Tait TB9100, Руководство по обслуживанию, Страница: 40 / 132

40 Network Circuitry TB9100 Reciter Service Manual
© Tait Electronics Limited January 2006
ensure that the execution units always have a continual flow of instructions.
Static branch prediction minimizes disruptions in the instruction flow
through the pipelines by pre-fetching instructions following on from a
program branch.
A full description of the operation and instruction set of the RISC core is
outside the scope of this document. It is recommended that the MPC866
user’s manual (reference 2) be consulted for comprehensive coverage of the
topic.
Memory Caches and
Memory Managers
Ordinarily, the fastest rate that the MPC can fetch data from external
memory is one 32-bit instruction per two memory clock cycles. However,
the CPU can execute two instructions per memory clock cycle since its
internal clock is configured to be twice the memory clock (see
“Clock
Generation” on page 42 and “MPC Configuration” on page 45 ).
Therefore, the CPU performance is restricted by the available transfer
bandwidth from its program memory.
To enable the CPU to run at full speed, internal cache memory, which
enables fetching of data or instructions without wait states, is provided.
The MPC859T includes separate instruction and data caches of 1kwords
(4kbytes) each.
The CPU implements a Harvard architecture internally, ie. there are
separate instruction and data caches, each with separate data and address
buses. This allows data and instruction fetches from the caches to occur
simultaneously. Although the internal structure of the instruction and data
caches differs, they operate in an essentially similar fashion.
Considering instruction fetches only, when the CPU fetches an instruction
word and that instruction is found within the instruction cache, then it can
be retrieved without delay. Should the required instruction not be in cache
then the cache issues a request to the SIU (see
“System Interface Unit
(SIU)” on page 42 ) to fetch that instruction from main memory. Actually, a
total of four words is fetched in a burst fetch (see “SDRAM Burst Cycles”
on page 64 ) since the cache memory is organized in lines of four words each,
which are updated together.
If there is spare space in the cache, the new instruction data is stored in the
cache. Otherwise, space must be cleared in the cache to enable the newly
fetched data to be retained. To accomplish this, cache entries are examined
to find the oldest entry using a least-recently-used (LRU) algorithm.
The oldest cache entry thus found is discarded to make way for the new
data.
The MPC provides memory management support for high-end operating
systems through separate instruction and data memory managers. These
translate the logical address from the RISC core into a real address used to
access external memory and peripherals. Separate instruction and data
memory managers support the Harvard architecture of the CPU but their