xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/PostRASchedulerList.cpp (revision 6be3386466ab79a84b48429ae66244f21526d3df)
1 //===----- SchedulePostRAList.cpp - list scheduler ------------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This implements a top-down list scheduler, using standard algorithms.
10 // The basic approach uses a priority queue of available nodes to schedule.
11 // One at a time, nodes are taken from the priority queue (thus in priority
12 // order), checked for legality to schedule, and emitted if legal.
13 //
14 // Nodes may not be legal to schedule either due to structural hazards (e.g.
15 // pipeline or resource constraints) or because an input to the instruction has
16 // not completed execution.
17 //
18 //===----------------------------------------------------------------------===//
19 
20 #include "llvm/ADT/Statistic.h"
21 #include "llvm/Analysis/AliasAnalysis.h"
22 #include "llvm/CodeGen/AntiDepBreaker.h"
23 #include "llvm/CodeGen/LatencyPriorityQueue.h"
24 #include "llvm/CodeGen/MachineDominators.h"
25 #include "llvm/CodeGen/MachineFunctionPass.h"
26 #include "llvm/CodeGen/MachineLoopInfo.h"
27 #include "llvm/CodeGen/MachineRegisterInfo.h"
28 #include "llvm/CodeGen/Passes.h"
29 #include "llvm/CodeGen/RegisterClassInfo.h"
30 #include "llvm/CodeGen/ScheduleDAGInstrs.h"
31 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
32 #include "llvm/CodeGen/SchedulerRegistry.h"
33 #include "llvm/CodeGen/TargetInstrInfo.h"
34 #include "llvm/CodeGen/TargetLowering.h"
35 #include "llvm/CodeGen/TargetPassConfig.h"
36 #include "llvm/CodeGen/TargetRegisterInfo.h"
37 #include "llvm/CodeGen/TargetSubtargetInfo.h"
38 #include "llvm/Config/llvm-config.h"
39 #include "llvm/InitializePasses.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/ErrorHandling.h"
43 #include "llvm/Support/raw_ostream.h"
44 using namespace llvm;
45 
46 #define DEBUG_TYPE "post-RA-sched"
47 
48 STATISTIC(NumNoops, "Number of noops inserted");
49 STATISTIC(NumStalls, "Number of pipeline stalls");
50 STATISTIC(NumFixedAnti, "Number of fixed anti-dependencies");
51 
52 // Post-RA scheduling is enabled with
53 // TargetSubtargetInfo.enablePostRAScheduler(). This flag can be used to
54 // override the target.
55 static cl::opt<bool>
56 EnablePostRAScheduler("post-RA-scheduler",
57                        cl::desc("Enable scheduling after register allocation"),
58                        cl::init(false), cl::Hidden);
59 static cl::opt<std::string>
60 EnableAntiDepBreaking("break-anti-dependencies",
61                       cl::desc("Break post-RA scheduling anti-dependencies: "
62                                "\"critical\", \"all\", or \"none\""),
63                       cl::init("none"), cl::Hidden);
64 
65 // If DebugDiv > 0 then only schedule MBB with (ID % DebugDiv) == DebugMod
66 static cl::opt<int>
67 DebugDiv("postra-sched-debugdiv",
68                       cl::desc("Debug control MBBs that are scheduled"),
69                       cl::init(0), cl::Hidden);
70 static cl::opt<int>
71 DebugMod("postra-sched-debugmod",
72                       cl::desc("Debug control MBBs that are scheduled"),
73                       cl::init(0), cl::Hidden);
74 
75 AntiDepBreaker::~AntiDepBreaker() { }
76 
77 namespace {
78   class PostRAScheduler : public MachineFunctionPass {
79     const TargetInstrInfo *TII = nullptr;
80     RegisterClassInfo RegClassInfo;
81 
82   public:
83     static char ID;
84     PostRAScheduler() : MachineFunctionPass(ID) {}
85 
86     void getAnalysisUsage(AnalysisUsage &AU) const override {
87       AU.setPreservesCFG();
88       AU.addRequired<AAResultsWrapperPass>();
89       AU.addRequired<TargetPassConfig>();
90       AU.addRequired<MachineDominatorTree>();
91       AU.addPreserved<MachineDominatorTree>();
92       AU.addRequired<MachineLoopInfo>();
93       AU.addPreserved<MachineLoopInfo>();
94       MachineFunctionPass::getAnalysisUsage(AU);
95     }
96 
97     MachineFunctionProperties getRequiredProperties() const override {
98       return MachineFunctionProperties().set(
99           MachineFunctionProperties::Property::NoVRegs);
100     }
101 
102     bool runOnMachineFunction(MachineFunction &Fn) override;
103 
104   private:
105     bool enablePostRAScheduler(
106         const TargetSubtargetInfo &ST, CodeGenOpt::Level OptLevel,
107         TargetSubtargetInfo::AntiDepBreakMode &Mode,
108         TargetSubtargetInfo::RegClassVector &CriticalPathRCs) const;
109   };
110   char PostRAScheduler::ID = 0;
111 
112   class SchedulePostRATDList : public ScheduleDAGInstrs {
113     /// AvailableQueue - The priority queue to use for the available SUnits.
