1 //===- MachinePipeliner.cpp - Machine Software Pipeliner Pass -------------===//
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 // An implementation of the Swing Modulo Scheduling (SMS) software pipeliner.
10 //
11 // This SMS implementation is a target-independent back-end pass. When enabled,
12 // the pass runs just prior to the register allocation pass, while the machine
13 // IR is in SSA form. If software pipelining is successful, then the original
14 // loop is replaced by the optimized loop. The optimized loop contains one or
15 // more prolog blocks, the pipelined kernel, and one or more epilog blocks. If
16 // the instructions cannot be scheduled in a given MII, we increase the MII by
17 // one and try again.
18 //
19 // The SMS implementation is an extension of the ScheduleDAGInstrs class. We
20 // represent loop carried dependences in the DAG as order edges to the Phi
21 // nodes. We also perform several passes over the DAG to eliminate unnecessary
22 // edges that inhibit the ability to pipeline. The implementation uses the
23 // DFAPacketizer class to compute the minimum initiation interval and the check
24 // where an instruction may be inserted in the pipelined schedule.
25 //
26 // In order for the SMS pass to work, several target specific hooks need to be
27 // implemented to get information about the loop structure and to rewrite
28 // instructions.
29 //
30 //===----------------------------------------------------------------------===//
31 
32 #include "llvm/ADT/ArrayRef.h"
33 #include "llvm/ADT/BitVector.h"
34 #include "llvm/ADT/DenseMap.h"
35 #include "llvm/ADT/MapVector.h"
36 #include "llvm/ADT/PriorityQueue.h"
37 #include "llvm/ADT/SetVector.h"
38 #include "llvm/ADT/SmallPtrSet.h"
39 #include "llvm/ADT/SmallSet.h"
40 #include "llvm/ADT/SmallVector.h"
41 #include "llvm/ADT/Statistic.h"
42 #include "llvm/ADT/iterator_range.h"
43 #include "llvm/Analysis/AliasAnalysis.h"
44 #include "llvm/Analysis/MemoryLocation.h"
45 #include "llvm/Analysis/ValueTracking.h"
46 #include "llvm/CodeGen/DFAPacketizer.h"
47 #include "llvm/CodeGen/LiveIntervals.h"
48 #include "llvm/CodeGen/MachineBasicBlock.h"
49 #include "llvm/CodeGen/MachineDominators.h"
50 #include "llvm/CodeGen/MachineFunction.h"
51 #include "llvm/CodeGen/MachineFunctionPass.h"
52 #include "llvm/CodeGen/MachineInstr.h"
53 #include "llvm/CodeGen/MachineInstrBuilder.h"
54 #include "llvm/CodeGen/MachineLoopInfo.h"
55 #include "llvm/CodeGen/MachineMemOperand.h"
56 #include "llvm/CodeGen/MachineOperand.h"
57 #include "llvm/CodeGen/MachinePipeliner.h"
58 #include "llvm/CodeGen/MachineRegisterInfo.h"
59 #include "llvm/CodeGen/ModuloSchedule.h"
60 #include "llvm/CodeGen/RegisterPressure.h"
61 #include "llvm/CodeGen/ScheduleDAG.h"
62 #include "llvm/CodeGen/ScheduleDAGMutation.h"
63 #include "llvm/CodeGen/TargetOpcodes.h"
64 #include "llvm/CodeGen/TargetRegisterInfo.h"
65 #include "llvm/CodeGen/TargetSubtargetInfo.h"
66 #include "llvm/Config/llvm-config.h"
67 #include "llvm/IR/Attributes.h"
68 #include "llvm/IR/DebugLoc.h"
69 #include "llvm/IR/Function.h"
70 #include "llvm/MC/LaneBitmask.h"
71 #include "llvm/MC/MCInstrDesc.h"
72 #include "llvm/MC/MCInstrItineraries.h"
73 #include "llvm/MC/MCRegisterInfo.h"
74 #include "llvm/Pass.h"
75 #include "llvm/Support/CommandLine.h"
76 #include "llvm/Support/Compiler.h"
77 #include "llvm/Support/Debug.h"
78 #include "llvm/Support/MathExtras.h"
79 #include "llvm/Support/raw_ostream.h"
80 #include <algorithm>
81 #include <cassert>
82 #include <climits>
83 #include <cstdint>
84 #include <deque>
85 #include <functional>
86 #include <iterator>
87 #include <map>
88 #include <memory>
89 #include <tuple>
90 #include <utility>
91 #include <vector>
92 
93 using namespace llvm;
94 
95 #define DEBUG_TYPE "pipeliner"
96 
97 STATISTIC(NumTrytoPipeline, "Number of loops that we attempt to pipeline");
98 STATISTIC(NumPipelined, "Number of loops software pipelined");
99 STATISTIC(NumNodeOrderIssues, "Number of node order issues found");
100 STATISTIC(NumFailBranch, "Pipeliner abort due to unknown branch");
101 STATISTIC(NumFailLoop, "Pipeliner abort due to unsupported loop");
102 STATISTIC(NumFailPreheader, "Pipeliner abort due to missing preheader");
103 STATISTIC(NumFailLargeMaxMII, "Pipeliner abort due to MaxMII too large");
104 STATISTIC(NumFailZeroMII, "Pipeliner abort due to zero MII");
105 STATISTIC(NumFailNoSchedule, "Pipeliner abort due to no schedule found");
106 STATISTIC(NumFailZeroStage, "Pipeliner abort due to zero stage");
107 STATISTIC(NumFailLargeMaxStage, "Pipeliner abort due to too many stages");
108 
109 /// A command line option to turn software pipelining on or off.
110 static cl::opt<bool> EnableSWP("enable-pipeliner", cl::Hidden, cl::init(true),
111                                cl::ZeroOrMore,
112                                cl::desc("Enable Software Pipelining"));
113 
114 /// A command line option to enable SWP at -Os.
115 static cl::opt<bool> EnableSWPOptSize("enable-pipeliner-opt-size",
116                                       cl::desc("Enable SWP at Os."), cl::Hidden,
117                                       cl::init(false));
118 
119 /// A command line argument to limit minimum initial interval for pipelining.
120 static cl::opt<int> SwpMaxMii("pipeliner-max-mii",
121                               cl::desc("Size limit for the MII."),
122                               cl::Hidden, cl::init(27));
123 
124 /// A command line argument to limit the number of stages in the pipeline.
125 static cl::opt<int>
126     SwpMaxStages("pipeliner-max-stages",
127                  cl::desc("Maximum stages allowed in the generated scheduled."),
128                  cl::Hidden, cl::init(3));
129 
130 /// A command line option to disable the pruning of chain dependences due to
131 /// an unrelated Phi.
132 static cl::opt<bool>
133     SwpPruneDeps("pipeliner-prune-deps",
134                  cl::desc("Prune dependences between unrelated Phi nodes."),
135                  cl::Hidden, cl::init(true));
136 
137 /// A command line option to disable the pruning of loop carried order
138 /// dependences.
139 static cl::opt<bool>
140     SwpPruneLoopCarried("pipeliner-prune-loop-carried",
141                         cl::desc("Prune loop carried order dependences."),
142                         cl::Hidden, cl::init(true));
143 
144 #ifndef NDEBUG
145 static cl::opt<int> SwpLoopLimit("pipeliner-max", cl::Hidden, cl::init(-1));
146 #endif
147 
148 static cl::opt<bool> SwpIgnoreRecMII("pipeliner-ignore-recmii",
149                                      cl::ReallyHidden, cl::init(false),
150                                      cl::ZeroOrMore, cl::desc("Ignore RecMII"));
151 
152 static cl::opt<bool> SwpShowResMask("pipeliner-show-mask", cl::Hidden,
153                                     cl::init(false));
154 static cl::opt<bool> SwpDebugResource("pipeliner-dbg-res", cl::Hidden,
155                                       cl::init(false));
156 
157 static cl::opt<bool> EmitTestAnnotations(
158     "pipeliner-annotate-for-testing", cl::Hidden, cl::init(false),
159     cl::desc("Instead of emitting the pipelined code, annotate instructions "
160              "with the generated schedule for feeding into the "
161              "-modulo-schedule-test pass"));
162 
163 static cl::opt<bool> ExperimentalCodeGen(
164     "pipeliner-experimental-cg", cl::Hidden, cl::init(false),
165     cl::desc(
166         "Use the experimental peeling code generator for software pipelining"));
167 
168 namespace llvm {
169 
170 // A command line option to enable the CopyToPhi DAG mutation.
171 cl::opt<bool>
172     SwpEnableCopyToPhi("pipeliner-enable-copytophi", cl::ReallyHidden,
173                        cl::init(true), cl::ZeroOrMore,
174                        cl::desc("Enable CopyToPhi DAG Mutation"));
175 
176 } // end namespace llvm
177 
178 unsigned SwingSchedulerDAG::Circuits::MaxPaths = 5;
179 char MachinePipeliner::ID = 0;
180 #ifndef NDEBUG
181 int MachinePipeliner::NumTries = 0;
182 #endif
183 char &llvm::MachinePipelinerID = MachinePipeliner::ID;
184 
185 INITIALIZE_PASS_BEGIN(MachinePipeliner, DEBUG_TYPE,
186                       "Modulo Software Pipelining", false, false)
187 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
188 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
189 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
190 INITIALIZE_PASS_DEPENDENCY(LiveIntervals)
191 INITIALIZE_PASS_END(MachinePipeliner, DEBUG_TYPE,
192                     "Modulo Software Pipelining", false, false)
193 
194 /// The "main" function for implementing Swing Modulo Scheduling.
195 bool MachinePipeliner::runOnMachineFunction(MachineFunction &mf) {
196   if (skipFunction(mf.getFunction()))
197     return false;
198 
199   if (!EnableSWP)
200     return false;
201 
202   if (mf.getFunction().getAttributes().hasAttribute(
203           AttributeList::FunctionIndex, Attribute::OptimizeForSize) &&
204       !EnableSWPOptSize.getPosition())
205     return false;
206 
207   if (!mf.getSubtarget().enableMachinePipeliner())
208     return false;
209 
210   // Cannot pipeline loops without instruction itineraries if we are using
211   // DFA for the pipeliner.
212   if (mf.getSubtarget().useDFAforSMS() &&
213       (!mf.getSubtarget().getInstrItineraryData() ||
214        mf.getSubtarget().getInstrItineraryData()->isEmpty()))
215     return false;
216 
217   MF = &mf;
218   MLI = &getAnalysis<MachineLoopInfo>();
219   MDT = &getAnalysis<MachineDominatorTree>();
220   ORE = &getAnalysis<MachineOptimizationRemarkEmitterPass>().getORE();
221   TII = MF->getSubtarget().getInstrInfo();
222   RegClassInfo.runOnMachineFunction(*MF);
223 
224   for (auto &L : *MLI)
225     scheduleLoop(*L);
226 
227   return false;
228 }
229 
230 /// Attempt to perform the SMS algorithm on the specified loop. This function is
231 /// the main entry point for the algorithm.  The function identifies candidate
232 /// loops, calculates the minimum initiation interval, and attempts to schedule
233 /// the loop.
234 bool MachinePipeliner::scheduleLoop(MachineLoop &L) {
235   bool Changed = false;
236   for (auto &InnerLoop : L)
237     Changed |= scheduleLoop(*InnerLoop);
238 
239 #ifndef NDEBUG
240   // Stop trying after reaching the limit (if any).
241   int Limit = SwpLoopLimit;
242   if (Limit >= 0) {
243     if (NumTries >= SwpLoopLimit)
244       return Changed;
245     NumTries++;
246   }
247 #endif
248 
249   setPragmaPipelineOptions(L);
250   if (!canPipelineLoop(L)) {
251     LLVM_DEBUG(dbgs() << "\n!!! Can not pipeline loop.\n");
252     ORE->emit([&]() {
253       return MachineOptimizationRemarkMissed(DEBUG_TYPE, "canPipelineLoop",
254                                              L.getStartLoc(), L.getHeader())
255              << "Failed to pipeline loop";
256     });
257 
258     return Changed;
259   }
260 
261   ++NumTrytoPipeline;
262 
263   Changed = swingModuloScheduler(L);
264 
265   return Changed;
266 }
267 
268 void MachinePipeliner::setPragmaPipelineOptions(MachineLoop &L) {
269   // Reset the pragma for the next loop in iteration.
270   disabledByPragma = false;
271 
272   MachineBasicBlock *LBLK = L.getTopBlock();
273 
274   if (LBLK == nullptr)
275     return;
276 
277   const BasicBlock *BBLK = LBLK->getBasicBlock();
278   if (BBLK == nullptr)
279     return;
280 
281   const Instruction *TI = BBLK->getTerminator();
282   if (TI == nullptr)
283     return;
284 
285   MDNode *LoopID = TI->getMetadata(LLVMContext::MD_loop);
286   if (LoopID == nullptr)
287     return;
288 
289   assert(LoopID->getNumOperands() > 0 && "requires atleast one operand");
290   assert(LoopID->getOperand(0) == LoopID && "invalid loop");
291 
292   for (unsigned i = 1, e = LoopID->getNumOperands(); i < e; ++i) {
293     MDNode *MD = dyn_cast<MDNode>(LoopID->getOperand(i));
294 
295     if (MD == nullptr)
296       continue;
297 
298     MDString *S = dyn_cast<MDString>(MD->getOperand(0));
299 
300     if (S == nullptr)
301       continue;
302 
303     if (S->getString() == "llvm.loop.pipeline.initiationinterval") {
304       assert(MD->getNumOperands() == 2 &&
305              "Pipeline initiation interval hint metadata should have two operands.");
306       II_setByPragma =
307           mdconst::extract<ConstantInt>(MD->getOperand(1))->getZExtValue();
308       assert(II_setByPragma >= 1 && "Pipeline initiation interval must be positive.");
309     } else if (S->getString() == "llvm.loop.pipeline.disable") {
310       disabledByPragma = true;
311     }
312   }
313 }
314 
315 /// Return true if the loop can be software pipelined.  The algorithm is
316 /// restricted to loops with a single basic block.  Make sure that the
317 /// branch in the loop can be analyzed.
318 bool MachinePipeliner::canPipelineLoop(MachineLoop &L) {
319   if (L.getNumBlocks() != 1) {
320     ORE->emit([&]() {
321       return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop",
322                                                L.getStartLoc(), L.getHeader())
323              << "Not a single basic block: "
324              << ore::NV("NumBlocks", L.getNumBlocks());
325     });
326     return false;
327   }
328 
329   if (disabledByPragma) {
330     ORE->emit([&]() {
331       return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop",
332                                                L.getStartLoc(), L.getHeader())
333              << "Disabled by Pragma.";
334     });
335     return false;
336   }
337 
338   // Check if the branch can't be understood because we can't do pipelining
339   // if that's the case.
340   LI.TBB = nullptr;
341   LI.FBB = nullptr;
342   LI.BrCond.clear();
343   if (TII->analyzeBranch(*L.getHeader(), LI.TBB, LI.FBB, LI.BrCond)) {
344     LLVM_DEBUG(dbgs() << "Unable to analyzeBranch, can NOT pipeline Loop\n");
345     NumFailBranch++;
346     ORE->emit([&]() {
347       return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop",
348                                                L.getStartLoc(), L.getHeader())
349              << "The branch can't be understood";
350     });
351     return false;
352   }
353 
354   LI.LoopInductionVar = nullptr;
355   LI.LoopCompare = nullptr;
356   if (!TII->analyzeLoopForPipelining(L.getTopBlock())) {
357     LLVM_DEBUG(dbgs() << "Unable to analyzeLoop, can NOT pipeline Loop\n");
358     NumFailLoop++;
359     ORE->emit([&]() {
360       return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop",
361                                                L.getStartLoc(), L.getHeader())
362              << "The loop structure is not supported";
363     });
364     return false;
365   }
366 
367   if (!L.getLoopPreheader()) {
368     LLVM_DEBUG(dbgs() << "Preheader not found, can NOT pipeline Loop\n");
369     NumFailPreheader++;
370     ORE->emit([&]() {
371       return MachineOptimizationRemarkAnalysis(DEBUG_TYPE, "canPipelineLoop",
372                                                L.getStartLoc(), L.getHeader())
373              << "No loop preheader found";
374     });
375     return false;
376   }
377 
378   // Remove any subregisters from inputs to phi nodes.
379   preprocessPhiNodes(*L.getHeader());
380   return true;
381 }
382 
383 void MachinePipeliner::preprocessPhiNodes(MachineBasicBlock &B) {
384   MachineRegisterInfo &MRI = MF->getRegInfo();
385   SlotIndexes &Slots = *getAnalysis<LiveIntervals>().getSlotIndexes();
386 
387   for (MachineInstr &PI : make_range(B.begin(), B.getFirstNonPHI())) {
388     MachineOperand &DefOp = PI.getOperand(0);
389     assert(DefOp.getSubReg() == 0);
390     auto *RC = MRI.getRegClass(DefOp.getReg());
391 
392     for (unsigned i = 1, n = PI.getNumOperands(); i != n; i += 2) {
393       MachineOperand &RegOp = PI.getOperand(i);
394       if (RegOp.getSubReg() == 0)
395         continue;
396 
397       // If the operand uses a subregister, replace it with a new register
398       // without subregisters, and generate a copy to the new register.
