Antonia Zhai

Compilers play a vital role in mapping high-level programming languages to underlying hardware and optimizing programs in the process. However, selecting suitable compiler optimization passes has long been known to be a complex task due to the vast number of optimizations and their interference. Additionally, generating accurate training data for such problems is costly, making data efficiency a crucial consideration for machine learning algorithms.

This group uses Data-efficient Compiler Optimization Selection (DeCOS), which performs a guided search in optimization spaces. To achieve data efficiency, DeCOS decomposes the straightforward machine-learning problem of optimizing a target program into producing a sequence of compiler optimization passes. In DeCOS, there are two separate mappings: one from the program features space to the policy space and another from the policy space to a subspace of the optimization space. This decomposition significantly reduces the size of the machine-learning problem and enables better data efficiency. DeCOS further improves data efficiency by using a more compact and accurate code representation enhanced by profiling data, deploying a set of exploration strategies to promote more promising searches, and using simulator-generated profiling data to reduce the noise level within the training data. In LLVM, by applying DeCOS to hot functions in MiBench, the researchers are able to achieve an average of 13.9% performance improvement compared to that of -O3 within less than eight hours of compilation, in comparison to multiple days of effort reported by prior work. The group aims to demonstrate that the DeCOS model trained using specific benchmarks and hardware platforms can be ported to different benchmarks and hardware platforms.