On Thursday 07 March 2002 09:31 pm, John M McIntosh wrote:

> -march=pentiumpro ?

That didn't help either (with -march=pentiumpro -mcpu=pentiumpro).

But I kept fiddling and got about 16-18% better speed on 0 tinyBenchmarks 
than the default 2.95.3 build.

Turns out that -funroll-loops was slowing down at least the tinyBenchmarks 
(though I don't know whether it helped (for instance) BitBlt).

With 3.0.3 and
-O3 -DLSB_FIRST=1 -fno-gcse -fomit-frame-pointer -fschedule-insns
'49230769 bytecodes/sec; 1510053 sends/sec'
'49382716 bytecodes/sec; 1510053 sends/sec'
'48967100 bytecodes/sec; 1518320 sends/sec'
'49344641 bytecodes/sec; 1514175 sends/sec'

Contrast with 2.95.3 and
-O2 -funroll-loops
 '42553191 bytecodes/sec; 1283021 sends/sec'
'42356055 bytecodes/sec; 1282033 sends/sec'

> >I think my warranty was already void because I fiddled with the knobs on
> > the back of the set.
>
> Oh sure if you're using gcc 3.0 then it's more like you've taken a
> screwdriver and removed those stickers that say warranty void if
> removed... And poked at the guts, beware about that voltage eh?

Hey, gcc 3.0 has been out since June of last year <g>... and 3.0.4 has been 
out for a couple of weeks now...

-- 
Ned Konz
currently: Stanwood, WA
email:     ned at bike-nomad.com
homepage:  http://bike-nomad.com