1 //===--- ScheduleDAGSDNodes.cpp - Implement the ScheduleDAGSDNodes class --===//
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 the ScheduleDAG class, which is a base class used by
11 // scheduling implementation classes.
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "pre-RA-sched"
16 #include "SDNodeDbgValue.h"
17 #include "ScheduleDAGSDNodes.h"
18 #include "InstrEmitter.h"
19 #include "llvm/CodeGen/SelectionDAG.h"
20 #include "llvm/Target/TargetMachine.h"
21 #include "llvm/Target/TargetInstrInfo.h"
22 #include "llvm/Target/TargetLowering.h"
23 #include "llvm/Target/TargetRegisterInfo.h"
24 #include "llvm/Target/TargetSubtarget.h"
25 #include "llvm/ADT/DenseMap.h"
26 #include "llvm/ADT/SmallPtrSet.h"
27 #include "llvm/ADT/SmallSet.h"
28 #include "llvm/ADT/SmallVector.h"
29 #include "llvm/ADT/Statistic.h"
30 #include "llvm/Support/Debug.h"
31 #include "llvm/Support/raw_ostream.h"
34 STATISTIC(LoadsClustered, "Number of loads clustered together");
36 ScheduleDAGSDNodes::ScheduleDAGSDNodes(MachineFunction &mf)
40 /// Run - perform scheduling.
42 void ScheduleDAGSDNodes::Run(SelectionDAG *dag, MachineBasicBlock *bb,
43 MachineBasicBlock::iterator insertPos) {
45 ScheduleDAG::Run(bb, insertPos);
48 /// NewSUnit - Creates a new SUnit and return a ptr to it.
50 SUnit *ScheduleDAGSDNodes::NewSUnit(SDNode *N) {
52 const SUnit *Addr = 0;
56 SUnits.push_back(SUnit(N, (unsigned)SUnits.size()));
57 assert((Addr == 0 || Addr == &SUnits[0]) &&
58 "SUnits std::vector reallocated on the fly!");
59 SUnits.back().OrigNode = &SUnits.back();
60 SUnit *SU = &SUnits.back();
61 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
62 if (N->isMachineOpcode() &&
63 N->getMachineOpcode() == TargetOpcode::IMPLICIT_DEF)
64 SU->SchedulingPref = Sched::None;
66 SU->SchedulingPref = TLI.getSchedulingPreference(N);
70 SUnit *ScheduleDAGSDNodes::Clone(SUnit *Old) {
71 SUnit *SU = NewSUnit(Old->getNode());
72 SU->OrigNode = Old->OrigNode;
73 SU->Latency = Old->Latency;
74 SU->isTwoAddress = Old->isTwoAddress;
75 SU->isCommutable = Old->isCommutable;
76 SU->hasPhysRegDefs = Old->hasPhysRegDefs;
77 SU->hasPhysRegClobbers = Old->hasPhysRegClobbers;
78 SU->SchedulingPref = Old->SchedulingPref;
83 /// CheckForPhysRegDependency - Check if the dependency between def and use of
84 /// a specified operand is a physical register dependency. If so, returns the
85 /// register and the cost of copying the register.
86 static void CheckForPhysRegDependency(SDNode *Def, SDNode *User, unsigned Op,
87 const TargetRegisterInfo *TRI,
88 const TargetInstrInfo *TII,
89 unsigned &PhysReg, int &Cost) {
90 if (Op != 2 || User->getOpcode() != ISD::CopyToReg)
93 unsigned Reg = cast<RegisterSDNode>(User->getOperand(1))->getReg();
94 if (TargetRegisterInfo::isVirtualRegister(Reg))
97 unsigned ResNo = User->getOperand(2).getResNo();
98 if (Def->isMachineOpcode()) {
99 const TargetInstrDesc &II = TII->get(Def->getMachineOpcode());
100 if (ResNo >= II.getNumDefs() &&
101 II.ImplicitDefs[ResNo - II.getNumDefs()] == Reg) {
103 const TargetRegisterClass *RC =
104 TRI->getPhysicalRegisterRegClass(Reg, Def->getValueType(ResNo));
105 Cost = RC->getCopyCost();
110 static void AddFlags(SDNode *N, SDValue Flag, bool AddFlag,
112 SmallVector<EVT, 4> VTs;
114 for (unsigned i = 0, e = N->getNumValues(); i != e; ++i)
115 VTs.push_back(N->getValueType(i));
118 VTs.push_back(MVT::Flag);
120 SmallVector<SDValue, 4> Ops;
121 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i)
122 Ops.push_back(N->getOperand(i));
127 SDVTList VTList = DAG->getVTList(&VTs[0], VTs.size());
128 MachineSDNode::mmo_iterator Begin = 0, End = 0;
129 MachineSDNode *MN = dyn_cast<MachineSDNode>(N);
131 // Store memory references.
