misched preparation: clarify ScheduleDAG and ScheduleDAGInstrs roles.

ScheduleDAG is responsible for the DAG: SUnits and SDeps. It provides target hooks for latency computation.

ScheduleDAGInstrs extends ScheduleDAG and defines the current scheduling region in terms of MachineInstr iterators. It has access to the target's scheduling itinerary data. ScheduleDAGInstrs provides the logic for building the ScheduleDAG for the sequence of MachineInstrs in the current region. Target's can implement highly custom schedulers by extending this class.

ScheduleDAGPostRATDList provides the driver and diagnostics for current postRA scheduling. It maintains a current Sequence of scheduled machine instructions and logic for splicing them into the block. During scheduling, it uses the ScheduleHazardRecognizer provided by the target.

Specific changes:
- Removed driver code from ScheduleDAG. clearDAG is the only interface needed.

- Added enterRegion/exitRegion hooks to ScheduleDAGInstrs to delimit the scope of each scheduling region and associated DAG. They should be used to setup and cleanup any region-specific state in addition to the DAG itself. This is necessary because we reuse the same ScheduleDAG object for the entire function. The target may extend these hooks to do things at regions boundaries, like bundle terminators. The hooks are called even if we decide not to schedule the region. So all instructions in a block are "covered" by these calls.

- Added ScheduleDAGInstrs::begin()/end() public API.

- Moved Sequence into the driver layer, which is specific to the scheduling algorithm.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@152208 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/CodeGen/ScheduleDAG.h

@@ -486,15 +486,11 @@ namespace llvm {
 
   class ScheduleDAG {
   public:
-    MachineBasicBlock *BB;          // The block in which to insert instructions
-    MachineBasicBlock::iterator InsertPos;// The position to insert instructions
     const TargetMachine &TM;              // Target processor
     const TargetInstrInfo *TII;           // Target instruction information
     const TargetRegisterInfo *TRI;        // Target processor register info
     MachineFunction &MF;                  // Machine function
     MachineRegisterInfo &MRI;             // Virtual/real register map
-    std::vector<SUnit*> Sequence;         // The schedule. Null SUnit*'s
-                                          // represent noop instructions.
     std::vector<SUnit> SUnits;            // The scheduling units.
     SUnit EntrySU;                        // Special node for the region entry.
     SUnit ExitSU;                         // Special node for the region exit.
@@ -509,6 +505,9 @@ namespace llvm {
 
     virtual ~ScheduleDAG();
 
+    /// clearDAG - clear the DAG state (between regions).
+    void clearDAG();
+
     /// getInstrDesc - Return the MCInstrDesc of this SUnit.
     /// Return NULL for SDNodes without a machine opcode.
     const MCInstrDesc *getInstrDesc(const SUnit *SU) const {
@@ -542,10 +541,6 @@ namespace llvm {
 #endif
 
   protected:
-    /// Run - perform scheduling.
-    ///
-    void Run(MachineBasicBlock *bb, MachineBasicBlock::iterator insertPos);
-
     /// ComputeLatency - Compute node latency.
     ///
     virtual void ComputeLatency(SUnit *SU) = 0;
@@ -556,11 +551,6 @@ namespace llvm {
     virtual void ComputeOperandLatency(SUnit *, SUnit *,
                                        SDep&) const { }
 
-    /// Schedule - Order nodes according to selected style, filling
-    /// in the Sequence member.
-    ///
-    virtual void Schedule() = 0;
-
     /// ForceUnitLatencies - Return true if all scheduling edges should be given
     /// a latency value of one.  The default is to return false; schedulers may
     /// override this as needed.

lib/CodeGen/DFAPacketizer.cpp

@@ -185,7 +185,9 @@ void VLIWPacketizerList::PacketizeMIs(MachineBasicBlock *MBB,
                                       MachineBasicBlock::iterator BeginItr,
                                       MachineBasicBlock::iterator EndItr) {
   DefaultVLIWScheduler *Scheduler = (DefaultVLIWScheduler *)SchedulerImpl;
-  Scheduler->Run(MBB, BeginItr, EndItr, MBB->size());
+  Scheduler->enterRegion(MBB, BeginItr, EndItr, MBB->size());
+  Scheduler->Schedule();
+  Scheduler->exitRegion();
 
