Reviews: Neural Code Comprehension: A Learnable Representation of Code Semantics

This paper proposes to use programming code embeddings for individual statements to be used in all kinds of program analysis tasks, such as classification of type of algorithms, placing of code on a heterogeneous cpu-gpu architecture and scheduling of programs on gpus. The novelty of the paper consists of the way the code embeddings are computed/trained. Instead of looking at code statements in high-level languages and following either data or control flow, the paper proposes to use statements at an intermediate-representation level (which are independent of the high-level language used) and take both data and control flow into account. As such, the proposed technique builds "contextual flow graphs" where the nodes are connected either through data or control flow edges. The nodes are variable or label identifiers.