Well, one would start from the symlink libc call, then as you say move on to the system call, then analyze the system call for "redundant" operations (libc call

Well, one would start from the symlink libc call, then as you say move on to the system call, then analyze the system call for "redundant" operations (libc calls usually have many, kernel calls much fewer - who needs error checking if all we care is speed?). Then compare it with non-portable raw fs modifications of a chosen filesystem. I bet there could be some tiny gains at each stage.