Actionable Program Analyses for Improving Software Performance

Nowadays, we have greater expectations of software than ever before. This is followed by constant pressure to run the same program on smaller and cheaper machines. To meet this demand, the application’s performance has become the essential concern in software development. Unfortunately, many applications still suffer from performance issues: coding or design errors that lead to performance degradation. However, finding performance issues is a challenging task: there is limited knowledge on how performance issues are discovered and fixed in practice, and current performance profilers report only where resources are spent, but not where resources are wasted. The goal of this dissertation is to investigate actionable performance analyses that help developers optimize their software by applying relatively simple code changes. To understand causes and fixes of performance issues in real-world software, we first present an empirical study of 98 issues in popular JavaScript projects. The study illustrates the prevalence of simple and recurring optimization patterns that lead to significant performance improvements. Then, to help developers optimize their code, we propose two actionable performance analyses that suggest optimizations based on reordering opportunities and method inlining. In this work, we focus on optimizations with four key properties. First, the optimizations are effective, that is, the changes suggested by the analysis lead to statistically significant performance improvements. Second, the optimizations are exploitable, that is, they are easy to understand and apply. Third, the optimizations are recurring, that is, they are applicable across multiple projects. Fourth, the optimizations are out-of-reach for compilers, that is, compilers can not guarantee that a code transformation preserves the original semantics. To reliably detect optimization opportunities and measure their performance benefits, the code must be executed with sufficient test inputs. The last contribution complements state-of-the-art test generation techniques by proposing a novel automated approach for generating effective tests for higher-order functions. We implement our techniques in practical tools and evaluate their effectiveness on a set of popular software systems. The empirical evaluation demonstrates the potential of actionable analyses in improving software performance through relatively simple optimization opportunities