what's the status of libNVVM?

what’s the status of the libNVVM library that ships with the cuda toolkit? Should one use it or the compiler that’s checked into LLVM?

it doesn’t seem like libNVVM is quite as up-to-date as the LLVM version. For example, you can’t compile IR code that contains trailing metadata specifiers on functions like

define void @mykernel(i64 %x) #1 {


with libLLVM but you can compile it with “llc -mcpu=sm_35 kernel.ll -o kernel.ptx”, so is there any reason to use libNVVM? is the quality of the compiled PTX code the same?