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 pred.
94 /// Updates the topological oredering 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 predecessors of
99 /// the current node M. Updates the topological oredering 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 oredering if required.
123 SUnit *CreateNewSUnit(SDNode *N) {
124 SUnit *NewNode = NewSUnit(N);
125 // Update the topologic ordering
126 if (NewNode->NodeNum >= Node2Index.size())
127 InitDAGTopologicalSorting();
131 /// CreateClone - Creates a new SUnit from old one.
132 /// Updates the topological oredering if required.
133 SUnit *CreateClone(SUnit *N) {
134 SUnit *NewNode = Clone(N);
135 // Update the topologic 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 for 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 is 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 fall 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 Algorithm first computes a topological ordering for a 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 should later get new topological indexes
181 /// by means of 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 a 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()-1; i != 0; --i) { // Ignore first node.
243 SUnit *SU = Sequence[i];
244 if (!SU || !SU->Node) continue;
245 if (SU->isCommutable) {
246 unsigned Opc = SU->Node->getTargetOpcode();
247 const TargetInstrDesc &TID = TII->get(Opc);
248 unsigned NumRes = TID.getNumDefs();
249 unsigned NumOps = TID.getNumOperands() - NumRes;
250 for (unsigned j = 0; j != NumOps; ++j) {
251 if (TID.getOperandConstraint(j+NumRes, TOI::TIED_TO) == -1)
254 SDNode *OpN = SU->Node->getOperand(j).Val;
255 SUnit *OpSU = isPassiveNode(OpN) ? NULL : SUnitMap[OpN][SU->InstanceNo];
256 if (OpSU && OperandSeen.count(OpSU) == 1) {
257 // Ok, so SU is not the last use of OpSU, but SU is two-address so
258 // it will clobber OpSU. Try to commute SU if no other source operands
260 bool DoCommute = true;
261 for (unsigned k = 0; k < NumOps; ++k) {
263 OpN = SU->Node->getOperand(k).Val;
264 OpSU = isPassiveNode(OpN) ? NULL : SUnitMap[OpN][SU->InstanceNo];
265 if (OpSU && OperandSeen.count(OpSU) == 1) {
272 CommuteSet.insert(SU->Node);
275 // Only look at the first use&def node for now.
280 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
283 OperandSeen.insert(I->Dep);
288 //===----------------------------------------------------------------------===//
289 // Bottom-Up Scheduling
290 //===----------------------------------------------------------------------===//
292 /// ReleasePred - Decrement the NumSuccsLeft count of a predecessor. Add it to
293 /// the AvailableQueue if the count reaches zero. Also update its cycle bound.
294 void ScheduleDAGRRList::ReleasePred(SUnit *PredSU, bool isChain,
296 // FIXME: the distance between two nodes is not always == the predecessor's
297 // latency. For example, the reader can very well read the register written
298 // by the predecessor later than the issue cycle. It also depends on the
299 // interrupt model (drain vs. freeze).
300 PredSU->CycleBound = std::max(PredSU->CycleBound, CurCycle + PredSU->Latency);
302 --PredSU->NumSuccsLeft;
305 if (PredSU->NumSuccsLeft < 0) {
306 cerr << "*** List scheduling failed! ***\n";
308 cerr << " has been released too many times!\n";
313 if (PredSU->NumSuccsLeft == 0) {
314 // EntryToken has to go last! Special case it here.
315 if (!PredSU->Node || PredSU->Node->getOpcode() != ISD::EntryToken) {
316 PredSU->isAvailable = true;
317 AvailableQueue->push(PredSU);
322 /// ScheduleNodeBottomUp - Add the node to the schedule. Decrement the pending
323 /// count of its predecessors. If a predecessor pending count is zero, add it to
324 /// the Available queue.
325 void ScheduleDAGRRList::ScheduleNodeBottomUp(SUnit *SU, unsigned CurCycle) {
326 DOUT << "*** Scheduling [" << CurCycle << "]: ";
327 DEBUG(SU->dump(&DAG));
328 SU->Cycle = CurCycle;
330 AvailableQueue->ScheduledNode(SU);
332 // Bottom up: release predecessors
333 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
335 ReleasePred(I->Dep, I->isCtrl, CurCycle);
337 // This is a physical register dependency and it's impossible or
338 // expensive to copy the register. Make sure nothing that can
339 // clobber the register is scheduled between the predecessor and
341 if (LiveRegs.insert(I->Reg)) {
342 LiveRegDefs[I->Reg] = I->Dep;
343 LiveRegCycles[I->Reg] = CurCycle;
348 // Release all the implicit physical register defs that are live.
349 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
352 if (LiveRegCycles[I->Reg] == I->Dep->Cycle) {
353 LiveRegs.erase(I->Reg);
354 assert(LiveRegDefs[I->Reg] == SU &&
355 "Physical register dependency violated?");
356 LiveRegDefs[I->Reg] = NULL;
357 LiveRegCycles[I->Reg] = 0;
362 SU->isScheduled = true;
365 /// CapturePred - This does the opposite of ReleasePred. Since SU is being
366 /// unscheduled, incrcease the succ left count of its predecessors. Remove
367 /// them from AvailableQueue if necessary.
368 void ScheduleDAGRRList::CapturePred(SUnit *PredSU, SUnit *SU, bool isChain) {
369 PredSU->CycleBound = 0;
370 for (SUnit::succ_iterator I = PredSU->Succs.begin(), E = PredSU->Succs.end();
374 PredSU->CycleBound = std::max(PredSU->CycleBound,
375 I->Dep->Cycle + PredSU->Latency);
378 if (PredSU->isAvailable) {
379 PredSU->isAvailable = false;
380 if (!PredSU->isPending)
381 AvailableQueue->remove(PredSU);
384 ++PredSU->NumSuccsLeft;
387 /// UnscheduleNodeBottomUp - Remove the node from the schedule, update its and
388 /// its predecessor states to reflect the change.
389 void ScheduleDAGRRList::UnscheduleNodeBottomUp(SUnit *SU) {
390 DOUT << "*** Unscheduling [" << SU->Cycle << "]: ";
391 DEBUG(SU->dump(&DAG));
393 AvailableQueue->UnscheduledNode(SU);
395 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
397 CapturePred(I->Dep, SU, I->isCtrl);
398 if (I->Cost < 0 && SU->Cycle == LiveRegCycles[I->Reg]) {
399 LiveRegs.erase(I->Reg);
400 assert(LiveRegDefs[I->Reg] == I->Dep &&
401 "Physical register dependency violated?");
402 LiveRegDefs[I->Reg] = NULL;
403 LiveRegCycles[I->Reg] = 0;
407 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
410 if (LiveRegs.insert(I->Reg)) {
411 assert(!LiveRegDefs[I->Reg] &&
412 "Physical register dependency violated?");
413 LiveRegDefs[I->Reg] = SU;
415 if (I->Dep->Cycle < LiveRegCycles[I->Reg])
416 LiveRegCycles[I->Reg] = I->Dep->Cycle;
421 SU->isScheduled = false;
422 SU->isAvailable = true;
423 AvailableQueue->push(SU);
426 /// IsReachable - Checks if SU is reachable from TargetSU.