399       Register NewReg = MRI.createVirtualRegister(RC);
400       MachineBasicBlock &PredB = *PI.getOperand(i+1).getMBB();
401       MachineBasicBlock::iterator At = PredB.getFirstTerminator();
402       const DebugLoc &DL = PredB.findDebugLoc(At);
403       auto Copy = BuildMI(PredB, At, DL, TII->get(TargetOpcode::COPY), NewReg)
404                     .addReg(RegOp.getReg(), getRegState(RegOp),
405                             RegOp.getSubReg());
406       Slots.insertMachineInstrInMaps(*Copy);
407       RegOp.setReg(NewReg);
408       RegOp.setSubReg(0);
409     }
410   }
411 }
412 
413 /// The SMS algorithm consists of the following main steps:
414 /// 1. Computation and analysis of the dependence graph.
415 /// 2. Ordering of the nodes (instructions).
416 /// 3. Attempt to Schedule the loop.
417 bool MachinePipeliner::swingModuloScheduler(MachineLoop &L) {
418   assert(L.getBlocks().size() == 1 && "SMS works on single blocks only.");
419 
420   SwingSchedulerDAG SMS(*this, L, getAnalysis<LiveIntervals>(), RegClassInfo,
421                         II_setByPragma);
422 
423   MachineBasicBlock *MBB = L.getHeader();
424   // The kernel should not include any terminator instructions.  These
425   // will be added back later.
426   SMS.startBlock(MBB);
427 
428   // Compute the number of 'real' instructions in the basic block by
429   // ignoring terminators.
430   unsigned size = MBB->size();
431   for (MachineBasicBlock::iterator I = MBB->getFirstTerminator(),
432                                    E = MBB->instr_end();
433        I != E; ++I, --size)
434     ;
435 
436   SMS.enterRegion(MBB, MBB->begin(), MBB->getFirstTerminator(), size);
437   SMS.schedule();
438   SMS.exitRegion();
439 
440   SMS.finishBlock();
441   return SMS.hasNewSchedule();
442 }
443 
444 void SwingSchedulerDAG::setMII(unsigned ResMII, unsigned RecMII) {
445   if (II_setByPragma > 0)
446     MII = II_setByPragma;
447   else
448     MII = std::max(ResMII, RecMII);
449 }
450 
451 void SwingSchedulerDAG::setMAX_II() {
452   if (II_setByPragma > 0)
453     MAX_II = II_setByPragma;
454   else
455     MAX_II = MII + 10;
456 }
457 
458 /// We override the schedule function in ScheduleDAGInstrs to implement the
459 /// scheduling part of the Swing Modulo Scheduling algorithm.
460 void SwingSchedulerDAG::schedule() {
461   AliasAnalysis *AA = &Pass.getAnalysis<AAResultsWrapperPass>().getAAResults();
462   buildSchedGraph(AA);
463   addLoopCarriedDependences(AA);
464   updatePhiDependences();
465   Topo.InitDAGTopologicalSorting();
466   changeDependences();
467   postprocessDAG();
468   LLVM_DEBUG(dump());
469 
470   NodeSetType NodeSets;
471   findCircuits(NodeSets);
472   NodeSetType Circuits = NodeSets;
473 
474   // Calculate the MII.
475   unsigned ResMII = calculateResMII();
476   unsigned RecMII = calculateRecMII(NodeSets);
477 
478   fuseRecs(NodeSets);
479 
480   // This flag is used for testing and can cause correctness problems.
481   if (SwpIgnoreRecMII)
482     RecMII = 0;
483 
484   setMII(ResMII, RecMII);
485   setMAX_II();
486 
487   LLVM_DEBUG(dbgs() << "MII = " << MII << " MAX_II = " << MAX_II
488                     << " (rec=" << RecMII << ", res=" << ResMII << ")\n");
489 
490   // Can't schedule a loop without a valid MII.
491   if (MII == 0) {
492     LLVM_DEBUG(dbgs() << "Invalid Minimal Initiation Interval: 0\n");
493     NumFailZeroMII++;
494     Pass.ORE->emit([&]() {
495       return MachineOptimizationRemarkAnalysis(
496                  DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader())
497              << "Invalid Minimal Initiation Interval: 0";
498     });
499     return;
500   }
501 
502   // Don't pipeline large loops.
503   if (SwpMaxMii != -1 && (int)MII > SwpMaxMii) {
504     LLVM_DEBUG(dbgs() << "MII > " << SwpMaxMii
505                       << ", we don't pipleline large loops\n");
506     NumFailLargeMaxMII++;
507     Pass.ORE->emit([&]() {
508       return MachineOptimizationRemarkAnalysis(
509                  DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader())
510              << "Minimal Initiation Interval too large: "
511              << ore::NV("MII", (int)MII) << " > "
512              << ore::NV("SwpMaxMii", SwpMaxMii) << "."
513              << "Refer to -pipeliner-max-mii.";
514     });
515     return;
516   }
517 
518   computeNodeFunctions(NodeSets);
519 
520   registerPressureFilter(NodeSets);
521 
522   colocateNodeSets(NodeSets);
523 
524   checkNodeSets(NodeSets);
525 
526   LLVM_DEBUG({
527     for (auto &I : NodeSets) {
528       dbgs() << "  Rec NodeSet ";
529       I.dump();
530     }
531   });
532 
533   llvm::stable_sort(NodeSets, std::greater<NodeSet>());
534 
535   groupRemainingNodes(NodeSets);
536 
537   removeDuplicateNodes(NodeSets);
538 
539   LLVM_DEBUG({
540     for (auto &I : NodeSets) {
541       dbgs() << "  NodeSet ";
542       I.dump();
543     }
544   });
545 
546   computeNodeOrder(NodeSets);
547 
548   // check for node order issues
549   checkValidNodeOrder(Circuits);
550 
551   SMSchedule Schedule(Pass.MF);
552   Scheduled = schedulePipeline(Schedule);
553 
554   if (!Scheduled){
555     LLVM_DEBUG(dbgs() << "No schedule found, return\n");
556     NumFailNoSchedule++;
557     Pass.ORE->emit([&]() {
558       return MachineOptimizationRemarkAnalysis(
559                  DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader())
560              << "Unable to find schedule";
561     });
562     return;
563   }
564 
565   unsigned numStages = Schedule.getMaxStageCount();
566   // No need to generate pipeline if there are no overlapped iterations.
567   if (numStages == 0) {
568     LLVM_DEBUG(dbgs() << "No overlapped iterations, skip.\n");
569     NumFailZeroStage++;
570     Pass.ORE->emit([&]() {
571       return MachineOptimizationRemarkAnalysis(
572                  DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader())
573              << "No need to pipeline - no overlapped iterations in schedule.";
574     });
575     return;
576   }
577   // Check that the maximum stage count is less than user-defined limit.
578   if (SwpMaxStages > -1 && (int)numStages > SwpMaxStages) {
579     LLVM_DEBUG(dbgs() << "numStages:" << numStages << ">" << SwpMaxStages
580                       << " : too many stages, abort\n");
581     NumFailLargeMaxStage++;
582     Pass.ORE->emit([&]() {
583       return MachineOptimizationRemarkAnalysis(
584                  DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader())
585              << "Too many stages in schedule: "
586              << ore::NV("numStages", (int)numStages) << " > "
587              << ore::NV("SwpMaxStages", SwpMaxStages)
588              << ". Refer to -pipeliner-max-stages.";
589     });
590     return;
591   }
592 
593   Pass.ORE->emit([&]() {
594     return MachineOptimizationRemark(DEBUG_TYPE, "schedule", Loop.getStartLoc(),
595                                      Loop.getHeader())
596            << "Pipelined succesfully!";
597   });
598 
599   // Generate the schedule as a ModuloSchedule.
600   DenseMap<MachineInstr *, int> Cycles, Stages;
601   std::vector<MachineInstr *> OrderedInsts;
602   for (int Cycle = Schedule.getFirstCycle(); Cycle <= Schedule.getFinalCycle();
603        ++Cycle) {
604     for (SUnit *SU : Schedule.getInstructions(Cycle)) {
605       OrderedInsts.push_back(SU->getInstr());
606       Cycles[SU->getInstr()] = Cycle;
607       Stages[SU->getInstr()] = Schedule.stageScheduled(SU);
608     }
609   }
610   DenseMap<MachineInstr *, std::pair<unsigned, int64_t>> NewInstrChanges;
611   for (auto &KV : NewMIs) {
612     Cycles[KV.first] = Cycles[KV.second];
613     Stages[KV.first] = Stages[KV.second];
614     NewInstrChanges[KV.first] = InstrChanges[getSUnit(KV.first)];
615   }
616 
617   ModuloSchedule MS(MF, &Loop, std::move(OrderedInsts), std::move(Cycles),
618                     std::move(Stages));
619   if (EmitTestAnnotations) {
620     assert(NewInstrChanges.empty() &&
621            "Cannot serialize a schedule with InstrChanges!");
622     ModuloScheduleTestAnnotater MSTI(MF, MS);
623     MSTI.annotate();
624     return;
625   }
626   // The experimental code generator can't work if there are InstChanges.
627   if (ExperimentalCodeGen && NewInstrChanges.empty()) {
628     PeelingModuloScheduleExpander MSE(MF, MS, &LIS);
629     MSE.expand();
630   } else {
631     ModuloScheduleExpander MSE(MF, MS, LIS, std::move(NewInstrChanges));
632     MSE.expand();
633     MSE.cleanup();
634   }
635   ++NumPipelined;
636 }
637 
638 /// Clean up after the software pipeliner runs.
639 void SwingSchedulerDAG::finishBlock() {
640   for (auto &KV : NewMIs)
641     MF.DeleteMachineInstr(KV.second);
642   NewMIs.clear();
643 
644   // Call the superclass.
645   ScheduleDAGInstrs::finishBlock();
646 }
647 
648 /// Return the register values for  the operands of a Phi instruction.
649 /// This function assume the instruction is a Phi.
650 static void getPhiRegs(MachineInstr &Phi, MachineBasicBlock *Loop,
651                        unsigned &InitVal, unsigned &LoopVal) {
652   assert(Phi.isPHI() && "Expecting a Phi.");
653 
654   InitVal = 0;
655   LoopVal = 0;
656   for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2)
657     if (Phi.getOperand(i + 1).getMBB() != Loop)
658       InitVal = Phi.getOperand(i).getReg();
659     else
660       LoopVal = Phi.getOperand(i).getReg();
661 
662   assert(InitVal != 0 && LoopVal != 0 && "Unexpected Phi structure.");
663 }
664 
665 /// Return the Phi register value that comes the loop block.
666 static unsigned getLoopPhiReg(MachineInstr &Phi, MachineBasicBlock *LoopBB) {
667   for (unsigned i = 1, e = Phi.getNumOperands(); i != e; i += 2)
668     if (Phi.getOperand(i + 1).getMBB() == LoopBB)
669       return Phi.getOperand(i).getReg();
670   return 0;
671 }
672 
673 /// Return true if SUb can be reached from SUa following the chain edges.
674 static bool isSuccOrder(SUnit *SUa, SUnit *SUb) {
675   SmallPtrSet<SUnit *, 8> Visited;
676   SmallVector<SUnit *, 8> Worklist;
677   Worklist.push_back(SUa);
678   while (!Worklist.empty()) {
679     const SUnit *SU = Worklist.pop_back_val();
680     for (auto &SI : SU->Succs) {
681       SUnit *SuccSU = SI.getSUnit();
682       if (SI.getKind() == SDep::Order) {
683         if (Visited.count(SuccSU))
684           continue;
685         if (SuccSU == SUb)
686           return true;
687         Worklist.push_back(SuccSU);
688         Visited.insert(SuccSU);
689       }
690     }
691   }
692   return false;
693 }
694 
695 /// Return true if the instruction causes a chain between memory
696 /// references before and after it.
697 static bool isDependenceBarrier(MachineInstr &MI, AliasAnalysis *AA) {
698   return MI.isCall() || MI.mayRaiseFPException() ||
699          MI.hasUnmodeledSideEffects() ||
700          (MI.hasOrderedMemoryRef() &&
701           (!MI.mayLoad() || !MI.isDereferenceableInvariantLoad(AA)));
702 }
703 
704 /// Return the underlying objects for the memory references of an instruction.
705 /// This function calls the code in ValueTracking, but first checks that the
706 /// instruction has a memory operand.
707 static void getUnderlyingObjects(const MachineInstr *MI,
708                                  SmallVectorImpl<const Value *> &Objs,
709                                  const DataLayout &DL) {
710   if (!MI->hasOneMemOperand())
711     return;
712   MachineMemOperand *MM = *MI->memoperands_begin();
713   if (!MM->getValue())
714     return;
715   GetUnderlyingObjects(MM->getValue(), Objs, DL);
716   for (const Value *V : Objs) {
717     if (!isIdentifiedObject(V)) {
718       Objs.clear();
719       return;
720     }
721     Objs.push_back(V);
722   }
723 }
724 
725 /// Add a chain edge between a load and store if the store can be an
726 /// alias of the load on a subsequent iteration, i.e., a loop carried
727 /// dependence. This code is very similar to the code in ScheduleDAGInstrs
728 /// but that code doesn't create loop carried dependences.
729 void SwingSchedulerDAG::addLoopCarriedDependences(AliasAnalysis *AA) {
730   MapVector<const Value *, SmallVector<SUnit *, 4>> PendingLoads;
731   Value *UnknownValue =
732     UndefValue::get(Type::getVoidTy(MF.getFunction().getContext()));
733   for (auto &SU : SUnits) {
734     MachineInstr &MI = *SU.getInstr();
735     if (isDependenceBarrier(MI, AA))
736       PendingLoads.clear();
737     else if (MI.mayLoad()) {
738       SmallVector<const Value *, 4> Objs;
739       getUnderlyingObjects(&MI, Objs, MF.getDataLayout());
740       if (Objs.empty())
741         Objs.push_back(UnknownValue);
742       for (auto V : Objs) {
743         SmallVector<SUnit *, 4> &SUs = PendingLoads[V];
744         SUs.push_back(&SU);
745       }
746     } else if (MI.mayStore()) {
747       SmallVector<const Value *, 4> Objs;
748       getUnderlyingObjects(&MI, Objs, MF.getDataLayout());
749       if (Objs.empty())
750         Objs.push_back(UnknownValue);
751       for (auto V : Objs) {
752         MapVector<const Value *, SmallVector<SUnit *, 4>>::iterator I =
753             PendingLoads.find(V);
754         if (I == PendingLoads.end())
755           continue;
756         for (auto Load : I->second) {
757           if (isSuccOrder(Load, &SU))
758             continue;
759           MachineInstr &LdMI = *Load->getInstr();
760           // First, perform the cheaper check that compares the base register.
761           // If they are the same and the load offset is less than the store
762           // offset, then mark the dependence as loop carried potentially.
763           const MachineOperand *BaseOp1, *BaseOp2;
764           int64_t Offset1, Offset2;
765           bool Offset1IsScalable, Offset2IsScalable;
766           if (TII->getMemOperandWithOffset(LdMI, BaseOp1, Offset1,
767                                            Offset1IsScalable, TRI) &&
768               TII->getMemOperandWithOffset(MI, BaseOp2, Offset2,
769                                            Offset2IsScalable, TRI)) {
770             if (BaseOp1->isIdenticalTo(*BaseOp2) &&
771                 Offset1IsScalable == Offset2IsScalable &&
772                 (int)Offset1 < (int)Offset2) {
773               assert(TII->areMemAccessesTriviallyDisjoint(LdMI, MI) &&
774                      "What happened to the chain edge?");
775               SDep Dep(Load, SDep::Barrier);
776               Dep.setLatency(1);
777               SU.addPred(Dep);
778               continue;
779             }
780           }
781           // Second, the more expensive check that uses alias analysis on the
782           // base registers. If they alias, and the load offset is less than
783           // the store offset, the mark the dependence as loop carried.
784           if (!AA) {
785             SDep Dep(Load, SDep::Barrier);
786             Dep.setLatency(1);
787             SU.addPred(Dep);
788             continue;
789           }
790           MachineMemOperand *MMO1 = *LdMI.memoperands_begin();
791           MachineMemOperand *MMO2 = *MI.memoperands_begin();
792           if (!MMO1->getValue() || !MMO2->getValue()) {
793             SDep Dep(Load, SDep::Barrier);
794             Dep.setLatency(1);
795             SU.addPred(Dep);
796             continue;
797           }
798           if (MMO1->getValue() == MMO2->getValue() &&
799               MMO1->getOffset() <= MMO2->getOffset()) {
800             SDep Dep(Load, SDep::Barrier);
801             Dep.setLatency(1);
802             SU.addPred(Dep);
803             continue;
804           }
805           AliasResult AAResult = AA->alias(
806               MemoryLocation(MMO1->getValue(), LocationSize::unknown(),
807                              MMO1->getAAInfo()),
808               MemoryLocation(MMO2->getValue(), LocationSize::unknown(),
809                              MMO2->getAAInfo()));
810 
811           if (AAResult != NoAlias) {
812             SDep Dep(Load, SDep::Barrier);
813             Dep.setLatency(1);
814             SU.addPred(Dep);
815           }
816         }
817       }
818     }
819   }
820 }
821 
822 /// Update the phi dependences to the DAG because ScheduleDAGInstrs no longer
823 /// processes dependences for PHIs. This function adds true dependences
824 /// from a PHI to a use, and a loop carried dependence from the use to the
825 /// PHI. The loop carried dependence is represented as an anti dependence
826 /// edge. This function also removes chain dependences between unrelated
827 /// PHIs.
828 void SwingSchedulerDAG::updatePhiDependences() {
829   SmallVector<SDep, 4> RemoveDeps;
830   const TargetSubtargetInfo &ST = MF.getSubtarget<TargetSubtargetInfo>();
831 
832   // Iterate over each DAG node.
833   for (SUnit &I : SUnits) {
834     RemoveDeps.clear();
835     // Set to true if the instruction has an operand defined by a Phi.
