Modern 3D engines usually contain some not-insignificant amount of assembly for performance reasons. It'd require almost completely retuning, probably, undoing years of ideas since ARM is an entirely different architecture.

Wouldn't hold your breath this early.

Modern 3D engines contain almost no assembly anymore these days, and for the rare bits that are still hand-coded assembly, platform-independent alternative codepaths always exist. We've been long past the point where raw CPU performance even matters for high-end games anymore, for some time already. Porting between e.g. PPC as used in consoles and x86 is business as usual, so I don't see why porting to ARM would be much more difficult.

The main problem right now is that ARM performance is simply not there yet. It might be somewhere in the future, but even that remains to be seen, it's not entirely clear how well ARM designs will keep scaling up as they start to approach midrange x86 chips. Right now, ARM is just starting to match Atom, which is dog-slow compared to Core i3/i5/i7, and basically useless for non-trivial gaming.

Most of the really low level stuff is probably in the graphics shaders, I think HLSL etc are capable of producing code that will run on most GPUs?

Assembly is mostly used to optimize rarely maintained vector/matrix SIMD operations and even then it's mostly wrapped in intrinsics.