================ @@ -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 _______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
