Aggressive Aggregation: a New Paradigm for Program Optimization

TL;DR

This paper introduces aggressive aggregation, a novel program optimization paradigm that decomposes code into optimal structures for enhanced optimization, significantly improving performance especially in loop-unrolled code.

Contribution

It presents a new decomposition-based optimization approach using Algebraic Decision Diagrams and Expression DAGs, enabling powerful optimizations like loop unrolling and redundancy elimination.

Findings

Achieved over an order of magnitude performance improvement on Fibonacci program

Supports loop unrolling as a first-class optimization technique

Effectively optimizes large program fragments from multiple loop unrollings

Abstract

In this paper, we propose a new paradigm for program optimization which is based on aggressive aggregation, i.e., on a partial evaluation-based decomposition of acyclic program fragments into a pair of computationally optimal structures: an Algebraic Decision Diagram (ADD) to capture conditional branching and a parallel assignment that refers to an Expression DAG (ED) which realizes redundancy-free computation. The point of this decomposition into, in fact, side-effect-free component structures allows for powerful optimization that semantically comprise effects traditionally aimed at by SSA form transformation, code specialization, common subexpression elimination, and (partial) redundancy elimination. We illustrate our approach along an optimization of the well-known iterative Fibonacci program, which, typically, is considered to lack any optimization potential. The point here is that our technique supports loop unrolling as a first class optimization technique and is tailored to optimally aggregate large program fragments, especially those resulting from multiple loop unrollings. For the Fibonacci program, this results in a performance improvement beyond an order of magnitude.