[ScheduleDAG] Make sure that addVRegUseDeps is called on subreg defs

[kparzysz]

Summary

A follow-up to D20102:

A def operand of a virtual register can actually read a register as well. Specifically, this happens in definitions of subregisters of a virtual register (unless the "undef" flag is present). This form of a use of a register should be accounted for in the construction of the dependence graph. In ScheduleDAGInstrs::buildSchedGraph, make sure that the function addVRegUseDeps is called for such defs.

[MatzeB]

That is the case I just discussed with Andy. We believe that the current code is fine because we get the output dependence edge anyway. So for now I added a comment to document this fact and avoid creating extra edges.

[kparzysz]

Ah, ok. I thought the original concern was that my first patch changed the existing behavior (which was then declared as not being the case), and that there was still an issue with the subregister definitions.

Abandoning.

lib/CodeGen/ScheduleDAGInstrs.cpp

for (unsigned j = 0, n = MI->getNumOperands(); j != n; ++j) {
       else {
         HasVRegDef = true;
         addVRegDefDeps(SU, j);
       }
     }
     // Now process all uses.
     for (unsigned j = 0, n = MI->getNumOperands(); j != n; ++j) {
       const MachineOperand &MO = MI->getOperand(j);
-      if (!MO.isReg() || !MO.isUse())
+      if (!MO.isReg())
         continue;
       unsigned Reg = MO.getReg();
       if (Reg == 0)
         continue;
 
-      if (TRI->isPhysicalRegister(Reg))
+      if (TRI->isPhysicalRegister(Reg)) {
+        if (MO.isUse())
           addPhysRegDeps(SU, j);
-      else if (MO.readsReg()) // ignore undef operands
+      } else if (MO.readsReg()) // ignore undef operands
         addVRegUseDeps(SU, j);
     }
 
     // If we haven't seen any uses in this scheduling region, create a
     // dependence edge to ExitSU to model the live-out latency. This is required
     // for vreg defs with no in-region use, and prefetches with no vreg def.
     //
     // FIXME: NumDataSuccs would be more precise than NumSuccs here. This