114     ///
115     LatencyPriorityQueue AvailableQueue;
116 
117     /// PendingQueue - This contains all of the instructions whose operands have
118     /// been issued, but their results are not ready yet (due to the latency of
119     /// the operation).  Once the operands becomes available, the instruction is
120     /// added to the AvailableQueue.
121     std::vector<SUnit*> PendingQueue;
122 
123     /// HazardRec - The hazard recognizer to use.
124     ScheduleHazardRecognizer *HazardRec;
125 
126     /// AntiDepBreak - Anti-dependence breaking object, or NULL if none
127     AntiDepBreaker *AntiDepBreak;
128 
129     /// AA - AliasAnalysis for making memory reference queries.
130     AliasAnalysis *AA;
131 
132     /// The schedule. Null SUnit*'s represent noop instructions.
133     std::vector<SUnit*> Sequence;
134 
135     /// Ordered list of DAG postprocessing steps.
136     std::vector<std::unique_ptr<ScheduleDAGMutation>> Mutations;
137 
138     /// The index in BB of RegionEnd.
139     ///
140     /// This is the instruction number from the top of the current block, not
141     /// the SlotIndex. It is only used by the AntiDepBreaker.
142     unsigned EndIndex;
143 
144   public:
145     SchedulePostRATDList(
146         MachineFunction &MF, MachineLoopInfo &MLI, AliasAnalysis *AA,
147         const RegisterClassInfo &,
148         TargetSubtargetInfo::AntiDepBreakMode AntiDepMode,
149         SmallVectorImpl<const TargetRegisterClass *> &CriticalPathRCs);
150 
151     ~SchedulePostRATDList() override;
152 
153     /// startBlock - Initialize register live-range state for scheduling in
154     /// this block.
155     ///
156     void startBlock(MachineBasicBlock *BB) override;
157 
158     // Set the index of RegionEnd within the current BB.
159     void setEndIndex(unsigned EndIdx) { EndIndex = EndIdx; }
160 
161     /// Initialize the scheduler state for the next scheduling region.
162     void enterRegion(MachineBasicBlock *bb,
163                      MachineBasicBlock::iterator begin,
164                      MachineBasicBlock::iterator end,
165                      unsigned regioninstrs) override;
166 
167     /// Notify that the scheduler has finished scheduling the current region.
168     void exitRegion() override;
169 
170     /// Schedule - Schedule the instruction range using list scheduling.
171     ///
172     void schedule() override;
173 
174     void EmitSchedule();
175 
176     /// Observe - Update liveness information to account for the current
177     /// instruction, which will not be scheduled.
178     ///
179     void Observe(MachineInstr &MI, unsigned Count);
180 
181     /// finishBlock - Clean up register live-range state.
182     ///
183     void finishBlock() override;
184 
185   private:
186     /// Apply each ScheduleDAGMutation step in order.
