Re: It'll be interesting to see how this is handled on Android when the time comes...
Given the thread size already, my comment will almost certainly never get read, but here goes anyway.
Kids today... Sigh. This isn't about physical memory and never was, it's about logical address space.
My background is post-ARM Acorn since 1996 (the company from which ARM originated) and RISC OS (the OS for which the ARM processor was originally created, via Arthur). See http://www.riscosopen.org/ and the Raspberry Pi for where things are with that these days. I'm one of the founder members of the not-for-profit RISC OS Open Limited.
When the OS moved from 26-bit (http://en.wikipedia.org/wiki/26-bit) to 32-bit addressing, one of the big benefits was increased available logical address space for applications. The overall OS memory map could make much better use of the 32-bit space available and applications could have considerably larger logical address space allocated.
Because of a dubious memory allocation API called "dynamic areas", RISC OS can still easily suffer logical address space exhaustion even if there is plenty of free physical RAM, just by asking for the reservation of dynamic areas - which are contiguous areas of reserved addresses - with the potential to grow to an indefinite physical size. This means that a large chunk of logical space has to be reserved on the off-chance that the application actually does try to extend that space to the maximum permitted size (IIRC the "maximum size" value used to be "available RAM" until 32-bit machines arrived with larger memory capacities - a 512MB RAM machine could've otherwise exhausted all virtual space in just a handful of unbounded dynamic area allocations, so application authors were encouraged to always specify an upper limit and the OS introduced a much-lower-than-RAM limit for those applications which said "as big as possible").
Under a 64-bit CPU, you have a *vast* address space you can use for logical addressing, so the memory map becomes very flexible and all sorts of constraints and limitations that don't necessarily have anything to do with physical RAM become lifted. There are usually performance benefits from increased register availability as others have correctly already pointed out but, as already pointed out too, the cost is the increased storage requirements and potential impacts on caches and memory accesses.
In any event, while it might not be *necessary* to have 64-bit address space in something as small scale as a phone, it can certainly be *advantageous* and there's no need to have anywhere near 4GB of actual physical RAM installed in order for it to be useful. One could, for example, memory-map the entire flash storage device into logical space without needing any of the additional special tricks required with 32-bit addressing.