in Toronto this month: International Symposium on Code Generation and Optimization

Lennart Sorensen lsorense-1wCw9BSqJbv44Nm34jS7GywD8/FfD2ys at public.gmane.org
Wed Apr 7 14:05:19 UTC 2010


On Tue, Apr 06, 2010 at 07:24:42PM -0400, D. Hugh Redelmeier wrote:
> Unix was an important enabling technology for the development of RISC
> microprocessors.  It meant that the people creating the micros didn't
> have to create an OS and attract users to a new OS.
> 
> Linux continues this tradition -- UNIX gave up almost 20 years ago.
> 
> Linux's second architecture was the Alpha if I remember correctly.

Yes it was.  We can thank DEC for doing a lot of the cleanup to make
linux portable.

> There were a few other OSes on new architectures:
> 
> - OpenVMS on Alpha
> 
> - Apollo's Domain/OS on their PRISM

I never was Apollo on anything other than m68k.  Hmm.

> - MS WinNT on MIPS

Well NT ran on MIPS, powerpc, x86 and alpha.

> The success rate seems low.  But then few of the Unix-based ones are
> still around either.
> 
> The worst things about x86 have been overcome:
> 
> - the 32-bit instruction set is OK; the 64-bit one even better

The 64bit is better.  The 32bit is not OK.  It's awful.

> - the overhead for the mediocre instruction set is handled by throwing
>   lots of cheap transistors at it.  At one point those transistors
>   could have been used for much better purposes.  Now its just a tiny
>   percent of the chip.

Yet every modern x86 doesn't execute x86 natively.  They all translate.
Perhaps that's a good indication that it would be better to write code
for the actual CPU rather than a legacy instruction set the CPU doesn't
even like.

> - x86 has decent code density which is actually useful.

RAM is cheap.  It really isn't a bit deal anymore.  ARM and powerpc are
not that bad.

> We do really need to have some new ideas to break through the current
> set of barriers.  Some way to use parallelism more effectively.  All
> ways seem to involve pain.  Several are specialized and we reject them
> because we're spoiled by how useful x86 is for so many applications.

Certainly writing parallel code is the big issue.  The instruction set
has nothing to do with that for the most part.

> Example: the IBM/Sony Cell is a novel design that is really good for
> some applications.  Most of us haven't bit so it will probably die.

IBM seems to have killed it.  Future powerpc designs will apparently
have SPE style cores in them though.

> Example: Sun's Niagara looked good.  Few took it up.  It will probably
> die.

Well it seems Sun too forever to finish it (lots of delays).  Eventually
it becomes irrelevant.

> Example: ARM or MIPS would make a useful notebook/netbook processor.
> The mass market seems to require MS Windows.  So the ARM notebook
> won't be important.  But the iPad is a disguised ARM notebook.

Well the chinese mips processor coming out this year is supposed to
be able to run x86 code translated by qemu at about 80% of native
performance.  Perhaps that will do.

I don't consider the ipad a computer.  A computer runs what I want it
to run.  The ipad does not, so it is not a computer.  It's an appliance.

-- 
Len Sorensen
--
The Toronto Linux Users Group.      Meetings: http://gtalug.org/
TLUG requests: Linux topics, No HTML, wrap text below 80 columns
How to UNSUBSCRIBE: http://gtalug.org/wiki/Mailing_lists





More information about the Legacy mailing list