1 //===--- ScheduleDAGSDNodes.cpp - Implement the ScheduleDAGSDNodes class --===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This implements the ScheduleDAG class, which is a base class used by
11 // scheduling implementation classes.
12 //
13 //===----------------------------------------------------------------------===//
14 
15 #include "ScheduleDAGSDNodes.h"
16 #include "InstrEmitter.h"
17 #include "SDNodeDbgValue.h"
18 #include "llvm/ADT/DenseMap.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/Statistic.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/SelectionDAG.h"
26 #include "llvm/CodeGen/TargetInstrInfo.h"
27 #include "llvm/CodeGen/TargetLowering.h"
28 #include "llvm/CodeGen/TargetRegisterInfo.h"
29 #include "llvm/CodeGen/TargetSubtargetInfo.h"
30 #include "llvm/Config/llvm-config.h"
31 #include "llvm/MC/MCInstrItineraries.h"
32 #include "llvm/Support/CommandLine.h"
33 #include "llvm/Support/Debug.h"
34 #include "llvm/Support/raw_ostream.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 
ScheduleDAGSDNodes(MachineFunction & mf)49 ScheduleDAGSDNodes::ScheduleDAGSDNodes(MachineFunction &mf)
50     : ScheduleDAG(mf), BB(nullptr), DAG(nullptr),
51       InstrItins(mf.getSubtarget().getInstrItineraryData()) {}
52 
53 /// Run - perform scheduling.
54 ///
Run(SelectionDAG * dag,MachineBasicBlock * bb)55 void ScheduleDAGSDNodes::Run(SelectionDAG *dag, MachineBasicBlock *bb) {
56   BB = bb;
57   DAG = dag;
58 
59   // Clear the scheduler's SUnit DAG.
60   ScheduleDAG::clearDAG();
61   Sequence.clear();
62 
63   // Invoke the target's selection of scheduler.
64   Schedule();
65 }
66 
67 /// NewSUnit - Creates a new SUnit and return a ptr to it.
68 ///
newSUnit(SDNode * N)69 SUnit *ScheduleDAGSDNodes::newSUnit(SDNode *N) {
70 #ifndef NDEBUG
71   const SUnit *Addr = nullptr;
72   if (!SUnits.empty())
73     Addr = &SUnits[0];
74 #endif
75   SUnits.emplace_back(N, (unsigned)SUnits.size());
76   assert((Addr == nullptr || Addr == &SUnits[0]) &&
77          "SUnits std::vector reallocated on the fly!");
78   SUnits.back().OrigNode = &SUnits.back();
79   SUnit *SU = &SUnits.back();
80   const TargetLowering &TLI = DAG->getTargetLoweringInfo();
81   if (!N ||
82       (N->isMachineOpcode() &&
83        N->getMachineOpcode() == TargetOpcode::IMPLICIT_DEF))
84     SU->SchedulingPref = Sched::None;
85   else
86     SU->SchedulingPref = TLI.getSchedulingPreference(N);
87   return SU;
88 }
89 
Clone(SUnit * Old)90 SUnit *ScheduleDAGSDNodes::Clone(SUnit *Old) {
91   SUnit *SU = newSUnit(Old->getNode());
92   SU->OrigNode = Old->OrigNode;
93   SU->Latency = Old->Latency;
94   SU->isVRegCycle = Old->isVRegCycle;
95   SU->isCall = Old->isCall;
96   SU->isCallOp = Old->isCallOp;
97   SU->isTwoAddress = Old->isTwoAddress;
98   SU->isCommutable = Old->isCommutable;
99   SU->hasPhysRegDefs = Old->hasPhysRegDefs;
100   SU->hasPhysRegClobbers = Old->hasPhysRegClobbers;
101   SU->isScheduleHigh = Old->isScheduleHigh;
102   SU->isScheduleLow = Old->isScheduleLow;
103   SU->SchedulingPref = Old->SchedulingPref;
104   Old->isCloned = true;
105   return SU;
106 }
107 
108 /// CheckForPhysRegDependency - Check if the dependency between def and use of
109 /// a specified operand is a physical register dependency. If so, returns the
110 /// register and the cost of copying the register.
CheckForPhysRegDependency(SDNode * Def,SDNode * User,unsigned Op,const TargetRegisterInfo * TRI,const TargetInstrInfo * TII,unsigned & PhysReg,int & Cost)111 static void CheckForPhysRegDependency(SDNode *Def, SDNode *User, unsigned Op,
112                                       const TargetRegisterInfo *TRI,
113                                       const TargetInstrInfo *TII,
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 (TargetRegisterInfo::isVirtualRegister(Reg))
120     return;
121 
122   unsigned ResNo = User->getOperand(2).getResNo();
123   if (Def->getOpcode() == ISD::CopyFromReg &&
124       cast<RegisterSDNode>(Def->getOperand(1))->getReg() == Reg) {
125     PhysReg = Reg;
126   } else if (Def->isMachineOpcode()) {
127     const MCInstrDesc &II = TII->get(Def->getMachineOpcode());
128     if (ResNo >= II.getNumDefs() &&
129         II.ImplicitDefs[ResNo - II.getNumDefs()] == Reg)
130       PhysReg = Reg;
131   }
132 
133   if (PhysReg != 0) {
134     const TargetRegisterClass *RC =
135         TRI->getMinimalPhysRegClass(Reg, Def->getSimpleValueType(ResNo));
136     Cost = RC->getCopyCost();
137   }
138 }
139 
140 // Helper for AddGlue to clone node operands.
