Re: Kernel Panic installing on SGI Indy
On Wed, 25 Feb 2009, Ralf Baechle wrote:
> > > You're getting this error message for the attempt to run a 64-bit kernel
> > > on certain very old R4000 revisions. So either run a 32-bit kernel,
> > > switch to newer version of the R4000 or to another MIPS CPU. For a
> > > processor with just 8k caches switching to a 32-bit kernel is a good
> > > idea to get best performance but we should probably enable the R4000
> > > workaround automatically in Kconfig for 64-bit Indy kernels.
> >
> > isn't there a gcc needed, which supports all workarounds ? What about
> > the performance impact of that workarounds ?
>
> They definately will have some performance impact but the only person
> who may have some experience with that is probably Maciej.
I haven't benchmarked the difference as when correctness is required the
performance hit is secondary. The hit is taken only if the workarounds
are actually activated -- there is nothing hardcoded in the patches and
the "-mno-daddi" and "-mdaddi" options switch the workarounds on and off
respectively. The default is "-mno-daddi" if "-mfix-r4000" or
"-mfix-r4400" was specified; "-mdaddi" otherwise.
Technically all 64-bit immediate additions are converted to a sequence of
an immediate load to a temporary register (which is really a 32-bit
addition to $zero) followed by a 64-bit register addition. Plus $t8 is
marked as fixed in GCC because the compiler is not (or at least wasn't at
the time I prepared the workaround) prepared to allocate a scratch
register for branches (an extra spare register is needed to load the
target address of an out-of-range branch being relaxed). It takes one
temporary register away from the pool available and will certainly affect
allocation in some cases and therefore performance.
> As I recall only a part of the workarounds needed is supported by an
> unpatched gcc but again Maciej will know more.
Correct -- the changes to work around the buggy DADDIU instructions
proved too invasive for long-term maintenance upstream. Two patches are
required, for binutils and GCC respectively. Some handcoded assembly code
which uses ".set noat" may not build anymore and require modifications.
I have made source RPM packages of patched tools available at:
ftp://ftp.linux-mips.org/pub/linux/mips/people/macro/SRPMS/ -- the newest
version of GCC is gcc-4.1.2-8.src.rpm and for binutils this is
binutils-2.18-4.src.rpm -- look for patches named *-mips-nodaddi-*.
Binary packages are available there in a neigbour directory too, in case
you wanted to try something quickly. ;)
Because of various more and less recent circumstances I haven't ported
the patches to newer versions of the respective tools (sorry), but I
intend to do so at some point in the future.
Let me know if I can be of assistance.
Maciej
Reply to: