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 
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 (Register::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   SDValue Chain;
203   unsigned NumOps = Node->getNumOperands();
204   if (Node->getOperand(NumOps-1).getValueType() == MVT::Other)
205     Chain = Node->getOperand(NumOps-1);
206   if (!Chain)
207     return;
208 
209   // Skip any load instruction that has a tied input. There may be an additional
210   // dependency requiring a different order than by increasing offsets, and the
211   // added glue may introduce a cycle.
212   auto hasTiedInput = [this](const SDNode *N) {
213     const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());
214     for (unsigned I = 0; I != MCID.getNumOperands(); ++I) {
215       if (MCID.getOperandConstraint(I, MCOI::TIED_TO) != -1)
216         return true;
217     }
218 
219     return false;
220   };
221 
222   // Look for other loads of the same chain. Find loads that are loading from
223   // the same base pointer and different offsets.
224   SmallPtrSet<SDNode*, 16> Visited;
225   SmallVector<int64_t, 4> Offsets;
226   DenseMap<long long, SDNode*> O2SMap;  // Map from offset to SDNode.
227   bool Cluster = false;
228   SDNode *Base = Node;
229 
230   if (hasTiedInput(Base))
231     return;
232 
233   // This algorithm requires a reasonably low use count before finding a match
234   // to avoid uselessly blowing up compile time in large blocks.
235   unsigned UseCount = 0;
236   for (SDNode::use_iterator I = Chain->use_begin(), E = Chain->use_end();
237        I != E && UseCount < 100; ++I, ++UseCount) {
238     if (I.getUse().getResNo() != Chain.getResNo())
239       continue;
240 
241     SDNode *User = *I;
242     if (User == Node || !Visited.insert(User).second)
243       continue;
244     int64_t Offset1, Offset2;
245     if (!TII->areLoadsFromSameBasePtr(Base, User, Offset1, Offset2) ||
246         Offset1 == Offset2 ||
247         hasTiedInput(User)) {
248       // FIXME: Should be ok if they addresses are identical. But earlier
249       // optimizations really should have eliminated one of the loads.
250       continue;
251     }
252     if (O2SMap.insert(std::make_pair(Offset1, Base)).second)
253       Offsets.push_back(Offset1);
254     O2SMap.insert(std::make_pair(Offset2, User));
255     Offsets.push_back(Offset2);
256     if (Offset2 < Offset1)
257       Base = User;
258     Cluster = true;
259     // Reset UseCount to allow more matches.
260     UseCount = 0;
261   }
262 
263   if (!Cluster)
264     return;
265 
266   // Sort them in increasing order.
267   llvm::sort(Offsets);
268 
269   // Check if the loads are close enough.
270   SmallVector<SDNode*, 4> Loads;
271   unsigned NumLoads = 0;
272   int64_t BaseOff = Offsets[0];
273   SDNode *BaseLoad = O2SMap[BaseOff];
274   Loads.push_back(BaseLoad);
275   for (unsigned i = 1, e = Offsets.size(); i != e; ++i) {
276     int64_t Offset = Offsets[i];
277     SDNode *Load = O2SMap[Offset];
278     if (!TII->shouldScheduleLoadsNear(BaseLoad, Load, BaseOff, Offset,NumLoads))
279       break; // Stop right here. Ignore loads that are further away.
280     Loads.push_back(Load);
281     ++NumLoads;
282   }
283 
284   if (NumLoads == 0)
285     return;
286 
287   // Cluster loads by adding MVT::Glue outputs and inputs. This also
288   // ensure they are scheduled in order of increasing addresses.
289   SDNode *Lead = Loads[0];
290   SDValue InGlue = SDValue(nullptr, 0);
291   if (AddGlue(Lead, InGlue, true, DAG))
292     InGlue = SDValue(Lead, Lead->getNumValues() - 1);
293   for (unsigned I = 1, E = Loads.size(); I != E; ++I) {
294     bool OutGlue = I < E - 1;
295     SDNode *Load = Loads[I];
296 
297     // If AddGlue fails, we could leave an unsused glue value. This should not
298     // cause any
299     if (AddGlue(Load, InGlue, OutGlue, DAG)) {
300       if (OutGlue)
301         InGlue = SDValue(Load, Load->getNumValues() - 1);
302 
303       ++LoadsClustered;
304     }
305     else if (!OutGlue && InGlue.getNode())
306       RemoveUnusedGlue(InGlue.getNode(), DAG);
307   }
308 }
309 
310 /// ClusterNodes - Cluster certain nodes which should be scheduled together.
311 ///
ClusterNodes()312 void ScheduleDAGSDNodes::ClusterNodes() {
313   for (SDNode &NI : DAG->allnodes()) {
314     SDNode *Node = &NI;
315     if (!Node || !Node->isMachineOpcode())
316       continue;
317 
318     unsigned Opc = Node->getMachineOpcode();
319     const MCInstrDesc &MCID = TII->get(Opc);
320     if (MCID.mayLoad())
321       // Cluster loads from "near" addresses into combined SUnits.
322       ClusterNeighboringLoads(Node);
323   }
324 }
325 
BuildSchedUnits()326 void ScheduleDAGSDNodes::BuildSchedUnits() {
327   // During scheduling, the NodeId field of SDNode is used to map SDNodes
328   // to their associated SUnits by holding SUnits table indices. A value
329   // of -1 means the SDNode does not yet have an associated SUnit.
330   unsigned NumNodes = 0;
331   for (SDNode &NI : DAG->allnodes()) {
332     NI.setNodeId(-1);
333     ++NumNodes;
334   }
335 
336   // Reserve entries in the vector for each of the SUnits we are creating.  This
337   // ensure that reallocation of the vector won't happen, so SUnit*'s won't get
338   // invalidated.
339   // FIXME: Multiply by 2 because we may clone nodes during scheduling.
340   // This is a temporary workaround.
341   SUnits.reserve(NumNodes * 2);
342 
343   // Add all nodes in depth first order.
344   SmallVector<SDNode*, 64> Worklist;
345   SmallPtrSet<SDNode*, 32> Visited;
346   Worklist.push_back(DAG->getRoot().getNode());
347   Visited.insert(DAG->getRoot().getNode());
348 
349   SmallVector<SUnit*, 8> CallSUnits;
350   while (!Worklist.empty()) {
351     SDNode *NI = Worklist.pop_back_val();
352 
353     // Add all operands to the worklist unless they've already been added.
354     for (const SDValue &Op : NI->op_values())
355       if (Visited.insert(Op.getNode()).second)
356         Worklist.push_back(Op.getNode());
357 
358     if (isPassiveNode(NI))  // Leaf node, e.g. a TargetImmediate.
359       continue;
360 
361     // If this node has already been processed, stop now.
362     if (NI->getNodeId() != -1) continue;
363 
364     SUnit *NodeSUnit = newSUnit(NI);
365 
366     // See if anything is glued to this node, if so, add them to glued
367     // nodes.  Nodes can have at most one glue input and one glue output.  Glue
368     // is required to be the last operand and result of a node.
369 
370     // Scan up to find glued preds.
371     SDNode *N = NI;
372     while (N->getNumOperands() &&
373            N->getOperand(N->getNumOperands()-1).getValueType() == MVT::Glue) {
374       N = N->getOperand(N->getNumOperands()-1).getNode();
375       assert(N->getNodeId() == -1 && "Node already inserted!");
376       N->setNodeId(NodeSUnit->NodeNum);
377       if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall())
378         NodeSUnit->isCall = true;
379     }
380 
381     // Scan down to find any glued succs.
382     N = NI;
383     while (N->getValueType(N->getNumValues()-1) == MVT::Glue) {
384       SDValue GlueVal(N, N->getNumValues()-1);
385 
386       // There are either zero or one users of the Glue result.
387       bool HasGlueUse = false;
388       for (SDNode::use_iterator UI = N->use_begin(), E = N->use_end();
389            UI != E; ++UI)
390         if (GlueVal.isOperandOf(*UI)) {
391           HasGlueUse = true;
392           assert(N->getNodeId() == -1 && "Node already inserted!");
393           N->setNodeId(NodeSUnit->NodeNum);
394           N = *UI;
395           if (N->isMachineOpcode() && TII->get(N->getMachineOpcode()).isCall())
396             NodeSUnit->isCall = true;
397           break;
398         }
399       if (!HasGlueUse) break;
400     }
401 
402     if (NodeSUnit->isCall)
403       CallSUnits.push_back(NodeSUnit);
404 
405     // Schedule zero-latency TokenFactor below any nodes that may increase the
406     // schedule height. Otherwise, ancestors of the TokenFactor may appear to
407     // have false stalls.
408     if (NI->getOpcode() == ISD::TokenFactor)
409       NodeSUnit->isScheduleLow = true;
410 
411     // If there are glue operands involved, N is now the bottom-most node
412     // of the sequence of nodes that are glued together.
413     // Update the SUnit.
414     NodeSUnit->setNode(N);
415     assert(N->getNodeId() == -1 && "Node already inserted!");
416     N->setNodeId(NodeSUnit->NodeNum);
417 
418     // Compute NumRegDefsLeft. This must be done before AddSchedEdges.