836     unsigned HasPhiUse = 0;
837     unsigned HasPhiDef = 0;
838     MachineInstr *MI = I.getInstr();
839     // Iterate over each operand, and we process the definitions.
840     for (MachineInstr::mop_iterator MOI = MI->operands_begin(),
841                                     MOE = MI->operands_end();
842          MOI != MOE; ++MOI) {
843       if (!MOI->isReg())
844         continue;
845       Register Reg = MOI->getReg();
846       if (MOI->isDef()) {
847         // If the register is used by a Phi, then create an anti dependence.
848         for (MachineRegisterInfo::use_instr_iterator
849                  UI = MRI.use_instr_begin(Reg),
850                  UE = MRI.use_instr_end();
851              UI != UE; ++UI) {
852           MachineInstr *UseMI = &*UI;
853           SUnit *SU = getSUnit(UseMI);
854           if (SU != nullptr && UseMI->isPHI()) {
855             if (!MI->isPHI()) {
856               SDep Dep(SU, SDep::Anti, Reg);
857               Dep.setLatency(1);
858               I.addPred(Dep);
859             } else {
860               HasPhiDef = Reg;
861               // Add a chain edge to a dependent Phi that isn't an existing
862               // predecessor.
863               if (SU->NodeNum < I.NodeNum && !I.isPred(SU))
864                 I.addPred(SDep(SU, SDep::Barrier));
865             }
866           }
867         }
868       } else if (MOI->isUse()) {
869         // If the register is defined by a Phi, then create a true dependence.
870         MachineInstr *DefMI = MRI.getUniqueVRegDef(Reg);
871         if (DefMI == nullptr)
872           continue;
873         SUnit *SU = getSUnit(DefMI);
874         if (SU != nullptr && DefMI->isPHI()) {
875           if (!MI->isPHI()) {
876             SDep Dep(SU, SDep::Data, Reg);
877             Dep.setLatency(0);
878             ST.adjustSchedDependency(SU, 0, &I, MI->getOperandNo(MOI), Dep);
879             I.addPred(Dep);
880           } else {
881             HasPhiUse = Reg;
882             // Add a chain edge to a dependent Phi that isn't an existing
883             // predecessor.
884             if (SU->NodeNum < I.NodeNum && !I.isPred(SU))
885               I.addPred(SDep(SU, SDep::Barrier));
886           }
887         }
888       }
889     }
890     // Remove order dependences from an unrelated Phi.
891     if (!SwpPruneDeps)
892       continue;
893     for (auto &PI : I.Preds) {
894       MachineInstr *PMI = PI.getSUnit()->getInstr();
895       if (PMI->isPHI() && PI.getKind() == SDep::Order) {
896         if (I.getInstr()->isPHI()) {
897           if (PMI->getOperand(0).getReg() == HasPhiUse)
898             continue;
899           if (getLoopPhiReg(*PMI, PMI->getParent()) == HasPhiDef)
900             continue;
901         }
902         RemoveDeps.push_back(PI);
903       }
904     }
905     for (int i = 0, e = RemoveDeps.size(); i != e; ++i)
906       I.removePred(RemoveDeps[i]);
907   }
908 }
909 
910 /// Iterate over each DAG node and see if we can change any dependences
911 /// in order to reduce the recurrence MII.
912 void SwingSchedulerDAG::changeDependences() {
913   // See if an instruction can use a value from the previous iteration.
914   // If so, we update the base and offset of the instruction and change
915   // the dependences.
916   for (SUnit &I : SUnits) {
917     unsigned BasePos = 0, OffsetPos = 0, NewBase = 0;
918     int64_t NewOffset = 0;
919     if (!canUseLastOffsetValue(I.getInstr(), BasePos, OffsetPos, NewBase,
920                                NewOffset))
921       continue;
922 
923     // Get the MI and SUnit for the instruction that defines the original base.
924     Register OrigBase = I.getInstr()->getOperand(BasePos).getReg();
925     MachineInstr *DefMI = MRI.getUniqueVRegDef(OrigBase);
926     if (!DefMI)
927       continue;
928     SUnit *DefSU = getSUnit(DefMI);
929     if (!DefSU)
930       continue;
931     // Get the MI and SUnit for the instruction that defins the new base.
932     MachineInstr *LastMI = MRI.getUniqueVRegDef(NewBase);
933     if (!LastMI)
934       continue;
935     SUnit *LastSU = getSUnit(LastMI);
936     if (!LastSU)
937       continue;
938 
939     if (Topo.IsReachable(&I, LastSU))
940       continue;
941 
942     // Remove the dependence. The value now depends on a prior iteration.
943     SmallVector<SDep, 4> Deps;
944     for (SUnit::pred_iterator P = I.Preds.begin(), E = I.Preds.end(); P != E;
945          ++P)
946       if (P->getSUnit() == DefSU)
947         Deps.push_back(*P);
948     for (int i = 0, e = Deps.size(); i != e; i++) {
949       Topo.RemovePred(&I, Deps[i].getSUnit());
950       I.removePred(Deps[i]);
951     }
952     // Remove the chain dependence between the instructions.
953     Deps.clear();
954     for (auto &P : LastSU->Preds)
955       if (P.getSUnit() == &I && P.getKind() == SDep::Order)
956         Deps.push_back(P);
957     for (int i = 0, e = Deps.size(); i != e; i++) {
958       Topo.RemovePred(LastSU, Deps[i].getSUnit());
959       LastSU->removePred(Deps[i]);
960     }
961 
962     // Add a dependence between the new instruction and the instruction
963     // that defines the new base.
964     SDep Dep(&I, SDep::Anti, NewBase);
965     Topo.AddPred(LastSU, &I);
966     LastSU->addPred(Dep);
967 
968     // Remember the base and offset information so that we can update the
969     // instruction during code generation.
970     InstrChanges[&I] = std::make_pair(NewBase, NewOffset);
971   }
972 }
973 
974 namespace {
975 
976 // FuncUnitSorter - Comparison operator used to sort instructions by
977 // the number of functional unit choices.
978 struct FuncUnitSorter {
979   const InstrItineraryData *InstrItins;
980   const MCSubtargetInfo *STI;
981   DenseMap<InstrStage::FuncUnits, unsigned> Resources;
982 
983   FuncUnitSorter(const TargetSubtargetInfo &TSI)
984       : InstrItins(TSI.getInstrItineraryData()), STI(&TSI) {}
985 
986   // Compute the number of functional unit alternatives needed
987   // at each stage, and take the minimum value. We prioritize the
988   // instructions by the least number of choices first.
989   unsigned minFuncUnits(const MachineInstr *Inst,
990                         InstrStage::FuncUnits &F) const {
991     unsigned SchedClass = Inst->getDesc().getSchedClass();
992     unsigned min = UINT_MAX;
993     if (InstrItins && !InstrItins->isEmpty()) {
994       for (const InstrStage &IS :
995            make_range(InstrItins->beginStage(SchedClass),
996                       InstrItins->endStage(SchedClass))) {
997         InstrStage::FuncUnits funcUnits = IS.getUnits();
998         unsigned numAlternatives = countPopulation(funcUnits);
999         if (numAlternatives < min) {
1000           min = numAlternatives;
1001           F = funcUnits;
1002         }
1003       }
1004       return min;
1005     }
1006     if (STI && STI->getSchedModel().hasInstrSchedModel()) {
1007       const MCSchedClassDesc *SCDesc =
1008           STI->getSchedModel().getSchedClassDesc(SchedClass);
1009       if (!SCDesc->isValid())
1010         // No valid Schedule Class Desc for schedClass, should be
1011         // Pseudo/PostRAPseudo
1012         return min;
1013 
1014       for (const MCWriteProcResEntry &PRE :
1015            make_range(STI->getWriteProcResBegin(SCDesc),
1016                       STI->getWriteProcResEnd(SCDesc))) {
1017         if (!PRE.Cycles)
1018           continue;
1019         const MCProcResourceDesc *ProcResource =
1020             STI->getSchedModel().getProcResource(PRE.ProcResourceIdx);
1021         unsigned NumUnits = ProcResource->NumUnits;
1022         if (NumUnits < min) {
1023           min = NumUnits;
1024           F = PRE.ProcResourceIdx;
1025         }
1026       }
1027       return min;
1028     }
1029     llvm_unreachable("Should have non-empty InstrItins or hasInstrSchedModel!");
1030   }
1031 
1032   // Compute the critical resources needed by the instruction. This
1033   // function records the functional units needed by instructions that
1034   // must use only one functional unit. We use this as a tie breaker
1035   // for computing the resource MII. The instrutions that require
1036   // the same, highly used, functional unit have high priority.
1037   void calcCriticalResources(MachineInstr &MI) {
1038     unsigned SchedClass = MI.getDesc().getSchedClass();
1039     if (InstrItins && !InstrItins->isEmpty()) {
1040       for (const InstrStage &IS :
1041            make_range(InstrItins->beginStage(SchedClass),
1042                       InstrItins->endStage(SchedClass))) {
1043         InstrStage::FuncUnits FuncUnits = IS.getUnits();
1044         if (countPopulation(FuncUnits) == 1)
1045           Resources[FuncUnits]++;
1046       }
1047       return;
1048     }
1049     if (STI && STI->getSchedModel().hasInstrSchedModel()) {
1050       const MCSchedClassDesc *SCDesc =
1051           STI->getSchedModel().getSchedClassDesc(SchedClass);
1052       if (!SCDesc->isValid())
1053         // No valid Schedule Class Desc for schedClass, should be
1054         // Pseudo/PostRAPseudo
1055         return;
1056 
1057       for (const MCWriteProcResEntry &PRE :
1058            make_range(STI->getWriteProcResBegin(SCDesc),
1059                       STI->getWriteProcResEnd(SCDesc))) {
1060         if (!PRE.Cycles)
1061           continue;
1062         Resources[PRE.ProcResourceIdx]++;
1063       }
1064       return;
1065     }
1066     llvm_unreachable("Should have non-empty InstrItins or hasInstrSchedModel!");
1067   }
1068 
1069   /// Return true if IS1 has less priority than IS2.
1070   bool operator()(const MachineInstr *IS1, const MachineInstr *IS2) const {
1071     InstrStage::FuncUnits F1 = 0, F2 = 0;
1072     unsigned MFUs1 = minFuncUnits(IS1, F1);
1073     unsigned MFUs2 = minFuncUnits(IS2, F2);
1074     if (MFUs1 == MFUs2)
1075       return Resources.lookup(F1) < Resources.lookup(F2);
1076     return MFUs1 > MFUs2;
1077   }
1078 };
1079 
1080 } // end anonymous namespace
1081 
1082 /// Calculate the resource constrained minimum initiation interval for the
1083 /// specified loop. We use the DFA to model the resources needed for
1084 /// each instruction, and we ignore dependences. A different DFA is created
1085 /// for each cycle that is required. When adding a new instruction, we attempt
1086 /// to add it to each existing DFA, until a legal space is found. If the
1087 /// instruction cannot be reserved in an existing DFA, we create a new one.
1088 unsigned SwingSchedulerDAG::calculateResMII() {
1089 
1090   LLVM_DEBUG(dbgs() << "calculateResMII:\n");
1091   SmallVector<ResourceManager*, 8> Resources;
1092   MachineBasicBlock *MBB = Loop.getHeader();
1093   Resources.push_back(new ResourceManager(&MF.getSubtarget()));
1094 
1095   // Sort the instructions by the number of available choices for scheduling,
1096   // least to most. Use the number of critical resources as the tie breaker.
1097   FuncUnitSorter FUS = FuncUnitSorter(MF.getSubtarget());
1098   for (MachineBasicBlock::iterator I = MBB->getFirstNonPHI(),
1099                                    E = MBB->getFirstTerminator();
1100        I != E; ++I)
1101     FUS.calcCriticalResources(*I);
1102   PriorityQueue<MachineInstr *, std::vector<MachineInstr *>, FuncUnitSorter>
1103       FuncUnitOrder(FUS);
1104 
1105   for (MachineBasicBlock::iterator I = MBB->getFirstNonPHI(),
1106                                    E = MBB->getFirstTerminator();
1107        I != E; ++I)
1108     FuncUnitOrder.push(&*I);
1109 
1110   while (!FuncUnitOrder.empty()) {
1111     MachineInstr *MI = FuncUnitOrder.top();
1112     FuncUnitOrder.pop();
1113     if (TII->isZeroCost(MI->getOpcode()))
1114       continue;
1115     // Attempt to reserve the instruction in an existing DFA. At least one
1116     // DFA is needed for each cycle.
1117     unsigned NumCycles = getSUnit(MI)->Latency;
1118     unsigned ReservedCycles = 0;
1119     SmallVectorImpl<ResourceManager *>::iterator RI = Resources.begin();
1120     SmallVectorImpl<ResourceManager *>::iterator RE = Resources.end();
1121     LLVM_DEBUG({
1122       dbgs() << "Trying to reserve resource for " << NumCycles
1123              << " cycles for \n";
1124       MI->dump();
1125     });
1126     for (unsigned C = 0; C < NumCycles; ++C)
1127       while (RI != RE) {
1128         if ((*RI)->canReserveResources(*MI)) {
1129           (*RI)->reserveResources(*MI);
1130           ++ReservedCycles;
1131           break;
1132         }
1133         RI++;
1134       }
1135     LLVM_DEBUG(dbgs() << "ReservedCycles:" << ReservedCycles
1136                       << ", NumCycles:" << NumCycles << "\n");
1137     // Add new DFAs, if needed, to reserve resources.
1138     for (unsigned C = ReservedCycles; C < NumCycles; ++C) {
1139       LLVM_DEBUG(if (SwpDebugResource) dbgs()
1140                  << "NewResource created to reserve resources"
1141                  << "\n");
1142       ResourceManager *NewResource = new ResourceManager(&MF.getSubtarget());
1143       assert(NewResource->canReserveResources(*MI) && "Reserve error.");
1144       NewResource->reserveResources(*MI);
1145       Resources.push_back(NewResource);
1146     }
1147   }
1148   int Resmii = Resources.size();
1149   LLVM_DEBUG(dbgs() << "Return Res MII:" << Resmii << "\n");
1150   // Delete the memory for each of the DFAs that were created earlier.
1151   for (ResourceManager *RI : Resources) {
1152     ResourceManager *D = RI;
1153     delete D;
1154   }
1155   Resources.clear();
1156   return Resmii;
1157 }
1158 
1159 /// Calculate the recurrence-constrainted minimum initiation interval.
1160 /// Iterate over each circuit.  Compute the delay(c) and distance(c)
1161 /// for each circuit. The II needs to satisfy the inequality
1162 /// delay(c) - II*distance(c) <= 0. For each circuit, choose the smallest
1163 /// II that satisfies the inequality, and the RecMII is the maximum
1164 /// of those values.
1165 unsigned SwingSchedulerDAG::calculateRecMII(NodeSetType &NodeSets) {
1166   unsigned RecMII = 0;
1167 
1168   for (NodeSet &Nodes : NodeSets) {
1169     if (Nodes.empty())
1170       continue;
1171 
1172     unsigned Delay = Nodes.getLatency();
1173     unsigned Distance = 1;
1174 
1175     // ii = ceil(delay / distance)
1176     unsigned CurMII = (Delay + Distance - 1) / Distance;
1177     Nodes.setRecMII(CurMII);
1178     if (CurMII > RecMII)
1179       RecMII = CurMII;
1180   }
1181 
1182   return RecMII;
1183 }
1184 
1185 /// Swap all the anti dependences in the DAG. That means it is no longer a DAG,
1186 /// but we do this to find the circuits, and then change them back.
1187 static void swapAntiDependences(std::vector<SUnit> &SUnits) {
1188   SmallVector<std::pair<SUnit *, SDep>, 8> DepsAdded;
1189   for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
1190     SUnit *SU = &SUnits[i];
1191     for (SUnit::pred_iterator IP = SU->Preds.begin(), EP = SU->Preds.end();
1192          IP != EP; ++IP) {
1193       if (IP->getKind() != SDep::Anti)
1194         continue;
1195       DepsAdded.push_back(std::make_pair(SU, *IP));
1196     }
1197   }
1198   for (SmallVector<std::pair<SUnit *, SDep>, 8>::iterator I = DepsAdded.begin(),
1199                                                           E = DepsAdded.end();
1200        I != E; ++I) {
1201     // Remove this anti dependency and add one in the reverse direction.
1202     SUnit *SU = I->first;
1203     SDep &D = I->second;
1204     SUnit *TargetSU = D.getSUnit();
1205     unsigned Reg = D.getReg();
1206     unsigned Lat = D.getLatency();
1207     SU->removePred(D);
1208     SDep Dep(SU, SDep::Anti, Reg);
1209     Dep.setLatency(Lat);
1210     TargetSU->addPred(Dep);
1211   }
1212 }
1213 
1214 /// Create the adjacency structure of the nodes in the graph.
1215 void SwingSchedulerDAG::Circuits::createAdjacencyStructure(
1216     SwingSchedulerDAG *DAG) {
1217   BitVector Added(SUnits.size());
1218   DenseMap<int, int> OutputDeps;
1219   for (int i = 0, e = SUnits.size(); i != e; ++i) {
1220     Added.reset();
1221     // Add any successor to the adjacency matrix and exclude duplicates.
1222     for (auto &SI : SUnits[i].Succs) {
1223       // Only create a back-edge on the first and last nodes of a dependence
1224       // chain. This records any chains and adds them later.
1225       if (SI.getKind() == SDep::Output) {
1226         int N = SI.getSUnit()->NodeNum;
1227         int BackEdge = i;
1228         auto Dep = OutputDeps.find(BackEdge);
1229         if (Dep != OutputDeps.end()) {
1230           BackEdge = Dep->second;
1231           OutputDeps.erase(Dep);
1232         }
1233         OutputDeps[N] = BackEdge;
1234       }
1235       // Do not process a boundary node, an artificial node.
