Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!uunet!seismo!rutgers!ames!amdcad!tim
From: tim@amdcad.AMD.COM (Tim Olson)
Newsgroups: comp.arch
Subject: Re: What with these Vector's anyways? [really string ops]
Message-ID: <17780@amdcad.AMD.COM>
Date: Fri, 31-Jul-87 12:02:06 EDT
Article-I.D.: amdcad.17780
Posted: Fri Jul 31 12:02:06 1987
Date-Received: Sun, 2-Aug-87 01:02:29 EDT
References: <218@astra.necisa.oz> <142700010@tiger.UUCP>
Reply-To: tim@amdcad.UUCP (Tim Olson)
Distribution: world
Organization: Advanced Micro Devices, Inc., Sunnyvale, Ca.
Lines: 38
Keywords: scalar vs. vectors, benchmarks

In article <279@diab.UUCP> pf@.UUCP (Per Fogelstrom) writes:
+-----
| Be serious! Do you really belive in Dhrystone? Okay i do admitt that we don't
| have anything better for the moment, but soon i hope.
+-----
I hope so, too.  Dhrystone has been "twisted" by the translation from
Ada to C (too much emphasis on string handling) and by the use of
globally-optimizing compilers (code elimination) so much that it now
should only be used as a *very* rough ballpark figure (unfortunately,
people tend to like single figures-of-merit).  There are on-going
attempts to define new benchmarks; let's hope that some of these
problems are being addressed.

+-----
| What this benchmark really tests is the compilers ability to remove code that
| does'nt really do anything. And of course the cpu's ability to handle strings.
| Just check the Am29000. The designers put in an instruction just to speed up
| this benchmark.... Now what, have we reaced a new point, where cpu's are
| designed for the bencmarks instead of vice versa, as it was before ????
+-----
No, we didn't put in the cpbyte instruction to "speed up Dhrystone" --
it was added because many internal programs that we ran showed that such
an instruction could be used to speed up C string handling (size
unknown, string terminated by a \0) in general.

Many processors have specialized instructions for strings -- Intel's 80x86
family (movs, cmps), AT&T's 32100 (strlen, strcmp), National's 32x32
family (movs, cmps), NEC's V70 (movc, cmpc), Fairchild's Clipper (movc,
cmpc macros).  However, most of these don't address "C-type" strings;
they require a character count as a parameter.

We were able to identify a single, simple instruction which increased
the performance of C string handling.  Others have also used this same
technique in software to also obtain performance increases.

	-- Tim Olson
	Advanced Micro Devices
	(tim@amdcad.amd.com)