Path: utzoo!attcan!uunet!tut.cis.ohio-state.edu!purdue!mentor.cc.purdue.edu!l.cc.purdue.edu!cik
From: cik@l.cc.purdue.edu (Herman Rubin)
Newsgroups: comp.arch
Subject: Re: Compiler Costs
Summary: Nondeterminism is unavoidable, and in many cases should not
	 be avoided even if possible.
Message-ID: <2323@l.cc.purdue.edu>
Date: 8 Jul 90 12:48:05 GMT
References: <1797@apctrc.UUCP> <3300143@m.cs.uiuc.edu>
Organization: Purdue University Statistics Department
Lines: 37

In article <3300143@m.cs.uiuc.edu>, gillies@m.cs.uiuc.edu writes:
> 
> It can be hard to find a speedup on CISC machines, since cacheing,
> alignment, scoreboarding, and even paging effects may be hard to
> predict.  For instance, the Motorola 68020 Manual admits that you can
> only hand-wave guesses at the execution time of an instruction stream
> -- this forces the dedicated tuner to run hoardes of benchmarks.
> 
> This is my greatest complaint about CISC architecture -- from an
> optimization standpoint, I sneer religiously at nondeterministic
> low-level hardware hacks in high-performance computer design (I don't
> even like caches, scoreboards, or paging -- they're just fads).

So the fact that a multiplication takes a variable length of time depending
on the numbers is a flaw in the machine?  I do not think so.  There are 
even cleaer examples.  I have algorithms for which a tool operation is
finding the spacing between the 1's in a random bit stream.  Of course
this is non-deterministic.  Also, it is even necessary to have planned
exceptions, which in some cases are rather likely.

There are certainly situations where a non-optimized CISC machine will
beat an optimized RISC machine with the same basic components.  I have
argued for some operations on machines which are deterministic, even,
but which will have hardware switches of essentially no cost, and would
have conditional transfers in RISC, or even in current CISC.

> Maybe it's because I think all future computers should run real-time
> operating systems, and because I do research in real-time system
> scheduling and optimization, that I hold these views.

There are many situations which can not be done in real time.  Also, the
operating system should be a very minor part of the software, and do only
what an operating system, and not user modules, must.
-- 
Herman Rubin, Dept. of Statistics, Purdue Univ., West Lafayette IN47907
Phone: (317)494-6054
hrubin@l.cc.purdue.edu (Internet, bitnet)	{purdue,pur-ee}!l.cc!cik(UUCP)