xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/SelectionDAG/ScheduleDAGSDNodes.cpp (revision 734e82fe33aa764367791a7d603b383996c6b40b)
1 //===--- ScheduleDAGSDNodes.cpp - Implement the ScheduleDAGSDNodes class --===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This implements the ScheduleDAG class, which is a base class used by
10 // scheduling implementation classes.
11 //
12 //===----------------------------------------------------------------------===//
13 
14 #include "ScheduleDAGSDNodes.h"
15 #include "InstrEmitter.h"
16 #include "SDNodeDbgValue.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SmallPtrSet.h"
19 #include "llvm/ADT/SmallSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/CodeGen/MachineInstrBuilder.h"
23 #include "llvm/CodeGen/MachineRegisterInfo.h"
24 #include "llvm/CodeGen/SelectionDAG.h"
25 #include "llvm/CodeGen/TargetInstrInfo.h"
26 #include "llvm/CodeGen/TargetLowering.h"
27 #include "llvm/CodeGen/TargetRegisterInfo.h"
28 #include "llvm/CodeGen/TargetSubtargetInfo.h"
29 #include "llvm/Config/llvm-config.h"
30 #include "llvm/MC/MCInstrItineraries.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/Target/TargetMachine.h"
35 using namespace llvm;
36 
37 #define DEBUG_TYPE "pre-RA-sched"
38 
39 STATISTIC(LoadsClustered, "Number of loads clustered together");
40 
41 // This allows the latency-based scheduler to notice high latency instructions
42 // without a target itinerary. The choice of number here has more to do with
43 // balancing scheduler heuristics than with the actual machine latency.
44 static cl::opt<int> HighLatencyCycles(
45   "sched-high-latency-cycles", cl::Hidden, cl::init(10),
46   cl::desc("Roughly estimate the number of cycles that 'long latency'"
47            "instructions take for targets with no itinerary"));
48 
49 ScheduleDAGSDNodes::ScheduleDAGSDNodes(MachineFunction &mf)
50     : ScheduleDAG(mf), InstrItins(mf.getSubtarget().getInstrItineraryData()) {}
51 
52 /// Run - perform scheduling.
53 ///
54 void ScheduleDAGSDNodes::Run(SelectionDAG *dag, MachineBasicBlock *bb) {
55   BB = bb;
56   DAG = dag;
57 
58   // Clear the scheduler's SUnit DAG.
59   ScheduleDAG::clearDAG();
60   Sequence.clear();
61 
62   // Invoke the target's selection of scheduler.
63   Schedule();
64 }
65 
66 /// NewSUnit - Creates a new SUnit and return a ptr to it.
67 ///
68 SUnit *ScheduleDAGSDNodes::newSUnit(SDNode *N) {
69 #ifndef NDEBUG
70   const SUnit *Addr = nullptr;
71   if (!SUnits.empty())
72     Addr = &SUnits[0];
73 #endif
74   SUnits.emplace_back(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();
80   if (!N ||
81       (N->isMachineOpcode() &&
82        N->getMachineOpcode() == TargetOpcode::IMPLICIT_DEF))
83     SU->SchedulingPref = Sched::None;
84   else
85     SU->SchedulingPref = TLI.getSchedulingPreference(N);
86   return SU;
87 }
88 
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;
104   return SU;
105 }
106 
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                                       const TargetLowering &TLI,
114                                       unsigned &PhysReg, int &Cost) {
115   if (Op != 2 || User->getOpcode() != ISD::CopyToReg)
116     return;
117 
118   unsigned Reg = cast<RegisterSDNode>(User->getOperand(1))->getReg();
119   if (TLI.checkForPhysRegDependency(Def, User, Op, TRI, TII, PhysReg, Cost))
120     return;
121 
122   if (Register::isVirtualRegister(Reg))
123     return;
124 
125   unsigned ResNo = User->getOperand(2).getResNo();
126   if (Def->getOpcode() == ISD::CopyFromReg &&
127       cast<RegisterSDNode>(Def->getOperand(1))->getReg() == Reg) {
128     PhysReg = Reg;
129   } else if (Def->isMachineOpcode()) {
130     const MCInstrDesc &II = TII->get(Def->getMachineOpcode());
131     if (ResNo >= II.getNumDefs() && II.hasImplicitDefOfPhysReg(Reg))
132       PhysReg = Reg;
133   }
134 
135   if (PhysReg != 0) {
136     const TargetRegisterClass *RC =
137         TRI->getMinimalPhysRegClass(Reg, Def->getSimpleValueType(ResNo));
138     Cost = RC->getCopyCost();
139   }
140 }
141 
142 // Helper for AddGlue to clone node operands.
143 static void CloneNodeWithValues(SDNode *N, SelectionDAG *DAG, ArrayRef<EVT> VTs,
144                                 SDValue ExtraOper = SDValue()) {
145   SmallVector<SDValue, 8> Ops(N->op_begin(), N->op_end());
146   if (ExtraOper.getNode())
147     Ops.push_back(ExtraOper);
148 
149   SDVTList VTList = DAG->getVTList(VTs);
150   MachineSDNode *MN = dyn_cast<MachineSDNode>(N);
151 
152   // Store memory references.
153   SmallVector<MachineMemOperand *, 2> MMOs;
154   if (MN)
155     MMOs.assign(MN->memoperands_begin(), MN->memoperands_end());
156 
157   DAG->MorphNodeTo(N, N->getOpcode(), VTList, Ops);
158 
159   // Reset the memory references
160   if (MN)
161     DAG->setNodeMemRefs(MN, MMOs);
162 }
163 
164 static bool AddGlue(SDNode *N, SDValue Glue, bool AddGlue, SelectionDAG *DAG) {
165   SDNode *GlueDestNode = Glue.getNode();
166 
167   // Don't add glue from a node to itself.
