[clang] [CUDA][HIP] Fix host/device context in concept (PR #67721)

[Yaxun Liu via cfe-commits]

================
@@ -176,3 +176,34 @@ Predefined Macros
    * - ``HIP_API_PER_THREAD_DEFAULT_STREAM``
      - Alias to ``__HIP_API_PER_THREAD_DEFAULT_STREAM__``. Deprecated.
 
+C++20 Concepts with HIP and CUDA
+--------------------------------
+
+In Clang, when working with HIP or CUDA, it's important to note that all constraints in C++20 concepts are assumed to be for the host side only. This behavior is consistent across both programming models, and developers should be aware of this assumption when writing code that utilizes C++20 concepts.
+
+Example:
+.. code-block:: c++
+
+   template <class T>
+   concept MyConcept = requires(T& obj) {
+     my_function(obj);  // Assumed to be a host-side requirement
----------------
yxsamliu wrote:

> If I understand correctly, normally a template is usable from either host or device (depending on whether it ends up calling any host-only or device-only function). 

No. template function in CUDA/HIP is subject to the same availability check regarding host/device attributes. e.g https://godbolt.org/z/ccTxhEhfo

constexpr template functions can be called by both host and device functions not because they are template functions, but because host device attributes are added to them implicitly. They are still subject to host/device availability checks.

https://github.com/llvm/llvm-project/pull/67721