133 Begin = MN->memoperands_begin();
134 End = MN->memoperands_end();
137 DAG->MorphNodeTo(N, N->getOpcode(), VTList, &Ops[0], Ops.size());
139 // Reset the memory references
141 MN->setMemRefs(Begin, End);
144 /// ClusterNeighboringLoads - Force nearby loads together by "flagging" them.
145 /// This function finds loads of the same base and different offsets. If the
146 /// offsets are not far apart (target specific), it add MVT::Flag inputs and
147 /// outputs to ensure they are scheduled together and in order. This
148 /// optimization may benefit some targets by improving cache locality.
149 void ScheduleDAGSDNodes::ClusterNeighboringLoads(SDNode *Node) {
151 unsigned NumOps = Node->getNumOperands();
152 if (Node->getOperand(NumOps-1).getValueType() == MVT::Other)
153 Chain = Node->getOperand(NumOps-1).getNode();
157 // Look for other loads of the same chain. Find loads that are loading from
158 // the same base pointer and different offsets.
159 SmallPtrSet<SDNode*, 16> Visited;
160 SmallVector<int64_t, 4> Offsets;
161 DenseMap<long long, SDNode*> O2SMap; // Map from offset to SDNode.
162 bool Cluster = false;
165 for (SDNode::use_iterator I = Chain->use_begin(), E = Chain->use_end();
168 if (User == Node || !Visited.insert(User))
170 int64_t Offset1, Offset2;
171 if (!TII->areLoadsFromSameBasePtr(Base, User, Offset1, Offset2) ||
173 // FIXME: Should be ok if they addresses are identical. But earlier
174 // optimizations really should have eliminated one of the loads.
176 if (O2SMap.insert(std::make_pair(Offset1, Base)).second)
177 Offsets.push_back(Offset1);
178 O2SMap.insert(std::make_pair(Offset2, User));
179 Offsets.push_back(Offset2);
180 if (Offset2 < Offset1) {
182 BaseOffset = Offset2;
184 BaseOffset = Offset1;
192 // Sort them in increasing order.
193 std::sort(Offsets.begin(), Offsets.end());
195 // Check if the loads are close enough.
196 SmallVector<SDNode*, 4> Loads;
197 unsigned NumLoads = 0;
198 int64_t BaseOff = Offsets[0];
199 SDNode *BaseLoad = O2SMap[BaseOff];
200 Loads.push_back(BaseLoad);
201 for (unsigned i = 1, e = Offsets.size(); i != e; ++i) {
202 int64_t Offset = Offsets[i];
203 SDNode *Load = O2SMap[Offset];
204 if (!TII->shouldScheduleLoadsNear(BaseLoad, Load, BaseOff, Offset,NumLoads))
205 break; // Stop right here. Ignore loads that are further away.
206 Loads.push_back(Load);
213 // Cluster loads by adding MVT::Flag outputs and inputs. This also
214 // ensure they are scheduled in order of increasing addresses.