CloneNodeWithValues(SDNode * N,SelectionDAG * DAG,ArrayRef<EVT> VTs,SDValue ExtraOper=SDValue ())141 static void CloneNodeWithValues(SDNode *N, SelectionDAG *DAG, ArrayRef<EVT> VTs,
142                                 SDValue ExtraOper = SDValue()) {
143   SmallVector<SDValue, 8> Ops(N->op_begin(), N->op_end());
144   if (ExtraOper.getNode())
145     Ops.push_back(ExtraOper);
146 
147   SDVTList VTList = DAG->getVTList(VTs);
148   MachineSDNode *MN = dyn_cast<MachineSDNode>(N);
149 
150   // Store memory references.
151   SmallVector<MachineMemOperand *, 2> MMOs;
152   if (MN)
153     MMOs.assign(MN->memoperands_begin(), MN->memoperands_end());
154 
155   DAG->MorphNodeTo(N, N->getOpcode(), VTList, Ops);
156 
157   // Reset the memory references
158   if (MN)
159     DAG->setNodeMemRefs(MN, MMOs);
160 }
161 
AddGlue(SDNode * N,SDValue Glue,bool AddGlue,SelectionDAG * DAG)162 static bool AddGlue(SDNode *N, SDValue Glue, bool AddGlue, SelectionDAG *DAG) {
163   SDNode *GlueDestNode = Glue.getNode();
164 
165   // Don't add glue from a node to itself.
166   if (GlueDestNode == N) return false;
167 
168   // Don't add a glue operand to something that already uses glue.
169   if (GlueDestNode &&
170       N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue) {
171     return false;
172   }
173   // Don't add glue to something that already has a glue value.
174   if (N->getValueType(N->getNumValues() - 1) == MVT::Glue) return false;
175 
176   SmallVector<EVT, 4> VTs(N->value_begin(), N->value_end());
177   if (AddGlue)
178     VTs.push_back(MVT::Glue);
179 
180   CloneNodeWithValues(N, DAG, VTs, Glue);
181 
182   return true;
183 }
184 
185 // Cleanup after unsuccessful AddGlue. Use the standard method of morphing the
186 // node even though simply shrinking the value list is sufficient.
RemoveUnusedGlue(SDNode * N,SelectionDAG * DAG)187 static void RemoveUnusedGlue(SDNode *N, SelectionDAG *DAG) {
188   assert((N->getValueType(N->getNumValues() - 1) == MVT::Glue &&
189           !N->hasAnyUseOfValue(N->getNumValues() - 1)) &&
190          "expected an unused glue value");
191 
192   CloneNodeWithValues(N, DAG,
193                       makeArrayRef(N->value_begin(), N->getNumValues() - 1));
194 }
195 
196 /// ClusterNeighboringLoads - Force nearby loads together by "gluing" them.
197 /// This function finds loads of the same base and different offsets. If the
198 /// offsets are not far apart (target specific), it add MVT::Glue inputs and
199 /// outputs to ensure they are scheduled together and in order. This
200 /// optimization may benefit some targets by improving cache locality.
ClusterNeighboringLoads(SDNode * Node)201 void ScheduleDAGSDNodes::ClusterNeighboringLoads(SDNode *Node) {
202   SDNode *Chain = nullptr;
203   unsigned NumOps = Node->getNumOperands();
204   if (Node->getOperand(NumOps-1).getValueType() == MVT::Other)
205     Chain = Node->getOperand(NumOps-1).getNode();
206   if (!Chain)
207     return;
208 
209   // Look for other loads of the same chain. Find loads that are loading from
210   // the same base pointer and different offsets.
211   SmallPtrSet<SDNode*, 16> Visited;
212   SmallVector<int64_t, 4> Offsets;
213   DenseMap<long long, SDNode*> O2SMap;  // Map from offset to SDNode.
214   bool Cluster = false;
215   SDNode *Base = Node;
216   // This algorithm requires a reasonably low use count before finding a match
217   // to avoid uselessly blowing up compile time in large blocks.
218   unsigned UseCount = 0;
219   for (SDNode::use_iterator I = Chain->use_begin(), E = Chain->use_end();
220        I != E && UseCount < 100; ++I, ++UseCount) {
221     SDNode *User = *I;
222     if (User == Node || !Visited.insert(User).second)
223       continue;
224     int64_t Offset1, Offset2;
225     if (!TII->areLoadsFromSameBasePtr(Base, User, Offset1, Offset2) ||
226         Offset1 == Offset2)
227       // FIXME: Should be ok if they addresses are identical. But earlier
228       // optimizations really should have eliminated one of the loads.
229       continue;
230     if (O2SMap.insert(std::make_pair(Offset1, Base)).second)
231       Offsets.push_back(Offset1);
232     O2SMap.insert(std::make_pair(Offset2, User));
233     Offsets.push_back(Offset2);
234     if (Offset2 < Offset1)
235       Base = User;
236     Cluster = true;
237     // Reset UseCount to allow more matches.
