1 /* { dg-do run } */
2 /* { dg-require-effective-target sse4 } */
3 /* { dg-options "-O2 -msse4.1" } */
4 /* { dg-skip-if "no M_PI" { vxworks_kernel } } */
5
6 #include "sse4_1-check.h"
7
8 #include <smmintrin.h>
9 #include <math.h>
10 #include <string.h>
11
12 #define NUM 64
13
14 static void
init_round(float * src)15 init_round (float *src)
16 {
17 int i, sign = 1;
18 float f = rand ();
19
20 for (i = 0; i < NUM; i++)
21 {
22 src[i] = (i + 1)* f * M_PI * sign;
23 if (i < (NUM / 2))
24 {
25 if ((i % 6) == 0)
26 f = f * src[i];
27 }
28 else if (i == (NUM / 2))
29 f = rand ();
30 else if ((i % 6) == 0)
31 f = 1 / (f * (i + 1) * src[i] * M_PI *sign);
32 sign = -sign;
33 }
34 }
35
36 static float
do_round(float f,int type)37 do_round (float f, int type)
38 {
39 unsigned short saved_cw, new_cw, clr_mask;
40 float ret;
41
42 if ((type & 4))
43 {
44 type = 0;
45 clr_mask = 0xFFFF;
46 }
47 else
48 {
49 type = 0x003F | ((type & 3) << 10);
50 clr_mask = ~0x0C3F;
51 }
52
53 __asm__ ("fnstcw %0" : "=m" (saved_cw));
54
55 new_cw = saved_cw & clr_mask;
56 new_cw |= type;
57
58 __asm__ ("fldcw %2\n\t"
59 "frndint\n\t"
60 "fldcw %3" : "=t" (ret)
61 : "0" (f), "m" (new_cw), "m" (saved_cw));
62 return ret;
63 }
64
65 static void
sse4_1_test(void)66 sse4_1_test (void)
67 {
68 int i, j;
69 float f;
70 union
71 {
72 __m128 x[NUM / 4];
73 float f[NUM];
74 } dst, src;
75
76 init_round (src.f);
77 memset (&dst, 0, NUM * sizeof(float));
78
79 for (i = 0; i < NUM / 4 ; i++)
80 dst.x[i] = _mm_round_ss (dst.x[i], src.x[i], _MM_FROUND_RINT);
81
82 for (i = 0; i < NUM; i += 4)
83 {
84 for (j = 0; j < 3; j++)
85 if (dst.f[i + j + 1] != 0.0)
86 abort ();
87
88 f = do_round (src.f[i], 0x04);
89 if (f != dst.f[i])
90 abort ();
91 }
92
93 for (i = 0; i < NUM / 4 ; i++)
94 dst.x[i] = _mm_round_ss (dst.x[i], src.x[i], _MM_FROUND_NEARBYINT);
95
96 for (i = 0; i < NUM; i += 4)
97 {
98 for (j = 0; j < 3; j++)
99 if (dst.f[i + j + 1] != 0.0)
100 abort ();
101
102 f = do_round (src.f[i], 0x0c);
103 if (f != dst.f[i])
104 abort ();
105 }
106 }
107