215 SDNode *Lead = Loads[0];
216 AddFlags(Lead, SDValue(0,0), true, DAG);
218 SDValue InFlag = SDValue(Lead, Lead->getNumValues()-1);
219 for (unsigned i = 1, e = Loads.size(); i != e; ++i) {
220 bool OutFlag = i < e-1;
221 SDNode *Load = Loads[i];
222 AddFlags(Load, InFlag, OutFlag, DAG);
225 InFlag = SDValue(Load, Load->getNumValues()-1);
231 /// ClusterNodes - Cluster certain nodes which should be scheduled together.
233 void ScheduleDAGSDNodes::ClusterNodes() {
234 for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
235 E = DAG->allnodes_end(); NI != E; ++NI) {
237 if (!Node || !Node->isMachineOpcode())
240 unsigned Opc = Node->getMachineOpcode();
241 const TargetInstrDesc &TID = TII->get(Opc);
243 // Cluster loads from "near" addresses into combined SUnits.
244 ClusterNeighboringLoads(Node);
248 void ScheduleDAGSDNodes::BuildSchedUnits() {
249 // During scheduling, the NodeId field of SDNode is used to map SDNodes
250 // to their associated SUnits by holding SUnits table indices. A value
251 // of -1 means the SDNode does not yet have an associated SUnit.
252 unsigned NumNodes = 0;
253 for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
254 E = DAG->allnodes_end(); NI != E; ++NI) {
259 // Reserve entries in the vector for each of the SUnits we are creating. This
260 // ensure that reallocation of the vector won't happen, so SUnit*'s won't get
262 // FIXME: Multiply by 2 because we may clone nodes during scheduling.
263 // This is a temporary workaround.
264 SUnits.reserve(NumNodes * 2);
266 // Add all nodes in depth first order.
267 SmallVector<SDNode*, 64> Worklist;
268 SmallPtrSet<SDNode*, 64> Visited;
269 Worklist.push_back(DAG->getRoot().getNode());
270 Visited.insert(DAG->getRoot().getNode());
272 while (!Worklist.empty()) {
273 SDNode *NI = Worklist.pop_back_val();
275 // Add all operands to the worklist unless they've already been added.
276 for (unsigned i = 0, e = NI->getNumOperands(); i != e; ++i)
277 if (Visited.insert(NI->getOperand(i).getNode()))
278 Worklist.push_back(NI->getOperand(i).getNode());
280 if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate.
283 // If this node has already been processed, stop now.
284 if (NI->getNodeId() != -1) continue;
286 SUnit *NodeSUnit = NewSUnit(NI);
288 // See if anything is flagged to this node, if so, add them to flagged
289 // nodes. Nodes can have at most one flag input and one flag output. Flags
290 // are required to be the last operand and result of a node.
292 // Scan up to find flagged preds.
294 while (N->getNumOperands() &&
295 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Flag) {
296 N = N->getOperand(N->getNumOperands()-1).getNode();
297 assert(N->getNodeId() == -1 && "Node already inserted!");
298 N->setNodeId(NodeSUnit->NodeNum);
301 // Scan down to find any flagged succs.
303 while (N->getValueType(N->getNumValues()-1) == MVT::Flag) {
304 SDValue FlagVal(N, N->getNumValues()-1);
306 // There are either zero or one users of the Flag result.
307 bool HasFlagUse = false;
308 for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
310 if (FlagVal.isOperandOf(*UI)) {
312 assert(N->getNodeId() == -1 && "Node already inserted!");
313 N->setNodeId(NodeSUnit->NodeNum);
317 if (!HasFlagUse) break;
320 // If there are flag operands involved, N is now the bottom-most node
321 // of the sequence of nodes that are flagged together.
323 NodeSUnit->setNode(N);
324 assert(N->getNodeId() == -1 && "Node already inserted!");
325 N->setNodeId(NodeSUnit->NodeNum);
327 // Assign the Latency field of NodeSUnit using target-provided information.
328 ComputeLatency(NodeSUnit);
332 void ScheduleDAGSDNodes::AddSchedEdges() {
333 const TargetSubtarget &ST = TM.getSubtarget<TargetSubtarget>();
335 // Check to see if the scheduler cares about latencies.
336 bool UnitLatencies = ForceUnitLatencies();
338 // Pass 2: add the preds, succs, etc.
