Title: Software Language Comprehension using a Program-Derived Semantics Graph

Abstract: Traditional code transformation structures, such as abstract syntax trees (ASTs), conteXtual flow graphs (XFGs), and more generally, compiler intermediate representations (IRs), may have limitations in extracting higher-order semantics from code. While work has already begun on higher-order semantics lifting (e.g., Aroma's simplified parse tree (SPT), verified lifting's lambda calculi, and Halide's intentional domain specific language (DSL)), research in this area is still immature. To continue to advance this research, we present the program-derived semantics graph, a new graphical structure to capture semantics of code. The PSG is designed to provide a single structure for capturing program semantics at multiple levels of abstraction. The PSG may be in a class of emerging structural representations that cannot be built from a traditional set of predefined rules and instead must be learned. In this paper, we describe the PSG and its fundamental structural differences compared to state-of-the-art structures. Although our exploration into the PSG is in its infancy, our early results and architectural analysis indicate it is a promising new research direction to automatically extract program semantics.