[Mlir-commits] [flang] [mlir] [flang][OpenMP] Implement HAS_DEVICE_ADDR clause (PR #128568)

[llvmlistbot at llvm.org]

agozillon wrote:

> This LGTM! Thank you for implementing this, and adding the always map type to the descriptor like it is downstream (or coming up with the same thought I did, and adding more credence to it!), will cut down my other PR set a bit :-)
> 
> However, worth mentioning it does in certain cases, cause issues (primarily some cases in UMT) for map, which I'll try to address in some future PRs in some manner, but not a problem for you to worry about! I would however ask you to add one or two runtime tests to the offload/test/offloading/fortran "suite", mainly just so that I am aware if any changes I (or anyone else make) break something runtime wise!
> 
> P.S. please do wait on approval from Sergio as well!

And as an aside to this, I have a feeling as it is just now, the always map type may break a few tests in the offloading suite (check-offload), I recollect having to implement this yet to land code blob for it: https://github.com/llvm/llvm-project/pull/119588 where we "cut around" the data pointer, so as not to map over it when remapping the descriptor. As unless we explicitly specify always, we should never enforce that mapping on the address portion or we'll cause problems in a lot of more complicated map cases! And even that isn't quite perfect, ran into a few edge cases in UMT, I could try and carve out the relevant bits and upstream it if you run into issues with any tests in check-offload, if that's helpful at all. P.S. the check-offload fortran tests are terrible in the sense that they don't always run depending on configuration, so do try to intentionally fail one to check if they're running!

https://github.com/llvm/llvm-project/pull/128568