1) use a worklist approach => better spreading of search work
2) use idea thread count for number of runnables => higher parallelism
3) avoid that the QRegularExpression is shared between threads (without this, e.g. 16 vs. 2 threads is no real difference)
4) deactive qmimedb atm => creates QRegularExpression internally => leads to same locking hell :/

with this, in a linux.git clone, searching for Linus takes:

1) with 2 threads: ~4.5 seconds
2) with 16 threads: ~0.5 seconds :P