Abstract: A cache and method of maintaining cache coherency in a data processing system are described. The data processing system includes a system memory, a plurality of processors, and a plurality of caches coupled to an interconnect. According to the method, a first data item is stored in a first of the caches in association with an address tag indicating an address of the first data item. A coherency indicator in the first cache is set to a first state that indicates that the address tag is valid and that the first data item is invalid. If, while the coherency indicator is set to the first state, the first cache receives a data transfer on the interconnect associated with the address indicated by the address tag, where the data transfer includes a second data item that is modified with respect to a corresponding data item in the system memory, the second data item is stored in the first cache in association with the address tag.

Abstract: A first data item is stored in a first cache in association with an address tag indicating an address of the data item. A coherency indicator in the first cache is set to a first state that indicates that the first data item is valid. In response to another cache indicating an intent to store to the address indicated by the address tag while the coherency indicator is set to the first state, the coherency indicator is updated to a second state that indicates that the address tag is valid and that the first data item is invalid. Thereafter, in response to detection of a remotely-sourced data transfer that is associated with the address indicated by the address tag and that includes a second data item, a determination is made, in response to a mode of operation of the first cache, whether or not to update the first cache.

Abstract: A cache coherency protocol uses a “Tagged” coherency state to track responsibility for writing a modified value back to system memory, allowing intervention of the value without immediately writing it back to system memory, thus increasing memory bandwidth. The Tagged state can migrate across the caches (horizontally) when assigned to a cache line that has most recently loaded the modified value. Historical states relating to the Tagged state may further be used. The invention may also be applied to a multi-processor computer system having clustered processing units, such that the Tagged state can be applied to one of the cache lines in each group of caches that support separate processing unit clusters. Priorities are assigned to different cache states, including the Tagged state, for responding to a request to access a corresponding memory block. Any tagged intervention response can be forwarded only to selected caches that could be affected by the intervention response, using cross-bars.

Abstract: The index field of an address maps to low order cache directory address lines. The remaining cache directory address line, the highest order line, is indexed by the parity of the address tag for the cache entry to be stored to or retrieved from the corresponding cache directory entry. Thus, even parity address tags are stored in cache directory locations with zero in the most significant index/address bit, while odd parity address tags are stored in cache directory locations with one in the most significant index/address bit. The opposite arrangement (msb 1=even parity; msb 0=odd parity) may also be employed, as may configurations in which parity supplies the least significant bit rather than the most significant bit. In any of these cases, even/odd parity is implied based on the location of the address tag within the cache directory.

Abstract: A method of controlling eviction of cache blocks to override eviction of a value which is reserved for a later operation. When a value is loaded into a cache of the processor and is reserved using a lwarx instruction, it sometimes is evicted from the cache due to the need to store other values in the cache set that the value is mapped to. The present invention provides a method of overriding eviction of reserved values by evicting a selected block of the cache which is a block other than the block containing the reserved value. The reserved value is indicated as being reserved by loading a memory address associated with the value into a reservation unit of the cache, and making a reservation flag in the reservation unit active.

Abstract: In evicting data from a first cache in a level other than the lowest in a multilevel cache hierarchy, data is written to the system bus and snooped back into a second cache on a lower level in the cache hierarchy. The need for a private data path between the two caches is thus eliminated, and the second cache memory need not be dual-ported. The reload path employed for updating the second cache is reused to snoop cast-outs off the system bus. As a result of the first cache evicting data via the system bus, the second cache never contains data which is modified (M) with respect to system memory and other devices in a multiprocessor system get updated earlier. The need for error correction code (ECC) checking is eliminated, together with the associated additional bits, and may be replaced by simple parity checking. The bus into the second cache thus requires fewer bits, consumes less area, and may be operated at a higher frequency.

Abstract: To avoid multiplexing within the critical address paths, the same address field is employed as a index to the cache directory and cache memory regardless of the cache memory size. An increase in cache memory size is supported by increasing associativity within the cache directory and memory, for example by increasing congruence classes from two members to four members. For the smaller cache size, an additional address “index” bit is employed to select one of multiple groups of address tags/data items within a cache directory or cache memory row by comparison to a bit forced to a logic 1.

