The last couple of commits have seen a ~2x regression on NanoSoldier in min wall time for the string join benchmark, graph. Locally, I actually see an improvement, but given the regression has persisted for a few runs I'm inclined to think it's real.