419     InitNumRegDefsLeft(NodeSUnit);
420 
421     // Assign the Latency field of NodeSUnit using target-provided information.
422     computeLatency(NodeSUnit);
423   }
424 
425   // Find all call operands.
426   while (!CallSUnits.empty()) {
427     SUnit *SU = CallSUnits.pop_back_val();
428     for (const SDNode *SUNode = SU->getNode(); SUNode;
429          SUNode = SUNode->getGluedNode()) {
430       if (SUNode->getOpcode() != ISD::CopyToReg)
431         continue;
432       SDNode *SrcN = SUNode->getOperand(2).getNode();
433       if (isPassiveNode(SrcN)) continue;   // Not scheduled.
434       SUnit *SrcSU = &SUnits[SrcN->getNodeId()];
435       SrcSU->isCallOp = true;
436     }
437   }
438 }
439 
AddSchedEdges()440 void ScheduleDAGSDNodes::AddSchedEdges() {
441   const TargetSubtargetInfo &ST = MF.getSubtarget();
442 
443   // Check to see if the scheduler cares about latencies.
444   bool UnitLatencies = forceUnitLatencies();
445 
446   // Pass 2: add the preds, succs, etc.
447   for (unsigned su = 0, e = SUnits.size(); su != e; ++su) {
448     SUnit *SU = &SUnits[su];
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()).getImplicitDefs()) {
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) continue;           // In the same group.
483 
484         EVT OpVT = N->getOperand(i).getValueType();
485         assert(OpVT != MVT::Glue && "Glued nodes should be in same sunit!");
486         bool isChain = OpVT == MVT::Other;
487 
488         unsigned PhysReg = 0;
489         int Cost = 1;
490         // Determine if this is a physical register dependency.
491         CheckForPhysRegDependency(OpN, N, i, TRI, TII, PhysReg, Cost);
492         assert((PhysReg == 0 || !isChain) &&
493                "Chain dependence via physreg data?");
494         // FIXME: See ScheduleDAGSDNodes::EmitCopyFromReg. For now, scheduler
495         // emits a copy from the physical register to a virtual register unless
496         // it requires a cross class copy (cost < 0). That means we are only
497         // treating "expensive to copy" register dependency as physical register
498         // dependency. This may change in the future though.
499         if (Cost >= 0 && !StressSched)
500           PhysReg = 0;
501 
502         // If this is a ctrl dep, latency is 1.
503         unsigned OpLatency = isChain ? 1 : OpSU->Latency;
504         // Special-case TokenFactor chains as zero-latency.
505         if(isChain && OpN->getOpcode() == ISD::TokenFactor)
506           OpLatency = 0;
507 
508         SDep Dep = isChain ? SDep(OpSU, SDep::Barrier)
509           : SDep(OpSU, SDep::Data, PhysReg);
510         Dep.setLatency(OpLatency);
511         if (!isChain && !UnitLatencies) {
512           computeOperandLatency(OpN, N, i, Dep);
513           ST.adjustSchedDependency(OpSU, DefIdx, SU, i, Dep);
514         }
515 
516         if (!SU->addPred(Dep) && !Dep.isCtrl() && OpSU->NumRegDefsLeft > 1) {
517           // Multiple register uses are combined in the same SUnit. For example,
518           // we could have a set of glued nodes with all their defs consumed by
519           // another set of glued nodes. Register pressure tracking sees this as
520           // a single use, so to keep pressure balanced we reduce the defs.
521           //
522           // We can't tell (without more book-keeping) if this results from
523           // glued nodes or duplicate operands. As long as we don't reduce
524           // NumRegDefsLeft to zero, we handle the common cases well.
525           --OpSU->NumRegDefsLeft;
526         }
527       }
528     }
529   }
530 }
531 
532 /// BuildSchedGraph - Build the SUnit graph from the selection dag that we
533 /// are input.  This SUnit graph is similar to the SelectionDAG, but
534 /// excludes nodes that aren't interesting to scheduling, and represents
535 /// glued together nodes with a single SUnit.
BuildSchedGraph(AAResults * AA)536 void ScheduleDAGSDNodes::BuildSchedGraph(AAResults *AA) {
537   // Cluster certain nodes which should be scheduled together.
538   ClusterNodes();
539   // Populate the SUnits array.
540   BuildSchedUnits();
541   // Compute all the scheduling dependencies between nodes.
542   AddSchedEdges();
543 }
544 
545 // Initialize NumNodeDefs for the current Node's opcode.
InitNodeNumDefs()546 void ScheduleDAGSDNodes::RegDefIter::InitNodeNumDefs() {
547   // Check for phys reg copy.