1236       // A back-edge is processed only if it goes to a Phi.
1237       if (SI.getSUnit()->isBoundaryNode() || SI.isArtificial() ||
1238           (SI.getKind() == SDep::Anti && !SI.getSUnit()->getInstr()->isPHI()))
1239         continue;
1240       int N = SI.getSUnit()->NodeNum;
1241       if (!Added.test(N)) {
1242         AdjK[i].push_back(N);
1243         Added.set(N);
1244       }
1245     }
1246     // A chain edge between a store and a load is treated as a back-edge in the
1247     // adjacency matrix.
1248     for (auto &PI : SUnits[i].Preds) {
1249       if (!SUnits[i].getInstr()->mayStore() ||
1250           !DAG->isLoopCarriedDep(&SUnits[i], PI, false))
1251         continue;
1252       if (PI.getKind() == SDep::Order && PI.getSUnit()->getInstr()->mayLoad()) {
1253         int N = PI.getSUnit()->NodeNum;
1254         if (!Added.test(N)) {
1255           AdjK[i].push_back(N);
1256           Added.set(N);
1257         }
1258       }
1259     }
1260   }
1261   // Add back-edges in the adjacency matrix for the output dependences.
1262   for (auto &OD : OutputDeps)
1263     if (!Added.test(OD.second)) {
1264       AdjK[OD.first].push_back(OD.second);
1265       Added.set(OD.second);
1266     }
1267 }
1268 
1269 /// Identify an elementary circuit in the dependence graph starting at the
1270 /// specified node.
1271 bool SwingSchedulerDAG::Circuits::circuit(int V, int S, NodeSetType &NodeSets,
1272                                           bool HasBackedge) {
1273   SUnit *SV = &SUnits[V];
1274   bool F = false;
1275   Stack.insert(SV);
1276   Blocked.set(V);
1277 
1278   for (auto W : AdjK[V]) {
1279     if (NumPaths > MaxPaths)
1280       break;
1281     if (W < S)
1282       continue;
1283     if (W == S) {
1284       if (!HasBackedge)
1285         NodeSets.push_back(NodeSet(Stack.begin(), Stack.end()));
1286       F = true;
1287       ++NumPaths;
1288       break;
1289     } else if (!Blocked.test(W)) {
1290       if (circuit(W, S, NodeSets,
1291                   Node2Idx->at(W) < Node2Idx->at(V) ? true : HasBackedge))
1292         F = true;
1293     }
1294   }
1295 
1296   if (F)
1297     unblock(V);
1298   else {
1299     for (auto W : AdjK[V]) {
1300       if (W < S)
1301         continue;
1302       if (B[W].count(SV) == 0)
1303         B[W].insert(SV);
1304     }
1305   }
1306   Stack.pop_back();
1307   return F;
1308 }
1309 
1310 /// Unblock a node in the circuit finding algorithm.
1311 void SwingSchedulerDAG::Circuits::unblock(int U) {
1312   Blocked.reset(U);
1313   SmallPtrSet<SUnit *, 4> &BU = B[U];
1314   while (!BU.empty()) {
1315     SmallPtrSet<SUnit *, 4>::iterator SI = BU.begin();
1316     assert(SI != BU.end() && "Invalid B set.");
1317     SUnit *W = *SI;
1318     BU.erase(W);
1319     if (Blocked.test(W->NodeNum))
1320       unblock(W->NodeNum);
1321   }
1322 }
1323 
1324 /// Identify all the elementary circuits in the dependence graph using
1325 /// Johnson's circuit algorithm.
1326 void SwingSchedulerDAG::findCircuits(NodeSetType &NodeSets) {
1327   // Swap all the anti dependences in the DAG. That means it is no longer a DAG,
1328   // but we do this to find the circuits, and then change them back.
1329   swapAntiDependences(SUnits);
1330 
1331   Circuits Cir(SUnits, Topo);
1332   // Create the adjacency structure.
1333   Cir.createAdjacencyStructure(this);
1334   for (int i = 0, e = SUnits.size(); i != e; ++i) {
1335     Cir.reset();
1336     Cir.circuit(i, i, NodeSets);
1337   }
1338 
1339   // Change the dependences back so that we've created a DAG again.
1340   swapAntiDependences(SUnits);
1341 }
1342 
1343 // Create artificial dependencies between the source of COPY/REG_SEQUENCE that
1344 // is loop-carried to the USE in next iteration. This will help pipeliner avoid
1345 // additional copies that are needed across iterations. An artificial dependence
1346 // edge is added from USE to SOURCE of COPY/REG_SEQUENCE.
1347 
1348 // PHI-------Anti-Dep-----> COPY/REG_SEQUENCE (loop-carried)
1349 // SRCOfCopY------True-Dep---> COPY/REG_SEQUENCE
1350 // PHI-------True-Dep------> USEOfPhi
1351 
1352 // The mutation creates
1353 // USEOfPHI -------Artificial-Dep---> SRCOfCopy
1354 
1355 // This overall will ensure, the USEOfPHI is scheduled before SRCOfCopy
1356 // (since USE is a predecessor), implies, the COPY/ REG_SEQUENCE is scheduled
1357 // late  to avoid additional copies across iterations. The possible scheduling
1358 // order would be
1359 // USEOfPHI --- SRCOfCopy---  COPY/REG_SEQUENCE.
1360 
1361 void SwingSchedulerDAG::CopyToPhiMutation::apply(ScheduleDAGInstrs *DAG) {
1362   for (SUnit &SU : DAG->SUnits) {
1363     // Find the COPY/REG_SEQUENCE instruction.
1364     if (!SU.getInstr()->isCopy() && !SU.getInstr()->isRegSequence())
1365       continue;
1366 
1367     // Record the loop carried PHIs.
1368     SmallVector<SUnit *, 4> PHISUs;
1369     // Record the SrcSUs that feed the COPY/REG_SEQUENCE instructions.
1370     SmallVector<SUnit *, 4> SrcSUs;
1371 
1372     for (auto &Dep : SU.Preds) {
1373       SUnit *TmpSU = Dep.getSUnit();
1374       MachineInstr *TmpMI = TmpSU->getInstr();
1375       SDep::Kind DepKind = Dep.getKind();
1376       // Save the loop carried PHI.
1377       if (DepKind == SDep::Anti && TmpMI->isPHI())
1378         PHISUs.push_back(TmpSU);
1379       // Save the source of COPY/REG_SEQUENCE.
1380       // If the source has no pre-decessors, we will end up creating cycles.
1381       else if (DepKind == SDep::Data && !TmpMI->isPHI() && TmpSU->NumPreds > 0)
1382         SrcSUs.push_back(TmpSU);
1383     }
1384 
1385     if (PHISUs.size() == 0 || SrcSUs.size() == 0)
1386       continue;
1387 
1388     // Find the USEs of PHI. If the use is a PHI or REG_SEQUENCE, push back this
1389     // SUnit to the container.
1390     SmallVector<SUnit *, 8> UseSUs;
1391     // Do not use iterator based loop here as we are updating the container.
1392     for (size_t Index = 0; Index < PHISUs.size(); ++Index) {
1393       for (auto &Dep : PHISUs[Index]->Succs) {
1394         if (Dep.getKind() != SDep::Data)
1395           continue;
1396 
1397         SUnit *TmpSU = Dep.getSUnit();
1398         MachineInstr *TmpMI = TmpSU->getInstr();
1399         if (TmpMI->isPHI() || TmpMI->isRegSequence()) {
1400           PHISUs.push_back(TmpSU);
1401           continue;
1402         }
1403         UseSUs.push_back(TmpSU);
1404       }
1405     }
1406 
1407     if (UseSUs.size() == 0)
1408       continue;
1409 
1410     SwingSchedulerDAG *SDAG = cast<SwingSchedulerDAG>(DAG);
1411     // Add the artificial dependencies if it does not form a cycle.
1412     for (auto I : UseSUs) {
1413       for (auto Src : SrcSUs) {
1414         if (!SDAG->Topo.IsReachable(I, Src) && Src != I) {
1415           Src->addPred(SDep(I, SDep::Artificial));
1416           SDAG->Topo.AddPred(Src, I);
1417         }
1418       }
1419     }
1420   }
1421 }
1422 
1423 /// Return true for DAG nodes that we ignore when computing the cost functions.
1424 /// We ignore the back-edge recurrence in order to avoid unbounded recursion
1425 /// in the calculation of the ASAP, ALAP, etc functions.
1426 static bool ignoreDependence(const SDep &D, bool isPred) {
1427   if (D.isArtificial())
1428     return true;
1429   return D.getKind() == SDep::Anti && isPred;
1430 }
1431 
1432 /// Compute several functions need to order the nodes for scheduling.
1433 ///  ASAP - Earliest time to schedule a node.
1434 ///  ALAP - Latest time to schedule a node.
1435 ///  MOV - Mobility function, difference between ALAP and ASAP.
1436 ///  D - Depth of each node.
1437 ///  H - Height of each node.
1438 void SwingSchedulerDAG::computeNodeFunctions(NodeSetType &NodeSets) {
1439   ScheduleInfo.resize(SUnits.size());
1440 
1441   LLVM_DEBUG({
1442     for (ScheduleDAGTopologicalSort::const_iterator I = Topo.begin(),
1443                                                     E = Topo.end();
1444          I != E; ++I) {
1445       const SUnit &SU = SUnits[*I];
1446       dumpNode(SU);
1447     }
1448   });
1449 
1450   int maxASAP = 0;
1451   // Compute ASAP and ZeroLatencyDepth.
1452   for (ScheduleDAGTopologicalSort::const_iterator I = Topo.begin(),
1453                                                   E = Topo.end();
1454        I != E; ++I) {
1455     int asap = 0;
1456     int zeroLatencyDepth = 0;
1457     SUnit *SU = &SUnits[*I];
1458     for (SUnit::const_pred_iterator IP = SU->Preds.begin(),
1459                                     EP = SU->Preds.end();
1460          IP != EP; ++IP) {
1461       SUnit *pred = IP->getSUnit();
1462       if (IP->getLatency() == 0)
1463         zeroLatencyDepth =
1464             std::max(zeroLatencyDepth, getZeroLatencyDepth(pred) + 1);
1465       if (ignoreDependence(*IP, true))
1466         continue;
1467       asap = std::max(asap, (int)(getASAP(pred) + IP->getLatency() -
1468                                   getDistance(pred, SU, *IP) * MII));
1469     }
1470     maxASAP = std::max(maxASAP, asap);
1471     ScheduleInfo[*I].ASAP = asap;
1472     ScheduleInfo[*I].ZeroLatencyDepth = zeroLatencyDepth;
1473   }
1474 
1475   // Compute ALAP, ZeroLatencyHeight, and MOV.
1476   for (ScheduleDAGTopologicalSort::const_reverse_iterator I = Topo.rbegin(),
1477                                                           E = Topo.rend();
1478        I != E; ++I) {
1479     int alap = maxASAP;
1480     int zeroLatencyHeight = 0;
1481     SUnit *SU = &SUnits[*I];
1482     for (SUnit::const_succ_iterator IS = SU->Succs.begin(),
1483                                     ES = SU->Succs.end();
1484          IS != ES; ++IS) {
1485       SUnit *succ = IS->getSUnit();
1486       if (IS->getLatency() == 0)
1487         zeroLatencyHeight =
1488             std::max(zeroLatencyHeight, getZeroLatencyHeight(succ) + 1);
1489       if (ignoreDependence(*IS, true))
1490         continue;
1491       alap = std::min(alap, (int)(getALAP(succ) - IS->getLatency() +
1492                                   getDistance(SU, succ, *IS) * MII));
1493     }
1494 
1495     ScheduleInfo[*I].ALAP = alap;
1496     ScheduleInfo[*I].ZeroLatencyHeight = zeroLatencyHeight;
1497   }
1498 
1499   // After computing the node functions, compute the summary for each node set.
1500   for (NodeSet &I : NodeSets)
1501     I.computeNodeSetInfo(this);
1502 
1503   LLVM_DEBUG({
1504     for (unsigned i = 0; i < SUnits.size(); i++) {
1505       dbgs() << "\tNode " << i << ":\n";
1506       dbgs() << "\t   ASAP = " << getASAP(&SUnits[i]) << "\n";
1507       dbgs() << "\t   ALAP = " << getALAP(&SUnits[i]) << "\n";
1508       dbgs() << "\t   MOV  = " << getMOV(&SUnits[i]) << "\n";
1509       dbgs() << "\t   D    = " << getDepth(&SUnits[i]) << "\n";
1510       dbgs() << "\t   H    = " << getHeight(&SUnits[i]) << "\n";
1511       dbgs() << "\t   ZLD  = " << getZeroLatencyDepth(&SUnits[i]) << "\n";
1512       dbgs() << "\t   ZLH  = " << getZeroLatencyHeight(&SUnits[i]) << "\n";
1513     }
1514   });
1515 }
1516 
1517 /// Compute the Pred_L(O) set, as defined in the paper. The set is defined
1518 /// as the predecessors of the elements of NodeOrder that are not also in
1519 /// NodeOrder.
1520 static bool pred_L(SetVector<SUnit *> &NodeOrder,
1521                    SmallSetVector<SUnit *, 8> &Preds,
1522                    const NodeSet *S = nullptr) {
1523   Preds.clear();
1524   for (SetVector<SUnit *>::iterator I = NodeOrder.begin(), E = NodeOrder.end();
1525        I != E; ++I) {
1526     for (SUnit::pred_iterator PI = (*I)->Preds.begin(), PE = (*I)->Preds.end();
1527          PI != PE; ++PI) {
1528       if (S && S->count(PI->getSUnit()) == 0)
1529         continue;
1530       if (ignoreDependence(*PI, true))
1531         continue;
1532       if (NodeOrder.count(PI->getSUnit()) == 0)
1533         Preds.insert(PI->getSUnit());
1534     }
1535     // Back-edges are predecessors with an anti-dependence.
1536     for (SUnit::const_succ_iterator IS = (*I)->Succs.begin(),
1537                                     ES = (*I)->Succs.end();
1538          IS != ES; ++IS) {
1539       if (IS->getKind() != SDep::Anti)
1540         continue;
1541       if (S && S->count(IS->getSUnit()) == 0)
1542         continue;
1543       if (NodeOrder.count(IS->getSUnit()) == 0)
1544         Preds.insert(IS->getSUnit());
1545     }
1546   }
1547   return !Preds.empty();
1548 }
1549 
1550 /// Compute the Succ_L(O) set, as defined in the paper. The set is defined
1551 /// as the successors of the elements of NodeOrder that are not also in
1552 /// NodeOrder.
1553 static bool succ_L(SetVector<SUnit *> &NodeOrder,
1554                    SmallSetVector<SUnit *, 8> &Succs,
1555                    const NodeSet *S = nullptr) {
1556   Succs.clear();
1557   for (SetVector<SUnit *>::iterator I = NodeOrder.begin(), E = NodeOrder.end();
1558        I != E; ++I) {
1559     for (SUnit::succ_iterator SI = (*I)->Succs.begin(), SE = (*I)->Succs.end();
1560          SI != SE; ++SI) {
1561       if (S && S->count(SI->getSUnit()) == 0)
1562         continue;
1563       if (ignoreDependence(*SI, false))
1564         continue;
1565       if (NodeOrder.count(SI->getSUnit()) == 0)
1566         Succs.insert(SI->getSUnit());
1567     }
1568     for (SUnit::const_pred_iterator PI = (*I)->Preds.begin(),
1569                                     PE = (*I)->Preds.end();
1570          PI != PE; ++PI) {
1571       if (PI->getKind() != SDep::Anti)
1572         continue;
1573       if (S && S->count(PI->getSUnit()) == 0)
1574         continue;
1575       if (NodeOrder.count(PI->getSUnit()) == 0)
1576         Succs.insert(PI->getSUnit());
1577     }
1578   }
1579   return !Succs.empty();
1580 }
1581 
1582 /// Return true if there is a path from the specified node to any of the nodes
1583 /// in DestNodes. Keep track and return the nodes in any path.
1584 static bool computePath(SUnit *Cur, SetVector<SUnit *> &Path,
1585                         SetVector<SUnit *> &DestNodes,
1586                         SetVector<SUnit *> &Exclude,
1587                         SmallPtrSet<SUnit *, 8> &Visited) {
1588   if (Cur->isBoundaryNode())
1589     return false;
1590   if (Exclude.count(Cur) != 0)
1591     return false;
1592   if (DestNodes.count(Cur) != 0)
1593     return true;
1594   if (!Visited.insert(Cur).second)
1595     return Path.count(Cur) != 0;
1596   bool FoundPath = false;
1597   for (auto &SI : Cur->Succs)
1598     FoundPath |= computePath(SI.getSUnit(), Path, DestNodes, Exclude, Visited);
1599   for (auto &PI : Cur->Preds)
1600     if (PI.getKind() == SDep::Anti)
1601       FoundPath |=
1602           computePath(PI.getSUnit(), Path, DestNodes, Exclude, Visited);
1603   if (FoundPath)
1604     Path.insert(Cur);
1605   return FoundPath;
1606 }
1607 
1608 /// Return true if Set1 is a subset of Set2.
1609 template <class S1Ty, class S2Ty> static bool isSubset(S1Ty &Set1, S2Ty &Set2) {
1610   for (typename S1Ty::iterator I = Set1.begin(), E = Set1.end(); I != E; ++I)
1611     if (Set2.count(*I) == 0)
1612       return false;
1613   return true;
1614 }
1615 
1616 /// Compute the live-out registers for the instructions in a node-set.
