OpenIoT & ELC Europe 2016

Feedback Driven Optimisation (FDO), also known as Profile Guided Optimisation (PGO) is a well known code optimisation technique, employed by compilers since mid XX century, yet not widely used in the wild these days. It relies on providing runtime-captured information about code execution (eg. "branch taken or not?") during next code compilation, improving quality of decisions made by compiler heuristics.

To be fair, there were good reasons for its demise which I hope to discuss, mainly time and complexity overhead and deployment difficulties, but there is some hope on the horizon, coming with new approach, called AutoFDO and originating at Google, based on statistical profiling (namely Linux perf + extra tools) and source code level attribution. I'll discuss existing support for it available in mainline GCC and LLVM and give examples of real-life, successful deployments.