[clang] [clang][CUDA] Allow `extern __shared__` on non-array types (PR #183879)

[Yaxun Liu via cfe-commits]

yxsamliu wrote:

Could you document this in either the CUDA guide or HIP documentation, and add a brief reference in the other? In non-RDC mode, `extern __shared__` scalars, structs, and fixed-size arrays get fixed shared storage, while unsized arrays use dynamic shared memory. The “dynamic shared memory overlay” wording seems misleading for the struct case.


https://github.com/llvm/llvm-project/pull/183879