Avoiding overly aggressive latency scheduling. If the two nodes share an

[oota-llvm.git] / lib / CodeGen / SelectionDAG / ScheduleDAGSDNodes.cpp

diff --git a/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp b/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp
index 06cf0530875504f54c8597a4b95a15e20629ab11..7d01bd31b960148517eaffc4a4f5654938ef8c47 100644 (file)
--- a/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp
+++ b/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp
@@ -34,8 +34,8 @@ using namespace llvm;
 STATISTIC(LoadsClustered, "Number of loads clustered together");
 
 ScheduleDAGSDNodes::ScheduleDAGSDNodes(MachineFunction &mf)
-  : ScheduleDAG(mf) {
-}
+  : ScheduleDAG(mf),
+    InstrItins(mf.getTarget().getInstrItineraryData()) {}
 
 /// Run - perform scheduling.
 ///
@@ -59,8 +59,9 @@ SUnit *ScheduleDAGSDNodes::NewSUnit(SDNode *N) {
   SUnits.back().OrigNode = &SUnits.back();
   SUnit *SU = &SUnits.back();
   const TargetLowering &TLI = DAG->getTargetLoweringInfo();
-  if (N->isMachineOpcode() &&
-      N->getMachineOpcode() == TargetOpcode::IMPLICIT_DEF)
+  if (!N ||
+      (N->isMachineOpcode() &&
+       N->getMachineOpcode() == TargetOpcode::IMPLICIT_DEF))
     SU->SchedulingPref = Sched::None;
   else
     SU->SchedulingPref = TLI.getSchedulingPreference(N);
@@ -428,8 +429,7 @@ void ScheduleDAGSDNodes::ComputeLatency(SUnit *SU) {
     return;
   }
 
-  const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
-  if (InstrItins.isEmpty()) {
+  if (!InstrItins || InstrItins->isEmpty()) {
     SU->Latency = 1;
     return;
   }
@@ -439,7 +439,7 @@ void ScheduleDAGSDNodes::ComputeLatency(SUnit *SU) {
   SU->Latency = 0;
   for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode())
     if (N->isMachineOpcode()) {
-      SU->Latency += InstrItins.
+      SU->Latency += InstrItins->
         getStageLatency(TII->get(N->getMachineOpcode()).getSchedClass());
     }
 }
@@ -450,32 +450,25 @@ void ScheduleDAGSDNodes::ComputeOperandLatency(SDNode *Def, SDNode *Use,
   if (ForceUnitLatencies())
     return;
 
-  const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
-  if (InstrItins.isEmpty())
-    return;
-  
   if (dep.getKind() != SDep::Data)
     return;
 
   unsigned DefIdx = Use->getOperand(OpIdx).getResNo();
-  if (Def->isMachineOpcode()) {
-    const TargetInstrDesc &II = TII->get(Def->getMachineOpcode());
-    if (DefIdx >= II.getNumDefs())
-      return;
-    int DefCycle = InstrItins.getOperandCycle(II.getSchedClass(), DefIdx);
-    if (DefCycle < 0)
-      return;
-    int UseCycle = 1;
-    if (Use->isMachineOpcode()) {
-      const unsigned UseClass = TII->get(Use->getMachineOpcode()).getSchedClass();
-      UseCycle = InstrItins.getOperandCycle(UseClass, OpIdx);
-    }
-    if (UseCycle >= 0) {
-      int Latency = DefCycle - UseCycle + 1;
-      if (Latency >= 0)
-        dep.setLatency(Latency);
-    }
+  if (Use->isMachineOpcode())
+    // Adjust the use operand index by num of defs.
+    OpIdx += TII->get(Use->getMachineOpcode()).getNumDefs();
+  int Latency = TII->getOperandLatency(InstrItins, Def, DefIdx, Use, OpIdx);
+  if (Latency > 1 && Use->getOpcode() == ISD::CopyToReg &&
+      !BB->succ_empty()) {
+    unsigned Reg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
+    if (TargetRegisterInfo::isVirtualRegister(Reg))
+      // This copy is a liveout value. It is likely coalesced, so reduce the
+      // latency so not to penalize the def.
+      // FIXME: need target specific adjustment here?
+      Latency = (Latency > 1) ? Latency - 1 : 1;
   }
+  if (Latency >= 0)
+    dep.setLatency(Latency);
 }
 
 void ScheduleDAGSDNodes::dumpNode(const SUnit *SU) const {