Re: CMUCL on Opterons??

Martin Cracauer <cracauer@xxxxxxxx> writes:

> This doesn't work.  The modes are not compatible enough to run a
> literal piece of 32 bit binary code in a 64 bit process.

What he says.  In my (ongoing) SBCL port to the AMD64 I've run into
two places so far where the x86 code generated by Python does
something utterly different on x86-64 -

1) x86-64 uses an extra ("REX") instruction prefix to select 64 bit
   modes and the extra registers.  These are bytes #x40-#x4f, which in
   the real x86 are forms of the INC and DEC instructions.

2) x86-64 has a RIP-relative addressing mode which is selected by a
   particular suffix (ModRM) byte hitherto used for (I think, it's
   been two weeks since I hacked on this) absolute indexing from zero.
   RIP (PC)-relative addressing actually ought to be useful for Python
   if it means we can lose a lot of the fixups in code components:
   fewer dirty pages and less frobbing when GC runs

It's also worth noting that the ABI for calling into C is
significantly different - it passes arguments in registers, almost
like a real computer - so if you want a Lisp that can do any kind of
IO, that bit needs sorting out too.

My opinion so far is that the x86-64 really is "a bag on the side of
the [x86]"[*], but, well, the x86 is already a pretty disgusting
accretion of hacks made tolerable only by its commodity price, so who
cares?  There's an Opteron board out now that can take 16Gb of memory
and costs around $450: it's hard to get sniffy about elegance with
toys like that to play with.


-dan


[*] For anyone who hasn't read "Soul of a New Machine", the reference
    is to the Data General Eagle: ``Ed DeCastro told the design
    engineers, "I don't want a bag on the side of the Eclipse, this is
    a 32 bit system with 16 bit backward compatibility, no mode
    bit."''            -- http://webpages.charter.net/dperr/bio.htm

-- 
"please make sure that the person is your friend before you confirm"

pgpvt10RWOKa8.pgp
Description: PGP signature