Advanced Computer Architecture, Imperial College 2001
31
More applns of prediction - cache bypass
•Suppose the cache line loaded by this instruction turns out never to be reused
•If this happens a lot, we could decide not to allocate data into the cache - so cache space is not used unnecessarily
•Often combined with hardware prefetching
•Eg in IBM Power3:
–hardware prefetch engine identifies fixed-stride access stream, initiates prefetching
–prefetched data is allocated directly into L1 cache but bypasses L2 cache
•(similar ideas in Pentium 4, Itanium etc).
•