427 bool ScheduleDAGRRList::IsReachable(SUnit *SU, SUnit *TargetSU) {
428 // If insertion of the edge SU->TargetSU would creates a cycle
429 // then there is a path from TargetSU to SU
430 int UpperBound, LowerBound;
431 LowerBound = Node2Index[TargetSU->NodeNum];
432 UpperBound = Node2Index[SU->NodeNum];
433 bool HasLoop = false;
434 // Is Ord(TargetSU) < Ord(SU) ?
435 if (LowerBound < UpperBound) {
437 // There may be a path from TargetSU to SU. Check for it.
438 DFS(TargetSU, UpperBound, HasLoop);
443 /// Allocate - assign the topological index to a node n
444 inline void ScheduleDAGRRList::Allocate(int n, int index) {
445 Node2Index[n] = index;
446 Index2Node[index] = n;
449 /// InitDAGTopologicalSorting - create the initial topological
450 /// ordering from the DAG to be scheduled.
451 void ScheduleDAGRRList::InitDAGTopologicalSorting() {
452 unsigned DAGSize = SUnits.size();
453 std::vector<unsigned> InDegree(DAGSize);
454 std::vector<SUnit*> WorkList;
455 WorkList.reserve(DAGSize);
456 std::vector<SUnit*> TopOrder;
457 TopOrder.reserve(DAGSize);
459 // Initialize the data structures
460 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
461 SUnit *SU = &SUnits[i];
462 int NodeNum = SU->NodeNum;
463 unsigned Degree = SU->Succs.size();
464 InDegree[NodeNum] = Degree;
466 // Is it a node without dependencies?
468 assert(SU->Succs.empty() && "SUnit should have no successors");
469 // Collect leaf nodes
470 WorkList.push_back(SU);
474 while (!WorkList.empty()) {
475 SUnit *SU = WorkList.back();
477 TopOrder.push_back(SU);
478 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
481 if (!--InDegree[SU->NodeNum])
482 // If all dependencies of the node are processed already,
483 // then the node can be computed now
484 WorkList.push_back(SU);
488 // Second pass, assign the actual topological order as node ids.
493 Index2Node.resize(DAGSize);
494 Node2Index.resize(DAGSize);
495 Visited.resize(DAGSize);
497 for (std::vector<SUnit*>::reverse_iterator TI = TopOrder.rbegin(),
498 TE = TopOrder.rend();TI != TE; ++TI) {
499 Allocate((*TI)->NodeNum, Id);
504 // Check correctness of the ordering
505 for (unsigned i = 0, e = DAGSize; i != e; ++i) {
506 SUnit *SU = &SUnits[i];
507 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
509 assert(Node2Index[SU->NodeNum] > Node2Index[I->Dep->NodeNum] &&
510 "Wrong topological sorting");
516 /// AddPred - adds edge from SUnit X to SUnit Y
517 /// Updates the topological oredering if required.
518 bool ScheduleDAGRRList::AddPred(SUnit *Y, SUnit *X, bool isCtrl, bool isSpecial,
519 unsigned PhyReg, int Cost) {
520 int UpperBound, LowerBound;
521 LowerBound = Node2Index[Y->NodeNum];
522 UpperBound = Node2Index[X->NodeNum];
523 bool HasLoop = false;
524 // Is Ord(X) < Ord(Y) ?
525 if (LowerBound < UpperBound) {
526 // Update the topological order
528 DFS(Y, UpperBound, HasLoop);
529 assert(!HasLoop && "Inserted edge creates a loop!");
530 // Recompute topological indexes
531 Shift(Visited, LowerBound, UpperBound);
533 // Now really insert the edge
534 return Y->addPred(X,isCtrl,isSpecial,PhyReg,Cost);
537 /// RemovePred - This removes the specified node N from preds of
538 /// the current node M. Updates the topological oredering if required
539 bool ScheduleDAGRRList::RemovePred(SUnit *M, SUnit *N,
540 bool isCtrl, bool isSpecial) {
541 // InitDAGTopologicalSorting();
542 return M->removePred(N, isCtrl, isSpecial);
545 /// DFS - make a DFS traversal to mark all nodes reachable from SU and and mark /// all nodes affected by the edge insertion. These nodes should later get new /// 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 is 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));
736 RemovePred(SU, ChainPred, true, false);
738 AddPred(LoadSU,ChainPred, true, false);
740 for (unsigned i = 0, e = LoadPreds.size(); i != e; ++i) {
741 SDep *Pred = &LoadPreds[i];
742 RemovePred(SU, Pred->Dep, Pred->isCtrl, Pred->isSpecial);
744 AddPred(LoadSU, Pred->Dep, Pred->isCtrl, Pred->isSpecial,
745 Pred->Reg, Pred->Cost);
748 for (unsigned i = 0, e = NodePreds.size(); i != e; ++i) {
749 SDep *Pred = &NodePreds[i];
750 RemovePred(SU, Pred->Dep, Pred->isCtrl, Pred->isSpecial);
751 AddPred(NewSU, Pred->Dep, Pred->isCtrl, Pred->isSpecial,
752 Pred->Reg, Pred->Cost);
754 for (unsigned i = 0, e = NodeSuccs.size(); i != e; ++i) {
755 SDep *Succ = &NodeSuccs[i];
756 RemovePred(Succ->Dep, SU, Succ->isCtrl, Succ->isSpecial);
757 AddPred(Succ->Dep, NewSU, Succ->isCtrl, Succ->isSpecial,
758 Succ->Reg, Succ->Cost);
760 for (unsigned i = 0, e = ChainSuccs.size(); i != e; ++i) {
761 SDep *Succ = &ChainSuccs[i];
762 RemovePred(Succ->Dep, SU, Succ->isCtrl, Succ->isSpecial);
764 AddPred(Succ->Dep, LoadSU, Succ->isCtrl, Succ->isSpecial,
765 Succ->Reg, Succ->Cost);
769 AddPred(NewSU, LoadSU, false, false);
773 AvailableQueue->addNode(LoadSU);
774 AvailableQueue->addNode(NewSU);
778 if (NewSU->NumSuccsLeft == 0) {
779 NewSU->isAvailable = true;
785 DOUT << "Duplicating SU # " << SU->NodeNum << "\n";
786 NewSU = CreateClone(SU);
788 // New SUnit has the exact same predecessors.
