Languages that need tagged memory can implement this in software, and on current arch they will still run as fast as in a CPU that had that trick implemented in

Languages that need tagged memory can implement this in software, and on current arch they will still run as fast as in a CPU that had that trick implemented in hardware. More likely faster, because the software solution enables much greater flexibility (e.g. look at Javascript VMs, they have various ways to represent values, and that's a single language). Also because compilers can often optimize out all value-representation overheads (by unboxing etc.). Stuff implemented in the CPU decoder (or even in microcode) can never afford to make that level of optimization.