1617 /// The live-out registers are those that are defined in the node-set,
1618 /// but not used. Except for use operands of Phis.
1619 static void computeLiveOuts(MachineFunction &MF, RegPressureTracker &RPTracker,
1620                             NodeSet &NS) {
1621   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
1622   MachineRegisterInfo &MRI = MF.getRegInfo();
1623   SmallVector<RegisterMaskPair, 8> LiveOutRegs;
1624   SmallSet<unsigned, 4> Uses;
1625   for (SUnit *SU : NS) {
1626     const MachineInstr *MI = SU->getInstr();
1627     if (MI->isPHI())
1628       continue;
1629     for (const MachineOperand &MO : MI->operands())
1630       if (MO.isReg() && MO.isUse()) {
1631         Register Reg = MO.getReg();
1632         if (Register::isVirtualRegister(Reg))
1633           Uses.insert(Reg);
1634         else if (MRI.isAllocatable(Reg))
1635           for (MCRegUnitIterator Units(Reg, TRI); Units.isValid(); ++Units)
1636             Uses.insert(*Units);
1637       }
1638   }
1639   for (SUnit *SU : NS)
1640     for (const MachineOperand &MO : SU->getInstr()->operands())
1641       if (MO.isReg() && MO.isDef() && !MO.isDead()) {
1642         Register Reg = MO.getReg();
1643         if (Register::isVirtualRegister(Reg)) {
1644           if (!Uses.count(Reg))
1645             LiveOutRegs.push_back(RegisterMaskPair(Reg,
1646                                                    LaneBitmask::getNone()));
1647         } else if (MRI.isAllocatable(Reg)) {
1648           for (MCRegUnitIterator Units(Reg, TRI); Units.isValid(); ++Units)
1649             if (!Uses.count(*Units))
1650               LiveOutRegs.push_back(RegisterMaskPair(*Units,
1651                                                      LaneBitmask::getNone()));
1652         }
1653       }
1654   RPTracker.addLiveRegs(LiveOutRegs);
1655 }
1656 
1657 /// A heuristic to filter nodes in recurrent node-sets if the register
1658 /// pressure of a set is too high.
1659 void SwingSchedulerDAG::registerPressureFilter(NodeSetType &NodeSets) {
1660   for (auto &NS : NodeSets) {
1661     // Skip small node-sets since they won't cause register pressure problems.
1662     if (NS.size() <= 2)
1663       continue;
1664     IntervalPressure RecRegPressure;
1665     RegPressureTracker RecRPTracker(RecRegPressure);
1666     RecRPTracker.init(&MF, &RegClassInfo, &LIS, BB, BB->end(), false, true);
1667     computeLiveOuts(MF, RecRPTracker, NS);
1668     RecRPTracker.closeBottom();
1669 
1670     std::vector<SUnit *> SUnits(NS.begin(), NS.end());
1671     llvm::sort(SUnits, [](const SUnit *A, const SUnit *B) {
1672       return A->NodeNum > B->NodeNum;
1673     });
1674 
1675     for (auto &SU : SUnits) {
1676       // Since we're computing the register pressure for a subset of the
1677       // instructions in a block, we need to set the tracker for each
1678       // instruction in the node-set. The tracker is set to the instruction
1679       // just after the one we're interested in.
1680       MachineBasicBlock::const_iterator CurInstI = SU->getInstr();
1681       RecRPTracker.setPos(std::next(CurInstI));
1682 
1683       RegPressureDelta RPDelta;
1684       ArrayRef<PressureChange> CriticalPSets;
1685       RecRPTracker.getMaxUpwardPressureDelta(SU->getInstr(), nullptr, RPDelta,
1686                                              CriticalPSets,
1687                                              RecRegPressure.MaxSetPressure);
1688       if (RPDelta.Excess.isValid()) {
1689         LLVM_DEBUG(
1690             dbgs() << "Excess register pressure: SU(" << SU->NodeNum << ") "
1691                    << TRI->getRegPressureSetName(RPDelta.Excess.getPSet())
1692                    << ":" << RPDelta.Excess.getUnitInc());
1693         NS.setExceedPressure(SU);
1694         break;
1695       }
1696       RecRPTracker.recede();
1697     }
1698   }
1699 }
1700 
1701 /// A heuristic to colocate node sets that have the same set of
1702 /// successors.
1703 void SwingSchedulerDAG::colocateNodeSets(NodeSetType &NodeSets) {
1704   unsigned Colocate = 0;
1705   for (int i = 0, e = NodeSets.size(); i < e; ++i) {
1706     NodeSet &N1 = NodeSets[i];
1707     SmallSetVector<SUnit *, 8> S1;
1708     if (N1.empty() || !succ_L(N1, S1))
1709       continue;
1710     for (int j = i + 1; j < e; ++j) {
1711       NodeSet &N2 = NodeSets[j];
1712       if (N1.compareRecMII(N2) != 0)
1713         continue;
1714       SmallSetVector<SUnit *, 8> S2;
1715       if (N2.empty() || !succ_L(N2, S2))
1716         continue;
1717       if (isSubset(S1, S2) && S1.size() == S2.size()) {
1718         N1.setColocate(++Colocate);
1719         N2.setColocate(Colocate);
1720         break;
1721       }
1722     }
1723   }
1724 }
1725 
1726 /// Check if the existing node-sets are profitable. If not, then ignore the
1727 /// recurrent node-sets, and attempt to schedule all nodes together. This is
1728 /// a heuristic. If the MII is large and all the recurrent node-sets are small,
1729 /// then it's best to try to schedule all instructions together instead of
1730 /// starting with the recurrent node-sets.
1731 void SwingSchedulerDAG::checkNodeSets(NodeSetType &NodeSets) {
1732   // Look for loops with a large MII.
1733   if (MII < 17)
1734     return;
1735   // Check if the node-set contains only a simple add recurrence.
1736   for (auto &NS : NodeSets) {
1737     if (NS.getRecMII() > 2)
1738       return;
1739     if (NS.getMaxDepth() > MII)
1740       return;
1741   }
1742   NodeSets.clear();
1743   LLVM_DEBUG(dbgs() << "Clear recurrence node-sets\n");
1744   return;
1745 }
1746 
1747 /// Add the nodes that do not belong to a recurrence set into groups
1748 /// based upon connected componenets.
1749 void SwingSchedulerDAG::groupRemainingNodes(NodeSetType &NodeSets) {
1750   SetVector<SUnit *> NodesAdded;
1751   SmallPtrSet<SUnit *, 8> Visited;
1752   // Add the nodes that are on a path between the previous node sets and
1753   // the current node set.
1754   for (NodeSet &I : NodeSets) {
1755     SmallSetVector<SUnit *, 8> N;
1756     // Add the nodes from the current node set to the previous node set.
1757     if (succ_L(I, N)) {
1758       SetVector<SUnit *> Path;
1759       for (SUnit *NI : N) {
1760         Visited.clear();
1761         computePath(NI, Path, NodesAdded, I, Visited);
1762       }
1763       if (!Path.empty())
1764         I.insert(Path.begin(), Path.end());
1765     }
1766     // Add the nodes from the previous node set to the current node set.
1767     N.clear();
1768     if (succ_L(NodesAdded, N)) {
1769       SetVector<SUnit *> Path;
1770       for (SUnit *NI : N) {
1771         Visited.clear();
1772         computePath(NI, Path, I, NodesAdded, Visited);
1773       }
1774       if (!Path.empty())
1775         I.insert(Path.begin(), Path.end());
1776     }
1777     NodesAdded.insert(I.begin(), I.end());
1778   }
1779 
1780   // Create a new node set with the connected nodes of any successor of a node
1781   // in a recurrent set.
1782   NodeSet NewSet;
1783   SmallSetVector<SUnit *, 8> N;
1784   if (succ_L(NodesAdded, N))
1785     for (SUnit *I : N)
1786       addConnectedNodes(I, NewSet, NodesAdded);
1787   if (!NewSet.empty())
1788     NodeSets.push_back(NewSet);
1789 
1790   // Create a new node set with the connected nodes of any predecessor of a node
1791   // in a recurrent set.
1792   NewSet.clear();
1793   if (pred_L(NodesAdded, N))
1794     for (SUnit *I : N)
1795       addConnectedNodes(I, NewSet, NodesAdded);
1796   if (!NewSet.empty())
1797     NodeSets.push_back(NewSet);
1798 
1799   // Create new nodes sets with the connected nodes any remaining node that
1800   // has no predecessor.
1801   for (unsigned i = 0; i < SUnits.size(); ++i) {
1802     SUnit *SU = &SUnits[i];
1803     if (NodesAdded.count(SU) == 0) {
1804       NewSet.clear();
1805       addConnectedNodes(SU, NewSet, NodesAdded);
1806       if (!NewSet.empty())
1807         NodeSets.push_back(NewSet);
1808     }
1809   }
1810 }
1811 
1812 /// Add the node to the set, and add all of its connected nodes to the set.
1813 void SwingSchedulerDAG::addConnectedNodes(SUnit *SU, NodeSet &NewSet,
1814                                           SetVector<SUnit *> &NodesAdded) {
1815   NewSet.insert(SU);
1816   NodesAdded.insert(SU);
1817   for (auto &SI : SU->Succs) {
1818     SUnit *Successor = SI.getSUnit();
1819     if (!SI.isArtificial() && NodesAdded.count(Successor) == 0)
1820       addConnectedNodes(Successor, NewSet, NodesAdded);
1821   }
1822   for (auto &PI : SU->Preds) {
1823     SUnit *Predecessor = PI.getSUnit();
1824     if (!PI.isArtificial() && NodesAdded.count(Predecessor) == 0)
1825       addConnectedNodes(Predecessor, NewSet, NodesAdded);
1826   }
1827 }
1828 
1829 /// Return true if Set1 contains elements in Set2. The elements in common
1830 /// are returned in a different container.
1831 static bool isIntersect(SmallSetVector<SUnit *, 8> &Set1, const NodeSet &Set2,
1832                         SmallSetVector<SUnit *, 8> &Result) {
1833   Result.clear();
1834   for (unsigned i = 0, e = Set1.size(); i != e; ++i) {
1835     SUnit *SU = Set1[i];
1836     if (Set2.count(SU) != 0)
1837       Result.insert(SU);
1838   }
1839   return !Result.empty();
1840 }
1841 
1842 /// Merge the recurrence node sets that have the same initial node.
1843 void SwingSchedulerDAG::fuseRecs(NodeSetType &NodeSets) {
1844   for (NodeSetType::iterator I = NodeSets.begin(), E = NodeSets.end(); I != E;
1845        ++I) {
1846     NodeSet &NI = *I;
1847     for (NodeSetType::iterator J = I + 1; J != E;) {
1848       NodeSet &NJ = *J;
1849       if (NI.getNode(0)->NodeNum == NJ.getNode(0)->NodeNum) {
1850         if (NJ.compareRecMII(NI) > 0)
1851           NI.setRecMII(NJ.getRecMII());
1852         for (NodeSet::iterator NII = J->begin(), ENI = J->end(); NII != ENI;
1853              ++NII)
1854           I->insert(*NII);
1855         NodeSets.erase(J);
1856         E = NodeSets.end();
1857       } else {
1858         ++J;
1859       }
1860     }
1861   }
1862 }
1863 
1864 /// Remove nodes that have been scheduled in previous NodeSets.
1865 void SwingSchedulerDAG::removeDuplicateNodes(NodeSetType &NodeSets) {
1866   for (NodeSetType::iterator I = NodeSets.begin(), E = NodeSets.end(); I != E;
1867        ++I)
1868     for (NodeSetType::iterator J = I + 1; J != E;) {
1869       J->remove_if([&](SUnit *SUJ) { return I->count(SUJ); });
1870 
1871       if (J->empty()) {
1872         NodeSets.erase(J);
1873         E = NodeSets.end();
1874       } else {
1875         ++J;
1876       }
1877     }
1878 }
1879 
1880 /// Compute an ordered list of the dependence graph nodes, which
1881 /// indicates the order that the nodes will be scheduled.  This is a
1882 /// two-level algorithm. First, a partial order is created, which
1883 /// consists of a list of sets ordered from highest to lowest priority.
1884 void SwingSchedulerDAG::computeNodeOrder(NodeSetType &NodeSets) {
1885   SmallSetVector<SUnit *, 8> R;
1886   NodeOrder.clear();
1887 
1888   for (auto &Nodes : NodeSets) {
1889     LLVM_DEBUG(dbgs() << "NodeSet size " << Nodes.size() << "\n");
1890     OrderKind Order;
1891     SmallSetVector<SUnit *, 8> N;
1892     if (pred_L(NodeOrder, N) && isSubset(N, Nodes)) {
1893       R.insert(N.begin(), N.end());
1894       Order = BottomUp;
1895       LLVM_DEBUG(dbgs() << "  Bottom up (preds) ");
1896     } else if (succ_L(NodeOrder, N) && isSubset(N, Nodes)) {
1897       R.insert(N.begin(), N.end());
1898       Order = TopDown;
1899       LLVM_DEBUG(dbgs() << "  Top down (succs) ");
1900     } else if (isIntersect(N, Nodes, R)) {
1901       // If some of the successors are in the existing node-set, then use the
1902       // top-down ordering.
1903       Order = TopDown;
1904       LLVM_DEBUG(dbgs() << "  Top down (intersect) ");
1905     } else if (NodeSets.size() == 1) {
1906       for (auto &N : Nodes)
1907         if (N->Succs.size() == 0)
1908           R.insert(N);
1909       Order = BottomUp;
1910       LLVM_DEBUG(dbgs() << "  Bottom up (all) ");
1911     } else {
1912       // Find the node with the highest ASAP.
1913       SUnit *maxASAP = nullptr;
1914       for (SUnit *SU : Nodes) {
1915         if (maxASAP == nullptr || getASAP(SU) > getASAP(maxASAP) ||
1916             (getASAP(SU) == getASAP(maxASAP) && SU->NodeNum > maxASAP->NodeNum))
1917           maxASAP = SU;
1918       }
1919       R.insert(maxASAP);
1920       Order = BottomUp;
1921       LLVM_DEBUG(dbgs() << "  Bottom up (default) ");
1922     }
1923 
1924     while (!R.empty()) {
1925       if (Order == TopDown) {
1926         // Choose the node with the maximum height.  If more than one, choose
1927         // the node wiTH the maximum ZeroLatencyHeight. If still more than one,
1928         // choose the node with the lowest MOV.
1929         while (!R.empty()) {
1930           SUnit *maxHeight = nullptr;
1931           for (SUnit *I : R) {
1932             if (maxHeight == nullptr || getHeight(I) > getHeight(maxHeight))
1933               maxHeight = I;
1934             else if (getHeight(I) == getHeight(maxHeight) &&
1935                      getZeroLatencyHeight(I) > getZeroLatencyHeight(maxHeight))
1936               maxHeight = I;
1937             else if (getHeight(I) == getHeight(maxHeight) &&
1938                      getZeroLatencyHeight(I) ==
1939                          getZeroLatencyHeight(maxHeight) &&
1940                      getMOV(I) < getMOV(maxHeight))
1941               maxHeight = I;
1942           }
1943           NodeOrder.insert(maxHeight);
1944           LLVM_DEBUG(dbgs() << maxHeight->NodeNum << " ");
1945           R.remove(maxHeight);
1946           for (const auto &I : maxHeight->Succs) {
1947             if (Nodes.count(I.getSUnit()) == 0)
1948               continue;
1949             if (NodeOrder.count(I.getSUnit()) != 0)
1950               continue;
1951             if (ignoreDependence(I, false))
1952               continue;
1953             R.insert(I.getSUnit());
1954           }
1955           // Back-edges are predecessors with an anti-dependence.
1956           for (const auto &I : maxHeight->Preds) {
1957             if (I.getKind() != SDep::Anti)
1958               continue;
1959             if (Nodes.count(I.getSUnit()) == 0)
1960               continue;
1961             if (NodeOrder.count(I.getSUnit()) != 0)
1962               continue;
1963             R.insert(I.getSUnit());
1964           }
1965         }
1966         Order = BottomUp;
1967         LLVM_DEBUG(dbgs() << "\n   Switching order to bottom up ");
1968         SmallSetVector<SUnit *, 8> N;
1969         if (pred_L(NodeOrder, N, &Nodes))
1970           R.insert(N.begin(), N.end());
1971       } else {
1972         // Choose the node with the maximum depth.  If more than one, choose
1973         // the node with the maximum ZeroLatencyDepth. If still more than one,
1974         // choose the node with the lowest MOV.
1975         while (!R.empty()) {
1976           SUnit *maxDepth = nullptr;
1977           for (SUnit *I : R) {
1978             if (maxDepth == nullptr || getDepth(I) > getDepth(maxDepth))
1979               maxDepth = I;
1980             else if (getDepth(I) == getDepth(maxDepth) &&
1981                      getZeroLatencyDepth(I) > getZeroLatencyDepth(maxDepth))
1982               maxDepth = I;
1983             else if (getDepth(I) == getDepth(maxDepth) &&
1984                      getZeroLatencyDepth(I) == getZeroLatencyDepth(maxDepth) &&
1985                      getMOV(I) < getMOV(maxDepth))
1986               maxDepth = I;
1987           }
1988           NodeOrder.insert(maxDepth);
1989           LLVM_DEBUG(dbgs() << maxDepth->NodeNum << " ");
1990           R.remove(maxDepth);
1991           if (Nodes.isExceedSU(maxDepth)) {
1992             Order = TopDown;
1993             R.clear();
1994             R.insert(Nodes.getNode(0));
1995             break;
1996           }
1997           for (const auto &I : maxDepth->Preds) {
1998             if (Nodes.count(I.getSUnit()) == 0)
1999               continue;
2000             if (NodeOrder.count(I.getSUnit()) != 0)
2001               continue;
2002             R.insert(I.getSUnit());
2003           }
2004           // Back-edges are predecessors with an anti-dependence.
