In model-driven low-code development, developers obtain the initial system implementation by generating the source code from models and then modify the generated code for custom-ization. In the subsequent development, the models may evolve so the code must be re-generated. How to merge the modified code with the newly generated code is an important issue. Existing model-driven development tools simply discard the code changed by developers or preserve developers' code based on some special annotations manually appended by developers. This paper pro-poses MerGen, a smart code merger for the generated code. Mer-Gen relies on universal unique identifiers that are associated with the generated code entities (i.e., types, fields, and methods) to pair the parts to be merged. Then, MerGen computes the digest of an entity in the normalized form to automatically determine whether the entity has been changed. Finally, MerGen uses a two-way re-factoring-based merging algorithm to merge the semantic code changes, rather than directly merging the code textually. We im-plement a prototype tool for the Eclipse Modeling Framework (EMF) and conduct a case study to evaluate the feasibility and the effectiveness of MerGen. The study results show that MerGen can effectively merge the modified code with the newly generated code compared with the default code merger of the EMF.

Similarity-based model matching is the foundation of model versioning. It pairs model elements based on a distance metric (e.g., edit distance). Because it is expensive to calculate the distance between two elements, a similarity-based matcher usually suffers from performance issues when the model size increases. This paper proposes a hash-based approach to accelerate similarity-based model matching. Firstly, we design a novel similarity-preserving hash function that maps a model element to a 64-bit hash value. If two elements are similar, their hashes are also very close. Secondly, we propose a 3-layer index structure and a query algorithm to quickly filter out impossible candidates for the element to be matched based on their hashes. For the remaining candidates, we employ the classical similarity-based matching algorithm to determine the final matches. Our approach has been realized and integrated into EMF Compare. The evaluation results show that our hash function is effective to preserve the similarity between model elements and our matching approach reduces 16%--72% of time costs while assuring the matching results consistent with EMF Compare.