5. Secondary Cache Interface

5.4 Secondary Cache Way Prediction Table

The primary caches read simultaneously from two separate tag arrays, corresponding to each way in the cache, and then select the data based on the result of two parallel tag compares.

The secondary cache does not use this implementation because it would either require too many pins to read in two full copies of the data and tags, or add latency to externally multiplex two banks of memory. Instead, a way prediction table is used to determine which way to read from first.

The way prediction table is internal to the processor and has 8K one-bit entries, each entry corresponding to a pair of secondary cache blocks. The bit entry indicates which way of the addressed set has been most-recently used (MRU). When the secondary cache is accessed, this prediction bit is used as an address bit; thus the two ways in the secondary cache are shared in the same SSRAM bank.

The secondary cache way prediction table is indexed with a subset of 11 to 13 bits of the physical address, based on both the secondary cache block size, and the secondary cache size, as shown in Table 5-3. "0 ||" indicates a zero bit concatenated to the address to pad the index out to a full 13-bits.

Table 5-3 Secondary Cache Way Prediction Table Index

Three states are possible in the way prediction table:

• the desired data is in the predicted way
• the desired data is in the non-predicted way
• the desired data is not in the secondary cache

SCTWay pin.

The three possible states are handled in the following manner:

• If, after reading the tags for both ways, it is discovered that the data exists in the predicted way, the processor continues normally.
• If the data exists in the non-predicted way, the processor accesses this non-predicted way in the secondary cache and updates the way prediction table to point to this way.

---

---

• If the access misses in both ways of the secondary cache, the data is fetched from the system interface. If the state of the predicted way is found to be invalid, the fetched data is placed in it and the MRU is unchanged. However, if the state of the predicted way is found to be valid then the fetched data is placed into the non-predicted way, and the way prediction table is updated to point to this way since it is now the most-recently-used. (See page 58 in Errata.)