Redundancy detection: A comparison of lexical and value based approaches

The methods for redundancy detection can be classified as lexical approaches and value based approaches. Lexical approaches discover redundancy among lexically identical expressions. Value based approaches discovers redundancy among expressions computing the same value. This paper studies the practical merits of these approaches by implementing both techniques in the LLVM Compiler framework and doing experimental comparison with the SPEC CPU 2006 benchmarks. As a lexical method, the classical Common subexpression elimination was implemented. The Simple Algorithm for Global Value Numbering by Saleena and Paleri, was chosen for the implementation of value based approach. Results with SPEC CPU 2006 benchmarks show that in a practical compiler the value based method discovers more redundancies than lexical method.