1 //===----- SchedulePostRAList.cpp - list scheduler ------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This implements a top-down list scheduler, using standard algorithms.
11 // The basic approach uses a priority queue of available nodes to schedule.
12 // One at a time, nodes are taken from the priority queue (thus in priority
13 // order), checked for legality to schedule, and emitted if legal.
15 // Nodes may not be legal to schedule either due to structural hazards (e.g.
16 // pipeline or resource constraints) or because an input to the instruction has
17 // not completed execution.
19 //===----------------------------------------------------------------------===//
21 #define DEBUG_TYPE "post-RA-sched"
22 #include "ScheduleDAGInstrs.h"
23 #include "llvm/CodeGen/Passes.h"
24 #include "llvm/CodeGen/LatencyPriorityQueue.h"
25 #include "llvm/CodeGen/SchedulerRegistry.h"
26 #include "llvm/CodeGen/MachineDominators.h"
27 #include "llvm/CodeGen/MachineFunctionPass.h"
28 #include "llvm/CodeGen/MachineLoopInfo.h"
29 #include "llvm/CodeGen/MachineRegisterInfo.h"
30 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
31 #include "llvm/Target/TargetMachine.h"
32 #include "llvm/Target/TargetInstrInfo.h"
33 #include "llvm/Target/TargetRegisterInfo.h"
34 #include "llvm/Support/Compiler.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/ADT/Statistic.h"
40 STATISTIC(NumNoops, "Number of noops inserted");
41 STATISTIC(NumStalls, "Number of pipeline stalls");
44 EnableAntiDepBreaking("break-anti-dependencies",
45 cl::desc("Break post-RA scheduling anti-dependencies"),
46 cl::init(true), cl::Hidden);
49 EnablePostRAHazardAvoidance("avoid-hazards",
50 cl::desc("Enable simple hazard-avoidance"),
51 cl::init(true), cl::Hidden);
54 class VISIBILITY_HIDDEN PostRAScheduler : public MachineFunctionPass {
57 PostRAScheduler() : MachineFunctionPass(&ID) {}
59 void getAnalysisUsage(AnalysisUsage &AU) const {
60 AU.addRequired<MachineDominatorTree>();
61 AU.addPreserved<MachineDominatorTree>();
62 AU.addRequired<MachineLoopInfo>();
63 AU.addPreserved<MachineLoopInfo>();
64 MachineFunctionPass::getAnalysisUsage(AU);
67 const char *getPassName() const {
68 return "Post RA top-down list latency scheduler";
71 bool runOnMachineFunction(MachineFunction &Fn);
73 char PostRAScheduler::ID = 0;
75 class VISIBILITY_HIDDEN SchedulePostRATDList : public ScheduleDAGInstrs {
76 /// AvailableQueue - The priority queue to use for the available SUnits.
78 LatencyPriorityQueue AvailableQueue;
80 /// PendingQueue - This contains all of the instructions whose operands have
81 /// been issued, but their results are not ready yet (due to the latency of
82 /// the operation). Once the operands becomes available, the instruction is
83 /// added to the AvailableQueue.
84 std::vector<SUnit*> PendingQueue;
86 /// Topo - A topological ordering for SUnits.
87 ScheduleDAGTopologicalSort Topo;
89 /// AllocatableSet - The set of allocatable registers.
90 /// We'll be ignoring anti-dependencies on non-allocatable registers,
91 /// because they may not be safe to break.
92 const BitVector AllocatableSet;
94 /// HazardRec - The hazard recognizer to use.
95 ScheduleHazardRecognizer *HazardRec;
98 SchedulePostRATDList(MachineFunction &MF,
99 const MachineLoopInfo &MLI,
100 const MachineDominatorTree &MDT,
101 ScheduleHazardRecognizer *HR)
102 : ScheduleDAGInstrs(MF, MLI, MDT), Topo(SUnits),
103 AllocatableSet(TRI->getAllocatableSet(MF)),
106 ~SchedulePostRATDList() {
113 void ReleaseSucc(SUnit *SU, SDep *SuccEdge);
114 void ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle);
115 void ListScheduleTopDown();
116 bool BreakAntiDependencies();
119 /// SimpleHazardRecognizer - A *very* simple hazard recognizer. It uses
120 /// a coarse classification and attempts to avoid that instructions of
121 /// a given class aren't grouped too densely together.