789 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
792 AddPred(NewSU, I->Dep, I->isCtrl, false, I->Reg, I->Cost);
793 NewSU->Depth = std::max(NewSU->Depth, I->Dep->Depth+1);
796 // Only copy scheduled successors. Cut them from old node's successor
797 // list and move them over.
798 SmallVector<std::pair<SUnit*, bool>, 4> DelDeps;
799 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
803 if (I->Dep->isScheduled) {
804 NewSU->Height = std::max(NewSU->Height, I->Dep->Height+1);
805 AddPred(I->Dep, NewSU, I->isCtrl, false, I->Reg, I->Cost);
806 DelDeps.push_back(std::make_pair(I->Dep, I->isCtrl));
809 for (unsigned i = 0, e = DelDeps.size(); i != e; ++i) {
810 SUnit *Succ = DelDeps[i].first;
811 bool isCtrl = DelDeps[i].second;
812 RemovePred(Succ, SU, isCtrl, false);
815 AvailableQueue->updateNode(SU);
816 AvailableQueue->addNode(NewSU);
822 /// InsertCCCopiesAndMoveSuccs - Insert expensive cross register class copies
823 /// and move all scheduled successors of the given SUnit to the last copy.
824 void ScheduleDAGRRList::InsertCCCopiesAndMoveSuccs(SUnit *SU, unsigned Reg,
825 const TargetRegisterClass *DestRC,
826 const TargetRegisterClass *SrcRC,
827 SmallVector<SUnit*, 2> &Copies) {
828 SUnit *CopyFromSU = CreateNewSUnit(NULL);
829 CopyFromSU->CopySrcRC = SrcRC;
830 CopyFromSU->CopyDstRC = DestRC;
831 CopyFromSU->Depth = SU->Depth;
832 CopyFromSU->Height = SU->Height;
834 SUnit *CopyToSU = CreateNewSUnit(NULL);
835 CopyToSU->CopySrcRC = DestRC;
836 CopyToSU->CopyDstRC = SrcRC;
838 // Only copy scheduled successors. Cut them from old node's successor
839 // list and move them over.
840 SmallVector<std::pair<SUnit*, bool>, 4> DelDeps;
841 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
845 if (I->Dep->isScheduled) {
846 CopyToSU->Height = std::max(CopyToSU->Height, I->Dep->Height+1);
847 AddPred(I->Dep, CopyToSU, I->isCtrl, false, I->Reg, I->Cost);
848 DelDeps.push_back(std::make_pair(I->Dep, I->isCtrl));
851 for (unsigned i = 0, e = DelDeps.size(); i != e; ++i) {
852 SUnit *Succ = DelDeps[i].first;
853 bool isCtrl = DelDeps[i].second;
854 RemovePred(Succ, SU, isCtrl, false);
857 AddPred(CopyFromSU, SU, false, false, Reg, -1);
858 AddPred(CopyToSU, CopyFromSU, false, false, Reg, 1);
860 AvailableQueue->updateNode(SU);
861 AvailableQueue->addNode(CopyFromSU);
862 AvailableQueue->addNode(CopyToSU);
863 Copies.push_back(CopyFromSU);
864 Copies.push_back(CopyToSU);
869 /// getPhysicalRegisterVT - Returns the ValueType of the physical register
870 /// definition of the specified node.
871 /// FIXME: Move to SelectionDAG?
872 static MVT::ValueType getPhysicalRegisterVT(SDNode *N, unsigned Reg,
873 const TargetInstrInfo *TII) {
874 const TargetInstrDesc &TID = TII->get(N->getTargetOpcode());
875 assert(TID.ImplicitDefs && "Physical reg def must be in implicit def list!");
876 unsigned NumRes = TID.getNumDefs();
877 for (const unsigned *ImpDef = TID.getImplicitDefs(); *ImpDef; ++ImpDef) {
882 return N->getValueType(NumRes);
885 /// DelayForLiveRegsBottomUp - Returns true if it is necessary to delay
886 /// scheduling of the given node to satisfy live physical register dependencies.
887 /// If the specific node is the last one that's available to schedule, do
888 /// whatever is necessary (i.e. backtracking or cloning) to make it possible.
889 bool ScheduleDAGRRList::DelayForLiveRegsBottomUp(SUnit *SU,
890 SmallVector<unsigned, 4> &LRegs){
891 if (LiveRegs.empty())
894 SmallSet<unsigned, 4> RegAdded;
895 // If this node would clobber any "live" register, then it's not ready.
896 for (SUnit::pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
899 unsigned Reg = I->Reg;
900 if (LiveRegs.count(Reg) && LiveRegDefs[Reg] != I->Dep) {
901 if (RegAdded.insert(Reg))
902 LRegs.push_back(Reg);
904 for (const unsigned *Alias = TRI->getAliasSet(Reg);
906 if (LiveRegs.count(*Alias) && LiveRegDefs[*Alias] != I->Dep) {
907 if (RegAdded.insert(*Alias))
908 LRegs.push_back(*Alias);
913 for (unsigned i = 0, e = SU->FlaggedNodes.size()+1; i != e; ++i) {
914 SDNode *Node = (i == 0) ? SU->Node : SU->FlaggedNodes[i-1];
915 if (!Node || !Node->isTargetOpcode())
917 const TargetInstrDesc &TID = TII->get(Node->getTargetOpcode());
918 if (!TID.ImplicitDefs)
920 for (const unsigned *Reg = TID.ImplicitDefs; *Reg; ++Reg) {
921 if (LiveRegs.count(*Reg) && LiveRegDefs[*Reg] != SU) {
922 if (RegAdded.insert(*Reg))
923 LRegs.push_back(*Reg);
925 for (const unsigned *Alias = TRI->getAliasSet(*Reg);
927 if (LiveRegs.count(*Alias) && LiveRegDefs[*Alias] != SU) {
928 if (RegAdded.insert(*Alias))
929 LRegs.push_back(*Alias);
933 return !LRegs.empty();
937 /// ListScheduleBottomUp - The main loop of list scheduling for bottom-up
939 void ScheduleDAGRRList::ListScheduleBottomUp() {
940 unsigned CurCycle = 0;
941 // Add root to Available queue.
942 SUnit *RootSU = SUnitMap[DAG.getRoot().Val].front();
943 RootSU->isAvailable = true;
944 AvailableQueue->push(RootSU);
946 // While Available queue is not empty, grab the node with the highest
947 // priority. If it is not ready put it back. Schedule the node.