168   if (GlueDestNode == N) return false;
169 
170   // Don't add a glue operand to something that already uses glue.
171   if (GlueDestNode &&
172       N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue) {
173     return false;
174   }
175   // Don't add glue to something that already has a glue value.
176   if (N->getValueType(N->getNumValues() - 1) == MVT::Glue) return false;
177 
178   SmallVector<EVT, 4> VTs(N->values());
179   if (AddGlue)
180     VTs.push_back(MVT::Glue);
181 
182   CloneNodeWithValues(N, DAG, VTs, Glue);
183 
184   return true;
185 }
186 
187 // Cleanup after unsuccessful AddGlue. Use the standard method of morphing the
188 // node even though simply shrinking the value list is sufficient.
189 static void RemoveUnusedGlue(SDNode *N, SelectionDAG *DAG) {
190   assert((N->getValueType(N->getNumValues() - 1) == MVT::Glue &&
191           !N->hasAnyUseOfValue(N->getNumValues() - 1)) &&
192          "expected an unused glue value");
193 
194   CloneNodeWithValues(N, DAG,
195                       ArrayRef(N->value_begin(), N->getNumValues() - 1));
196 }
197 
198 /// ClusterNeighboringLoads - Force nearby loads together by "gluing" them.
199 /// This function finds loads of the same base and different offsets. If the
200 /// offsets are not far apart (target specific), it add MVT::Glue inputs and
201 /// outputs to ensure they are scheduled together and in order. This
202 /// optimization may benefit some targets by improving cache locality.
203 void ScheduleDAGSDNodes::ClusterNeighboringLoads(SDNode *Node) {
204   SDValue Chain;
205   unsigned NumOps = Node->getNumOperands();
206   if (Node->getOperand(NumOps-1).getValueType() == MVT::Other)
207     Chain = Node->getOperand(NumOps-1);
208   if (!Chain)
209     return;
210 
211   // Skip any load instruction that has a tied input. There may be an additional
212   // dependency requiring a different order than by increasing offsets, and the
213   // added glue may introduce a cycle.
214   auto hasTiedInput = [this](const SDNode *N) {
215     const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
216     for (unsigned I = 0; I != MCID.getNumOperands(); ++I) {
217       if (MCID.getOperandConstraint(I, MCOI::TIED_TO) != -1)
218         return true;
219     }
220 
221     return false;
222   };
223 
224   // Look for other loads of the same chain. Find loads that are loading from
225   // the same base pointer and different offsets.
226   SmallPtrSet<SDNode*, 16> Visited;
227   SmallVector<int64_t, 4> Offsets;
228   DenseMap<long long, SDNode*> O2SMap;  // Map from offset to SDNode.
229   bool Cluster = false;
230   SDNode *Base = Node;
231 
232   if (hasTiedInput(Base))
233     return;
234 
235   // This algorithm requires a reasonably low use count before finding a match
236   // to avoid uselessly blowing up compile time in large blocks.
237   unsigned UseCount = 0;
238   for (SDNode::use_iterator I = Chain->use_begin(), E = Chain->use_end();
239        I != E && UseCount < 100; ++I, ++UseCount) {
240     if (I.getUse().getResNo() != Chain.getResNo())
241       continue;
242 
243     SDNode *User = *I;
244     if (User == Node || !Visited.insert(User).second)
245       continue;
246     int64_t Offset1, Offset2;
247     if (!TII->areLoadsFromSameBasePtr(Base, User, Offset1, Offset2) ||
248         Offset1 == Offset2 ||
249         hasTiedInput(User)) {
250       // FIXME: Should be ok if they addresses are identical. But earlier
251       // optimizations really should have eliminated one of the loads.
252       continue;
253     }
254     if (O2SMap.insert(std::make_pair(Offset1, Base)).second)
255       Offsets.push_back(Offset1);
256     O2SMap.insert(std::make_pair(Offset2, User));
257     Offsets.push_back(Offset2);
258     if (Offset2 < Offset1)
259       Base = User;
260     Cluster = true;
261     // Reset UseCount to allow more matches.
262     UseCount = 0;
263   }
264 
265   if (!Cluster)
266     return;
267 
268   // Sort them in increasing order.
269   llvm::sort(Offsets);
270 
271   // Check if the loads are close enough.
272   SmallVector<SDNode*, 4> Loads;
273   unsigned NumLoads = 0;
274   int64_t BaseOff = Offsets[0];
275   SDNode *BaseLoad = O2SMap[BaseOff];
276   Loads.push_back(BaseLoad);
277   for (unsigned i = 1, e = Offsets.size(); i != e; ++i) {
278     int64_t Offset = Offsets[i];
279     SDNode *Load = O2SMap[Offset];
280     if (!TII->shouldScheduleLoadsNear(BaseLoad, Load, BaseOff, Offset,NumLoads))
281       break; // Stop right here. Ignore loads that are further away.
282     Loads.push_back(Load);
283     ++NumLoads;
284   }
285 
286   if (NumLoads == 0)
287     return;
288 
289   // Cluster loads by adding MVT::Glue outputs and inputs. This also
290   // ensure they are scheduled in order of increasing addresses.
291   SDNode *Lead = Loads[0];
292   SDValue InGlue;
293   if (AddGlue(Lead, InGlue, true, DAG))
294     InGlue = SDValue(Lead, Lead->getNumValues() - 1);
295   for (unsigned I = 1, E = Loads.size(); I != E; ++I) {
296     bool OutGlue = I < E - 1;
297     SDNode *Load = Loads[I];
298 
299     // If AddGlue fails, we could leave an unsused glue value. This should not
300     // cause any
301     if (AddGlue(Load, InGlue, OutGlue, DAG)) {
302       if (OutGlue)
303         InGlue = SDValue(Load, Load->getNumValues() - 1);
304 
305       ++LoadsClustered;
306     }
307     else if (!OutGlue && InGlue.getNode())
308       RemoveUnusedGlue(InGlue.getNode(), DAG);
309   }
310 }
311 
312 /// ClusterNodes - Cluster certain nodes which should be scheduled together.
