Partial Dead Code Elimination Using Extended Value Graph

Abstract

This paper presents an efficient and effective code optimization algorithm for eliminating partially dead assignments, which become redundant on execution of specific program paths. It is one of the most aggressive compiling techniques, including invariant code motion from loop bodies. Since the traditional techniques proposed to this optimization would produce the second-order effects such as sinking-sinking effects, they should be repeatedly applied to eliminate dead code completely, paying higher computation cost. Furthermore, there is a restriction that assignments sunk to a join point on ow of control must be lexically identical.

Our technique proposed here can eliminate possibly more dead assignments without the restriction at join points, using an explicit representation of data dependence relations within a program in a form of SSA (Static Single Assignment). Such representation called Extended Value Graph (EVG), shows the computationally equivalent structure among assignments before and after moving them on the control ow graph. We can get the final result directly by once application of this technique, because it can capture the second-order effects as the main effects, based on EVG.