This patch prevents DBG_VALUE instructions from influencing LiveRegUnits::stepBackward(). It mirrors rL327862 / D43850 that made the same changes for LivePhysRegs.
What motivated this patch was a problem in an out-of-tree target where the PrologEpilogInserter does register scavenging and decides it has to spill the scavenged register that LiveRegUnits finds live as it has a DBG_VALUE use. Since it's out-of-tree and a bit involved, I don't have a testcase to present here.