search for: 506d770f

So, taking PR25526 as context, and after reading the internet, it seems to me that creating an atomic LL/SC loop at the IR level -- as is the modern way to do it in LLVM -- is effectively impossible to do correctly. Nor, for that matter, was it correct to create such a loop at isel time, as the implementation prior to r205525 did (and, as some targets still do, not having been updated yet to use