Source Matching and Rewriting for MLIR Using String-Based Automata

A typical compiler flow relies on a uni-directional sequence of translation/optimization steps that lower the program abstract representation, making it hard to preserve higher-level information across each transformation step. On other hand, modern ISA extensions and hardware accelerators can benefit from compiler’s ability detect raise idioms acceleration instructions or optimized library calls. Although recent works based Multi-Level IR (MLIR) have been proposed for code raising, they rely specialized languages, recompilation, in-depth dialect knowledge. This article presents Source Matching Rewriting (SMR), user-oriented source-code-based approach MLIR idiom matching rewriting does not require expert’s intervention. SMR uses two-phase automaton-based DAG-matching algorithm inspired by early work tree-pattern matching. First, Control-Dependency Graph (CDG) is matched against program’s CDG rule out fragments do control-flow structure similar desired idiom. Second, candidate previous phase their Data-Dependency Graphs (DDGs) constructed DDG. Experimental results show effectively match Fortran (FIR) C (CIL) programs while raising them as BLAS calls improve performance. Additional experiments also performance improvements when using enable replacement in areas like approximate computing acceleration.