So, I'm not 100% sure yet, I need time to rebuild 4.14.28 with Michel's patch
and run it for a bit, but at this point I'm about 90% certain that a patch in
4.14.29 introduces a regression, as described in bug 106529. And from reading
the changelog and looking at the code, it does seem to me that this patch is
the most likely culprit.

Short version of what's wrong: on a multi-monitor setup, if the primary display
is on DP and this code toggles it off/on, the window manager doesn't take it
well and you end up with windows all over the place. 

(I don't know if this is the same codepath as is used with amdgpu.dc=1 on more
recent kernels, though I do now the same bug occurs on 4.16+ with dc=1.)