339 for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
340 SUnit *SU = &SUnits[su];
341 SDNode *MainNode = SU->getNode();
343 if (MainNode->isMachineOpcode()) {
344 unsigned Opc = MainNode->getMachineOpcode();
345 const TargetInstrDesc &TID = TII->get(Opc);
346 for (unsigned i = 0; i != TID.getNumOperands(); ++i) {
347 if (TID.getOperandConstraint(i, TOI::TIED_TO) != -1) {
348 SU->isTwoAddress = true;
352 if (TID.isCommutable())
353 SU->isCommutable = true;
356 // Find all predecessors and successors of the group.
357 for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode()) {
358 if (N->isMachineOpcode() &&
359 TII->get(N->getMachineOpcode()).getImplicitDefs()) {
360 SU->hasPhysRegClobbers = true;
361 unsigned NumUsed = InstrEmitter::CountResults(N);
362 while (NumUsed != 0 && !N->hasAnyUseOfValue(NumUsed - 1))
363 --NumUsed; // Skip over unused values at the end.
364 if (NumUsed > TII->get(N->getMachineOpcode()).getNumDefs())
365 SU->hasPhysRegDefs = true;
368 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
369 SDNode *OpN = N->getOperand(i).getNode();
370 if (isPassiveNode(OpN)) continue; // Not scheduled.
371 SUnit *OpSU = &SUnits[OpN->getNodeId()];
372 assert(OpSU && "Node has no SUnit!");
373 if (OpSU == SU) continue; // In the same group.
375 EVT OpVT = N->getOperand(i).getValueType();
376 assert(OpVT != MVT::Flag && "Flagged nodes should be in same sunit!");
377 bool isChain = OpVT == MVT::Other;
379 unsigned PhysReg = 0;
381 // Determine if this is a physical register dependency.
382 CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost);
383 assert((PhysReg == 0 || !isChain) &&
384 "Chain dependence via physreg data?");
385 // FIXME: See ScheduleDAGSDNodes::EmitCopyFromReg. For now, scheduler
386 // emits a copy from the physical register to a virtual register unless
387 // it requires a cross class copy (cost < 0). That means we are only
388 // treating "expensive to copy" register dependency as physical register
389 // dependency. This may change in the future though.
393 // If this is a ctrl dep, latency is 1.
394 unsigned OpLatency = isChain ? 1 : OpSU->Latency;
395 const SDep &dep = SDep(OpSU, isChain ? SDep::Order : SDep::Data,
397 if (!isChain && !UnitLatencies) {
398 ComputeOperandLatency(OpN, N, i, const_cast<SDep &>(dep));
399 ST.adjustSchedDependency(OpSU, SU, const_cast<SDep &>(dep));
408 /// BuildSchedGraph - Build the SUnit graph from the selection dag that we
409 /// are input. This SUnit graph is similar to the SelectionDAG, but
410 /// excludes nodes that aren't interesting to scheduling, and represents
411 /// flagged together nodes with a single SUnit.
412 void ScheduleDAGSDNodes::BuildSchedGraph(AliasAnalysis *AA) {
413 // Cluster certain nodes which should be scheduled together.
415 // Populate the SUnits array.
417 // Compute all the scheduling dependencies between nodes.
421 void ScheduleDAGSDNodes::ComputeLatency(SUnit *SU) {
422 // Check to see if the scheduler cares about latencies.
423 if (ForceUnitLatencies()) {
428 const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
429 if (InstrItins.isEmpty()) {
434 // Compute the latency for the node. We use the sum of the latencies for
435 // all nodes flagged together into this SUnit.
437 for (SDNode *N = SU->getNode(); N; N = N->getFlaggedNode())
438 if (N->isMachineOpcode()) {
439 SU->Latency += InstrItins.
440 getStageLatency(TII->get(N->getMachineOpcode()).getSchedClass());
444 void ScheduleDAGSDNodes::ComputeOperandLatency(SDNode *Def, SDNode *Use,
445 unsigned OpIdx, SDep& dep) const{
446 // Check to see if the scheduler cares about latencies.