313 ///
314 void ScheduleDAGSDNodes::ClusterNodes() {
315   for (SDNode &NI : DAG->allnodes()) {
316     SDNode *Node = &NI;
317     if (!Node || !Node->isMachineOpcode())
318       continue;
319 
320     unsigned Opc = Node->getMachineOpcode();
321     const MCInstrDesc &MCID = TII->get(Opc);
322     if (MCID.mayLoad())
323       // Cluster loads from "near" addresses into combined SUnits.
324       ClusterNeighboringLoads(Node);
325   }
326 }
327 
328 void ScheduleDAGSDNodes::BuildSchedUnits() {
329   // During scheduling, the NodeId field of SDNode is used to map SDNodes
330   // to their associated SUnits by holding SUnits table indices. A value
331   // of -1 means the SDNode does not yet have an associated SUnit.
332   unsigned NumNodes = 0;
333   for (SDNode &NI : DAG->allnodes()) {
334     NI.setNodeId(-1);
335     ++NumNodes;
336   }
337 
338   // Reserve entries in the vector for each of the SUnits we are creating.  This
339   // ensure that reallocation of the vector won't happen, so SUnit*'s won't get
340   // invalidated.
341   // FIXME: Multiply by 2 because we may clone nodes during scheduling.
342   // This is a temporary workaround.
343   SUnits.reserve(NumNodes * 2);
344 
345   // Add all nodes in depth first order.
346   SmallVector<SDNode*, 64> Worklist;
347   SmallPtrSet<SDNode*, 32> Visited;
348   Worklist.push_back(DAG->getRoot().getNode());
349   Visited.insert(DAG->getRoot().getNode());
350 
351   SmallVector<SUnit*, 8> CallSUnits;
352   while (!Worklist.empty()) {
353     SDNode *NI = Worklist.pop_back_val();
354 
355     // Add all operands to the worklist unless they've already been added.
356     for (const SDValue &Op : NI->op_values())
357       if (Visited.insert(Op.getNode()).second)
358         Worklist.push_back(Op.getNode());
359 
360     if (isPassiveNode(NI))  // Leaf node, e.g. a TargetImmediate.
361       continue;
362 
363     // If this node has already been processed, stop now.
364     if (NI->getNodeId() != -1) continue;
365 
366     SUnit *NodeSUnit = newSUnit(NI);
367 
368     // See if anything is glued to this node, if so, add them to glued
369     // nodes.  Nodes can have at most one glue input and one glue output.  Glue
370     // is required to be the last operand and result of a node.
371 
372     // Scan up to find glued preds.
373     SDNode *N = NI;
374     while (N->getNumOperands() &&
375            N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue) {
376       N = N->getOperand(N->getNumOperands()-1).getNode();
377       assert(N->getNodeId() == -1 && "Node already inserted!");
378       N->setNodeId(NodeSUnit->NodeNum);
379       if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall())
380         NodeSUnit->isCall = true;
381     }
382 
383     // Scan down to find any glued succs.
384     N = NI;
385     while (N->getValueType(N->getNumValues()-1) == MVT::Glue) {
386       SDValue GlueVal(N, N->getNumValues()-1);
387 
388       // There are either zero or one users of the Glue result.
389       bool HasGlueUse = false;
390       for (SDNode *U : N->uses())
391         if (GlueVal.isOperandOf(U)) {
392           HasGlueUse = true;
393           assert(N->getNodeId() == -1 && "Node already inserted!");
394           N->setNodeId(NodeSUnit->NodeNum);
395           N = U;
396           if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall())
397             NodeSUnit->isCall = true;
398           break;
399         }
400       if (!HasGlueUse) break;
401     }
402 
403     if (NodeSUnit->isCall)
404       CallSUnits.push_back(NodeSUnit);
405 
406     // Schedule zero-latency TokenFactor below any nodes that may increase the
407     // schedule height. Otherwise, ancestors of the TokenFactor may appear to
408     // have false stalls.
409     if (NI->getOpcode() == ISD::TokenFactor)
410       NodeSUnit->isScheduleLow = true;
411 
412     // If there are glue operands involved, N is now the bottom-most node
413     // of the sequence of nodes that are glued together.
414     // Update the SUnit.
415     NodeSUnit->setNode(N);
416     assert(N->getNodeId() == -1 && "Node already inserted!");
417     N->setNodeId(NodeSUnit->NodeNum);
418 
419     // Compute NumRegDefsLeft. This must be done before AddSchedEdges.
420     InitNumRegDefsLeft(NodeSUnit);
421 
422     // Assign the Latency field of NodeSUnit using target-provided information.
423     computeLatency(NodeSUnit);
424   }
425 
426   // Find all call operands.
427   while (!CallSUnits.empty()) {
428     SUnit *SU = CallSUnits.pop_back_val();
429     for (const SDNode *SUNode = SU->getNode(); SUNode;
430          SUNode = SUNode->getGluedNode()) {
431       if (SUNode->getOpcode() != ISD::CopyToReg)
432         continue;
433       SDNode *SrcN = SUNode->getOperand(2).getNode();
434       if (isPassiveNode(SrcN)) continue;   // Not scheduled.
435       SUnit *SrcSU = &SUnits[SrcN->getNodeId()];
436       SrcSU->isCallOp = true;
437     }
438   }
439 }
440 
441 void ScheduleDAGSDNodes::AddSchedEdges() {
442   const TargetSubtargetInfo &ST = MF.getSubtarget();
443 
444   // Check to see if the scheduler cares about latencies.
