Re: Patch: hoist more loop invariants

[Andreas Jaeger <aj at suse dot de>]

Andreas Jaeger <aj at suse dot de> writes:

> Richard Henderson <rth at redhat dot com> writes:
>
>> On Fri, Feb 28, 2003 at 08:48:47AM +0100, Andreas Jaeger wrote:
>>> So, it's just one point which is not really meaningfull.  Or should I
>>> have used different compiler options for this?
>>
>> I wouldn't expect this to show up much on x86.  Indeed, it is
>> comforting to know that we don't regress too much due to 
>> over-committing registers.
>>
>> Can you test on another system?  ia64, alpha, or x86-64?
>
> I could test on x86-64 but cannot post complete results ;-(.  But I'll
> do a run there and report some "tentative numbers",

Ok, here're the numbers on same random x86-64 hardware in percentages.

Significant changes (improvements in SPEC numbers, 100% is GCC without
the patch):

SPECint
-O2
vpr: -1%
crafty: -2%
eon: -1 %
gap: -3%
vortex: +2%
bzip: +1%
twolf: -1%
Whole SPECint: -0,4%


-O3 -funroll-all-loops -fprefetch-loop-arrays:
vpr: +0,4%
crafty: -1%
eon: +6%
gap: -2%
vortex: +5%
twolf: +2%
Whole SPECint: +1 %

SPECfp:
-O2:
wupwise: -2%
mgrid: +1%
mesa: -1%
art: +2%
sixtrack: +1%
Change for all C and F77 programs in SPECfp: -0,5%

-O3 -funroll-all-loops -fprefetch-loop-arrays:
wupwise: +0,4%
mgrid: -3%
mesa: -3%
sixtrack: -1%
Change for all C and F77 programs in SPECfp: -0,7%

Andreas
-- 
 Andreas Jaeger
  SuSE Labs aj at suse dot de
   private aj at arthur dot inka dot de
    http://www.suse.de/~aj