447 if (ForceUnitLatencies())
450 const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
451 if (InstrItins.isEmpty())
454 if (dep.getKind() != SDep::Data)
457 unsigned DefIdx = Use->getOperand(OpIdx).getResNo();
458 if (Def->isMachineOpcode()) {
459 const TargetInstrDesc &II = TII->get(Def->getMachineOpcode());
460 if (DefIdx >= II.getNumDefs())
462 int DefCycle = InstrItins.getOperandCycle(II.getSchedClass(), DefIdx);
466 if (Use->isMachineOpcode()) {
467 const unsigned UseClass = TII->get(Use->getMachineOpcode()).getSchedClass();
468 UseCycle = InstrItins.getOperandCycle(UseClass, OpIdx);
471 int Latency = DefCycle - UseCycle + 1;
473 dep.setLatency(Latency);
478 void ScheduleDAGSDNodes::dumpNode(const SUnit *SU) const {
479 if (!SU->getNode()) {
480 dbgs() << "PHYS REG COPY\n";
484 SU->getNode()->dump(DAG);
486 SmallVector<SDNode *, 4> FlaggedNodes;
487 for (SDNode *N = SU->getNode()->getFlaggedNode(); N; N = N->getFlaggedNode())
488 FlaggedNodes.push_back(N);
489 while (!FlaggedNodes.empty()) {
491 FlaggedNodes.back()->dump(DAG);
493 FlaggedNodes.pop_back();
499 bool operator()(const std::pair<unsigned, MachineInstr*> &A,
500 const std::pair<unsigned, MachineInstr*> &B) {
501 return A.first < B.first;
506 // ProcessSourceNode - Process nodes with source order numbers. These are added
507 // to a vector which EmitSchedule use to determine how to insert dbg_value
508 // instructions in the right order.
509 static void ProcessSourceNode(SDNode *N, SelectionDAG *DAG,
510 InstrEmitter &Emitter,
511 DenseMap<SDValue, unsigned> &VRBaseMap,
512 SmallVector<std::pair<unsigned, MachineInstr*>, 32> &Orders,
513 SmallSet<unsigned, 8> &Seen) {
514 unsigned Order = DAG->GetOrdering(N);
515 if (!Order || !Seen.insert(Order))
518 MachineBasicBlock *BB = Emitter.getBlock();
519 if (BB->empty() || BB->back().isPHI()) {
520 // Did not insert any instruction.
521 Orders.push_back(std::make_pair(Order, (MachineInstr*)0));
525 Orders.push_back(std::make_pair(Order, &BB->back()));
526 if (!N->getHasDebugValue())
528 // Opportunistically insert immediate dbg_value uses, i.e. those with source
529 // order number right after the N.
530 MachineBasicBlock::iterator InsertPos = Emitter.getInsertPos();
531 SmallVector<SDDbgValue*,2> &DVs = DAG->GetDbgValues(N);
532 for (unsigned i = 0, e = DVs.size(); i != e; ++i) {
533 if (DVs[i]->isInvalidated())
535 unsigned DVOrder = DVs[i]->getOrder();
536 if (DVOrder == ++Order) {
537 MachineInstr *DbgMI = Emitter.EmitDbgValue(DVs[i], VRBaseMap);
539 Orders.push_back(std::make_pair(DVOrder, DbgMI));
540 BB->insert(InsertPos, DbgMI);
542 DVs[i]->setIsInvalidated();
548 /// EmitSchedule - Emit the machine code in scheduled order.
549 MachineBasicBlock *ScheduleDAGSDNodes::EmitSchedule() {
550 InstrEmitter Emitter(BB, InsertPos);
551 DenseMap<SDValue, unsigned> VRBaseMap;
552 DenseMap<SUnit*, unsigned> CopyVRBaseMap;
553 SmallVector<std::pair<unsigned, MachineInstr*>, 32> Orders;
554 SmallSet<unsigned, 8> Seen;
555 bool HasDbg = DAG->hasDebugValues();
557 // If this is the first BB, emit byval parameter dbg_value's.
