Something like the approach used in transmeta crusoe cores? Anyway, it would be quite interesting to investigate (if it's not done already) what is the main performance bottleneck in arm->x86 emulators and if any reasonably simple extension to the arm instruction set can result in more efficient x86 emulation