Function inlining versus function cloning

Abstract

Function inlining is a powerful high level optimization which eliminates call cost and increases the chances of other optimizations taking effect due to the breaking down of the call boundaries. Goal-directed function cloning on the other hand strives to create specialized copies of functions to aid interprocedural optimizations. Both of these program transformations create additional code. In the traditional optimizers the problem of choosing which functions to inline is modelled as a 0--1 knapsack problem with a given budget. Function cloning is also treated likewise. However, the usual modelling considers these two as separate knapsack problems with their own respective budgets for code growth. This creates problems as independent budgets can give rise to excessive cloning or inlining and throttling opportunities for one or the other. To circumvent this problem we propose to formulate the inlining and cloning site selection problem as a unified multiple-choice knapsack problem. Multiple-choice knapsack is a well-studied problem in the domain of linear programming. In this paper we would look into a greedy algorithm to solve such knapsack problems and its application to the problem at hand.