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 "ScheduleDAGSDNodes.h"
17 #include "InstrEmitter.h"
18 #include "SDNodeDbgValue.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallSet.h"
22 #include "llvm/ADT/SmallVector.h"
23 #include "llvm/ADT/Statistic.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/CodeGen/SelectionDAG.h"
27 #include "llvm/MC/MCInstrItineraries.h"
28 #include "llvm/Support/CommandLine.h"
29 #include "llvm/Support/Debug.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include "llvm/Target/TargetInstrInfo.h"
32 #include "llvm/Target/TargetLowering.h"
33 #include "llvm/Target/TargetMachine.h"
34 #include "llvm/Target/TargetRegisterInfo.h"
35 #include "llvm/Target/TargetSubtargetInfo.h"
38 STATISTIC(LoadsClustered, "Number of loads clustered together");
40 // This allows latency based scheduler to notice high latency instructions
41 // without a target itinerary. The choise if number here has more to do with
42 // balancing scheduler heursitics than with the actual machine latency.
43 static cl::opt<int> HighLatencyCycles(
44 "sched-high-latency-cycles", cl::Hidden, cl::init(10),
45 cl::desc("Roughly estimate the number of cycles that 'long latency'"
46 "instructions take for targets with no itinerary"));
48 ScheduleDAGSDNodes::ScheduleDAGSDNodes(MachineFunction &mf)
49 : ScheduleDAG(mf), BB(nullptr), DAG(nullptr),
50 InstrItins(mf.getTarget().getInstrItineraryData()) {}
52 /// Run - perform scheduling.
54 void ScheduleDAGSDNodes::Run(SelectionDAG *dag, MachineBasicBlock *bb) {
58 // Clear the scheduler's SUnit DAG.
59 ScheduleDAG::clearDAG();
62 // Invoke the target's selection of scheduler.
66 /// NewSUnit - Creates a new SUnit and return a ptr to it.
68 SUnit *ScheduleDAGSDNodes::newSUnit(SDNode *N) {
70 const SUnit *Addr = nullptr;
74 SUnits.push_back(SUnit(N, (unsigned)SUnits.size()));
75 assert((Addr == nullptr || Addr == &SUnits[0]) &&
76 "SUnits std::vector reallocated on the fly!");
77 SUnits.back().OrigNode = &SUnits.back();
78 SUnit *SU = &SUnits.back();
79 const TargetLowering &TLI = DAG->getTargetLoweringInfo();
81 (N->isMachineOpcode() &&
82 N->getMachineOpcode() == TargetOpcode::IMPLICIT_DEF))
83 SU->SchedulingPref = Sched::None;
85 SU->SchedulingPref = TLI.getSchedulingPreference(N);
89 SUnit *ScheduleDAGSDNodes::Clone(SUnit *Old) {
90 SUnit *SU = newSUnit(Old->getNode());
91 SU->OrigNode = Old->OrigNode;
92 SU->Latency = Old->Latency;
93 SU->isVRegCycle = Old->isVRegCycle;
94 SU->isCall = Old->isCall;
95 SU->isCallOp = Old->isCallOp;
96 SU->isTwoAddress = Old->isTwoAddress;
97 SU->isCommutable = Old->isCommutable;
98 SU->hasPhysRegDefs = Old->hasPhysRegDefs;
99 SU->hasPhysRegClobbers = Old->hasPhysRegClobbers;
100 SU->isScheduleHigh = Old->isScheduleHigh;
101 SU->isScheduleLow = Old->isScheduleLow;
102 SU->SchedulingPref = Old->SchedulingPref;
103 Old->isCloned = true;
107 /// CheckForPhysRegDependency - Check if the dependency between def and use of
108 /// a specified operand is a physical register dependency. If so, returns the
109 /// register and the cost of copying the register.
110 static void CheckForPhysRegDependency(SDNode *Def, SDNode *User, unsigned Op,
111 const TargetRegisterInfo *TRI,
112 const TargetInstrInfo *TII,
113 unsigned &PhysReg, int &Cost) {
114 if (Op != 2 || User->getOpcode() != ISD::CopyToReg)
117 unsigned Reg = cast<RegisterSDNode>(User->getOperand(1))->getReg();
118 if (TargetRegisterInfo::isVirtualRegister(Reg))
121 unsigned ResNo = User->getOperand(2).getResNo();
122 if (Def->isMachineOpcode()) {
123 const MCInstrDesc &II = TII->get(Def->getMachineOpcode());
124 if (ResNo >= II.getNumDefs() &&
125 II.ImplicitDefs[ResNo - II.getNumDefs()] == Reg) {
127 const TargetRegisterClass *RC =
128 TRI->getMinimalPhysRegClass(Reg, Def->getValueType(ResNo));
129 Cost = RC->getCopyCost();
134 // Helper for AddGlue to clone node operands.
