mk.conf:

OBJECT_FMT=ELF
USE_EGCS=1
USE_EGCS_HAIFA=1
I387_LIBM?=yes

: sean.ebone.net ; cd /usr/src/share/mk
: sean.ebone.net ; diff sys.mk /usr/share/mk/sys.mk
28c28
< CFLAGS?=      -O2
---
> CFLAGS?=      -O3 -march=i686 -mfancy-math-387
: sean.ebone.net ; sed 1q sys.mk
#       $NetBSD: sys.mk,v 1.54 1999/03/10 14:06:14 mycroft Exp $

I use the same cc options elsewhere where they're easy to set. :)

I also munge xsrc/xc/config/cf/xf86site.def habitually, including
#define DefaultGcc2i386Opt      -O3 -march=i686 -mfancy-math-387

Works great, I've been tracking -current with these settings for
a long time now, and have yet to see a problem unrelated to 
managment of -march= differences (I have a 486, a 586 and a 686...).

I don't build a kernel with -march=i686 since that has failed in the
distant past, perhaps as long ago as pre-egcs.  I also don't use 
-O3 on the kernel, also because of ancient suspicious behaviour.

I also munge xsrc/xc/config/cf/xf86site.def habitually, including
#define DefaultGcc2i386Opt	-O3 -march=i686 -mfancy-math-387

Unfortunately I haven't been in a position to compare the shipped
C compilation option/setting with the ones I use on an Intel box
in a long time, but in the days when egcs and then the haifa
scheduler were first introduced, the qualitative performance
improvement was really good.

	Sean.