GCC trunk SPEC2000 performance

Janis Johnson janis187@us.ibm.com
Thu Jun 20 12:44:00 GMT 2002


On Thu, Jun 20, 2002 at 11:16:50AM -0700, Richard Henderson wrote:
> On Wed, Jun 19, 2002 at 10:43:40PM -0400, David Edelsohn wrote:
> > 2) "peak" (-O3 -funroll-all-loops -fprefetch=loop-arrays) produces slower
> > code than base (-O2) for many tests.
> 
> Peak cannot be considered "peak" with -funroll-all-loops
> as opposed to -funroll-loops.  That forces the compiler 
> to unroll a loop against its better judgement.
> 
> It's good for correctness testing, however.

-fprefetch-loop-arrays needs work to actually improve most programs.
Honza's description of it for 3.1 is that it's experimental, which is
still quite accurate.

Janis



More information about the Gcc mailing list