445   bool UnitLatencies = forceUnitLatencies();
446 
447   // Pass 2: add the preds, succs, etc.
448   for (SUnit &SU : SUnits) {
449     SDNode *MainNode = SU.getNode();
450 
451     if (MainNode->isMachineOpcode()) {
452       unsigned Opc = MainNode->getMachineOpcode();
453       const MCInstrDesc &MCID = TII->get(Opc);
454       for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
455         if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
456           SU.isTwoAddress = true;
457           break;
458         }
459       }
460       if (MCID.isCommutable())
461         SU.isCommutable = true;
462     }
463 
464     // Find all predecessors and successors of the group.
465     for (SDNode *N = SU.getNode(); N; N = N->getGluedNode()) {
466       if (N->isMachineOpcode() &&
467           !TII->get(N->getMachineOpcode()).implicit_defs().empty()) {
468         SU.hasPhysRegClobbers = true;
469         unsigned NumUsed = InstrEmitter::CountResults(N);
470         while (NumUsed != 0 && !N->hasAnyUseOfValue(NumUsed - 1))
471           --NumUsed;    // Skip over unused values at the end.
472         if (NumUsed > TII->get(N->getMachineOpcode()).getNumDefs())
473           SU.hasPhysRegDefs = true;
474       }
475 
476       for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
477         SDNode *OpN = N->getOperand(i).getNode();
478         unsigned DefIdx = N->getOperand(i).getResNo();
479         if (isPassiveNode(OpN)) continue;   // Not scheduled.
480         SUnit *OpSU = &SUnits[OpN->getNodeId()];
481         assert(OpSU && "Node has no SUnit!");
482         if (OpSU == &SU)
483           continue; // In the same group.
484 
485         EVT OpVT = N->getOperand(i).getValueType();
486         assert(OpVT != MVT::Glue && "Glued nodes should be in same sunit!");
487         bool isChain = OpVT == MVT::Other;
488 
489         unsigned PhysReg = 0;
490         int Cost = 1;
491         // Determine if this is a physical register dependency.
492         const TargetLowering &TLI = DAG->getTargetLoweringInfo();
493         CheckForPhysRegDependency(OpN, N, i, TRI, TII, TLI, PhysReg, Cost);
494         assert((PhysReg == 0 || !isChain) &&
495                "Chain dependence via physreg data?");
496         // FIXME: See ScheduleDAGSDNodes::EmitCopyFromReg. For now, scheduler
497         // emits a copy from the physical register to a virtual register unless
498         // it requires a cross class copy (cost < 0). That means we are only
499         // treating "expensive to copy" register dependency as physical register
500         // dependency. This may change in the future though.
501         if (Cost >= 0 && !StressSched)
502           PhysReg = 0;
503 
504         // If this is a ctrl dep, latency is 1.
505         unsigned OpLatency = isChain ? 1 : OpSU->Latency;
506         // Special-case TokenFactor chains as zero-latency.
507         if(isChain && OpN->getOpcode() == ISD::TokenFactor)
508           OpLatency = 0;
509 
510         SDep Dep = isChain ? SDep(OpSU, SDep::Barrier)
511           : SDep(OpSU, SDep::Data, PhysReg);
512         Dep.setLatency(OpLatency);
513         if (!isChain && !UnitLatencies) {
514           computeOperandLatency(OpN, N, i, Dep);
515           ST.adjustSchedDependency(OpSU, DefIdx, &SU, i, Dep);
516         }
517 
518         if (!SU.addPred(Dep) && !Dep.isCtrl() && OpSU->NumRegDefsLeft > 1) {
519           // Multiple register uses are combined in the same SUnit. For example,
520           // we could have a set of glued nodes with all their defs consumed by
521           // another set of glued nodes. Register pressure tracking sees this as
522           // a single use, so to keep pressure balanced we reduce the defs.
523           //
524           // We can't tell (without more book-keeping) if this results from
525           // glued nodes or duplicate operands. As long as we don't reduce
526           // NumRegDefsLeft to zero, we handle the common cases well.
527           --OpSU->NumRegDefsLeft;
528         }
529       }
530     }
531   }
532 }
533 
534 /// BuildSchedGraph - Build the SUnit graph from the selection dag that we
535 /// are input.  This SUnit graph is similar to the SelectionDAG, but
536 /// excludes nodes that aren't interesting to scheduling, and represents
537 /// glued together nodes with a single SUnit.
538 void ScheduleDAGSDNodes::BuildSchedGraph(AAResults *AA) {
539   // Cluster certain nodes which should be scheduled together.
540   ClusterNodes();
541   // Populate the SUnits array.
542   BuildSchedUnits();
543   // Compute all the scheduling dependencies between nodes.
544   AddSchedEdges();
545 }
546 
547 // Initialize NumNodeDefs for the current Node's opcode.
548 void ScheduleDAGSDNodes::RegDefIter::InitNodeNumDefs() {
549   // Check for phys reg copy.
550   if (!Node)
551     return;
552 
553   if (!Node->isMachineOpcode()) {
554     if (Node->getOpcode() == ISD::CopyFromReg)
555       NodeNumDefs = 1;
556     else
557       NodeNumDefs = 0;
558     return;
559   }
560   unsigned POpc = Node->getMachineOpcode();
561   if (POpc == TargetOpcode::IMPLICIT_DEF) {
562     // No register need be allocated for this.
563     NodeNumDefs = 0;
564     return;
565   }
566   if (POpc == TargetOpcode::PATCHPOINT &&
567       Node->getValueType(0) == MVT::Other) {
568     // PATCHPOINT is defined to have one result, but it might really have none
569     // if we're not using CallingConv::AnyReg. Don't mistake the chain for a
570     // real definition.
