Re: debian ppc64
Sven Luther wrote:
On Mon, Dec 20, 2004 at 01:59:44PM -0800, Joaquin Menchaca wrote:
Benjamin Herrenschmidt wrote:
Oh. There's no ppc64? Aaaah. I assumed it was 64-bit. I am interested
in getting ppc64 distribution. Though, then if there's no real
performance gain, then I guess this is a moot point.
On Sun, 2004-12-19 at 13:50 +0100, Andreas Jochens wrote:
It is of course true that some other distributions put 32 bit libraries
in /usr/lib and 64 bit libraries in /usr/lib64. I will not follow that
path because for a native 64 bit distribution this will lead to an
almost empty /usr/lib directory while all the native libraries are
in /usr/lib64. In other words this would mean to change the name of
the well established '/usr/lib' hierarchy to '/usr/lib64'.
This would be ugly, at least IMHO.
I'm still rather unsure of the interest of having the ppc64 distribution
be native... Unlike amd64, as I wrote several times, there is no benefit
as in more registers etc... All you'll get is bigger/slower code...
I vaguely recall long ago in another proccessor architecture, Alpha,
there was a handbook about building optimized code for that platform to
get significant performance gains, like how for loops are created and
64-bit aligned data, etc. I don't know if any of that would be
applicable for the G5 architecture. Still, I can't for the life of me
imagine 32bit code running like/similar speeds of 64bit code, and gain
no perfomance. Hmpf! :-(
Well, as i recall, the powerpc instruction set was designed with 64bit in
mind, and the original spec had a 64bit extension. It took a long time to go
from the ill fated ppc620 which never happened to the power3 and finally the
G5 and later power4/5 chips, but still it was designed for 64bit back then.
Actually, as i recall, the 64bit code should be slower, since all pointers are
now 64bit, and thus you have to transfer double amount of code from the ram
and so on. The only reason amd64 benefit from the 64bitness is the intrinsic
deficiency of the x86 instruction set, and its lame number of usable registers
(8 of which at least 2 are not usable as i understand), which are doubled for
amd64 to 16 (compared to 32 for both ppc32 and ppc64). I doubt that you had
any 64bit optimization stuff for alpha, since alpha where always 64bit from
the start, and seem to be a dead CPU now that intel and compaq have killed it.
That clears up a lot. There actually was a lot of optimization needed,
but this was because of how lammo compilers generated the dysfunctional
assembler. Most still think in 32-bit mindset and ignore generic
On the Intel/Compaq, it's a shame. Sigh! Compaq had a good thing
going, but deplorable mismanaged their business, forcing Tandem to use
Alphas instead of MIPS, while killing the Alphas at the same time for
instance. Surprisingly, they're still kicking, don't know how. I even
had job inquiries for Digital (Tru64) UNIX Administrator.
I still cross my fingers for a revival of that architecture, now that
the whole IA64 bombed. HP might have some sense to dump PA-RISC and the
ever buggy HP-UX. (I'll duck out know, in case anyone is in love with
PS - When Compaq purchased Tandem, you know what the joke names for the
merged companies were?!? :-)
"Tampaq" and "Condem". Though, I don't think the
executives took the employees advice. :-)