Addressing JavaScript JIT Engines Performance Quirks: A Crowdsourced Adaptive Compiler

JavaScript has long outpaced its original target applications, being used not only for coding complex web clients, but also web servers, game development and even desktop applications. The most appealing advantage of moving applications to JavaScript is its capability to run the same code in a large number of different devices. It is not surprising that many compilers target JavaScript as an intermediate language. However, writing optimizations and analyses passes for a compiler that emits JavaScript is challenging: a long time spent in optimizing the code in a certain way can be excellent for some browsers, but a futile effort for others. For example, we show that applying JavaScript code optimizations in a tablet with Windows 8 and Internet Explorer 11 increased performance by, on average, 5 times, while running in a desktop with Windows 7 and Firefox decreased performance by 20%. Such a scenario demands a radical new solution for the traditional compiler optimization flow. This paper proposes collecting web clients performance data to build a crowdsourced compiler flag suggestion system in the cloud that helps the compiler perform the appropriate optimizations for each client platform. Since this information comes from crowdsourcing rather than manual investigations, fruitless or harmful optimizations are automatically discarded. Our approach is based on live measurements done while clients use the application on real platforms, proposing a new paradigm on how optimizations are tested.