[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: AMD response about Debian x86-64 port

* Michael Banck <mbanck@gmx.net> [030110 06:26]:
> Hi,
> I've talked to AMDs Open Source Relation Engineer, Marc Miller about a
> possible Debian x86-64 port, see his (quite positive, IMHO) answer
> below. AFAIK, single developers accepted NDAs for porting purposes in
> the past, am I mistaken? I'd hope such an NDA would be time limited,
> e.g. until the Opteron is officially available. Another issue would be
> hosting a possible hardware loan, I've heard on irc that Brainfood might
> be able to do it in a way acceptable to AMD.

Great work.  Thanks, Michael.

> I am definetly lacking the knowledge, experience and time to do a port,
> but I guess/hope there are couple of us who would be interested in
> working on a x86-64 port, even under an NDA? Marco d'Itri already said
> he would generally be interested. It might be wise to form a porting
> team and strategy before everybody bombs Marc with individual requests.

I am definately interested in joining this team!

> I'd suspect that AMD would be more willing to share/give away hardware
> if they knew that we were strongly committed to a port. Anyway, we could
> use the time being without access to hardware to discuss matters about
> the port, e.g. if a mixed x86/x86-64 userspace is feasable or even
> possible or if we could cross-compile some important toolchain stuff
> in the meantime.

I will assume that if we do any work we must start with a 64-bit kernel.

From what I understand, the hardware provides a "64-bit long mode" in
which the kernel could run and allow for 32-bit applications to execute
without recompilation (this later mode is referred to as compatibility
submode of long mode).  However, recalling Andi Kleen's talk at OLS two
years ago, I believe that a recompile will still be needed for these
32-bit apps due to an ABI change on the x86-64 platform.

I am pretty sure all the changes in user-space would be localized to
glibc -- as it implements the ABI's API to the rest of user application.  
That is not too bad.

However, since we are already recompiling applications on the x86-64 to
run in 32-bit more, we should recompile them in 64-bit mode to improve
performance (wider registers, 8 additional general purpose registers,
etc).  There may be some apps that will (initially) not work in 64-bit
mode, and for these we use 32-bit mode.  The only snag with this is that
we now need two glibc's on the system: the 32-bit one and the 64-bit


				WebSig: http://www.jukie.net/~bart/sig/

Attachment: pgpJWkEZQtdFU.pgp
Description: PGP signature

Reply to: