> > The most likely answer is that the NetBSD kernel source code accesses memory 
> > not entirely in a type-safe fashion, and higher optimization levels will 
> > expose these problems because the compiler's assumptions about which
> > variables can be changed via pointer, structure, or union member
> > dereferencing are not correct.
> 
> I don't think so. Kernels compile with "-march=p4" and "-O2" work fine
> for me on two differnt systems. My guess is a code generation bug in GCC,
> IIRC at least "-march=pentium" is broken.

Seconded.  I have booting problems on Transmeta TM-5800 based laptop.
-fno-strict-aliasing or -fno-rename-registers do not help.
Working kernel results only when -march=pentium removed.

Offtopic question: is there any way to run TM-5800 in native mode?