I am developing backend for a new architecture, which has many Intrinsics with multiple outputs. However, I found TableGen doesn’t support such Intrinsics during Instruction Selection.
For example, load intrinsic with address updated.
def int_xxx_load_addr_updated: Intrinsic<[llvm_i32_ty, llvm_ptr_ty], [llvm_ptr_ty, llvm_i32_ty], >;
NVPTX has the same situation in the trunk, but it handles all intrinsics in source code which introduced a huge chunk of repetitive codes.
I have a patch for this issue, which is waiting for review. https://reviews.llvm.org/D32280
Any suggestions are highly appreciated.