Global Program Optimization: Register Allocation of Static Scalar Objects

Performing certain optimizations on whole programs offers huge potential for code improvement. To exploit this possibility, we implemented a framework that allows sophisticated analysis and transformation of machine code, during or after linkage. This paper presents one of the many applications of our framework: the allocation of global scalar variables into registers. We present the algorithms that select the variables and remove the load and store instructions associated with them. The simulations of the unscheduled code of our benchmarks show that up to 10% of the execution cycles is saved by applying this technique.