[clang] [clang][AMDGPU] Don't define feature macros on host code (PR #83558)

[Yaxun Liu via cfe-commits]

yxsamliu wrote:

When clang does host compilation, it essentially makes an assumption that the generated IR for host does not depend on the assumed GPU arch, or, the generated IR may be affected by assumed GPU arch, but it won't affect the program output. This is true in most cases. For example, in the case of rocprim, it needs to see __AMDGCN_WAVEFRONT_SIZE  to be able to parse device functions in host compilation, but that does not affect what host IR is generated. That is the reason why we can assume a default GPU arch in host compilation.

I think a more practical approach to void __AMDGCN_WAVEFRONT_SIZE  being used in host functions is to define it as a clang constexpr builtin that is evaluated to an integer constant in device functions or variables, but will be diagnosed when used in host functions or variables. warpSize should also be defined as a device constexpr variable to avoid it being used in host code.

https://github.com/llvm/llvm-project/pull/83558