948 SmallVector<SUnit*, 4> NotReady;
949 while (!AvailableQueue->empty()) {
950 bool Delayed = false;
951 DenseMap<SUnit*, SmallVector<unsigned, 4> > LRegsMap;
952 SUnit *CurSU = AvailableQueue->pop();
954 if (CurSU->CycleBound <= CurCycle) {
955 SmallVector<unsigned, 4> LRegs;
956 if (!DelayForLiveRegsBottomUp(CurSU, LRegs))
959 LRegsMap.insert(std::make_pair(CurSU, LRegs));
962 CurSU->isPending = true; // This SU is not in AvailableQueue right now.
963 NotReady.push_back(CurSU);
964 CurSU = AvailableQueue->pop();
967 // All candidates are delayed due to live physical reg dependencies.
968 // Try backtracking, code duplication, or inserting cross class copies
970 if (Delayed && !CurSU) {
971 for (unsigned i = 0, e = NotReady.size(); i != e; ++i) {
972 SUnit *TrySU = NotReady[i];
973 SmallVector<unsigned, 4> &LRegs = LRegsMap[TrySU];
975 // Try unscheduling up to the point where it's safe to schedule
977 unsigned LiveCycle = CurCycle;
978 for (unsigned j = 0, ee = LRegs.size(); j != ee; ++j) {
979 unsigned Reg = LRegs[j];
980 unsigned LCycle = LiveRegCycles[Reg];
981 LiveCycle = std::min(LiveCycle, LCycle);
983 SUnit *OldSU = Sequence[LiveCycle];
984 if (!WillCreateCycle(TrySU, OldSU)) {
985 BacktrackBottomUp(TrySU, LiveCycle, CurCycle);
986 // Force the current node to be scheduled before the node that
987 // requires the physical reg dep.
988 if (OldSU->isAvailable) {
989 OldSU->isAvailable = false;
990 AvailableQueue->remove(OldSU);
992 AddPred(TrySU, OldSU, true, true);
993 // If one or more successors has been unscheduled, then the current
994 // node is no longer avaialable. Schedule a successor that's now
995 // available instead.
996 if (!TrySU->isAvailable)
997 CurSU = AvailableQueue->pop();
1000 TrySU->isPending = false;
1001 NotReady.erase(NotReady.begin()+i);
1008 // Can't backtrack. Try duplicating the nodes that produces these
1009 // "expensive to copy" values to break the dependency. In case even
1010 // that doesn't work, insert cross class copies.
1011 SUnit *TrySU = NotReady[0];
1012 SmallVector<unsigned, 4> &LRegs = LRegsMap[TrySU];
1013 assert(LRegs.size() == 1 && "Can't handle this yet!");
1014 unsigned Reg = LRegs[0];
1015 SUnit *LRDef = LiveRegDefs[Reg];
1016 SUnit *NewDef = CopyAndMoveSuccessors(LRDef);
1018 // Issue expensive cross register class copies.
1019 MVT::ValueType VT = getPhysicalRegisterVT(LRDef->Node, Reg, TII);
1020 const TargetRegisterClass *RC =
1021 TRI->getPhysicalRegisterRegClass(Reg, VT);
1022 const TargetRegisterClass *DestRC = TRI->getCrossCopyRegClass(RC);
1024 assert(false && "Don't know how to copy this physical register!");
1027 SmallVector<SUnit*, 2> Copies;
1028 InsertCCCopiesAndMoveSuccs(LRDef, Reg, DestRC, RC, Copies);
1029 DOUT << "Adding an edge from SU # " << TrySU->NodeNum
1030 << " to SU #" << Copies.front()->NodeNum << "\n";
1031 AddPred(TrySU, Copies.front(), true, true);
1032 NewDef = Copies.back();
1035 DOUT << "Adding an edge from SU # " << NewDef->NodeNum
1036 << " to SU #" << TrySU->NodeNum << "\n";
1037 LiveRegDefs[Reg] = NewDef;
1038 AddPred(NewDef, TrySU, true, true);
1039 TrySU->isAvailable = false;
1044 assert(false && "Unable to resolve live physical register dependencies!");
1049 // Add the nodes that aren't ready back onto the available list.
1050 for (unsigned i = 0, e = NotReady.size(); i != e; ++i) {
1051 NotReady[i]->isPending = false;
1052 // May no longer be available due to backtracking.
1053 if (NotReady[i]->isAvailable)
1054 AvailableQueue->push(NotReady[i]);
1059 Sequence.push_back(0);
1061 ScheduleNodeBottomUp(CurSU, CurCycle);
1062 Sequence.push_back(CurSU);
1067 // Add entry node last
1068 if (DAG.getEntryNode().Val != DAG.getRoot().Val) {
1069 SUnit *Entry = SUnitMap[DAG.getEntryNode().Val].front();
1070 Sequence.push_back(Entry);
1073 // Reverse the order if it is bottom up.
1074 std::reverse(Sequence.begin(), Sequence.end());
1078 // Verify that all SUnits were scheduled.
1079 bool AnyNotSched = false;
1080 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
1081 if (SUnits[i].NumSuccsLeft != 0) {
1083 cerr << "*** List scheduling failed! ***\n";
1084 SUnits[i].dump(&DAG);
1085 cerr << "has not been scheduled!\n";
1089 assert(!AnyNotSched);
1093 //===----------------------------------------------------------------------===//
1094 // Top-Down Scheduling
1095 //===----------------------------------------------------------------------===//
1097 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to
1098 /// the AvailableQueue if the count reaches zero. Also update its cycle bound.
1099 void ScheduleDAGRRList::ReleaseSucc(SUnit *SuccSU, bool isChain,
1100 unsigned CurCycle) {
1101 // FIXME: the distance between two nodes is not always == the predecessor's
1102 // latency. For example, the reader can very well read the register written
1103 // by the predecessor later than the issue cycle. It also depends on the
1104 // interrupt model (drain vs. freeze).
1105 SuccSU->CycleBound = std::max(SuccSU->CycleBound, CurCycle + SuccSU->Latency);
1107 --SuccSU->NumPredsLeft;
1110 if (SuccSU->NumPredsLeft < 0) {
1111 cerr << "*** List scheduling failed! ***\n";
1113 cerr << " has been released too many times!\n";
1118 if (SuccSU->NumPredsLeft == 0) {
1119 SuccSU->isAvailable = true;
1120 AvailableQueue->push(SuccSU);
1125 /// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending
1126 /// count of its successors. If a successor pending count is zero, add it to
1127 /// the Available queue.
