I looked at issues, and one person filed issue 74 in June. It is an issue I have seen. I added a comment with another failure pathway.
I would like to take one of your CUDA examples and put it through your toolchain and get out LLVM IR. Could you please demonstrate how you think this should work?
None of the CUDA tests pass, by the way, when I do the check after the build. I can provide more details if I think there is a point -- like someone actively monitoring these issues.