1 //===----- ScheduleDAGList.cpp - Reg pressure reduction list scheduler ----===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements bottom-up and top-down register pressure reduction list
11 // schedulers, using standard algorithms. The basic approach uses a priority
12 // queue of available nodes to schedule. One at a time, nodes are taken from
13 // the priority queue (thus in priority order), checked for legality to
14 // schedule, and emitted if legal.
16 //===----------------------------------------------------------------------===//
18 #define DEBUG_TYPE "pre-RA-sched"
19 #include "llvm/CodeGen/ScheduleDAG.h"
20 #include "llvm/CodeGen/SchedulerRegistry.h"
21 #include "llvm/Target/TargetRegisterInfo.h"
22 #include "llvm/Target/TargetData.h"
23 #include "llvm/Target/TargetMachine.h"
24 #include "llvm/Target/TargetInstrInfo.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/Compiler.h"
27 #include "llvm/ADT/SmallPtrSet.h"
28 #include "llvm/ADT/SmallSet.h"
29 #include "llvm/ADT/Statistic.h"
32 #include "llvm/Support/CommandLine.h"
35 STATISTIC(NumBacktracks, "Number of times scheduler backtracked");
36 STATISTIC(NumUnfolds, "Number of nodes unfolded");
37 STATISTIC(NumDups, "Number of duplicated nodes");
38 STATISTIC(NumCCCopies, "Number of cross class copies");
40 static RegisterScheduler
41 burrListDAGScheduler("list-burr",
42 " Bottom-up register reduction list scheduling",
43 createBURRListDAGScheduler);
44 static RegisterScheduler
45 tdrListrDAGScheduler("list-tdrr",
46 " Top-down register reduction list scheduling",
47 createTDRRListDAGScheduler);
50 //===----------------------------------------------------------------------===//
51 /// ScheduleDAGRRList - The actual register reduction list scheduler
52 /// implementation. This supports both top-down and bottom-up scheduling.
54 class VISIBILITY_HIDDEN ScheduleDAGRRList : public ScheduleDAG {
56 /// isBottomUp - This is true if the scheduling problem is bottom-up, false if
60 /// AvailableQueue - The priority queue to use for the available SUnits.
61 SchedulingPriorityQueue *AvailableQueue;
63 /// LiveRegs / LiveRegDefs - A set of physical registers and their definition
64 /// that are "live". These nodes must be scheduled before any other nodes that
65 /// modifies the registers can be scheduled.
66 SmallSet<unsigned, 4> LiveRegs;
67 std::vector<SUnit*> LiveRegDefs;
68 std::vector<unsigned> LiveRegCycles;
71 ScheduleDAGRRList(SelectionDAG &dag, MachineBasicBlock *bb,
72 const TargetMachine &tm, bool isbottomup,
73 SchedulingPriorityQueue *availqueue)
74 : ScheduleDAG(dag, bb, tm), isBottomUp(isbottomup),
75 AvailableQueue(availqueue) {
78 ~ScheduleDAGRRList() {
79 delete AvailableQueue;
84 /// IsReachable - Checks if SU is reachable from TargetSU.
85 bool IsReachable(SUnit *SU, SUnit *TargetSU);
87 /// willCreateCycle - Returns true if adding an edge from SU to TargetSU will
89 bool WillCreateCycle(SUnit *SU, SUnit *TargetSU);
91 /// AddPred - This adds the specified node X as a predecessor of
92 /// the current node Y if not already.
93 /// This returns true if this is a new predecessor.
94 /// Updates the topological ordering if required.
95 bool AddPred(SUnit *Y, SUnit *X, bool isCtrl, bool isSpecial,
96 unsigned PhyReg = 0, int Cost = 1);
98 /// RemovePred - This removes the specified node N from the predecessors of
99 /// the current node M. Updates the topological ordering if required.
100 bool RemovePred(SUnit *M, SUnit *N, bool isCtrl, bool isSpecial);
103 void ReleasePred(SUnit*, bool, unsigned);
104 void ReleaseSucc(SUnit*, bool isChain, unsigned);
105 void CapturePred(SUnit*, SUnit*, bool);
106 void ScheduleNodeBottomUp(SUnit*, unsigned);
107 void ScheduleNodeTopDown(SUnit*, unsigned);
108 void UnscheduleNodeBottomUp(SUnit*);
109 void BacktrackBottomUp(SUnit*, unsigned, unsigned&);
110 SUnit *CopyAndMoveSuccessors(SUnit*);
111 void InsertCCCopiesAndMoveSuccs(SUnit*, unsigned,
112 const TargetRegisterClass*,
113 const TargetRegisterClass*,
114 SmallVector<SUnit*, 2>&);
115 bool DelayForLiveRegsBottomUp(SUnit*, SmallVector<unsigned, 4>&);
116 void ListScheduleTopDown();
117 void ListScheduleBottomUp();
118 void CommuteNodesToReducePressure();
121 /// CreateNewSUnit - Creates a new SUnit and returns a pointer to it.
122 /// Updates the topological ordering if required.
123 SUnit *CreateNewSUnit(SDNode *N) {
124 SUnit *NewNode = NewSUnit(N);
125 // Update the topological ordering.
126 if (NewNode->NodeNum >= Node2Index.size())
127 InitDAGTopologicalSorting();
131 /// CreateClone - Creates a new SUnit from an existing one.
132 /// Updates the topological ordering if required.
133 SUnit *CreateClone(SUnit *N) {
134 SUnit *NewNode = Clone(N);
135 // Update the topological ordering.
136 if (NewNode->NodeNum >= Node2Index.size())
137 InitDAGTopologicalSorting();
141 /// Functions for preserving the topological ordering
142 /// even after dynamic insertions of new edges.
143 /// This allows a very fast implementation of IsReachable.
147 The idea of the algorithm is taken from
148 "Online algorithms for managing the topological order of
149 a directed acyclic graph" by David J. Pearce and Paul H.J. Kelly
150 This is the MNR algorithm, which was first introduced by
151 A. Marchetti-Spaccamela, U. Nanni and H. Rohnert in
152 "Maintaining a topological order under edge insertions".
154 Short description of the algorithm:
156 Topological ordering, ord, of a DAG maps each node to a topological
157 index so that for all edges X->Y it is the case that ord(X) < ord(Y).
159 This means that if there is a path from the node X to the node Z,
160 then ord(X) < ord(Z).
162 This property can be used to check for reachability of nodes:
163 if Z is reachable from X, then an insertion of the edge Z->X would
166 The algorithm first computes a topological ordering for the DAG by initializing
167 the Index2Node and Node2Index arrays and then tries to keep the ordering
168 up-to-date after edge insertions by reordering the DAG.
170 On insertion of the edge X->Y, the algorithm first marks by calling DFS the
171 nodes reachable from Y, and then shifts them using Shift to lie immediately
172 after X in Index2Node.
175 /// InitDAGTopologicalSorting - create the initial topological
176 /// ordering from the DAG to be scheduled.
177 void InitDAGTopologicalSorting();
179 /// DFS - make a DFS traversal and mark all nodes affected by the
180 /// edge insertion. These nodes will later get new topological indexes
181 /// by means of the Shift method.
182 void DFS(SUnit *SU, int UpperBound, bool& HasLoop);
184 /// Shift - reassign topological indexes for the nodes in the DAG
185 /// to preserve the topological ordering.
186 void Shift(BitVector& Visited, int LowerBound, int UpperBound);
188 /// Allocate - assign the topological index to the node n.
189 void Allocate(int n, int index);
191 /// Index2Node - Maps topological index to the node number.
192 std::vector<int> Index2Node;
193 /// Node2Index - Maps the node number to its topological index.
194 std::vector<int> Node2Index;
195 /// Visited - a set of nodes visited during a DFS traversal.
198 } // end anonymous namespace
201 /// Schedule - Schedule the DAG using list scheduling.
202 void ScheduleDAGRRList::Schedule() {
203 DOUT << "********** List Scheduling **********\n";
205 LiveRegDefs.resize(TRI->getNumRegs(), NULL);
206 LiveRegCycles.resize(TRI->getNumRegs(), 0);
208 // Build scheduling units.
211 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
212 SUnits[su].dumpAll(&DAG));
215 InitDAGTopologicalSorting();
217 AvailableQueue->initNodes(SUnitMap, SUnits);
219 // Execute the actual scheduling loop Top-Down or Bottom-Up as appropriate.