238     UseCount = 0;
239   }
240 
241   if (!Cluster)
242     return;
243 
244   // Sort them in increasing order.
245   llvm::sort(Offsets);
246 
247   // Check if the loads are close enough.
248   SmallVector<SDNode*, 4> Loads;
249   unsigned NumLoads = 0;
250   int64_t BaseOff = Offsets[0];
251   SDNode *BaseLoad = O2SMap[BaseOff];
252   Loads.push_back(BaseLoad);
253   for (unsigned i = 1, e = Offsets.size(); i != e; ++i) {
254     int64_t Offset = Offsets[i];
255     SDNode *Load = O2SMap[Offset];
256     if (!TII->shouldScheduleLoadsNear(BaseLoad, Load, BaseOff, Offset,NumLoads))
257       break; // Stop right here. Ignore loads that are further away.
258     Loads.push_back(Load);
259     ++NumLoads;
260   }
261 
262   if (NumLoads == 0)
263     return;
264 
265   // Cluster loads by adding MVT::Glue outputs and inputs. This also
266   // ensure they are scheduled in order of increasing addresses.
267   SDNode *Lead = Loads[0];
268   SDValue InGlue = SDValue(nullptr, 0);
269   if (AddGlue(Lead, InGlue, true, DAG))
270     InGlue = SDValue(Lead, Lead->getNumValues() - 1);
271   for (unsigned I = 1, E = Loads.size(); I != E; ++I) {
272     bool OutGlue = I < E - 1;
273     SDNode *Load = Loads[I];
274 
275     // If AddGlue fails, we could leave an unsused glue value. This should not
276     // cause any
277     if (AddGlue(Load, InGlue, OutGlue, DAG)) {
278       if (OutGlue)
279         InGlue = SDValue(Load, Load->getNumValues() - 1);
280 
281       ++LoadsClustered;
282     }
283     else if (!OutGlue && InGlue.getNode())
284       RemoveUnusedGlue(InGlue.getNode(), DAG);
285   }
286 }
287 
288 /// ClusterNodes - Cluster certain nodes which should be scheduled together.
289 ///
ClusterNodes()290 void ScheduleDAGSDNodes::ClusterNodes() {
291   for (SDNode &NI : DAG->allnodes()) {
292     SDNode *Node = &NI;
293     if (!Node || !Node->isMachineOpcode())
294       continue;
295 
296     unsigned Opc = Node->getMachineOpcode();
297     const MCInstrDesc &MCID = TII->get(Opc);
298     if (MCID.mayLoad())
299       // Cluster loads from "near" addresses into combined SUnits.
300       ClusterNeighboringLoads(Node);
301   }
302 }
303 
BuildSchedUnits()304 void ScheduleDAGSDNodes::BuildSchedUnits() {
305   // During scheduling, the NodeId field of SDNode is used to map SDNodes
306   // to their associated SUnits by holding SUnits table indices. A value
307   // of -1 means the SDNode does not yet have an associated SUnit.
308   unsigned NumNodes = 0;
309   for (SDNode &NI : DAG->allnodes()) {
310     NI.setNodeId(-1);
311     ++NumNodes;
312   }
313 
314   // Reserve entries in the vector for each of the SUnits we are creating.  This
315   // ensure that reallocation of the vector won't happen, so SUnit*'s won't get
316   // invalidated.
317   // FIXME: Multiply by 2 because we may clone nodes during scheduling.
318   // This is a temporary workaround.
319   SUnits.reserve(NumNodes * 2);
320 
321   // Add all nodes in depth first order.
322   SmallVector<SDNode*, 64> Worklist;
323   SmallPtrSet<SDNode*, 32> Visited;
324   Worklist.push_back(DAG->getRoot().getNode());
325   Visited.insert(DAG->getRoot().getNode());
326 
327   SmallVector<SUnit*, 8> CallSUnits;
328   while (!Worklist.empty()) {
329     SDNode *NI = Worklist.pop_back_val();
330 
331     // Add all operands to the worklist unless they've already been added.
332     for (const SDValue &Op : NI->op_values())
333       if (Visited.insert(Op.getNode()).second)
334         Worklist.push_back(Op.getNode());
335 
336     if (isPassiveNode(NI))  // Leaf node, e.g. a TargetImmediate.
337       continue;
338 
339     // If this node has already been processed, stop now.
340     if (NI->getNodeId() != -1) continue;
341 
342     SUnit *NodeSUnit = newSUnit(NI);
343 
344     // See if anything is glued to this node, if so, add them to glued
345     // nodes.  Nodes can have at most one glue input and one glue output.  Glue
346     // is required to be the last operand and result of a node.
347 
348     // Scan up to find glued preds.
349     SDNode *N = NI;
350     while (N->getNumOperands() &&
351            N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue) {
352       N = N->getOperand(N->getNumOperands()-1).getNode();
353       assert(N->getNodeId() == -1 && "Node already inserted!");
354       N->setNodeId(NodeSUnit->NodeNum);
355       if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall())
356         NodeSUnit->isCall = true;
357     }
358 
359     // Scan down to find any glued succs.