548   if (!Node)
549     return;
550 
551   if (!Node->isMachineOpcode()) {
552     if (Node->getOpcode() == ISD::CopyFromReg)
553       NodeNumDefs = 1;
554     else
555       NodeNumDefs = 0;
556     return;
557   }
558   unsigned POpc = Node->getMachineOpcode();
559   if (POpc == TargetOpcode::IMPLICIT_DEF) {
560     // No register need be allocated for this.
561     NodeNumDefs = 0;
562     return;
563   }
564   if (POpc == TargetOpcode::PATCHPOINT &&
565       Node->getValueType(0) == MVT::Other) {
566     // PATCHPOINT is defined to have one result, but it might really have none
567     // if we're not using CallingConv::AnyReg. Don't mistake the chain for a
568     // real definition.
569     NodeNumDefs = 0;
570     return;
571   }
572   unsigned NRegDefs = SchedDAG->TII->get(Node->getMachineOpcode()).getNumDefs();
573   // Some instructions define regs that are not represented in the selection DAG
574   // (e.g. unused flags). See tMOVi8. Make sure we don't access past NumValues.
575   NodeNumDefs = std::min(Node->getNumValues(), NRegDefs);
576   DefIdx = 0;
577 }
578 
579 // Construct a RegDefIter for this SUnit and find the first valid value.
RegDefIter(const SUnit * SU,const ScheduleDAGSDNodes * SD)580 ScheduleDAGSDNodes::RegDefIter::RegDefIter(const SUnit *SU,
581                                            const ScheduleDAGSDNodes *SD)
582   : SchedDAG(SD), Node(SU->getNode()), DefIdx(0), NodeNumDefs(0) {
583   InitNodeNumDefs();
584   Advance();
585 }
586 
587 // Advance to the next valid value defined by the SUnit.
Advance()588 void ScheduleDAGSDNodes::RegDefIter::Advance() {
589   for (;Node;) { // Visit all glued nodes.
590     for (;DefIdx < NodeNumDefs; ++DefIdx) {
591       if (!Node->hasAnyUseOfValue(DefIdx))
592         continue;
593       ValueType = Node->getSimpleValueType(DefIdx);
594       ++DefIdx;
595       return; // Found a normal regdef.
596     }
597     Node = Node->getGluedNode();
598     if (!Node) {
599       return; // No values left to visit.
600     }
601     InitNodeNumDefs();
602   }
603 }
604 
InitNumRegDefsLeft(SUnit * SU)605 void ScheduleDAGSDNodes::InitNumRegDefsLeft(SUnit *SU) {
606   assert(SU->NumRegDefsLeft == 0 && "expect a new node");
607   for (RegDefIter I(SU, this); I.IsValid(); I.Advance()) {
608     assert(SU->NumRegDefsLeft < USHRT_MAX && "overflow is ok but unexpected");
609     ++SU->NumRegDefsLeft;
610   }
611 }
612 
computeLatency(SUnit * SU)613 void ScheduleDAGSDNodes::computeLatency(SUnit *SU) {
614   SDNode *N = SU->getNode();
615 
616   // TokenFactor operands are considered zero latency, and some schedulers
617   // (e.g. Top-Down list) may rely on the fact that operand latency is nonzero
618   // whenever node latency is nonzero.
619   if (N && N->getOpcode() == ISD::TokenFactor) {
620     SU->Latency = 0;
621     return;
622   }
623 
624   // Check to see if the scheduler cares about latencies.
625   if (forceUnitLatencies()) {
626     SU->Latency = 1;
627     return;
628   }
629 
630   if (!InstrItins || InstrItins->isEmpty()) {
631     if (N && N->isMachineOpcode() &&
632         TII->isHighLatencyDef(N->getMachineOpcode()))
633       SU->Latency = HighLatencyCycles;
634     else
635       SU->Latency = 1;
636     return;
637   }
638 
639   // Compute the latency for the node.  We use the sum of the latencies for
640   // all nodes glued together into this SUnit.
641   SU->Latency = 0;
642   for (SDNode *N = SU->getNode(); N; N = N->getGluedNode())
643     if (N->isMachineOpcode())
644       SU->Latency += TII->getInstrLatency(InstrItins, N);
645 }
646 
computeOperandLatency(SDNode * Def,SDNode * Use,unsigned OpIdx,SDep & dep) const647 void ScheduleDAGSDNodes::computeOperandLatency(SDNode *Def, SDNode *Use,
648                                                unsigned OpIdx, SDep& dep) const{
649   // Check to see if the scheduler cares about latencies.
