Re: New architecture ppc64el on Debian
On Sun, Sep 22, 2013 at 06:08:12PM +0900, Joel Rees wrote:
> I suppose I'm not really one privileged to express these reactions,
> but I tend to be a loose cannon sometimes.
> (Checking my calendar. Nope, I haven't suddenly slipped back to the
> early 1990s.)
> (Still puzzled.)
Well I see patches for linux on powerpc in little endian mode (it
is software selectable on most powerpc chips as far as I know) were
submitted in August of this year, and patches to qemu were done around
the same time to support emulating little endian mode powerpc.
> So, pardon me for being dense, but is the build process really going
> to be all that much more difficult than putting a compiler flag for
> least significant byte first in the make files and typing make?
Well you take a compiler that generates code for powerpc in little
endian mode, and then you build stuff. Then you go fix all the code
where people incorrectly assumed powerpc meant big endian. I imagine
there is quite a bit of that around. Also anything with powerpc assembly
code might need fixing.
> I suppose, along with lsb-first, there are some changes in the soc
> peripherals and some special tuning of the instruction set. I'm
> assuming the latter is taken care of by the compiler, which I assume
> you already have had running and debugged for the newly tuned parts
> for at least a few months?
Certainly drivers can require some fixes.
> Not that it's my problem to worry about at all.
> My impression is that the sooner you can have the new target hardware
> available to donate to the build farms the better, as someone else has
> also mentioned.
Hardware has existed for years. It just isn't the mode if is normally used in.
> And I'm wondering if this "little-endian" hardware is hardwired
> "little-endian", and why the fuss?
I don't think any powerpc was ever hardwired for little endian. Some a
big endian only.
> You've had several people hint. I've hinted, too. It seems like there
> is something that needs to be explained.
> Why the fuss?
> It's not like endian-ness is a barrier to performance. Is there some
> Intel-fomented rumor floating around data centers with BOFH admins,
> that the endian-ness is the Power architecture's Achilles' heel or
> Hmm. I suppose it might have something to do with graphics chips which
> seem to be all x86-oriented and are now being used as dedicated
> numeric array processors, per this commentary in LWN:
> <http://lwn.net/Articles/408845/>. That's going the way of the dodo in
> about three to five years, but I suppose there may be some purpose in
> reactionary marketing now.
It does appear to have to do with support for graphics cards that were
designed to assume little endian. Given x86 and arm seem to run almost
exclusively in little endian mode these days (not that x86 even has an
option not to unlike arm), perhaps IBM has decided that to get better
performance and support for running high performance software using
graphics cards, it makes sense to get little endian support going in
linux for their big powerpc boxes.
> Well, anyway, the devs are the ones to make up their minds on how much
> resource to put in on this, so I suppose I should shut up. But it sure
> seems like Intel leads a lot of useless market churn. If the
> predictions about global climate change turn out to be true, when the
> amount of excess waste heat generated by using x86 in infrastructure
> is calculated, do we sue Intel for the part it plays in wasting the
> earth through excess hardware turnover and energy use?
Well you could always complain about IBM choosing the x86 in the first
place, or complain that people bought IBM's machines in the first place
at a time when lots of other machines were using m68k chips instead.
Seems a bit late to fix it now though. So perhaps this is a sensible
move on IBM's part.