Optimization Principles
Traditional compilation optimization can only predict the execution behavior of programs through static program analysis and heuristic algorithms. By collecting program runtime information, PGO can accurately determine the cold, hot, and execution probability of code. In this way, PGO can efficiently optimize cold and hot partitioning, branch prediction, function rearrangement, register allocation, vectorization, and function inline, improving the cache hit ratio, branch hit ratio, and data parallelism, and reducing the pressure on the register.
The typical optimization principles are described as follows:
Hot/Cold Partitioning
Cold branches are removed to aggregate hot code and improve the cache hit ratio.
Function Rearrangement
Code section functions are rearranged to aggregate hotspot functions and reduce iTLB and iCache miss rates.
Branch Prediction
Branch sequence is adjusted to reduce the branch miss rate.
Function Inline
Feedback-based inline: global analysis, precise inline, optimized call stack, and better memory allocation.
Switch Optimization
Structure branches are adjusted to reduce jumps and the branch miss rate.