187     void postprocessDAG();
188 
189     void ReleaseSucc(SUnit *SU, SDep *SuccEdge);
190     void ReleaseSuccessors(SUnit *SU);
191     void ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle);
192     void ListScheduleTopDown();
193 
194     void dumpSchedule() const;
195     void emitNoop(unsigned CurCycle);
196   };
197 }
198 
199 char &llvm::PostRASchedulerID = PostRAScheduler::ID;
200 
201 INITIALIZE_PASS(PostRAScheduler, DEBUG_TYPE,
202                 "Post RA top-down list latency scheduler", false, false)
203 
204 SchedulePostRATDList::SchedulePostRATDList(
205     MachineFunction &MF, MachineLoopInfo &MLI, AliasAnalysis *AA,
206     const RegisterClassInfo &RCI,
207     TargetSubtargetInfo::AntiDepBreakMode AntiDepMode,
208     SmallVectorImpl<const TargetRegisterClass *> &CriticalPathRCs)
209     : ScheduleDAGInstrs(MF, &MLI), AA(AA), EndIndex(0) {
210 
211   const InstrItineraryData *InstrItins =
212       MF.getSubtarget().getInstrItineraryData();
213   HazardRec =
214       MF.getSubtarget().getInstrInfo()->CreateTargetPostRAHazardRecognizer(
215           InstrItins, this);
216   MF.getSubtarget().getPostRAMutations(Mutations);
217 
218   assert((AntiDepMode == TargetSubtargetInfo::ANTIDEP_NONE ||
219           MRI.tracksLiveness()) &&
220          "Live-ins must be accurate for anti-dependency breaking");
221   AntiDepBreak = ((AntiDepMode == TargetSubtargetInfo::ANTIDEP_ALL)
222                       ? createAggressiveAntiDepBreaker(MF, RCI, CriticalPathRCs)
223                       : ((AntiDepMode == TargetSubtargetInfo::ANTIDEP_CRITICAL)
224                              ? createCriticalAntiDepBreaker(MF, RCI)
225                              : nullptr));
226 }
227 
228 SchedulePostRATDList::~SchedulePostRATDList() {
229   delete HazardRec;
230   delete AntiDepBreak;
231 }
232 
233 /// Initialize state associated with the next scheduling region.
234 void SchedulePostRATDList::enterRegion(MachineBasicBlock *bb,
235                  MachineBasicBlock::iterator begin,
236                  MachineBasicBlock::iterator end,
237                  unsigned regioninstrs) {
238   ScheduleDAGInstrs::enterRegion(bb, begin, end, regioninstrs);
239   Sequence.clear();
240 }
241 
242 /// Print the schedule before exiting the region.
243 void SchedulePostRATDList::exitRegion() {
244   LLVM_DEBUG({
245     dbgs() << "*** Final schedule ***\n";
246     dumpSchedule();
247     dbgs() << '\n';
248   });
249   ScheduleDAGInstrs::exitRegion();
250 }
251 
252 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
253 /// dumpSchedule - dump the scheduled Sequence.
254 LLVM_DUMP_METHOD void SchedulePostRATDList::dumpSchedule() const {
255   for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
256     if (SUnit *SU = Sequence[i])
257       dumpNode(*SU);
258     else
259       dbgs() << "**** NOOP ****\n";
260   }
261 }
262 #endif
263 
264 bool PostRAScheduler::enablePostRAScheduler(
265     const TargetSubtargetInfo &ST,
266     CodeGenOpt::Level OptLevel,
267     TargetSubtargetInfo::AntiDepBreakMode &Mode,
268     TargetSubtargetInfo::RegClassVector &CriticalPathRCs) const {
269   Mode = ST.getAntiDepBreakMode();
270   ST.getCriticalPathRCs(CriticalPathRCs);
271 
272   // Check for explicit enable/disable of post-ra scheduling.
273   if (EnablePostRAScheduler.getPosition() > 0)
274     return EnablePostRAScheduler;
275 
276   return ST.enablePostRAScheduler() &&
277          OptLevel >= ST.getOptLevelToEnablePostRAScheduler();
278 }
279 
280 bool PostRAScheduler::runOnMachineFunction(MachineFunction &Fn) {
281   if (skipFunction(Fn.getFunction()))
282     return false;
283 
284   TII = Fn.getSubtarget().getInstrInfo();
285   MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();
286   AliasAnalysis *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
287   TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
288 
289   RegClassInfo.runOnMachineFunction(Fn);
290 
291   TargetSubtargetInfo::AntiDepBreakMode AntiDepMode =
292     TargetSubtargetInfo::ANTIDEP_NONE;
293   SmallVector<const TargetRegisterClass*, 4> CriticalPathRCs;
294 
295   // Check that post-RA scheduling is enabled for this target.