221 ListScheduleBottomUp();
223 ListScheduleTopDown();
225 AvailableQueue->releaseState();
227 CommuteNodesToReducePressure();
229 DOUT << "*** Final schedule ***\n";
230 DEBUG(dumpSchedule());
233 // Emit in scheduled order
237 /// CommuteNodesToReducePressure - If a node is two-address and commutable, and
238 /// it is not the last use of its first operand, add it to the CommuteSet if
239 /// possible. It will be commuted when it is translated to a MI.
240 void ScheduleDAGRRList::CommuteNodesToReducePressure() {
241 SmallPtrSet<SUnit*, 4> OperandSeen;
242 for (unsigned i = Sequence.size(); i != 0; ) {
244 SUnit *SU = Sequence[i];
245 if (!SU || !SU->Node) continue;
246 if (SU->isCommutable) {
247 unsigned Opc = SU->Node->getTargetOpcode();
248 const TargetInstrDesc &TID = TII->get(Opc);
249 unsigned NumRes = TID.getNumDefs();
250 unsigned NumOps = TID.getNumOperands() - NumRes;
251 for (unsigned j = 0; j != NumOps; ++j) {
252 if (TID.getOperandConstraint(j+NumRes, TOI::TIED_TO) == -1)
255 SDNode *OpN = SU->Node->getOperand(j).Val;
256 SUnit *OpSU = isPassiveNode(OpN) ? NULL : SUnitMap[OpN][SU->InstanceNo];
257 if (OpSU && OperandSeen.count(OpSU) == 1) {
258 // Ok, so SU is not the last use of OpSU, but SU is two-address so
259 // it will clobber OpSU. Try to commute SU if no other source operands
261 bool DoCommute = true;
262 for (unsigned k = 0; k < NumOps; ++k) {
264 OpN = SU->Node->getOperand(k).Val;
265 OpSU = isPassiveNode(OpN) ? NULL : SUnitMap[OpN][SU->InstanceNo];
266 if (OpSU && OperandSeen.count(OpSU) == 1) {
273 CommuteSet.insert(SU->Node);
276 // Only look at the first use&def node for now.
281 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
284 OperandSeen.insert(I->Dep);
289 //===----------------------------------------------------------------------===//
290 // Bottom-Up Scheduling
291 //===----------------------------------------------------------------------===//
293 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. Add it to
294 /// the AvailableQueue if the count reaches zero. Also update its cycle bound.
295 void ScheduleDAGRRList::ReleasePred(SUnit *PredSU, bool isChain,
297 // FIXME: the distance between two nodes is not always == the predecessor's
298 // latency. For example, the reader can very well read the register written
299 // by the predecessor later than the issue cycle. It also depends on the
300 // interrupt model (drain vs. freeze).
301 PredSU->CycleBound = std::max(PredSU->CycleBound, CurCycle + PredSU->Latency);
303 --PredSU->NumSuccsLeft;
306 if (PredSU->NumSuccsLeft < 0) {
307 cerr << "*** List scheduling failed! ***\n";
309 cerr << " has been released too many times!\n";
314 if (PredSU->NumSuccsLeft == 0) {
315 PredSU->isAvailable = true;
316 AvailableQueue->push(PredSU);
320 /// ScheduleNodeBottomUp - Add the node to the schedule. Decrement the pending
321 /// count of its predecessors. If a predecessor pending count is zero, add it to
322 /// the Available queue.
323 void ScheduleDAGRRList::ScheduleNodeBottomUp(SUnit *SU, unsigned CurCycle) {
324 DOUT << "*** Scheduling [" << CurCycle << "]: ";
325 DEBUG(SU->dump(&DAG));
326 SU->Cycle = CurCycle;
328 AvailableQueue->ScheduledNode(SU);
330 // Bottom up: release predecessors
331 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
333 ReleasePred(I->Dep, I->isCtrl, CurCycle);
335 // This is a physical register dependency and it's impossible or
336 // expensive to copy the register. Make sure nothing that can
337 // clobber the register is scheduled between the predecessor and
339 if (LiveRegs.insert(I->Reg)) {
340 LiveRegDefs[I->Reg] = I->Dep;
341 LiveRegCycles[I->Reg] = CurCycle;
346 // Release all the implicit physical register defs that are live.
347 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
350 if (LiveRegCycles[I->Reg] == I->Dep->Cycle) {
351 LiveRegs.erase(I->Reg);
352 assert(LiveRegDefs[I->Reg] == SU &&
353 "Physical register dependency violated?");
354 LiveRegDefs[I->Reg] = NULL;
355 LiveRegCycles[I->Reg] = 0;
360 SU->isScheduled = true;
363 /// CapturePred - This does the opposite of ReleasePred. Since SU is being
364 /// unscheduled, incrcease the succ left count of its predecessors. Remove
365 /// them from AvailableQueue if necessary.
366 void ScheduleDAGRRList::CapturePred(SUnit *PredSU, SUnit *SU, bool isChain) {
367 PredSU->CycleBound = 0;
368 for (SUnit::succ_iterator I = PredSU->Succs.begin(), E = PredSU->Succs.end();
372 PredSU->CycleBound = std::max(PredSU->CycleBound,
373 I->Dep->Cycle + PredSU->Latency);
376 if (PredSU->isAvailable) {
377 PredSU->isAvailable = false;
378 if (!PredSU->isPending)
379 AvailableQueue->remove(PredSU);
382 ++PredSU->NumSuccsLeft;
385 /// UnscheduleNodeBottomUp - Remove the node from the schedule, update its and
386 /// its predecessor states to reflect the change.
387 void ScheduleDAGRRList::UnscheduleNodeBottomUp(SUnit *SU) {
388 DOUT << "*** Unscheduling [" << SU->Cycle << "]: ";
389 DEBUG(SU->dump(&DAG));
391 AvailableQueue->UnscheduledNode(SU);
393 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
395 CapturePred(I->Dep, SU, I->isCtrl);
396 if (I->Cost < 0 && SU->Cycle == LiveRegCycles[I->Reg]) {
397 LiveRegs.erase(I->Reg);
398 assert(LiveRegDefs[I->Reg] == I->Dep &&
399 "Physical register dependency violated?");
400 LiveRegDefs[I->Reg] = NULL;
401 LiveRegCycles[I->Reg] = 0;
405 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
408 if (LiveRegs.insert(I->Reg)) {
409 assert(!LiveRegDefs[I->Reg] &&
410 "Physical register dependency violated?");
411 LiveRegDefs[I->Reg] = SU;
413 if (I->Dep->Cycle < LiveRegCycles[I->Reg])
414 LiveRegCycles[I->Reg] = I->Dep->Cycle;
419 SU->isScheduled = false;
420 SU->isAvailable = true;
421 AvailableQueue->push(SU);
424 /// IsReachable - Checks if SU is reachable from TargetSU.
425 bool ScheduleDAGRRList::IsReachable(SUnit *SU, SUnit *TargetSU) {
426 // If insertion of the edge SU->TargetSU would create a cycle
427 // then there is a path from TargetSU to SU.
428 int UpperBound, LowerBound;
429 LowerBound = Node2Index[TargetSU->NodeNum];
430 UpperBound = Node2Index[SU->NodeNum];
431 bool HasLoop = false;
432 // Is Ord(TargetSU) < Ord(SU) ?
433 if (LowerBound < UpperBound) {
435 // There may be a path from TargetSU to SU. Check for it.
436 DFS(TargetSU, UpperBound, HasLoop);
441 /// Allocate - assign the topological index to the node n.
442 inline void ScheduleDAGRRList::Allocate(int n, int index) {
443 Node2Index[n] = index;
444 Index2Node[index] = n;
447 /// InitDAGTopologicalSorting - create the initial topological
448 /// ordering from the DAG to be scheduled.
449 void ScheduleDAGRRList::InitDAGTopologicalSorting() {
450 unsigned DAGSize = SUnits.size();
451 std::vector<unsigned> InDegree(DAGSize);
452 std::vector<SUnit*> WorkList;
453 WorkList.reserve(DAGSize);
454 std::vector<SUnit*> TopOrder;
455 TopOrder.reserve(DAGSize);
457 // Initialize the data structures.
458 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
459 SUnit *SU = &SUnits[i];
460 int NodeNum = SU->NodeNum;
461 unsigned Degree = SU->Succs.size();
462 InDegree[NodeNum] = Degree;
464 // Is it a node without dependencies?
