A solution to the can or cannot problem of learning based compilation

Modern compilers explore various large and complex transformation spaces in an iterative manner in search for high performance for a given program. Machine learning techniques have recently been used by compilers to capture features of a given program and find out useful heuristics from their prior experience with similar programs. However, we point out a can/cannot pitfall for learning-based compilation, in that a compiler may not have sufficient experience to deal with arbitrary programs encountered. Its success relies heavily on the training examples chosen. To tackle this pitfall, we use reverse K-nearest neighbor (RKNN) algorithm to help a compiler to decide whether to use existing prior experience directly, or turn to launch an optimization space search for outlier programs instead. Preliminary experimental results are given to demonstrate its effectiveness.