[clang] [llvm] [AMDGPU] Add initial support for VGPR as memory (PR #205435)

[Matt Arsenault via llvm-commits]

================
@@ -3604,6 +3604,26 @@ An error will be given if:
   }];
 }
 
+def AMDGPUVGPRDocs : Documentation {
+  let Category = DocCatAMDGPUAttributes;
+  let Content = [{
+This attribute requests that a kernel-local variable be allocated in the
+"VGPR as memory" address space (``addrspace(13)``) on the AMDGPU target,
+so that accesses with statically known indices lower to vector register
+copies instead of scratch memory traffic.
+
+Clang supports the ``__attribute__((amdgpu_vgpr))`` or
+``[[clang::amdgpu_vgpr]]`` attribute in HIP/CUDA. It may only be applied to
+local variables declared in a ``__global__`` (kernel) function; applying it to
+a variable in a ``__device__`` or host function, or outside HIP/CUDA, is an
+error.
+
+Known limitation: the request is only honored with optimizations enabled. At
+``-O0`` the variable falls back to ordinary (scratch) memory and a warning is
+emitted.
----------------
arsenm wrote:

This should not be a request. There's much less point in having this address space if the semantics are the same as stack migratable to VGPRs. We already do that today.
This should be an explicit allocation, not treated like a stack slot. I thought @nhaehnle was working on reworking this 

https://github.com/llvm/llvm-project/pull/205435