09-11-2013, 06:08 AM
[Unknown], for once you're actually at least partly wrong 
http://www.arm.com/files/downloads/ARMv8...ecture.pdf
ARM64 doubles the amount of GPR registers, which means that we can probably almost entirely ditch the regalloc and use a static mapping instead.
It has an improved NEON unit, also with more registers.
In addition, we can do the same memory mirror trick we do on x86-64, saving one instruction per memory access.
We will have to write a separate new JIT though as ARM64 is quite different from ARM32.
http://www.arm.com/files/downloads/ARMv8...ecture.pdf
ARM64 doubles the amount of GPR registers, which means that we can probably almost entirely ditch the regalloc and use a static mapping instead.
It has an improved NEON unit, also with more registers.
In addition, we can do the same memory mirror trick we do on x86-64, saving one instruction per memory access.
We will have to write a separate new JIT though as ARM64 is quite different from ARM32.