International Business Machines Corporation

Kin Shing Chan

Dwain Alan Hicks

Shih-Hsiung Stephen Tung

Michael John Mayfield

Hung Qui Le

Chiao-Mei Chuang

Alessandro Marchioro

Dung Q Nguyen

Sang Hoo Dhong

Pradip Bose

Peichun Peter Liu

Rajinder Paul Singh

Ravindra K Nair

Pei-Chun Peter Liu

Robert Eric Wasmuth

A method of retiring operations to a cache. Initially, a first operation is queued in a stack such as the store queue of a retire unit. The first operation is then copied, in a first transfer, to a latch referred to as the miss latch in response to a resource conflict that prevents the first operation from accessing the cache. The first operation is maintained in the stack for the duration of the resource conflict. When the resource conflict is resolved, the cache is accessed, in a first cache access, with the first operation from the stack. Preferably, the first operation is removed from the stack when the resource conflict is resolved and the first cache access is initiated. In the preferred embodiment, the first operation is maintained in the miss latch until the first cache access results in a cache hit. One embodiment of the invention further includes accessing the cache, in a first miss access, with the first operation from the miss latch in response to a cache miss that resulted from the first cache access. In a presently preferred embodiment, a second access is executed to access the cache with a second operation queued in the stack in response to a cache hit resulting from the first cache access. The first and second cache accesses preferably occur in consecutive cycles. Typically, the first and second operations are store operations that are queued in the stack in program order. In one embodiment the first operation is removed from the stack upon resolving of the resource conflict.

Efficient store machine in cache based microprocessor

A method and apparatus for increasing instruction level parallelism using a buffer pointer assignment scheme is implemented whereby rename buffers are assigned during dispatch even though the physical rename registers may not yet be available. These virtual rename buffers are assigned by a buffer pointer assignment table. A virtual bit implemented along with each of the physical rename registers is flipped when an instruction corresponding to the entry stored within a particular physical rename register is completed and the result written to the architected register. Thus, at dispatch time, rename registers are assigned as if there were more rename buffers than there existed physical rename registers.

Methods and apparatus for exploiting virtual buffers to increase instruction parallelism in a pipelined processor

A deallocation pipelining circuit for use in a cache memory subsystem. The pipelining circuit is configured to initiate a storeback buffer (SBB) transfer of first line data stored in a first line of a cache memory array if the deallocation pipelining circuit detects a cache miss signal corresponding to the first line and identifies the first line data as modified data. The deallocation pipelining circuit is configured to issue a storeback request signal to a bus interface unit after the completion of the SBB transfer. The circuit initiates a bus interface unit transfer of the first line data after receiving a data acknowledge signal from the bus interface unit. The pipelining circuit is still further configured to deallocate the first line of the cache memory after receiving a request acknowledge signal from the bus interface unit. This deallocation of the first line of the cache memory occurs regardless of a completion status of the bus interface unit transfer whereby a pending fill of the first cache line may proceed prior to completion of the bus interface unit transfer. In one embodiment, the storeback buffer includes first and second segments for storing first and second segment data respectively. In this embodiment, the deallocation pipelining circuit is able to detect the completion of the transfer of the first segment data during the bus interface unit transfer and preferably configured to initiate an SBB transfer of second line data from a second line in the cache memory array in response to the completion of the first segment data transfer. In this manner, the initiation of the second line SBB transfer precedes the completion of the first line bus interface unit transfer.

Pipelined cache memory deallocation and storeback

A cache has an array for holding data or instruction values, a buffer connected to the array, and means for accessing the buffer to retrieve a value for a processing unit. The accessing means uses wires having a pitch which is substantially equal to a wire pitch of the cache array. Multiplexers can be used with a plurality of such buffers to create a common output path. The cache can be interleaved, with the array being a first subarray, and the buffer being a first buffer, and further comprising a second subarray and a second buffer, wherein the first and second buffers separate the first and second subarrays. The invention can be applied to a store-back buffer as well as a reload buffer.

