2 /* { dg-require-effective-target sse4 } */
3 /* { dg-options "-O2 -msse4.1" } */
4 /* { dg-skip-if "no M_PI" { vxworks_kernel } } */
6 #include "sse4_1-check.h"
15 init_round (float *src)
20 for (i = 0; i < NUM; i++)
22 src[i] = (i + 1)* f * M_PI * sign;
28 else if (i == (NUM / 2))
30 else if ((i % 6) == 0)
31 f = 1 / (f * (i + 1) * src[i] * M_PI *sign);
37 do_round (float f, int type)
39 short saved_cw, new_cw, clr_mask;
49 type = 0x003F | ((type & 3) << 10);
53 __asm__ ("flds %0" : : "m" (*&f));
55 __asm__ ("fstcw %0" : "=m" (*&saved_cw));
56 new_cw = saved_cw & clr_mask;
58 __asm__ ("fldcw %0" : : "m" (*&new_cw));
61 "fstps %0\n" : "=m" (*&ret));
62 __asm__ ("fldcw %0" : : "m" (*&saved_cw));
78 memset (&dst, 0, NUM * sizeof(float));
80 for (i = 0; i < NUM / 4 ; i++)
81 dst.x[i] = _mm_round_ss (dst.x[i], src.x[i], _MM_FROUND_RINT);
83 for (i = 0; i < NUM; i += 4)
85 for (j = 0; j < 3; j++)
86 if (dst.f[i + j + 1] != 0.0)
89 f = do_round (src.f[i], 0x04);
94 for (i = 0; i < NUM / 4 ; i++)
95 dst.x[i] = _mm_round_ss (dst.x[i], src.x[i], _MM_FROUND_NEARBYINT);
97 for (i = 0; i < NUM; i += 4)
99 for (j = 0; j < 3; j++)
100 if (dst.f[i + j + 1] != 0.0)
103 f = do_round (src.f[i], 0x0c);