Abstract: A data processing system and method of maintaining cache coherency in a data processing system are described. The data processing system includes a plurality of caches and a plurality of processors grouped into at least first and second clusters, where each of the first and second clusters has at least one upper level cache and at least one lower level cache. According to the method, a first data item in the upper level cache of the first cluster is stored in association with an address tag indicating a particular address. A coherency indicator in the upper level cache of the first cluster is set to a first state that indicates that the address tag is valid and that the first data item is invalid. Similarly, in the upper level cache of the second cluster, a second data item is stored in association with an address tag indicating the particular address. In addition, a coherency indicator in the upper level cache of the second cluster is set to the first state.

Abstract: A method of evicting a cache block from a congruence class in a cache of a multi-processor computer system. After a cache miss, one of the cache blocks in the congruence class is selected for eviction based on the cache coherency states of the cache blocks. Any block having an Invalid state is preferably selected but, if such a block is not present, then one is preferably selected that has an invalid-type state, such as the new Hover state. If there are many blocks in the Hover state, then the least recently used is deallocated. If neither of these types of blocks are present, then a block is preferably selected for deallocation which is in the Modified state. This intelligent approach to victim selection generally improves cache performance.

Abstract: A method of managing and speculatively issuing architectural operations in a computer system is disclosed. A first architectural operation at a first coherency granule size is issued and translated into a large-scale architectural operation. The first architectural operation can be a first cache instruction directed to a memory block, and the translating results in a page-level cache instruction being issued which is directed to a page that includes the memory block. The large-scale architectural operation is transmitted to a system bus of the computer system. A system bus history table may be used to store a record of the large-scale architectural operations. The history table then can be used to filter out any later architectural operation that is subsumed by the large-scale architectural operation. The history table monitors the computer system to ensure that the large-scale architectural operations recorded in the table are still valid.

Abstract: Depending on a processor or instruction mode, a data cache block store (dcbst) or equivalent instruction is treated differently. A coherency maintenance mode for the instruction, in which the instruction is utilized to maintain coherency between bifurcated data and instruction caches, may be entered by setting bits in a processor register or by setting hint bits within the instruction. In the coherency maintenance mode, the instruction both pushes modified data to system memory and invalidates the cache entry in instruction caches. Subsequent instruction cache block invalidate (icbi) or equivalent instructions targeting the same cache location are no-oped when issued by a processor following a data cache block store or equivalent instruction executed in coherency maintenance mode. Execution of the data cache clock store instruction in coherency maintenance mode results in a novel system bus operation being initiated on the system bus.

Abstract: A register associated with the architected logic queue of a memory-coherent device within a multiprocessor system contains a flag set whenever an architected operation—one which might affect the storage hierarchy as perceived by other devices within the system—is posted in the snoop queue of a remote snooping device. The flag remains set and is reset only when a synchronization instruction (such as the “sync” instruction supported by the PowerPC™ family of devices) is received from a local processor. The state of the flag thus provides historical information regarding architected operations which may be pending in other devices within the system after being snooped from the system bus. This historical information is utilized to determine whether a synchronization operation should be presented on the system bus, allowing unnecessary synchronization operations to be filtered and additional system bus cycles made available for other purposes.

Abstract: A method of managing and speculatively issuing architectural operations in a computer system. A first architectural operation is snooped and translated into a plurality of granular architectural operations to effect a large-scale architectural operation. The first architectural operation can be a first cache instruction directed to a memory block, and a plurality of cache instructions are issued which are directed to memory blocks contained in a page associated with the memory block. The granular architectural operations are transmitted to a processor bus of the computer system. A processor bus history table may be used to store a record of the large-scale architectural operation. The history table then can filter out any later architectural operation that is subsumed by the large-scale architectural operation. The history table monitors the processor bus to ensure that the large-scale architectural operations recorded in the table are still valid.