122 class SimpleHazardRecognizer : public ScheduleHazardRecognizer {
123 /// Class - A simple classification for SUnits.
128 /// Window - The Class values of the most recently issued
132 /// getClass - Classify the given SUnit.
133 Class getClass(const SUnit *SU) {
134 const MachineInstr *MI = SU->getInstr();
135 const TargetInstrDesc &TID = MI->getDesc();
143 /// Step - Rotate the existing entries in Window and insert the
144 /// given class value in position as the most recent.
146 std::copy(Window+1, array_endof(Window), Window);
147 Window[array_lengthof(Window)-1] = C;
151 SimpleHazardRecognizer() : Window() {}
153 virtual HazardType getHazardType(SUnit *SU) {
154 Class C = getClass(SU);
158 for (unsigned i = 0; i != array_lengthof(Window); ++i)
161 if (Score > array_lengthof(Window) * 2)
166 virtual void EmitInstruction(SUnit *SU) {
170 virtual void AdvanceCycle() {
176 bool PostRAScheduler::runOnMachineFunction(MachineFunction &Fn) {
177 DOUT << "PostRAScheduler\n";
179 const MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();
180 const MachineDominatorTree &MDT = getAnalysis<MachineDominatorTree>();
181 ScheduleHazardRecognizer *HR = EnablePostRAHazardAvoidance ?
182 new SimpleHazardRecognizer :
183 new ScheduleHazardRecognizer();
185 SchedulePostRATDList Scheduler(Fn, MLI, MDT, HR);
187 // Loop over all of the basic blocks
188 for (MachineFunction::iterator MBB = Fn.begin(), MBBe = Fn.end();
189 MBB != MBBe; ++MBB) {
190 // Schedule each sequence of instructions not interrupted by a label
191 // or anything else that effectively needs to shut down scheduling.
192 MachineBasicBlock::iterator Current = MBB->end(), Top = MBB->begin();
193 for (MachineBasicBlock::iterator I = Current; I != Top; ) {
194 MachineInstr *MI = --I;
195 if (MI->getDesc().isTerminator() || MI->isLabel()) {
196 Scheduler.Run(0, MBB, next(I), Current);
197 Scheduler.EmitSchedule();
202 Scheduler.Run(0, MBB, Top, Current);
203 Scheduler.EmitSchedule();
209 /// Schedule - Schedule the DAG using list scheduling.
210 void SchedulePostRATDList::Schedule() {
211 DOUT << "********** List Scheduling **********\n";
213 // Build the scheduling graph.
216 if (EnableAntiDepBreaking) {
217 if (BreakAntiDependencies()) {
218 // We made changes. Update the dependency graph.
219 // Theoretically we could update the graph in place:
220 // When a live range is changed to use a different register, remove
221 // the def's anti-dependence *and* output-dependence edges due to
222 // that register, and add new anti-dependence and output-dependence
223 // edges based on the next live range of the register.
229 AvailableQueue.initNodes(SUnits);
231 ListScheduleTopDown();
233 AvailableQueue.releaseState();
236 /// getInstrOperandRegClass - Return register class of the operand of an
237 /// instruction of the specified TargetInstrDesc.
238 static const TargetRegisterClass*
239 getInstrOperandRegClass(const TargetRegisterInfo *TRI,
240 const TargetInstrInfo *TII, const TargetInstrDesc &II,
242 if (Op >= II.getNumOperands())
244 if (II.OpInfo[Op].isLookupPtrRegClass())
245 return TII->getPointerRegClass();
246 return TRI->getRegClass(II.OpInfo[Op].RegClass);
249 /// CriticalPathStep - Return the next SUnit after SU on the bottom-up
251 static SDep *CriticalPathStep(SUnit *SU) {
253 unsigned NextDepth = 0;
254 // Find the predecessor edge with the greatest depth.
255 for (SUnit::pred_iterator P = SU->Preds.begin(), PE = SU->Preds.end();
257 SUnit *PredSU = P->getSUnit();
258 unsigned PredLatency = P->getLatency();
259 unsigned PredTotalLatency = PredSU->getDepth() + PredLatency;
260 // In the case of a latency tie, prefer an anti-dependency edge over
261 // other types of edges.
262 if (NextDepth < PredTotalLatency ||
263 (NextDepth == PredTotalLatency && P->getKind() == SDep::Anti)) {
264 NextDepth = PredTotalLatency;
271 /// BreakAntiDependencies - Identifiy anti-dependencies along the critical path
272 /// of the ScheduleDAG and break them by renaming registers.