1128 void ScheduleDAGRRList::ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle) {
1129 DOUT << "*** Scheduling [" << CurCycle << "]: ";
1130 DEBUG(SU->dump(&DAG));
1131 SU->Cycle = CurCycle;
1133 AvailableQueue->ScheduledNode(SU);
1135 // Top down: release successors
1136 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1138 ReleaseSucc(I->Dep, I->isCtrl, CurCycle);
1139 SU->isScheduled = true;
1142 /// ListScheduleTopDown - The main loop of list scheduling for top-down
1144 void ScheduleDAGRRList::ListScheduleTopDown() {
1145 unsigned CurCycle = 0;
1146 SUnit *Entry = SUnitMap[DAG.getEntryNode().Val].front();
1148 // All leaves to Available queue.
1149 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
1150 // It is available if it has no predecessors.
1151 if (SUnits[i].Preds.empty() && &SUnits[i] != Entry) {
1152 AvailableQueue->push(&SUnits[i]);
1153 SUnits[i].isAvailable = true;
1157 // Emit the entry node first.
1158 ScheduleNodeTopDown(Entry, CurCycle);
1159 Sequence.push_back(Entry);
1162 // While Available queue is not empty, grab the node with the highest
1163 // priority. If it is not ready put it back. Schedule the node.
1164 std::vector<SUnit*> NotReady;
1165 while (!AvailableQueue->empty()) {
1166 SUnit *CurSU = AvailableQueue->pop();
1167 while (CurSU && CurSU->CycleBound > CurCycle) {
1168 NotReady.push_back(CurSU);
1169 CurSU = AvailableQueue->pop();
1172 // Add the nodes that aren't ready back onto the available list.
1173 AvailableQueue->push_all(NotReady);
1177 Sequence.push_back(0);
1179 ScheduleNodeTopDown(CurSU, CurCycle);
1180 Sequence.push_back(CurSU);
1187 // Verify that all SUnits were scheduled.
1188 bool AnyNotSched = false;
1189 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
1190 if (!SUnits[i].isScheduled) {
1192 cerr << "*** List scheduling failed! ***\n";
1193 SUnits[i].dump(&DAG);
1194 cerr << "has not been scheduled!\n";
1198 assert(!AnyNotSched);
1204 //===----------------------------------------------------------------------===//
1205 // RegReductionPriorityQueue Implementation
1206 //===----------------------------------------------------------------------===//
1208 // This is a SchedulingPriorityQueue that schedules using Sethi Ullman numbers
1209 // to reduce register pressure.
1213 class RegReductionPriorityQueue;
1215 /// Sorting functions for the Available queue.
1216 struct bu_ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> {
1217 RegReductionPriorityQueue<bu_ls_rr_sort> *SPQ;
1218 bu_ls_rr_sort(RegReductionPriorityQueue<bu_ls_rr_sort> *spq) : SPQ(spq) {}
1219 bu_ls_rr_sort(const bu_ls_rr_sort &RHS) : SPQ(RHS.SPQ) {}
1221 bool operator()(const SUnit* left, const SUnit* right) const;
1224 struct td_ls_rr_sort : public std::binary_function<SUnit*, SUnit*, bool> {
1225 RegReductionPriorityQueue<td_ls_rr_sort> *SPQ;
1226 td_ls_rr_sort(RegReductionPriorityQueue<td_ls_rr_sort> *spq) : SPQ(spq) {}
1227 td_ls_rr_sort(const td_ls_rr_sort &RHS) : SPQ(RHS.SPQ) {}
1229 bool operator()(const SUnit* left, const SUnit* right) const;
1231 } // end anonymous namespace
1233 static inline bool isCopyFromLiveIn(const SUnit *SU) {
1234 SDNode *N = SU->Node;
1235 return N && N->getOpcode() == ISD::CopyFromReg &&
1236 N->getOperand(N->getNumOperands()-1).getValueType() != MVT::Flag;
1241 class VISIBILITY_HIDDEN RegReductionPriorityQueue
1242 : public SchedulingPriorityQueue {
1243 std::priority_queue<SUnit*, std::vector<SUnit*>, SF> Queue;
1246 RegReductionPriorityQueue() :
1249 virtual void initNodes(DenseMap<SDNode*, std::vector<SUnit*> > &sumap,
1250 std::vector<SUnit> &sunits) {}
1252 virtual void addNode(const SUnit *SU) {}
1254 virtual void updateNode(const SUnit *SU) {}
1256 virtual void releaseState() {}
1258 virtual unsigned getNodePriority(const SUnit *SU) const {
1262 unsigned size() const { return Queue.size(); }
1264 bool empty() const { return Queue.empty(); }
1266 void push(SUnit *U) {
1269 void push_all(const std::vector<SUnit *> &Nodes) {
1270 for (unsigned i = 0, e = Nodes.size(); i != e; ++i)
1271 Queue.push(Nodes[i]);
1275 if (empty()) return NULL;
1276 SUnit *V = Queue.top();
1281 /// remove - This is a really inefficient way to remove a node from a
1282 /// priority queue. We should roll our own heap to make this better or
1284 void remove(SUnit *SU) {
1285 std::vector<SUnit*> Temp;
1287 assert(!Queue.empty() && "Not in queue!");
1288 while (Queue.top() != SU) {
1289 Temp.push_back(Queue.top());
1291 assert(!Queue.empty() && "Not in queue!");
1294 // Remove the node from the PQ.
1297 // Add all the other nodes back.
1298 for (unsigned i = 0, e = Temp.size(); i != e; ++i)
1299 Queue.push(Temp[i]);
1304 class VISIBILITY_HIDDEN BURegReductionPriorityQueue
1305 : public RegReductionPriorityQueue<SF> {
1306 // SUnitMap SDNode to SUnit mapping (n -> n).
1307 DenseMap<SDNode*, std::vector<SUnit*> > *SUnitMap;
1309 // SUnits - The SUnits for the current graph.
1310 const std::vector<SUnit> *SUnits;
1312 // SethiUllmanNumbers - The SethiUllman number for each node.
1313 std::vector<unsigned> SethiUllmanNumbers;
1315 const TargetInstrInfo *TII;
1316 const TargetRegisterInfo *TRI;
1317 ScheduleDAGRRList *scheduleDAG;
1319 explicit BURegReductionPriorityQueue(const TargetInstrInfo *tii,
1320 const TargetRegisterInfo *tri)
1321 : TII(tii), TRI(tri), scheduleDAG(NULL) {}
1323 void initNodes(DenseMap<SDNode*, std::vector<SUnit*> > &sumap,
1324 std::vector<SUnit> &sunits) {
1327 // Add pseudo dependency edges for two-address nodes.