650   if (forceUnitLatencies())
651     return;
652 
653   if (dep.getKind() != SDep::Data)
654     return;
655 
656   unsigned DefIdx = Use->getOperand(OpIdx).getResNo();
657   if (Use->isMachineOpcode())
658     // Adjust the use operand index by num of defs.
659     OpIdx += TII->get(Use->getMachineOpcode()).getNumDefs();
660   int Latency = TII->getOperandLatency(InstrItins, Def, DefIdx, Use, OpIdx);
661   if (Latency > 1 && Use->getOpcode() == ISD::CopyToReg &&
662       !BB->succ_empty()) {
663     unsigned Reg = cast<RegisterSDNode>(Use->getOperand(1))->getReg();
664     if (Register::isVirtualRegister(Reg))
665       // This copy is a liveout value. It is likely coalesced, so reduce the
666       // latency so not to penalize the def.
667       // FIXME: need target specific adjustment here?
668       Latency = (Latency > 1) ? Latency - 1 : 1;
669   }
670   if (Latency >= 0)
671     dep.setLatency(Latency);
672 }
673 
dumpNode(const SUnit & SU) const674 void ScheduleDAGSDNodes::dumpNode(const SUnit &SU) const {
675 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
676   dumpNodeName(SU);
677   dbgs() << ": ";
678 
679   if (!SU.getNode()) {
680     dbgs() << "PHYS REG COPY\n";
681     return;
682   }
683 
684   SU.getNode()->dump(DAG);
685   dbgs() << "\n";
686   SmallVector<SDNode *, 4> GluedNodes;
687   for (SDNode *N = SU.getNode()->getGluedNode(); N; N = N->getGluedNode())
688     GluedNodes.push_back(N);
689   while (!GluedNodes.empty()) {
690     dbgs() << "    ";
691     GluedNodes.back()->dump(DAG);
692     dbgs() << "\n";
693     GluedNodes.pop_back();
694   }
695 #endif
696 }
697 
dump() const698 void ScheduleDAGSDNodes::dump() const {
699 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
700   if (EntrySU.getNode() != nullptr)
701     dumpNodeAll(EntrySU);
702   for (const SUnit &SU : SUnits)
703     dumpNodeAll(SU);
704   if (ExitSU.getNode() != nullptr)
705     dumpNodeAll(ExitSU);
706 #endif
707 }
708 
709 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
dumpSchedule() const710 void ScheduleDAGSDNodes::dumpSchedule() const {
711   for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
712     if (SUnit *SU = Sequence[i])
713       dumpNode(*SU);
714     else
715       dbgs() << "**** NOOP ****\n";
716   }
717 }
718 #endif
719 
720 #ifndef NDEBUG
721 /// VerifyScheduledSequence - Verify that all SUnits were scheduled and that
722 /// their state is consistent with the nodes listed in Sequence.
723 ///
VerifyScheduledSequence(bool isBottomUp)724 void ScheduleDAGSDNodes::VerifyScheduledSequence(bool isBottomUp) {
725   unsigned ScheduledNodes = ScheduleDAG::VerifyScheduledDAG(isBottomUp);
726   unsigned Noops = 0;
727   for (unsigned i = 0, e = Sequence.size(); i != e; ++i)
728     if (!Sequence[i])
729       ++Noops;
730   assert(Sequence.size() - Noops == ScheduledNodes &&
731          "The number of nodes scheduled doesn't match the expected number!");
732 }
733 #endif // NDEBUG
734 
735 /// ProcessSDDbgValues - Process SDDbgValues associated with this node.
736 static void
ProcessSDDbgValues(SDNode * N,SelectionDAG * DAG,InstrEmitter & Emitter,SmallVectorImpl<std::pair<unsigned,MachineInstr * >> & Orders,DenseMap<SDValue,Register> & VRBaseMap,unsigned Order)737 ProcessSDDbgValues(SDNode *N, SelectionDAG *DAG, InstrEmitter &Emitter,
738                    SmallVectorImpl<std::pair<unsigned, MachineInstr*> > &Orders,
739                    DenseMap<SDValue, Register> &VRBaseMap, unsigned Order) {
740   if (!N->getHasDebugValue())
741     return;
742 
743   // Opportunistically insert immediate dbg_value uses, i.e. those with the same
744   // source order number as N.
745   MachineBasicBlock *BB = Emitter.getBlock();
746   MachineBasicBlock::iterator InsertPos = Emitter.getInsertPos();
747   for (auto DV : DAG->GetDbgValues(N)) {
748     if (DV->isEmitted())
749       continue;
750     unsigned DVOrder = DV->getOrder();
751     if (!Order || DVOrder == Order) {
752       MachineInstr *DbgMI = Emitter.EmitDbgValue(DV, VRBaseMap);
753       if (DbgMI) {
754         Orders.push_back({DVOrder, DbgMI});
755         BB->insert(InsertPos, DbgMI);
756       }
757     }
758   }
759 }
760 
761 // ProcessSourceNode - Process nodes with source order numbers. These are added
762 // to a vector which EmitSchedule uses to determine how to insert dbg_value
763 // instructions in the right order.