558 if (HasDbg && BB->getParent()->begin() == MachineFunction::iterator(BB)) {
559 SDDbgInfo::DbgIterator PDI = DAG->ByvalParmDbgBegin();
560 SDDbgInfo::DbgIterator PDE = DAG->ByvalParmDbgEnd();
561 for (; PDI != PDE; ++PDI) {
562 MachineInstr *DbgMI= Emitter.EmitDbgValue(*PDI, VRBaseMap);
564 BB->push_back(DbgMI);
568 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
569 SUnit *SU = Sequence[i];
571 // Null SUnit* is a noop.
576 // For pre-regalloc scheduling, create instructions corresponding to the
577 // SDNode and any flagged SDNodes and append them to the block.
578 if (!SU->getNode()) {
580 EmitPhysRegCopy(SU, CopyVRBaseMap);
584 SmallVector<SDNode *, 4> FlaggedNodes;
585 for (SDNode *N = SU->getNode()->getFlaggedNode(); N;
586 N = N->getFlaggedNode())
587 FlaggedNodes.push_back(N);
588 while (!FlaggedNodes.empty()) {
589 SDNode *N = FlaggedNodes.back();
590 Emitter.EmitNode(FlaggedNodes.back(), SU->OrigNode != SU, SU->isCloned,
592 // Remember the source order of the inserted instruction.
594 ProcessSourceNode(N, DAG, Emitter, VRBaseMap, Orders, Seen);
595 FlaggedNodes.pop_back();
597 Emitter.EmitNode(SU->getNode(), SU->OrigNode != SU, SU->isCloned,
599 // Remember the source order of the inserted instruction.
601 ProcessSourceNode(SU->getNode(), DAG, Emitter, VRBaseMap, Orders,
605 // Insert all the dbg_values which have not already been inserted in source
608 MachineBasicBlock::iterator BBBegin = BB->empty() ? BB->end() : BB->begin();
609 while (BBBegin != BB->end() && BBBegin->isPHI())
612 // Sort the source order instructions and use the order to insert debug
614 std::sort(Orders.begin(), Orders.end(), OrderSorter());
616 SDDbgInfo::DbgIterator DI = DAG->DbgBegin();
617 SDDbgInfo::DbgIterator DE = DAG->DbgEnd();
618 // Now emit the rest according to source order.
619 unsigned LastOrder = 0;
620 MachineInstr *LastMI = 0;
621 for (unsigned i = 0, e = Orders.size(); i != e && DI != DE; ++i) {
622 unsigned Order = Orders[i].first;
623 MachineInstr *MI = Orders[i].second;
624 // Insert all SDDbgValue's whose order(s) are before "Order".
627 MachineBasicBlock *MIBB = MI->getParent();
629 unsigned LastDIOrder = 0;
632 (*DI)->getOrder() >= LastOrder && (*DI)->getOrder() < Order; ++DI) {
634 assert((*DI)->getOrder() >= LastDIOrder &&
635 "SDDbgValue nodes must be in source order!");
636 LastDIOrder = (*DI)->getOrder();
638 if ((*DI)->isInvalidated())
640 MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap);
643 // Insert to start of the BB (after PHIs).
644 BB->insert(BBBegin, DbgMI);
646 MachineBasicBlock::iterator Pos = MI;
647 MIBB->insert(llvm::next(Pos), DbgMI);
654 // Add trailing DbgValue's before the terminator. FIXME: May want to add
655 // some of them before one or more conditional branches?
657 MachineBasicBlock *InsertBB = Emitter.getBlock();
658 MachineBasicBlock::iterator Pos= Emitter.getBlock()->getFirstTerminator();
659 if (!(*DI)->isInvalidated()) {
660 MachineInstr *DbgMI= Emitter.EmitDbgValue(*DI, VRBaseMap);
662 InsertBB->insert(Pos, DbgMI);
668 BB = Emitter.getBlock();
669 InsertPos = Emitter.getInsertPos();