Path: utzoo!attcan!uunet!hsi!mfci!cutler
From: cutler@mfci.UUCP (Ben Cutler)
Newsgroups: comp.arch
Subject: Re: VLIW Architecture
Keywords: VLIW
Message-ID: <1068@m3.mfci.UUCP>
Date: 8 Oct 89 04:06:56 GMT
References: <251FCB3F.12366@maccs.dcss.mcmaster.ca> <1050@m3.mfci.UUCP> <13050@pur-ee.UUCP> <1630@l.cc.purdue.edu> <1989Oct5.025841.2046@esegue.segue.boston.ma.us> <3449@alliant.Alliant.COM>
Sender: news@mfci.UUCP
Reply-To: cutler@mfci.UUCP (Ben Cutler)
Organization: Multiflow Computer Inc., Branford Ct. 06405
Lines: 44
Summary:
Expires:
Sender:
Followup-To:
Distribution:

In article <3449@alliant.Alliant.COM> lewitt@Alliant.COM (Martin Lewitt) writes:
>Ouch!!  Why don't the Floating Point System's AP120B and FPS-164 processors
>qualify as VLIW?  They were commercial products introduced in 1975 and 1981
>respectively:
>
>   [ lots of deleted text ]
>
>Let's get the history right.
>       1) The first commercially available VLIW machine was the FPS-164,
>          not the Multiflow (by 6 years).

VLIWs and Trace Scheduling (TM) Compiler technology go hand-in-hand.  Ellis
describes why the FPS-164 doesn't qualify as a good compilation target in
his thesis:

``Many micro-programmed architectures have so-called `hot spots', latches,
that will hold a value for only one cycle or until another value displaces it.
For example, the outputs of the multiplier and adder pipelines in the FPS-164
[ APAL64 Programmer's Guide, FPS 1982 ] are latched and can be fed back to the
register banks or directly to the inputs of the pipelines.  There isn't
enough bandwidth into the register banks, so to keep the pipelines full the
programmer is often forced to use the latched pipeline outputs directly as
pipeline inputs.  But the programmer must be very careful to synchronize the
pipelines' inputs and outputs, since the next value coming out of the
pipeline will overwrite the previous latched value.''

Ellis goes on to describe how hot spots are unpleasant artifacts for a
compiler to deal with, how it isn't always possible to move a value in a
latch out of the way and into a register bank, or if you can do so by making
room in the register bank, that in turn might require displacing a value in
another latch, and so on.  If you want to do a good job on complex loops,
a backtracking compiler (blech) may be in order.

According to Ellis, ''Every value producing operation and every data transfer
reads its operands from a register bank and delivers its result to a register
bank.''  Ellis notes that, ``Architectures with hot spots are easy to build,
but building compilers for them is hard. (It took years to build the FPS-164
compiler.)  Better to build the compiler and hardware in parallel, avoiding
hardware features that can't be used easily by the compiler.''

If you want to know more, then read the thesis, which is extremely
well-written and informative.

( My apologies for any typos in the above quotations. )