Re: Strange optimization results

[kevin diggs <diggskevin38@xxxxxxxxx>]

2011/4/11 Дмитрий Оксенчук <oksenchuk89@xxxxxxxxx>:
>
> Yes, I run tests 3 times and this is average results. I used time
> utility. Time for create and run this program without loop is 0.002s,
> so this time is not significant. System busy was little and the same
> in all tests.
>
Wouldn't it be better to use something like getrusage()? Like in the
attached bit of silliness?

kevin

#include <stdio.h>
#include <math.h>
#include <stdlib.h>
#include <limits.h>
#include <sys/resource.h>

volatile inline double m_pi_minus_fldpi(void)
{
double diff,pi=M_PI;

	asm(	"fldpi\n\t"
		"fsubp\n"
		:"=t" (diff)
		:"0" (pi)
	);

	return diff;
}

volatile inline float xa(const float a, const float b)
{
	return (a*M_PI)+1.0/b;
}

volatile inline float xb(const float a, const float b)
{
float c;
	/*
	 * (a*M_PI)+1.0/b
	 */
	asm(	"fldpi\n\t"
		"fmulp\n\t"
		"fld1\n\t"
		"fdivp %%st,%%st(2)\n\t"
		"faddp"
		:"=t" (c)
		:"0" (a), "u" (b)
		:"st(1)"
	);

	return c;
}

int main1(int argc, char *argv[])
{
float a,b,res0,res1;

	if(argc<3) {
		fprintf(stderr,__FILE__"`%s()-%d:  usage:  %s clock a b\n",
			__func__,__LINE__,argv[0]);

		return -1;
	}

	a=(float)atof(argv[1]);
	b=(float)atof(argv[2]);

	printf("argv[1]=%s, argv[2]=%s\n",argv[1],argv[2]);

	res0=xa(a,b);
	res1=xb(a,b);

	printf("a=%f, b=%f, xa()=%f, xb()=%f\n",a,b,res0,res1);

	return 0;
}

#define LOOPS INT_MAX

void printResults(struct rusage *after, struct rusage *before, int loops,
	int clock)
{
double loopsf,utime;
long sec,usec;


	sec=after->ru_utime.tv_sec-before->ru_utime.tv_sec;
	usec=after->ru_utime.tv_usec-before->ru_utime.tv_usec;

	if(usec<0)
	{
		sec--;
		usec+=1000000;
	}

	printf("\t%d.%06du, %d.%06ds\n",sec,usec,
		after->ru_stime.tv_sec,after->ru_stime.tv_usec);

	loopsf=(double) loops;
	utime=sec+usec/1000000.0;

	printf("\t%f loops/sec\n",loopsf/utime);

	if(clock>0) printf("\t%f cycles/loop\n",clock/loopsf*utime);
}

int main(int argc, char *argv[])
{
struct rusage ru_b,ru_a;
int clock=0,argc0;
volatile int m=1,n,o;
volatile unsigned int l;
volatile float c,d;
float a,b;
double diff;
char *argv0[4];

	a=1.0;
	b=2.2;
	argv0[0]=argv[0];
	argv0[1]="1.0";
	argv0[2]="2.2";
	argv0[3]=NULL;
	argc0=3;

	if(argc>1)
		clock=atoi(argv[1]);

	if(argc>3) {
		a=(float)atof(argv[2]);
		b=(float)atof(argv[3]);
		argv0[1]=argv[2];
		argv0[2]=argv[3];
	}

	getrusage(RUSAGE_SELF,&ru_b);

	for(l=LOOPS;l;l--) ;

	getrusage(RUSAGE_SELF,&ru_a);

	printf("Empty loop:\n");
	printResults(&ru_a,&ru_b,LOOPS,clock);

	getrusage(RUSAGE_SELF,&ru_b);

	for(l=LOOPS;l;l--) c=xa(a,b);

	getrusage(RUSAGE_SELF,&ru_a);

	printf("xa()=%g (straight c):\n",c);
	printResults(&ru_a,&ru_b,LOOPS,clock);

	getrusage(RUSAGE_SELF,&ru_b);

	for(l=LOOPS;l;l--) d=xb(a,b);

	getrusage(RUSAGE_SELF,&ru_a);

	printf("xb()=%g (embedded asm):\n",d);
	printResults(&ru_a,&ru_b,LOOPS,clock);

	getrusage(RUSAGE_SELF,&ru_b);

	n=0x100000;
	o=0xff;

	for(l=LOOPS;l;l--) m=n*o;

	getrusage(RUSAGE_SELF,&ru_a);

	printf("m = 0x100000 x 0xff:\n");
	printResults(&ru_a,&ru_b,LOOPS,clock);

	getrusage(RUSAGE_SELF,&ru_b);

	for(l=LOOPS;l;l--) m=o*0;

	getrusage(RUSAGE_SELF,&ru_a);

	printf("m = 0xff x 0x100000:\n");
	printResults(&ru_a,&ru_b,LOOPS,clock);

	(void) main1(argc0,argv0);

	diff=m_pi_minus_fldpi();
	printf("m_pi_minus_fldpi()=%g\n",diff);

	return 0;
}