From: Andrew Haley <aph@xxxxxxxxxxx>
To: "cino hilliard" <hillcino368@xxxxxxxxxxx>
CC: gcc-help@xxxxxxxxxxx
Subject: Re: Free Lunch
Date: Tue, 18 Apr 2006 11:14:37 +0100
There's a good reason for this. You can discover it by reading gcc's
documentation. See ``Deleting "empty" loops''.
Ok here it is.
Historically, GCC has not deleted ?empty? loops under the assumption that
the most
likely reason you would put one in a program is to have a delay, so deleting
them will
not make real programs run any faster.
However, the rationale here is that optimization of a nonempty loop cannot
produce
an empty one. This held for carefully written C compiled with less powerful
optimizers
but is not always the case for carefully written C++ or with more powerful
optimizers.
Thus GCC will remove operations from loops whenever it can determine those
operations
are not externally visible (apart from the time taken to execute them, of
course).
As GCC improves, it will remove the loop itself. Indeed, with
?-funroll-loops? small
loops can already be removed, so leaving an empty non-unrolled loop is both
suboptimal
and inconsistent.
Be aware of this when performing timing tests, for instance the following
loop can be
completely removed, provided some_expression can provably not change any
global
state.
{
int sum = 0;
int ix;
for (ix = 0; ix != 10000; ix++)
sum += some_expression;
}
Now based on this, I modifed the code shown below to produce the following
timing.
2000000000
Sec = 1.640625
2000000001
Sec = 1.000000
So count1 loop which has the assignment statement z=x takes 1.64 sec while
count 2 which
which has 3 assignment statements
z=j;
y+=1;
s=y+1;
Yet count 2 executes in .64 sec secs less time. That is the free lunch.
The best I can do in just an empty loop in assembly language A386 is 2.47.
Some of the time
is probably the timer and output routines.
On my p3 1 ghz system I get
count1 4.171875 sec
count2 4.187500 sec1
a small price for 2 billion loops of
y+=1;
s=y+1;
The gcc assembly code appears to produce 2 bill loops for both functions
Also if we change the type to unsigned long and use 4000000000UL, we get
4000000000
Sec = 3.234375
4000000001
Sec = 0.000000
If we relax the -O3 optimization switch, the free lunch goes BY By as we
get,
4000000000
Sec = 8.125000
4000000001
Sec = 17.531250
Similarly for type long we get
2000000000
Sec = 4.031250
2000000001
Sec = 8.828125
And the free lunch goes BY By.
To Optimize or not to Optimize, that is the question.
Dog gone it TANSTAAFL!
Have fun,
Cino
***********************************code********************************
// Free Lunch
//> Executing: c:\context\ConExec.exe "gc2.bat" COUNTgcc
//C:\gcc\examples>f:\gcc\bin\g++ COUNTgcc.c -o COUNTgcc.exe -s -O3
-mtune=pentium4
//C:\gcc\examples>rem g++ -c -g -Wa,-a,-ad COUNTgcc.c > COUNTgcc.lst
//C:\gcc\examples>g++ -S -c COUNTgcc.c
//C:\gcc\examples>g++ -dumpversion
//3.4.4
//> Execution finished.
//System: p4 2.53 ghz xp pro
#include <windows.h>
#include <stdio.h>
#define timer GetTickCount()/1000.0
float t1,t2;
long count1(long);
long count2(long);
int main()
{
long j1,j2;
t1=timer;
j1 = count1(2000000000);
t2=timer;
printf("%u\n",j1);
printf("%s%f\n"," Sec = ",t2-t1);
printf("\n");
t1=timer;
j2 = count2(2000000000);
t2=timer;
printf("%u\n",j2);
printf("%s%f\n"," Sec = ",t2-t1);
getchar();
return 0;
}
long count1(long y)
{
long x,z;
for(x=1;x<=y;x++)
{
z=x;
}
return z;
}
long count2(long i)
{
long j,y=0,s=0,z;
for(j=1;j<=i;j++)
{
z=j;
y+=1;
s=y+1;
}
return s;
}