135 static void CloneNodeWithValues(SDNode *N, SelectionDAG *DAG,
136 SmallVectorImpl<EVT> &VTs,
137 SDValue ExtraOper = SDValue()) {
138 SmallVector<SDValue, 4> Ops;
139 for (unsigned I = 0, E = N->getNumOperands(); I != E; ++I)
140 Ops.push_back(N->getOperand(I));
142 if (ExtraOper.getNode())
143 Ops.push_back(ExtraOper);
145 SDVTList VTList = DAG->getVTList(VTs);
146 MachineSDNode::mmo_iterator Begin = nullptr, End = nullptr;
147 MachineSDNode *MN = dyn_cast<MachineSDNode>(N);
149 // Store memory references.
151 Begin = MN->memoperands_begin();
152 End = MN->memoperands_end();
155 DAG->MorphNodeTo(N, N->getOpcode(), VTList, &Ops[0], Ops.size());
157 // Reset the memory references
159 MN->setMemRefs(Begin, End);
162 static bool AddGlue(SDNode *N, SDValue Glue, bool AddGlue, SelectionDAG *DAG) {
163 SmallVector<EVT, 4> VTs;
164 SDNode *GlueDestNode = Glue.getNode();
166 // Don't add glue from a node to itself.
167 if (GlueDestNode == N) return false;
169 // Don't add a glue operand to something that already uses glue.
171 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue) {
174 // Don't add glue to something that already has a glue value.
175 if (N->getValueType(N->getNumValues() - 1) == MVT::Glue) return false;
177 for (unsigned I = 0, E = N->getNumValues(); I != E; ++I)
178 VTs.push_back(N->getValueType(I));
181 VTs.push_back(MVT::Glue);
183 CloneNodeWithValues(N, DAG, VTs, Glue);
188 // Cleanup after unsuccessful AddGlue. Use the standard method of morphing the
189 // node even though simply shrinking the value list is sufficient.
190 static void RemoveUnusedGlue(SDNode *N, SelectionDAG *DAG) {
191 assert((N->getValueType(N->getNumValues() - 1) == MVT::Glue &&
192 !N->hasAnyUseOfValue(N->getNumValues() - 1)) &&
193 "expected an unused glue value");
195 SmallVector<EVT, 4> VTs;
196 for (unsigned I = 0, E = N->getNumValues()-1; I != E; ++I)
197 VTs.push_back(N->getValueType(I));
199 CloneNodeWithValues(N, DAG, VTs);
202 /// ClusterNeighboringLoads - Force nearby loads together by "gluing" them.
203 /// This function finds loads of the same base and different offsets. If the
204 /// offsets are not far apart (target specific), it add MVT::Glue inputs and
205 /// outputs to ensure they are scheduled together and in order. This
206 /// optimization may benefit some targets by improving cache locality.
207 void ScheduleDAGSDNodes::ClusterNeighboringLoads(SDNode *Node) {
208 SDNode *Chain = nullptr;
209 unsigned NumOps = Node->getNumOperands();
210 if (Node->getOperand(NumOps-1).getValueType() == MVT::Other)
211 Chain = Node->getOperand(NumOps-1).getNode();
215 // Look for other loads of the same chain. Find loads that are loading from
216 // the same base pointer and different offsets.
217 SmallPtrSet<SDNode*, 16> Visited;
218 SmallVector<int64_t, 4> Offsets;
219 DenseMap<long long, SDNode*> O2SMap; // Map from offset to SDNode.
220 bool Cluster = false;
222 // This algorithm requires a reasonably low use count before finding a match
223 // to avoid uselessly blowing up compile time in large blocks.
