As memory system performance becomes an increasingly dominant factor in overall system performance, it is important to optimize programs for memory related operations. This paper concerns static analysis to detect redundant memory operations and enable other compiler transformations to remove such redundant operations.

We present an extended global value numbering algorithm to detect redundant memory operations. The key of the new algorithm is a novel SSA-based representation for memory state which allows accurate reasoning about memory state. Using this representation, the algorithm can trace values through memory operations to detect equivalence in the same way that it traces them through register-based scalar operations. Thus it discovers both redundancy involving scalar values and redundancy involving memory operations. The redundancy relation detected by the algorithm can then be used by traditional redundancy elimination transformations to remove those redundant operations.

Experiments using a suite of realistic applications demonstrate the algorithm is powerful and fast. In practice, it has essentially linear time complexity.