Host/Device interfaces: integration header vs host compiler.

[bader]

See these two threads: https://reviews.llvm.org/D99190?id=333780#inline-935611 and https://reviews.llvm.org/D99190?id=333780#inline-935593.

In nutshell, @Naghasan suggests we consider using a host compiler to emit the data from the integration header. OpenMP-offload uses such approach for binding offloaded code with the host application. I assume CUDA and HIP compilers are doing something similar. I was going to check if we can apply this approach to SYCL and gather all pros and cons for both approaches.

Originally posted by @bader in #3471 (comment)

[bader]

Reply from @erichkeane

I don't think we can do it in host mode for 2 reasons:
1- the information about the actual kernels themselves is only guaranteed to be present during device compile
2- We don't necessarily control the host compiler. The SYCL design permits arbitrary host compilers, so doing this during integration header isn't particularly possible.