764 static void
ProcessSourceNode(SDNode * N,SelectionDAG * DAG,InstrEmitter & Emitter,DenseMap<SDValue,Register> & VRBaseMap,SmallVectorImpl<std::pair<unsigned,MachineInstr * >> & Orders,SmallSet<Register,8> & Seen,MachineInstr * NewInsn)765 ProcessSourceNode(SDNode *N, SelectionDAG *DAG, InstrEmitter &Emitter,
766                   DenseMap<SDValue, Register> &VRBaseMap,
767                   SmallVectorImpl<std::pair<unsigned, MachineInstr *>> &Orders,
768                   SmallSet<Register, 8> &Seen, MachineInstr *NewInsn) {
769   unsigned Order = N->getIROrder();
770   if (!Order || Seen.count(Order)) {
771     // Process any valid SDDbgValues even if node does not have any order
772     // assigned.
773     ProcessSDDbgValues(N, DAG, Emitter, Orders, VRBaseMap, 0);
774     return;
775   }
776 
777   // If a new instruction was generated for this Order number, record it.
778   // Otherwise, leave this order number unseen: we will either find later
779   // instructions for it, or leave it unseen if there were no instructions at
780   // all.
781   if (NewInsn) {
782     Seen.insert(Order);
783     Orders.push_back({Order, NewInsn});
784   }
785 
786   // Even if no instruction was generated, a Value may have become defined via
787   // earlier nodes. Try to process them now.
788   ProcessSDDbgValues(N, DAG, Emitter, Orders, VRBaseMap, Order);
789 }
790 
791 void ScheduleDAGSDNodes::
EmitPhysRegCopy(SUnit * SU,DenseMap<SUnit *,Register> & VRBaseMap,MachineBasicBlock::iterator InsertPos)792 EmitPhysRegCopy(SUnit *SU, DenseMap<SUnit*, Register> &VRBaseMap,
793                 MachineBasicBlock::iterator InsertPos) {
794   for (SUnit::const_pred_iterator I = SU->Preds.begin(), E = SU->Preds.end();
795        I != E; ++I) {
796     if (I->isCtrl()) continue;  // ignore chain preds
797     if (I->getSUnit()->CopyDstRC) {
798       // Copy to physical register.
799       DenseMap<SUnit*, Register>::iterator VRI = VRBaseMap.find(I->getSUnit());
800       assert(VRI != VRBaseMap.end() && "Node emitted out of order - late");
801       // Find the destination physical register.
802       Register Reg;
803       for (SUnit::const_succ_iterator II = SU->Succs.begin(),
804              EE = SU->Succs.end(); II != EE; ++II) {
805         if (II->isCtrl()) continue;  // ignore chain preds
806         if (II->getReg()) {
807           Reg = II->getReg();
808           break;
809         }
810       }
811       BuildMI(*BB, InsertPos, DebugLoc(), TII->get(TargetOpcode::COPY), Reg)
812         .addReg(VRI->second);
813     } else {
814       // Copy from physical register.
815       assert(I->getReg() && "Unknown physical register!");
816       Register VRBase = MRI.createVirtualRegister(SU->CopyDstRC);
817       bool isNew = VRBaseMap.insert(std::make_pair(SU, VRBase)).second;
818       (void)isNew; // Silence compiler warning.
819       assert(isNew && "Node emitted out of order - early");
820       BuildMI(*BB, InsertPos, DebugLoc(), TII->get(TargetOpcode::COPY), VRBase)
821         .addReg(I->getReg());
822     }
823     break;
824   }
825 }
826 
827 /// EmitSchedule - Emit the machine code in scheduled order. Return the new
828 /// InsertPos and MachineBasicBlock that contains this insertion
829 /// point. ScheduleDAGSDNodes holds a BB pointer for convenience, but this does
830 /// not necessarily refer to returned BB. The emitter may split blocks.
831 MachineBasicBlock *ScheduleDAGSDNodes::
EmitSchedule(MachineBasicBlock::iterator & InsertPos)832 EmitSchedule(MachineBasicBlock::iterator &InsertPos) {
833   InstrEmitter Emitter(BB, InsertPos);
834   DenseMap<SDValue, Register> VRBaseMap;
835   DenseMap<SUnit*, Register> CopyVRBaseMap;
836   SmallVector<std::pair<unsigned, MachineInstr*>, 32> Orders;
837   SmallSet<Register, 8> Seen;
838   bool HasDbg = DAG->hasDebugValues();
839 
840   // Emit a node, and determine where its first instruction is for debuginfo.
