🚀 go-pugleaf

   As the caches start to get as smart as we have discussed they
begin to look more and more similar to a peice of fast local
memory or large register set.  A good example of this can be
seen with the Cray-2 which has a 2k (I think maybe more) of very
fast local memory.  The compiler in all of its wisdom makes
decisions about what it wants to store and hopefully does a
better job for vectors and other pathological cases than normal
cache management does.  The only problem with this scheme as I
see it is that the program must know exactly how much local memory
it has in advance (although that might also be a problem for
smart compilers using smart caches) and so programs must be
recompiled for machines with different size local memories.
Does anyone else have any more semi-rational thoughts on this
and where caches might go in the future?

Mark Levison
levisonm@qucis.queensu.ca
Queen's University
Kingston, Ont.
Canada