2005           for (const auto &I : maxDepth->Succs) {
2006             if (I.getKind() != SDep::Anti)
2007               continue;
2008             if (Nodes.count(I.getSUnit()) == 0)
2009               continue;
2010             if (NodeOrder.count(I.getSUnit()) != 0)
2011               continue;
2012             R.insert(I.getSUnit());
2013           }
2014         }
2015         Order = TopDown;
2016         LLVM_DEBUG(dbgs() << "\n   Switching order to top down ");
2017         SmallSetVector<SUnit *, 8> N;
2018         if (succ_L(NodeOrder, N, &Nodes))
2019           R.insert(N.begin(), N.end());
2020       }
2021     }
2022     LLVM_DEBUG(dbgs() << "\nDone with Nodeset\n");
2023   }
2024 
2025   LLVM_DEBUG({
2026     dbgs() << "Node order: ";
2027     for (SUnit *I : NodeOrder)
2028       dbgs() << " " << I->NodeNum << " ";
2029     dbgs() << "\n";
2030   });
2031 }
2032 
2033 /// Process the nodes in the computed order and create the pipelined schedule
2034 /// of the instructions, if possible. Return true if a schedule is found.
2035 bool SwingSchedulerDAG::schedulePipeline(SMSchedule &Schedule) {
2036 
2037   if (NodeOrder.empty()){
2038     LLVM_DEBUG(dbgs() << "NodeOrder is empty! abort scheduling\n" );
2039     return false;
2040   }
2041 
2042   bool scheduleFound = false;
2043   unsigned II = 0;
2044   // Keep increasing II until a valid schedule is found.
2045   for (II = MII; II <= MAX_II && !scheduleFound; ++II) {
2046     Schedule.reset();
2047     Schedule.setInitiationInterval(II);
2048     LLVM_DEBUG(dbgs() << "Try to schedule with " << II << "\n");
2049 
2050     SetVector<SUnit *>::iterator NI = NodeOrder.begin();
2051     SetVector<SUnit *>::iterator NE = NodeOrder.end();
2052     do {
2053       SUnit *SU = *NI;
2054 
2055       // Compute the schedule time for the instruction, which is based
2056       // upon the scheduled time for any predecessors/successors.
2057       int EarlyStart = INT_MIN;
2058       int LateStart = INT_MAX;
2059       // These values are set when the size of the schedule window is limited
2060       // due to chain dependences.
2061       int SchedEnd = INT_MAX;
2062       int SchedStart = INT_MIN;
2063       Schedule.computeStart(SU, &EarlyStart, &LateStart, &SchedEnd, &SchedStart,
2064                             II, this);
2065       LLVM_DEBUG({
2066         dbgs() << "\n";
2067         dbgs() << "Inst (" << SU->NodeNum << ") ";
2068         SU->getInstr()->dump();
2069         dbgs() << "\n";
2070       });
2071       LLVM_DEBUG({
2072         dbgs() << format("\tes: %8x ls: %8x me: %8x ms: %8x\n", EarlyStart,
2073                          LateStart, SchedEnd, SchedStart);
2074       });
2075 
2076       if (EarlyStart > LateStart || SchedEnd < EarlyStart ||
2077           SchedStart > LateStart)
2078         scheduleFound = false;
2079       else if (EarlyStart != INT_MIN && LateStart == INT_MAX) {
2080         SchedEnd = std::min(SchedEnd, EarlyStart + (int)II - 1);
2081         scheduleFound = Schedule.insert(SU, EarlyStart, SchedEnd, II);
2082       } else if (EarlyStart == INT_MIN && LateStart != INT_MAX) {
2083         SchedStart = std::max(SchedStart, LateStart - (int)II + 1);
2084         scheduleFound = Schedule.insert(SU, LateStart, SchedStart, II);
2085       } else if (EarlyStart != INT_MIN && LateStart != INT_MAX) {
2086         SchedEnd =
2087             std::min(SchedEnd, std::min(LateStart, EarlyStart + (int)II - 1));
2088         // When scheduling a Phi it is better to start at the late cycle and go
2089         // backwards. The default order may insert the Phi too far away from
2090         // its first dependence.
2091         if (SU->getInstr()->isPHI())
2092           scheduleFound = Schedule.insert(SU, SchedEnd, EarlyStart, II);
2093         else
2094           scheduleFound = Schedule.insert(SU, EarlyStart, SchedEnd, II);
2095       } else {
2096         int FirstCycle = Schedule.getFirstCycle();
2097         scheduleFound = Schedule.insert(SU, FirstCycle + getASAP(SU),
2098                                         FirstCycle + getASAP(SU) + II - 1, II);
2099       }
2100       // Even if we find a schedule, make sure the schedule doesn't exceed the
2101       // allowable number of stages. We keep trying if this happens.
2102       if (scheduleFound)
2103         if (SwpMaxStages > -1 &&
2104             Schedule.getMaxStageCount() > (unsigned)SwpMaxStages)
2105           scheduleFound = false;
2106 
2107       LLVM_DEBUG({
2108         if (!scheduleFound)
2109           dbgs() << "\tCan't schedule\n";
2110       });
2111     } while (++NI != NE && scheduleFound);
2112 
2113     // If a schedule is found, check if it is a valid schedule too.
2114     if (scheduleFound)
2115       scheduleFound = Schedule.isValidSchedule(this);
2116   }
2117 
2118   LLVM_DEBUG(dbgs() << "Schedule Found? " << scheduleFound << " (II=" << II
2119                     << ")\n");
2120 
2121   if (scheduleFound) {
2122     Schedule.finalizeSchedule(this);
2123     Pass.ORE->emit([&]() {
2124       return MachineOptimizationRemarkAnalysis(
2125                  DEBUG_TYPE, "schedule", Loop.getStartLoc(), Loop.getHeader())
2126              << "Schedule found with Initiation Interval: " << ore::NV("II", II)
2127              << ", MaxStageCount: "
2128              << ore::NV("MaxStageCount", Schedule.getMaxStageCount());
2129     });
2130   } else
2131     Schedule.reset();
2132 
2133   return scheduleFound && Schedule.getMaxStageCount() > 0;
2134 }
2135 
2136 /// Return true if we can compute the amount the instruction changes
2137 /// during each iteration. Set Delta to the amount of the change.
2138 bool SwingSchedulerDAG::computeDelta(MachineInstr &MI, unsigned &Delta) {
2139   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
2140   const MachineOperand *BaseOp;
2141   int64_t Offset;
2142   bool OffsetIsScalable;
2143   if (!TII->getMemOperandWithOffset(MI, BaseOp, Offset, OffsetIsScalable, TRI))
2144     return false;
2145 
2146   // FIXME: This algorithm assumes instructions have fixed-size offsets.
2147   if (OffsetIsScalable)
2148     return false;
2149 
2150   if (!BaseOp->isReg())
2151     return false;
2152 
2153   Register BaseReg = BaseOp->getReg();
2154 
2155   MachineRegisterInfo &MRI = MF.getRegInfo();
2156   // Check if there is a Phi. If so, get the definition in the loop.
2157   MachineInstr *BaseDef = MRI.getVRegDef(BaseReg);
2158   if (BaseDef && BaseDef->isPHI()) {
2159     BaseReg = getLoopPhiReg(*BaseDef, MI.getParent());
2160     BaseDef = MRI.getVRegDef(BaseReg);
2161   }
2162   if (!BaseDef)
2163     return false;
2164 
2165   int D = 0;
2166   if (!TII->getIncrementValue(*BaseDef, D) && D >= 0)
2167     return false;
2168 
2169   Delta = D;
2170   return true;
2171 }
2172 
2173 /// Check if we can change the instruction to use an offset value from the
2174 /// previous iteration. If so, return true and set the base and offset values
2175 /// so that we can rewrite the load, if necessary.
2176 ///   v1 = Phi(v0, v3)
2177 ///   v2 = load v1, 0
2178 ///   v3 = post_store v1, 4, x
2179 /// This function enables the load to be rewritten as v2 = load v3, 4.
2180 bool SwingSchedulerDAG::canUseLastOffsetValue(MachineInstr *MI,
2181                                               unsigned &BasePos,
2182                                               unsigned &OffsetPos,
2183                                               unsigned &NewBase,
2184                                               int64_t &Offset) {
2185   // Get the load instruction.
2186   if (TII->isPostIncrement(*MI))
2187     return false;
2188   unsigned BasePosLd, OffsetPosLd;
2189   if (!TII->getBaseAndOffsetPosition(*MI, BasePosLd, OffsetPosLd))
2190     return false;
2191   Register BaseReg = MI->getOperand(BasePosLd).getReg();
2192 
2193   // Look for the Phi instruction.
2194   MachineRegisterInfo &MRI = MI->getMF()->getRegInfo();
2195   MachineInstr *Phi = MRI.getVRegDef(BaseReg);
2196   if (!Phi || !Phi->isPHI())
2197     return false;
2198   // Get the register defined in the loop block.
2199   unsigned PrevReg = getLoopPhiReg(*Phi, MI->getParent());
2200   if (!PrevReg)
2201     return false;
2202 
2203   // Check for the post-increment load/store instruction.
2204   MachineInstr *PrevDef = MRI.getVRegDef(PrevReg);
2205   if (!PrevDef || PrevDef == MI)
2206     return false;
2207 
2208   if (!TII->isPostIncrement(*PrevDef))
2209     return false;
2210 
2211   unsigned BasePos1 = 0, OffsetPos1 = 0;
2212   if (!TII->getBaseAndOffsetPosition(*PrevDef, BasePos1, OffsetPos1))
2213     return false;
2214 
2215   // Make sure that the instructions do not access the same memory location in
2216   // the next iteration.
2217   int64_t LoadOffset = MI->getOperand(OffsetPosLd).getImm();
2218   int64_t StoreOffset = PrevDef->getOperand(OffsetPos1).getImm();
2219   MachineInstr *NewMI = MF.CloneMachineInstr(MI);
2220   NewMI->getOperand(OffsetPosLd).setImm(LoadOffset + StoreOffset);
2221   bool Disjoint = TII->areMemAccessesTriviallyDisjoint(*NewMI, *PrevDef);
2222   MF.DeleteMachineInstr(NewMI);
2223   if (!Disjoint)
2224     return false;
2225 
2226   // Set the return value once we determine that we return true.
2227   BasePos = BasePosLd;
2228   OffsetPos = OffsetPosLd;
2229   NewBase = PrevReg;
2230   Offset = StoreOffset;
2231   return true;
2232 }
2233 
2234 /// Apply changes to the instruction if needed. The changes are need
2235 /// to improve the scheduling and depend up on the final schedule.
2236 void SwingSchedulerDAG::applyInstrChange(MachineInstr *MI,
2237                                          SMSchedule &Schedule) {
2238   SUnit *SU = getSUnit(MI);
2239   DenseMap<SUnit *, std::pair<unsigned, int64_t>>::iterator It =
2240       InstrChanges.find(SU);
2241   if (It != InstrChanges.end()) {
2242     std::pair<unsigned, int64_t> RegAndOffset = It->second;
2243     unsigned BasePos, OffsetPos;
2244     if (!TII->getBaseAndOffsetPosition(*MI, BasePos, OffsetPos))
2245       return;
2246     Register BaseReg = MI->getOperand(BasePos).getReg();
2247     MachineInstr *LoopDef = findDefInLoop(BaseReg);
2248     int DefStageNum = Schedule.stageScheduled(getSUnit(LoopDef));
2249     int DefCycleNum = Schedule.cycleScheduled(getSUnit(LoopDef));
2250     int BaseStageNum = Schedule.stageScheduled(SU);
2251     int BaseCycleNum = Schedule.cycleScheduled(SU);
2252     if (BaseStageNum < DefStageNum) {
2253       MachineInstr *NewMI = MF.CloneMachineInstr(MI);
2254       int OffsetDiff = DefStageNum - BaseStageNum;
2255       if (DefCycleNum < BaseCycleNum) {
2256         NewMI->getOperand(BasePos).setReg(RegAndOffset.first);
2257         if (OffsetDiff > 0)
2258           --OffsetDiff;
2259       }
2260       int64_t NewOffset =
2261           MI->getOperand(OffsetPos).getImm() + RegAndOffset.second * OffsetDiff;
2262       NewMI->getOperand(OffsetPos).setImm(NewOffset);
2263       SU->setInstr(NewMI);
2264       MISUnitMap[NewMI] = SU;
2265       NewMIs[MI] = NewMI;
2266     }
2267   }
2268 }
2269 
2270 /// Return the instruction in the loop that defines the register.
2271 /// If the definition is a Phi, then follow the Phi operand to
2272 /// the instruction in the loop.
2273 MachineInstr *SwingSchedulerDAG::findDefInLoop(unsigned Reg) {
2274   SmallPtrSet<MachineInstr *, 8> Visited;
2275   MachineInstr *Def = MRI.getVRegDef(Reg);
2276   while (Def->isPHI()) {
2277     if (!Visited.insert(Def).second)
2278       break;
2279     for (unsigned i = 1, e = Def->getNumOperands(); i < e; i += 2)
2280       if (Def->getOperand(i + 1).getMBB() == BB) {
2281         Def = MRI.getVRegDef(Def->getOperand(i).getReg());
2282         break;
2283       }
2284   }
2285   return Def;
2286 }
2287 
2288 /// Return true for an order or output dependence that is loop carried
2289 /// potentially. A dependence is loop carried if the destination defines a valu
2290 /// that may be used or defined by the source in a subsequent iteration.
2291 bool SwingSchedulerDAG::isLoopCarriedDep(SUnit *Source, const SDep &Dep,
2292                                          bool isSucc) {
2293   if ((Dep.getKind() != SDep::Order && Dep.getKind() != SDep::Output) ||
2294       Dep.isArtificial())
2295     return false;
2296 
2297   if (!SwpPruneLoopCarried)
2298     return true;
2299 
2300   if (Dep.getKind() == SDep::Output)
2301     return true;
2302 
2303   MachineInstr *SI = Source->getInstr();
2304   MachineInstr *DI = Dep.getSUnit()->getInstr();
2305   if (!isSucc)
2306     std::swap(SI, DI);
2307   assert(SI != nullptr && DI != nullptr && "Expecting SUnit with an MI.");
2308 
2309   // Assume ordered loads and stores may have a loop carried dependence.
2310   if (SI->hasUnmodeledSideEffects() || DI->hasUnmodeledSideEffects() ||
2311       SI->mayRaiseFPException() || DI->mayRaiseFPException() ||
2312       SI->hasOrderedMemoryRef() || DI->hasOrderedMemoryRef())
2313     return true;
2314 
2315   // Only chain dependences between a load and store can be loop carried.
2316   if (!DI->mayStore() || !SI->mayLoad())
2317     return false;
2318 
2319   unsigned DeltaS, DeltaD;
2320   if (!computeDelta(*SI, DeltaS) || !computeDelta(*DI, DeltaD))
2321     return true;
2322 
2323   const MachineOperand *BaseOpS, *BaseOpD;
2324   int64_t OffsetS, OffsetD;
2325   bool OffsetSIsScalable, OffsetDIsScalable;
2326   const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
2327   if (!TII->getMemOperandWithOffset(*SI, BaseOpS, OffsetS, OffsetSIsScalable,
2328                                     TRI) ||
2329       !TII->getMemOperandWithOffset(*DI, BaseOpD, OffsetD, OffsetDIsScalable,
2330                                     TRI))
2331     return true;
2332 
2333   assert(!OffsetSIsScalable && !OffsetDIsScalable &&
2334          "Expected offsets to be byte offsets");
2335 
2336   if (!BaseOpS->isIdenticalTo(*BaseOpD))
2337     return true;
2338 
2339   // Check that the base register is incremented by a constant value for each
2340   // iteration.
2341   MachineInstr *Def = MRI.getVRegDef(BaseOpS->getReg());
2342   if (!Def || !Def->isPHI())
2343     return true;
2344   unsigned InitVal = 0;
2345   unsigned LoopVal = 0;
2346   getPhiRegs(*Def, BB, InitVal, LoopVal);
2347   MachineInstr *LoopDef = MRI.getVRegDef(LoopVal);
2348   int D = 0;
2349   if (!LoopDef || !TII->getIncrementValue(*LoopDef, D))
2350     return true;
2351 
2352   uint64_t AccessSizeS = (*SI->memoperands_begin())->getSize();
2353   uint64_t AccessSizeD = (*DI->memoperands_begin())->getSize();
2354 
2355   // This is the main test, which checks the offset values and the loop
2356   // increment value to determine if the accesses may be loop carried.
2357   if (AccessSizeS == MemoryLocation::UnknownSize ||
2358       AccessSizeD == MemoryLocation::UnknownSize)
2359     return true;
2360 
2361   if (DeltaS != DeltaD || DeltaS < AccessSizeS || DeltaD < AccessSizeD)
2362     return true;
2363 
2364   return (OffsetS + (int64_t)AccessSizeS < OffsetD + (int64_t)AccessSizeD);
2365 }
2366 
2367 void SwingSchedulerDAG::postprocessDAG() {
2368   for (auto &M : Mutations)
2369     M->apply(this);
2370 }
2371 
2372 /// Try to schedule the node at the specified StartCycle and continue
2373 /// until the node is schedule or the EndCycle is reached.  This function
2374 /// returns true if the node is scheduled.  This routine may search either
2375 /// forward or backward for a place to insert the instruction based upon
2376 /// the relative values of StartCycle and EndCycle.
