/* { dg-do run } */
|
/* { dg-do run } */
|
/* { dg-require-effective-target sse4 } */
|
/* { dg-require-effective-target sse4 } */
|
/* { dg-options "-O2 -msse4.1" } */
|
/* { dg-options "-O2 -msse4.1" } */
|
/* { dg-skip-if "no M_PI" { vxworks_kernel } } */
|
/* { dg-skip-if "no M_PI" { vxworks_kernel } } */
|
|
|
#include "sse4_1-check.h"
|
#include "sse4_1-check.h"
|
|
|
#include <smmintrin.h>
|
#include <smmintrin.h>
|
#include <math.h>
|
#include <math.h>
|
#include <string.h>
|
#include <string.h>
|
|
|
#define NUM 64
|
#define NUM 64
|
|
|
static void
|
static void
|
init_round (double *src)
|
init_round (double *src)
|
{
|
{
|
int i, sign = 1;
|
int i, sign = 1;
|
double d = rand ();
|
double d = rand ();
|
|
|
for (i = 0; i < NUM; i++)
|
for (i = 0; i < NUM; i++)
|
{
|
{
|
src[i] = (i + 1)* d * M_PI * sign;
|
src[i] = (i + 1)* d * M_PI * sign;
|
if (i < (NUM / 2))
|
if (i < (NUM / 2))
|
{
|
{
|
if ((i % 6) == 0)
|
if ((i % 6) == 0)
|
d = d * src[i];
|
d = d * src[i];
|
}
|
}
|
else if (i == (NUM / 2))
|
else if (i == (NUM / 2))
|
d = rand ();
|
d = rand ();
|
else if ((i % 6) == 0)
|
else if ((i % 6) == 0)
|
d = 1 / (d * (i + 1) * src[i] * M_PI *sign);
|
d = 1 / (d * (i + 1) * src[i] * M_PI *sign);
|
sign = -sign;
|
sign = -sign;
|
}
|
}
|
}
|
}
|
|
|
static double
|
static double
|
do_round (double f, int type)
|
do_round (double f, int type)
|
{
|
{
|
short saved_cw, new_cw, clr_mask;
|
short saved_cw, new_cw, clr_mask;
|
double ret;
|
double ret;
|
|
|
if ((type & 4))
|
if ((type & 4))
|
{
|
{
|
type = 0;
|
type = 0;
|
clr_mask = 0xFFFF;
|
clr_mask = 0xFFFF;
|
}
|
}
|
else
|
else
|
{
|
{
|
type = 0x003F | ((type & 3) << 10);
|
type = 0x003F | ((type & 3) << 10);
|
clr_mask = ~0x0C3F;
|
clr_mask = ~0x0C3F;
|
}
|
}
|
|
|
__asm__ ("fldl %0" : : "m" (*&f));
|
__asm__ ("fldl %0" : : "m" (*&f));
|
|
|
__asm__ ("fstcw %0" : "=m" (*&saved_cw));
|
__asm__ ("fstcw %0" : "=m" (*&saved_cw));
|
new_cw = saved_cw & clr_mask;
|
new_cw = saved_cw & clr_mask;
|
new_cw |= type;
|
new_cw |= type;
|
__asm__ ("fldcw %0" : : "m" (*&new_cw));
|
__asm__ ("fldcw %0" : : "m" (*&new_cw));
|
|
|
__asm__ ("frndint\n"
|
__asm__ ("frndint\n"
|
"fstpl %0\n" : "=m" (*&ret));
|
"fstpl %0\n" : "=m" (*&ret));
|
__asm__ ("fldcw %0" : : "m" (*&saved_cw));
|
__asm__ ("fldcw %0" : : "m" (*&saved_cw));
|
return ret;
|
return ret;
|
}
|
}
|
|
|
static void
|
static void
|
sse4_1_test (void)
|
sse4_1_test (void)
|
{
|
{
|
int i;
|
int i;
|
double f;
|
double f;
|
union
|
union
|
{
|
{
|
__m128d x[NUM / 2];
|
__m128d x[NUM / 2];
|
double d[NUM];
|
double d[NUM];
|
} dst, src;
|
} dst, src;
|
|
|
init_round (src.d);
|
init_round (src.d);
|
memset (&dst, 0, NUM * sizeof(double));
|
memset (&dst, 0, NUM * sizeof(double));
|
|
|
for (i = 0; i < NUM / 2 ; i++)
|
for (i = 0; i < NUM / 2 ; i++)
|
dst.x[i] = _mm_round_sd (dst.x[i], src.x[i], _MM_FROUND_TRUNC);
|
dst.x[i] = _mm_round_sd (dst.x[i], src.x[i], _MM_FROUND_TRUNC);
|
|
|
for (i = 0; i < NUM; i += 2)
|
for (i = 0; i < NUM; i += 2)
|
{
|
{
|
if (dst.d[i + 1] != 0.0)
|
if (dst.d[i + 1] != 0.0)
|
abort ();
|
abort ();
|
|
|
f = do_round (src.d[i], 0x03);
|
f = do_round (src.d[i], 0x03);
|
if (f != dst.d[i])
|
if (f != dst.d[i])
|
abort ();
|
abort ();
|
}
|
}
|
}
|
}
|
|
|