Even Rosetta still gives up 10%+ efficiency compared to a native compilation of the same program. I’m not saying it’s not viable, but in a resource constrained (especially battery-constrained) device 10% is a lot.