I think it is a mistake to add architecture-specific #ifdef to MI
code rather than chasing and fixing the MD issue.

I also have an issue with the idea that this is one architecture
without a problem.  I strongly suspect uvm's behaviour has changed
in a dangerous way which is _already_ _also_ affecting other architectures,
and hiding that is bad.

I suspect the root cause is memory shortage in some layer, triggering
buggy retry sequences.