296   // This may upgrade the AntiDepMode.
297   if (!enablePostRAScheduler(Fn.getSubtarget(), PassConfig->getOptLevel(),
298                              AntiDepMode, CriticalPathRCs))
299     return false;
300 
301   // Check for antidep breaking override...
302   if (EnableAntiDepBreaking.getPosition() > 0) {
303     AntiDepMode = (EnableAntiDepBreaking == "all")
304       ? TargetSubtargetInfo::ANTIDEP_ALL
305       : ((EnableAntiDepBreaking == "critical")
306          ? TargetSubtargetInfo::ANTIDEP_CRITICAL
307          : TargetSubtargetInfo::ANTIDEP_NONE);
308   }
309 
310   LLVM_DEBUG(dbgs() << "PostRAScheduler\n");
311 
312   SchedulePostRATDList Scheduler(Fn, MLI, AA, RegClassInfo, AntiDepMode,
313                                  CriticalPathRCs);
314 
315   // Loop over all of the basic blocks
316   for (auto &MBB : Fn) {
317 #ifndef NDEBUG
318     // If DebugDiv > 0 then only schedule MBB with (ID % DebugDiv) == DebugMod
319     if (DebugDiv > 0) {
320       static int bbcnt = 0;
321       if (bbcnt++ % DebugDiv != DebugMod)
322         continue;
323       dbgs() << "*** DEBUG scheduling " << Fn.getName() << ":"
324              << printMBBReference(MBB) << " ***\n";
325     }
326 #endif
327 
328     // Initialize register live-range state for scheduling in this block.
329     Scheduler.startBlock(&MBB);
330 
331     // Schedule each sequence of instructions not interrupted by a label
332     // or anything else that effectively needs to shut down scheduling.
333     MachineBasicBlock::iterator Current = MBB.end();
334     unsigned Count = MBB.size(), CurrentCount = Count;
335     for (MachineBasicBlock::iterator I = Current; I != MBB.begin();) {
336       MachineInstr &MI = *std::prev(I);
337       --Count;
338       // Calls are not scheduling boundaries before register allocation, but
339       // post-ra we don't gain anything by scheduling across calls since we
340       // don't need to worry about register pressure.
341       if (MI.isCall() || TII->isSchedulingBoundary(MI, &MBB, Fn)) {
342         Scheduler.enterRegion(&MBB, I, Current, CurrentCount - Count);
343         Scheduler.setEndIndex(CurrentCount);
344         Scheduler.schedule();
345         Scheduler.exitRegion();
346         Scheduler.EmitSchedule();
347         Current = &MI;
348         CurrentCount = Count;
349         Scheduler.Observe(MI, CurrentCount);
350       }
351       I = MI;
352       if (MI.isBundle())
353         Count -= MI.getBundleSize();
354     }
355     assert(Count == 0 && "Instruction count mismatch!");
356     assert((MBB.begin() == Current || CurrentCount != 0) &&
357            "Instruction count mismatch!");
358     Scheduler.enterRegion(&MBB, MBB.begin(), Current, CurrentCount);
359     Scheduler.setEndIndex(CurrentCount);
360     Scheduler.schedule();
361     Scheduler.exitRegion();
362     Scheduler.EmitSchedule();
363 
364     // Clean up register live-range state.
365     Scheduler.finishBlock();
366 
367     // Update register kills
368     Scheduler.fixupKills(MBB);
369   }
370 
371   return true;
372 }
373 
374 /// StartBlock - Initialize register live-range state for scheduling in
375 /// this block.
376 ///
377 void SchedulePostRATDList::startBlock(MachineBasicBlock *BB) {
378   // Call the superclass.
379   ScheduleDAGInstrs::startBlock(BB);
380 
381   // Reset the hazard recognizer and anti-dep breaker.
382   HazardRec->Reset();
383   if (AntiDepBreak)
384     AntiDepBreak->StartBlock(BB);
385 }
386 
387 /// Schedule - Schedule the instruction range using list scheduling.
