We currently virtually allocate 8GiB because odd corner cases potentially allow accesses to go above 4GiB. Those would always trap, so we need the memory to be mapped. This lowers the number of "fast" signaling memory we can allocate, it would be better to figure out which accesses could go above 4GiB, bounds check those, and all the other ones can stay entirely unchecked.
Note: maybe we want 4GiB + some small redzone for small offsets.
*** This bug has been marked as a duplicate of bug 170628 ***