Multi-objective quality metric for optimizing compilers

To obtain fast-executing computer code, optimizing compilers are used. The resulting code and the compilation process itself have several characteristics. The most important of them is the resulting code execution time, but the compilation time or the code size are also important for compiler users and should be minimized. The research goal of this paper is to provide a tool that allows to explicitly control compilation characteristics and allow smooth tuning between all of them. That is, the idea is to construct a mathematical multi-objective criterion, or a compilation quality functional, for comparing two compilers using a test suite. We show that the problem of choosing such a criterion for an arbitrary number of compilation characteristics has an exact solution that follows from user's evident expectations. The key findings are: (i) the functional must be multiplicative in its components, not additive; and (ii) it must be evaluated on the entire compilation task as a whole, not as a sum of independent per-procedure optima. We prove that an additive per-procedure functional does not reach the Pareto border for the whole task, while our multiplicative functional does. As an example of constructed compilation quality functional application we illustrate its minimum points used for a posteriori selection of the optimizing sequences.