[CIR][CUDA] Add initial support for name-mangling CUDA declarations

[AdUhTkJm]

The name-mangling for normal CUDA device/host functions are typically the same as normal C++ functions.

However, special care is taken when -fgpu-rdc is specified, which requires __device__ function names to be unique across complication unit in order to avoid symbol collisions.

[bcardosolopes]

This is missing a testcase. See clang/test/CIR for examples

[AdUhTkJm]

Thank you. However, as there is no CUDA support yet, when I specify -fcuda-is-device, clang will crash after AST is parsed. In this case, how should I test the mangled name?

[bcardosolopes]

... as there is no CUDA support yet, when I specify -fcuda-is-device, clang will crash after AST is parsed. In this case, how should I test the mangled name?

One strategy is to first focus on getting an empty file to build with all the flags you need, once that lands you resume this PR (so that you can finally check for the mangled name output).

[AdUhTkJm]

I accidentally overwrote the branch of this PR, and I couldn't change it back. Sorry for that, I'll close this one and open up another PR when I'm ready.