Feedback Driven Optimization: Improving Program Performance Through Compilation

This talk explains feedback-directed and profile-guided optimization, including AutoFDO’s use of statistical profiling to guide compiler decisions. It discusses GCC and LLVM support, practical challenges, and deployment examples.

Syllabus

Introduction
Terminology
FDO Manual
Limit Instrumentation Based FDO
Example Program
Swap Optimization
Swap Optimization Results
Profile Optimization Results
Problems with FDO
Feedback Directed Optimization
Sampling Profile Quality
Non Perfect Condition
Where is it being used