360     N = NI;
361     while (N->getValueType(N->getNumValues()-1) == MVT::Glue) {
362       SDValue GlueVal(N, N->getNumValues()-1);
363 
364       // There are either zero or one users of the Glue result.
365       bool HasGlueUse = false;
366       for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
367            UI != E; ++UI)
368         if (GlueVal.isOperandOf(*UI)) {
369           HasGlueUse = true;
370           assert(N->getNodeId() == -1 && "Node already inserted!");
371           N->setNodeId(NodeSUnit->NodeNum);
372           N = *UI;
373           if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall())
374             NodeSUnit->isCall = true;
375           break;
376         }
377       if (!HasGlueUse) break;
378     }
379 
380     if (NodeSUnit->isCall)
381       CallSUnits.push_back(NodeSUnit);
382 
383     // Schedule zero-latency TokenFactor below any nodes that may increase the
384     // schedule height. Otherwise, ancestors of the TokenFactor may appear to
385     // have false stalls.
386     if (NI->getOpcode() == ISD::TokenFactor)
387       NodeSUnit->isScheduleLow = true;
388 
389     // If there are glue operands involved, N is now the bottom-most node
390     // of the sequence of nodes that are glued together.
391     // Update the SUnit.
392     NodeSUnit->setNode(N);
393     assert(N->getNodeId() == -1 && "Node already inserted!");
394     N->setNodeId(NodeSUnit->NodeNum);
395 
396     // Compute NumRegDefsLeft. This must be done before AddSchedEdges.
397     InitNumRegDefsLeft(NodeSUnit);
398 
399     // Assign the Latency field of NodeSUnit using target-provided information.
400     computeLatency(NodeSUnit);
401   }
402 
403   // Find all call operands.
404   while (!CallSUnits.empty()) {
405     SUnit *SU = CallSUnits.pop_back_val();
406     for (const SDNode *SUNode = SU->getNode(); SUNode;
407          SUNode = SUNode->getGluedNode()) {
408       if (SUNode->getOpcode() != ISD::CopyToReg)
409         continue;
410       SDNode *SrcN = SUNode->getOperand(2).getNode();
411       if (isPassiveNode(SrcN)) continue;   // Not scheduled.
412       SUnit *SrcSU = &SUnits[SrcN->getNodeId()];
413       SrcSU->isCallOp = true;
414     }
415   }
416 }
417 
AddSchedEdges()418 void ScheduleDAGSDNodes::AddSchedEdges() {
419   const TargetSubtargetInfo &ST = MF.getSubtarget();
420 
421   // Check to see if the scheduler cares about latencies.
422   bool UnitLatencies = forceUnitLatencies();
423 
424   // Pass 2: add the preds, succs, etc.
425   for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
426     SUnit *SU = &SUnits[su];
427     SDNode *MainNode = SU->getNode();
428 
429     if (MainNode->isMachineOpcode()) {
430       unsigned Opc = MainNode->getMachineOpcode();
431       const MCInstrDesc &MCID = TII->get(Opc);
432       for (unsigned i = 0; i != MCID.getNumOperands(); ++i) {
433         if (MCID.getOperandConstraint(i, MCOI::TIED_TO) != -1) {
434           SU->isTwoAddress = true;
435           break;
436         }
437       }
438       if (MCID.isCommutable())
439         SU->isCommutable = true;
440     }
441 
442     // Find all predecessors and successors of the group.
443     for (SDNode *N = SU->getNode(); N; N = N->getGluedNode()) {
444       if (N->isMachineOpcode() &&
445           TII->get(N->getMachineOpcode()).getImplicitDefs()) {
446         SU->hasPhysRegClobbers = true;
447         unsigned NumUsed = InstrEmitter::CountResults(N);
448         while (NumUsed != 0 && !N->hasAnyUseOfValue(NumUsed - 1))
449           --NumUsed;    // Skip over unused values at the end.
450         if (NumUsed > TII->get(N->getMachineOpcode()).getNumDefs())
451           SU->hasPhysRegDefs = true;
452       }
453 
454       for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
455         SDNode *OpN = N->getOperand(i).getNode();
456         if (isPassiveNode(OpN)) continue;   // Not scheduled.
457         SUnit *OpSU = &SUnits[OpN->getNodeId()];
458         assert(OpSU && "Node has no SUnit!");
459         if (OpSU == SU) continue;           // In the same group.
460 
461         EVT OpVT = N->getOperand(i).getValueType();
462         assert(OpVT != MVT::Glue && "Glued nodes should be in same sunit!");
463         bool isChain = OpVT == MVT::Other;
464 
465         unsigned PhysReg = 0;
466         int Cost = 1;
467         // Determine if this is a physical register dependency.
468         CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost);
469         assert((PhysReg == 0 || !isChain) &&
470                "Chain dependence via physreg data?");
471         // FIXME: See ScheduleDAGSDNodes::EmitCopyFromReg. For now, scheduler
472         // emits a copy from the physical register to a virtual register unless
473         // it requires a cross class copy (cost < 0). That means we are only
474         // treating "expensive to copy" register dependency as physical register
475         // dependency. This may change in the future though.
476         if (Cost >= 0 && !StressSched)
477           PhysReg = 0;
478 
479         // If this is a ctrl dep, latency is 1.
