Using the LLVM framework for static performanceprediction with embedding of intermediate representation

We consider a method for mapping programs to the space of vector embeddings to create of empirical estimation of programs’ performance at compilation time. This method is based on transformation series of the initial intermediate representation (IR) such as instrumentation with artificial instructions in a compiler’s pass, transformation of instrumented IR in a multidimensional vector with IR2Vec [1] and dimensionality reduction using t-SNE [2]. The D1 cache miss ratio is considered as performance metric. A heuristic separation criterion of programs with more or less cache miss ratio in 2D-space is given. The experimental validation of the criterion is performed in synthetic tests.