571     NodeNumDefs = 0;
572     return;
573   }
574   unsigned NRegDefs = SchedDAG->TII->get(Node->getMachineOpcode()).getNumDefs();
575   // Some instructions define regs that are not represented in the selection DAG
576   // (e.g. unused flags). See tMOVi8. Make sure we don't access past NumValues.
577   NodeNumDefs = std::min(Node->getNumValues(), NRegDefs);
578   DefIdx = 0;
579 }
580 
581 // Construct a RegDefIter for this SUnit and find the first valid value.
582 ScheduleDAGSDNodes::RegDefIter::RegDefIter(const SUnit *SU,
583                                            const ScheduleDAGSDNodes *SD)
584     : SchedDAG(SD), Node(SU->getNode()) {
585   InitNodeNumDefs();
586   Advance();
587 }
588 
589 // Advance to the next valid value defined by the SUnit.
590 void ScheduleDAGSDNodes::RegDefIter::Advance() {
591   for (;Node;) { // Visit all glued nodes.
592     for (;DefIdx < NodeNumDefs; ++DefIdx) {
593       if (!Node->hasAnyUseOfValue(DefIdx))
594         continue;
595       ValueType = Node->getSimpleValueType(DefIdx);
596       ++DefIdx;
597       return; // Found a normal regdef.
598     }
599     Node = Node->getGluedNode();
600     if (!Node) {
601       return; // No values left to visit.
602     }
603     InitNodeNumDefs();
604   }
605 }
606 
607 void ScheduleDAGSDNodes::InitNumRegDefsLeft(SUnit *SU) {
608   assert(SU->NumRegDefsLeft == 0 && "expect a new node");
609   for (RegDefIter I(SU, this); I.IsValid(); I.Advance()) {
610     assert(SU->NumRegDefsLeft < USHRT_MAX && "overflow is ok but unexpected");
611     ++SU->NumRegDefsLeft;
612   }
613 }
614 
615 void ScheduleDAGSDNodes::computeLatency(SUnit *SU) {
616   SDNode *N = SU->getNode();
617 
618   // TokenFactor operands are considered zero latency, and some schedulers
619   // (e.g. Top-Down list) may rely on the fact that operand latency is nonzero
620   // whenever node latency is nonzero.
621   if (N && N->getOpcode() == ISD::TokenFactor) {
622     SU->Latency = 0;
623     return;
624   }
625 
626   // Check to see if the scheduler cares about latencies.
627   if (forceUnitLatencies()) {
628     SU->Latency = 1;
629     return;
630   }
631 
632   if (!InstrItins || InstrItins->isEmpty()) {
633     if (N && N->isMachineOpcode() &&
634         TII->isHighLatencyDef(N->getMachineOpcode()))
635       SU->Latency = HighLatencyCycles;
636     else
637       SU->Latency = 1;
638     return;
639   }
640 
641   // Compute the latency for the node.  We use the sum of the latencies for
642   // all nodes glued together into this SUnit.
643   SU->Latency = 0;
644   for (SDNode *N = SU->getNode(); N; N = N->getGluedNode())
645     if (N->isMachineOpcode())
646       SU->Latency += TII->getInstrLatency(InstrItins, N);
647 }
648 
649 void ScheduleDAGSDNodes::computeOperandLatency(SDNode *Def, SDNode *Use,
650                                                unsigned OpIdx, SDep& dep) const{
651   // Check to see if the scheduler cares about latencies.
652   if (forceUnitLatencies())
653     return;
654 
655   if (dep.getKind() != SDep::Data)
656     return;
657 
658   unsigned DefIdx = Use->getOperand(OpIdx).getResNo();
659   if (Use->isMachineOpcode())
660     // Adjust the use operand index by num of defs.
661     OpIdx += TII->get(Use->getMachineOpcode()).getNumDefs();
662   int Latency = TII->getOperandLatency(InstrItins, Def, DefIdx, Use, OpIdx);
663   if (Latency > 1 && Use->getOpcode() == ISD::CopyToReg &&
664       !BB->succ_empty()) {
665     unsigned Reg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
666     if (Register::isVirtualRegister(Reg))
667       // This copy is a liveout value. It is likely coalesced, so reduce the
668       // latency so not to penalize the def.
669       // FIXME: need target specific adjustment here?
670       Latency = (Latency > 1) ? Latency - 1 : 1;
671   }
672   if (Latency >= 0)
673     dep.setLatency(Latency);
674 }
675 
676 void ScheduleDAGSDNodes::dumpNode(const SUnit &SU) const {
677 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
678   dumpNodeName(SU);
679   dbgs() << ": ";
680 
681   if (!SU.getNode()) {
682     dbgs() << "PHYS REG COPY\n";
683     return;
684   }
685 
686   SU.getNode()->dump(DAG);
687   dbgs() << "\n";
688   SmallVector<SDNode *, 4> GluedNodes;
689   for (SDNode *N = SU.getNode()->getGluedNode(); N; N = N->getGluedNode())
690     GluedNodes.push_back(N);
691   while (!GluedNodes.empty()) {
692     dbgs() << "    ";
693     GluedNodes.back()->dump(DAG);
694     dbgs() << "\n";
695     GluedNodes.pop_back();
696   }
697 #endif
698 }
699 
700 void ScheduleDAGSDNodes::dump() const {
701 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
702   if (EntrySU.getNode() != nullptr)
703     dumpNodeAll(EntrySU);
704   for (const SUnit &SU : SUnits)
705     dumpNodeAll(SU);
706   if (ExitSU.getNode() != nullptr)
707     dumpNodeAll(ExitSU);
708 #endif
709 }
710 
711 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
712 void ScheduleDAGSDNodes::dumpSchedule() const {
713   for (const SUnit *SU : Sequence) {
714     if (SU)
715       dumpNode(*SU);
716     else
717       dbgs() << "**** NOOP ****\n";
718   }
719 }
720 #endif
721 
722 #ifndef NDEBUG
723 /// VerifyScheduledSequence - Verify that all SUnits were scheduled and that
724 /// their state is consistent with the nodes listed in Sequence.