480         unsigned OpLatency = isChain ? 1 : OpSU->Latency;
481         // Special-case TokenFactor chains as zero-latency.
482         if(isChain && OpN->getOpcode() == ISD::TokenFactor)
483           OpLatency = 0;
484 
485         SDep Dep = isChain ? SDep(OpSU, SDep::Barrier)
486           : SDep(OpSU, SDep::Data, PhysReg);
487         Dep.setLatency(OpLatency);
488         if (!isChain && !UnitLatencies) {
489           computeOperandLatency(OpN, N, i, Dep);
490           ST.adjustSchedDependency(OpSU, SU, Dep);
491         }
492 
493         if (!SU->addPred(Dep) && !Dep.isCtrl() && OpSU->NumRegDefsLeft > 1) {
494           // Multiple register uses are combined in the same SUnit. For example,
495           // we could have a set of glued nodes with all their defs consumed by
496           // another set of glued nodes. Register pressure tracking sees this as
497           // a single use, so to keep pressure balanced we reduce the defs.
498           //
499           // We can't tell (without more book-keeping) if this results from
500           // glued nodes or duplicate operands. As long as we don't reduce
501           // NumRegDefsLeft to zero, we handle the common cases well.
502           --OpSU->NumRegDefsLeft;
503         }
504       }
505     }
506   }
507 }
508 
509 /// BuildSchedGraph - Build the SUnit graph from the selection dag that we
510 /// are input.  This SUnit graph is similar to the SelectionDAG, but
511 /// excludes nodes that aren't interesting to scheduling, and represents
512 /// glued together nodes with a single SUnit.
BuildSchedGraph(AliasAnalysis * AA)513 void ScheduleDAGSDNodes::BuildSchedGraph(AliasAnalysis *AA) {
514   // Cluster certain nodes which should be scheduled together.
515   ClusterNodes();
516   // Populate the SUnits array.
517   BuildSchedUnits();
518   // Compute all the scheduling dependencies between nodes.
519   AddSchedEdges();
520 }
521 
522 // Initialize NumNodeDefs for the current Node's opcode.
InitNodeNumDefs()523 void ScheduleDAGSDNodes::RegDefIter::InitNodeNumDefs() {
524   // Check for phys reg copy.
525   if (!Node)
526     return;
527 
528   if (!Node->isMachineOpcode()) {
529     if (Node->getOpcode() == ISD::CopyFromReg)
530       NodeNumDefs = 1;
531     else
532       NodeNumDefs = 0;
533     return;
534   }
535   unsigned POpc = Node->getMachineOpcode();
536   if (POpc == TargetOpcode::IMPLICIT_DEF) {
537     // No register need be allocated for this.
538     NodeNumDefs = 0;
539     return;
540   }
541   if (POpc == TargetOpcode::PATCHPOINT &&
542       Node->getValueType(0) == MVT::Other) {
543     // PATCHPOINT is defined to have one result, but it might really have none
544     // if we're not using CallingConv::AnyReg. Don't mistake the chain for a
545     // real definition.
546     NodeNumDefs = 0;
547     return;
548   }
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);
553   DefIdx = 0;
554 }
555 
556 // Construct a RegDefIter for this SUnit and find the first valid value.
RegDefIter(const SUnit * SU,const ScheduleDAGSDNodes * SD)557 ScheduleDAGSDNodes::RegDefIter::RegDefIter(const SUnit *SU,
558                                            const ScheduleDAGSDNodes *SD)
559   : SchedDAG(SD), Node(SU->getNode()), DefIdx(0), NodeNumDefs(0) {
560   InitNodeNumDefs();
561   Advance();
562 }
563 
564 // Advance to the next valid value defined by the SUnit.
Advance()565 void ScheduleDAGSDNodes::RegDefIter::Advance() {
566   for (;Node;) { // Visit all glued nodes.
567     for (;DefIdx < NodeNumDefs; ++DefIdx) {
568       if (!Node->hasAnyUseOfValue(DefIdx))
569         continue;
570       ValueType = Node->getSimpleValueType(DefIdx);
571       ++DefIdx;
572       return; // Found a normal regdef.
573     }
574     Node = Node->getGluedNode();
575     if (!Node) {
576       return; // No values left to visit.
577     }
578     InitNodeNumDefs();
579   }
580 }
581 
InitNumRegDefsLeft(SUnit * SU)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;
587   }
588 }
589 
computeLatency(SUnit * SU)590 void ScheduleDAGSDNodes::computeLatency(SUnit *SU) {
591   SDNode *N = SU->getNode();
592 
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) {
597     SU->Latency = 0;
598     return;
599   }
600 
601   // Check to see if the scheduler cares about latencies.
602   if (forceUnitLatencies()) {
603     SU->Latency = 1;
604     return;
605   }
606 
607   if (!InstrItins || InstrItins->isEmpty()) {
608     if (N && N->isMachineOpcode() &&
609         TII->isHighLatencyDef(N->getMachineOpcode()))
610       SU->Latency = HighLatencyCycles;
611     else
612       SU->Latency = 1;
613     return;
614   }
615 
616   // Compute the latency for the node.  We use the sum of the latencies for
617   // all nodes glued together into this SUnit.