274 bool SchedulePostRATDList::BreakAntiDependencies() {
275 // The code below assumes that there is at least one instruction,
276 // so just duck out immediately if the block is empty.
277 if (SUnits.empty()) return false;
279 // Find the node at the bottom of the critical path.
281 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
282 SUnit *SU = &SUnits[i];
283 if (!Max || SU->getDepth() + SU->Latency > Max->getDepth() + Max->Latency)
287 DOUT << "Critical path has total latency "
288 << (Max->getDepth() + Max->Latency) << "\n";
290 // Track progress along the critical path through the SUnit graph as we walk
292 SUnit *CriticalPathSU = Max;
293 MachineInstr *CriticalPathMI = CriticalPathSU->getInstr();
295 // For live regs that are only used in one register class in a live range,
296 // the register class. If the register is not live, the corresponding value
297 // is null. If the register is live but used in multiple register classes,
298 // the corresponding value is -1 casted to a pointer.
299 const TargetRegisterClass *
300 Classes[TargetRegisterInfo::FirstVirtualRegister] = {};
302 // Map registers to all their references within a live range.
303 std::multimap<unsigned, MachineOperand *> RegRefs;
305 // The index of the most recent kill (proceding bottom-up), or ~0u if
306 // the register is not live.
307 unsigned KillIndices[TargetRegisterInfo::FirstVirtualRegister];
308 std::fill(KillIndices, array_endof(KillIndices), ~0u);
309 // The index of the most recent complete def (proceding bottom up), or ~0u if
310 // the register is live.
311 unsigned DefIndices[TargetRegisterInfo::FirstVirtualRegister];
312 std::fill(DefIndices, array_endof(DefIndices), BB->size());
314 // Determine the live-out physregs for this block.
315 if (BB->back().getDesc().isReturn())
316 // In a return block, examine the function live-out regs.
317 for (MachineRegisterInfo::liveout_iterator I = MRI.liveout_begin(),
318 E = MRI.liveout_end(); I != E; ++I) {
320 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
321 KillIndices[Reg] = BB->size();
322 DefIndices[Reg] = ~0u;
323 // Repeat, for all aliases.
324 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
325 unsigned AliasReg = *Alias;
326 Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
327 KillIndices[AliasReg] = BB->size();
328 DefIndices[AliasReg] = ~0u;
332 // In a non-return block, examine the live-in regs of all successors.
333 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
334 SE = BB->succ_end(); SI != SE; ++SI)
335 for (MachineBasicBlock::livein_iterator I = (*SI)->livein_begin(),
336 E = (*SI)->livein_end(); I != E; ++I) {
338 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
339 KillIndices[Reg] = BB->size();
340 DefIndices[Reg] = ~0u;
341 // Repeat, for all aliases.
342 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
343 unsigned AliasReg = *Alias;
344 Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
345 KillIndices[AliasReg] = BB->size();
346 DefIndices[AliasReg] = ~0u;
350 // Consider callee-saved registers as live-out, since we're running after
351 // prologue/epilogue insertion so there's no way to add additional
354 // TODO: If the callee saves and restores these, then we can potentially
355 // use them between the save and the restore. To do that, we could scan
356 // the exit blocks to see which of these registers are defined.
357 // Alternatively, callee-saved registers that aren't saved and restored
358 // could be marked live-in in every block.
359 for (const unsigned *I = TRI->getCalleeSavedRegs(); *I; ++I) {
361 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
362 KillIndices[Reg] = BB->size();
363 DefIndices[Reg] = ~0u;
364 // Repeat, for all aliases.
365 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
366 unsigned AliasReg = *Alias;
367 Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
368 KillIndices[AliasReg] = BB->size();
369 DefIndices[AliasReg] = ~0u;
373 // Consider this pattern:
382 // There are three anti-dependencies here, and without special care,
383 // we'd break all of them using the same register:
392 // because at each anti-dependence, B is the first register that
393 // isn't A which is free. This re-introduces anti-dependencies
394 // at all but one of the original anti-dependencies that we were
395 // trying to break. To avoid this, keep track of the most recent
396 // register that each register was replaced with, avoid avoid
397 // using it to repair an anti-dependence on the same register.
398 // This lets us produce this:
407 // This still has an anti-dependence on B, but at least it isn't on the
408 // original critical path.