466 assert(SU->Succs.empty() && "SUnit should have no successors");
467 // Collect leaf nodes.
468 WorkList.push_back(SU);
472 while (!WorkList.empty()) {
473 SUnit *SU = WorkList.back();
475 TopOrder.push_back(SU);
476 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
479 if (!--InDegree[SU->NodeNum])
480 // If all dependencies of the node are processed already,
481 // then the node can be computed now.
482 WorkList.push_back(SU);
486 // Second pass, assign the actual topological order as node ids.
491 Index2Node.resize(DAGSize);
492 Node2Index.resize(DAGSize);
493 Visited.resize(DAGSize);
495 for (std::vector<SUnit*>::reverse_iterator TI = TopOrder.rbegin(),
496 TE = TopOrder.rend();TI != TE; ++TI) {
497 Allocate((*TI)->NodeNum, Id);
502 // Check correctness of the ordering
503 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
504 SUnit *SU = &SUnits[i];
505 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
507 assert(Node2Index[SU->NodeNum] > Node2Index[I->Dep->NodeNum] &&
508 "Wrong topological sorting");
514 /// AddPred - adds an edge from SUnit X to SUnit Y.
515 /// Updates the topological ordering if required.
516 bool ScheduleDAGRRList::AddPred(SUnit *Y, SUnit *X, bool isCtrl, bool isSpecial,
517 unsigned PhyReg, int Cost) {
518 int UpperBound, LowerBound;
519 LowerBound = Node2Index[Y->NodeNum];
520 UpperBound = Node2Index[X->NodeNum];
521 bool HasLoop = false;
522 // Is Ord(X) < Ord(Y) ?
523 if (LowerBound < UpperBound) {
524 // Update the topological order.
526 DFS(Y, UpperBound, HasLoop);
527 assert(!HasLoop && "Inserted edge creates a loop!");
528 // Recompute topological indexes.
529 Shift(Visited, LowerBound, UpperBound);
531 // Now really insert the edge.
532 return Y->addPred(X, isCtrl, isSpecial, PhyReg, Cost);
535 /// RemovePred - This removes the specified node N from the predecessors of
536 /// the current node M. Updates the topological ordering if required.
537 bool ScheduleDAGRRList::RemovePred(SUnit *M, SUnit *N,
538 bool isCtrl, bool isSpecial) {
539 // InitDAGTopologicalSorting();
540 return M->removePred(N, isCtrl, isSpecial);
543 /// DFS - Make a DFS traversal to mark all nodes reachable from SU and mark
544 /// all nodes affected by the edge insertion. These nodes will later get new
545 /// topological indexes by means of the Shift method.
546 void ScheduleDAGRRList::DFS(SUnit *SU, int UpperBound, bool& HasLoop) {
547 std::vector<SUnit*> WorkList;
548 WorkList.reserve(SUnits.size());
550 WorkList.push_back(SU);
551 while (!WorkList.empty()) {
552 SU = WorkList.back();
554 Visited.set(SU->NodeNum);
555 for (int I = SU->Succs.size()-1; I >= 0; --I) {
556 int s = SU->Succs[I].Dep->NodeNum;
557 if (Node2Index[s] == UpperBound) {
561 // Visit successors if not already and in affected region.
562 if (!Visited.test(s) && Node2Index[s] < UpperBound) {
563 WorkList.push_back(SU->Succs[I].Dep);
569 /// Shift - Renumber the nodes so that the topological ordering is
571 void ScheduleDAGRRList::Shift(BitVector& Visited, int LowerBound,
577 for (i = LowerBound; i <= UpperBound; ++i) {
578 // w is node at topological index i.
579 int w = Index2Node[i];
580 if (Visited.test(w)) {
586 Allocate(w, i - shift);
590 for (unsigned j = 0; j < L.size(); ++j) {
591 Allocate(L[j], i - shift);
597 /// WillCreateCycle - Returns true if adding an edge from SU to TargetSU will
599 bool ScheduleDAGRRList::WillCreateCycle(SUnit *SU, SUnit *TargetSU) {
600 if (IsReachable(TargetSU, SU))
602 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
604 if (I->Cost < 0 && IsReachable(TargetSU, I->Dep))
609 /// BacktrackBottomUp - Backtrack scheduling to a previous cycle specified in
610 /// BTCycle in order to schedule a specific node. Returns the last unscheduled
611 /// SUnit. Also returns if a successor is unscheduled in the process.
612 void ScheduleDAGRRList::BacktrackBottomUp(SUnit *SU, unsigned BtCycle,
613 unsigned &CurCycle) {
615 while (CurCycle > BtCycle) {
616 OldSU = Sequence.back();
618 if (SU->isSucc(OldSU))
619 // Don't try to remove SU from AvailableQueue.
620 SU->isAvailable = false;
621 UnscheduleNodeBottomUp(OldSU);
626 if (SU->isSucc(OldSU)) {
627 assert(false && "Something is wrong!");
634 /// CopyAndMoveSuccessors - Clone the specified node and move its scheduled
635 /// successors to the newly created node.
636 SUnit *ScheduleDAGRRList::CopyAndMoveSuccessors(SUnit *SU) {
637 if (SU->FlaggedNodes.size())
640 SDNode *N = SU->Node;
645 bool TryUnfold = false;
646 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i) {
647 MVT::ValueType VT = N->getValueType(i);
650 else if (VT == MVT::Other)
653 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
654 const SDOperand &Op = N->getOperand(i);
655 MVT::ValueType VT = Op.Val->getValueType(Op.ResNo);
661 SmallVector<SDNode*, 4> NewNodes;
662 if (!TII->unfoldMemoryOperand(DAG, N, NewNodes))
665 DOUT << "Unfolding SU # " << SU->NodeNum << "\n";
666 assert(NewNodes.size() == 2 && "Expected a load folding node!");
669 SDNode *LoadNode = NewNodes[0];
670 unsigned NumVals = N->getNumValues();
671 unsigned OldNumVals = SU->Node->getNumValues();
672 for (unsigned i = 0; i != NumVals; ++i)
673 DAG.ReplaceAllUsesOfValueWith(SDOperand(SU->Node, i), SDOperand(N, i));
674 DAG.ReplaceAllUsesOfValueWith(SDOperand(SU->Node, OldNumVals-1),
675 SDOperand(LoadNode, 1));
677 SUnit *NewSU = CreateNewSUnit(N);
678 SUnitMap[N].push_back(NewSU);
679 const TargetInstrDesc &TID = TII->get(N->getTargetOpcode());
680 for (unsigned i = 0; i != TID.getNumOperands(); ++i) {
681 if (TID.getOperandConstraint(i, TOI::TIED_TO) != -1) {
682 NewSU->isTwoAddress = true;
686 if (TID.isCommutable())
687 NewSU->isCommutable = true;
688 // FIXME: Calculate height / depth and propagate the changes?
689 NewSU->Depth = SU->Depth;
690 NewSU->Height = SU->Height;
691 ComputeLatency(NewSU);
693 // LoadNode may already exist. This can happen when there is another
694 // load from the same location and producing the same type of value
695 // but it has different alignment or volatileness.