725 ///
726 void ScheduleDAGSDNodes::VerifyScheduledSequence(bool isBottomUp) {
727   unsigned ScheduledNodes = ScheduleDAG::VerifyScheduledDAG(isBottomUp);
728   unsigned Noops = llvm::count(Sequence, nullptr);
729   assert(Sequence.size() - Noops == ScheduledNodes &&
730          "The number of nodes scheduled doesn't match the expected number!");
731 }
732 #endif // NDEBUG
733 
734 /// ProcessSDDbgValues - Process SDDbgValues associated with this node.
735 static void
736 ProcessSDDbgValues(SDNode *N, SelectionDAG *DAG, InstrEmitter &Emitter,
737                    SmallVectorImpl<std::pair<unsigned, MachineInstr*> > &Orders,
738                    DenseMap<SDValue, Register> &VRBaseMap, unsigned Order) {
739   if (!N->getHasDebugValue())
740     return;
741 
742   /// Returns true if \p DV has any VReg operand locations which don't exist in
743   /// VRBaseMap.
744   auto HasUnknownVReg = [&VRBaseMap](SDDbgValue *DV) {
745     for (const SDDbgOperand &L : DV->getLocationOps()) {
746       if (L.getKind() == SDDbgOperand::SDNODE &&
747           VRBaseMap.count({L.getSDNode(), L.getResNo()}) == 0)
748         return true;
749     }
750     return false;
751   };
752 
753   // Opportunistically insert immediate dbg_value uses, i.e. those with the same
754   // source order number as N.
755   MachineBasicBlock *BB = Emitter.getBlock();
756   MachineBasicBlock::iterator InsertPos = Emitter.getInsertPos();
757   for (auto *DV : DAG->GetDbgValues(N)) {
758     if (DV->isEmitted())
759       continue;
760     unsigned DVOrder = DV->getOrder();
761     if (Order != 0 && DVOrder != Order)
762       continue;
763     // If DV has any VReg location operands which haven't been mapped then
764     // either that node is no longer available or we just haven't visited the
765     // node yet. In the former case we should emit an undef dbg_value, but we
766     // can do it later. And for the latter we'll want to wait until all
767     // dependent nodes have been visited.
768     if (!DV->isInvalidated() && HasUnknownVReg(DV))
769       continue;
770     MachineInstr *DbgMI = Emitter.EmitDbgValue(DV, VRBaseMap);
771     if (!DbgMI)
772       continue;
773     Orders.push_back({DVOrder, DbgMI});
774     BB->insert(InsertPos, DbgMI);
775   }
776 }
777 
778 // ProcessSourceNode - Process nodes with source order numbers. These are added
779 // to a vector which EmitSchedule uses to determine how to insert dbg_value
780 // instructions in the right order.
781 static void
782 ProcessSourceNode(SDNode *N, SelectionDAG *DAG, InstrEmitter &Emitter,
783                   DenseMap<SDValue, Register> &VRBaseMap,
784                   SmallVectorImpl<std::pair<unsigned, MachineInstr *>> &Orders,
785                   SmallSet<Register, 8> &Seen, MachineInstr *NewInsn) {
786   unsigned Order = N->getIROrder();
787   if (!Order || Seen.count(Order)) {
788     // Process any valid SDDbgValues even if node does not have any order
789     // assigned.
790     ProcessSDDbgValues(N, DAG, Emitter, Orders, VRBaseMap, 0);
791     return;
792   }
793 
794   // If a new instruction was generated for this Order number, record it.
795   // Otherwise, leave this order number unseen: we will either find later
796   // instructions for it, or leave it unseen if there were no instructions at
797   // all.
798   if (NewInsn) {
799     Seen.insert(Order);
800     Orders.push_back({Order, NewInsn});
801   }
802 
803   // Even if no instruction was generated, a Value may have become defined via
804   // earlier nodes. Try to process them now.
805   ProcessSDDbgValues(N, DAG, Emitter, Orders, VRBaseMap, Order);
806 }
807 
808 void ScheduleDAGSDNodes::
809 EmitPhysRegCopy(SUnit *SU, DenseMap<SUnit*, Register> &VRBaseMap,
810                 MachineBasicBlock::iterator InsertPos) {
811   for (const SDep &Pred : SU->Preds) {
812     if (Pred.isCtrl())
813       continue; // ignore chain preds
814     if (Pred.getSUnit()->CopyDstRC) {
815       // Copy to physical register.
816       DenseMap<SUnit *, Register>::iterator VRI =
817           VRBaseMap.find(Pred.getSUnit());
818       assert(VRI != VRBaseMap.end() && "Node emitted out of order - late");
819       // Find the destination physical register.
820       Register Reg;
821       for (const SDep &Succ : SU->Succs) {
822         if (Succ.isCtrl())
823           continue; // ignore chain preds
824         if (Succ.getReg()) {
825           Reg = Succ.getReg();
826           break;
827         }
828       }
829       BuildMI(*BB, InsertPos, DebugLoc(), TII->get(TargetOpcode::COPY), Reg)
830         .addReg(VRI->second);
831     } else {
832       // Copy from physical register.
