2 /* { dg-require-effective-target sse4 } */
3 /* { dg-options "-O2 -msse4.1" } */
5 #include "sse4_1-check.h"
14 init_round (float *src)
19 for (i = 0; i < NUM; i++)
21 src[i] = (i + 1)* f * M_PI * sign;
27 else if (i == (NUM / 2))
29 else if ((i % 6) == 0)
30 f = 1 / (f * (i + 1) * src[i] * M_PI *sign);
36 do_round (float f, int type)
38 short saved_cw, new_cw, clr_mask;
48 type = 0x003F | ((type & 3) << 10);
52 __asm__ ("flds %0" : : "m" (*&f));
54 __asm__ ("fstcw %0" : "=m" (*&saved_cw));
55 new_cw = saved_cw & clr_mask;
57 __asm__ ("fldcw %0" : : "m" (*&new_cw));
60 "fstps %0\n" : "=m" (*&ret));
61 __asm__ ("fldcw %0" : : "m" (*&saved_cw));
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)
88 f = do_round (src.f[i], 0x04);
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)
102 f = do_round (src.f[i], 0x0c);