Merged with mainline at revision 128810.
[official-gcc.git] / gcc / testsuite / gcc.target / i386 / sse4_1-roundss-4.c
blobdc7f34edb5b4147c33d3ef80e943598d7d03601b
1 /* { dg-do run } */
2 /* { dg-require-effective-target sse4 } */
3 /* { dg-options "-O2 -msse4.1" } */
5 #include "sse4_1-check.h"
7 #include <smmintrin.h>
8 #include <math.h>
9 #include <string.h>
11 #define NUM 64
13 static void
14 init_round (float *src)
16 int i, sign = 1;
17 float f = rand ();
19 for (i = 0; i < NUM; i++)
21 src[i] = (i + 1)* f * M_PI * sign;
22 if (i < (NUM / 2))
24 if ((i % 6) == 0)
25 f = f * src[i];
27 else if (i == (NUM / 2))
28 f = rand ();
29 else if ((i % 6) == 0)
30 f = 1 / (f * (i + 1) * src[i] * M_PI *sign);
31 sign = -sign;
35 static float
36 do_round (float f, int type)
38 short saved_cw, new_cw, clr_mask;
39 float ret;
41 if ((type & 4))
43 type = 0;
44 clr_mask = 0xFFFF;
46 else
48 type = 0x003F | ((type & 3) << 10);
49 clr_mask = ~0x0C3F;
52 __asm__ ("flds %0" : : "m" (*&f));
54 __asm__ ("fstcw %0" : "=m" (*&saved_cw));
55 new_cw = saved_cw & clr_mask;
56 new_cw |= type;
57 __asm__ ("fldcw %0" : : "m" (*&new_cw));
59 __asm__ ("frndint\n"
60 "fstps %0\n" : "=m" (*&ret));
61 __asm__ ("fldcw %0" : : "m" (*&saved_cw));
62 return ret;
65 static void
66 sse4_1_test (void)
68 int i, j;
69 float f;
70 union
72 __m128 x[NUM / 4];
73 float f[NUM];
74 } dst, src;
76 init_round (src.f);
77 memset (&dst, 0, NUM * sizeof(float));
79 for (i = 0; i < NUM / 4 ; i++)
80 dst.x[i] = _mm_round_ss (dst.x[i], src.x[i], _MM_FROUND_RINT);
82 for (i = 0; i < NUM; i += 4)
84 for (j = 0; j < 3; j++)
85 if (dst.f[i + j + 1] != 0.0)
86 abort ();
88 f = do_round (src.f[i], 0x04);
89 if (f != dst.f[i])
90 abort ();
93 for (i = 0; i < NUM / 4 ; i++)
94 dst.x[i] = _mm_round_ss (dst.x[i], src.x[i], _MM_FROUND_NEARBYINT);
96 for (i = 0; i < NUM; i += 4)
98 for (j = 0; j < 3; j++)
99 if (dst.f[i + j + 1] != 0.0)
100 abort ();
102 f = do_round (src.f[i], 0x0c);
103 if (f != dst.f[i])
104 abort ();