One of the ways that custom instruction set extensions can improve over software execution is through the use of hardware structures that have been optimized at the arithmetic level. Arithmetic hardware, in many cases, can be partitioned into networks of full-adders, separated by other logic that is better expressed using other types of logic gates. In this paper we present a novel logic synthesis technique that optimizes networks of full adders and is intended for use in the context of custom instruction set synthesis. Unlike earlier work (e.g., Three Greedy Approach , ) our approach does not require any prior knowledge about the functionality of the circuit. The proposed technique automatically infers the use of carry-save arithmetic, when appropriate, and suppresses its use when unfavorable. Our approach reduces the critical path delay through networks of full adders, when compared to the Three Greedy Approach, and in some cases, reduces the cell area as well.