While profiling tst_bench_qobject, I found that the disconnect() call dominates the connect/disconnect tests (and we can't really separate them in the benchmark, because otherwise we'd be testing the performance of the connection singly-linked list inside QObject, and not mere connect()).

It turns out that disconnect() unconditionally normalizes the signature, assuming, apparently, that disconnect() performance doesn't matter much. It surely matters less than connect() performance, but it doesn't not matter.

So, figure out a way to introduce a connect()-style fast-path that doesn't normalize.