Abstract: An integrated circuit comprises a semiconductor substrate having integrated circuitry formed therein. According to the present invention, the integrated circuitry includes a plurality of subcircuits, including first and second subcircuits that concurrently operate at diverse first and second frequencies, respectively. According to one embodiment, the integrated circuit has a clock signal that alternates between an active state and an inactive state at a third frequency and is broadcast to all of the subcircuits. In this embodiment, at least one subcircuit among the plurality of subcircuits, for example, a processor, operates in response to the clock signal at the third frequency, which is higher than the first frequency. According to another embodiment, the subcircuits each communicate with at least one other subcircuit via a latch-to-latch interface.

Abstract: To avoid multiplexing within the critical address path, the same field from an address is employed to index rows within a cache directory and memory regardless of the size of the cache memory. Depending on the size of the cache memory being employed, different address bits (such as Add[12] or Add[25] are employed as a "late select" for the last stage of multiplexing within the cache directory and cache memory. Since smaller address tag fields are employed for the larger cache memory size, the extra address tag bit is forced to a logic 1 within the cache directory and compared to a logic 1 by address tag comparators at the output of the cache directory.

Abstract: A method of avoiding deadlocks in cache coherency protocol for a multi-processor computer system, by loading a memory value into a plurality of cache blocks, assigning a first coherency state having a higher collision priority to only one of the cache blocks, and assigning one or more additional coherency states having lower collision priorities to all of the remaining cache blocks. Different system bus codes can be used to indicate the priority of conflicting requests (e.g., DClaim operations) to modify the memory value. The invention also allows folding or elimination of redundant DClaim operations, and can be applied in a global versus local manner within a multi-processor computer system having processing units grouped into at least two clusters.

Abstract: A system and method for transferring bus operations in a processing system which includes at least one processor, the method and system include issuing a plurality of ordered bus operations by the at least one processor, wherein the plurality of bus operations include a first bus operation and a second bus operation, wherein the second bus operation is issued next after the first bus operation is issued. It also determines if a first response for the first bus operation has been received by the at least one processor prior to issuing the second bus operation, wherein the first response indicates that the first bus operation can be transferred.

Abstract: A cache coherency protocol uses a "Tagged" coherency state to track responsibility for writing a modified value back to system memory, allowing intervention of the value without immediately writing it back to system memory, thus increasing memory bandwidth. The Tagged state can migrate across the caches (horizontally) when assigned to a cache line that has most recently loaded the modified value. Historical states relating to the Tagged state may further be used. The invention may also be applied to a multi-processor computer system having clustered processing units, such that the Tagged state can be applied to one of the cache lines in each group of caches that support separate processing unit clusters. Priorities are assigned to different cache states, including the Tagged state, for responding to a request to access a corresponding memory block. Any tagged intervention response can be forwarded only to selected caches that could be affected by the intervention response, using cross-bars.

Abstract: When a device snooping the system bus of a multiprocessor system detects an operation requesting data which is resident within a local memory in a coherency state requiring the data to be sourced from the device, the device attempts a intervention. If the intervention is impeded by a second device asserting a retry, the device sets a flag to provide historical information regarding the failed intervention. On a subsequent snoop hit to the same cache location, if the device again asserts an intervention and the snooped operation is again retried, the device undertakes an action to alter the coherency state of the requested cache item towards an ultimate coherency state expected to be the result of the original operation requesting the cache item. In the case where the requested cache item includes modified data resident in the device's local memory, the action may include a push operation writing the requested cache item to system memory.

Abstract: System and method for selectively adapting the burst mode writeback from cache to main memory consistent with the extent of a cache line actually modified by the processor and at the granularity of the bus connecting the cache to main memory. A cache controller speculatively reads a cache line with each address issued by this processor. When the address is related to a read cycle of the processor, the data is forwarded to the processor. When the address is related to a write cycle of the processor, the data read from the cache is compared to the write data from the processor to detect changes at a granularity consist with the size of the system data bus. The cache line stored in the cache upon such writing is marked at the granularity of the system data bus with tag bits to indicate which portions have been modified. Upon deallocation, the tag bits stored in the cache directory identify those portions of the cache lines requiring transmission back to main memory as an aspect of the burst writeback operation.