Very few of today's apps actually benefit from having more than 3GB of address space. Nothing on my desktop PC is using more than 300MB at the moment. All else being equal, 32 bits will be more efficient, particularly on portable devices which tend to have far less memory bandwidth.
Of course, the performance gains have really nothing to do with 64-bit addressing alone, but ARM took the opportunity to almost completely redesign the ISA, and A64 is in general a better and faster ISA than A32. So all else is not equal.
I actually find this 64-bit-has-better-ISA-than-32-bit situation annoying. I would like to have better ISA, without making pointers larger. There is x32(https://sites.google.com/site/x32abi/), will there be something similar for ARM?
mmapping files without blowing out your address space works a lot better in a 64-bit process (and system libraries love to mmap things like fonts, which can quickly take up a bunch of address space).
I just took a look at the address space usage on my Mac, and the spotlight indexer seems to be using about 1.5GB of address space (and 350MB of actual RAM, for some reason?), and a number of other built-in things (WindowServer, Apple80211Agent, Dashboard) are all north of the 500MB address space mark.
But I think most stuff that runs off of soldered-on DDR3L and flash RAM (e.g., apps on phone and tablet OSes) may be better off sticking with 32 bit addresses for another generation.
OK, yeah the fundamental number type of Javascript is the 64-bit floating point double so I could see how the ability of a JIT to emit code to handle 56 bit integers might be useful. But it would still be a tradeoff against the cost of doubling the size of every pointer.
Very few of today's apps actually benefit from having more than 3GB of address space. Nothing on my desktop PC is using more than 300MB at the moment. All else being equal, 32 bits will be more efficient, particularly on portable devices which tend to have far less memory bandwidth.