2377 bool SMSchedule::insert(SUnit *SU, int StartCycle, int EndCycle, int II) {
2378   bool forward = true;
2379   LLVM_DEBUG({
2380     dbgs() << "Trying to insert node between " << StartCycle << " and "
2381            << EndCycle << " II: " << II << "\n";
2382   });
2383   if (StartCycle > EndCycle)
2384     forward = false;
2385 
2386   // The terminating condition depends on the direction.
2387   int termCycle = forward ? EndCycle + 1 : EndCycle - 1;
2388   for (int curCycle = StartCycle; curCycle != termCycle;
2389        forward ? ++curCycle : --curCycle) {
2390 
2391     // Add the already scheduled instructions at the specified cycle to the
2392     // DFA.
2393     ProcItinResources.clearResources();
2394     for (int checkCycle = FirstCycle + ((curCycle - FirstCycle) % II);
2395          checkCycle <= LastCycle; checkCycle += II) {
2396       std::deque<SUnit *> &cycleInstrs = ScheduledInstrs[checkCycle];
2397 
2398       for (std::deque<SUnit *>::iterator I = cycleInstrs.begin(),
2399                                          E = cycleInstrs.end();
2400            I != E; ++I) {
2401         if (ST.getInstrInfo()->isZeroCost((*I)->getInstr()->getOpcode()))
2402           continue;
2403         assert(ProcItinResources.canReserveResources(*(*I)->getInstr()) &&
2404                "These instructions have already been scheduled.");
2405         ProcItinResources.reserveResources(*(*I)->getInstr());
2406       }
2407     }
2408     if (ST.getInstrInfo()->isZeroCost(SU->getInstr()->getOpcode()) ||
2409         ProcItinResources.canReserveResources(*SU->getInstr())) {
2410       LLVM_DEBUG({
2411         dbgs() << "\tinsert at cycle " << curCycle << " ";
2412         SU->getInstr()->dump();
2413       });
2414 
2415       ScheduledInstrs[curCycle].push_back(SU);
2416       InstrToCycle.insert(std::make_pair(SU, curCycle));
2417       if (curCycle > LastCycle)
2418         LastCycle = curCycle;
2419       if (curCycle < FirstCycle)
2420         FirstCycle = curCycle;
2421       return true;
2422     }
2423     LLVM_DEBUG({
2424       dbgs() << "\tfailed to insert at cycle " << curCycle << " ";
2425       SU->getInstr()->dump();
2426     });
2427   }
2428   return false;
2429 }
2430 
2431 // Return the cycle of the earliest scheduled instruction in the chain.
2432 int SMSchedule::earliestCycleInChain(const SDep &Dep) {
2433   SmallPtrSet<SUnit *, 8> Visited;
2434   SmallVector<SDep, 8> Worklist;
2435   Worklist.push_back(Dep);
2436   int EarlyCycle = INT_MAX;
2437   while (!Worklist.empty()) {
2438     const SDep &Cur = Worklist.pop_back_val();
2439     SUnit *PrevSU = Cur.getSUnit();
2440     if (Visited.count(PrevSU))
2441       continue;
2442     std::map<SUnit *, int>::const_iterator it = InstrToCycle.find(PrevSU);
2443     if (it == InstrToCycle.end())
2444       continue;
2445     EarlyCycle = std::min(EarlyCycle, it->second);
2446     for (const auto &PI : PrevSU->Preds)
2447       if (PI.getKind() == SDep::Order || PI.getKind() == SDep::Output)
2448         Worklist.push_back(PI);
2449     Visited.insert(PrevSU);
2450   }
2451   return EarlyCycle;
2452 }
2453 
2454 // Return the cycle of the latest scheduled instruction in the chain.
2455 int SMSchedule::latestCycleInChain(const SDep &Dep) {
2456   SmallPtrSet<SUnit *, 8> Visited;
2457   SmallVector<SDep, 8> Worklist;
2458   Worklist.push_back(Dep);
2459   int LateCycle = INT_MIN;
2460   while (!Worklist.empty()) {
2461     const SDep &Cur = Worklist.pop_back_val();
2462     SUnit *SuccSU = Cur.getSUnit();
2463     if (Visited.count(SuccSU))
2464       continue;
2465     std::map<SUnit *, int>::const_iterator it = InstrToCycle.find(SuccSU);
2466     if (it == InstrToCycle.end())
2467       continue;
2468     LateCycle = std::max(LateCycle, it->second);
2469     for (const auto &SI : SuccSU->Succs)
2470       if (SI.getKind() == SDep::Order || SI.getKind() == SDep::Output)
2471         Worklist.push_back(SI);
2472     Visited.insert(SuccSU);
2473   }
2474   return LateCycle;
2475 }
2476 
2477 /// If an instruction has a use that spans multiple iterations, then
2478 /// return true. These instructions are characterized by having a back-ege
2479 /// to a Phi, which contains a reference to another Phi.
2480 static SUnit *multipleIterations(SUnit *SU, SwingSchedulerDAG *DAG) {
2481   for (auto &P : SU->Preds)
2482     if (DAG->isBackedge(SU, P) && P.getSUnit()->getInstr()->isPHI())
2483       for (auto &S : P.getSUnit()->Succs)
2484         if (S.getKind() == SDep::Data && S.getSUnit()->getInstr()->isPHI())
2485           return P.getSUnit();
2486   return nullptr;
2487 }
2488 
2489 /// Compute the scheduling start slot for the instruction.  The start slot
2490 /// depends on any predecessor or successor nodes scheduled already.
2491 void SMSchedule::computeStart(SUnit *SU, int *MaxEarlyStart, int *MinLateStart,
2492                               int *MinEnd, int *MaxStart, int II,
2493                               SwingSchedulerDAG *DAG) {
2494   // Iterate over each instruction that has been scheduled already.  The start
2495   // slot computation depends on whether the previously scheduled instruction
2496   // is a predecessor or successor of the specified instruction.
2497   for (int cycle = getFirstCycle(); cycle <= LastCycle; ++cycle) {
2498 
2499     // Iterate over each instruction in the current cycle.
2500     for (SUnit *I : getInstructions(cycle)) {
2501       // Because we're processing a DAG for the dependences, we recognize
2502       // the back-edge in recurrences by anti dependences.
2503       for (unsigned i = 0, e = (unsigned)SU->Preds.size(); i != e; ++i) {
2504         const SDep &Dep = SU->Preds[i];
2505         if (Dep.getSUnit() == I) {
2506           if (!DAG->isBackedge(SU, Dep)) {
2507             int EarlyStart = cycle + Dep.getLatency() -
2508                              DAG->getDistance(Dep.getSUnit(), SU, Dep) * II;
2509             *MaxEarlyStart = std::max(*MaxEarlyStart, EarlyStart);
2510             if (DAG->isLoopCarriedDep(SU, Dep, false)) {
2511               int End = earliestCycleInChain(Dep) + (II - 1);
2512               *MinEnd = std::min(*MinEnd, End);
2513             }
2514           } else {
2515             int LateStart = cycle - Dep.getLatency() +
2516                             DAG->getDistance(SU, Dep.getSUnit(), Dep) * II;
2517             *MinLateStart = std::min(*MinLateStart, LateStart);
2518           }
2519         }
2520         // For instruction that requires multiple iterations, make sure that
2521         // the dependent instruction is not scheduled past the definition.
2522         SUnit *BE = multipleIterations(I, DAG);
2523         if (BE && Dep.getSUnit() == BE && !SU->getInstr()->isPHI() &&
2524             !SU->isPred(I))
2525           *MinLateStart = std::min(*MinLateStart, cycle);
2526       }
2527       for (unsigned i = 0, e = (unsigned)SU->Succs.size(); i != e; ++i) {
2528         if (SU->Succs[i].getSUnit() == I) {
2529           const SDep &Dep = SU->Succs[i];
2530           if (!DAG->isBackedge(SU, Dep)) {
2531             int LateStart = cycle - Dep.getLatency() +
2532                             DAG->getDistance(SU, Dep.getSUnit(), Dep) * II;
2533             *MinLateStart = std::min(*MinLateStart, LateStart);
2534             if (DAG->isLoopCarriedDep(SU, Dep)) {
2535               int Start = latestCycleInChain(Dep) + 1 - II;
2536               *MaxStart = std::max(*MaxStart, Start);
2537             }
2538           } else {
2539             int EarlyStart = cycle + Dep.getLatency() -
2540                              DAG->getDistance(Dep.getSUnit(), SU, Dep) * II;
2541             *MaxEarlyStart = std::max(*MaxEarlyStart, EarlyStart);
2542           }
2543         }
2544       }
2545     }
2546   }
2547 }
2548 
2549 /// Order the instructions within a cycle so that the definitions occur
2550 /// before the uses. Returns true if the instruction is added to the start
2551 /// of the list, or false if added to the end.
2552 void SMSchedule::orderDependence(SwingSchedulerDAG *SSD, SUnit *SU,
2553                                  std::deque<SUnit *> &Insts) {
2554   MachineInstr *MI = SU->getInstr();
2555   bool OrderBeforeUse = false;
2556   bool OrderAfterDef = false;
2557   bool OrderBeforeDef = false;
2558   unsigned MoveDef = 0;
2559   unsigned MoveUse = 0;
2560   int StageInst1 = stageScheduled(SU);
2561 
2562   unsigned Pos = 0;
2563   for (std::deque<SUnit *>::iterator I = Insts.begin(), E = Insts.end(); I != E;
2564        ++I, ++Pos) {
2565     for (unsigned i = 0, e = MI->getNumOperands(); i < e; ++i) {
2566       MachineOperand &MO = MI->getOperand(i);
2567       if (!MO.isReg() || !Register::isVirtualRegister(MO.getReg()))
2568         continue;
2569 
2570       Register Reg = MO.getReg();
2571       unsigned BasePos, OffsetPos;
2572       if (ST.getInstrInfo()->getBaseAndOffsetPosition(*MI, BasePos, OffsetPos))
2573         if (MI->getOperand(BasePos).getReg() == Reg)
2574           if (unsigned NewReg = SSD->getInstrBaseReg(SU))
2575             Reg = NewReg;
2576       bool Reads, Writes;
2577       std::tie(Reads, Writes) =
2578           (*I)->getInstr()->readsWritesVirtualRegister(Reg);
2579       if (MO.isDef() && Reads && stageScheduled(*I) <= StageInst1) {
2580         OrderBeforeUse = true;
2581         if (MoveUse == 0)
2582           MoveUse = Pos;
2583       } else if (MO.isDef() && Reads && stageScheduled(*I) > StageInst1) {
2584         // Add the instruction after the scheduled instruction.
2585         OrderAfterDef = true;
2586         MoveDef = Pos;
2587       } else if (MO.isUse() && Writes && stageScheduled(*I) == StageInst1) {
2588         if (cycleScheduled(*I) == cycleScheduled(SU) && !(*I)->isSucc(SU)) {
2589           OrderBeforeUse = true;
2590           if (MoveUse == 0)
2591             MoveUse = Pos;
2592         } else {
2593           OrderAfterDef = true;
2594           MoveDef = Pos;
2595         }
2596       } else if (MO.isUse() && Writes && stageScheduled(*I) > StageInst1) {
2597         OrderBeforeUse = true;
2598         if (MoveUse == 0)
2599           MoveUse = Pos;
2600         if (MoveUse != 0) {
2601           OrderAfterDef = true;
2602           MoveDef = Pos - 1;
2603         }
2604       } else if (MO.isUse() && Writes && stageScheduled(*I) < StageInst1) {
2605         // Add the instruction before the scheduled instruction.
2606         OrderBeforeUse = true;
2607         if (MoveUse == 0)
2608           MoveUse = Pos;
2609       } else if (MO.isUse() && stageScheduled(*I) == StageInst1 &&
2610                  isLoopCarriedDefOfUse(SSD, (*I)->getInstr(), MO)) {
2611         if (MoveUse == 0) {
2612           OrderBeforeDef = true;
2613           MoveUse = Pos;
2614         }
2615       }
2616     }
2617     // Check for order dependences between instructions. Make sure the source
2618     // is ordered before the destination.
2619     for (auto &S : SU->Succs) {
2620       if (S.getSUnit() != *I)
2621         continue;
2622       if (S.getKind() == SDep::Order && stageScheduled(*I) == StageInst1) {
2623         OrderBeforeUse = true;
2624         if (Pos < MoveUse)
2625           MoveUse = Pos;
2626       }
2627       // We did not handle HW dependences in previous for loop,
2628       // and we normally set Latency = 0 for Anti deps,
2629       // so may have nodes in same cycle with Anti denpendent on HW regs.
2630       else if (S.getKind() == SDep::Anti && stageScheduled(*I) == StageInst1) {
2631         OrderBeforeUse = true;
2632         if ((MoveUse == 0) || (Pos < MoveUse))
2633           MoveUse = Pos;
2634       }
2635     }
2636     for (auto &P : SU->Preds) {
2637       if (P.getSUnit() != *I)
2638         continue;
2639       if (P.getKind() == SDep::Order && stageScheduled(*I) == StageInst1) {
2640         OrderAfterDef = true;
2641         MoveDef = Pos;
2642       }
2643     }
2644   }
2645 
2646   // A circular dependence.
2647   if (OrderAfterDef && OrderBeforeUse && MoveUse == MoveDef)
2648     OrderBeforeUse = false;
2649 
2650   // OrderAfterDef takes precedences over OrderBeforeDef. The latter is due
2651   // to a loop-carried dependence.
2652   if (OrderBeforeDef)
2653     OrderBeforeUse = !OrderAfterDef || (MoveUse > MoveDef);
2654 
2655   // The uncommon case when the instruction order needs to be updated because
2656   // there is both a use and def.
2657   if (OrderBeforeUse && OrderAfterDef) {
2658     SUnit *UseSU = Insts.at(MoveUse);
2659     SUnit *DefSU = Insts.at(MoveDef);
2660     if (MoveUse > MoveDef) {
2661       Insts.erase(Insts.begin() + MoveUse);
2662       Insts.erase(Insts.begin() + MoveDef);
2663     } else {
2664       Insts.erase(Insts.begin() + MoveDef);
2665       Insts.erase(Insts.begin() + MoveUse);
2666     }
2667     orderDependence(SSD, UseSU, Insts);
2668     orderDependence(SSD, SU, Insts);
2669     orderDependence(SSD, DefSU, Insts);
2670     return;
2671   }
2672   // Put the new instruction first if there is a use in the list. Otherwise,
2673   // put it at the end of the list.
2674   if (OrderBeforeUse)
2675     Insts.push_front(SU);
2676   else
2677     Insts.push_back(SU);
2678 }
2679 
2680 /// Return true if the scheduled Phi has a loop carried operand.
2681 bool SMSchedule::isLoopCarried(SwingSchedulerDAG *SSD, MachineInstr &Phi) {
2682   if (!Phi.isPHI())
2683     return false;
2684   assert(Phi.isPHI() && "Expecting a Phi.");
2685   SUnit *DefSU = SSD->getSUnit(&Phi);
2686   unsigned DefCycle = cycleScheduled(DefSU);
2687   int DefStage = stageScheduled(DefSU);
2688 
2689   unsigned InitVal = 0;
2690   unsigned LoopVal = 0;
2691   getPhiRegs(Phi, Phi.getParent(), InitVal, LoopVal);
2692   SUnit *UseSU = SSD->getSUnit(MRI.getVRegDef(LoopVal));
2693   if (!UseSU)
2694     return true;
2695   if (UseSU->getInstr()->isPHI())
2696     return true;
2697   unsigned LoopCycle = cycleScheduled(UseSU);
2698   int LoopStage = stageScheduled(UseSU);
2699   return (LoopCycle > DefCycle) || (LoopStage <= DefStage);
2700 }
2701 
2702 /// Return true if the instruction is a definition that is loop carried
2703 /// and defines the use on the next iteration.
2704 ///        v1 = phi(v2, v3)
2705 ///  (Def) v3 = op v1
2706 ///  (MO)   = v1
2707 /// If MO appears before Def, then then v1 and v3 may get assigned to the same
2708 /// register.
2709 bool SMSchedule::isLoopCarriedDefOfUse(SwingSchedulerDAG *SSD,
2710                                        MachineInstr *Def, MachineOperand &MO) {
2711   if (!MO.isReg())
2712     return false;
2713   if (Def->isPHI())
2714     return false;
2715   MachineInstr *Phi = MRI.getVRegDef(MO.getReg());
2716   if (!Phi || !Phi->isPHI() || Phi->getParent() != Def->getParent())
2717     return false;
2718   if (!isLoopCarried(SSD, *Phi))
2719     return false;
2720   unsigned LoopReg = getLoopPhiReg(*Phi, Phi->getParent());
2721   for (unsigned i = 0, e = Def->getNumOperands(); i != e; ++i) {
2722     MachineOperand &DMO = Def->getOperand(i);
2723     if (!DMO.isReg() || !DMO.isDef())
2724       continue;
2725     if (DMO.getReg() == LoopReg)
2726       return true;
2727   }
2728   return false;
2729 }
2730 
2731 // Check if the generated schedule is valid. This function checks if
2732 // an instruction that uses a physical register is scheduled in a
2733 // different stage than the definition. The pipeliner does not handle
2734 // physical register values that may cross a basic block boundary.
