Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!mnetor!seismo!lll-lcc!styx!ames!ptsfa!ihnp4!ihdev!pdg
From: pdg@ihdev.ATT.COM (Joe Isuzu)
Newsgroups: comp.arch,comp.sys.nsc.32k
Subject: Re: Branch prediction in the 532
Message-ID: <1372@ihdev.ATT.COM>
Date: Wed, 6-May-87 13:42:46 EDT
Article-I.D.: ihdev.1372
Posted: Wed May  6 13:42:46 1987
Date-Received: Sat, 9-May-87 01:08:55 EDT
References: <324@dumbo.UUCP> <809@killer.UUCP>
Reply-To: pdg@ihdev.UUCP (Joe Isuzu)
Organization: American Nasal Amputation Centre
Lines: 24
Xref: mnetor comp.arch:1199 comp.sys.nsc.32k:132

In article <809@killer.UUCP> jfh@killer.UUCP (John Haugh) writes:
>Craig stated that, like the rest of us, he was raised to believe that
>backwards branches make nice branches after while loops and such, and
>forward branches make nice branches before if-thens.  Has anyone bothered
>to think about adding branch instructions that tell the CPU explicitly
>whether or not they expect to be taken?  For example,
>
>	MBLEQ - most of the time branch less or equal
>	SBLEQ - seldom branch less or equal

The AT&T CRISP processor does this, with instructions like

	ifFjmpn, ifFjmpy (if false, expect to jump no or yes)
	ifTjmpn, ifTjmpy (if true, expect to jump no or yes)

These take 0 cycles for a correct prediction, and a max of 3 cycles
for an incorrect prediction (-1 cycle per instruction between the
compare and the branch).

For more info, see any of the CRISP articles from IEEE Compcon '87.
-- 

Paul Guthrie
ihnp4!ihdev!pdg			This Brain left intentionally blank.