618   SU->Latency = 0;
619   for (SDNode *N = SU->getNode(); N; N = N->getGluedNode())
620     if (N->isMachineOpcode())
621       SU->Latency += TII->getInstrLatency(InstrItins, N);
622 }
623 
computeOperandLatency(SDNode * Def,SDNode * Use,unsigned OpIdx,SDep & dep) const624 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())
628     return;
629 
630   if (dep.getKind() != SDep::Data)
631     return;
632 
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 &&
639       !BB->succ_empty()) {
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;
646   }
647   if (Latency >= 0)
648     dep.setLatency(Latency);
649 }
650 
dumpNode(const SUnit & SU) const651 void ScheduleDAGSDNodes::dumpNode(const SUnit &SU) const {
652 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
653   dumpNodeName(SU);
654   dbgs() << ": ";
655 
656   if (!SU.getNode()) {
657     dbgs() << "PHYS REG COPY\n";
658     return;
659   }
660 
661   SU.getNode()->dump(DAG);
662   dbgs() << "\n";
663   SmallVector<SDNode *, 4> GluedNodes;
664   for (SDNode *N = SU.getNode()->getGluedNode(); N; N = N->getGluedNode())
665     GluedNodes.push_back(N);
666   while (!GluedNodes.empty()) {
667     dbgs() << "    ";
668     GluedNodes.back()->dump(DAG);
669     dbgs() << "\n";
670     GluedNodes.pop_back();
671   }
672 #endif
673 }
674 
dump() const675 void ScheduleDAGSDNodes::dump() const {
676 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
677   if (EntrySU.getNode() != nullptr)
678     dumpNodeAll(EntrySU);
679   for (const SUnit &SU : SUnits)
680     dumpNodeAll(SU);
681   if (ExitSU.getNode() != nullptr)
682     dumpNodeAll(ExitSU);
683 #endif
684 }
685 
686 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dumpSchedule() const687 void ScheduleDAGSDNodes::dumpSchedule() const {
688   for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
689     if (SUnit *SU = Sequence[i])
690       dumpNode(*SU);
691     else
692       dbgs() << "**** NOOP ****\n";
693   }
694 }
695 #endif
696 
697 #ifndef NDEBUG
698 /// VerifyScheduledSequence - Verify that all SUnits were scheduled and that
699 /// their state is consistent with the nodes listed in Sequence.
700 ///
VerifyScheduledSequence(bool isBottomUp)701 void ScheduleDAGSDNodes::VerifyScheduledSequence(bool isBottomUp) {
702   unsigned ScheduledNodes = ScheduleDAG::VerifyScheduledDAG(isBottomUp);
703   unsigned Noops = 0;
704   for (unsigned i = 0, e = Sequence.size(); i != e; ++i)
705     if (!Sequence[i])
706       ++Noops;
707   assert(Sequence.size() - Noops == ScheduledNodes &&
708          "The number of nodes scheduled doesn't match the expected number!");
709 }
710 #endif // NDEBUG
711 
712 /// ProcessSDDbgValues - Process SDDbgValues associated with this node.
713 static void
ProcessSDDbgValues(SDNode * N,SelectionDAG * DAG,InstrEmitter & Emitter,SmallVectorImpl<std::pair<unsigned,MachineInstr * >> & Orders,DenseMap<SDValue,unsigned> & VRBaseMap,unsigned Order)714 ProcessSDDbgValues(SDNode *N, SelectionDAG *DAG, InstrEmitter &Emitter,
715                    SmallVectorImpl<std::pair<unsigned, MachineInstr*> > &Orders,
716                    DenseMap<SDValue, unsigned> &VRBaseMap, unsigned Order) {
717   if (!N->getHasDebugValue())
718     return;
719 
720   // Opportunistically insert immediate dbg_value uses, i.e. those with the same
721   // source order number as N.
722   MachineBasicBlock *BB = Emitter.getBlock();
723   MachineBasicBlock::iterator InsertPos = Emitter.getInsertPos();
724   for (auto DV : DAG->GetDbgValues(N)) {
725     if (DV->isEmitted())
726       continue;
727     unsigned DVOrder = DV->getOrder();
728     if (!Order || DVOrder == Order) {
729       MachineInstr *DbgMI = Emitter.EmitDbgValue(DV, VRBaseMap);
730       if (DbgMI) {
731         Orders.push_back({DVOrder, DbgMI});
732         BB->insert(InsertPos, DbgMI);
733       }
734     }
735   }
736 }
737 
738 // ProcessSourceNode - Process nodes with source order numbers. These are added
739 // to a vector which EmitSchedule uses to determine how to insert dbg_value
740 // instructions in the right order.
741 static void
ProcessSourceNode(SDNode * N,SelectionDAG * DAG,InstrEmitter & Emitter,DenseMap<SDValue,unsigned> & VRBaseMap,SmallVectorImpl<std::pair<unsigned,MachineInstr * >> & Orders,SmallSet<unsigned,8> & Seen)742 ProcessSourceNode(SDNode *N, SelectionDAG *DAG, InstrEmitter &Emitter,
743                   DenseMap<SDValue, unsigned> &VRBaseMap,
744                   SmallVectorImpl<std::pair<unsigned, MachineInstr*> > &Orders,
745                   SmallSet<unsigned, 8> &Seen) {
746   unsigned Order = N->getIROrder();
747   if (!Order || !Seen.insert(Order).second) {
748     // Process any valid SDDbgValues even if node does not have any order
749     // assigned.