   // Remember scheduling units.
   SUnits = Scheduler->SUnits;

lib/CodeGen/MachineScheduler.cpp

@@ -281,15 +281,17 @@ bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) {
         if (TII->isSchedulingBoundary(llvm::prior(I), MBB, *MF))
           break;
       }
-      if (I == RegionEnd) {
-        // Skip empty scheduling regions.
-        RegionEnd = llvm::prior(RegionEnd);
-        --RemainingCount;
-        continue;
-      }
-      // Skip regions with one instruction.
-      if (I == llvm::prior(RegionEnd)) {
+      // Notify the scheduler of the region, even if we may skip scheduling
+      // it. Perhaps it still needs to be bundled.
+      Scheduler->enterRegion(MBB, I, RegionEnd, RemainingCount);
+
+      // Skip empty scheduling regions (0 or 1 schedulable instructions).
+      if (I == RegionEnd || I == llvm::prior(RegionEnd)) {
         RegionEnd = llvm::prior(RegionEnd);
+        if (I != RegionEnd)
+          --RemainingCount;
+        // Close the current region. Bundle the terminator if needed.
+        Scheduler->exitRegion();
         continue;
       }
       DEBUG(dbgs() << "MachineScheduling " << MF->getFunction()->getName()
@@ -300,8 +302,12 @@ bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) {
 
       // Inform ScheduleDAGInstrs of the region being scheduled. It calls back
       // to our Schedule() method.
-      Scheduler->Run(MBB, I, RegionEnd, MBB->size());
-      RegionEnd = Scheduler->Begin;
+      Scheduler->Schedule();
+      Scheduler->exitRegion();
+
+      // Scheduling has invalidated the current iterator 'I'. Ask the
+      // scheduler for the top of it's scheduled region.
+      RegionEnd = Scheduler->begin();
     }
     assert(RemainingCount == 0 && "Instruction count mismatch!");
     Scheduler->FinishBlock();

lib/CodeGen/PostRASchedulerList.cpp

@@ -127,6 +127,9 @@ namespace {
     /// LiveRegs - true if the register is live.
     BitVector LiveRegs;
 
+    /// The schedule. Null SUnit*'s represent noop instructions.
+    std::vector<SUnit*> Sequence;
+
   public:
     SchedulePostRATDList(
       MachineFunction &MF, MachineLoopInfo &MLI, MachineDominatorTree &MDT,
@@ -141,6 +144,15 @@ namespace {
     ///
     void StartBlock(MachineBasicBlock *BB);
 
+    /// Initialize the scheduler state for the next scheduling region.
+    virtual void enterRegion(MachineBasicBlock *bb,
+                             MachineBasicBlock::iterator begin,
+                             MachineBasicBlock::iterator end,
+                             unsigned endcount);
+
+    /// Notify that the scheduler has finished scheduling the current region.
+    virtual void exitRegion();
+
     /// Schedule - Schedule the instruction range using list scheduling.
     ///
     void Schedule();
@@ -206,6 +218,25 @@ SchedulePostRATDList::~SchedulePostRATDList() {
   delete AntiDepBreak;
 }
 
