International Business Machines Corporation

Thomas L Jeremiah

Guy L Guthrie

William John Starke

Bartholomew Blaner

Jeffrey A Stuecheli

Stamatis Vassiliadis

William Lloyd McNeil

Phillip G Williams

Piyush Chunilal Patel

Charles Ray Johns

Hugh Shen

Brian Mitchell Bass

Richard James Eickemeyer

Robert Alan Cargnoni

David James Shippy

Thuong Quang Truong

Stephen J Powell

Steven Lee Gregor

Gregory Scott Still

Alvan Wing Ng

Raymond J Eberhard

Michael James Mack

Philip G Williams

Harmony Lynn Helterhoff

Charles Robert Moore

Lawrence Jones

Carroll J Dick

Bruce J Ditmyer

Jeffrey A Steucheli

Robert Albert Healey

Karl F Pezdirtz

A system and method for tracking core load requests and providing arbitration and ordering of requests. When a core interface unit (CIU) receives a load operation from the processor core, a new entry in allocated in a queue of the CIU. In response to allocating the new entry in the queue, the CIU detects contention between the load request and another memory access request. In response to detecting contention, the load request may be suspended until the contention is resolved. Received load requests may be stored in the queue and tracked using a least recently used (LRU) mechanism. The load request may then be processed when the load request resides in a least recently used entry in the load request queue. CIU may also suspend issuing an instruction unless a read claim (RC) machine is available. In another embodiment, CIU may issue stored load requests in a specific priority order.

Load request scheduling in a cache hierarchy

A data processing system includes a processor core having an associated upper level cache and a lower level victim cache. In response to a memory access request of the processor core that specifies a non-modifying access to a target coherency granule, a determination is made whether the memory access request hits or misses in a directory of the lower level victim cache. In response to determining that the memory access request hits in the lower level victim cache in a data-valid coherence state, the lower level victim cache provides the target coherency granule of the memory access request to the upper level cache. The lower level victim cache preserves the target coherency granule in the lower level victim cache in a shared coherence state if the memory access request is of a first type and invalidates the target coherency granule if the memory access request is of a second type.

Victim cache replacement

A second lower level cache receives an LCO command issued by a first lower level cache on an interconnect fabric. The LCO command indicates an address of a victim cache line to be castout from the first lower level cache and indicates that the second lower level cache is an intended destination of the victim cache line. The second lower level cache determines whether to accept the victim cache line from the first lower level cache based at least in part on the address of the victim cache line indicated by the LCO command. In response to determining not to accept the victim cache line, the second lower level cache provides a coherence response to the LCO command refusing the identified victim cache line. In response to determining to accept the victim cache line, the second lower level cache updates an entry corresponding to the identified victim cache line.

Empirically based dynamic control of acceptance of victim cache lateral castouts

A data processing system includes a processor core having an associated upper level cache and a lower level victim cache. In response to a memory access request of the processor core, the lower level cache victim determines whether the memory access request hits or misses in the directory of the lower level victim cache, and the upper level cache determines whether a castout from the upper level cache is to be performed and selects a victim coherency granule for eviction from the upper level cache. In response to determining that a castout from the upper level cache is to be performed, the upper level cache evicts the selected victim coherency granule. In the eviction, the upper level cache reads out the victim coherency granule from the data array of the upper level cache only in response to an indication that the memory access request misses in the directory of the lower level victim cache.

A computer implemented method, a processor chip, a data processing system, and computer program product in a data processing system process information in a store cache of a data processing system. The store cache receives a first entry that includes a first address indicating a first segment of a cache line. The store cache then receives a second entry including a second address indicating a second segment of the cache line. Responsive to the first segment not being equal to the second segment, the first entry is chained to the second entry.

Chaining multiple smaller store queue entries for more efficient store queue usage

A cache memory includes a cache array including a plurality of congruence classes each containing a plurality of cache lines, where each cache line belongs to one of multiple classes which include at least a first class and a second class. The cache memory also includes a cache directory of the cache array that indicates class membership. The cache memory further includes a cache controller that selects a victim cache line for eviction from a congruence class. If the congruence class contains a cache line belonging to the second class, the cache controller preferentially selects as the victim cache line a cache line of the congruence class belonging to the second class based upon access order. If the congruence class contains no cache line belonging to the second class, the cache controller selects as the victim cache line a cache line belonging to the first class based upon access order.

Victim cache line selection

The present invention utilizes the good qualities of a single address concentrator (AC), without any extra chips or wires, and distributes the AC function among the various chips, making use of the fact that each chip in the system has a copy of the AC function therein. Using the distributed address concentrator function, each chip will handle approximately one-fourth of the command traffic and the average latency of servicing the commands will be approximately the same across each chip in the system.

Distributed address arbitration scheme for symmetrical multiprocessor system

A processor that efficiently obtains target path instructions in the presence of tight program loops includes at least one execution unit for executing instructions and instruction sequencing logic that supplies instructions to the at least one execution unit for execution. The instruction sequencing logic includes an instruction fetch buffer and a branch prediction unit including a branch target cache. In response to prediction of a branch instruction as taken, the branch target cache causes multiple copies of a target instruction group to be loaded into the instruction fetch buffer under the assumption that the branch instruction is a member of the target instruction group. Thereafter, the branch target cache causes all but one of the multiple copies to be canceled from the instruction fetch buffer prior to dispatch if the branch instruction does not belong to the target instruction group. Thus, the branch target cache can meet the instruction fetch cycle time of the processor even for the worst case condition in which the branch instruction is within the target instruction group.