388 ///
389 void SchedulePostRATDList::schedule() {
390   // Build the scheduling graph.
391   buildSchedGraph(AA);
392 
393   if (AntiDepBreak) {
394     unsigned Broken =
395       AntiDepBreak->BreakAntiDependencies(SUnits, RegionBegin, RegionEnd,
396                                           EndIndex, DbgValues);
397 
398     if (Broken != 0) {
399       // We made changes. Update the dependency graph.
400       // Theoretically we could update the graph in place:
401       // When a live range is changed to use a different register, remove
402       // the def's anti-dependence *and* output-dependence edges due to
403       // that register, and add new anti-dependence and output-dependence
404       // edges based on the next live range of the register.
405       ScheduleDAG::clearDAG();
406       buildSchedGraph(AA);
407 
408       NumFixedAnti += Broken;
409     }
410   }
411 
412   postprocessDAG();
413 
414   LLVM_DEBUG(dbgs() << "********** List Scheduling **********\n");
415   LLVM_DEBUG(dump());
416 
417   AvailableQueue.initNodes(SUnits);
418   ListScheduleTopDown();
419   AvailableQueue.releaseState();
420 }
421 
422 /// Observe - Update liveness information to account for the current
423 /// instruction, which will not be scheduled.
424 ///
425 void SchedulePostRATDList::Observe(MachineInstr &MI, unsigned Count) {
426   if (AntiDepBreak)
427     AntiDepBreak->Observe(MI, Count, EndIndex);
428 }
429 
430 /// FinishBlock - Clean up register live-range state.
431 ///
432 void SchedulePostRATDList::finishBlock() {
433   if (AntiDepBreak)
434     AntiDepBreak->FinishBlock();
435 
436   // Call the superclass.
437   ScheduleDAGInstrs::finishBlock();
438 }
439 
440 /// Apply each ScheduleDAGMutation step in order.
441 void SchedulePostRATDList::postprocessDAG() {
442   for (auto &M : Mutations)
443     M->apply(this);
444 }
445 
446 //===----------------------------------------------------------------------===//
447 //  Top-Down Scheduling
448 //===----------------------------------------------------------------------===//
449 
450 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to
451 /// the PendingQueue if the count reaches zero.
452 void SchedulePostRATDList::ReleaseSucc(SUnit *SU, SDep *SuccEdge) {
453   SUnit *SuccSU = SuccEdge->getSUnit();
454 
455   if (SuccEdge->isWeak()) {
456     --SuccSU->WeakPredsLeft;
457     return;
458   }
459 #ifndef NDEBUG
460   if (SuccSU->NumPredsLeft == 0) {
461     dbgs() << "*** Scheduling failed! ***\n";
462     dumpNode(*SuccSU);
463     dbgs() << " has been released too many times!\n";
464     llvm_unreachable(nullptr);
465   }
466 #endif
467   --SuccSU->NumPredsLeft;
468 
469   // Standard scheduler algorithms will recompute the depth of the successor
470   // here as such:
471   //   SuccSU->setDepthToAtLeast(SU->getDepth() + SuccEdge->getLatency());
472   //
473   // However, we lazily compute node depth instead. Note that
474   // ScheduleNodeTopDown has already updated the depth of this node which causes
475   // all descendents to be marked dirty. Setting the successor depth explicitly
476   // here would cause depth to be recomputed for all its ancestors. If the
477   // successor is not yet ready (because of a transitively redundant edge) then
478   // this causes depth computation to be quadratic in the size of the DAG.
479 
480   // If all the node's predecessors are scheduled, this node is ready
481   // to be scheduled. Ignore the special ExitSU node.
482   if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
483     PendingQueue.push_back(SuccSU);
484 }
485 
486 /// ReleaseSuccessors - Call ReleaseSucc on each of SU's successors.
487 void SchedulePostRATDList::ReleaseSuccessors(SUnit *SU) {
488   for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
489        I != E; ++I) {
490     ReleaseSucc(SU, &*I);
491   }
492 }
493 
494 /// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending
495 /// count of its successors. If a successor pending count is zero, add it to
496 /// the Available queue.