750     ProcessSDDbgValues(N, DAG, Emitter, Orders, VRBaseMap, 0);
751     return;
752   }
753 
754   MachineBasicBlock *BB = Emitter.getBlock();
755   auto IP = Emitter.getInsertPos();
756   if (IP == BB->begin() || BB->back().isPHI() ||
757       // Fast-isel may have inserted some instructions, in which case the
758       // BB->back().isPHI() test will not fire when we want it to.
759       std::prev(IP)->isPHI()) {
760     // Did not insert any instruction.
761     Orders.push_back({Order, (MachineInstr *)nullptr});
762     return;
763   }
764 
765   Orders.push_back({Order, &*std::prev(IP)});
766   ProcessSDDbgValues(N, DAG, Emitter, Orders, VRBaseMap, Order);
767 }
768 
769 void ScheduleDAGSDNodes::
EmitPhysRegCopy(SUnit * SU,DenseMap<SUnit *,unsigned> & VRBaseMap,MachineBasicBlock::iterator InsertPos)770 EmitPhysRegCopy(SUnit *SU, DenseMap<SUnit*, unsigned> &VRBaseMap,
771                 MachineBasicBlock::iterator InsertPos) {
772   for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
773        I != E; ++I) {
774     if (I->isCtrl()) continue;  // ignore chain preds
775     if (I->getSUnit()->CopyDstRC) {
776       // Copy to physical register.
777       DenseMap<SUnit*, unsigned>::iterator VRI = VRBaseMap.find(I->getSUnit());
778       assert(VRI != VRBaseMap.end() && "Node emitted out of order - late");
779       // Find the destination physical register.
780       unsigned Reg = 0;
781       for (SUnit::const_succ_iterator II = SU->Succs.begin(),
782              EE = SU->Succs.end(); II != EE; ++II) {
783         if (II->isCtrl()) continue;  // ignore chain preds
784         if (II->getReg()) {
785           Reg = II->getReg();
786           break;
787         }
788       }
789       BuildMI(*BB, InsertPos, DebugLoc(), TII->get(TargetOpcode::COPY), Reg)
790         .addReg(VRI->second);
791     } else {
792       // Copy from physical register.
793       assert(I->getReg() && "Unknown physical register!");
794       unsigned VRBase = MRI.createVirtualRegister(SU->CopyDstRC);
795       bool isNew = VRBaseMap.insert(std::make_pair(SU, VRBase)).second;
796       (void)isNew; // Silence compiler warning.
797       assert(isNew && "Node emitted out of order - early");
798       BuildMI(*BB, InsertPos, DebugLoc(), TII->get(TargetOpcode::COPY), VRBase)
799         .addReg(I->getReg());
800     }
801     break;
802   }
803 }
804 
805 /// EmitSchedule - Emit the machine code in scheduled order. Return the new
806 /// InsertPos and MachineBasicBlock that contains this insertion
807 /// point. ScheduleDAGSDNodes holds a BB pointer for convenience, but this does
808 /// not necessarily refer to returned BB. The emitter may split blocks.
809 MachineBasicBlock *ScheduleDAGSDNodes::
EmitSchedule(MachineBasicBlock::iterator & InsertPos)810 EmitSchedule(MachineBasicBlock::iterator &InsertPos) {
811   InstrEmitter Emitter(BB, InsertPos);
812   DenseMap<SDValue, unsigned> VRBaseMap;
813   DenseMap<SUnit*, unsigned> CopyVRBaseMap;
814   SmallVector<std::pair<unsigned, MachineInstr*>, 32> Orders;
815   SmallSet<unsigned, 8> Seen;
816   bool HasDbg = DAG->hasDebugValues();
817 
818   // If this is the first BB, emit byval parameter dbg_value's.
819   if (HasDbg && BB->getParent()->begin() == MachineFunction::iterator(BB)) {
820     SDDbgInfo::DbgIterator PDI = DAG->ByvalParmDbgBegin();
821     SDDbgInfo::DbgIterator PDE = DAG->ByvalParmDbgEnd();
822     for (; PDI != PDE; ++PDI) {
823       MachineInstr *DbgMI= Emitter.EmitDbgValue(*PDI, VRBaseMap);
824       if (DbgMI) {
825         BB->insert(InsertPos, DbgMI);
826         // We re-emit the dbg_value closer to its use, too, after instructions
827         // are emitted to the BB.
828         (*PDI)->clearIsEmitted();
829       }
830     }
831   }
832 
833   for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
834     SUnit *SU = Sequence[i];
835     if (!SU) {
836       // Null SUnit* is a noop.
837       TII->insertNoop(*Emitter.getBlock(), InsertPos);
838       continue;
839     }
840 
841     // For pre-regalloc scheduling, create instructions corresponding to the
842     // SDNode and any glued SDNodes and append them to the block.
843     if (!SU->getNode()) {
844       // Emit a copy.
