Interactive Composition of Compiler Optimizations

Conventional compilers provide limited external control over the optimizations they automatically apply to attain high performance. Consequently, these optimizations have become increasingly ineffective due to the difficulty of understanding the higher-level semantics of the user applications. This paper presents a framework that provides interactive fine-grained control of compiler optimizations to external users as part of an integrated program development environment. Through a source-level optimization specification language and a Graphical User Interface (GUI), users can interactively select regions within their source code as targets of optimization and then explicitly compose and configure how each optimization should be applied to maximize performance. The optimization specifications can then be downloaded and fed into a backend transformation engine, which empirically tunes the optimization configurations on varying architectures. When used to optimize a collection of matrix and stencil kernels, our framework was able to attain 1.84X/3.83X speedup on average compared with using icc/gcc alone.