[Openmp-commits] [llvm] [openmp] [OpenMP] Change build of OpenMP device runtime to be a separate runtime (PR #136729)

[Joseph Huber via Openmp-commits]

jhuber6 wrote:

> > As I understand, we already have a pretty strong tendency toward the former. We have right now `flang-rt`, `compiler-rt`, `libclc`, and `openmp`.
> 
> My understanding (which might be incorrect), is that `flang-rt` and `compiler-ft` are host-only libraries, `libclc` is device-only, and `openmp` has both host and device components with the location of the device-only component being the crux of this discussion. A policy of using top level directories for host-only RTs and the host portions of RTs that span host/device, and placing device-only RT libraries under `offload` makes sense to me. However...

I don't really like to make a distinction between 'host' and 'device' here. As shown by the `libc` project, we should be able to treat the GPU as just another target. OpenMP is a little special here because it does enforce different semantics on the host vs. device, but everything else is just some flavor of compiling some utility functions for that target. Wasn't OpenCL designed with execution on CPUs in mind as well? It's probably easier to think of just having some utility library that works correctly w/ cross-compiling.

https://github.com/llvm/llvm-project/pull/136729