1328 AddPseudoTwoAddrDeps();
1329 // Calculate node priorities.
1330 CalculateSethiUllmanNumbers();
1333 void addNode(const SUnit *SU) {
1334 SethiUllmanNumbers.resize(SUnits->size(), 0);
1335 CalcNodeSethiUllmanNumber(SU);
1338 void updateNode(const SUnit *SU) {
1339 SethiUllmanNumbers[SU->NodeNum] = 0;
1340 CalcNodeSethiUllmanNumber(SU);
1343 void releaseState() {
1345 SethiUllmanNumbers.clear();
1348 unsigned getNodePriority(const SUnit *SU) const {
1349 assert(SU->NodeNum < SethiUllmanNumbers.size());
1350 unsigned Opc = SU->Node ? SU->Node->getOpcode() : 0;
1351 if (Opc == ISD::CopyFromReg && !isCopyFromLiveIn(SU))
1352 // CopyFromReg should be close to its def because it restricts
1353 // allocation choices. But if it is a livein then perhaps we want it
1354 // closer to its uses so it can be coalesced.
1356 else if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
1357 // CopyToReg should be close to its uses to facilitate coalescing and
1360 else if (Opc == TargetInstrInfo::EXTRACT_SUBREG ||
1361 Opc == TargetInstrInfo::INSERT_SUBREG)
1362 // EXTRACT_SUBREG / INSERT_SUBREG should be close to its use to
1363 // facilitate coalescing.
1365 else if (SU->NumSuccs == 0)
1366 // If SU does not have a use, i.e. it doesn't produce a value that would
1367 // be consumed (e.g. store), then it terminates a chain of computation.
1368 // Give it a large SethiUllman number so it will be scheduled right
1369 // before its predecessors that it doesn't lengthen their live ranges.
1371 else if (SU->NumPreds == 0)
1372 // If SU does not have a def, schedule it close to its uses because it
1373 // does not lengthen any live ranges.
1376 return SethiUllmanNumbers[SU->NodeNum];
1379 void setScheduleDAG(ScheduleDAGRRList *scheduleDag) {
1380 scheduleDAG = scheduleDag;
1384 bool canClobber(const SUnit *SU, const SUnit *Op);
1385 void AddPseudoTwoAddrDeps();
1386 void CalculateSethiUllmanNumbers();
1387 unsigned CalcNodeSethiUllmanNumber(const SUnit *SU);
1392 class VISIBILITY_HIDDEN TDRegReductionPriorityQueue
1393 : public RegReductionPriorityQueue<SF> {
1394 // SUnitMap SDNode to SUnit mapping (n -> n).
1395 DenseMap<SDNode*, std::vector<SUnit*> > *SUnitMap;
1397 // SUnits - The SUnits for the current graph.
1398 const std::vector<SUnit> *SUnits;
1400 // SethiUllmanNumbers - The SethiUllman number for each node.
1401 std::vector<unsigned> SethiUllmanNumbers;
1404 TDRegReductionPriorityQueue() {}
1406 void initNodes(DenseMap<SDNode*, std::vector<SUnit*> > &sumap,
1407 std::vector<SUnit> &sunits) {
1410 // Calculate node priorities.
1411 CalculateSethiUllmanNumbers();
1414 void addNode(const SUnit *SU) {
1415 SethiUllmanNumbers.resize(SUnits->size(), 0);
1416 CalcNodeSethiUllmanNumber(SU);
1419 void updateNode(const SUnit *SU) {
1420 SethiUllmanNumbers[SU->NodeNum] = 0;
1421 CalcNodeSethiUllmanNumber(SU);
1424 void releaseState() {
1426 SethiUllmanNumbers.clear();
1429 unsigned getNodePriority(const SUnit *SU) const {
1430 assert(SU->NodeNum < SethiUllmanNumbers.size());
1431 return SethiUllmanNumbers[SU->NodeNum];
1435 void CalculateSethiUllmanNumbers();
1436 unsigned CalcNodeSethiUllmanNumber(const SUnit *SU);
1440 /// closestSucc - Returns the scheduled cycle of the successor which is
1441 /// closet to the current cycle.
1442 static unsigned closestSucc(const SUnit *SU) {
1443 unsigned MaxCycle = 0;
1444 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1446 unsigned Cycle = I->Dep->Cycle;
1447 // If there are bunch of CopyToRegs stacked up, they should be considered
1448 // to be at the same position.
1449 if (I->Dep->Node && I->Dep->Node->getOpcode() == ISD::CopyToReg)
1450 Cycle = closestSucc(I->Dep)+1;
1451 if (Cycle > MaxCycle)
1457 /// calcMaxScratches - Returns an cost estimate of the worse case requirement
1458 /// for scratch registers. Live-in operands and live-out results don't count
1459 /// since they are "fixed".
1460 static unsigned calcMaxScratches(const SUnit *SU) {
1461 unsigned Scratches = 0;
1462 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1464 if (I->isCtrl) continue; // ignore chain preds
1465 if (!I->Dep->Node || I->Dep->Node->getOpcode() != ISD::CopyFromReg)
1468 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1470 if (I->isCtrl) continue; // ignore chain succs
1471 if (!I->Dep->Node || I->Dep->Node->getOpcode() != ISD::CopyToReg)
1478 bool bu_ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const {
1479 // There used to be a special tie breaker here that looked for
1480 // two-address instructions and preferred the instruction with a
1481 // def&use operand. The special case triggered diagnostics when
1482 // _GLIBCXX_DEBUG was enabled because it broke the strict weak
1483 // ordering that priority_queue requires. It didn't help much anyway
1484 // because AddPseudoTwoAddrDeps already covers many of the cases
1485 // where it would have applied. In addition, it's counter-intuitive
1486 // that a tie breaker would be the first thing attempted. There's a
1487 // "real" tie breaker below that is the operation of last resort.
1488 // The fact that the "special tie breaker" would trigger when there
1489 // wasn't otherwise a tie is what broke the strict weak ordering
1492 unsigned LPriority = SPQ->getNodePriority(left);
1493 unsigned RPriority = SPQ->getNodePriority(right);
1494 if (LPriority != RPriority)
1495 return LPriority > RPriority;
1497 // Try schedule def + use closer when Sethi-Ullman numbers are the same.
1502 // and the following instructions are both ready.
1506 // Then schedule t2 = op first.
1513 // This creates more short live intervals.
1514 unsigned LDist = closestSucc(left);
1515 unsigned RDist = closestSucc(right);
1517 return LDist < RDist;
1519 // Intuitively, it's good to push down instructions whose results are
1520 // liveout so their long live ranges won't conflict with other values
1521 // which are needed inside the BB. Further prioritize liveout instructions
1522 // by the number of operands which are calculated within the BB.
