Fine-Grain Register Allocation Based on a Global Spill Costs Analysis

Abstract

A graph-coloring approach is widely used for register allocation, but its efficiency is limited because its formulation is too abstracted to use information about program context. This paper proposes a new register allocation technique that improves the efficiency by using information about the flow of variable references of a program. In the new approach, register allocation is performed at every reference of a variable in the order of the variable reference flow. For each reference, the costs of various possible register allocations are estimated by tracing a possible instruction sequence resulting from register allocations. A cost model is formulated to reduce the scope of the trace. Experimental results show that the proposed approach reduces spill code by an average of 34.3% and 17.8% in 8 benchmarks when compared to the Briggs’ allocator and the interference region spilling allocator, respectively.