224 unsigned UseCount = 0;
225 for (SDNode::use_iterator I = Chain->use_begin(), E = Chain->use_end();
226 I != E && UseCount < 100; ++I, ++UseCount) {
228 if (User == Node || !Visited.insert(User))
230 int64_t Offset1, Offset2;
231 if (!TII->areLoadsFromSameBasePtr(Base, User, Offset1, Offset2) ||
233 // FIXME: Should be ok if they addresses are identical. But earlier
234 // optimizations really should have eliminated one of the loads.
236 if (O2SMap.insert(std::make_pair(Offset1, Base)).second)
237 Offsets.push_back(Offset1);
238 O2SMap.insert(std::make_pair(Offset2, User));
239 Offsets.push_back(Offset2);
240 if (Offset2 < Offset1)
243 // Reset UseCount to allow more matches.
250 // Sort them in increasing order.
251 std::sort(Offsets.begin(), Offsets.end());
253 // Check if the loads are close enough.
254 SmallVector<SDNode*, 4> Loads;
255 unsigned NumLoads = 0;
256 int64_t BaseOff = Offsets[0];
257 SDNode *BaseLoad = O2SMap[BaseOff];
258 Loads.push_back(BaseLoad);
259 for (unsigned i = 1, e = Offsets.size(); i != e; ++i) {
260 int64_t Offset = Offsets[i];
261 SDNode *Load = O2SMap[Offset];
262 if (!TII->shouldScheduleLoadsNear(BaseLoad, Load, BaseOff, Offset,NumLoads))
263 break; // Stop right here. Ignore loads that are further away.
264 Loads.push_back(Load);
271 // Cluster loads by adding MVT::Glue outputs and inputs. This also
272 // ensure they are scheduled in order of increasing addresses.
273 SDNode *Lead = Loads[0];
274 SDValue InGlue = SDValue(nullptr, 0);
275 if (AddGlue(Lead, InGlue, true, DAG))
276 InGlue = SDValue(Lead, Lead->getNumValues() - 1);
277 for (unsigned I = 1, E = Loads.size(); I != E; ++I) {
278 bool OutGlue = I < E - 1;
279 SDNode *Load = Loads[I];
281 // If AddGlue fails, we could leave an unsused glue value. This should not
283 if (AddGlue(Load, InGlue, OutGlue, DAG)) {
285 InGlue = SDValue(Load, Load->getNumValues() - 1);
289 else if (!OutGlue && InGlue.getNode())
290 RemoveUnusedGlue(InGlue.getNode(), DAG);
294 /// ClusterNodes - Cluster certain nodes which should be scheduled together.
296 void ScheduleDAGSDNodes::ClusterNodes() {
297 for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
298 E = DAG->allnodes_end(); NI != E; ++NI) {
300 if (!Node || !Node->isMachineOpcode())
303 unsigned Opc = Node->getMachineOpcode();
304 const MCInstrDesc &MCID = TII->get(Opc);
306 // Cluster loads from "near" addresses into combined SUnits.
307 ClusterNeighboringLoads(Node);
311 void ScheduleDAGSDNodes::BuildSchedUnits() {
312 // During scheduling, the NodeId field of SDNode is used to map SDNodes
313 // to their associated SUnits by holding SUnits table indices. A value
314 // of -1 means the SDNode does not yet have an associated SUnit.
315 unsigned NumNodes = 0;
316 for (SelectionDAG::allnodes_iterator NI = DAG->allnodes_begin(),
317 E = DAG->allnodes_end(); NI != E; ++NI) {
322 // Reserve entries in the vector for each of the SUnits we are creating. This
323 // ensure that reallocation of the vector won't happen, so SUnit*'s won't get
325 // FIXME: Multiply by 2 because we may clone nodes during scheduling.
326 // This is a temporary workaround.
327 SUnits.reserve(NumNodes * 2);
329 // Add all nodes in depth first order.
330 SmallVector<SDNode*, 64> Worklist;
331 SmallPtrSet<SDNode*, 64> Visited;
332 Worklist.push_back(DAG->getRoot().getNode());
333 Visited.insert(DAG->getRoot().getNode());
335 SmallVector<SUnit*, 8> CallSUnits;
336 while (!Worklist.empty()) {
337 SDNode *NI = Worklist.pop_back_val();
339 // Add all operands to the worklist unless they've already been added.
340 for (unsigned i = 0, e = NI->getNumOperands(); i != e; ++i)
341 if (Visited.insert(NI->getOperand(i).getNode()))
342 Worklist.push_back(NI->getOperand(i).getNode());
344 if (isPassiveNode(NI)) // Leaf node, e.g. a TargetImmediate.
