摘要:
A prefetching mechanism for a memory hierarchy which includes at least two levels of storage, with L1 being a high-speed low-capacity memory, and L2 being a low-speed high-capacity memory, with the units of L2 and L1 being blocks and sub-blocks respectively, with each block containing several sub-blocks in consecutive addresses. Each sub-block is provided an additional bit, called a r-bit, which indicates that the sub-block has been previously stored in L1 when the bit is 1, and has not been previously stored in L1 when the bit is 0. Initially when a block is loaded into L2 each of the r-bits in the sub-block are set to 0. When a sub-block is transferred from L1 to L2, its r-bit is then set to 1 in the L2 block, to indicate its previous storage in L1. When the CPU references a given sub-block which is not present in L1, and has to be fetched from L2 to L1, the remaining sub-blocks in this block having r-bits set to 1 are prefetched to L1. This prefetching of the other sub-blocks having r-bits set to 1 results in a more efficient utilization of the L1 storage capacity and results in a highter hit ratio.
摘要:
A cache memory system develops an optimum sequence for transferring data values between a main memory and a line buffer internal to the cache. At the end of a line transfer, the data in the line buffer is written into the cache memory as a block. Following an initial cache miss, the cache memory system monitors the sequence of data requests received for data in the line that is being read in from main memory. If the sequence being used to read in the data causes the processor to wait for a specific data value in the line, a new sequence is generated in which the specific data value is read at an earlier time in the transfer cycle. This sequence is associated with the instruction that caused the first miss and is used for subsequent misses caused by the instruction. If, in the process of handling a first miss related to a specific instruction, a second miss occurs which is caused by the same instruction but which is for data in a different line of memory, the sequence associated with the instruction is marked as an ephemeral miss. Data transferred to the line buffer in response to an ephemeral miss is not stored in the cache memory and limited to that portion of the line accessed within the line buffer.
摘要:
A method and apparatus for controlling access to its general purpose registers (GPRs) by a high end machine configuration including a plurality of execution units within a single CPU. The invention allows up to "N" execution units to be concurrently executing up to "N" instructions using the GPR sequentially or different GPR's concurrently as either SINK or SOURCE while at the same time preserving the logical integrity of the data supplied to the execution units. The use of the invention allows a higher degree of parallelism in the execution of the instructions than would otherwise be possible if only sequential operations were performed.A series of special purpose tags are associated with each GPR and execution unit. These tags are used together with control circuitry both within the GPR's, within the individual execution units and within the instruction decode unit, which permit the multiple use of the registers to be accomplished while maintaining the requisite logical integrity.
摘要:
A high speed buffer store arrangement for use in a data processing system having multiple cache buffer storage units in a hierarchial arrangement permits fast transfer of wide data blocks. On each cache chip, input and output latches are integrated thus avoiding separate intermediate buffering. Input and output latches are interconnected by 64-byte wide data buses so that data blocks can be shifted rapidly from one cache hierarchy level to another and back. Chip-internal feedback connections from output to input latches allow data blocks to be selectively reentered into a cache after reading. An additional register array is provided so that data blocks can be furnished again after transfer from cache to main memory or CPU without accessing the respective cache. Wide data blocks can be transferred within one cycle, thus tying up caches much less in transfer operations, so that they have increased availability.
摘要:
A prefetching mechanism for a system having a cache has, in addition to the normal cache directory, a two-level shadow directory. When an information block is accessed, a parent identifier derived from the block address is stored in a first level of the shadow directory. The address of a subsequently accessed block is stored in the second level of the shadow directory, in a position associated with the first-level position of the respective parent identifier.With each access to an information block, a check is made whether the respective parent identifier is already stored in the first level of the shadow directory. If it is found, then a descendant address from the associated second-level position is used to prefetch an information block to the cache if it is not already resident therein. This mechanism avoids, with a high probability, the occurrence of cache misses.
