Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!watmath!clyde!burl!ulysses!allegra!alice!grigg!dmr
From: dmr@grigg.UUCP
Newsgroups: net.arch
Subject: Strange architecture proposal
Message-ID: <1105@grigg.UUCP>
Date: Fri, 15-Feb-85 01:48:55 EST
Article-I.D.: grigg.1105
Posted: Fri Feb 15 01:48:55 1985
Date-Received: Sat, 16-Feb-85 06:04:48 EST
Lines: 34

Radford Neal's suggestion, in which memory addresses encode the
size of the addressed object, is a good one, but it isn't new.
So far as I know, David Wheeler, of Cambridge University, was
the first to suggest it several years ago.  I don't know if it
was ever published.  We learned of it when he was visiting
Bell Labs, and have since called it "the Wheeler device."

Most of Martindale's objections to the idea aren't important.
Different data types of the same size:  so long as you accept
that floating add is an operation different from fixed add, and
(if you like) that the same is true for move, the fact that different
data types may be the same size doesn't matter.

The plan doesn't contemplate extending the idea to big data objects,
just the usual byte-to-quadword range.  

It does indeed require that objects be aligned on a boundary divisible
by their size.  There has been debate about this, but I don't think
it is an onerous requirement.

Neal worries about registers.  I don't think he should.  It all works
out if you just make register references use their full width; that is,
the Wheeler device is a gadget that fiddles with data going between
the CPU and memory, and has nothing to do with registers.

It deserves to be tried.

The potential advantage is that you can use the redundancy provided
by requiring alignment of operands to reduce the number of bits needed
for the opcode and/or address specification, while still providing
orthogonality of references.  The disadvantage is more complicated
decoding and the need to be careful when moving pointers around.

	Dennis Ritchie