[llvm] [OpenMP][Flang][Clang][Offload] Prevent shadow pointer write-back to null host pointers (PR #197106)

[via llvm-commits]

agozillon wrote:

So, this might not be the best fix for this issue, I am not the most runtime savvy, so I'm more than happy to take suggestions for other adjustments to solve this problem :-) Perhaps the UMT benchmark is even currently doing something not so legal, by breaking the maps of a larger object up like this (as I am fairly certain there's some wording to that affect in the OpenMP specification, but I'm unsure if it's applicable in this scenario). I will mention that this adjustment to the runtime has been part of our downstream branch for about a year now and it's not been flagged up as a problem by any of our testing infrastructure or client issues.

Please do feel free to add other reviewers if you think there's someone that would be good as an additional reviewer! 

https://github.com/llvm/llvm-project/pull/197106