Synthesizing certified code

Abstract

Code certification is a lightweight approach for formally demonstrating software quality. Its basic idea is to require code producers to provide formal proofs that their code satisfies certain quality properties. These proofs serve as certificates that can be checked independently. Since code certification uses the same underlying technology as program verification, it requires detailed annotations (e.g., loop invariants) to make the proofs possible. However, manually adding annotations to the code is time-consuming and error-prone. We address this problem by combining code certification with automatic program synthesis. Given a high-level specification, our approach simultaneously generates code and all annotations required to certify the generated code. We describe a certification extension of AutoBayes, a synthesis tool for automatically generating data analysis programs. Based on built-in domain knowledge, proof annotations are added and used to generate proof obligations that are discharged by the automated theorem prover E-SETHEO. We demonstrate our approach by certifying operator- and memory-safety on a data-classification program. For this program, our approach was faster and more precise than PolySpace, a commercial static analysis tool.