410 // TODO: If we tracked more than one register here, we could potentially
411 // fix that remaining critical edge too. This is a little more involved,
412 // because unlike the most recent register, less recent registers should
413 // still be considered, though only if no other registers are available.
414 unsigned LastNewReg[TargetRegisterInfo::FirstVirtualRegister] = {};
416 // Attempt to break anti-dependence edges on the critical path. Walk the
417 // instructions from the bottom up, tracking information about liveness
418 // as we go to help determine which registers are available.
419 bool Changed = false;
420 unsigned Count = SUnits.size() - 1;
421 for (MachineBasicBlock::iterator I = End, E = Begin;
423 MachineInstr *MI = --I;
425 // After regalloc, IMPLICIT_DEF instructions aren't safe to treat as
426 // dependence-breaking. In the case of an INSERT_SUBREG, the IMPLICIT_DEF
427 // is left behind appearing to clobber the super-register, while the
428 // subregister needs to remain live. So we just ignore them.
429 if (MI->getOpcode() == TargetInstrInfo::IMPLICIT_DEF)
432 // Check if this instruction has a dependence on the critical path that
433 // is an anti-dependence that we may be able to break. If it is, set
434 // AntiDepReg to the non-zero register associated with the anti-dependence.
436 // We limit our attention to the critical path as a heuristic to avoid
437 // breaking anti-dependence edges that aren't going to significantly
438 // impact the overall schedule. There are a limited number of registers
439 // and we want to save them for the important edges.
441 // TODO: Instructions with multiple defs could have multiple
442 // anti-dependencies. The current code here only knows how to break one
443 // edge per instruction. Note that we'd have to be able to break all of
444 // the anti-dependencies in an instruction in order to be effective.
445 unsigned AntiDepReg = 0;
446 if (MI == CriticalPathMI) {
447 if (SDep *Edge = CriticalPathStep(CriticalPathSU)) {
448 SUnit *NextSU = Edge->getSUnit();
450 // Only consider anti-dependence edges.
451 if (Edge->getKind() == SDep::Anti) {
452 AntiDepReg = Edge->getReg();
453 assert(AntiDepReg != 0 && "Anti-dependence on reg0?");
454 // Don't break anti-dependencies on non-allocatable registers.
455 if (!AllocatableSet.test(AntiDepReg))
458 // If the SUnit has other dependencies on the SUnit that it
459 // anti-depends on, don't bother breaking the anti-dependency
460 // since those edges would prevent such units from being
461 // scheduled past each other regardless.
463 // Also, if there are dependencies on other SUnits with the
464 // same register as the anti-dependency, don't attempt to
466 for (SUnit::pred_iterator P = CriticalPathSU->Preds.begin(),
467 PE = CriticalPathSU->Preds.end(); P != PE; ++P)
468 if (P->getSUnit() == NextSU ?
469 (P->getKind() != SDep::Anti || P->getReg() != AntiDepReg) :
470 (P->getKind() == SDep::Data && P->getReg() == AntiDepReg)) {
476 CriticalPathSU = NextSU;
477 CriticalPathMI = CriticalPathSU->getInstr();
479 // We've reached the end of the critical path.
485 // Scan the register operands for this instruction and update
486 // Classes and RegRefs.
487 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
488 MachineOperand &MO = MI->getOperand(i);
489 if (!MO.isReg()) continue;
490 unsigned Reg = MO.getReg();
491 if (Reg == 0) continue;
492 const TargetRegisterClass *NewRC =
493 getInstrOperandRegClass(TRI, TII, MI->getDesc(), i);
495 // If this instruction has a use of AntiDepReg, breaking it
497 if (MO.isUse() && AntiDepReg == Reg)
500 // For now, only allow the register to be changed if its register
501 // class is consistent across all uses.
502 if (!Classes[Reg] && NewRC)
503 Classes[Reg] = NewRC;
504 else if (!NewRC || Classes[Reg] != NewRC)
505 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
507 // Now check for aliases.
508 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
509 // If an alias of the reg is used during the live range, give up.
510 // Note that this allows us to skip checking if AntiDepReg
511 // overlaps with any of the aliases, among other things.
512 unsigned AliasReg = *Alias;
513 if (Classes[AliasReg]) {
514 Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
515 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
519 // If we're still willing to consider this register, note the reference.
