Differential testing solidity compiler through deep contract manipulation and mutation

Solidity, the language utilized for developing smart contracts, has been gaining increased importance in blockchain system. Ensuring bug-free of its accompanying language compiler, which converts the contract source codes into executables finally deployed on the blockchain, is thus of paramount importance. This study presents DeSCDT, a Deep learning-based Solidity Compiler Differential Testing approach, to explore possible defects in Solidity compiler. At the core lies a well-behaving deep contract generator following the Transformer architecture and learnt with diverse contract code. From an initial seed pool of contracts carefully picked through semantic encoding and clustering, the generator is capable of stably producing highly syntactic-valid and functional-rich smart contracts, with three meticulously formulated generation strategies and a set of mutation operations. Subsequently, in the meantime of compiling these generated contracts to trigger compiler crashes, a differential testing environment is set up to explore misoptimization bugs, by observing the inconsistencies between the outcomes and the aspects including opcode size of the optimized and non-optimized bytecodes. For the experiments, the syntactic validity and diversity of the contracts generated with DeSCDT, as well as its ability in discovering compiler defects, are investigated. The findings indicate that DeSCDT can effectively generate syntactically correct contracts with a pass rate of 90.8