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 14static void 15init_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 36static float 37do_round (float f, int type) 38{ 39 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__ ("flds %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 "fstps %0\n" : "=m" (*&ret)); 62 __asm__ ("fldcw %0" : : "m" (*&saved_cw)); 63 return ret; 64} 65 66static void 67sse4_1_test (void) 68{ 69 int i, j; 70 float f; 71 union 72 { 73 __m128 x[NUM / 4]; 74 float f[NUM]; 75 } dst, src; 76 77 init_round (src.f); 78 memset (&dst, 0, NUM * sizeof(float)); 79 80 for (i = 0; i < NUM / 4 ; i++) 81 dst.x[i] = _mm_round_ss (dst.x[i], src.x[i], _MM_FROUND_RINT); 82 83 for (i = 0; i < NUM; i += 4) 84 { 85 for (j = 0; j < 3; j++) 86 if (dst.f[i + j + 1] != 0.0) 87 abort (); 88 89 f = do_round (src.f[i], 0x04); 90 if (f != dst.f[i]) 91 abort (); 92 } 93 94 for (i = 0; i < NUM / 4 ; i++) 95 dst.x[i] = _mm_round_ss (dst.x[i], src.x[i], _MM_FROUND_NEARBYINT); 96 97 for (i = 0; i < NUM; i += 4) 98 { 99 for (j = 0; j < 3; j++) 100 if (dst.f[i + j + 1] != 0.0) 101 abort (); 102 103 f = do_round (src.f[i], 0x0c); 104 if (f != dst.f[i]) 105 abort (); 106 } 107} 108