At very least having a big fat decoder costs you power and space, measuring the tradeoff in raw decoder performance is probably dependent on being an engineer working on it at Intel or AMD (or Apple). At very least, Zen 3 is a cutting edge microarchitecture on a cutting edge node with lots of transistors to play with, and it still decodes only half as many instructions per cycle as M1 - an unbelievably wide CPU, which is probably the way forward.

The "x86 tax" (if it exists, I guess) is usually estimated at somewhere on the order of 5% to 10% - which is a lot of money at scale but probably not enough for a total rethink.