As the only observable behaviour of this_thread::yield is forward progress, because of the as-if rule, the compiler doesn't actually need to replace the loop, when running on a runtime that guarantees preemption. That's the case when std::threads are backed by kernel threads. On a M:N implementation, then yes, a yield would need to be added, but that would be desirable.

Interestingly, posix realtime FIFO scheduling doesn't preempt even on kernel thread based implementations, so one reading of the standard would require yield on this case. But that can actually be potentially catastrophic as FIFO scheduling is expected to be deterministic. But realtime scheduling is already beyond the standard: I doubt gcc and clang will do the transformation by default.

In practice the equivalence is necessary to make some obscure corner of the memory model work and prevent some undesirable optimizations; I expect that in practice the compilers, if they implement this at all, will provide an opt-in flag, but they will optimize as-if the call was there.