347 // If this node has already been processed, stop now.
348 if (NI->getNodeId() != -1) continue;
350 SUnit *NodeSUnit = newSUnit(NI);
352 // See if anything is glued to this node, if so, add them to glued
353 // nodes. Nodes can have at most one glue input and one glue output. Glue
354 // is required to be the last operand and result of a node.
356 // Scan up to find glued preds.
358 while (N->getNumOperands() &&
359 N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue) {
360 N = N->getOperand(N->getNumOperands()-1).getNode();
361 assert(N->getNodeId() == -1 && "Node already inserted!");
362 N->setNodeId(NodeSUnit->NodeNum);
363 if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall())
364 NodeSUnit->isCall = true;
367 // Scan down to find any glued succs.
369 while (N->getValueType(N->getNumValues()-1) == MVT::Glue) {
370 SDValue GlueVal(N, N->getNumValues()-1);
372 // There are either zero or one users of the Glue result.
373 bool HasGlueUse = false;
374 for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
376 if (GlueVal.isOperandOf(*UI)) {
378 assert(N->getNodeId() == -1 && "Node already inserted!");
379 N->setNodeId(NodeSUnit->NodeNum);
381 if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall())
382 NodeSUnit->isCall = true;
385 if (!HasGlueUse) break;
388 if (NodeSUnit->isCall)
389 CallSUnits.push_back(NodeSUnit);
391 // Schedule zero-latency TokenFactor below any nodes that may increase the
392 // schedule height. Otherwise, ancestors of the TokenFactor may appear to
393 // have false stalls.
394 if (NI->getOpcode() == ISD::TokenFactor)
395 NodeSUnit->isScheduleLow = true;
397 // If there are glue operands involved, N is now the bottom-most node
398 // of the sequence of nodes that are glued together.
400 NodeSUnit->setNode(N);
401 assert(N->getNodeId() == -1 && "Node already inserted!");
402 N->setNodeId(NodeSUnit->NodeNum);
404 // Compute NumRegDefsLeft. This must be done before AddSchedEdges.
405 InitNumRegDefsLeft(NodeSUnit);
407 // Assign the Latency field of NodeSUnit using target-provided information.
408 computeLatency(NodeSUnit);
411 // Find all call operands.
412 while (!CallSUnits.empty()) {
413 SUnit *SU = CallSUnits.pop_back_val();
414 for (const SDNode *SUNode = SU->getNode(); SUNode;
415 SUNode = SUNode->getGluedNode()) {
416 if (SUNode->getOpcode() != ISD::CopyToReg)
418 SDNode *SrcN = SUNode->getOperand(2).getNode();
419 if (isPassiveNode(SrcN)) continue; // Not scheduled.
420 SUnit *SrcSU = &SUnits[SrcN->getNodeId()];
421 SrcSU->isCallOp = true;
426 void ScheduleDAGSDNodes::AddSchedEdges() {
427 const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
429 // Check to see if the scheduler cares about latencies.
430 bool UnitLatencies = forceUnitLatencies();
432 // Pass 2: add the preds, succs, etc.
433 for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
434 SUnit *SU = &SUnits[su];
435 SDNode *MainNode = SU->getNode();
437 if (MainNode->isMachineOpcode()) {
438 unsigned Opc = MainNode->getMachineOpcode();
439 const MCInstrDesc &MCID = TII->get(Opc);
440 for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
441 if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
442 SU->isTwoAddress = true;
446 if (MCID.isCommutable())
447 SU->isCommutable = true;
450 // Find all predecessors and successors of the group.
451 for (SDNode *N = SU->getNode(); N; N = N->getGluedNode()) {
452 if (N->isMachineOpcode() &&
453 TII->get(N->getMachineOpcode()).getImplicitDefs()) {
454 SU->hasPhysRegClobbers = true;
455 unsigned NumUsed = InstrEmitter::CountResults(N);
456 while (NumUsed != 0 && !N->hasAnyUseOfValue(NumUsed - 1))
457 --NumUsed; // Skip over unused values at the end.
458 if (NumUsed > TII->get(N->getMachineOpcode()).getNumDefs())
459 SU->hasPhysRegDefs = true;
462 for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
463 SDNode *OpN = N->getOperand(i).getNode();
464 if (isPassiveNode(OpN)) continue; // Not scheduled.