摘要:
A store through cache environment managed exclusively grants exclusivity on a large granularity basis. A cross-invalidate is realized for all changed lines via a single transmission when exclusivity is released. A dynamic table that operates in conjunction with a directory look-aside table (DLAT) determines a number of pages that can be held exclusive simultaneously. For adequate operating speed, the special table must be either fully associative or at least set associative. Alternatively, the table can be incorporated into the DLAT. Each DLAT entry is also extended to include a set of "resident" bits and a "valid nonresident" bit. When exclusively is released, the set of local change bits is broadcast to all processors. Upon receipt of such broadcast, the appropriate action is to change the "valid nonresident" indication to read-only and to clear residence bits whose corresponding local change bit is set.
摘要:
A branch history table (BHT) is substantially improved by dividing it into two parts: an active area, and a backup area. The active area contains entries for a small number of branches which the processor can encounter in the near future and the backup area contains all other branch entries. Means are provided to bring entries from the backup area into the active area ahead of when the processor will use those entries. When entries are no longer needed they are removed from the active area and put into the backup area if not already there. New entries for the near future are brought in, so that the active area, though small, will almost always contain the branch information needed by the processor.The small size of the active area allows it to be fast and to be optimally located in the processor layout. The backup area can be located outside the critical part of the layout and can therefore be made larger than would be practicable for a standard BHT.
摘要:
A multi-prediction branch prediction mechanism predicts each conditional branch at least twice, first during the instruction-fetch phase of the pipeline and then again during the decode phase of the pipeline. The mechanism uses at least two different branch prediction mechanisms, each a separate and independent mechanism from the other. A set of rules are used to resolve those instances as to when the predictions differ.
摘要:
A fast queue mechanism is provided which keeps a queue of changes (i.e. store actions) issued by each processor, which queue is accessible by all processors. When any processor issues a store action to a line of memory in the queue, the old data is overwritten with the new data. If the queue does not currently have a corresponding entry, a new entry is activated. Room for the new entry is made by selecting some existing entry, either the oldest or the least recently used, to be removed. An entry that is to be removed is first used to update the line corresponding to it in main memory. After the changes held in the entry to be removed are applied to the old value of the line (from main memory) and the updated value is put back into main memory, the entry in the queue is removed by marking it "empty". When a processor accesses a line of data not in its cache, a cache miss occurs and it is necessary to fetch the line from main memory. Such fetches are monitored by the queue mechanism to see if it is holding changes to the line being fetched. If so, the changes are applied to the line coming from main memory before the line is sent to the requesting processor. After a new entry is made in the queue mechanism, other store actions to the same entry by any processor may occur and usually a number of store actions will occur to the entry before it is removed to make room for another.
摘要:
Methods and apparatus are described for processing branch instructions using a history based branch prediction mechanism (such as a branch history table) in combination with a data dependent branch table (DDBT), where the branch instructions can vary in both outcome and test operand location. The novel methods and apparatus are sensitive to branch mispredictions and to operand addresses used by the DDBT, to identify irrelevant DDBT entries. Irrelevant DDBT entries are identified within the prediction mechanism using state bits which, when set, indicate that: (1) a given entry in the prediction mechanism was updated by the DDBT and (2) subsequent to such update a misprediction occurred making further DDBT updates irrelevant. Once a DDBT entry is determined to be irrelevant, it is prevented from updating the prediction mechanism. The invention also provides methods and apparatus for locating and removing irrelevant entries from the DDBT. The update packet, sent by the DDBT to the history based prediction mechanism, is expanded to include the test operand address actually used by the DDBT. If the state bits indicate the update is irrelevant, then the operand address can be used to locate and delete the offending DDBT entry since the DDBT is organized based on operand addresses. Additionally, the invention provides for inhibiting creation of further DDBT entries when a Branch Wrong Guess event occurs subsequent to a DDBT update to a given prediction mechanism entry.