Path: utzoo!attcan!uunet!husc6!mailrus!ames!amdahl!pyramid!prls!mips!mash
From: mash@mips.COM (John Mashey)
Newsgroups: comp.arch
Subject: Re: integer multiplies (Was: architecture/implementation -- MIPS)
Message-ID: <2242@winchester.mips.COM>
Date: 25 May 88 07:45:44 GMT
References: <2231@gumby.mips.COM> <3006@phoenix.Princeton.EDU> <2241@gumby.mips.COM>
Reply-To: mash@winchester.UUCP (John Mashey)
Organization: MIPS Computer Systems, Sunnyvale, CA
Lines: 47

In article <2241@gumby.mips.COM> earl@mips.COM (Earl Killian) writes:
>In article <3006@phoenix.Princeton.EDU> lgy@pupthy2.PRINCETON.EDU (Larry Yaffe) writes:

>       I'm curious and a little surprised about the relative speed of
>   integer and floating point multiplies....

Here are the dynamic frequencies of integer and FP multiplies, as a percentage
of instructions (NOT instruction cycles, which is much more complex to 
describe, given instruction overlap), in a quick sample of programs.
No science presumed: I just grabbed a few programs believed to be OK:

Program		Integer		FP
as1		 .02%		 0
ccom		 .08%		 0
espresso	<.01%		<.01%
gnuchess	 .06%		0
hspice		 .02%		2.4%
linpackd*	 .13%		8.2%
nroff		 .45%		0
whetd*		1.42%		7.38%
timberwolf	 .31%		.45%

As a gross cycle estimate, multiply the Integer # by 10, and the FP one by 5.
(* = benchmark, above)

Conclusions:
1) If a program uses FP in any serious fashion,  FP mult > integer mult.

2) A few programs (nroff, whetd, and timberwolf) use integer mult enough
that it would be nice to be faster, i.e., a 10-cycle multiply causes
>2% of the instruction cycles.

3) Many programs don't execute integer mult enough to be noticable.

Of course, our compilers do strength reduction, common subexpression
elimination, etc, and convert mulitplies by constants into shifts/adds/subs;
i.e., they've squished out a lot of actual multiplies.

This is perfect example where your choice of benchmarks tells
completely different answers to the question: "should we have an integer
mult instruction, and if so, how much silicon are we willing to devote to
making it X cycles."
-- 
-john mashey	DISCLAIMER: <generic disclaimer, I speak for me only, etc>
UUCP: 	{ames,decwrl,prls,pyramid}!mips!mash  OR  mash@mips.com
DDD:  	408-991-0253 or 408-720-1700, x253
USPS: 	MIPS Computer Systems, 930 E. Arques, Sunnyvale, CA 94086