+/// Initialize state associated with the next scheduling region.
+void SchedulePostRATDList::enterRegion(MachineBasicBlock *bb,
+                 MachineBasicBlock::iterator begin,
+                 MachineBasicBlock::iterator end,
+                 unsigned endcount) {
+  ScheduleDAGInstrs::enterRegion(bb, begin, end, endcount);
+  Sequence.clear();
+}
+
+/// Print the schedule before exiting the region.
+void SchedulePostRATDList::exitRegion() {
+  DEBUG({
+      dbgs() << "*** Final schedule ***\n";
+      dumpSchedule();
+      dbgs() << '\n';
+    });
+  ScheduleDAGInstrs::exitRegion();
+}
+
 /// dumpSchedule - dump the scheduled Sequence.
 void SchedulePostRATDList::dumpSchedule() const {
   for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
@@ -282,7 +313,9 @@ bool PostRAScheduler::runOnMachineFunction(MachineFunction &Fn) {
       // post-ra we don't gain anything by scheduling across calls since we
       // don't need to worry about register pressure.
       if (MI->isCall() || TII->isSchedulingBoundary(MI, MBB, Fn)) {
-        Scheduler.Run(MBB, I, Current, CurrentCount);
+        Scheduler.enterRegion(MBB, I, Current, CurrentCount);
+        Scheduler.Schedule();
+        Scheduler.exitRegion();
         Scheduler.EmitSchedule();
         Current = MI;
         CurrentCount = Count - 1;
@@ -296,7 +329,9 @@ bool PostRAScheduler::runOnMachineFunction(MachineFunction &Fn) {
     assert(Count == 0 && "Instruction count mismatch!");
     assert((MBB->begin() == Current || CurrentCount != 0) &&
            "Instruction count mismatch!");
-    Scheduler.Run(MBB, MBB->begin(), Current, CurrentCount);
+    Scheduler.enterRegion(MBB, MBB->begin(), Current, CurrentCount);
+    Scheduler.Schedule();
+    Scheduler.exitRegion();
     Scheduler.EmitSchedule();
 
     // Clean up register live-range state.
@@ -340,10 +375,7 @@ void SchedulePostRATDList::Schedule() {
       // the def's anti-dependence *and* output-dependence edges due to
       // that register, and add new anti-dependence and output-dependence
       // edges based on the next live range of the register.
-      SUnits.clear();
-      Sequence.clear();
-      EntrySU = SUnit();
-      ExitSU = SUnit();
+      ScheduleDAG::clearDAG();
       BuildSchedGraph(AA);
 
       NumFixedAnti += Broken;
@@ -357,12 +389,6 @@ void SchedulePostRATDList::Schedule() {
   AvailableQueue.initNodes(SUnits);
   ListScheduleTopDown();
   AvailableQueue.releaseState();
-
-  DEBUG({
-      dbgs() << "*** Final schedule ***\n";
-      dumpSchedule();
-      dbgs() << '\n';
-    });
 }
 
 /// Observe - Update liveness information to account for the current

lib/CodeGen/ScheduleDAG.cpp

@@ -46,27 +46,19 @@ ScheduleDAG::ScheduleDAG(MachineFunction &mf)
 
 ScheduleDAG::~ScheduleDAG() {}
 
+/// Clear the DAG state (e.g. between scheduling regions).
+void ScheduleDAG::clearDAG() {
+  SUnits.clear();
+  EntrySU = SUnit();
+  ExitSU = SUnit();
+}
+
 /// getInstrDesc helper to handle SDNodes.
 const MCInstrDesc *ScheduleDAG::getNodeDesc(const SDNode *Node) const {
   if (!Node || !Node->isMachineOpcode()) return NULL;
   return &TII->get(Node->getMachineOpcode());
 }
 
-/// Run - perform scheduling.
-///
-void ScheduleDAG::Run(MachineBasicBlock *bb,
-                      MachineBasicBlock::iterator insertPos) {
-  BB = bb;
-  InsertPos = insertPos;
-
-  SUnits.clear();
-  Sequence.clear();
-  EntrySU = SUnit();
-  ExitSU = SUnit();
-
-  Schedule();
-}
-
 /// addPred - This adds the specified edge as a pred of the current node if
 /// not already.  It also adds the current node as a successor of the
 /// specified node.

lib/CodeGen/ScheduleDAGInstrs.cpp

@@ -46,22 +46,6 @@ ScheduleDAGInstrs::ScheduleDAGInstrs(MachineFunction &mf,
          "Virtual registers must be removed prior to PostRA scheduling");
 }
 
