[RFC] Address space preserving bitcasts

[vchuravy]

@maleadt worked towards supporting the CUDA/PTX backend in llvm in Julia. He presented his work in https://groups.google.com/forum/#!topic/julia-dev/doQUqQabC8A, because he will not be able to spend much time on it in the coming months.

I personally would like to see support for this in Julia and planned to work on something similar for the OpenCL/SPIR backend in llvm.

In order to integrate @maleadt work with upstream Julia I plan to submit PRs for separated concerns so that they can be properly discussed.

First up is address space preserving bitcasts.

@maleadt do you have time to explain shortly the rational behind this?

Cherry-picked from:
6c4fbcc and 855bb43

[maleadt]

In PTX land pointers live in different address spaces, depending on the type of memory you want to address. This is quote important, because the behaviour of each address space is vastly different (for example, thread-shared memory).

Julia does not care about address spaces, and makes everything live in the default one (0). Worse, the global pointer type objects (jl_pvalue_llvmt etc) are defined as living in addrspace 0. These objects are used when bitcasting jl_value_ts, regardless of the address space of those values. In older LLVM versions (pre 3.5, IIRC) this was silently ignored. Recent versions however croak when the address spaces within a bitcast don't match.

The preserving_bitcast function makes sure that, when casting pointer values, the address space is preserved by first casting the type object to the address space of the value, and then bitcasting the value with the proper type.

Note that this still lacks support for setting the address space of pointers from within Julia, which is a requirement to make this functionality useful. I was thinking of parameterizing the Ptr type to include the address space, but didn't get too far.

[Keno]

LGTM after rebase.

[vchuravy]

rebased

[vchuravy]

Superseded by #17193