Time- and size-efficient supercompilation

Abstract: Intermediate structures such as lists and higher-order functions are very common in most styles of functional programming. While allowing the programmer to write clear and concise programs, the creation and destruction of these structures impose a run time overhead which is not negligible. Supercompilation algorithms is a family of program transformations that remove these intermediate structures in an automated fashion, thereby improving program performance.While there has been plenty of work on supercompilation algorithms that remove intermediate structures for languages with call-by-name semantics, no investigations have been performed for call-by-value languages. It has been suggested that existing call-by-name algorithms could be applied to call-by-value programs, possibly introducing termination in the program. This hides looping bugs from the programmer, and changes the behaviour of a program depending on whether it is optimized or not.We present positive supercompilation algorithms for higher-order call-by-value and callby-name languages that preserves termination properties of the programs they optimize. We prove the call-by-value algorithm correct and compare it to existing call-by-name transformations. Our results show that deforestation-like transformations are both possible and useful for call-by-value languages, with speedups up to an order of magnitude for certain benchmarks.We also suggest to speculatively supercompile expressions and discard the result if it turned out bad. To test this approach we implemented the call-by-name algorithm in GHC and performed measurements on the standard nofib benchmark suite. We manage to supercompile large parts of the imaginary and spectral parts of nofib in a matter of seconds while keeping the binary size increase below 5%.Our algorithms are particularly important in the context of embedded systems where resources are scarce. By both removing intermediate structures and performing program specialization the footprint of programs can shrink considerably without any manual intervention by the programmer.