696 bool isNewLoad = true;
698 DenseMap<SDNode*, std::vector<SUnit*> >::iterator SMI =
699 SUnitMap.find(LoadNode);
700 if (SMI != SUnitMap.end()) {
701 LoadSU = SMI->second.front();
704 LoadSU = CreateNewSUnit(LoadNode);
705 SUnitMap[LoadNode].push_back(LoadSU);
707 LoadSU->Depth = SU->Depth;
708 LoadSU->Height = SU->Height;
709 ComputeLatency(LoadSU);
712 SUnit *ChainPred = NULL;
713 SmallVector<SDep, 4> ChainSuccs;
714 SmallVector<SDep, 4> LoadPreds;
715 SmallVector<SDep, 4> NodePreds;
716 SmallVector<SDep, 4> NodeSuccs;
717 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
721 else if (I->Dep->Node && I->Dep->Node->isOperandOf(LoadNode))
722 LoadPreds.push_back(SDep(I->Dep, I->Reg, I->Cost, false, false));
724 NodePreds.push_back(SDep(I->Dep, I->Reg, I->Cost, false, false));
726 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
729 ChainSuccs.push_back(SDep(I->Dep, I->Reg, I->Cost,
730 I->isCtrl, I->isSpecial));
732 NodeSuccs.push_back(SDep(I->Dep, I->Reg, I->Cost,
733 I->isCtrl, I->isSpecial));
737 RemovePred(SU, ChainPred, true, false);
739 AddPred(LoadSU, ChainPred, true, false);
741 for (unsigned i = 0, e = LoadPreds.size(); i != e; ++i) {
742 SDep *Pred = &LoadPreds[i];
743 RemovePred(SU, Pred->Dep, Pred->isCtrl, Pred->isSpecial);
745 AddPred(LoadSU, Pred->Dep, Pred->isCtrl, Pred->isSpecial,
746 Pred->Reg, Pred->Cost);
749 for (unsigned i = 0, e = NodePreds.size(); i != e; ++i) {
750 SDep *Pred = &NodePreds[i];
751 RemovePred(SU, Pred->Dep, Pred->isCtrl, Pred->isSpecial);
752 AddPred(NewSU, Pred->Dep, Pred->isCtrl, Pred->isSpecial,
753 Pred->Reg, Pred->Cost);
755 for (unsigned i = 0, e = NodeSuccs.size(); i != e; ++i) {
756 SDep *Succ = &NodeSuccs[i];
757 RemovePred(Succ->Dep, SU, Succ->isCtrl, Succ->isSpecial);
758 AddPred(Succ->Dep, NewSU, Succ->isCtrl, Succ->isSpecial,
759 Succ->Reg, Succ->Cost);
761 for (unsigned i = 0, e = ChainSuccs.size(); i != e; ++i) {
762 SDep *Succ = &ChainSuccs[i];
763 RemovePred(Succ->Dep, SU, Succ->isCtrl, Succ->isSpecial);
765 AddPred(Succ->Dep, LoadSU, Succ->isCtrl, Succ->isSpecial,
766 Succ->Reg, Succ->Cost);
770 AddPred(NewSU, LoadSU, false, false);
774 AvailableQueue->addNode(LoadSU);
775 AvailableQueue->addNode(NewSU);
779 if (NewSU->NumSuccsLeft == 0) {
780 NewSU->isAvailable = true;
786 DOUT << "Duplicating SU # " << SU->NodeNum << "\n";
787 NewSU = CreateClone(SU);
789 // New SUnit has the exact same predecessors.
790 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
793 AddPred(NewSU, I->Dep, I->isCtrl, false, I->Reg, I->Cost);
794 NewSU->Depth = std::max(NewSU->Depth, I->Dep->Depth+1);
797 // Only copy scheduled successors. Cut them from old node's successor
798 // list and move them over.
799 SmallVector<std::pair<SUnit*, bool>, 4> DelDeps;
800 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
804 if (I->Dep->isScheduled) {
805 NewSU->Height = std::max(NewSU->Height, I->Dep->Height+1);
806 AddPred(I->Dep, NewSU, I->isCtrl, false, I->Reg, I->Cost);
807 DelDeps.push_back(std::make_pair(I->Dep, I->isCtrl));
810 for (unsigned i = 0, e = DelDeps.size(); i != e; ++i) {
811 SUnit *Succ = DelDeps[i].first;
812 bool isCtrl = DelDeps[i].second;
813 RemovePred(Succ, SU, isCtrl, false);
816 AvailableQueue->updateNode(SU);
817 AvailableQueue->addNode(NewSU);
823 /// InsertCCCopiesAndMoveSuccs - Insert expensive cross register class copies
824 /// and move all scheduled successors of the given SUnit to the last copy.
825 void ScheduleDAGRRList::InsertCCCopiesAndMoveSuccs(SUnit *SU, unsigned Reg,
826 const TargetRegisterClass *DestRC,
827 const TargetRegisterClass *SrcRC,
828 SmallVector<SUnit*, 2> &Copies) {
829 SUnit *CopyFromSU = CreateNewSUnit(NULL);
830 CopyFromSU->CopySrcRC = SrcRC;
831 CopyFromSU->CopyDstRC = DestRC;
832 CopyFromSU->Depth = SU->Depth;
833 CopyFromSU->Height = SU->Height;
835 SUnit *CopyToSU = CreateNewSUnit(NULL);
836 CopyToSU->CopySrcRC = DestRC;
837 CopyToSU->CopyDstRC = SrcRC;
839 // Only copy scheduled successors. Cut them from old node's successor
840 // list and move them over.
841 SmallVector<std::pair<SUnit*, bool>, 4> DelDeps;
842 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
846 if (I->Dep->isScheduled) {
847 CopyToSU->Height = std::max(CopyToSU->Height, I->Dep->Height+1);
848 AddPred(I->Dep, CopyToSU, I->isCtrl, false, I->Reg, I->Cost);
849 DelDeps.push_back(std::make_pair(I->Dep, I->isCtrl));
852 for (unsigned i = 0, e = DelDeps.size(); i != e; ++i) {
853 SUnit *Succ = DelDeps[i].first;
854 bool isCtrl = DelDeps[i].second;
855 RemovePred(Succ, SU, isCtrl, false);
858 AddPred(CopyFromSU, SU, false, false, Reg, -1);
859 AddPred(CopyToSU, CopyFromSU, false, false, Reg, 1);
861 AvailableQueue->updateNode(SU);
862 AvailableQueue->addNode(CopyFromSU);
863 AvailableQueue->addNode(CopyToSU);
864 Copies.push_back(CopyFromSU);
865 Copies.push_back(CopyToSU);
870 /// getPhysicalRegisterVT - Returns the ValueType of the physical register
871 /// definition of the specified node.
872 /// FIXME: Move to SelectionDAG?
873 static MVT::ValueType getPhysicalRegisterVT(SDNode *N, unsigned Reg,
874 const TargetInstrInfo *TII) {
875 const TargetInstrDesc &TID = TII->get(N->getTargetOpcode());
876 assert(TID.ImplicitDefs && "Physical reg def must be in implicit def list!");
877 unsigned NumRes = TID.getNumDefs();
878 for (const unsigned *ImpDef = TID.getImplicitDefs(); *ImpDef; ++ImpDef) {
883 return N->getValueType(NumRes);
886 /// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay
887 /// scheduling of the given node to satisfy live physical register dependencies.
888 /// If the specific node is the last one that's available to schedule, do
889 /// whatever is necessary (i.e. backtracking or cloning) to make it possible.
890 bool ScheduleDAGRRList::DelayForLiveRegsBottomUp(SUnit *SU,
891 SmallVector<unsigned, 4> &LRegs){
892 if (LiveRegs.empty())
895 SmallSet<unsigned, 4> RegAdded;
896 // If this node would clobber any "live" register, then it's not ready.
897 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
900 unsigned Reg = I->Reg;
901 if (LiveRegs.count(Reg) && LiveRegDefs[Reg] != I->Dep) {
902 if (RegAdded.insert(Reg))
903 LRegs.push_back(Reg);
905 for (const unsigned *Alias = TRI->getAliasSet(Reg);
907 if (LiveRegs.count(*Alias) && LiveRegDefs[*Alias] != I->Dep) {
908 if (RegAdded.insert(*Alias))
909 LRegs.push_back(*Alias);
914 for (unsigned i = 0, e = SU->FlaggedNodes.size()+1; i != e; ++i) {
915 SDNode *Node = (i == 0) ? SU->Node : SU->FlaggedNodes[i-1];
916 if (!Node || !Node->isTargetOpcode())
918 const TargetInstrDesc &TID = TII->get(Node->getTargetOpcode());
919 if (!TID.ImplicitDefs)
921 for (const unsigned *Reg = TID.ImplicitDefs; *Reg; ++Reg) {
922 if (LiveRegs.count(*Reg) && LiveRegDefs[*Reg] != SU) {
923 if (RegAdded.insert(*Reg))
924 LRegs.push_back(*Reg);
926 for (const unsigned *Alias = TRI->getAliasSet(*Reg);
928 if (LiveRegs.count(*Alias) && LiveRegDefs[*Alias] != SU) {
929 if (RegAdded.insert(*Alias))
930 LRegs.push_back(*Alias);
934 return !LRegs.empty();
938 /// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
940 void ScheduleDAGRRList::ListScheduleBottomUp() {
941 unsigned CurCycle = 0;
942 // Add root to Available queue.
