Relay-Version: version B 2.10 5/3/83; site utzoo.UUCP
Path: utzoo!utgpu!water!watnot!watmath!clyde!rutgers!sri-spam!sri-unix!hplabs!decwrl!pyramid!oliveb!intelca!mipos3!cpocd2!howard
From: howard@cpocd2.UUCP
Newsgroups: comp.arch
Subject: Re: RISC register windows
Message-ID: <451@cpocd2.UUCP>
Date: Wed, 25-Feb-87 20:14:18 EST
Article-I.D.: cpocd2.451
Posted: Wed Feb 25 20:14:18 1987
Date-Received: Sat, 28-Feb-87 04:19:16 EST
References: <1881@homxc.UUCP> <898@moscom.UUCP> <476@mntgfx.MENTOR.COM> <430@cpocd2.UUCP> <1093@uwmacc.UUCP> <443@cpocd2.UUCP> <5746@amdahl.26 Feb 87 01:14:18 GMT
Reply-To: howard@cpocd2.UUCP (Howard A. Landman)
Organization: Intel Corp. ASIC Services Organization, Chandler AZ
Lines: 47

In article <1335@Shasta.STANFORD.EDU> andy@Shasta.UUCP (Andy Freeman) writes:
>The sliding register window had an area for parameters passed
>to the procedure, locals, and parameters passed to other procedures.
>There were also a set of global registers.
>
>One argument for fixed window sizes is that register access
>shouldn't require an addition.  Bit-wise or is one thing (or the
>register number with the current base to determine the register);
>addition is an expensive operation to put on the register access
>data path.  This basically restricts window sizes to powers of two.

Actually we didn't even use bitwise manipulation, we just built double
decoders for the overlapped registers so that two addresses would both
activate the same register.  You can easily see this in the chip photos,
in the upper left, as light and dark bands in the regfile decoder.  In
retrospect I'm not sure that was the best method.  Dave Johannsen once
suggested to me an alternate window structure in which the overlap could
be implemented with a "normal" RAM and only 2 extra gates!

>Then again, I seem to remember that the window size wasn't a power
>of two on the Berkeley RISC, which surprised me.  Maybe I'm wrong.

You are.  In RISC-I, say, in each window you have 6 "high" and 6 "low"
registers which overlap with the next window, 10 "local" ones which
don't, and 10 "global" ones which are the same in all windows.  That
adds up to 32.  However, the 10 global registers only exist in one
place, not in each window, so each window has 22 registers which are
non-global; this may be where you got your non-power-of-2-idea.  On
the other hand, adding an extra window costs 16 registers, because
either the high or the low (I forget which) already exist in the
previous window, so we're back to a power of 2 again.

The numbers in RISC-II are different, but could be plugged into the
preceding paragraph without making it untrue (I think).

Let H = # of high registers, M = # of local registers, L = # of low registers,
G = # of global registers, R = # of registers seen in each window, W = # of
windows, P = # of physical registers.  The following must hold for 2 or more
windows, assuming circular overflow (top-window high = bottom-window low):

	H = L
	R = H + M + L + G
	P = W*(M+L) + G
-- 

	Howard A. Landman
	...!intelca!mipos3!cpocd2!howard