The overhead appears to be neglectible. Using TSS had a similar overhead, with the disadvantages that it uses more storage (one cache per thread vs one cache), has less cache reuse, and that Windows/MSVC doesn't support automatic TSS cleanup, so it requires manual cleanup work (which is non-trivial). Note that the mutex approach may yield more overhead in multi-threaded applications (in case of contention). Currently, the mutex also locks during the whole nameprep sequence, which is not strictly necessary.