1523 unsigned LScratch = calcMaxScratches(left);
1524 unsigned RScratch = calcMaxScratches(right);
1525 if (LScratch != RScratch)
1526 return LScratch > RScratch;
1528 if (left->Height != right->Height)
1529 return left->Height > right->Height;
1531 if (left->Depth != right->Depth)
1532 return left->Depth < right->Depth;
1534 if (left->CycleBound != right->CycleBound)
1535 return left->CycleBound > right->CycleBound;
1537 // FIXME: No strict ordering.
1541 template<class SF> bool
1542 BURegReductionPriorityQueue<SF>::canClobber(const SUnit *SU, const SUnit *Op) {
1543 if (SU->isTwoAddress) {
1544 unsigned Opc = SU->Node->getTargetOpcode();
1545 const TargetInstrDesc &TID = TII->get(Opc);
1546 unsigned NumRes = TID.getNumDefs();
1547 unsigned NumOps = TID.getNumOperands() - NumRes;
1548 for (unsigned i = 0; i != NumOps; ++i) {
1549 if (TID.getOperandConstraint(i+NumRes, TOI::TIED_TO) != -1) {
1550 SDNode *DU = SU->Node->getOperand(i).Val;
1551 if ((*SUnitMap).find(DU) != (*SUnitMap).end() &&
1552 Op == (*SUnitMap)[DU][SU->InstanceNo])
1561 /// hasCopyToRegUse - Return true if SU has a value successor that is a
1563 static bool hasCopyToRegUse(SUnit *SU) {
1564 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1566 if (I->isCtrl) continue;
1567 SUnit *SuccSU = I->Dep;
1568 if (SuccSU->Node && SuccSU->Node->getOpcode() == ISD::CopyToReg)
1574 /// canClobberPhysRegDefs - True if SU would clobber one of SuccSU's
1575 /// physical register def.
1576 static bool canClobberPhysRegDefs(SUnit *SuccSU, SUnit *SU,
1577 const TargetInstrInfo *TII,
1578 const TargetRegisterInfo *TRI) {
1579 SDNode *N = SuccSU->Node;
1580 unsigned NumDefs = TII->get(N->getTargetOpcode()).getNumDefs();
1581 const unsigned *ImpDefs = TII->get(N->getTargetOpcode()).getImplicitDefs();
1584 const unsigned *SUImpDefs =
1585 TII->get(SU->Node->getTargetOpcode()).getImplicitDefs();
1588 for (unsigned i = NumDefs, e = N->getNumValues(); i != e; ++i) {
1589 MVT::ValueType VT = N->getValueType(i);
1590 if (VT == MVT::Flag || VT == MVT::Other)
1592 unsigned Reg = ImpDefs[i - NumDefs];
1593 for (;*SUImpDefs; ++SUImpDefs) {
1594 unsigned SUReg = *SUImpDefs;
1595 if (TRI->regsOverlap(Reg, SUReg))
1602 /// AddPseudoTwoAddrDeps - If two nodes share an operand and one of them uses
1603 /// it as a def&use operand. Add a pseudo control edge from it to the other
1604 /// node (if it won't create a cycle) so the two-address one will be scheduled
1605 /// first (lower in the schedule). If both nodes are two-address, favor the
1606 /// one that has a CopyToReg use (more likely to be a loop induction update).
1607 /// If both are two-address, but one is commutable while the other is not
1608 /// commutable, favor the one that's not commutable.
1610 void BURegReductionPriorityQueue<SF>::AddPseudoTwoAddrDeps() {
1611 for (unsigned i = 0, e = SUnits->size(); i != e; ++i) {
1612 SUnit *SU = (SUnit *)&((*SUnits)[i]);
1613 if (!SU->isTwoAddress)
1616 SDNode *Node = SU->Node;
1617 if (!Node || !Node->isTargetOpcode() || SU->FlaggedNodes.size() > 0)
1620 unsigned Opc = Node->getTargetOpcode();
1621 const TargetInstrDesc &TID = TII->get(Opc);
1622 unsigned NumRes = TID.getNumDefs();
1623 unsigned NumOps = TID.getNumOperands() - NumRes;
1624 for (unsigned j = 0; j != NumOps; ++j) {
1625 if (TID.getOperandConstraint(j+NumRes, TOI::TIED_TO) != -1) {
1626 SDNode *DU = SU->Node->getOperand(j).Val;
1627 if ((*SUnitMap).find(DU) == (*SUnitMap).end())
1629 SUnit *DUSU = (*SUnitMap)[DU][SU->InstanceNo];
1630 if (!DUSU) continue;
1631 for (SUnit::succ_iterator I = DUSU->Succs.begin(),E = DUSU->Succs.end();
1633 if (I->isCtrl) continue;
1634 SUnit *SuccSU = I->Dep;
1637 // Be conservative. Ignore if nodes aren't at roughly the same
1638 // depth and height.
1639 if (SuccSU->Height < SU->Height && (SU->Height - SuccSU->Height) > 1)
1641 if (!SuccSU->Node || !SuccSU->Node->isTargetOpcode())
1643 // Don't constrain nodes with physical register defs if the
1644 // predecessor can clobber them.
1645 if (SuccSU->hasPhysRegDefs) {
1646 if (canClobberPhysRegDefs(SuccSU, SU, TII, TRI))
1649 // Don't constraint extract_subreg / insert_subreg these may be
1650 // coalesced away. We don't them close to their uses.
1651 unsigned SuccOpc = SuccSU->Node->getTargetOpcode();
1652 if (SuccOpc == TargetInstrInfo::EXTRACT_SUBREG ||
1653 SuccOpc == TargetInstrInfo::INSERT_SUBREG)
1655 if ((!canClobber(SuccSU, DUSU) ||
1656 (hasCopyToRegUse(SU) && !hasCopyToRegUse(SuccSU)) ||
1657 (!SU->isCommutable && SuccSU->isCommutable)) &&
1658 !scheduleDAG->IsReachable(SuccSU, SU)) {
1659 DOUT << "Adding an edge from SU # " << SU->NodeNum
1660 << " to SU #" << SuccSU->NodeNum << "\n";
1661 scheduleDAG->AddPred(SU, SuccSU, true, true);
1669 /// CalcNodeSethiUllmanNumber - Priority is the Sethi Ullman number.
1670 /// Smaller number is the higher priority.
