Re: [LUG] GCC debug info and some other garbage.

[Andrew Rogers <andrew@xxxxxxxxxxxxxxxx>]

Simon Waters wrote:

> -----BEGIN PGP SIGNED MESSAGE-----
> Hash: SHA1
>
> Andrew Rogers wrote:
>  
>
> > A friend has compiled his program on both GCC 2.96 and MS
> >    
> Visual C++ (I
>  
>
> > don't know which release). He claims that Visual C++ gives a
> >    
> two fold
>  
>
> > speed increase to his program. As I did not want to believe
> >    
> this I set
>  
>
> > about investigating the cause of this apparent speed increase.
> >    
>
> As a maintainer of a chess program we have a keen interest in
> performance.
>
> For C code both Intel and Microsoft compilers outperform GCC on
> Intel hardware for most benchmarks, although not usually by 50%.
>
> Most real world applications are closer, and in some cases high
> level optimisations in GCC will result in code that performs
> commercial compilers.
>
> GCC's real strength is it works pretty well everywhere, not that
> it produces the fastest code, which in compilers usually goes to
> those who know the hardware best - (read Intels compiler
> produces fast code).
>
> Now what is the program doing would be my first question,
> especially what system routines, as unless it is producing the
> same native system calls it may not be a direct comparison.
The program is used to test a turbo code, this is part of our satellite 
communications research at the University of Plymouth. Essentially it 
makes billions of data sequences and calculates the decoding performance 
for each. There are very few system calls compared to the calculation 
that goes on. Our programs sometimes need a month to run, that's why 
performance and compiler optimisation is critical.

>  
>
> > I can't seem to find a way to remove this debugging info and
> > there are no -g switches in my GCC command line.
> >    
>
> man strip ?
strip removes some of the stuff, but unfortunately this doesn't lead to 
an increase in performance.

> What does this information look like, what does "file
> <executable>" return.
>
>  
>
> > Slighty seperate issue, I compiled a program with -O3
> >    
> (optimisation) on
>  
>
> > both GCC 2.96 and GCC 3.2 but they both execute at exactly the
> >    
> same
>  
>
> > speed. I thought that GCC 3.2 had significantly improved
> >    
> optimisation.
>
> GCC 3 executes GNU Chess with about 20% improvement in nodes per
> second on same hardware (Cyrix 166MHz).
>
> You have to have code that the new optimisations work on.....
>
>  
>
> > Is there any new optimisation command line switches in GCC 3.2
> >    
> that I
>  
>
> > may have missed?
> >    
>
> I'd suggest profiling and looking at what is actually consuming
> CPU, and compare between compilers if possible.
I will my friend to do this. I don't have MS Visual C++ so I can't do 
these tests.

While I was looking at the executable with khexedit I noticed a section 
of text that looks like this:

GCC: (GNU) 2.96 20000731 (Mandrake Linux 8.2 2.96-0.76mdk)GCC: (GNU) 
2.96 20000731 (Mandrake Linux 8.2 2.96-0.76mdk)GCC: (GNU) 2.96 20000731 
(Mandrake Linux 8.2 2.96-0.76mdk)GCC: (GNU) 2.96 20000731 (Mandrake 
Linux 8.2 2.96-0.76mdk)GCC: (GNU) 2.96 20000731 (Mandrake Linux 8.2 
2.96-0.76mdk)GCC: (GNU) 2.96 20000731 (Mandrake Linux 8.2 
2.96-0.76mdk)GCC: (GNU) 2.96 20000731 (Mandrake Linux 8.2 
2.96-0.76mdk)GCC: (GNU) 2.96 20000731 (Mandrake Linux 8.2 
2.96-0.76mdk)GCC: (GNU) 2.96 20000731 (Mandrake Linux 8.2 
2.96-0.76mdk)GCC: (GNU) 2.96 20000731 (Mandrake Linux 8.2 
2.96-0.76mdk)GCC: (GNU) 2.96 20000731 (Mandrake Linux 8.2 
2.96-0.76mdk)GCC: (GNU) 2.96 20000731 (Mandrake Linux 8.2 2.96-0.76mdk)

As you can see its repeated about ten times. strip does not remove this 
so I removed it with khexedit. This did not cause the program to crash 
although it may be wrong to remove this stuff.

Regards
Andrew Rogers


--
The Mailing List for the Devon & Cornwall LUG
Mail majordomo@xxxxxxxxxxxx with "unsubscribe list" in the
message body to unsubscribe.