Shared Learning Table for Load Value and Address Prediction

Solution Overview

Problem

Existing load value prediction techniques in computer processors face challenges due to high costs in circuit area and power consumption, as well as substantial drawbacks from mispredictions, which reduce the performance benefits of speculative execution.

Innovation Solution

A shared prediction learning table is used for both load value and load address prediction, reducing the area and power impact by storing a hash of the load value instead of the actual value and employing a probing load to retrieve the predicted value, thereby improving performance and reducing power consumption.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Speed

If traditional load value prediction is implemented, then load execution speed is improved, but circuit area and power consumption increase substantially

Engineering Contradiction:

Improveload execution speedVSAvoidcircuit area

Core Design Contradiction:

SpeedVSArea of stationary object

Solution Approach 1:

The patent combines load value prediction and load address prediction into a single shared learning table structure. The learning table stores both predicted load values and predicted load addresses using the same entry format, allowing both prediction functions to share the same hardware resources including the table storage, hash function units, and confidence counter logic, thereby reducing overall circuit area while maintaining both prediction capabilities

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The shared learning table entry is designed to serve multiple functions simultaneously. Each entry contains fields that can be used for either value prediction or address prediction depending on the prediction type being performed. The same table structure and associated logic circuitry handle both prediction types, making the hardware universal and reducing the need for separate dedicated circuits for each prediction function

Inventive Principle:

Principle #6Universality (Multi-functionality)

2Speed

If traditional load value prediction is implemented, then load execution speed is improved, but power consumption increases substantially

Engineering Contradiction:

Improveload execution speedVSAvoidpower consumption

Core Design Contradiction:

SpeedVSUse of energy by stationary object

Solution Approach 1:

The patent combines load value prediction and load address prediction into a single shared learning table structure. The learning table stores both predicted load values and predicted load addresses using the same entry format, allowing both prediction functions to share the same hardware resources including the table storage, hash function units, and confidence counter logic, thereby reducing overall circuit area while maintaining both prediction capabilities

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The shared learning table entry is designed to serve multiple functions simultaneously. Each entry contains fields that can be used for either value prediction or address prediction depending on the prediction type being performed. The same table structure and associated logic circuitry handle both prediction types, making the hardware universal and reducing the need for separate dedicated circuits for each prediction function

Inventive Principle:

Principle #6Universality (Multi-functionality)

3Productivity

If load value prediction is implemented, then speculative execution is improved, but misprediction costs increase

Engineering Contradiction:

Improvespeculative execution performanceVSAvoidmisprediction accuracy

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent implements feedback mechanisms where the actual load value and actual load address are compared against the predicted values after the load operation completes. The confidence counters for both value prediction and address prediction are updated based on whether the predictions were correct. This feedback allows the system to adaptively adjust prediction confidence levels and improve accuracy over time, reducing misprediction rates while maintaining high speculative execution performance

Inventive Principle:

Principle #23Feedback

Data Source

PatentUS20240362027A1Shared Learning Table for Load Value Prediction and Load Address Prediction

Publication Date: 2024.10.31 APPLE INC

AI summary

Techniques are disclosed relating to load value prediction. In some embodiments, a processor includes load address prediction circuitry and load value prediction circuitry. Training circuitry may train loads in a given entry, and may include a first entry configured to store first predicted load address information and a confidence indication of confidence that the first predicted load address information is correct and a second entry configured to store first predicted load value information and a confidence indication of confidence that the first predicted load value information is correct (note a given entry may be configured to load or value prediction at different times). Control circuitry may, in response to an entry in the training circuitry reaching a threshold level of confidence, allocate a corresponding entry in either the load value prediction circuitry or the load address prediction circuitry.