465 SUnit *OpSU = &SUnits[OpN->getNodeId()];
466 assert(OpSU && "Node has no SUnit!");
467 if (OpSU == SU) continue; // In the same group.
469 EVT OpVT = N->getOperand(i).getValueType();
470 assert(OpVT != MVT::Glue && "Glued nodes should be in same sunit!");
471 bool isChain = OpVT == MVT::Other;
473 unsigned PhysReg = 0;
475 // Determine if this is a physical register dependency.
476 CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost);
477 assert((PhysReg == 0 || !isChain) &&
478 "Chain dependence via physreg data?");
479 // FIXME: See ScheduleDAGSDNodes::EmitCopyFromReg. For now, scheduler
480 // emits a copy from the physical register to a virtual register unless
481 // it requires a cross class copy (cost < 0). That means we are only
482 // treating "expensive to copy" register dependency as physical register
483 // dependency. This may change in the future though.
484 if (Cost >= 0 && !StressSched)
487 // If this is a ctrl dep, latency is 1.
488 unsigned OpLatency = isChain ? 1 : OpSU->Latency;
489 // Special-case TokenFactor chains as zero-latency.
490 if(isChain && OpN->getOpcode() == ISD::TokenFactor)
493 SDep Dep = isChain ? SDep(OpSU, SDep::Barrier)
494 : SDep(OpSU, SDep::Data, PhysReg);
495 Dep.setLatency(OpLatency);
496 if (!isChain && !UnitLatencies) {
497 computeOperandLatency(OpN, N, i, Dep);
498 ST.adjustSchedDependency(OpSU, SU, Dep);
501 if (!SU->addPred(Dep) && !Dep.isCtrl() && OpSU->NumRegDefsLeft > 1) {
502 // Multiple register uses are combined in the same SUnit. For example,
503 // we could have a set of glued nodes with all their defs consumed by
504 // another set of glued nodes. Register pressure tracking sees this as
505 // a single use, so to keep pressure balanced we reduce the defs.
507 // We can't tell (without more book-keeping) if this results from
508 // glued nodes or duplicate operands. As long as we don't reduce
509 // NumRegDefsLeft to zero, we handle the common cases well.
510 --OpSU->NumRegDefsLeft;
517 /// BuildSchedGraph - Build the SUnit graph from the selection dag that we
518 /// are input. This SUnit graph is similar to the SelectionDAG, but
519 /// excludes nodes that aren't interesting to scheduling, and represents
520 /// glued together nodes with a single SUnit.
521 void ScheduleDAGSDNodes::BuildSchedGraph(AliasAnalysis *AA) {
522 // Cluster certain nodes which should be scheduled together.
524 // Populate the SUnits array.
526 // Compute all the scheduling dependencies between nodes.
530 // Initialize NumNodeDefs for the current Node's opcode.
531 void ScheduleDAGSDNodes::RegDefIter::InitNodeNumDefs() {
532 // Check for phys reg copy.
536 if (!Node->isMachineOpcode()) {
537 if (Node->getOpcode() == ISD::CopyFromReg)
543 unsigned POpc = Node->getMachineOpcode();
544 if (POpc == TargetOpcode::IMPLICIT_DEF) {
545 // No register need be allocated for this.
549 unsigned NRegDefs = SchedDAG->TII->get(Node->getMachineOpcode()).getNumDefs();
550 // Some instructions define regs that are not represented in the selection DAG
551 // (e.g. unused flags). See tMOVi8. Make sure we don't access past NumValues.
552 NodeNumDefs = std::min(Node->getNumValues(), NRegDefs);
556 // Construct a RegDefIter for this SUnit and find the first valid value.
557 ScheduleDAGSDNodes::RegDefIter::RegDefIter(const SUnit *SU,
558 const ScheduleDAGSDNodes *SD)
559 : SchedDAG(SD), Node(SU->getNode()), DefIdx(0), NodeNumDefs(0) {
564 // Advance to the next valid value defined by the SUnit.
565 void ScheduleDAGSDNodes::RegDefIter::Advance() {
566 for (;Node;) { // Visit all glued nodes.
567 for (;DefIdx < NodeNumDefs; ++DefIdx) {
568 if (!Node->hasAnyUseOfValue(DefIdx))
570 ValueType = Node->getSimpleValueType(DefIdx);
572 return; // Found a normal regdef.
