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

[Matt Arsenault via llvm-commits]

================
@@ -3710,6 +3710,9 @@ def err_attribute_argument_invalid : Error<
 def err_attribute_amdgpu_flat_work_group_size_mismatch : Error<
   "'amdgpu_flat_work_group_size' attribute must match "
   "'reqd_work_group_size' product">;
+def err_amdgpu_vgpr_not_kernel_local : Error<
+  "%0 attribute can only be applied to local variables in "
+  "'__global__' (kernel) functions">;
----------------
arsenm wrote:

This should be implemented the same way as LDS, in which case this restriction shouldn't be needed. Even if you reject this use in the frontend, the backend still needs to handle IR with direct references to the global variable in functions. IPO can introduce these even if they were originally only referenced from kernels 

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