GraalVM PGO optimizations for the non-executed code on the PGO training phase

[zamazan4ik]

What is the current optimization strategy in GraalVM PGO implementation for the code that wasn't executed during the profiling phase? Is it optimized for size or speed? Is there a reason to implement some alternative to -fprofile-partial-training flag from GCC (link)? LLVM PGO implementation has a similar request too.