574 Node = Node->getGluedNode();
576 return; // No values left to visit.
582 void ScheduleDAGSDNodes::InitNumRegDefsLeft(SUnit *SU) {
583 assert(SU->NumRegDefsLeft == 0 && "expect a new node");
584 for (RegDefIter I(SU, this); I.IsValid(); I.Advance()) {
585 assert(SU->NumRegDefsLeft < USHRT_MAX && "overflow is ok but unexpected");
586 ++SU->NumRegDefsLeft;
590 void ScheduleDAGSDNodes::computeLatency(SUnit *SU) {
591 SDNode *N = SU->getNode();
593 // TokenFactor operands are considered zero latency, and some schedulers
594 // (e.g. Top-Down list) may rely on the fact that operand latency is nonzero
595 // whenever node latency is nonzero.
596 if (N && N->getOpcode() == ISD::TokenFactor) {
601 // Check to see if the scheduler cares about latencies.
602 if (forceUnitLatencies()) {
607 if (!InstrItins || InstrItins->isEmpty()) {
608 if (N && N->isMachineOpcode() &&
609 TII->isHighLatencyDef(N->getMachineOpcode()))
610 SU->Latency = HighLatencyCycles;
616 // Compute the latency for the node. We use the sum of the latencies for
617 // all nodes glued together into this SUnit.
619 for (SDNode *N = SU->getNode(); N; N = N->getGluedNode())
620 if (N->isMachineOpcode())
621 SU->Latency += TII->getInstrLatency(InstrItins, N);
624 void ScheduleDAGSDNodes::computeOperandLatency(SDNode *Def, SDNode *Use,
625 unsigned OpIdx, SDep& dep) const{
626 // Check to see if the scheduler cares about latencies.
627 if (forceUnitLatencies())
630 if (dep.getKind() != SDep::Data)
633 unsigned DefIdx = Use->getOperand(OpIdx).getResNo();
634 if (Use->isMachineOpcode())
635 // Adjust the use operand index by num of defs.
636 OpIdx += TII->get(Use->getMachineOpcode()).getNumDefs();
637 int Latency = TII->getOperandLatency(InstrItins, Def, DefIdx, Use, OpIdx);
638 if (Latency > 1 && Use->getOpcode() == ISD::CopyToReg &&
640 unsigned Reg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
641 if (TargetRegisterInfo::isVirtualRegister(Reg))
642 // This copy is a liveout value. It is likely coalesced, so reduce the
643 // latency so not to penalize the def.
644 // FIXME: need target specific adjustment here?
645 Latency = (Latency > 1) ? Latency - 1 : 1;
648 dep.setLatency(Latency);
651 void ScheduleDAGSDNodes::dumpNode(const SUnit *SU) const {
652 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
653 if (!SU->getNode()) {
654 dbgs() << "PHYS REG COPY\n";
658 SU->getNode()->dump(DAG);
660 SmallVector<SDNode *, 4> GluedNodes;
661 for (SDNode *N = SU->getNode()->getGluedNode(); N; N = N->getGluedNode())
662 GluedNodes.push_back(N);
663 while (!GluedNodes.empty()) {
665 GluedNodes.back()->dump(DAG);
667 GluedNodes.pop_back();
672 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
673 void ScheduleDAGSDNodes::dumpSchedule() const {
674 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
675 if (SUnit *SU = Sequence[i])
678 dbgs() << "**** NOOP ****\n";
684 /// VerifyScheduledSequence - Verify that all SUnits were scheduled and that
685 /// their state is consistent with the nodes listed in Sequence.
687 void ScheduleDAGSDNodes::VerifyScheduledSequence(bool isBottomUp) {
688 unsigned ScheduledNodes = ScheduleDAG::VerifyScheduledDAG(isBottomUp);
690 for (unsigned i = 0, e = Sequence.size(); i != e; ++i)
693 assert(Sequence.size() - Noops == ScheduledNodes &&
694 "The number of nodes scheduled doesn't match the expected number!");
698 /// ProcessSDDbgValues - Process SDDbgValues associated with this node.
700 ProcessSDDbgValues(SDNode *N, SelectionDAG *DAG, InstrEmitter &Emitter,
701 SmallVectorImpl<std::pair<unsigned, MachineInstr*> > &Orders,
702 DenseMap<SDValue, unsigned> &VRBaseMap, unsigned Order) {
703 if (!N->getHasDebugValue())
706 // Opportunistically insert immediate dbg_value uses, i.e. those with source
707 // order number right after the N.