833       assert(Pred.getReg() && "Unknown physical register!");
834       Register VRBase = MRI.createVirtualRegister(SU->CopyDstRC);
835       bool isNew = VRBaseMap.insert(std::make_pair(SU, VRBase)).second;
836       (void)isNew; // Silence compiler warning.
837       assert(isNew && "Node emitted out of order - early");
838       BuildMI(*BB, InsertPos, DebugLoc(), TII->get(TargetOpcode::COPY), VRBase)
839           .addReg(Pred.getReg());
840     }
841     break;
842   }
843 }
844 
845 /// EmitSchedule - Emit the machine code in scheduled order. Return the new
846 /// InsertPos and MachineBasicBlock that contains this insertion
847 /// point. ScheduleDAGSDNodes holds a BB pointer for convenience, but this does
848 /// not necessarily refer to returned BB. The emitter may split blocks.
849 MachineBasicBlock *ScheduleDAGSDNodes::
850 EmitSchedule(MachineBasicBlock::iterator &InsertPos) {
851   InstrEmitter Emitter(DAG->getTarget(), BB, InsertPos);
852   DenseMap<SDValue, Register> VRBaseMap;
853   DenseMap<SUnit*, Register> CopyVRBaseMap;
854   SmallVector<std::pair<unsigned, MachineInstr*>, 32> Orders;
855   SmallSet<Register, 8> Seen;
856   bool HasDbg = DAG->hasDebugValues();
857 
858   // Emit a node, and determine where its first instruction is for debuginfo.
859   // Zero, one, or multiple instructions can be created when emitting a node.
860   auto EmitNode =
861       [&](SDNode *Node, bool IsClone, bool IsCloned,
862           DenseMap<SDValue, Register> &VRBaseMap) -> MachineInstr * {
863     // Fetch instruction prior to this, or end() if nonexistant.
864     auto GetPrevInsn = [&](MachineBasicBlock::iterator I) {
865       if (I == BB->begin())
866         return BB->end();
867       else
868         return std::prev(Emitter.getInsertPos());
869     };
870 
871     MachineBasicBlock::iterator Before = GetPrevInsn(Emitter.getInsertPos());
872     Emitter.EmitNode(Node, IsClone, IsCloned, VRBaseMap);
873     MachineBasicBlock::iterator After = GetPrevInsn(Emitter.getInsertPos());
874 
875     // If the iterator did not change, no instructions were inserted.
876     if (Before == After)
877       return nullptr;
878 
879     MachineInstr *MI;
880     if (Before == BB->end()) {
881       // There were no prior instructions; the new ones must start at the
882       // beginning of the block.
883       MI = &Emitter.getBlock()->instr_front();
884     } else {
885       // Return first instruction after the pre-existing instructions.
886       MI = &*std::next(Before);
887     }
888 
889     if (MI->isCandidateForCallSiteEntry() &&
890         DAG->getTarget().Options.EmitCallSiteInfo)
891       MF.addCallArgsForwardingRegs(MI, DAG->getCallSiteInfo(Node));
892 
893     if (DAG->getNoMergeSiteInfo(Node)) {
894       MI->setFlag(MachineInstr::MIFlag::NoMerge);
895     }
896 
897     if (MDNode *MD = DAG->getPCSections(Node))
898       MI->setPCSections(MF, MD);
899 
900     return MI;
901   };
902 
903   // If this is the first BB, emit byval parameter dbg_value's.
904   if (HasDbg && BB->getParent()->begin() == MachineFunction::iterator(BB)) {
905     SDDbgInfo::DbgIterator PDI = DAG->ByvalParmDbgBegin();
906     SDDbgInfo::DbgIterator PDE = DAG->ByvalParmDbgEnd();
907     for (; PDI != PDE; ++PDI) {
908       MachineInstr *DbgMI= Emitter.EmitDbgValue(*PDI, VRBaseMap);
909       if (DbgMI) {
910         BB->insert(InsertPos, DbgMI);
911         // We re-emit the dbg_value closer to its use, too, after instructions
912         // are emitted to the BB.
913         (*PDI)->clearIsEmitted();
914       }
915     }
916   }
917 
918   for (SUnit *SU : Sequence) {
919     if (!SU) {
920       // Null SUnit* is a noop.
921       TII->insertNoop(*Emitter.getBlock(), InsertPos);
922       continue;
923     }
924 
925     // For pre-regalloc scheduling, create instructions corresponding to the
926     // SDNode and any glued SDNodes and append them to the block.
927     if (!SU->getNode()) {
928       // Emit a copy.
929       EmitPhysRegCopy(SU, CopyVRBaseMap, InsertPos);
930       continue;
931     }
932 
933     SmallVector<SDNode *, 4> GluedNodes;
934     for (SDNode *N = SU->getNode()->getGluedNode(); N; N = N->getGluedNode())
935       GluedNodes.push_back(N);
936     while (!GluedNodes.empty()) {
937       SDNode *N = GluedNodes.back();
938       auto NewInsn = EmitNode(N, SU->OrigNode != SU, SU->isCloned, VRBaseMap);
939       // Remember the source order of the inserted instruction.
940       if (HasDbg)
941         ProcessSourceNode(N, DAG, Emitter, VRBaseMap, Orders, Seen, NewInsn);
942 
943       if (MDNode *MD = DAG->getHeapAllocSite(N))
944         if (NewInsn && NewInsn->isCall())
945           NewInsn->setHeapAllocMarker(MF, MD);
946 
947       GluedNodes.pop_back();
948     }
949     auto NewInsn =
950         EmitNode(SU->getNode(), SU->OrigNode != SU, SU->isCloned, VRBaseMap);
951     // Remember the source order of the inserted instruction.