-/// Run - perform scheduling.
-///
-void ScheduleDAGInstrs::Run(MachineBasicBlock *bb,
-                            MachineBasicBlock::iterator begin,
-                            MachineBasicBlock::iterator end,
-                            unsigned endcount) {
-  BB = bb;
-  Begin = begin;
-  InsertPosIndex = endcount;
-
-  // Check to see if the scheduler cares about latencies.
-  UnitLatencies = ForceUnitLatencies();
-
-  ScheduleDAG::Run(bb, end);
-}
-
 /// getUnderlyingObjectFromInt - This is the function that does the work of
 /// looking through basic ptrtoint+arithmetic+inttoptr sequences.
 static const Value *getUnderlyingObjectFromInt(const Value *V) {
@@ -148,6 +132,10 @@ void ScheduleDAGInstrs::StartBlock(MachineBasicBlock *BB) {
       LoopRegs.VisitLoop(ML);
 }
 
+void ScheduleDAGInstrs::FinishBlock() {
+  // Nothing to do.
+}
+
 /// Initialize the map with the number of registers.
 void ScheduleDAGInstrs::Reg2SUnitsMap::setRegLimit(unsigned Limit) {
   PhysRegSet.setUniverse(Limit);
@@ -162,6 +150,31 @@ void ScheduleDAGInstrs::Reg2SUnitsMap::clear() {
   PhysRegSet.clear();
 }
 
+/// Initialize the DAG and common scheduler state for the current scheduling
+/// region. This does not actually create the DAG, only clears it. The
+/// scheduling driver may call BuildSchedGraph multiple times per scheduling
+/// region.
+void ScheduleDAGInstrs::enterRegion(MachineBasicBlock *bb,
+                                    MachineBasicBlock::iterator begin,
+                                    MachineBasicBlock::iterator end,
+                                    unsigned endcount) {
+  BB = bb;
+  Begin = begin;
+  InsertPos = end;
+  InsertPosIndex = endcount;
+
+  // Check to see if the scheduler cares about latencies.
+  UnitLatencies = ForceUnitLatencies();
+
+  ScheduleDAG::clearDAG();
+}
+
+/// Close the current scheduling region. Don't clear any state in case the
+/// driver wants to refer to the previous scheduling region.
+void ScheduleDAGInstrs::exitRegion() {
+  // Nothing to do.
+}
+
 /// AddSchedBarrierDeps - Add dependencies from instructions in the current
 /// list of instructions being scheduled to scheduling barrier by adding
 /// the exit SU to the register defs and use list. This is because we want to
@@ -715,10 +728,6 @@ void ScheduleDAGInstrs::BuildSchedGraph(AliasAnalysis *AA) {
   MISUnitMap.clear();
 }
 
-void ScheduleDAGInstrs::FinishBlock() {
-  // Nothing to do.
-}
-
 void ScheduleDAGInstrs::ComputeLatency(SUnit *SU) {
   // Compute the latency for the node.
   if (!InstrItins || InstrItins->isEmpty()) {

lib/CodeGen/ScheduleDAGInstrs.h

@@ -101,6 +101,7 @@ namespace llvm {
   /// ScheduleDAGInstrs - A ScheduleDAG subclass for scheduling lists of
   /// MachineInstrs.
   class LLVM_LIBRARY_VISIBILITY ScheduleDAGInstrs : public ScheduleDAG {
+  protected:
     const MachineLoopInfo &MLI;
     const MachineDominatorTree &MDT;
     const MachineFrameInfo *MFI;
@@ -112,6 +113,23 @@ namespace llvm {
     /// Live Intervals provides reaching defs in preRA scheduling.
     LiveIntervals *LIS;
 
