Avoid address space casts.

[maleadt]

This PR changes how we work with and store address-space pointers: instead of casting them to and from generic pointers, directly work with their in-address space representation. That means we don't have to addrspacecast anymore, just inttoptr ... addrspace(...)* directly.

For WMMA, this requires 1.5's Base.AddrSpacePtr, because otherwise we select intrinsics for the generic case which leads to illegal memory accesses (cc @thomasfaingnaert).

Fixes #641.

[thomasfaingnaert]

Will this completely eliminate all addrspacecasts in kernels? LLVM doesn't seem to handle them too well. In particular, I've noticed issues with the LSV and SeparateConstOffsetFromGEP passes.

For vectorisation in #629, I use the CUDA C++ approach of casting to float4* before loading/storing, which I'm not a big fan of. I suspect that, with this change, I can just call unsafe_load and unsafe_store! directly, so long as I use the correct alignment. The LSV pass will then handle the rest.

(Also, since AddrSpacePtr was backported to 1.4.1, the WMMA tests should still pass for non-debug 1.4.1 I believe, so 1.5 is too strict a requirement)

[maleadt]

Will this completely eliminate all addrspacecasts in kernels?

It should, yes.

(Also, since AddrSpacePtr was backported to 1.4.1, the WMMA tests should still pass for non-debug 1.4.1 I believe, so 1.5 is too strict a requirement)

Ah yes, good call, it's only the return value that's problematic on 1.4.