841   // Zero, one, or multiple instructions can be created when emitting a node.
842   auto EmitNode =
843       [&](SDNode *Node, bool IsClone, bool IsCloned,
844           DenseMap<SDValue, Register> &VRBaseMap) -> MachineInstr * {
845     // Fetch instruction prior to this, or end() if nonexistant.
846     auto GetPrevInsn = [&](MachineBasicBlock::iterator I) {
847       if (I == BB->begin())
848         return BB->end();
849       else
850         return std::prev(Emitter.getInsertPos());
851     };
852 
853     MachineBasicBlock::iterator Before = GetPrevInsn(Emitter.getInsertPos());
854     Emitter.EmitNode(Node, IsClone, IsCloned, VRBaseMap);
855     MachineBasicBlock::iterator After = GetPrevInsn(Emitter.getInsertPos());
856 
857     // If the iterator did not change, no instructions were inserted.
858     if (Before == After)
859       return nullptr;
860 
861     MachineInstr *MI;
862     if (Before == BB->end()) {
863       // There were no prior instructions; the new ones must start at the
864       // beginning of the block.
865       MI = &Emitter.getBlock()->instr_front();
866     } else {
867       // Return first instruction after the pre-existing instructions.
868       MI = &*std::next(Before);
869     }
870 
871     if (MI->isCandidateForCallSiteEntry() &&
872         DAG->getTarget().Options.EmitCallSiteInfo)
873       MF.addCallArgsForwardingRegs(MI, DAG->getSDCallSiteInfo(Node));
874 
875     if (DAG->getNoMergeSiteInfo(Node)) {
876       MI->setFlag(MachineInstr::MIFlag::NoMerge);
877     }
878 
879     return MI;
880   };
881 
882   // If this is the first BB, emit byval parameter dbg_value's.
883   if (HasDbg && BB->getParent()->begin() == MachineFunction::iterator(BB)) {
884     SDDbgInfo::DbgIterator PDI = DAG->ByvalParmDbgBegin();
885     SDDbgInfo::DbgIterator PDE = DAG->ByvalParmDbgEnd();
886     for (; PDI != PDE; ++PDI) {
887       MachineInstr *DbgMI= Emitter.EmitDbgValue(*PDI, VRBaseMap);
888       if (DbgMI) {
889         BB->insert(InsertPos, DbgMI);
890         // We re-emit the dbg_value closer to its use, too, after instructions
891         // are emitted to the BB.
892         (*PDI)->clearIsEmitted();
893       }
894     }
895   }
896 
897   for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
898     SUnit *SU = Sequence[i];
899     if (!SU) {
900       // Null SUnit* is a noop.
901       TII->insertNoop(*Emitter.getBlock(), InsertPos);
902       continue;
903     }
904 
905     // For pre-regalloc scheduling, create instructions corresponding to the
906     // SDNode and any glued SDNodes and append them to the block.
907     if (!SU->getNode()) {
908       // Emit a copy.
909       EmitPhysRegCopy(SU, CopyVRBaseMap, InsertPos);
910       continue;
911     }
912 
913     SmallVector<SDNode *, 4> GluedNodes;
914     for (SDNode *N = SU->getNode()->getGluedNode(); N; N = N->getGluedNode())
915       GluedNodes.push_back(N);
916     while (!GluedNodes.empty()) {
917       SDNode *N = GluedNodes.back();
918       auto NewInsn = EmitNode(N, SU->OrigNode != SU, SU->isCloned, VRBaseMap);
919       // Remember the source order of the inserted instruction.
920       if (HasDbg)
921         ProcessSourceNode(N, DAG, Emitter, VRBaseMap, Orders, Seen, NewInsn);
922 
923       if (MDNode *MD = DAG->getHeapAllocSite(N))
924         if (NewInsn && NewInsn->isCall())
925           NewInsn->setHeapAllocMarker(MF, MD);
926 
927       GluedNodes.pop_back();
928     }
929     auto NewInsn =
930         EmitNode(SU->getNode(), SU->OrigNode != SU, SU->isCloned, VRBaseMap);
931     // Remember the source order of the inserted instruction.
932     if (HasDbg)
933       ProcessSourceNode(SU->getNode(), DAG, Emitter, VRBaseMap, Orders, Seen,
934                         NewInsn);
935 
936     if (MDNode *MD = DAG->getHeapAllocSite(SU->getNode())) {
937       if (NewInsn && NewInsn->isCall())
938         NewInsn->setHeapAllocMarker(MF, MD);
939     }
940   }
941 
942   // Insert all the dbg_values which have not already been inserted in source
943   // order sequence.