943 if (!SUnits.empty()) {
944 SUnit *RootSU = SUnitMap[DAG.getRoot().Val].front();
945 assert(RootSU->Succs.empty() && "Graph root shouldn't have successors!");
946 RootSU->isAvailable = true;
947 AvailableQueue->push(RootSU);
950 // While Available queue is not empty, grab the node with the highest
951 // priority. If it is not ready put it back. Schedule the node.
952 SmallVector<SUnit*, 4> NotReady;
953 while (!AvailableQueue->empty()) {
954 bool Delayed = false;
955 DenseMap<SUnit*, SmallVector<unsigned, 4> > LRegsMap;
956 SUnit *CurSU = AvailableQueue->pop();
958 if (CurSU->CycleBound <= CurCycle) {
959 SmallVector<unsigned, 4> LRegs;
960 if (!DelayForLiveRegsBottomUp(CurSU, LRegs))
963 LRegsMap.insert(std::make_pair(CurSU, LRegs));
966 CurSU->isPending = true; // This SU is not in AvailableQueue right now.
967 NotReady.push_back(CurSU);
968 CurSU = AvailableQueue->pop();
971 // All candidates are delayed due to live physical reg dependencies.
972 // Try backtracking, code duplication, or inserting cross class copies
974 if (Delayed && !CurSU) {
975 for (unsigned i = 0, e = NotReady.size(); i != e; ++i) {
976 SUnit *TrySU = NotReady[i];
977 SmallVector<unsigned, 4> &LRegs = LRegsMap[TrySU];
979 // Try unscheduling up to the point where it's safe to schedule
981 unsigned LiveCycle = CurCycle;
982 for (unsigned j = 0, ee = LRegs.size(); j != ee; ++j) {
983 unsigned Reg = LRegs[j];
984 unsigned LCycle = LiveRegCycles[Reg];
985 LiveCycle = std::min(LiveCycle, LCycle);
987 SUnit *OldSU = Sequence[LiveCycle];
988 if (!WillCreateCycle(TrySU, OldSU)) {
989 BacktrackBottomUp(TrySU, LiveCycle, CurCycle);
990 // Force the current node to be scheduled before the node that
991 // requires the physical reg dep.
992 if (OldSU->isAvailable) {
993 OldSU->isAvailable = false;
994 AvailableQueue->remove(OldSU);
996 AddPred(TrySU, OldSU, true, true);
997 // If one or more successors has been unscheduled, then the current
998 // node is no longer avaialable. Schedule a successor that's now
999 // available instead.
1000 if (!TrySU->isAvailable)
1001 CurSU = AvailableQueue->pop();
1004 TrySU->isPending = false;
1005 NotReady.erase(NotReady.begin()+i);
1012 // Can't backtrack. Try duplicating the nodes that produces these
1013 // "expensive to copy" values to break the dependency. In case even
1014 // that doesn't work, insert cross class copies.
1015 SUnit *TrySU = NotReady[0];
1016 SmallVector<unsigned, 4> &LRegs = LRegsMap[TrySU];
1017 assert(LRegs.size() == 1 && "Can't handle this yet!");
1018 unsigned Reg = LRegs[0];
1019 SUnit *LRDef = LiveRegDefs[Reg];
1020 SUnit *NewDef = CopyAndMoveSuccessors(LRDef);
1022 // Issue expensive cross register class copies.
1023 MVT::ValueType VT = getPhysicalRegisterVT(LRDef->Node, Reg, TII);
1024 const TargetRegisterClass *RC =
1025 TRI->getPhysicalRegisterRegClass(Reg, VT);
1026 const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
1028 assert(false && "Don't know how to copy this physical register!");
1031 SmallVector<SUnit*, 2> Copies;
1032 InsertCCCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
1033 DOUT << "Adding an edge from SU # " << TrySU->NodeNum
1034 << " to SU #" << Copies.front()->NodeNum << "\n";
1035 AddPred(TrySU, Copies.front(), true, true);
1036 NewDef = Copies.back();
1039 DOUT << "Adding an edge from SU # " << NewDef->NodeNum
1040 << " to SU #" << TrySU->NodeNum << "\n";
1041 LiveRegDefs[Reg] = NewDef;
1042 AddPred(NewDef, TrySU, true, true);
1043 TrySU->isAvailable = false;
1048 assert(false && "Unable to resolve live physical register dependencies!");
1053 // Add the nodes that aren't ready back onto the available list.
1054 for (unsigned i = 0, e = NotReady.size(); i != e; ++i) {
1055 NotReady[i]->isPending = false;
1056 // May no longer be available due to backtracking.
1057 if (NotReady[i]->isAvailable)
1058 AvailableQueue->push(NotReady[i]);
1063 Sequence.push_back(0);
1065 ScheduleNodeBottomUp(CurSU, CurCycle);
1066 Sequence.push_back(CurSU);
1071 // Reverse the order if it is bottom up.
1072 std::reverse(Sequence.begin(), Sequence.end());
1076 // Verify that all SUnits were scheduled.
1077 bool AnyNotSched = false;
1078 unsigned DeadNodes = 0;
1079 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
1080 if (!SUnits[i].isScheduled) {
1081 if (SUnits[i].NumPreds == 0 && SUnits[i].NumSuccs == 0) {
1086 cerr << "*** List scheduling failed! ***\n";
1087 SUnits[i].dump(&DAG);
1088 cerr << "has not been scheduled!\n";
1091 if (SUnits[i].NumSuccsLeft != 0) {
1093 cerr << "*** List scheduling failed! ***\n";
1094 SUnits[i].dump(&DAG);
1095 cerr << "has successors left!\n";
1099 assert(!AnyNotSched);
1100 assert(Sequence.size() + DeadNodes == SUnits.size() &&
1101 "The number of nodes scheduled doesn't match the expected number!");
1105 //===----------------------------------------------------------------------===//
1106 // Top-Down Scheduling
1107 //===----------------------------------------------------------------------===//
1109 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to
1110 /// the AvailableQueue if the count reaches zero. Also update its cycle bound.
1111 void ScheduleDAGRRList::ReleaseSucc(SUnit *SuccSU, bool isChain,
1112 unsigned CurCycle) {
1113 // FIXME: the distance between two nodes is not always == the predecessor's
1114 // latency. For example, the reader can very well read the register written
1115 // by the predecessor later than the issue cycle. It also depends on the
1116 // interrupt model (drain vs. freeze).
1117 SuccSU->CycleBound = std::max(SuccSU->CycleBound, CurCycle + SuccSU->Latency);
1119 --SuccSU->NumPredsLeft;
1122 if (SuccSU->NumPredsLeft < 0) {
1123 cerr << "*** List scheduling failed! ***\n";
1125 cerr << " has been released too many times!\n";
1130 if (SuccSU->NumPredsLeft == 0) {
1131 SuccSU->isAvailable = true;
1132 AvailableQueue->push(SuccSU);
1137 /// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending
1138 /// count of its successors. If a successor pending count is zero, add it to
1139 /// the Available queue.
1140 void ScheduleDAGRRList::ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle) {
1141 DOUT << "*** Scheduling [" << CurCycle << "]: ";
1142 DEBUG(SU->dump(&DAG));
1143 SU->Cycle = CurCycle;
1145 AvailableQueue->ScheduledNode(SU);
1147 // Top down: release successors
1148 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1150 ReleaseSucc(I->Dep, I->isCtrl, CurCycle);
1151 SU->isScheduled = true;
1154 /// ListScheduleTopDown - The main loop of list scheduling for top-down
1156 void ScheduleDAGRRList::ListScheduleTopDown() {
1157 unsigned CurCycle = 0;
1159 // All leaves to Available queue.
1160 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
1161 // It is available if it has no predecessors.
1162 if (SUnits[i].Preds.empty()) {
1163 AvailableQueue->push(&SUnits[i]);
1164 SUnits[i].isAvailable = true;
1168 // While Available queue is not empty, grab the node with the highest
1169 // priority. If it is not ready put it back. Schedule the node.
1170 std::vector<SUnit*> NotReady;
1171 while (!AvailableQueue->empty()) {
1172 SUnit *CurSU = AvailableQueue->pop();
1173 while (CurSU && CurSU->CycleBound > CurCycle) {
1174 NotReady.push_back(CurSU);
1175 CurSU = AvailableQueue->pop();
1178 // Add the nodes that aren't ready back onto the available list.
