International Business Machines Corporation

Robert Alan Philhower

Michael K Gschwind

David James Shippy

Jonathan James DeMent

Ronald P Hall

Christopher Michael Abernathy

Raymond Cheung Yeung

Nicholas R Orzol

David S Levitan

Balaram Sinharoy

Ravi Nair

Bruce Martin Fleischer

Tai Anh Cao

Douglas Hooker Bradley

Ghavam G Shahidi

Alper Buyuktosunoglu

Pradip Bose

Hans Mikael Jacobson

Gi-Joon Nam

Richard James Eickemeyer

Paul Gerard Villarrubia

Gustavo Enrique Tellez

Christopher J Berry

Maarten Jakob Boersma

Lakshmi Narasimha Reddy

Michael H Wood

Christian Zoellin

Chen-Yong Cher

Kurt Alan Feiste

Michael Joseph Genden

Jentje Leenstra

Terry I Chappell

Barbara A Chappell

David J Widiger

Jinwook Jung

Gary Stephen Ditlow

Wolfram Sauer

Michael P Beakes

Barry Lee Dorfman

Vinod Narayanan

George Anthony Halasz

Wai Yin Wong

Aspects of the invention include receiving, by a processor, a plurality of instructions at an instruction pipeline. The processor can further determine an operand bit field size for each of the received plurality of instructions. The processor can further compare the operand bit field size of at least a subset of the received instructions to a predetermined threshold. The processor can further fuse at least two of the received instructions that have an operand bit field size that meets the predetermined threshold. The processor can further perform an execution stage within the instruction pipeline to execute the received instructions, including the fused instructions.

Dynamic fusion based on operand size

For each of a plurality of source-sink pairs, a corresponding interconnect layer is selected having a reach length which permits propagation of a required signal within a required sink cycle delay. For a first clock cycle, a movable region for a first latch is located as a first plurality of overlapped regions one reach length from a source and the required sink cycle delay number of reach lengths from each one of the sinks; and the first plurality of overlapped regions is represented as nodes on a first cycle level of a topology search graph. Analogous actions are carried out for a second clock cycle of the required sink cycle delay. A latch tree is created based on the topology search graph, wherein a required number of latches is minimized, and at each of the cycle levels, all sinks of source nodes selected at a previous level are covered.

Multi-cycle latch tree synthesis

A processor-implemented method is provided. The processor-implemented includes reading, by a processor, an instruction stream by fetching instructions from an instruction cache of the processor. The processor then executes a branch prediction operation based on a context of the instruction stream and an index when one of the instructions includes a branch instruction. The branch prediction operation output a prediction and a context. The processor then compares the context of the instruction stream and the context from the branch prediction operation to determine whether to execute a stop fetch.

Tagging target branch predictors with context with index modification and late stop fetch on tag mismatch

Aspects of the invention include a computer-implemented method for executing one or more instructions by a processing unit. The method includes receiving, by an instruction fetch unit (IFU), a request to fetch an instruction for execution, wherein the instruction includes an effective address (EA). The IFU can further access an instruction cache directory (I-directory) using the EA of the requested instruction to determine whether the EA of the requested instruction matches an EA stored in an associated instruction cache (I-cache). An instruction cache (I-cache) can output the requested instruction in response to or based at least in part on determining that the requested instruction EA matches an entry in the I-cache. A decode unit can decode the requested instruction output by the I-cache.

Effective address based instruction fetch unit for out of order processors

A computer processor includes a branch prediction unit that includes a local branch predictor and a global branch predictor. Managing power consumption in such a computer processor includes, for each of a plurality of branch instructions: performing, by the local branch predictor, a local branch prediction; performing, by each of the global branch predictors, a global branch prediction; determining to utilize the local branch prediction over the global branch predictions as a branch prediction for the branch instruction; incrementing a value of a counter; determining whether the value of the counter exceeds a predetermined threshold; and if the value of the counter exceeds the predetermined threshold, powering down at least one of the global branch predictors and configuring the branch prediction unit to bypass the powered down global branch predictor for branch predictions of subsequent branch instructions.

Power management of branch predictors in a computer processor

An information handling system employs a processor that includes a thread priority controller. An issue unit in the processor sends branch issue information to the thread priority controller when a branch instruction of an instruction thread issues. In one embodiment, if the branch issue information indicates low confidence in a branch prediction for the branch instruction, the thread priority controller speculatively increases or boosts the priority of the instruction thread containing this low confidence branch instruction. In the manner, should a branch redirect actually occur due to a mispredict, a fetcher is ready to access a redirect address in a memory array sooner than would otherwise be possible.

