NYI: Aggressive Scheduling for Instructor-Level Processors

本课题研究指令级并行处理器（ILP）的主动指令调度技术。ILP处理器，如超标量和VLIW处理器，在通用和科学计算机系统中都得到了有效的应用。对于通用系统，分支指令的高频率需要积极的全局调度技术来支持推测性和有条件执行的指令。对于科学系统，循环调度技术，如软件流水线已经发展到积极调度内循环。本项目探索的四个领域是：(1)研究与软件流水线技术和模调度相关的问题。这包括调度while循环，处理具有跨国依赖关系的循环，以及消除统一的起始间隔限制；(2)形式化一个操作级中间程序表示，允许应用众所周知的局部调度技术来实现全局调度代码。该方法已成功地应用于模调度，并推广到无环调度；(3)开发一种任务粒度调整方法，在概念上类似于向量化，用于基于超标量/VLIW的多处理器系统中的软件流水线；(4)通过研究与调度多路分支和部分预测体系结构相关的硬件和编译器设计复杂性，分析现有处理器超标量实现的控制体系结构。

In this project, aggressive instruction scheduling techniques for instruction-level parallel (ILP) processors are investigated. ILP processors, such as superscalar and VLIW processors, are used effectively in both general purpose and scientific computer systems. For general purpose systems, the high frequency of branch instructions creates the need for aggressive global scheduling techniques that support speculatively and conditionally executed instructions. For scientific systems, cyclic scheduling techniques such as software pipelining have been developed to aggressively schedule inner loops. The four areas that are explored in this project are: (1) Investigate issues associated with the software pipelining technique and Modulo scheduling. These include scheduling while loops, handling loops with cross- iter\-ation dependencies and removing the uniform initiation interval restriction; (2) Formalize an operation-level intermediate program representation that allows the application of well-known local scheduling techniques to achieve globally scheduled code. This approach has been successfully applied to Modulo Scheduling and is extended to acyclic scheduling; (3) Develop a task granularity adjustment methodology, similar in concept to vectorization, for software pipelining in a superscalar/VLIW based multiple processor system; and (4) Analyze the control architectures for superscalar implementations of existing processors by studying the hardware and compiler design complexities associated with scheduling multi-way branching and partial predication architectures.