Integrated cache buffers

An interleaved data cache array which is divided into two subarrays is provided for utilization within a data processing system. Each subarray includes a plurality of cache lines wherein each cache line includes a selected block of data, a parity field, a content addressable field containing a portion of an effective address (ECAM) for the selected block of data, a second content addressable field contains a real address (RCAM) for the selected block of data and a data status field. Separate effective address ports (EA) and a real address port (RA) permit parallel access to the cache without conflict in separate subarrays and a subarray arbitration logic circuit is provided for attempted simultaneous access of a single subarray by both the effective address port (EA) and the real address port (RA). A normal word line is provided and activated by either the effective address port or the real address port through the subarray arbitration. An existing Real Address (RA) cache snoop port is used to check whether a pre-fetching stream's line access is a true cache hit or not. The snoop read access uses a (33-bit) real address to access the data cache without occupying a data port during testing of the pre-fetching stream hits. Therefore, the two Effective Address (EA) accesses and a RCAM snoop access can access the data cache simultaneously thereby increasing pre-fetching performance.

Method and system for pre-fetch cache interrogation using snoop port

Method and apparatus for reconstructing the address of the next instruction to be completed in a pipelined processor

In response to dispatching a condition register modifying instruction to an execution unit, a condition register rename buffer is associated with such a condition register modifying instruction. The instruction is then executed in the execution unit. Following the execution of the condition register modifying instruction, condition register data is set in the condition register rename buffer to reflect the result of such instruction execution. Additionally, an indicator is set to indicate the condition register data is valid. At the time for completing the condition register modifying instruction, the condition register data is transferred from the condition register rename buffer to the architected condition register, thereby permitting condition register modifying instructions to be dispatched, executed, and finished before the condition register is available to complete each condition register modifying instruction.

Method and system for buffering condition code data in a data processing

In response to reloading an instruction from main memory for storing in an instruction cache in a superscalar data processing system, a particular instruction category in which the instruction belongs is selected from multiple instruction categories. Types of data processing system resources required for instruction execution and a quantity of each type of resource required are determined. Thereafter, a plurality of decode bits are calculated, wherein the decode bits represent a particular instruction category in which the instruction belongs and the type and quantity of each data processing system resource required for execution of the instruction. Thereafter, the instruction and the predecode bits are stored in instruction cache. The predecode bits enable the dispatch unit to efficiently, and without fully decoding the instruction at dispatch time, select an execution unit for executing the instruction and determine if the data processing system resources required for execution of the instruction are available before the dispatch unit dispatches the instruction.

Methods and system for predecoding instructions in a superscalar data

A method and system are disclosed for tracking the allocation of resources within a processor having multiple execution units which support speculative execution of instructions. The processor includes a resource counter including a first counter and a second counter and a number of resources, wherein one or more of the resources are allocated to each of a number of instructions dispatched for execution to the execution units. In response to dispatching an instruction among the plurality of instructions to one of the execution units for execution, the first counter is incremented once for each of the resources allocated to the instruction, and if the instruction is a first instruction within a speculative execution path, the second counter is loaded with a value of the first counter prior to incrementing the first counter. In response to completion of a particular instruction among the number of instructions dispatched to one of the multiple execution units, the first and the second counters are decremented once for each resource allocated to the particular instruction. In response to a refutation of the speculative execution path, a value of the second counter is transferred to the first counter, such that the resource counter tracks a number of the plurality of resources allocated to the plurality of instructions.

Method and system for tracking resource allocation within a processor

Logic circuitry provides a fast resolution of conditional branch instructions in a high-performance superscalar processor. The logic circuitry facilities early (fast) resolution of a subset of conditional branches located within the first position of the primary instruction buffer within the processor enabling the overall branch processing logic to bypass history table-based prediction logic for such branches without crossing the cycle boundary. Thus, penalties associated with possible mispredictions for this subset of conditional branches are avoided.

Method and system for reducing average branch resolution time and

A fast adder/subtracter using a decoder and shifting function instead of conventional full-adders is disclosed. The circuit is optimized for the addition of multiple operands up to 4-5 binary bits in magnitude. Using this method a subtraction operation can be performed at no added cost with respect to addition (compared to the conventional method requiring complementing one of the operands). Addition and subtraction of multiple operands is implemented by simple multiple shift operations. The multiple shift operations can be implemented as a chain of series NMOS pulldown devices with a precharged load providing considerable speed advantage over conventional solutions. Fast overflow detection may be implemented by or-ing the higher order bits in the shifter.

Growing community of inventors

Austin, TX, United States of America

Kin Shing Chan