Path: utzoo!utgpu!attcan!uunet!seismo!sundc!pitstop!sun!quintus!ok
From: ok@quintus
Newsgroups: comp.arch
Subject: more about division
Message-ID: <628@quintus.UUCP>
Date: 4 Nov 88 12:18:21 GMT
Sender: news@quintus.UUCP
Reply-To: ok@quintus ()
Organization: Quintus Computer Systems, Inc.
Lines: 23


Just another little data point on the RISC/division/COBOL topic.
I mentioned that one of the things you need for COBOL is fast text->binary
and fast binary->text conversion, and hinted that I have some ideas about
how to do them fast without multiply/divide/remainder.  (No big deal; all
it takes is tables and a twisted mind.)  I haven't done the 32-bit version
yet, but I have done a 20-bit one (generates 7 digits).
		Fast	Slow	Slow/Fast
MC68020		15.6	76	 4.9	
Whizzbang	 3.4	51	15.2
(times in microseconds per conversion)  The Whizzbang is a RISC which has
no hardware support for division, not even a divide step.  The Fast and
Slow times include subroutine call, but the Fast version is small enough
to do in-line.  It uses bit-field extraction, adds, word loads, and byte
stores.

What's really interesting about the example is that the Slow version takes
41% of the total program time on the 68020, 44% on the Whizzbang.  The rest
of the time was going in I/O.  So although the Whizzbang is about 4.6 times
as fast on RISCy instructions, it ended up running the program only 1.6 times
faster.  Oddly enough, C compilation times on the two machines also exhibit
about a 1.6 ratio.  Perhaps the next generation of workstations should have
more effort put into speeding up I/O than into speeding up the CPU?