Performance-Aware Code Size Optimization of Generic Functions through Automatic Implementation of Dynamic Dispatch

Sammanfattning: Monomorphization and dynamic dispatch are two common techniques for implementing polymorphism in statically typed programming languages. Function templates in C++ use the former technique to enable algorithms written as generic functions to be efficiently reused with multiple different data types by producing a separate function instantiation for each invocation that uses a unique permutation of argument types. This avoids the overhead of indirection associated with dynamic dispatch and allows the generated code of each instantiation to be optimized by the compiler for its specific concrete types, which typically yields great improvements in runtime performance over any dynamic approach. The disadvantage of this implementation, compared to the type-erased generics found in many other programming languages, is that careless over-use of templates with many different argument types can lead to an excessive amount of redundant code being generated for the same function. This increase in code size may increase the binary size of the final program and reduce the amount of useful code that can fit into the processor's instruction cache during execution, reducing code locality and thereby potentially reducing performance. Monomorphization can also increase compilation time due to the increase in generated code that needs to be compiled and optimized. This thesis presents a heuristic-based approach to generic programming that allows function templates to be automatically converted to use dynamic dispatch in scenarios where the resulting negative impact on runtime performance is predicted to be low. The thesis project includes the development of a proof of concept plugin for the Clang compiler frontend that can be used to compile existing C++ projects with the conversions applied. The design of a heuristic function for determining whether a given function template should use monomorphization or dynamic dispatch based on statically known metrics is proposed based on the results of an experiment. This heuristic is shown to achieve a small general improvement in program size across a set of open-source C++ projects when they are compiled using the plugin. The key findings from the experiment and from the development of the plugin are summarized with a general strategy for how the approach can be integrated into the design of future programming languages to promote more extensive use of generic programming in performance-sensitive code while avoiding regressions in program size and compilation time.