Support 39-bit virtual address space

[caizixian]

For most RISC-V boards and low-end aarch64 boards, the MMU might only support 39-bit virtual memory (such as sv39). Currently, MMTk assumes 48-bit virtual address space, which has been true for x86_64. MMTk should support 39-bit virtual address space, and performance comprise is acceptable.

[caizixian]

@steveblackburn as discussed, we should sit down together at some point and draft a plan.

[qinsoon]

@wenyuzhao is working on compressed pointer which uses 38 bit address space. Does that solve this issue as well?

[qinsoon]

#899 allows bindings to specify the address space to use. I assume we can close this issue now.

[caizixian]

Let's keep this open. There are a bunch of metadata stuff we need to fix as well, e.g., I had to hack the core here caizixian@ceb37f1

[qinsoon]

There is no concrete task in this issue we can work on. I will keep this issue open for a while, but I generally suggest opening new issues or PRs if you find any bug.

[caizixian]

I think this is the appropriate metaissue. The core currently doesn't support 39-bit virtual address space fully because the core still maps metadata above 39-bit, and will fail due to mmap failure on those machines.

[caizixian]

We might need to create a child issue for the potential side metadata refactoring required, but we should keep this metaissue open until the underlying problems are all resolved.