Predicting Dynamic Properties of Heap Allocations using Neural Networks Trained on Static Code: An Intellectual Abstract

Memory allocators and runtime systems can leverage dynamic properties of heap allocations – such as object lifetimes, hotness or access correlations – to improve performance and resource consumption. A significant amount of work has focused on approaches that collect this information in performance profiles and then use it in new memory allocator or runtime designs, both offline (e.g., in ahead-of-time compilers) and online (e.g., in JIT compilers). This is a special instance of profile-guided optimization.