497 void SchedulePostRATDList::ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle) {
498   LLVM_DEBUG(dbgs() << "*** Scheduling [" << CurCycle << "]: ");
499   LLVM_DEBUG(dumpNode(*SU));
500 
501   Sequence.push_back(SU);
502   assert(CurCycle >= SU->getDepth() &&
503          "Node scheduled above its depth!");
504   SU->setDepthToAtLeast(CurCycle);
505 
506   ReleaseSuccessors(SU);
507   SU->isScheduled = true;
508   AvailableQueue.scheduledNode(SU);
509 }
510 
511 /// emitNoop - Add a noop to the current instruction sequence.
512 void SchedulePostRATDList::emitNoop(unsigned CurCycle) {
513   LLVM_DEBUG(dbgs() << "*** Emitting noop in cycle " << CurCycle << '\n');
514   HazardRec->EmitNoop();
515   Sequence.push_back(nullptr);   // NULL here means noop
516   ++NumNoops;
517 }
518 
519 /// ListScheduleTopDown - The main loop of list scheduling for top-down
520 /// schedulers.
521 void SchedulePostRATDList::ListScheduleTopDown() {
522   unsigned CurCycle = 0;
523 
524   // We're scheduling top-down but we're visiting the regions in
525   // bottom-up order, so we don't know the hazards at the start of a
526   // region. So assume no hazards (this should usually be ok as most
527   // blocks are a single region).
528   HazardRec->Reset();
529 
530   // Release any successors of the special Entry node.
531   ReleaseSuccessors(&EntrySU);
532 
533   // Add all leaves to Available queue.
534   for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
535     // It is available if it has no predecessors.
536     if (!SUnits[i].NumPredsLeft && !SUnits[i].isAvailable) {
537       AvailableQueue.push(&SUnits[i]);
538       SUnits[i].isAvailable = true;
539     }
540   }
541 
542   // In any cycle where we can't schedule any instructions, we must
543   // stall or emit a noop, depending on the target.
544   bool CycleHasInsts = false;
545 
546   // While Available queue is not empty, grab the node with the highest
547   // priority. If it is not ready put it back.  Schedule the node.
548   std::vector<SUnit*> NotReady;
549   Sequence.reserve(SUnits.size());
550   while (!AvailableQueue.empty() || !PendingQueue.empty()) {
551     // Check to see if any of the pending instructions are ready to issue.  If
552     // so, add them to the available queue.
553     unsigned MinDepth = ~0u;
554     for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
555       if (PendingQueue[i]->getDepth() <= CurCycle) {
556         AvailableQueue.push(PendingQueue[i]);
557         PendingQueue[i]->isAvailable = true;
558         PendingQueue[i] = PendingQueue.back();
559         PendingQueue.pop_back();
560         --i; --e;
561       } else if (PendingQueue[i]->getDepth() < MinDepth)
562         MinDepth = PendingQueue[i]->getDepth();
563     }
564 
565     LLVM_DEBUG(dbgs() << "\n*** Examining Available\n";
566                AvailableQueue.dump(this));
567 
568     SUnit *FoundSUnit = nullptr, *NotPreferredSUnit = nullptr;
569     bool HasNoopHazards = false;
570     while (!AvailableQueue.empty()) {
571       SUnit *CurSUnit = AvailableQueue.pop();
572 
573       ScheduleHazardRecognizer::HazardType HT =
574         HazardRec->getHazardType(CurSUnit, 0/*no stalls*/);
575       if (HT == ScheduleHazardRecognizer::NoHazard) {
576         if (HazardRec->ShouldPreferAnother(CurSUnit)) {
577           if (!NotPreferredSUnit) {
578             // If this is the first non-preferred node for this cycle, then
579             // record it and continue searching for a preferred node. If this
580             // is not the first non-preferred node, then treat it as though
581             // there had been a hazard.
582             NotPreferredSUnit = CurSUnit;
583             continue;
584           }
585         } else {
586           FoundSUnit = CurSUnit;
587           break;
588         }
589       }
590 
591       // Remember if this is a noop hazard.
592       HasNoopHazards |= HT == ScheduleHazardRecognizer::NoopHazard;
593 
594       NotReady.push_back(CurSUnit);
595     }
596 
597     // If we have a non-preferred node, push it back onto the available list.
