Register Reverse Rematerialization

Reversible computing could be in more or less long term mandatory for minimizing heat dissipation inherent to computing. It aims at keeping all information on input and intermediate values available at any step of the computation. Rematerialization in register allocation amounts to recomputing values instead of spilling them in memory when registers run out. In this paper we detail a heuristic algorithm for performing reverse register materialization and we use the high memory demanding LQCD (Lattice Quantum ChromoDynamics) application to demonstrate that important gains of up to 33% on register pressure can be obtained. This in turn enables an increase in Instruction-Level Parallelism and Thread-Level Parallelism. We demonstrate a 16.8% (statically timed) gain over a basic LQCD computation. Basic ideas of the algorithm and experimental results were already presented in a poster of another conference.