This is the mail archive of the
mailing list for the GCC project.
Re: program executin times differ between gcc versions
- From: Andrew Haley <aph-gcc at littlepinkcloud dot COM>
- To: "Paul Moore" <slesuser at gmail dot com>
- Cc: gcc-help at gcc dot gnu dot org
- Date: Thu, 1 Nov 2007 13:28:36 +0000
- Subject: Re: program executin times differ between gcc versions
- References: <firstname.lastname@example.org>
Paul Moore writes:
> Dear List,
> If this isn't the right place to ask, please advise.
> I've come across different program execution times based on the gcc
> version used to compile C code. I understand that this is expected
> between compiler generations, but I expected to see better results
> with newer compiler generations. In fact, the results were worse.
> The program/shared libraries involved are
> - coded in C
> - compiled as version 1: gcc 3.3.3 on SLES 9
> as version 2: gcc 4.1.2 on SLES 10 SP1
> - gcc flags used for both compilation were identical:
> gcc -c -DLINUX -D_GNU_SOURCE -fPIC -funsigned-char -m32 -g
> The application is single threaded, CPU bound. The utilization of one
> CPU core is in both cases near or on 100%.
> The averaged test run times were
> - 4m07s for the program compiled with gcc 3.3.3
> - 5m24s for the program compiled with gcc 4.1.2
> The test runs were performed on the same system, a 8 CPU AMD Opteron
> with 2.6 GHz each, SLES 10 SP1. The program does pattern matching and
> format generation, nothing fancy. No fp calculations or similar math
> stuff is involved.
> Using optimization flags is not possible due to the fact that large
> parts of the application failed in -O test runs and would need to be
> fixed/tested extensively first. Profiling with -pg/gprof has not been
> produced usable results as shared libraries are not being tracked.
> Is there any explanation for this or am I missing something?
gcc 4.x generates code in a very different way from 3.x. It's quite
likely that at -O0 there will me more temporaries and more
Using optimization is the only way to increase performance.