AMDGPU/SI: Update colorEndsAccordingToDependencies

[axeldavy]

Summary

This patch adds a comment to the
colorEndsAccordingToDependencies coloring pass of the SI scheduler
and disables it when there is no reserved block.

[vpykhtin]

lib/Target/AMDGPU/SIMachineScheduler.cpp

840

max_element can return end iterator only if the range is empty, so could you place an assert for this before these comparisons?

[axeldavy]

lib/Target/AMDGPU/SIMachineScheduler.cpp

840

The two vectors are supposed to be the size of DAGSize.
Should I assert their size, and that DAGSize is >= 1 ?

[vpykhtin]

lib/Target/AMDGPU/SIMachineScheduler.cpp

840

It would look a bit paranoid :-) but I think it would make more sence than just asserting its not empty.

[axeldavy]

Add assert over the initialization of Current*ReservedDependencyColoring and DAGSIZE

[vpykhtin]

LGTM.

lib/Target/AMDGPU/SIMachineScheduler.cpp

for (unsigned i = 0, e = DAGSize; i != e; ++i) {
     }
   }
 }
 
 void SIScheduleBlockCreator::colorEndsAccordingToDependencies() {
   unsigned DAGSize = DAG->SUnits.size();
   std::vector<int> PendingColoring = CurrentColoring;
 
+  assert(DAGSize >= 1 &&
+         CurrentBottomUpReservedDependencyColoring.size() == DAGSize &&
+         CurrentTopDownReservedDependencyColoring.size() == DAGSize);
+  // If there is no reserved block at all, do nothing. We don't want
+  // everything in one block.
+  if (*std::max_element(CurrentBottomUpReservedDependencyColoring.begin(),
+                        CurrentBottomUpReservedDependencyColoring.end()) == 0 &&
+      *std::max_element(CurrentTopDownReservedDependencyColoring.begin(),
+                        CurrentTopDownReservedDependencyColoring.end()) == 0)
+    return;
+
   for (unsigned SUNum : DAG->BottomUpIndex2SU) {
     SUnit *SU = &DAG->SUnits[SUNum];
     std::set<unsigned> SUColors;
     std::set<unsigned> SUColorsPending;
 
     if (CurrentColoring[SU->NodeNum] <= (int)DAGSize)
       continue;
 
     if (CurrentBottomUpReservedDependencyColoring[SU->NodeNum] > 0 ||
         CurrentTopDownReservedDependencyColoring[SU->NodeNum] > 0)
       continue;
 
     for (SDep& SuccDep : SU->Succs) {
       SUnit *Succ = SuccDep.getSUnit();
       if (SuccDep.isWeak() || Succ->NodeNum >= DAGSize)
         continue;
       if (CurrentBottomUpReservedDependencyColoring[Succ->NodeNum] > 0 ||
           CurrentTopDownReservedDependencyColoring[Succ->NodeNum] > 0)
         SUColors.insert(CurrentColoring[Succ->NodeNum]);
       SUColorsPending.insert(PendingColoring[Succ->NodeNum]);
     }
+    // If there is only one child/parent block, and that block
+    // is not among the ones we are removing in this path, then
+    // merge the instruction to that block
     if (SUColors.size() == 1 && SUColorsPending.size() == 1)
       PendingColoring[SU->NodeNum] = *SUColors.begin();
     else // TODO: Attribute new colors depending on color
          // combination of children.
       PendingColoring[SU->NodeNum] = NextNonReservedID++;
   }
   CurrentColoring = PendingColoring;
 }