使用离线profile提升JavaScript程序的类型可预测性

发布时间：2019-01-10 18:12

【摘要】：主流的JavaScript引擎(如V8)都使用了类型特化技术来优化程序性能,这种技术需要在运行时预测各个程序点的变量的类型集合。但JavaScript语言的动态性会降低程序的类型可预测性,使得引擎对类型预测失败(导致优化代码发生Deoptimization,即脱优化)或者产生较差的预测结果(导致生成低质量的代码),对程序性能造成负面影响。本文通过分析Octane、Kraken和SunSpider测试集中的程序,发现了提高程序类型可预测性的策略,并据此设计实现了基于离线profile的优化方案。本文的主要贡献如下:(1)分析这些测试集中的应用程序在运行过程中的行为,提出通过提高类型关联度来提升类型可预测性的策略。本文从48个应用程序中总结出了 6种会导致Deoptimization的模式,并分析了它们和引擎类型系统的关联。分析结果表明,可以通过增加类型之间的联系(称为类型关联度)来提升类型可预测性。(2)根据引擎为每个属性访问点的对象所预测的类型集合,分析每个集合中类型之间的关联,提出通过调整对象布局来提高类型关联度的策略。通过分析预测的类型集合大小大于1的属性访问点,以被访问属性为桥梁,将这些类型之间的关系分为4类,从而提出通过调整对象布局来增加类型关联度的策略。(3)提出了基于离线Profile的方案来调整对象布局,并在V8中进行了实现。方案分成三个阶段:插桩、Profiling和优化。插桩阶段会在程序中根据(2)中所提到的4种特点,识别程序点并安插收集变量类型信息的代码;Profiling阶段执行被插桩的代码收集这些程序点中的类型信息,并进一步生成对属性布局的调整方案(即属性调整参照)。优化阶段会在启动时读取属性调整参照信息,再生成调整对象布局的优化代码以便增加类型之间的关联度,提高类型可预测性。实验显示,本文的方案在所有测试集上共减少22%的Deoptimization,而且运行速度提升最高可达6%。本文提出了通过增加类型关联度来提高程序类型可预测性的策略,并提出基于离线profile的方案来调整对象布局,以增加类型关联度。此外,本文对常见代码模式的分析总结,可供JavaScript程序员和语言引擎优化工作者参考。
[Abstract]:Mainstream JavaScript engines, such as V8, use type-specific techniques to optimize program performance, which requires predicting the type set of variables at run time for each program point. But the dynamic nature of the JavaScript language reduces the type predictability of the program, causing the engine to fail to predict the type (causing the optimization code to be Deoptimization, or de-optimized) or to produce poor prediction results (resulting in the generation of low-quality code). It can have a negative effect on the procedure. Based on the analysis of programs in Octane,Kraken and SunSpider test sets, a strategy to improve the predictability of program types is found, and an optimization scheme based on offline profile is designed and implemented. The main contributions of this paper are as follows: (1) analyze the behavior of the applications in these test sets during the running process and propose a strategy to improve the predictability of the types by improving the correlation degree of the types. This paper summarizes six patterns that lead to Deoptimization from 48 applications and analyzes their association with engine type systems. The analysis results show that type predictability can be enhanced by increasing the relationship between types (called type correlation). (2) based on the type set predicted by the engine for each attribute access point object, This paper analyzes the relationship between types in each set and proposes a strategy to improve the correlation of types by adjusting the layout of objects. By analyzing the attribute access points whose size of the predicted type set is greater than 1, the relationship between these types is divided into four categories, taking the visited attributes as the bridge. Thus, the strategy of increasing type correlation by adjusting object layout is put forward. (3) A scheme based on offline Profile is proposed to adjust object layout, which is implemented in V8. The scheme is divided into three stages: pile insertion, Profiling and optimization. According to the four characteristics mentioned in (2), the program points will be identified and the code of collecting variable type information will be inserted in the program. The Profiling phase executes the code of the inserted pile to collect the type information in these program points and further generate the adjustment scheme for the property layout (that is, the attribute adjustment reference). In the optimization phase, the attribute adjustment reference information is read at startup, and the optimization code is generated to adjust the layout of the object in order to increase the correlation between the types and improve the predictability of the types. The experimental results show that the proposed scheme reduces 22% of Deoptimization, on all test sets and increases the running speed by up to 6%. This paper proposes a strategy to improve the predictability of program types by increasing the type correlation degree, and proposes a scheme based on offline profile to adjust the object layout to increase the type correlation degree. In addition, the analysis of common code patterns can be used as a reference for JavaScript programmers and language engine optimizers.
【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP312.2

【相似文献】