Dynamic taint analysis is a powerful technique to detect memory corruption attacks. However, with typical overheads of an order of magnitude, current implementations are not suitable for most production systems. The research question we address in this paper is whether the slow-down is a fundamental speed barrier, or an artifact of bolting information flow tracking on emulators really not designed for it. In other words, we designed a new type of emulator from scratch with the goal of removing superfluous instructions to propagate taint. The results are very promising. The emulator, known as Minemu, incurs a slowdown of 1.5x-3x for real and complex applications and 2.4x for SPEC INT2006, while tracking taint at byte level granularity. Minemu's performance is significantly better than that of existing systems, despite the fact that we have not applied some of their optimizations yet. We believe that the new design may be suitable for certain classes of applications in production systems.