+    /// State specific to the current scheduling region.
+    ///
+
+    // The block in which to insert instructions
+    MachineBasicBlock *BB;
+
+    // The beginning of the range to
+    // be scheduled. The range extends
+    // to InsertPos.
+    MachineBasicBlock::iterator Begin;
+
+    // The position to insert instructions
+    MachineBasicBlock::iterator InsertPos;
+
+    // The index in BB of InsertPos.
+    unsigned InsertPosIndex;
+
     /// After calling BuildSchedGraph, each machine instruction in the current
     /// scheduling region is mapped to an SUnit.
     DenseMap<MachineInstr*, SUnit*> MISUnitMap;
@@ -209,11 +227,6 @@ namespace llvm {
     MachineInstr *FirstDbgValue;
 
   public:
-    MachineBasicBlock::iterator Begin;    // The beginning of the range to
-                                          // be scheduled. The range extends
-                                          // to InsertPos.
-    unsigned InsertPosIndex;              // The index in BB of InsertPos.
-
     explicit ScheduleDAGInstrs(MachineFunction &mf,
                                const MachineLoopInfo &mli,
                                const MachineDominatorTree &mdt,
@@ -222,6 +235,12 @@ namespace llvm {
 
     virtual ~ScheduleDAGInstrs() {}
 
+    /// begin - Return an iterator to the top of the current scheduling region.
+    MachineBasicBlock::iterator begin() const { return Begin; }
+
+    /// end - Return an iterator to the bottom of the current scheduling region.
+    MachineBasicBlock::iterator end() const { return InsertPos; }
+
     /// NewSUnit - Creates a new SUnit and return a ptr to it.
     ///
     SUnit *NewSUnit(MachineInstr *MI) {
@@ -235,13 +254,22 @@ namespace llvm {
       return &SUnits.back();
     }
 
-
-    /// Run - perform scheduling.
+    /// StartBlock - Prepare to perform scheduling in the given block.
     ///
-    void Run(MachineBasicBlock *bb,
-             MachineBasicBlock::iterator begin,
-             MachineBasicBlock::iterator end,
-             unsigned endindex);
+    virtual void StartBlock(MachineBasicBlock *BB);
+
+    /// FinishBlock - Clean up after scheduling in the given block.
+    ///
+    virtual void FinishBlock();
+
+    /// Initialize the scheduler state for the next scheduling region.
+    virtual void enterRegion(MachineBasicBlock *bb,
+                             MachineBasicBlock::iterator begin,
+                             MachineBasicBlock::iterator end,
+                             unsigned endcount);
+
+    /// Notify that the scheduler has finished scheduling the current region.
+    virtual void exitRegion();
 
     /// BuildSchedGraph - Build SUnits from the MachineBasicBlock that we are
     /// input.
@@ -266,19 +294,11 @@ namespace llvm {
     virtual void ComputeOperandLatency(SUnit *Def, SUnit *Use,
                                        SDep& dep) const;
 
-    /// StartBlock - Prepare to perform scheduling in the given block.
-    ///
-    virtual void StartBlock(MachineBasicBlock *BB);
-
     /// Schedule - Order nodes according to selected style, filling
     /// in the Sequence member.
     ///
     virtual void Schedule() = 0;
 
-    /// FinishBlock - Clean up after scheduling in the given block.
-    ///
-    virtual void FinishBlock();
-
     virtual void dumpNode(const SUnit *SU) const;
 
     /// Return a label for a DAG node that points to an instruction.

lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp

@@ -46,15 +46,21 @@ static cl::opt<int> HighLatencyCycles(
            "instructions take for targets with no itinerary"));
 
 ScheduleDAGSDNodes::ScheduleDAGSDNodes(MachineFunction &mf)
-  : ScheduleDAG(mf),
+  : ScheduleDAG(mf), BB(0), DAG(0),
     InstrItins(mf.getTarget().getInstrItineraryData()) {}
 
 /// Run - perform scheduling.
 ///
-void ScheduleDAGSDNodes::Run(SelectionDAG *dag, MachineBasicBlock *bb,
-                             MachineBasicBlock::iterator insertPos) {
+void ScheduleDAGSDNodes::Run(SelectionDAG *dag, MachineBasicBlock *bb) {
+  BB = bb;
   DAG = dag;
-  ScheduleDAG::Run(bb, insertPos);
+
+  // Clear the scheduler's SUnit DAG.
+  ScheduleDAG::clearDAG();
+  Sequence.clear();
+
+  // Invoke the target's selection of scheduler.
+  Schedule();
 }
 
