Memory usage improvement using runtime alias detection

Tofte and Talpin proposed a novel method for memory management called region inference in typed, higher-order languages[7]. In their method, memory is composed of blocks called regions, and when to allocate and deallocate regions are determined automatically by the compiler using a type system, and source programs are translated into target programs with region annotations. Their method realizes fairly economical use of memory for many programs. However, there are some kinds of programs which are not friendly to region inference. When executed on region-based systems, these programs can cause large amount of memory leak and in the worst case, they cannot finish their execution because of memory shortage. Therefore, some people proposed methods to improve memory usage of Tofte/Talpin system and others proposed different region inference systems. In this paper, we propose another technique to improve memory usage of Tofte/Talpin system. Our technique adds some changes to Storage Mode Analysis, which is a succeeding phase of their system, and delays some of decisions till runtime as to whether or not it is possible to overwrite existing objects. Our method is especially useful for a program compiled separately, where we cannot see the contexts in which top-level funtions are called. We implemented this technique to MLKit and confirmed that the amount of memory used during execution is reduced for some programs.