Method and apparatus for thread priority control in a multi-threaded processor based upon branch issue information including branch confidence information

An information handling system includes a processor that throttles an instruction fetcher whenever a group of instructions in a branch instruction queue together exhibits a confidence in the accuracy of branch predictions of branch instructions therein that is less than a first predetermined threshold confidence threshold. In one embodiment, the processor includes a fetch throttle controller that inhibits fetch throttling by the instruction fetcher when confidence in the accuracy of a branch prediction for a particular currently issued branch instruction exhibits less than a second predetermined threshold confidence threshold.

Method and apparatus for inhibiting fetch throttling when a processor encounters a low confidence branch instruction in an information handling system

An information handling system includes a processor with an array power management controller. The array power management controller gates off a memory array, such as a cache, to conserve power whenever a group of instructions in a branch instruction queue together as a group exhibits a confidence in the accuracy of branch predictions of branch instructions therein that is less than a first predetermined threshold confidence threshold. In one embodiment of the information handling system, the array power management controller speculatively inhibits the gating off of the memory array when confidence in the accuracy of a branch prediction for a particular currently issued branch instruction exhibits less than a second predetermined threshold confidence threshold. In this manner, the array power management controller again allows access to the memory array in the event a branch redirect is likely.

Method and apparatus for controlling memory array gating when a processor executes a low confidence branch instruction in an information handling system

Tracking the order of issued instructions using a counter is presented. In one embodiment, a saturating, decrementing counter is used. The counter is initialized to a value that corresponds to the processor's commit point. Instructions are issued from a first issue queue to one or more execution units and one or more second issue queues. After being issued by the first issue queue, the counter associated with each instruction is decremented during each instruction cycle until the instruction is executed by one of the execution units. Once the counter reaches zero it will be completed by the execution unit. If a flush condition occurs, instructions with counters equal to zero are maintained (i.e., not flushed or invalidated), while other instructions in the pipeline are invalidated based upon their counter values.

Time-of-life counter for handling instruction flushes from a queue

A method, system, and computer program product are provided, for maintaining a path history register of register indirect branches. A set of bits is generated based on a set of target address bits using a hit selection and/or a hash function operation, and the generated set of bits is inserted into a path history register by shifting bits in the path history register and/or applying a hash operation, information corresponding to prior history is removed from the path history register, using a shift out operation and/or a hash operation. The path, history register is used to maintain a recent target, table and generate register-indirect branch target address predictions based on path history correlation between register-indirect branches captured by the path history register.

Method, system, and computer program product for path-correlated indirect address predictions

A method of obtaining data, comprising at least one sector, for use by at least a first thread wherein each processor cycle is allocated to at least one thread, includes the steps of: requesting data for at least a first thread; upon receipt of at least a first sector of the data, determining whether the at least first sector is aligned with the at least first thread, wherein a given sector is aligned with a given thread when a processor cycle in which the given sector will be written is allocated to the given thread; responsive to a determination that the at least first sector is aligned with the at least first thread, bypassing the at least first sector, wherein bypassing a sector comprises reading the sector while it is being written; and responsive to a determination that the at least first sector is not aligned with the at least first thread, delaying the writing of the at least first sector until the occurrence of a processor cycle allocated to the at least first thread by retaining the at least first sector in at least one alignment register, thereby permitting the at least first sector to be bypassed.

Alignment of cache fetch return data relative to a thread

Dynamic power management in a processor design is presented. A pipeline stage's stall detection logic detects a stall condition, and sends a signal to idle detection logic to gate off the pipeline's register clocks. The stall detection logic also monitors a downstream pipeline stage's stall condition, and instructs the idle detection logic to gate off the pipeline stage's registers when the downstream pipeline stage is in a stall condition as well. In addition, when the pipeline stage's stall detection logic detects a stall condition, either from the downstream pipeline stage or from its own pipeline units, the pipeline stage's stall detection logic informs an upstream pipeline stage to gate off its clocks and thus, conserve more power.

Dynamic power management in a processor design

An information handling system includes a processor that throttles the instruction fetcher whenever the inaccuracy, or lack of confidence, in branch predictions for branch instructions stored in a branch instruction queue exceeds a predetermined threshold confidence level of inaccuracy or error. In this manner, fetch operations slow down to conserve processor power when it is likely that the processor will mispredict the outcome of branch instructions. Fetch operations return to full speed when it is likely that the processor will correctly predict the outcome of branch instructions.

Method and apparatus for conserving power by throttling instruction fetching when a processor encounters low confidence branches in an information handling system

Growing community of inventors

Valley Cottage, NY, United States of America

Robert Alan Philhower