Path: utzoo!utgpu!news-server.csri.toronto.edu!clyde.concordia.ca!uunet!tut.cis.ohio-state.edu!rutgers!umn-d-ub!cs.umn.edu!mike From: mike@cs.umn.edu (Mike Haertel) Newsgroups: comp.lang.misc Subject: Re: (Register windows considered harmful (Was: function calls) Message-ID: <1990Mar25.153123.27724@cs.umn.edu> Date: 25 Mar 90 15:31:23 GMT References: <6283@crdgw1.crd.ge.com> <14288@lambda.UUCP> Distribution: comp Organization: Free Software Foundation Lines: 25 In article gumby@Gang-of-Four.Stanford.EDU (David Vinayak Wallace) writes: >Can someone point me to any real comparison numbers on the performance >of such machines vs reg. window machines? The point in defense of reg >windows is usually that they are cheap to make (in chip landscape) and >that the space used by the directory would better be spent on more >registers themselves. However I've never seen real data on their >performance. There is a paper by David Wall of DECWRL, "Register Windows vs. Register Allocation", that is probably what you're looking for. You can order it from them. Basically, they demonstrate that a compiler using intelligent register allocation (they actually have an interprocedural allocator that they run during the link phase) can get performance as good as, and often better than, register windows, even for the kinds of programs for which register windows are at their best. And, as you note, although register windows are very good for some things, they kill you if you have lexically nested procedures, heavily recursive procedures, or lots of other things that are not in the programming model they try to address. -- Mike Haertel "We are trying to support small memory machines." -- Larry McVoy