Subject: Re: 240/260 CPU hacking....
To: None <port-pmax@netbsd.org>
From: Toru Nishimura <nisimura@itc.aist-nara.ac.jp>
List: port-pmax
Date: 03/28/2001 11:55:16
> It's really all quite ingenious how they designed the R4x00 to be a
> chameleon, it's a 32bit and 64bit machine all at once. The more I
> read, the more respect I have for it.
R4x00 is basically true 64bit architecture with top 32bit off from
64bit addressing (I'm talking in software sense).
Bad story is, the blue print was fine endorsed by deep insights, while
implementation was fluctuated. The introduction of EXL bit is a
clever idea, however, interference between SR(IM), SR(IE) and SR(EXL)
bits is rather a jeopardy which rusts the brightness with costly
software work around. Many implementation of processors require
special treatments to avoid possible frauds when TLB miss or cache
miss are interacted badly.
In sum, R4000/R4400 ran slower than well tuned R3000(A). Computional
power per R3000 CPU cycle was better, indeed. It was the introduction
of QED R4600 _implementation_ which brought break thru for R4x00
architecture.
Tohru Nishimura