598     // If we did not find a preferred node, then schedule this first
599     // non-preferred node.
600     if (NotPreferredSUnit) {
601       if (!FoundSUnit) {
602         LLVM_DEBUG(
603             dbgs() << "*** Will schedule a non-preferred instruction...\n");
604         FoundSUnit = NotPreferredSUnit;
605       } else {
606         AvailableQueue.push(NotPreferredSUnit);
607       }
608 
609       NotPreferredSUnit = nullptr;
610     }
611 
612     // Add the nodes that aren't ready back onto the available list.
613     if (!NotReady.empty()) {
614       AvailableQueue.push_all(NotReady);
615       NotReady.clear();
616     }
617 
618     // If we found a node to schedule...
619     if (FoundSUnit) {
620       // If we need to emit noops prior to this instruction, then do so.
621       unsigned NumPreNoops = HazardRec->PreEmitNoops(FoundSUnit);
622       for (unsigned i = 0; i != NumPreNoops; ++i)
623         emitNoop(CurCycle);
624 
625       // ... schedule the node...
626       ScheduleNodeTopDown(FoundSUnit, CurCycle);
627       HazardRec->EmitInstruction(FoundSUnit);
628       CycleHasInsts = true;
629       if (HazardRec->atIssueLimit()) {
630         LLVM_DEBUG(dbgs() << "*** Max instructions per cycle " << CurCycle
631                           << '\n');
632         HazardRec->AdvanceCycle();
633         ++CurCycle;
634         CycleHasInsts = false;
635       }
636     } else {
637       if (CycleHasInsts) {
638         LLVM_DEBUG(dbgs() << "*** Finished cycle " << CurCycle << '\n');
639         HazardRec->AdvanceCycle();
640       } else if (!HasNoopHazards) {
641         // Otherwise, we have a pipeline stall, but no other problem,
642         // just advance the current cycle and try again.
643         LLVM_DEBUG(dbgs() << "*** Stall in cycle " << CurCycle << '\n');
644         HazardRec->AdvanceCycle();
645         ++NumStalls;
646       } else {
647         // Otherwise, we have no instructions to issue and we have instructions
648         // that will fault if we don't do this right.  This is the case for
649         // processors without pipeline interlocks and other cases.
650         emitNoop(CurCycle);
651       }
652 
653       ++CurCycle;
654       CycleHasInsts = false;
655     }
656   }
657 
658 #ifndef NDEBUG
659   unsigned ScheduledNodes = VerifyScheduledDAG(/*isBottomUp=*/false);
660   unsigned Noops = 0;
661   for (unsigned i = 0, e = Sequence.size(); i != e; ++i)
662     if (!Sequence[i])
663       ++Noops;
664   assert(Sequence.size() - Noops == ScheduledNodes &&
665          "The number of nodes scheduled doesn't match the expected number!");
666 #endif // NDEBUG
667 }
668 
669 // EmitSchedule - Emit the machine code in scheduled order.
670 void SchedulePostRATDList::EmitSchedule() {
671   RegionBegin = RegionEnd;
672 
673   // If first instruction was a DBG_VALUE then put it back.
674   if (FirstDbgValue)
675     BB->splice(RegionEnd, BB, FirstDbgValue);
676 
677   // Then re-insert them according to the given schedule.
678   for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
679     if (SUnit *SU = Sequence[i])
680       BB->splice(RegionEnd, BB, SU->getInstr());
681     else
682       // Null SUnit* is a noop.
683       TII->insertNoop(*BB, RegionEnd);
684 
685     // Update the Begin iterator, as the first instruction in the block
686     // may have been scheduled later.
687     if (i == 0)
688       RegionBegin = std::prev(RegionEnd);
689   }
690 
691   // Reinsert any remaining debug_values.
692   for (std::vector<std::pair<MachineInstr *, MachineInstr *> >::iterator
693          DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) {
694     std::pair<MachineInstr *, MachineInstr *> P = *std::prev(DI);
695     MachineInstr *DbgValue = P.first;
696     MachineBasicBlock::iterator OrigPrivMI = P.second;
697     BB->splice(++OrigPrivMI, BB, DbgValue);
698   }
699   DbgValues.clear();
700   FirstDbgValue = nullptr;
701 }
702