Using combined profiling to decide when Thread Level Speculation is profitable

Thread Level Speculation (TLS) speculatively executes parts of a program in parallel. Statically determined may dependences between store-load pairs prevent the compiler from speculatively executing parts of programs (e.g loop iterations or functions). If a compiler can determine that the probability of a may dependence occurring at runtime is low, then it can use TLS to execute the loop in parallel. This research will develop a may dependence profiling framework that is able to capture the effect of different inputs on the dependence behaviour of the program during runtime, using a technique called Combined Profiling (CP) [1]. The dependence profiling will be made efficient using the output from static analysis. TLS code generation strategies will be implemented in a version of the LLVM compiler that will generate code for the hardware support for TLS in the IBM BlueGene/Q (BG/Q) machine.