Re: [Mingw-users] GCC 4.3.0 20080502 (alpha) Released - Please help us test it!

[Stéphane Larouche <stephane.larouche@xxxxxxxxxx>]

Aaron W. LaFramboise <aaron77thyme@...> writes:
> This is a problem in general on i386 due to the way that relatively 
> simple operations (such as assignment) can cause the fp size to change. 
>   This is a very difficult problem to solve, and is present in most 
> compilers based on i386 chips.  The best workaround is careful coding to 
> avoid inadvertent sensitivity or instability.

This is no really a problem for me, for all practical purposes the result is the
same, the difference being very small. But it still puzzles me who thought
computers provided a single answer when given the same set of commands.

> > I will try to find some time during the weekend to generate a
> > minimal example that reproduces the problem.
> 
> Please let me know how this goes for you.

The difference seems to be related, in part, to the use of complex numbers.
Here's some (quite stupid) code:

---------- Beginning of test.cpp ----------

#include <cstdio>
#include <cmath>
#include <ctime>
#include <complex>

using namespace std;

const complex<double> j = complex<double>(0.0, 1.0);

int main ()
{
	long i, nb = 10000000;
	complex<double> x, answer = 1.0;
	clock_t start, end;

	start = clock();

	for(i = 0; i < nb; i++)
	{
		x = double(i)/double(nb) * M_PI - j * double(i)/double(nb) * M_PI;
		answer *= sin(x)*sin(x) + cos(x)*cos(x);
	}

	end = clock();

	printf("Answer: %.16f + %.16fj\n", real(answer), imag(answer));
	printf("number of clock ticks: %i\n", end-start);
}
 
---------- End of test.cpp ----------

I compile it using:
g++ -c test.cpp
g++ -o test.exe test.o

The number of clock ticks is:
- GCC 3.4.5: 77481
- GCC 4.2.1-dw2: 85532
- GCC 4.3.0 alpha: 86995

If I add the -O3 flag, the number of clock ticks is:
- GCC 3.4.5: 75168
- GCC 4.2.1-dw2: 36792
- GCC 4.3.0 alpha: 42741

If I replace the complex<double> by double (and adjust the code consequently),
the number of clock ticks is,  for all versions, approximately 17200 when using
no flags, and approximately 8600 when using the -O3 flag.

I'll continue investigating since this explains why 4.3.0 alpha is slower than
4.2.1, but not why it is slower than 3.4.5 when using the -O3 flag.

I hope this helps,

Stéphane


-------------------------------------------------------------------------
This SF.net email is sponsored by the 2008 JavaOne(SM) Conference 
Don't miss this year's exciting event. There's still time to save $100. 
Use priority code J8TL2D2. 
http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone
_______________________________________________
MinGW-users mailing list
MinGW-users@xxxxxxxxxxxxxxxxxxxxx

You may change your MinGW Account Options or unsubscribe at:
https://lists.sourceforge.net/lists/listinfo/mingw-users