Add seeds.
I think we're reaching the point where it is hard to increase fuzzing coverage significantly via fuzzing along, so to reach significant additional coverage going forward I think we'll have to add more harnesses to master.
Review welcome for the following fuzzing harnesses awaiting review: https://github.com/bitcoin/bitcoin/pulls/practicalswift :)
Do you think there's any benefit to running a minimization tool like afl-cmin or afl-tmin on the seeds? It can give faster startup times if there are a ton of redundant seeds, though I wonder if there's case to be made for a huge pool of diverse inputs.
For example, running afl-cmin on the addrman_deserialize seeds trims them from 714 files to 253 files.
@Crypt-iQ Yes, corpus minimization is good! The corpus is trimmed using libFuzzer's -use_value_profile=1 -merge=1 both locally before I submit the PR, and sometimes also before merge when MarcoFalke runs the merge script.
That means that only coverage increasing inputs should be added (with some hopefully rare exceptions with non-deterministic harnesses where -merge=1 is not stable for obvious reasons) :)
Note that -use_value_profile=1 -merge=1 and -merge=1 will give different results when minimizing the corpus.
Ah I see, that explains why afl-cmin trimmed it so much! I wonder if one minimization algorithm has a benefit over the other... might be worth comparing
@Crypt-iQ I believe libFuzzer's -merge=1 combined with -use_value_profile=1 is the current gold standard. It is what ClusterFuzz (and thus OSS-Fuzz) is using: see corpus_pruning_task.py.