Re: linux 2.1.117 built with egcs-1.1 19980823

[Andi Kleen <ak at muc dot de>]

On Tue, Aug 25, 1998 at 06:15:50AM +0200, David S. Miller wrote:
>    Date: Mon, 24 Aug 1998 22:07:49 -0600
>    From: Jeffrey A Law <law@hurl.cygnus.com>
> 
>    It would probably be to our advantage if you could select one of
>    the datapoints from the bloat-o-meter chart and analyze it.  Even
>    superficial like "more spills than before" would be worth knowing.
> 
> Finding this wouldn't surprise me as in other areas the Haifa
> scheduler has been shown to extend register lifetimes unduly.

My tests don't use the Haifa scheduler (sorry, I should have said that)
The i386 port still defaults to the old scheduler, and I didn't enable
it explicitely because people told me earlier that the i386 port is not
tuned for Haifa yet.

> 
> (Andi, as a fun experiment, rebuild the kernel with
>  "-fno-schedule-insns -fschedule-insns2", this turns off
>  the first scheduling pass, yet keeps the second pass enabled which
>  runs after register allocation and therefore can't increase
>  register pressure and create more spills)

I tried this, it seems to cut off another 4K:

Here is the summary so far:

 text    data     bss     dec     hex filename
933404   95000  132368 1160772  11b644 vmlinux-2.7.2
2.7.2.3 options: -malign-{functions,loops,jumps}=2 -fno-strength-reduce -O2
		 -fomit-frame-pointer -m486
955025   99676  133412 1188113  122111 vmlinux
egcs-1.1-19980816 / no haifa options:
-O2 -fomit-frame-pointer -fno-exceptions -fno-schedule-insns -fschedule-insns2 -mpentiumpro -malign-{loops,functions,jumps}=2 -fno-gcse

I'll look at some more function diffs later today to see if I can spot 
something obvious.

-Andi