Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Posting-Version: version B 2.10.2 9/18/84 +MMDF+MULTI+2.11; site wcwvax.UUCP
Path: utzoo!watmath!clyde!burl!ulysses!allegra!mit-eddie!think!harvard!seismo!mcvax!ukc!qtlon!wcwvax!ian
From: ian@wcwvax.UUCP (Ian Kemmish)
Newsgroups: net.micro.68k,net.micro.16k
Subject: Restart vs. saving state on page faults
Message-ID: <330@wcwvax.UUCP>
Date: Wed, 28-Aug-85 07:08:01 EDT
Article-I.D.: wcwvax.330
Posted: Wed Aug 28 07:08:01 1985
Date-Received: Fri, 6-Sep-85 03:38:37 EDT
Organization: Whitechapel Computer Works, London
Lines: 19
Xref: watmath net.micro.68k:1095 net.micro.16k:383

Dave Trissl @ Motorola says the state-saving approach means
the processor doesn't have to wait until it knows a write has
succeeded without a page fault before starting on the next
instruction in the pipeline.  In contrast, instruction restart
implies waiting until the write works (and therefore that the
pipeline isn't as much use as it sounds).

Until last week, I would have agreed.

BUT . . . I was running some timing tests on my line-drawing algorithm
for the Whitechapel MG-1 (uses a NS32016).  Basically, it's just a
tarted-up Bresenham algorithm with an unwound inner loop of one
read-modify-write instruction followed by three register-register
instructions.   I was trying to special-case nearly-vertical lines
by having another loop with two of the register-register instructions
removed.   After allowing for the time taken to test for the special case,
I got a speedup of a massive 10% inside the loop.  That sounds like a
pretty efficient pipeline to me, folks!   (No, I don't know how they do
it - I just know what I measured.   Anyone else done any similar experiments?)