Reducing cache misses by snarfing writebacks in non-inclusive memory

Abstract

A non-inclusive multi-level cache memory system is optimized by removing a first cache content from a first cache, so as to provide cache space in the first cache. In response to a cache miss in the first and second caches, the removed first cache content is stored in a second cache. All cache contents that are stored in the second cache are limited to have read-only attributes so that if any copies of the cache contents in the second cache exist in the cache memory system, a processor or equivalent device must seek permission to access the location in which that copy exists, ensuring cache coherency. If the first cache content is required by a processor (e.g., when a cache hit occurs in the second cache for the first cache content), room is again made available, if required, in the first cache by selecting a second cache content from the first cache and moving it to the second cache. The first cache content is then moved from the second cache to the first cache, rendering the first cache available for write access. Limiting the second cache to read-only access reduces the number of status bits per tag that are required to maintain cache coherency. In a cache memory system using a MOESI protocol, the number of status bits per tag is reduced to a single bit for the second cache, reducing tag overhead and minimizing silicon real estate used when placed on-chip to improve cache bandwidth.