On Fri, Feb 04, 2005 at 03:19:49PM +0000, Dominic Sweetman wrote:
> Only some CPUs suffer from aliases. A 4Kbyte direct-mapped cache must
> be alias free, because all the virtual index bits are the same (being
> in-page) as the physical address bits. That's true but irrelvant,
> since there aren't any 4Kbyte caches: but what's slightly less obvious
> is that a 16Kbyte 4-way set-associative cache is also alias free.
I had dark memory of some el cheapo CPU having 4k caches.
> 24K's "AR" bit trick applies *only* to the D-cache, and only to a
> 32Kbyte cache. (But then most alias problems are D-cache aliases, and
> 32Kbyte happens to be the most popular size for a 24K cache - so this
> is a trick worth doing).
>
> Note that I-cache aliases are not completely harmless; sometimes you
> want to invalidate any I-cache copies of some data, and if it's
> aliased you may miss some of them. Shared libraries are generally
> aligned to some large page-size multiple - so multiple text images are
> usually the same colour, and don't matter. You can get problems with
> trampolines and stuff.
Linux computes the necessary alignment on the fly. The method used is
not strictly correct because as you say it should account for possible
I-cache aliases also.
Seems it's cache day again today ;-)
Ralf
|