I have never gotten the point of this trick. Don't basically all processors have some equivalent of XCHG (swap contents of two registers)? Even the 8086 had it back in 1978. (The z80 doesn't have it, but it is a couple of years older than 8086.)

I tried searching the LLVM source for any mention of the trick, but couldn't find anything. So I can't see it having any relevance to modern CPUs.