i : 4 * i) + 1 + (i & 3) + i - a[i]) >= 0.0001f)
|
|
68 abort ();
|
|
69 else
|
|
70 a[i] = 131.25;
|
|
71 f2 (16);
|
|
72 for (i = 0; i < 64; i++)
|
|
73 if (fabsf (((i & 2) ? -4 * i : 4 * i) + 1 + (i & 1) + i - a[i]) >= 0.0001f)
|
|
74 abort ();
|
|
75 else
|
|
76 a[i] = 131.25;
|
|
77 f3 ();
|
|
78 for (i = 0; i < 64; i++)
|
|
79 if (fabsf (((i & 2) ? -4 * i : 4 * i) + 1 + i - a[i]) >= 0.0001f)
|
|
80 abort ();
|
|
81 else
|
|
82 a[i] = 131.25;
|
|
83 f4 (10);
|
|
84 for (i = 0; i < 60; i++)
|
|
85 if (fabsf (((i & 2) ? -4 * i : 4 * i) + 1 + (i % 3) + i - a[i]) >= 0.0001f)
|
|
86 abort ();
|
|
87 return 0;
|
|
88 }
|
|
89
|
|
90 int
|
|
91 main ()
|
|
92 {
|
|
93 check_vect ();
|
|
94 return main1 ();
|
|
95 }
|
|
96
|
|
97 /* { dg-final { scan-tree-dump-times "vectorized 1 loops" 4 "vect" { target { vect_call_copysignf && vect_call_sqrtf } } } } */
|
|
98 /* { dg-final { scan-tree-dump-times "vectorizing stmts using SLP" 3 "vect" { target { vect_call_copysignf && vect_call_sqrtf } } } } */
|