Global optimization by suppression of partial redundancies

Abstract

The elimination of redundant computations and the moving of invariant computations out of loops are often done separately, with invariants moved outward loop by loop. We propose to do both at once and to move each expression directly to the entrance of the outermost loop in which it is invariant. This is done by solving a more general problem, i.e. the elimination of computations performed twice on a given execution path. Such computations are termed partially redundant. Moreover, the algorithm does not require any graphical information or restrictions on the shape of the program graph. Testing this algorithm has shown that its execution cost is nearly linear with the size of the program, and that it leads to a smaller optimizer that requires less execution time.

Basic-block graphs: Living dinosaurs?

“…Both of these characteristics hardly arise in practice. E.g., Morel and Renvoise report that they never observed more than 3 iterations in their experiments [20], while Dhamdhere reports a number of 5 [3,4]. Typical DFA-probtems requiring lattices with longer chains (than e.g.…”

An Improved Algorithm for Redundancy Detection Using Global Value Numbering

“…The additional argument ant_exp is the set of expressions that are anticipable at the join point. We assume that the set of anticipable expressions are precomputed at every program point using the classical anticipability analysis [9,10]. T is the set of all program variables …”

“…Most of the efficient GVN algorithms that followed Kildall concentrate on detecting equalities among variables and hence, are limited in their ability to identify value-based redundancies [3,7]. Rosen et al [2] is an attempt at using GVN for partial redundancy elimination (PRE) [8][9][10]. Briggs et al [11] give a comparison of the different techniques for value numbering with respect to their use in different kinds of optimizations.…”

A Fresh Look at PRE as a Maximum Flow Problem

“…PRE problems come in two flavours: classic PRE and speculative PRE. Classic PRE, as described in the seminal work [22], inserts a computation at a point only if the point is safe (or down-safe) for the computation, i.e., only if the computation is fully anticipatable at the point. On the other hand, speculative PRE may insert a computation at a point even if the computation is partially but not necessarily fully anticipatable at the point.…”