Profile-Guided Optimization (PGO) for fuzzing optimization

[zamazan4ik]

Is your feature request related to a problem? Please describe.
An idea of how to possibly improve fuzzing performance.

Describe the solution you'd like
I am doing research on how Profile-Guided Optimization (PGO) helps with improving the performance of different software (link).

One of the interesting areas to me is optimizing fuzzing. Did anyone try to apply PGO on fuzzing? E.g. try to speed up the generation of new inputs (and correspondingly generate more corpuses per CPU time quantity) or something similar? I didn't see before such PGO applications. I want to highlight that I am interested at first in optimizing fuzzing codebase itself, not fuzzed targets (because target loads are too different and usually they easily can be optimized with PGO - see my results).

If you have any ideas, thoughts and/or existing numbers - would be awesome if you can share them. Thanks in advance! Maybe some already existing research from FuzzBench - who knows...

[andreafioraldi]

Hi, PGO require special instrumentation and sometimes tedious compiler setups, we never tried it.
On the other hand, AFL++ is optimized to reduce allocations cost and the map walking cost (the second most important bottleneck in a fuzzer, after the target execution) by hand using perf.