1179 AvailableQueue->push_all(NotReady);
1183 Sequence.push_back(0);
1185 ScheduleNodeTopDown(CurSU, CurCycle);
1186 Sequence.push_back(CurSU);
1193 // Verify that all SUnits were scheduled.
1194 bool AnyNotSched = false;
1195 unsigned DeadNodes = 0;
1196 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
1197 if (!SUnits[i].isScheduled) {
1198 if (SUnits[i].NumPreds == 0 && SUnits[i].NumSuccs == 0) {
1203 cerr << "*** List scheduling failed! ***\n";
1204 SUnits[i].dump(&DAG);
1205 cerr << "has not been scheduled!\n";
1208 if (SUnits[i].NumPredsLeft != 0) {
1210 cerr << "*** List scheduling failed! ***\n";
1211 SUnits[i].dump(&DAG);
1212 cerr << "has predecessors left!\n";
1216 assert(!AnyNotSched);
1217 assert(Sequence.size() + DeadNodes == SUnits.size() &&
1218 "The number of nodes scheduled doesn't match the expected number!");
1224 //===----------------------------------------------------------------------===//
1225 // RegReductionPriorityQueue Implementation
1226 //===----------------------------------------------------------------------===//
1228 // This is a SchedulingPriorityQueue that schedules using Sethi Ullman numbers
1229 // to reduce register pressure.
1233 class RegReductionPriorityQueue;
1235 /// Sorting functions for the Available queue.
1236 struct bu_ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> {
1237 RegReductionPriorityQueue<bu_ls_rr_sort> *SPQ;
1238 bu_ls_rr_sort(RegReductionPriorityQueue<bu_ls_rr_sort> *spq) : SPQ(spq) {}
1239 bu_ls_rr_sort(const bu_ls_rr_sort &RHS) : SPQ(RHS.SPQ) {}
1241 bool operator()(const SUnit* left, const SUnit* right) const;
1244 struct td_ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> {
1245 RegReductionPriorityQueue<td_ls_rr_sort> *SPQ;
1246 td_ls_rr_sort(RegReductionPriorityQueue<td_ls_rr_sort> *spq) : SPQ(spq) {}
1247 td_ls_rr_sort(const td_ls_rr_sort &RHS) : SPQ(RHS.SPQ) {}
1249 bool operator()(const SUnit* left, const SUnit* right) const;
1251 } // end anonymous namespace
1253 static inline bool isCopyFromLiveIn(const SUnit *SU) {
1254 SDNode *N = SU->Node;
1255 return N && N->getOpcode() == ISD::CopyFromReg &&
1256 N->getOperand(N->getNumOperands()-1).getValueType() != MVT::Flag;
1261 class VISIBILITY_HIDDEN RegReductionPriorityQueue
1262 : public SchedulingPriorityQueue {
1263 std::priority_queue<SUnit*, std::vector<SUnit*>, SF> Queue;
1266 RegReductionPriorityQueue() :
1269 virtual void initNodes(DenseMap<SDNode*, std::vector<SUnit*> > &sumap,
1270 std::vector<SUnit> &sunits) {}
1272 virtual void addNode(const SUnit *SU) {}
1274 virtual void updateNode(const SUnit *SU) {}
1276 virtual void releaseState() {}
1278 virtual unsigned getNodePriority(const SUnit *SU) const {
1282 unsigned size() const { return Queue.size(); }
1284 bool empty() const { return Queue.empty(); }
1286 void push(SUnit *U) {
1289 void push_all(const std::vector<SUnit *> &Nodes) {
1290 for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
1291 Queue.push(Nodes[i]);
1295 if (empty()) return NULL;
1296 SUnit *V = Queue.top();
1301 /// remove - This is a really inefficient way to remove a node from a
1302 /// priority queue. We should roll our own heap to make this better or
1304 void remove(SUnit *SU) {
1305 std::vector<SUnit*> Temp;
1307 assert(!Queue.empty() && "Not in queue!");
1308 while (Queue.top() != SU) {
1309 Temp.push_back(Queue.top());
1311 assert(!Queue.empty() && "Not in queue!");
1314 // Remove the node from the PQ.
1317 // Add all the other nodes back.
1318 for (unsigned i = 0, e = Temp.size(); i != e; ++i)
1319 Queue.push(Temp[i]);
1324 class VISIBILITY_HIDDEN BURegReductionPriorityQueue
1325 : public RegReductionPriorityQueue<SF> {
1326 // SUnitMap SDNode to SUnit mapping (n -> n).
1327 DenseMap<SDNode*, std::vector<SUnit*> > *SUnitMap;
1329 // SUnits - The SUnits for the current graph.
1330 const std::vector<SUnit> *SUnits;
1332 // SethiUllmanNumbers - The SethiUllman number for each node.
1333 std::vector<unsigned> SethiUllmanNumbers;
1335 const TargetInstrInfo *TII;
1336 const TargetRegisterInfo *TRI;
1337 ScheduleDAGRRList *scheduleDAG;
1339 explicit BURegReductionPriorityQueue(const TargetInstrInfo *tii,
1340 const TargetRegisterInfo *tri)
1341 : TII(tii), TRI(tri), scheduleDAG(NULL) {}
1343 void initNodes(DenseMap<SDNode*, std::vector<SUnit*> > &sumap,
1344 std::vector<SUnit> &sunits) {
1347 // Add pseudo dependency edges for two-address nodes.
1348 AddPseudoTwoAddrDeps();
1349 // Calculate node priorities.
1350 CalculateSethiUllmanNumbers();
1353 void addNode(const SUnit *SU) {
1354 SethiUllmanNumbers.resize(SUnits->size(), 0);
1355 CalcNodeSethiUllmanNumber(SU);
1358 void updateNode(const SUnit *SU) {
1359 SethiUllmanNumbers[SU->NodeNum] = 0;
1360 CalcNodeSethiUllmanNumber(SU);
1363 void releaseState() {
1365 SethiUllmanNumbers.clear();
1368 unsigned getNodePriority(const SUnit *SU) const {
1369 assert(SU->NodeNum < SethiUllmanNumbers.size());
1370 unsigned Opc = SU->Node ? SU->Node->getOpcode() : 0;
1371 if (Opc == ISD::CopyFromReg && !isCopyFromLiveIn(SU))
1372 // CopyFromReg should be close to its def because it restricts
1373 // allocation choices. But if it is a livein then perhaps we want it
1374 // closer to its uses so it can be coalesced.
1376 else if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
1377 // CopyToReg should be close to its uses to facilitate coalescing and
1380 else if (Opc == TargetInstrInfo::EXTRACT_SUBREG ||
1381 Opc == TargetInstrInfo::INSERT_SUBREG)
1382 // EXTRACT_SUBREG / INSERT_SUBREG should be close to its use to
1383 // facilitate coalescing.
1385 else if (SU->NumSuccs == 0)
1386 // If SU does not have a use, i.e. it doesn't produce a value that would
1387 // be consumed (e.g. store), then it terminates a chain of computation.
1388 // Give it a large SethiUllman number so it will be scheduled right
1389 // before its predecessors that it doesn't lengthen their live ranges.
1391 else if (SU->NumPreds == 0)
1392 // If SU does not have a def, schedule it close to its uses because it
1393 // does not lengthen any live ranges.
1396 return SethiUllmanNumbers[SU->NodeNum];
1399 void setScheduleDAG(ScheduleDAGRRList *scheduleDag) {
1400 scheduleDAG = scheduleDag;
1404 bool canClobber(const SUnit *SU, const SUnit *Op);
1405 void AddPseudoTwoAddrDeps();
1406 void CalculateSethiUllmanNumbers();
1407 unsigned CalcNodeSethiUllmanNumber(const SUnit *SU);
1412 class VISIBILITY_HIDDEN TDRegReductionPriorityQueue
1413 : public RegReductionPriorityQueue<SF> {
1414 // SUnitMap SDNode to SUnit mapping (n -> n).
1415 DenseMap<SDNode*, std::vector<SUnit*> > *SUnitMap;
1417 // SUnits - The SUnits for the current graph.
1418 const std::vector<SUnit> *SUnits;
1420 // SethiUllmanNumbers - The SethiUllman number for each node.
1421 std::vector<unsigned> SethiUllmanNumbers;
1424 TDRegReductionPriorityQueue() {}
1426 void initNodes(DenseMap<SDNode*, std::vector<SUnit*> > &sumap,
1427 std::vector<SUnit> &sunits) {
1430 // Calculate node priorities.
