18 Sep
2015
18 Sep
'15
5:46 a.m.
-----Original Message-----
From: cctalk [mailto:cctalk-bounces at classiccmp.org] On Behalf Of Liam
Proven
Sent: 18 September 2015 13:23
To: General Discussion: On-Topic and Off-Topic Posts
<cctalk at classiccmp.org>
Subject: Re: OT: x86 machine code [Was: Re: Self modifying code, lambda
calculus - Re: ENIAC programming]
On 18 September 2015 at 13:06, Pontus Pihlgren <pontus at update.uu.se>
wrote:
Don't know about the Intel chips but the IBM Mainframe chips have an intermediate
level of code they call "millicode" as described for example here:-
http://ibmsystemsmag.com/mainframe/administrator/performance/millicode_roge…
and basically you don't actually want to by-pass this code because by doing so you
by-pass the optimisations and so run more slowly....
I've been told this more than a few times and
read it in various places.
It always make me wonder, could we not allow a mode in modern Intel
processors that lets us bypass the x86 code emulation/translation and
run "directly on the metal" (if there were such a thing).
The purpose, of course, would be to gain performance. Certainly this
would already have been done if there was any significant gain to be
had?
I think not, because the "RISC core interpreting
x86 instructions" is a fairly
gross over-simplification, as best as I have been able to determine. Yes, all
C21 x86 chips borrow lots of design principles from RISC, but they are CISC
chips executing CISC code -- just doing a lot of fancy on-the-fly optimisations.
There isn't an underlying separate different microcode.
That has been done, though. It was the Transmeta Crusoe line of processors.
These were not directly x86-compatible at all: they had their own instruction
set, and during boot, loaded a translation layer on top which executed x86
code via a sort of optimising compiler/interpreter with JIT.
The purpose was to achieve very low power consumption, for portables.
The performance was not as good as native x86 even at the time -- although
nearly -- but the processors used significantly less power.
They were not RISC, though: they were VLIW underneath.
I always thought it was *the* critical mistake of Transmeta not to at least
release the native instruction set. If they could also execute Motorola 680x0
code, or PowerPC code, or Alpha code, or any other discontinued (or
effectively discontinued) instruction set, they would still have a market
today.
--
Liam Proven ? Profile: http://lproven.livejournal.com/profile
Email: lproven at cix.co.uk ? GMail/G+/Twitter/Flickr/Facebook: lproven
MSN: lproven at hotmail.com ? Skype/AIM/Yahoo/LinkedIn: liamproven
Cell/Mobiles: +44 7939-087884 (UK) ? +420 702 829 053 (?R)