[Top] [All Lists]

Re: CVS linux

To: "Maciej W. Rozycki" <>
Subject: Re: CVS linux
From: Ralf Baechle <>
Date: Thu, 3 Apr 2003 17:42:19 +0200
In-reply-to: <>; from on Thu, Apr 03, 2003 at 04:48:21PM +0200
Original-recipient: rfc822;
References: <> <>
User-agent: Mutt/
On Thu, Apr 03, 2003 at 04:48:21PM +0200, Maciej W. Rozycki wrote:

> > I know of one machine where changing the size of the cacheline is supposed
> > not to work, that's the MIPS Magnum 4000 and it's close relatives.
>  Hmm, why -- is such a change observable externally in any way?  Of
> course you can't switch the other way if the s-cache uses a line width of
> 16 bytes.  Maybe that's the case with the Magnum?

It's a hardware problem with the memory controller I was told by one of
it's developpers.  That forced them to run the machine with an different
line size for D-cache and I-Cache.  There's various revs of the Magnum's
memory controller and only one of them got all the cases right ...

Maybe DECstation and other SGI hardware got that better?

> > Anyway, I put the check there for the unlikely case there are broken
> > systems out there.  In practice I assume vendors were aware of this
> > problem and the check is purely theoretical and for paranoid correctness's
> > sake.
>  Still V3.0 should be taken into account.


> > It seems like changing the configuration to larger cache lines where
> > possible should improve performance somewhat.  If all the cache code is
>  Why?  It isn't that obvious especially as a p-cache miss costs a single
> cycle only.

During my recent work on the cache code I found the execution time of
cache flushing code to be quite a bit higher than previously assumed so
larger lines would help reducing that also.

> > working truly correct we also should no longer see VCE exceptions on
> > R4000SC processors - the reason why Indys are still a valuable test tool.
>  As are DECstations which use the opposite endianness -- so you can test
> code both ways.

A bunch of evaluation boards that support running in the other endianess
and way exceed the performance of any R4000-based platform.  Just having
to flip a switch on the board is very handy.


<Prev in Thread] Current Thread [Next in Thread>