It's hard to say and would probably vary significantly depending on the workload. 64-bit pointers could theoretically take twice as much space, whereas TLB management, since the MIPS has a software TLB, is also expensive because the TLB interrupt handler needs to run, which means more icache usage (and also has effects on the instruction pipeline, comparable to any other interrupt.)
According to this[1] paper the fact, that you effectively double your data cache with 32bit pointers leads to a performance increase of up to 30% in real world applications. I don't believe the comparatively rare context switches caused by TLB management interrupts cause a speed penalty that high. Hence I still don't understand why people want desperately go to 64bit. Just because some kernel programmers find it too stressful to bother with some minor memory management details?