952     if (HasDbg)
953       ProcessSourceNode(SU->getNode(), DAG, Emitter, VRBaseMap, Orders, Seen,
954                         NewInsn);
955 
956     if (MDNode *MD = DAG->getHeapAllocSite(SU->getNode())) {
957       if (NewInsn && NewInsn->isCall())
958         NewInsn->setHeapAllocMarker(MF, MD);
959     }
960   }
961 
962   // Insert all the dbg_values which have not already been inserted in source
963   // order sequence.
964   if (HasDbg) {
965     MachineBasicBlock::iterator BBBegin = BB->getFirstNonPHI();
966 
967     // Sort the source order instructions and use the order to insert debug
968     // values. Use stable_sort so that DBG_VALUEs are inserted in the same order
969     // regardless of the host's implementation fo std::sort.
970     llvm::stable_sort(Orders, less_first());
971     std::stable_sort(DAG->DbgBegin(), DAG->DbgEnd(),
972                      [](const SDDbgValue *LHS, const SDDbgValue *RHS) {
973                        return LHS->getOrder() < RHS->getOrder();
974                      });
975 
976     SDDbgInfo::DbgIterator DI = DAG->DbgBegin();
977     SDDbgInfo::DbgIterator DE = DAG->DbgEnd();
978     // Now emit the rest according to source order.
979     unsigned LastOrder = 0;
980     for (unsigned i = 0, e = Orders.size(); i != e && DI != DE; ++i) {
981       unsigned Order = Orders[i].first;
982       MachineInstr *MI = Orders[i].second;
983       // Insert all SDDbgValue's whose order(s) are before "Order".
984       assert(MI);
985       for (; DI != DE; ++DI) {
986         if ((*DI)->getOrder() < LastOrder || (*DI)->getOrder() >= Order)
987           break;
988         if ((*DI)->isEmitted())
989           continue;
990 
991         MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap);
992         if (DbgMI) {
993           if (!LastOrder)
994             // Insert to start of the BB (after PHIs).
995             BB->insert(BBBegin, DbgMI);
996           else {
997             // Insert at the instruction, which may be in a different
998             // block, if the block was split by a custom inserter.
999             MachineBasicBlock::iterator Pos = MI;
1000             MI->getParent()->insert(Pos, DbgMI);
1001           }
1002         }
1003       }
1004       LastOrder = Order;
1005     }
1006     // Add trailing DbgValue's before the terminator. FIXME: May want to add
1007     // some of them before one or more conditional branches?
1008     SmallVector<MachineInstr*, 8> DbgMIs;
1009     for (; DI != DE; ++DI) {
1010       if ((*DI)->isEmitted())
1011         continue;
1012       assert((*DI)->getOrder() >= LastOrder &&
1013              "emitting DBG_VALUE out of order");
1014       if (MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap))
1015         DbgMIs.push_back(DbgMI);
1016     }
1017 
1018     MachineBasicBlock *InsertBB = Emitter.getBlock();
1019     MachineBasicBlock::iterator Pos = InsertBB->getFirstTerminator();
1020     InsertBB->insert(Pos, DbgMIs.begin(), DbgMIs.end());
1021 
1022     SDDbgInfo::DbgLabelIterator DLI = DAG->DbgLabelBegin();
1023     SDDbgInfo::DbgLabelIterator DLE = DAG->DbgLabelEnd();
1024     // Now emit the rest according to source order.
1025     LastOrder = 0;
1026     for (const auto &InstrOrder : Orders) {
1027       unsigned Order = InstrOrder.first;
1028       MachineInstr *MI = InstrOrder.second;
1029       if (!MI)
1030         continue;
1031 
1032       // Insert all SDDbgLabel's whose order(s) are before "Order".
1033       for (; DLI != DLE &&
1034              (*DLI)->getOrder() >= LastOrder && (*DLI)->getOrder() < Order;
1035              ++DLI) {
1036         MachineInstr *DbgMI = Emitter.EmitDbgLabel(*DLI);
1037         if (DbgMI) {
1038           if (!LastOrder)
1039             // Insert to start of the BB (after PHIs).
1040             BB->insert(BBBegin, DbgMI);
1041           else {
1042             // Insert at the instruction, which may be in a different
1043             // block, if the block was split by a custom inserter.
1044             MachineBasicBlock::iterator Pos = MI;
1045             MI->getParent()->insert(Pos, DbgMI);
1046           }
1047         }
1048       }
1049       if (DLI == DLE)
1050         break;
1051 
1052       LastOrder = Order;
1053     }
1054   }
1055 
1056   InsertPos = Emitter.getInsertPos();
1057   // In some cases, DBG_VALUEs might be inserted after the first terminator,
1058   // which results in an invalid MBB. If that happens, move the DBG_VALUEs
1059   // before the first terminator.
1060   MachineBasicBlock *InsertBB = Emitter.getBlock();
1061   auto FirstTerm = InsertBB->getFirstTerminator();
1062   if (FirstTerm != InsertBB->end()) {
1063     assert(!FirstTerm->isDebugValue() &&
1064            "first terminator cannot be a debug value");
1065     for (MachineInstr &MI : make_early_inc_range(
1066              make_range(std::next(FirstTerm), InsertBB->end()))) {
1067       // Only scan up to insertion point.
1068       if (&MI == InsertPos)
1069         break;
1070 
1071       if (!MI.isDebugValue())
1072         continue;
1073 
1074       // The DBG_VALUE was referencing a value produced by a terminator. By
1075       // moving the DBG_VALUE, the referenced value also needs invalidating.
1076       MI.getOperand(0).ChangeToRegister(0, false);
1077       MI.moveBefore(&*FirstTerm);
1078     }
1079   }
1080   return InsertBB;
1081 }
1082 
1083 /// Return the basic block label.
1084 std::string ScheduleDAGSDNodes::getDAGName() const {
1085   return "sunit-dag." + BB->getFullName();
1086 }
1087