Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP Path: utzoo!utgpu!water!watmath!clyde!rutgers!mit-eddie!bloom-beacon!oberon!cit-vax!ucla-cs!marc From: marc@ucla-cs.UUCP Newsgroups: comp.arch Subject: Re: register windows Message-ID: <8467@shemp.UCLA.EDU> Date: Sat, 3-Oct-87 13:07:00 EDT Article-I.D.: shemp.8467 Posted: Sat Oct 3 13:07:00 1987 Date-Received: Sun, 4-Oct-87 02:36:04 EDT References: <1242@k.gp.cs.cmu.edu> Sender: root@CS.UCLA.EDU Reply-To: marc@CS.UCLA.EDU (Marc Tremblay) Distribution: na Organization: UCLA Computer Science Department Lines: 35 In article <1242@k.gp.cs.cmu.edu> lindsay@k.gp.cs.cmu.edu (Donald Lindsay) writes: >Register windows interest me. > >Lose#1: more state to be saved/restored at task switches. (Caches can > be just declared invalid, or have tags.) >Lose#2: more hardware that can be on the ALU critical path. (Longer R bus, > deeper R decode ). We designed a register file which almost eliminates this problem. By using a "Shift-Register File", the data bus only goes through a single window of registers. In our case it is 16 registers. Adding more windows to the file, does not lenghten the data bus. >There are some ways to ways to fix things up. For example, instead of a >call sliding by (say) 8 registers, slide by a program-supplied distance. >The average slide might then be (say) 3. This economy helps lose#1, and >changes lose#2 in very implementation-specific ways. To solve some of this problem, we make use of variable-size windows. Our windows sizes vary in increments of 4 (8, 12, 16) so that they can fit a procedure call/return approprietly. In this way very few registers are wasted because of bigger-than-necessary-windows. That is the reason while we claim that our file of 64 registers performs just as well as a normal 128 register file. For more information you can read: 1) "A Reduced Register File for RISC Architectures", M. Huguet & T. Lang, Computer Architecture News (June 1985). 2) "VLSI Implementation of a Shift-Register File", M. Tremblay & T. Lang, 20th Hawaian Int. Conf. on System Sciences, pp 112-121 (January 1987). Marc Tremblay marc@CS.UCLA.EDU ...!(ihnp4,ucbvax)!ucla-cs!marc Computer Science Department, UCLA