Demand-driven Inlining Heuristics in a Region-based Optimizing Compiler for ILP Architectures

While scientists and engineers have been taking on the challenge of converting their applications into parallel programs, compilers are charged with identifying opportunities and transforming codes to exploit the available instruction-level parallelism (ILP) offered by today’s uniprocessors, even within a parallel computing environment. Region-based compilation repartitions a program into more desirable compilation units for optimization and ILP scheduling. With region-based compilation, the compiler can control problem size and complexity by controlling region size and contents, and expose interprocedural scheduling and optimization opportunities without interprocedural analysis or large function bodies. However, heuristics play a key role in determining when it is most beneficial to inline functions during region formation. This paper presents and experimentally compares a set of heuristics for demand-driven inlining performed during region formation in terms of static code growth, compile-time memory, and run-time performance.