944   if (HasDbg) {
945     MachineBasicBlock::iterator BBBegin = BB->getFirstNonPHI();
946 
947     // Sort the source order instructions and use the order to insert debug
948     // values. Use stable_sort so that DBG_VALUEs are inserted in the same order
949     // regardless of the host's implementation fo std::sort.
950     llvm::stable_sort(Orders, less_first());
951     std::stable_sort(DAG->DbgBegin(), DAG->DbgEnd(),
952                      [](const SDDbgValue *LHS, const SDDbgValue *RHS) {
953                        return LHS->getOrder() < RHS->getOrder();
954                      });
955 
956     SDDbgInfo::DbgIterator DI = DAG->DbgBegin();
957     SDDbgInfo::DbgIterator DE = DAG->DbgEnd();
958     // Now emit the rest according to source order.
959     unsigned LastOrder = 0;
960     for (unsigned i = 0, e = Orders.size(); i != e && DI != DE; ++i) {
961       unsigned Order = Orders[i].first;
962       MachineInstr *MI = Orders[i].second;
963       // Insert all SDDbgValue's whose order(s) are before "Order".
964       assert(MI);
965       for (; DI != DE; ++DI) {
966         if ((*DI)->getOrder() < LastOrder || (*DI)->getOrder() >= Order)
967           break;
968         if ((*DI)->isEmitted())
969           continue;
970 
971         MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap);
972         if (DbgMI) {
973           if (!LastOrder)
974             // Insert to start of the BB (after PHIs).
975             BB->insert(BBBegin, DbgMI);
976           else {
977             // Insert at the instruction, which may be in a different
978             // block, if the block was split by a custom inserter.
979             MachineBasicBlock::iterator Pos = MI;
980             MI->getParent()->insert(Pos, DbgMI);
981           }
982         }
983       }
984       LastOrder = Order;
985     }
986     // Add trailing DbgValue's before the terminator. FIXME: May want to add
987     // some of them before one or more conditional branches?
988     SmallVector<MachineInstr*, 8> DbgMIs;
989     for (; DI != DE; ++DI) {
990       if ((*DI)->isEmitted())
991         continue;
992       assert((*DI)->getOrder() >= LastOrder &&
993              "emitting DBG_VALUE out of order");
994       if (MachineInstr *DbgMI = Emitter.EmitDbgValue(*DI, VRBaseMap))
995         DbgMIs.push_back(DbgMI);
996     }
997 
998     MachineBasicBlock *InsertBB = Emitter.getBlock();
999     MachineBasicBlock::iterator Pos = InsertBB->getFirstTerminator();
1000     InsertBB->insert(Pos, DbgMIs.begin(), DbgMIs.end());
1001 
1002     SDDbgInfo::DbgLabelIterator DLI = DAG->DbgLabelBegin();
1003     SDDbgInfo::DbgLabelIterator DLE = DAG->DbgLabelEnd();
1004     // Now emit the rest according to source order.
1005     LastOrder = 0;
1006     for (const auto &InstrOrder : Orders) {
1007       unsigned Order = InstrOrder.first;
1008       MachineInstr *MI = InstrOrder.second;
1009       if (!MI)
1010         continue;
1011 
1012       // Insert all SDDbgLabel's whose order(s) are before "Order".
1013       for (; DLI != DLE &&
1014              (*DLI)->getOrder() >= LastOrder && (*DLI)->getOrder() < Order;
1015              ++DLI) {
1016         MachineInstr *DbgMI = Emitter.EmitDbgLabel(*DLI);
1017         if (DbgMI) {
1018           if (!LastOrder)
1019             // Insert to start of the BB (after PHIs).
1020             BB->insert(BBBegin, DbgMI);
1021           else {
1022             // Insert at the instruction, which may be in a different
1023             // block, if the block was split by a custom inserter.
1024             MachineBasicBlock::iterator Pos = MI;
1025             MI->getParent()->insert(Pos, DbgMI);
1026           }
1027         }
1028       }
1029       if (DLI == DLE)
1030         break;
1031 
1032       LastOrder = Order;
1033     }
1034   }
1035 
1036   InsertPos = Emitter.getInsertPos();
1037   return Emitter.getBlock();
1038 }
1039 
1040 /// Return the basic block label.
getDAGName() const1041 std::string ScheduleDAGSDNodes::getDAGName() const {
1042   return "sunit-dag." + BB->getFullName();
1043 }
1044