1431 CalculateSethiUllmanNumbers();
1434 void addNode(const SUnit *SU) {
1435 SethiUllmanNumbers.resize(SUnits->size(), 0);
1436 CalcNodeSethiUllmanNumber(SU);
1439 void updateNode(const SUnit *SU) {
1440 SethiUllmanNumbers[SU->NodeNum] = 0;
1441 CalcNodeSethiUllmanNumber(SU);
1444 void releaseState() {
1446 SethiUllmanNumbers.clear();
1449 unsigned getNodePriority(const SUnit *SU) const {
1450 assert(SU->NodeNum < SethiUllmanNumbers.size());
1451 return SethiUllmanNumbers[SU->NodeNum];
1455 void CalculateSethiUllmanNumbers();
1456 unsigned CalcNodeSethiUllmanNumber(const SUnit *SU);
1460 /// closestSucc - Returns the scheduled cycle of the successor which is
1461 /// closet to the current cycle.
1462 static unsigned closestSucc(const SUnit *SU) {
1463 unsigned MaxCycle = 0;
1464 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1466 unsigned Cycle = I->Dep->Cycle;
1467 // If there are bunch of CopyToRegs stacked up, they should be considered
1468 // to be at the same position.
1469 if (I->Dep->Node && I->Dep->Node->getOpcode() == ISD::CopyToReg)
1470 Cycle = closestSucc(I->Dep)+1;
1471 if (Cycle > MaxCycle)
1477 /// calcMaxScratches - Returns an cost estimate of the worse case requirement
1478 /// for scratch registers. Live-in operands and live-out results don't count
1479 /// since they are "fixed".
1480 static unsigned calcMaxScratches(const SUnit *SU) {
1481 unsigned Scratches = 0;
1482 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1484 if (I->isCtrl) continue; // ignore chain preds
1485 if (!I->Dep->Node || I->Dep->Node->getOpcode() != ISD::CopyFromReg)
1488 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1490 if (I->isCtrl) continue; // ignore chain succs
1491 if (!I->Dep->Node || I->Dep->Node->getOpcode() != ISD::CopyToReg)
1498 bool bu_ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const {
1499 // There used to be a special tie breaker here that looked for
1500 // two-address instructions and preferred the instruction with a
1501 // def&use operand. The special case triggered diagnostics when
1502 // _GLIBCXX_DEBUG was enabled because it broke the strict weak
1503 // ordering that priority_queue requires. It didn't help much anyway
1504 // because AddPseudoTwoAddrDeps already covers many of the cases
1505 // where it would have applied. In addition, it's counter-intuitive
1506 // that a tie breaker would be the first thing attempted. There's a
1507 // "real" tie breaker below that is the operation of last resort.
1508 // The fact that the "special tie breaker" would trigger when there
1509 // wasn't otherwise a tie is what broke the strict weak ordering
1512 unsigned LPriority = SPQ->getNodePriority(left);
1513 unsigned RPriority = SPQ->getNodePriority(right);
1514 if (LPriority != RPriority)
1515 return LPriority > RPriority;
1517 // Try schedule def + use closer when Sethi-Ullman numbers are the same.
1522 // and the following instructions are both ready.
1526 // Then schedule t2 = op first.
1533 // This creates more short live intervals.
1534 unsigned LDist = closestSucc(left);
1535 unsigned RDist = closestSucc(right);
1537 return LDist < RDist;
1539 // Intuitively, it's good to push down instructions whose results are
1540 // liveout so their long live ranges won't conflict with other values
1541 // which are needed inside the BB. Further prioritize liveout instructions
1542 // by the number of operands which are calculated within the BB.
1543 unsigned LScratch = calcMaxScratches(left);
1544 unsigned RScratch = calcMaxScratches(right);
1545 if (LScratch != RScratch)
1546 return LScratch > RScratch;
1548 if (left->Height != right->Height)
1549 return left->Height > right->Height;
1551 if (left->Depth != right->Depth)
1552 return left->Depth < right->Depth;
1554 if (left->CycleBound != right->CycleBound)
1555 return left->CycleBound > right->CycleBound;
1557 // FIXME: No strict ordering.
1561 template<class SF> bool
1562 BURegReductionPriorityQueue<SF>::canClobber(const SUnit *SU, const SUnit *Op) {
1563 if (SU->isTwoAddress) {
1564 unsigned Opc = SU->Node->getTargetOpcode();
1565 const TargetInstrDesc &TID = TII->get(Opc);
1566 unsigned NumRes = TID.getNumDefs();
1567 unsigned NumOps = TID.getNumOperands() - NumRes;
1568 for (unsigned i = 0; i != NumOps; ++i) {
1569 if (TID.getOperandConstraint(i+NumRes, TOI::TIED_TO) != -1) {
1570 SDNode *DU = SU->Node->getOperand(i).Val;
1571 if ((*SUnitMap).find(DU) != (*SUnitMap).end() &&
1572 Op == (*SUnitMap)[DU][SU->InstanceNo])
1581 /// hasCopyToRegUse - Return true if SU has a value successor that is a
1583 static bool hasCopyToRegUse(SUnit *SU) {
1584 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1586 if (I->isCtrl) continue;
1587 SUnit *SuccSU = I->Dep;
1588 if (SuccSU->Node && SuccSU->Node->getOpcode() == ISD::CopyToReg)
1594 /// canClobberPhysRegDefs - True if SU would clobber one of SuccSU's
1595 /// physical register def.
1596 static bool canClobberPhysRegDefs(SUnit *SuccSU, SUnit *SU,
1597 const TargetInstrInfo *TII,
1598 const TargetRegisterInfo *TRI) {
1599 SDNode *N = SuccSU->Node;
1600 unsigned NumDefs = TII->get(N->getTargetOpcode()).getNumDefs();
1601 const unsigned *ImpDefs = TII->get(N->getTargetOpcode()).getImplicitDefs();
1604 const unsigned *SUImpDefs =
1605 TII->get(SU->Node->getTargetOpcode()).getImplicitDefs();
1608 for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
1609 MVT::ValueType VT = N->getValueType(i);
1610 if (VT == MVT::Flag || VT == MVT::Other)
1612 unsigned Reg = ImpDefs[i - NumDefs];
1613 for (;*SUImpDefs; ++SUImpDefs) {
1614 unsigned SUReg = *SUImpDefs;
1615 if (TRI->regsOverlap(Reg, SUReg))
1622 /// AddPseudoTwoAddrDeps - If two nodes share an operand and one of them uses
1623 /// it as a def&use operand. Add a pseudo control edge from it to the other
1624 /// node (if it won't create a cycle) so the two-address one will be scheduled
1625 /// first (lower in the schedule). If both nodes are two-address, favor the
1626 /// one that has a CopyToReg use (more likely to be a loop induction update).
1627 /// If both are two-address, but one is commutable while the other is not
1628 /// commutable, favor the one that's not commutable.
1630 void BURegReductionPriorityQueue<SF>::AddPseudoTwoAddrDeps() {
1631 for (unsigned i = 0, e = SUnits->size(); i != e; ++i) {
1632 SUnit *SU = (SUnit *)&((*SUnits)[i]);
1633 if (!SU->isTwoAddress)
1636 SDNode *Node = SU->Node;
1637 if (!Node || !Node->isTargetOpcode() || SU->FlaggedNodes.size() > 0)
1640 unsigned Opc = Node->getTargetOpcode();
1641 const TargetInstrDesc &TID = TII->get(Opc);
1642 unsigned NumRes = TID.getNumDefs();
1643 unsigned NumOps = TID.getNumOperands() - NumRes;
1644 for (unsigned j = 0; j != NumOps; ++j) {
1645 if (TID.getOperandConstraint(j+NumRes, TOI::TIED_TO) != -1) {
1646 SDNode *DU = SU->Node->getOperand(j).Val;
1647 if ((*SUnitMap).find(DU) == (*SUnitMap).end())
1649 SUnit *DUSU = (*SUnitMap)[DU][SU->InstanceNo];
1650 if (!DUSU) continue;
1651 for (SUnit::succ_iterator I = DUSU->Succs.begin(),E = DUSU->Succs.end();
1653 if (I->isCtrl) continue;
1654 SUnit *SuccSU = I->Dep;
1657 // Be conservative. Ignore if nodes aren't at roughly the same
1658 // depth and height.
