Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!uunet!husc6!bbn!rochester!PT.CS.CMU.EDU!K.GP.CS.CMU.EDU!lindsay
From: lindsay@K.GP.CS.CMU.EDU (Donald Lindsay)
Newsgroups: comp.arch
Subject: Re: register windows
Message-ID: <199@PT.CS.CMU.EDU>
Date: Sun, 18-Oct-87 19:38:13 EDT
Article-I.D.: PT.199
Posted: Sun Oct 18 19:38:13 1987
Date-Received: Mon, 19-Oct-87 02:00:19 EDT
Sender: netnews@PT.CS.CMU.EDU
Organization: Carnegie-Mellon University, CS/RI
Lines: 31
Keywords: register windows

References:


It occurs to me that register-window machines tend to remember data which a
cache would (correctly) overwrite.

For example, assume that a program is half a dozen calls deep. What is the
average amount of execution before the bottom (oldest) window is uncovered ?

A reasonable answer: more than the average amount of execution between
system calls, or more than the average amount of execution between
interrupts, or more than the average amount of execution between task
switches, or all of the above.

If this dead weight is sent to memory more than once (by the inevitible
second system call, or second interrupt) then considerable traffic could
occur.  This leads to the idea that a register-window machine should spill
its elderly data to memory, perhaps at the first interrupt, perhaps at the
first system call, or perhaps by a "trickle" mechanism. It shouldn't recall
these values back from memory until some (much) later time - say, when their
slot is returned-to (or when the one above them is returned-to) (or when
return-from-interrupt notices that they are almost uncovered).

It would be interesting to see simulation results on this. There is an
excellent article in the September issue of (IEEE) Computer, by Flynn and
others, which reports simulation results on various other register models.
I consider the article flawed by its emphasis on counting memory traffic,
while ignoring the topic of syscalls and interrupts. If they would factor
these in, then the ideas presented above could be tested.

Yes, I know that interrupt frequency is system-dependant (and so on). 
My point remains. Our understanding of register issues still needs work.
-- 
	Don		lindsay@k.gp.cs.cmu.edu    CMU Computer Science