708 MachineBasicBlock *BB = Emitter.getBlock();
709 MachineBasicBlock::iterator InsertPos = Emitter.getInsertPos();
710 ArrayRef<SDDbgValue*> DVs = DAG->GetDbgValues(N);
711 for (unsigned i = 0, e = DVs.size(); i != e; ++i) {
712 if (DVs[i]->isInvalidated())
714 unsigned DVOrder = DVs[i]->getOrder();
715 if (!Order || DVOrder == ++Order) {
716 MachineInstr *DbgMI = Emitter.EmitDbgValue(DVs[i], VRBaseMap);
718 Orders.push_back(std::make_pair(DVOrder, DbgMI));
719 BB->insert(InsertPos, DbgMI);
721 DVs[i]->setIsInvalidated();
726 // ProcessSourceNode - Process nodes with source order numbers. These are added
727 // to a vector which EmitSchedule uses to determine how to insert dbg_value
728 // instructions in the right order.
730 ProcessSourceNode(SDNode *N, SelectionDAG *DAG, InstrEmitter &Emitter,
731 DenseMap<SDValue, unsigned> &VRBaseMap,
732 SmallVectorImpl<std::pair<unsigned, MachineInstr*> > &Orders,
733 SmallSet<unsigned, 8> &Seen) {
734 unsigned Order = N->getIROrder();
735 if (!Order || !Seen.insert(Order)) {
736 // Process any valid SDDbgValues even if node does not have any order
738 ProcessSDDbgValues(N, DAG, Emitter, Orders, VRBaseMap, 0);
742 MachineBasicBlock *BB = Emitter.getBlock();
743 if (Emitter.getInsertPos() == BB->begin() || BB->back().isPHI() ||
744 // Fast-isel may have inserted some instructions, in which case the
745 // BB->back().isPHI() test will not fire when we want it to.
746 std::prev(Emitter.getInsertPos())->isPHI()) {
747 // Did not insert any instruction.
748 Orders.push_back(std::make_pair(Order, (MachineInstr*)nullptr));
752 Orders.push_back(std::make_pair(Order, std::prev(Emitter.getInsertPos())));
753 ProcessSDDbgValues(N, DAG, Emitter, Orders, VRBaseMap, Order);
756 void ScheduleDAGSDNodes::
757 EmitPhysRegCopy(SUnit *SU, DenseMap<SUnit*, unsigned> &VRBaseMap,
758 MachineBasicBlock::iterator InsertPos) {
759 for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
761 if (I->isCtrl()) continue; // ignore chain preds
762 if (I->getSUnit()->CopyDstRC) {
763 // Copy to physical register.
764 DenseMap<SUnit*, unsigned>::iterator VRI = VRBaseMap.find(I->getSUnit());
765 assert(VRI != VRBaseMap.end() && "Node emitted out of order - late");
766 // Find the destination physical register.
768 for (SUnit::const_succ_iterator II = SU->Succs.begin(),
769 EE = SU->Succs.end(); II != EE; ++II) {
770 if (II->isCtrl()) continue; // ignore chain preds
776 BuildMI(*BB, InsertPos, DebugLoc(), TII->get(TargetOpcode::COPY), Reg)
777 .addReg(VRI->second);
779 // Copy from physical register.
780 assert(I->getReg() && "Unknown physical register!");
781 unsigned VRBase = MRI.createVirtualRegister(SU->CopyDstRC);
782 bool isNew = VRBaseMap.insert(std::make_pair(SU, VRBase)).second;
783 (void)isNew; // Silence compiler warning.
784 assert(isNew && "Node emitted out of order - early");
785 BuildMI(*BB, InsertPos, DebugLoc(), TII->get(TargetOpcode::COPY), VRBase)
786 .addReg(I->getReg());
792 /// EmitSchedule - Emit the machine code in scheduled order. Return the new
793 /// InsertPos and MachineBasicBlock that contains this insertion
794 /// point. ScheduleDAGSDNodes holds a BB pointer for convenience, but this does
795 /// not necessarily refer to returned BB. The emitter may split blocks.