845       EmitPhysRegCopy(SU, CopyVRBaseMap, InsertPos);
846       continue;
847     }
848 
849     SmallVector<SDNode *, 4> GluedNodes;
850     for (SDNode *N = SU->getNode()->getGluedNode(); N; N = N->getGluedNode())
851       GluedNodes.push_back(N);
852     while (!GluedNodes.empty()) {
853       SDNode *N = GluedNodes.back();
854       Emitter.EmitNode(N, SU->OrigNode != SU, SU->isCloned, VRBaseMap);
855       // Remember the source order of the inserted instruction.
856       if (HasDbg)
857         ProcessSourceNode(N, DAG, Emitter, VRBaseMap, Orders, Seen);
858       GluedNodes.pop_back();
859     }
860     Emitter.EmitNode(SU->getNode(), SU->OrigNode != SU, SU->isCloned,
861                      VRBaseMap);
862     // Remember the source order of the inserted instruction.
863     if (HasDbg)
864       ProcessSourceNode(SU->getNode(), DAG, Emitter, VRBaseMap, Orders,
865                         Seen);
866   }
867 
868   // Insert all the dbg_values which have not already been inserted in source
869   // order sequence.
870   if (HasDbg) {
871     MachineBasicBlock::iterator BBBegin = BB->getFirstNonPHI();
872 
873     // Sort the source order instructions and use the order to insert debug
874     // values. Use stable_sort so that DBG_VALUEs are inserted in the same order
875     // regardless of the host's implementation fo std::sort.
876     std::stable_sort(Orders.begin(), Orders.end(), less_first());
877     std::stable_sort(DAG->DbgBegin(), DAG->DbgEnd(),
878                      [](const SDDbgValue *LHS, const SDDbgValue *RHS) {
879                        return LHS->getOrder() < RHS->getOrder();
880                      });
881 
882     SDDbgInfo::DbgIterator DI = DAG->DbgBegin();
883     SDDbgInfo::DbgIterator DE = DAG->DbgEnd();
884     // Now emit the rest according to source order.
885     unsigned LastOrder = 0;
886     for (unsigned i = 0, e = Orders.size(); i != e && DI != DE; ++i) {
887       unsigned Order = Orders[i].first;
888       MachineInstr *MI = Orders[i].second;
889       // Insert all SDDbgValue's whose order(s) are before "Order".
890       if (!MI)
891         continue;
892       for (; DI != DE; ++DI) {
893         if ((*DI)->getOrder() < LastOrder || (*DI)->getOrder() >= Order)
894           break;
895         if ((*DI)->isEmitted())
896           continue;
897 
898         MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap);
899         if (DbgMI) {
900           if (!LastOrder)
901             // Insert to start of the BB (after PHIs).
902             BB->insert(BBBegin, DbgMI);
903           else {
904             // Insert at the instruction, which may be in a different
905             // block, if the block was split by a custom inserter.
906             MachineBasicBlock::iterator Pos = MI;
907             MI->getParent()->insert(Pos, DbgMI);
908           }
909         }
910       }
911       LastOrder = Order;
912     }
913     // Add trailing DbgValue's before the terminator. FIXME: May want to add
914     // some of them before one or more conditional branches?
915     SmallVector<MachineInstr*, 8> DbgMIs;
916     for (; DI != DE; ++DI) {
917       if ((*DI)->isEmitted())
918         continue;
919       assert((*DI)->getOrder() >= LastOrder &&
920              "emitting DBG_VALUE out of order");
921       if (MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap))
922         DbgMIs.push_back(DbgMI);
923     }
924 
925     MachineBasicBlock *InsertBB = Emitter.getBlock();
926     MachineBasicBlock::iterator Pos = InsertBB->getFirstTerminator();
927     InsertBB->insert(Pos, DbgMIs.begin(), DbgMIs.end());
928 
929     SDDbgInfo::DbgLabelIterator DLI = DAG->DbgLabelBegin();
930     SDDbgInfo::DbgLabelIterator DLE = DAG->DbgLabelEnd();
931     // Now emit the rest according to source order.
932     LastOrder = 0;
933     for (const auto &InstrOrder : Orders) {
934       unsigned Order = InstrOrder.first;
935       MachineInstr *MI = InstrOrder.second;
936       if (!MI)
937         continue;
938 
939       // Insert all SDDbgLabel's whose order(s) are before "Order".
940       for (; DLI != DLE &&
941              (*DLI)->getOrder() >= LastOrder && (*DLI)->getOrder() < Order;
942              ++DLI) {
943         MachineInstr *DbgMI = Emitter.EmitDbgLabel(*DLI);
944         if (DbgMI) {
945           if (!LastOrder)
946             // Insert to start of the BB (after PHIs).
947             BB->insert(BBBegin, DbgMI);
948           else {
949             // Insert at the instruction, which may be in a different
950             // block, if the block was split by a custom inserter.
951             MachineBasicBlock::iterator Pos = MI;
952             MI->getParent()->insert(Pos, DbgMI);
953           }
954         }
955       }
956       if (DLI == DLE)
957         break;
958 
959       LastOrder = Order;
960     }
961   }
962 
963   InsertPos = Emitter.getInsertPos();
964   return Emitter.getBlock();
965 }
966 
967 /// Return the basic block label.
getDAGName() const968 std::string ScheduleDAGSDNodes::getDAGName() const {
969   return "sunit-dag." + BB->getFullName();
970 }
971