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(double * src)15 init_round (double *src)
16 {
17 int i, sign = 1;
18 double d = rand ();
19
20 for (i = 0; i < NUM; i++)
21 {
22 src[i] = (i + 1)* d * M_PI * sign;
23 if (i < (NUM / 2))
24 {
25 if ((i % 6) == 0)
26 d = d * src[i];
27 }
28 else if (i == (NUM / 2))
29 d = rand ();
30 else if ((i % 6) == 0)
31 d = 1 / (d * (i + 1) * src[i] * M_PI *sign);
32 sign = -sign;
33 }
34 }
35
36 static double
do_round(double f,int type)37 do_round (double f, int type)
38 {
39 short saved_cw, new_cw, clr_mask;
40 double 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__ ("fldl %0" : : "m" (*&f));
54
55 __asm__ ("fstcw %0" : "=m" (*&saved_cw));
56 new_cw = saved_cw & clr_mask;
57 new_cw |= type;
58 __asm__ ("fldcw %0" : : "m" (*&new_cw));
59
60 __asm__ ("frndint\n"
61 "fstpl %0\n" : "=m" (*&ret));
62 __asm__ ("fldcw %0" : : "m" (*&saved_cw));
63 return ret;
64 }
65
66 static void
sse4_1_test(void)67 sse4_1_test (void)
68 {
69 int i;
70 double f;
71 union
72 {
73 __m128d x[NUM / 2];
74 double d[NUM];
75 } dst, src;
76
77 init_round (src.d);
78 memset (&dst, 0, NUM * sizeof(double));
79
80 for (i = 0; i < NUM / 2 ; i++)
81 dst.x[i] = _mm_round_sd (dst.x[i], src.x[i], _MM_FROUND_TRUNC);
82
83 for (i = 0; i < NUM; i += 2)
84 {
85 if (dst.d[i + 1] != 0.0)
86 abort ();
87
88 f = do_round (src.d[i], 0x03);
89 if (f != dst.d[i])
90 abort ();
91 }
92 }
93