[clang] [flang] [llvm] [Clang][AMDGPU] Expose buffer load lds as a clang builtin (PR #132048)

[Juan Manuel Martinez Caamaño via llvm-commits]

================
@@ -162,6 +162,8 @@ BUILTIN(__builtin_amdgcn_raw_buffer_load_b64, "V2UiQbiiIi", "n")
 BUILTIN(__builtin_amdgcn_raw_buffer_load_b96, "V3UiQbiiIi", "n")
 BUILTIN(__builtin_amdgcn_raw_buffer_load_b128, "V4UiQbiiIi", "n")
 
+TARGET_BUILTIN(__builtin_amdgcn_raw_ptr_buffer_load_lds, "vQbv*3IUiiiIiIi", "t", "lds-buffer-load-insts")
----------------
jmmartinez wrote:

Before gfx950: 1, 2, or 4.
On gfx950: 12 and 16 are also supported.

The main advantage of taking the size is that it matches the `__builtin_amdgcn_global_load_lds` and the LLVM IR intrinsic (I assume we can update the later more freely than clang's builtin signatures).

IMHO, having one version per size would have been better.

https://github.com/llvm/llvm-project/pull/132048