1659 if (SuccSU->Height < SU->Height && (SU->Height - SuccSU->Height) > 1)
1661 if (!SuccSU->Node || !SuccSU->Node->isTargetOpcode())
1663 // Don't constrain nodes with physical register defs if the
1664 // predecessor can clobber them.
1665 if (SuccSU->hasPhysRegDefs) {
1666 if (canClobberPhysRegDefs(SuccSU, SU, TII, TRI))
1669 // Don't constraint extract_subreg / insert_subreg these may be
1670 // coalesced away. We don't them close to their uses.
1671 unsigned SuccOpc = SuccSU->Node->getTargetOpcode();
1672 if (SuccOpc == TargetInstrInfo::EXTRACT_SUBREG ||
1673 SuccOpc == TargetInstrInfo::INSERT_SUBREG)
1675 if ((!canClobber(SuccSU, DUSU) ||
1676 (hasCopyToRegUse(SU) && !hasCopyToRegUse(SuccSU)) ||
1677 (!SU->isCommutable && SuccSU->isCommutable)) &&
1678 !scheduleDAG->IsReachable(SuccSU, SU)) {
1679 DOUT << "Adding an edge from SU # " << SU->NodeNum
1680 << " to SU #" << SuccSU->NodeNum << "\n";
1681 scheduleDAG->AddPred(SU, SuccSU, true, true);
1689 /// CalcNodeSethiUllmanNumber - Priority is the Sethi Ullman number.
1690 /// Smaller number is the higher priority.
1692 unsigned BURegReductionPriorityQueue<SF>::
1693 CalcNodeSethiUllmanNumber(const SUnit *SU) {
1694 unsigned &SethiUllmanNumber = SethiUllmanNumbers[SU->NodeNum];
1695 if (SethiUllmanNumber != 0)
1696 return SethiUllmanNumber;
1699 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1701 if (I->isCtrl) continue; // ignore chain preds
1702 SUnit *PredSU = I->Dep;
1703 unsigned PredSethiUllman = CalcNodeSethiUllmanNumber(PredSU);
1704 if (PredSethiUllman > SethiUllmanNumber) {
1705 SethiUllmanNumber = PredSethiUllman;
1707 } else if (PredSethiUllman == SethiUllmanNumber && !I->isCtrl)
1711 SethiUllmanNumber += Extra;
1713 if (SethiUllmanNumber == 0)
1714 SethiUllmanNumber = 1;
1716 return SethiUllmanNumber;
1719 /// CalculateSethiUllmanNumbers - Calculate Sethi-Ullman numbers of all
1720 /// scheduling units.
1722 void BURegReductionPriorityQueue<SF>::CalculateSethiUllmanNumbers() {
1723 SethiUllmanNumbers.assign(SUnits->size(), 0);
1725 for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
1726 CalcNodeSethiUllmanNumber(&(*SUnits)[i]);
1729 /// LimitedSumOfUnscheduledPredsOfSuccs - Compute the sum of the unscheduled
1730 /// predecessors of the successors of the SUnit SU. Stop when the provided
1731 /// limit is exceeded.
1732 static unsigned LimitedSumOfUnscheduledPredsOfSuccs(const SUnit *SU,
1735 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1737 SUnit *SuccSU = I->Dep;
1738 for (SUnit::const_pred_iterator II = SuccSU->Preds.begin(),
1739 EE = SuccSU->Preds.end(); II != EE; ++II) {
1740 SUnit *PredSU = II->Dep;
1741 if (!PredSU->isScheduled)
1751 bool td_ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const {
1752 unsigned LPriority = SPQ->getNodePriority(left);
1753 unsigned RPriority = SPQ->getNodePriority(right);
1754 bool LIsTarget = left->Node && left->Node->isTargetOpcode();
1755 bool RIsTarget = right->Node && right->Node->isTargetOpcode();
1756 bool LIsFloater = LIsTarget && left->NumPreds == 0;
1757 bool RIsFloater = RIsTarget && right->NumPreds == 0;
1758 unsigned LBonus = (LimitedSumOfUnscheduledPredsOfSuccs(left,1) == 1) ? 2 : 0;
1759 unsigned RBonus = (LimitedSumOfUnscheduledPredsOfSuccs(right,1) == 1) ? 2 : 0;
1761 if (left->NumSuccs == 0 && right->NumSuccs != 0)
1763 else if (left->NumSuccs != 0 && right->NumSuccs == 0)
1770 if (left->NumSuccs == 1)
1772 if (right->NumSuccs == 1)
1775 if (LPriority+LBonus != RPriority+RBonus)
1776 return LPriority+LBonus < RPriority+RBonus;
1778 if (left->Depth != right->Depth)
1779 return left->Depth < right->Depth;
1781 if (left->NumSuccsLeft != right->NumSuccsLeft)
1782 return left->NumSuccsLeft > right->NumSuccsLeft;
1784 if (left->CycleBound != right->CycleBound)
1785 return left->CycleBound > right->CycleBound;
1787 // FIXME: No strict ordering.
1791 /// CalcNodeSethiUllmanNumber - Priority is the Sethi Ullman number.
1792 /// Smaller number is the higher priority.
1794 unsigned TDRegReductionPriorityQueue<SF>::
1795 CalcNodeSethiUllmanNumber(const SUnit *SU) {
1796 unsigned &SethiUllmanNumber = SethiUllmanNumbers[SU->NodeNum];
1797 if (SethiUllmanNumber != 0)
1798 return SethiUllmanNumber;
1800 unsigned Opc = SU->Node ? SU->Node->getOpcode() : 0;
1801 if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
1802 SethiUllmanNumber = 0xffff;
1803 else if (SU->NumSuccsLeft == 0)
1804 // If SU does not have a use, i.e. it doesn't produce a value that would
1805 // be consumed (e.g. store), then it terminates a chain of computation.
1806 // Give it a small SethiUllman number so it will be scheduled right before
1807 // its predecessors that it doesn't lengthen their live ranges.
1808 SethiUllmanNumber = 0;
1809 else if (SU->NumPredsLeft == 0 &&
1810 (Opc != ISD::CopyFromReg || isCopyFromLiveIn(SU)))
1811 SethiUllmanNumber = 0xffff;
1814 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1816 if (I->isCtrl) continue; // ignore chain preds
1817 SUnit *PredSU = I->Dep;
1818 unsigned PredSethiUllman = CalcNodeSethiUllmanNumber(PredSU);
1819 if (PredSethiUllman > SethiUllmanNumber) {
1820 SethiUllmanNumber = PredSethiUllman;
1822 } else if (PredSethiUllman == SethiUllmanNumber && !I->isCtrl)
1826 SethiUllmanNumber += Extra;
1829 return SethiUllmanNumber;
1832 /// CalculateSethiUllmanNumbers - Calculate Sethi-Ullman numbers of all
1833 /// scheduling units.
1835 void TDRegReductionPriorityQueue<SF>::CalculateSethiUllmanNumbers() {
1836 SethiUllmanNumbers.assign(SUnits->size(), 0);
1838 for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
1839 CalcNodeSethiUllmanNumber(&(*SUnits)[i]);
1842 //===----------------------------------------------------------------------===//
1843 // Public Constructor Functions
1844 //===----------------------------------------------------------------------===//
1846 llvm::ScheduleDAG* llvm::createBURRListDAGScheduler(SelectionDAGISel *IS,
1848 MachineBasicBlock *BB) {
1849 const TargetInstrInfo *TII = DAG->getTarget().getInstrInfo();
1850 const TargetRegisterInfo *TRI = DAG->getTarget().getRegisterInfo();
1852 BURegReductionPriorityQueue<bu_ls_rr_sort> *priorityQueue =
1853 new BURegReductionPriorityQueue<bu_ls_rr_sort>(TII, TRI);
1855 ScheduleDAGRRList * scheduleDAG =
1856 new ScheduleDAGRRList(*DAG, BB, DAG->getTarget(), true, priorityQueue);
1857 priorityQueue->setScheduleDAG(scheduleDAG);
1861 llvm::ScheduleDAG* llvm::createTDRRListDAGScheduler(SelectionDAGISel *IS,
1863 MachineBasicBlock *BB) {
1864 return new ScheduleDAGRRList(*DAG, BB, DAG->getTarget(), false,
1865 new TDRegReductionPriorityQueue<td_ls_rr_sort>());