Combining analyses, combining optimizations

Reviewer: Eddy Bevers

Many compilers nowadays use static analyses to generate optimized code. A problem that often occurs is how to decide in what order these analyses should be performed: the results of one analysis can be of use to another. Sometimes combining two analyses leads to better results than running the two separately, in any order. In this paper, the authors start from analyses that are described in the monotone analysis framework [1]. They propose a model for combining such frameworks and for reasoning about the properties of the combination, such as its monotonicity and its run time complexity. In particular, the structure of the combination gives insight into whether the combination yields strictly better results than the separate analyses. The ideas in this paper are sound and well presented. They are illustrated with plenty of examples. In particular, the combined framework for constant propagation, value numbering, and unreachable-code elimination analyses gives a good insight into the problem and shows the practical use of the work. This research paper is primarily intended for researchers in compiler technology, but is also of use to compiler developers. Some mathematical background, especially in domain theory, is required.