Branch target cache and method for efficiently obtaining target path instructions for tight program loops

A cache system provides for accessing set associative caches with no increase in critical path delay, for reducing the latency penalty for cache accesses, for reducing snoop busy time, and for responding to MRU misses and cache misses. The cache array is accessed by multiplexing two most-recently-used (MRU) arrays which are addressed and accessed substantially in parallel with effective address generation, the outputs of which MRU arrays are generated, one by assuming a carryin of zero, and the other by assuming a carryin of one to the least significant bit of the portion of the effective addressed used to access the MRU arrays. The hit rate in the MRU array is improved by hashing within an adder the adder's input operands with predetermined additional operand bits.

Cache address generation with and without carry-in

A cache system provides for accessing set associative caches with no increase in critical path delay, for reducing the latency penalty for cache accesses, for reducing snoop busy time, and for responding to MRU misses and cache misses. A multiway cache includes a single array partitioned into a plurality of cache slots and a directory, both directory and cache slots connected to the same data bus. A first cache slot is selected and accessed; and then corresponding data is accessed from alternate slots while searching said directory, thereby reducing the latency penalty for cache access.

Multiway associative external microprocessor cache

A certain class of computer has been previously described which has improved performance through the analysis of instructions comprising the computer's control program and appending control information to the instructions in the form of tags. One such computer analyzes instruction cache lines as they are loaded into the cache to create the tags. A disadvantage of that design is the inability to create control information for portions of the cache line whose control tags depend on instructions in another cache line as well as the line being loaded. A method and apparatus is described herein which facilitates creation of control tags based on instructions which reside in different cache lines. The method permits a more complete analysis to be performed, thereby improving processor performance.

Cross-cache-line compounding algorithm for scism processors

An instruction processor system for decoding compound instructions created from a series of base instructions of a scalar machine, the processor generating a series of compound instructions with an instruction format text having appended control bits in the instruction format text enabling the execution of the compound instruction format text in said instruction processor with a compounding facility which fetches and decodes compound instructions which can be executed as compounded and single instructions by the arithmetic and logic units of the instruction processor while preserving intact the scalar execution of the base instructions of a scalar machine which were originally in storage. The system nullifies any execution of a member instruction unit of a compound instruction upon occurrence of possible conditions, such as branch, which would affect the correctness of recording results of execution of the member instruction unit portion based upon the interrelationship of member units of the compound instruction with other instructions. The resultant series of compounded instructions generally executes in a faster manner than the original format which is preserved due to the parallel nature of the compounded instruction stream which is executed.

System for executing scalar instructions in parallel based on control

A system for compounding instructions across cache line boundaries transfers an instruction line from a relatively slow memory to a instruction compounding unit if there is a miss for an instruction in that line in the instruction cache. At the same time the numerically preceding instruction in cache is transferred to the instruction compounding unit and instructions from the two lines are compounded. If a numerically preceding cache line has been compounded with a cache line that has been deleted and then replaced, compounding tags for the numerically preceding cache line are deleted.

An apparatus for generating microcode in a scalable compound instruction set machine operates in response to compounding information indicating that two or more adjacent instructions are to be executed in parallel. Separate and independent microcode is held in control store for each possible instruction in a group. Microcode sequences for each instruction of a group of instructions to be executed in parallel are merged in response to the compounding information into a single microinstruction sequence.

Microcode generation for a scalable compound instruction set machine

An instruction processing unit (IPU) and a storage array, a storage-to-instruction-processing-unit interface, including a hardware accelerator for cross-boundary storage access with a cross-boundary buffer for providing residual read and write data in support of high speed block concurrent accessing of multi-word operands of a computer system. A cross-boundary buffer (CBB) is used, coupled to a write rotating shifter, a write merger (WMERGE) and a write merge controller (WMCTL) which is coupled for an input to said control register (CREG) for sequencing data transmitted on the data bus for merger with data contained in the cross-boundary buffer (CBB) by the write merger before it is latched in a data bus out register, and for simultaneously also latching the data in the cross-boundary buffer (CBB), and for writing data from the data bus out register into the storage array in the next clock cycle of the instruction processor at the doubleword address addressed. The cross-boundary buffer (CCB) is also coupled to a read rotating shifter (RROTATE), a read merger (RMERGE) and a read merge controller which responds to control instruction sequencing. The storage-to-instruction-processing-unit interface operates on multiple words, with residues from a second and subsequent accesses allowing continuation of the accessing process beyond two memory words. The hardware can repeat a second microword until an operand of arbitrary length is transferred. The interface permits efficient data transfer to be interrupted and resumed at a desired point, for efficient execution of Load Multiple and Store Multiple operations.

Growing community of inventors

Endwell, NY, United States of America

Thomas L Jeremiah