1672 unsigned BURegReductionPriorityQueue<SF>::
1673 CalcNodeSethiUllmanNumber(const SUnit *SU) {
1674 unsigned &SethiUllmanNumber = SethiUllmanNumbers[SU->NodeNum];
1675 if (SethiUllmanNumber != 0)
1676 return SethiUllmanNumber;
1679 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1681 if (I->isCtrl) continue; // ignore chain preds
1682 SUnit *PredSU = I->Dep;
1683 unsigned PredSethiUllman = CalcNodeSethiUllmanNumber(PredSU);
1684 if (PredSethiUllman > SethiUllmanNumber) {
1685 SethiUllmanNumber = PredSethiUllman;
1687 } else if (PredSethiUllman == SethiUllmanNumber && !I->isCtrl)
1691 SethiUllmanNumber += Extra;
1693 if (SethiUllmanNumber == 0)
1694 SethiUllmanNumber = 1;
1696 return SethiUllmanNumber;
1699 /// CalculateSethiUllmanNumbers - Calculate Sethi-Ullman numbers of all
1700 /// scheduling units.
1702 void BURegReductionPriorityQueue<SF>::CalculateSethiUllmanNumbers() {
1703 SethiUllmanNumbers.assign(SUnits->size(), 0);
1705 for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
1706 CalcNodeSethiUllmanNumber(&(*SUnits)[i]);
1709 static unsigned SumOfUnscheduledPredsOfSuccs(const SUnit *SU) {
1711 for (SUnit::const_succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1713 SUnit *SuccSU = I->Dep;
1714 for (SUnit::const_pred_iterator II = SuccSU->Preds.begin(),
1715 EE = SuccSU->Preds.end(); II != EE; ++II) {
1716 SUnit *PredSU = II->Dep;
1717 if (!PredSU->isScheduled)
1727 bool td_ls_rr_sort::operator()(const SUnit *left, const SUnit *right) const {
1728 unsigned LPriority = SPQ->getNodePriority(left);
1729 unsigned RPriority = SPQ->getNodePriority(right);
1730 bool LIsTarget = left->Node && left->Node->isTargetOpcode();
1731 bool RIsTarget = right->Node && right->Node->isTargetOpcode();
1732 bool LIsFloater = LIsTarget && left->NumPreds == 0;
1733 bool RIsFloater = RIsTarget && right->NumPreds == 0;
1734 unsigned LBonus = (SumOfUnscheduledPredsOfSuccs(left) == 1) ? 2 : 0;
1735 unsigned RBonus = (SumOfUnscheduledPredsOfSuccs(right) == 1) ? 2 : 0;
1737 if (left->NumSuccs == 0 && right->NumSuccs != 0)
1739 else if (left->NumSuccs != 0 && right->NumSuccs == 0)
1746 if (left->NumSuccs == 1)
1748 if (right->NumSuccs == 1)
1751 if (LPriority+LBonus != RPriority+RBonus)
1752 return LPriority+LBonus < RPriority+RBonus;
1754 if (left->Depth != right->Depth)
1755 return left->Depth < right->Depth;
1757 if (left->NumSuccsLeft != right->NumSuccsLeft)
1758 return left->NumSuccsLeft > right->NumSuccsLeft;
1760 if (left->CycleBound != right->CycleBound)
1761 return left->CycleBound > right->CycleBound;
1763 // FIXME: No strict ordering.
1767 /// CalcNodeSethiUllmanNumber - Priority is the Sethi Ullman number.
1768 /// Smaller number is the higher priority.
1770 unsigned TDRegReductionPriorityQueue<SF>::
1771 CalcNodeSethiUllmanNumber(const SUnit *SU) {
1772 unsigned &SethiUllmanNumber = SethiUllmanNumbers[SU->NodeNum];
1773 if (SethiUllmanNumber != 0)
1774 return SethiUllmanNumber;
1776 unsigned Opc = SU->Node ? SU->Node->getOpcode() : 0;
1777 if (Opc == ISD::TokenFactor || Opc == ISD::CopyToReg)
1778 SethiUllmanNumber = 0xffff;
1779 else if (SU->NumSuccsLeft == 0)
1780 // If SU does not have a use, i.e. it doesn't produce a value that would
1781 // be consumed (e.g. store), then it terminates a chain of computation.
1782 // Give it a small SethiUllman number so it will be scheduled right before
1783 // its predecessors that it doesn't lengthen their live ranges.
1784 SethiUllmanNumber = 0;
1785 else if (SU->NumPredsLeft == 0 &&
1786 (Opc != ISD::CopyFromReg || isCopyFromLiveIn(SU)))
1787 SethiUllmanNumber = 0xffff;
1790 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
1792 if (I->isCtrl) continue; // ignore chain preds
1793 SUnit *PredSU = I->Dep;
1794 unsigned PredSethiUllman = CalcNodeSethiUllmanNumber(PredSU);
1795 if (PredSethiUllman > SethiUllmanNumber) {
1796 SethiUllmanNumber = PredSethiUllman;
1798 } else if (PredSethiUllman == SethiUllmanNumber && !I->isCtrl)
1802 SethiUllmanNumber += Extra;
1805 return SethiUllmanNumber;
1808 /// CalculateSethiUllmanNumbers - Calculate Sethi-Ullman numbers of all
1809 /// scheduling units.
1811 void TDRegReductionPriorityQueue<SF>::CalculateSethiUllmanNumbers() {
1812 SethiUllmanNumbers.assign(SUnits->size(), 0);
1814 for (unsigned i = 0, e = SUnits->size(); i != e; ++i)
1815 CalcNodeSethiUllmanNumber(&(*SUnits)[i]);
1818 //===----------------------------------------------------------------------===//
1819 // Public Constructor Functions
1820 //===----------------------------------------------------------------------===//
1822 llvm::ScheduleDAG* llvm::createBURRListDAGScheduler(SelectionDAGISel *IS,
1824 MachineBasicBlock *BB) {
1825 const TargetInstrInfo *TII = DAG->getTarget().getInstrInfo();
1826 const TargetRegisterInfo *TRI = DAG->getTarget().getRegisterInfo();
1828 BURegReductionPriorityQueue<bu_ls_rr_sort> *priorityQueue =
1829 new BURegReductionPriorityQueue<bu_ls_rr_sort>(TII, TRI);
1831 ScheduleDAGRRList * scheduleDAG =
1832 new ScheduleDAGRRList(*DAG, BB, DAG->getTarget(), true, priorityQueue);
1833 priorityQueue->setScheduleDAG(scheduleDAG);
1837 llvm::ScheduleDAG* llvm::createTDRRListDAGScheduler(SelectionDAGISel *IS,
1839 MachineBasicBlock *BB) {
1840 return new ScheduleDAGRRList(*DAG, BB, DAG->getTarget(), false,
1841 new TDRegReductionPriorityQueue<td_ls_rr_sort>());