2735 bool SMSchedule::isValidSchedule(SwingSchedulerDAG *SSD) {
2736   for (int i = 0, e = SSD->SUnits.size(); i < e; ++i) {
2737     SUnit &SU = SSD->SUnits[i];
2738     if (!SU.hasPhysRegDefs)
2739       continue;
2740     int StageDef = stageScheduled(&SU);
2741     assert(StageDef != -1 && "Instruction should have been scheduled.");
2742     for (auto &SI : SU.Succs)
2743       if (SI.isAssignedRegDep())
2744         if (Register::isPhysicalRegister(SI.getReg()))
2745           if (stageScheduled(SI.getSUnit()) != StageDef)
2746             return false;
2747   }
2748   return true;
2749 }
2750 
2751 /// A property of the node order in swing-modulo-scheduling is
2752 /// that for nodes outside circuits the following holds:
2753 /// none of them is scheduled after both a successor and a
2754 /// predecessor.
2755 /// The method below checks whether the property is met.
2756 /// If not, debug information is printed and statistics information updated.
2757 /// Note that we do not use an assert statement.
2758 /// The reason is that although an invalid node oder may prevent
2759 /// the pipeliner from finding a pipelined schedule for arbitrary II,
2760 /// it does not lead to the generation of incorrect code.
2761 void SwingSchedulerDAG::checkValidNodeOrder(const NodeSetType &Circuits) const {
2762 
2763   // a sorted vector that maps each SUnit to its index in the NodeOrder
2764   typedef std::pair<SUnit *, unsigned> UnitIndex;
2765   std::vector<UnitIndex> Indices(NodeOrder.size(), std::make_pair(nullptr, 0));
2766 
2767   for (unsigned i = 0, s = NodeOrder.size(); i < s; ++i)
2768     Indices.push_back(std::make_pair(NodeOrder[i], i));
2769 
2770   auto CompareKey = [](UnitIndex i1, UnitIndex i2) {
2771     return std::get<0>(i1) < std::get<0>(i2);
2772   };
2773 
2774   // sort, so that we can perform a binary search
2775   llvm::sort(Indices, CompareKey);
2776 
2777   bool Valid = true;
2778   (void)Valid;
2779   // for each SUnit in the NodeOrder, check whether
2780   // it appears after both a successor and a predecessor
2781   // of the SUnit. If this is the case, and the SUnit
2782   // is not part of circuit, then the NodeOrder is not
2783   // valid.
2784   for (unsigned i = 0, s = NodeOrder.size(); i < s; ++i) {
2785     SUnit *SU = NodeOrder[i];
2786     unsigned Index = i;
2787 
2788     bool PredBefore = false;
2789     bool SuccBefore = false;
2790 
2791     SUnit *Succ;
2792     SUnit *Pred;
2793     (void)Succ;
2794     (void)Pred;
2795 
2796     for (SDep &PredEdge : SU->Preds) {
2797       SUnit *PredSU = PredEdge.getSUnit();
2798       unsigned PredIndex = std::get<1>(
2799           *llvm::lower_bound(Indices, std::make_pair(PredSU, 0), CompareKey));
2800       if (!PredSU->getInstr()->isPHI() && PredIndex < Index) {
2801         PredBefore = true;
2802         Pred = PredSU;
2803         break;
2804       }
2805     }
2806 
2807     for (SDep &SuccEdge : SU->Succs) {
2808       SUnit *SuccSU = SuccEdge.getSUnit();
2809       // Do not process a boundary node, it was not included in NodeOrder,
2810       // hence not in Indices either, call to std::lower_bound() below will
2811       // return Indices.end().
2812       if (SuccSU->isBoundaryNode())
2813         continue;
2814       unsigned SuccIndex = std::get<1>(
2815           *llvm::lower_bound(Indices, std::make_pair(SuccSU, 0), CompareKey));
2816       if (!SuccSU->getInstr()->isPHI() && SuccIndex < Index) {
2817         SuccBefore = true;
2818         Succ = SuccSU;
2819         break;
2820       }
2821     }
2822 
2823     if (PredBefore && SuccBefore && !SU->getInstr()->isPHI()) {
2824       // instructions in circuits are allowed to be scheduled
2825       // after both a successor and predecessor.
2826       bool InCircuit = llvm::any_of(
2827           Circuits, [SU](const NodeSet &Circuit) { return Circuit.count(SU); });
2828       if (InCircuit)
2829         LLVM_DEBUG(dbgs() << "In a circuit, predecessor ";);
2830       else {
2831         Valid = false;
2832         NumNodeOrderIssues++;
2833         LLVM_DEBUG(dbgs() << "Predecessor ";);
2834       }
2835       LLVM_DEBUG(dbgs() << Pred->NodeNum << " and successor " << Succ->NodeNum
2836                         << " are scheduled before node " << SU->NodeNum
2837                         << "\n";);
2838     }
2839   }
2840 
2841   LLVM_DEBUG({
2842     if (!Valid)
2843       dbgs() << "Invalid node order found!\n";
2844   });
2845 }
2846 
2847 /// Attempt to fix the degenerate cases when the instruction serialization
2848 /// causes the register lifetimes to overlap. For example,
2849 ///   p' = store_pi(p, b)
2850 ///      = load p, offset
2851 /// In this case p and p' overlap, which means that two registers are needed.
2852 /// Instead, this function changes the load to use p' and updates the offset.
2853 void SwingSchedulerDAG::fixupRegisterOverlaps(std::deque<SUnit *> &Instrs) {
2854   unsigned OverlapReg = 0;
2855   unsigned NewBaseReg = 0;
2856   for (SUnit *SU : Instrs) {
2857     MachineInstr *MI = SU->getInstr();
2858     for (unsigned i = 0, e = MI->getNumOperands(); i < e; ++i) {
2859       const MachineOperand &MO = MI->getOperand(i);
2860       // Look for an instruction that uses p. The instruction occurs in the
2861       // same cycle but occurs later in the serialized order.
2862       if (MO.isReg() && MO.isUse() && MO.getReg() == OverlapReg) {
2863         // Check that the instruction appears in the InstrChanges structure,
2864         // which contains instructions that can have the offset updated.
2865         DenseMap<SUnit *, std::pair<unsigned, int64_t>>::iterator It =
2866           InstrChanges.find(SU);
2867         if (It != InstrChanges.end()) {
2868           unsigned BasePos, OffsetPos;
2869           // Update the base register and adjust the offset.
2870           if (TII->getBaseAndOffsetPosition(*MI, BasePos, OffsetPos)) {
2871             MachineInstr *NewMI = MF.CloneMachineInstr(MI);
2872             NewMI->getOperand(BasePos).setReg(NewBaseReg);
2873             int64_t NewOffset =
2874                 MI->getOperand(OffsetPos).getImm() - It->second.second;
2875             NewMI->getOperand(OffsetPos).setImm(NewOffset);
2876             SU->setInstr(NewMI);
2877             MISUnitMap[NewMI] = SU;
2878             NewMIs[MI] = NewMI;
2879           }
2880         }
2881         OverlapReg = 0;
2882         NewBaseReg = 0;
2883         break;
2884       }
2885       // Look for an instruction of the form p' = op(p), which uses and defines
2886       // two virtual registers that get allocated to the same physical register.
2887       unsigned TiedUseIdx = 0;
2888       if (MI->isRegTiedToUseOperand(i, &TiedUseIdx)) {
2889         // OverlapReg is p in the example above.
2890         OverlapReg = MI->getOperand(TiedUseIdx).getReg();
2891         // NewBaseReg is p' in the example above.
2892         NewBaseReg = MI->getOperand(i).getReg();
2893         break;
2894       }
2895     }
2896   }
2897 }
2898 
2899 /// After the schedule has been formed, call this function to combine
2900 /// the instructions from the different stages/cycles.  That is, this
2901 /// function creates a schedule that represents a single iteration.
2902 void SMSchedule::finalizeSchedule(SwingSchedulerDAG *SSD) {
2903   // Move all instructions to the first stage from later stages.
2904   for (int cycle = getFirstCycle(); cycle <= getFinalCycle(); ++cycle) {
2905     for (int stage = 1, lastStage = getMaxStageCount(); stage <= lastStage;
2906          ++stage) {
2907       std::deque<SUnit *> &cycleInstrs =
2908           ScheduledInstrs[cycle + (stage * InitiationInterval)];
2909       for (std::deque<SUnit *>::reverse_iterator I = cycleInstrs.rbegin(),
2910                                                  E = cycleInstrs.rend();
2911            I != E; ++I)
2912         ScheduledInstrs[cycle].push_front(*I);
2913     }
2914   }
2915 
2916   // Erase all the elements in the later stages. Only one iteration should
2917   // remain in the scheduled list, and it contains all the instructions.
2918   for (int cycle = getFinalCycle() + 1; cycle <= LastCycle; ++cycle)
2919     ScheduledInstrs.erase(cycle);
2920 
2921   // Change the registers in instruction as specified in the InstrChanges
2922   // map. We need to use the new registers to create the correct order.
2923   for (int i = 0, e = SSD->SUnits.size(); i != e; ++i) {
2924     SUnit *SU = &SSD->SUnits[i];
2925     SSD->applyInstrChange(SU->getInstr(), *this);
2926   }
2927 
2928   // Reorder the instructions in each cycle to fix and improve the
2929   // generated code.
2930   for (int Cycle = getFirstCycle(), E = getFinalCycle(); Cycle <= E; ++Cycle) {
2931     std::deque<SUnit *> &cycleInstrs = ScheduledInstrs[Cycle];
2932     std::deque<SUnit *> newOrderPhi;
2933     for (unsigned i = 0, e = cycleInstrs.size(); i < e; ++i) {
2934       SUnit *SU = cycleInstrs[i];
2935       if (SU->getInstr()->isPHI())
2936         newOrderPhi.push_back(SU);
2937     }
2938     std::deque<SUnit *> newOrderI;
2939     for (unsigned i = 0, e = cycleInstrs.size(); i < e; ++i) {
2940       SUnit *SU = cycleInstrs[i];
2941       if (!SU->getInstr()->isPHI())
2942         orderDependence(SSD, SU, newOrderI);
2943     }
2944     // Replace the old order with the new order.
2945     cycleInstrs.swap(newOrderPhi);
2946     cycleInstrs.insert(cycleInstrs.end(), newOrderI.begin(), newOrderI.end());
2947     SSD->fixupRegisterOverlaps(cycleInstrs);
2948   }
2949 
2950   LLVM_DEBUG(dump(););
2951 }
2952 
2953 void NodeSet::print(raw_ostream &os) const {
2954   os << "Num nodes " << size() << " rec " << RecMII << " mov " << MaxMOV
2955      << " depth " << MaxDepth << " col " << Colocate << "\n";
2956   for (const auto &I : Nodes)
2957     os << "   SU(" << I->NodeNum << ") " << *(I->getInstr());
2958   os << "\n";
2959 }
2960 
2961 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
2962 /// Print the schedule information to the given output.
2963 void SMSchedule::print(raw_ostream &os) const {
2964   // Iterate over each cycle.
2965   for (int cycle = getFirstCycle(); cycle <= getFinalCycle(); ++cycle) {
2966     // Iterate over each instruction in the cycle.
2967     const_sched_iterator cycleInstrs = ScheduledInstrs.find(cycle);
2968     for (SUnit *CI : cycleInstrs->second) {
2969       os << "cycle " << cycle << " (" << stageScheduled(CI) << ") ";
2970       os << "(" << CI->NodeNum << ") ";
2971       CI->getInstr()->print(os);
2972       os << "\n";
2973     }
2974   }
2975 }
2976 
2977 /// Utility function used for debugging to print the schedule.
2978 LLVM_DUMP_METHOD void SMSchedule::dump() const { print(dbgs()); }
2979 LLVM_DUMP_METHOD void NodeSet::dump() const { print(dbgs()); }
2980 
2981 #endif
2982 
2983 void ResourceManager::initProcResourceVectors(
2984     const MCSchedModel &SM, SmallVectorImpl<uint64_t> &Masks) {
2985   unsigned ProcResourceID = 0;
2986 
2987   // We currently limit the resource kinds to 64 and below so that we can use
2988   // uint64_t for Masks
2989   assert(SM.getNumProcResourceKinds() < 64 &&
2990          "Too many kinds of resources, unsupported");
2991   // Create a unique bitmask for every processor resource unit.
2992   // Skip resource at index 0, since it always references 'InvalidUnit'.
2993   Masks.resize(SM.getNumProcResourceKinds());
2994   for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) {
2995     const MCProcResourceDesc &Desc = *SM.getProcResource(I);
2996     if (Desc.SubUnitsIdxBegin)
2997       continue;
2998     Masks[I] = 1ULL << ProcResourceID;
2999     ProcResourceID++;
3000   }
3001   // Create a unique bitmask for every processor resource group.
3002   for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) {
3003     const MCProcResourceDesc &Desc = *SM.getProcResource(I);
3004     if (!Desc.SubUnitsIdxBegin)
3005       continue;
3006     Masks[I] = 1ULL << ProcResourceID;
3007     for (unsigned U = 0; U < Desc.NumUnits; ++U)
3008       Masks[I] |= Masks[Desc.SubUnitsIdxBegin[U]];
3009     ProcResourceID++;
3010   }
3011   LLVM_DEBUG({
3012     if (SwpShowResMask) {
3013       dbgs() << "ProcResourceDesc:\n";
3014       for (unsigned I = 1, E = SM.getNumProcResourceKinds(); I < E; ++I) {
3015         const MCProcResourceDesc *ProcResource = SM.getProcResource(I);
3016         dbgs() << format(" %16s(%2d): Mask: 0x%08x, NumUnits:%2d\n",
3017                          ProcResource->Name, I, Masks[I],
3018                          ProcResource->NumUnits);
3019       }
3020       dbgs() << " -----------------\n";
3021     }
3022   });
3023 }
3024 
3025 bool ResourceManager::canReserveResources(const MCInstrDesc *MID) const {
3026 
3027   LLVM_DEBUG({
3028     if (SwpDebugResource)
3029       dbgs() << "canReserveResources:\n";
3030   });
3031   if (UseDFA)
3032     return DFAResources->canReserveResources(MID);
3033 
3034   unsigned InsnClass = MID->getSchedClass();
3035   const MCSchedClassDesc *SCDesc = SM.getSchedClassDesc(InsnClass);
3036   if (!SCDesc->isValid()) {
3037     LLVM_DEBUG({
3038       dbgs() << "No valid Schedule Class Desc for schedClass!\n";
3039       dbgs() << "isPseduo:" << MID->isPseudo() << "\n";
3040     });
3041     return true;
3042   }
3043 
3044   const MCWriteProcResEntry *I = STI->getWriteProcResBegin(SCDesc);
3045   const MCWriteProcResEntry *E = STI->getWriteProcResEnd(SCDesc);
3046   for (; I != E; ++I) {
3047     if (!I->Cycles)
3048       continue;
3049     const MCProcResourceDesc *ProcResource =
3050         SM.getProcResource(I->ProcResourceIdx);
3051     unsigned NumUnits = ProcResource->NumUnits;
3052     LLVM_DEBUG({
3053       if (SwpDebugResource)
3054         dbgs() << format(" %16s(%2d): Count: %2d, NumUnits:%2d, Cycles:%2d\n",
3055                          ProcResource->Name, I->ProcResourceIdx,
3056                          ProcResourceCount[I->ProcResourceIdx], NumUnits,
3057                          I->Cycles);
3058     });
3059     if (ProcResourceCount[I->ProcResourceIdx] >= NumUnits)
3060       return false;
3061   }
3062   LLVM_DEBUG(if (SwpDebugResource) dbgs() << "return true\n\n";);
3063   return true;
3064 }
3065 
3066 void ResourceManager::reserveResources(const MCInstrDesc *MID) {
3067   LLVM_DEBUG({
3068     if (SwpDebugResource)
3069       dbgs() << "reserveResources:\n";
3070   });
3071   if (UseDFA)
3072     return DFAResources->reserveResources(MID);
3073 
3074   unsigned InsnClass = MID->getSchedClass();
3075   const MCSchedClassDesc *SCDesc = SM.getSchedClassDesc(InsnClass);
3076   if (!SCDesc->isValid()) {
3077     LLVM_DEBUG({
3078       dbgs() << "No valid Schedule Class Desc for schedClass!\n";
3079       dbgs() << "isPseduo:" << MID->isPseudo() << "\n";
3080     });
3081     return;
3082   }
3083   for (const MCWriteProcResEntry &PRE :
3084        make_range(STI->getWriteProcResBegin(SCDesc),
3085                   STI->getWriteProcResEnd(SCDesc))) {
3086     if (!PRE.Cycles)
3087       continue;
3088     ++ProcResourceCount[PRE.ProcResourceIdx];
3089     LLVM_DEBUG({
3090       if (SwpDebugResource) {
3091         const MCProcResourceDesc *ProcResource =
3092             SM.getProcResource(PRE.ProcResourceIdx);
3093         dbgs() << format(" %16s(%2d): Count: %2d, NumUnits:%2d, Cycles:%2d\n",
3094                          ProcResource->Name, PRE.ProcResourceIdx,
3095                          ProcResourceCount[PRE.ProcResourceIdx],
3096                          ProcResource->NumUnits, PRE.Cycles);
3097       }
3098     });
3099   }
3100   LLVM_DEBUG({
3101     if (SwpDebugResource)
3102       dbgs() << "reserveResources: done!\n\n";
3103   });
3104 }
3105 
3106 bool ResourceManager::canReserveResources(const MachineInstr &MI) const {
3107   return canReserveResources(&MI.getDesc());
3108 }
3109 
3110 void ResourceManager::reserveResources(const MachineInstr &MI) {
3111   return reserveResources(&MI.getDesc());
3112 }
3113 
3114 void ResourceManager::clearResources() {
3115   if (UseDFA)
3116     return DFAResources->clearResources();
3117   std::fill(ProcResourceCount.begin(), ProcResourceCount.end(), 0);
3118 }
3119