Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP Path: utzoo!mnetor!seismo!rutgers!ames!ucbcad!ucbvax!decvax!decwrl!pyramid!prls!mips!hansen From: hansen@mips.UUCP (Craig Hansen) Newsgroups: comp.arch,comp.sys.nsc.32k Subject: Branch prediction in the 532 Message-ID: <324@dumbo.UUCP> Date: Tue, 21-Apr-87 02:21:29 EST Article-I.D.: dumbo.324 Posted: Tue Apr 21 02:21:29 1987 Date-Received: Wed, 22-Apr-87 05:37:36 EST Lines: 34 Xref: mnetor comp.arch:1037 comp.sys.nsc.32k:102 I've been reading two publications from National, one the infamous "Series 32000: NS32532 Competitive Comparison" that lists the eight separate patent applications we've already heard about, the other titled "Series 32000: NS32532 Technical Introduction." The feature that most caught my eye was that of branch prediction, for which I quote from the "Technical Introduction": The selection is based on the branch condition and direction; forward branches are predicted as taken, backward branches as non-taken. Well, I don't see how the branch prediction can depend on the branch condition, since waiting for the condition corresponds to no branch prediction at all. (Certainly novel and non-obvious...) But when I was a young sprout, they always told me that backward branches often represented loops and were most frequently taken. Forward branches often represent conditional expressions, and are more ambivalent. As far as I can see from the description, there is no prediction based on past behaviour from previous invocations of the branch instruction. If the description in National's literature is entirely accurate, the branch prediction scheme used is seriously broken and should predict the wrong answer more often than the right answer. Then again, if National hasn't seen silicon yet, and uses the opportunity to fix it now, I may have just saved them a few tens of thousands of dollars in wafer costs and hundreds in otherwise lost revenue.... -- Craig Hansen Manager, Architecture Development MIPS Computer Systems, Inc. ...decwrl!mips!hansen