796 MachineBasicBlock *ScheduleDAGSDNodes::
797 EmitSchedule(MachineBasicBlock::iterator &InsertPos) {
798 InstrEmitter Emitter(BB, InsertPos);
799 DenseMap<SDValue, unsigned> VRBaseMap;
800 DenseMap<SUnit*, unsigned> CopyVRBaseMap;
801 SmallVector<std::pair<unsigned, MachineInstr*>, 32> Orders;
802 SmallSet<unsigned, 8> Seen;
803 bool HasDbg = DAG->hasDebugValues();
805 // If this is the first BB, emit byval parameter dbg_value's.
806 if (HasDbg && BB->getParent()->begin() == MachineFunction::iterator(BB)) {
807 SDDbgInfo::DbgIterator PDI = DAG->ByvalParmDbgBegin();
808 SDDbgInfo::DbgIterator PDE = DAG->ByvalParmDbgEnd();
809 for (; PDI != PDE; ++PDI) {
810 MachineInstr *DbgMI= Emitter.EmitDbgValue(*PDI, VRBaseMap);
812 BB->insert(InsertPos, DbgMI);
816 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
817 SUnit *SU = Sequence[i];
819 // Null SUnit* is a noop.
820 TII->insertNoop(*Emitter.getBlock(), InsertPos);
824 // For pre-regalloc scheduling, create instructions corresponding to the
825 // SDNode and any glued SDNodes and append them to the block.
826 if (!SU->getNode()) {
828 EmitPhysRegCopy(SU, CopyVRBaseMap, InsertPos);
832 SmallVector<SDNode *, 4> GluedNodes;
833 for (SDNode *N = SU->getNode()->getGluedNode(); N; N = N->getGluedNode())
834 GluedNodes.push_back(N);
835 while (!GluedNodes.empty()) {
836 SDNode *N = GluedNodes.back();
837 Emitter.EmitNode(GluedNodes.back(), SU->OrigNode != SU, SU->isCloned,
839 // Remember the source order of the inserted instruction.
841 ProcessSourceNode(N, DAG, Emitter, VRBaseMap, Orders, Seen);
842 GluedNodes.pop_back();
844 Emitter.EmitNode(SU->getNode(), SU->OrigNode != SU, SU->isCloned,
846 // Remember the source order of the inserted instruction.
848 ProcessSourceNode(SU->getNode(), DAG, Emitter, VRBaseMap, Orders,
852 // Insert all the dbg_values which have not already been inserted in source
855 MachineBasicBlock::iterator BBBegin = BB->getFirstNonPHI();
857 // Sort the source order instructions and use the order to insert debug
859 std::sort(Orders.begin(), Orders.end(), less_first());
861 SDDbgInfo::DbgIterator DI = DAG->DbgBegin();
862 SDDbgInfo::DbgIterator DE = DAG->DbgEnd();
863 // Now emit the rest according to source order.
864 unsigned LastOrder = 0;
865 for (unsigned i = 0, e = Orders.size(); i != e && DI != DE; ++i) {
866 unsigned Order = Orders[i].first;
867 MachineInstr *MI = Orders[i].second;
868 // Insert all SDDbgValue's whose order(s) are before "Order".
872 (*DI)->getOrder() >= LastOrder && (*DI)->getOrder() < Order; ++DI) {
873 if ((*DI)->isInvalidated())
875 MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap);
878 // Insert to start of the BB (after PHIs).
879 BB->insert(BBBegin, DbgMI);
881 // Insert at the instruction, which may be in a different
882 // block, if the block was split by a custom inserter.
883 MachineBasicBlock::iterator Pos = MI;
884 MI->getParent()->insert(Pos, DbgMI);
890 // Add trailing DbgValue's before the terminator. FIXME: May want to add
891 // some of them before one or more conditional branches?
892 SmallVector<MachineInstr*, 8> DbgMIs;
894 if (!(*DI)->isInvalidated())
895 if (MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap))
896 DbgMIs.push_back(DbgMI);
900 MachineBasicBlock *InsertBB = Emitter.getBlock();
901 MachineBasicBlock::iterator Pos = InsertBB->getFirstTerminator();
902 InsertBB->insert(Pos, DbgMIs.begin(), DbgMIs.end());
905 InsertPos = Emitter.getInsertPos();
906 return Emitter.getBlock();
909 /// Return the basic block label.
910 std::string ScheduleDAGSDNodes::getDAGName() const {
911 return "sunit-dag." + BB->getFullName();