 /// NewSUnit - Creates a new SUnit and return a ptr to it.
@@ -752,7 +758,8 @@ EmitPhysRegCopy(SUnit *SU, DenseMap<SUnit*, unsigned> &VRBaseMap,
 /// InsertPos and MachineBasicBlock that contains this insertion
 /// point. ScheduleDAGSDNodes holds a BB pointer for convenience, but this does
 /// not necessarily refer to returned BB. The emitter may split blocks.
-MachineBasicBlock *ScheduleDAGSDNodes::EmitSchedule() {
+MachineBasicBlock *ScheduleDAGSDNodes::
+EmitSchedule(MachineBasicBlock::iterator &InsertPos) {
   InstrEmitter Emitter(BB, InsertPos);
   DenseMap<SDValue, unsigned> VRBaseMap;
   DenseMap<SUnit*, unsigned> CopyVRBaseMap;
@@ -860,9 +867,8 @@ MachineBasicBlock *ScheduleDAGSDNodes::EmitSchedule() {
     }
   }
 
-  BB = Emitter.getBlock();
   InsertPos = Emitter.getInsertPos();
-  return BB;
+  return Emitter.getBlock();
 }
 
 /// Return the basic block label.

lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.h

@@ -35,17 +35,20 @@ namespace llvm {
   ///
   class ScheduleDAGSDNodes : public ScheduleDAG {
   public:
+    MachineBasicBlock *BB;
     SelectionDAG *DAG;                    // DAG of the current basic block
     const InstrItineraryData *InstrItins;
 
+    /// The schedule. Null SUnit*'s represent noop instructions.
+    std::vector<SUnit*> Sequence;
+
     explicit ScheduleDAGSDNodes(MachineFunction &mf);
 
     virtual ~ScheduleDAGSDNodes() {}
 
     /// Run - perform scheduling.
     ///
-    void Run(SelectionDAG *dag, MachineBasicBlock *bb,
-             MachineBasicBlock::iterator insertPos);
+    void Run(SelectionDAG *dag, MachineBasicBlock *bb);
 
     /// isPassiveNode - Return true if the node is a non-scheduled leaf.
     ///
@@ -104,8 +107,6 @@ namespace llvm {
     virtual void ComputeOperandLatency(SDNode *Def, SDNode *Use,
                                        unsigned OpIdx, SDep& dep) const;
 
-    virtual MachineBasicBlock *EmitSchedule();
-
     /// Schedule - Order nodes according to selected style, filling
     /// in the Sequence member.
     ///

lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp

@@ -673,7 +673,7 @@ void SelectionDAGISel::CodeGenAndEmitDAG() {
   {
     NamedRegionTimer T("Instruction Scheduling", GroupName,
                        TimePassesIsEnabled);
-    Scheduler->Run(CurDAG, FuncInfo->MBB, FuncInfo->InsertPt);
+    Scheduler->Run(CurDAG, FuncInfo->MBB);
   }
 
   if (ViewSUnitDAGs) Scheduler->viewGraph();
@@ -684,8 +684,9 @@ void SelectionDAGISel::CodeGenAndEmitDAG() {
   {
     NamedRegionTimer T("Instruction Creation", GroupName, TimePassesIsEnabled);
 
-    LastMBB = FuncInfo->MBB = Scheduler->EmitSchedule();
-    FuncInfo->InsertPt = Scheduler->InsertPos;
+    // FuncInfo->InsertPt is passed by reference and set to the end of the
+    // scheduled instructions.
+    LastMBB = FuncInfo->MBB = Scheduler->EmitSchedule(FuncInfo->InsertPt);
   }
 
   // If the block was split, make sure we update any references that are used to