On Sun, Oct 07, 2001 at 03:39:10AM +0200, Ralf Baechle wrote:
> On Fri, Oct 05, 2001 at 10:48:15AM -0700, Jun Sun wrote:
>
> > If your cpu does not have ll/sc instruction, you might be suffering the famous
> > sysmips() problem. The latest kernel should get you going.
> >
> > There is also FPU emulation bug which may cause this problem, but that only
> > happens on heavy context switches and FPU usages.
>
> I've checked in a major bundle of FPU emu fixes last week. The kernel
> fp code should now produce accurate results and handle exceptions and
> the Flush to Zero bit as per spec.
>
This particular problem seems still there. Anybody can try the following
test program on a CPU *without* fpu.
Jun
/* mipel-linux-gcc -O -o kfpe_chk3 kfpe_chk3.c */
#include <stdio.h>
#include <math.h>
#include <signal.h>
#include <sys/time.h>
double count;
int dummy;
struct itimerval ival;
int icnt=0;
static void alarm_sig_handler(int signum)
{
int i,j;
double d;
if( icnt++ > 500 ){
icnt = 0;
/* printf("count = %f\n", count);*/
}
for( i = 0, d = 0.0; d < 1.0; i++, d+=0.1 ){
count = d;
}
setitimer(ITIMER_REAL, &ival, NULL);
// alarm(1);
}
int main()
{
int i,j;
double d;
signal(SIGALRM, alarm_sig_handler);
ival.it_interval.tv_sec = 0;
ival.it_interval.tv_usec = 0;
ival.it_value.tv_sec = 0;
ival.it_value.tv_usec = 10000;
setitimer(ITIMER_REAL, &ival, NULL);
// alarm(1);
for(;;){
for( i = 0, d = 0.0; d < 10000.0; i++, d+=0.1 ){
count = d;
}
signal(SIGALRM, alarm_sig_handler);
}
return 0;
}