520 if (Classes[Reg] != reinterpret_cast<TargetRegisterClass *>(-1))
521 RegRefs.insert(std::make_pair(Reg, &MO));
524 // Determine AntiDepReg's register class, if it is live and is
525 // consistently used within a single class.
526 const TargetRegisterClass *RC = AntiDepReg != 0 ? Classes[AntiDepReg] : 0;
527 assert((AntiDepReg == 0 || RC != NULL) &&
528 "Register should be live if it's causing an anti-dependence!");
529 if (RC == reinterpret_cast<TargetRegisterClass *>(-1))
532 // Look for a suitable register to use to break the anti-depenence.
534 // TODO: Instead of picking the first free register, consider which might
536 if (AntiDepReg != 0) {
537 for (TargetRegisterClass::iterator R = RC->allocation_order_begin(MF),
538 RE = RC->allocation_order_end(MF); R != RE; ++R) {
539 unsigned NewReg = *R;
540 // Don't replace a register with itself.
541 if (NewReg == AntiDepReg) continue;
542 // Don't replace a register with one that was recently used to repair
543 // an anti-dependence with this AntiDepReg, because that would
544 // re-introduce that anti-dependence.
545 if (NewReg == LastNewReg[AntiDepReg]) continue;
546 // If NewReg is dead and NewReg's most recent def is not before
547 // AntiDepReg's kill, it's safe to replace AntiDepReg with NewReg.
548 assert(((KillIndices[AntiDepReg] == ~0u) != (DefIndices[AntiDepReg] == ~0u)) &&
549 "Kill and Def maps aren't consistent for AntiDepReg!");
550 assert(((KillIndices[NewReg] == ~0u) != (DefIndices[NewReg] == ~0u)) &&
551 "Kill and Def maps aren't consistent for NewReg!");
552 if (KillIndices[NewReg] == ~0u &&
553 Classes[NewReg] != reinterpret_cast<TargetRegisterClass *>(-1) &&
554 KillIndices[AntiDepReg] <= DefIndices[NewReg]) {
555 DOUT << "Breaking anti-dependence edge on "
556 << TRI->getName(AntiDepReg)
557 << " with " << RegRefs.count(AntiDepReg) << " references"
558 << " using " << TRI->getName(NewReg) << "!\n";
560 // Update the references to the old register to refer to the new
562 std::pair<std::multimap<unsigned, MachineOperand *>::iterator,
563 std::multimap<unsigned, MachineOperand *>::iterator>
564 Range = RegRefs.equal_range(AntiDepReg);
565 for (std::multimap<unsigned, MachineOperand *>::iterator
566 Q = Range.first, QE = Range.second; Q != QE; ++Q)
567 Q->second->setReg(NewReg);
569 // We just went back in time and modified history; the
570 // liveness information for the anti-depenence reg is now
571 // inconsistent. Set the state as if it were dead.
572 Classes[NewReg] = Classes[AntiDepReg];
573 DefIndices[NewReg] = DefIndices[AntiDepReg];
574 KillIndices[NewReg] = KillIndices[AntiDepReg];
576 Classes[AntiDepReg] = 0;
577 DefIndices[AntiDepReg] = KillIndices[AntiDepReg];
578 KillIndices[AntiDepReg] = ~0u;
580 RegRefs.erase(AntiDepReg);
582 LastNewReg[AntiDepReg] = NewReg;
589 // Proceding upwards, registers that are defed but not used in this
590 // instruction are now dead.
591 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
592 MachineOperand &MO = MI->getOperand(i);
593 if (!MO.isReg()) continue;
594 unsigned Reg = MO.getReg();
595 if (Reg == 0) continue;
596 if (!MO.isDef()) continue;
597 // Ignore two-addr defs.
598 if (MI->isRegReDefinedByTwoAddr(i)) continue;
600 DefIndices[Reg] = Count;
601 KillIndices[Reg] = ~0u;
604 // Repeat, for all subregs.
605 for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
607 unsigned SubregReg = *Subreg;
608 DefIndices[SubregReg] = Count;
609 KillIndices[SubregReg] = ~0u;
610 Classes[SubregReg] = 0;
611 RegRefs.erase(SubregReg);
613 // Conservatively mark super-registers as unusable.
614 for (const unsigned *Super = TRI->getSuperRegisters(Reg);
616 unsigned SuperReg = *Super;
617 Classes[SuperReg] = reinterpret_cast<TargetRegisterClass *>(-1);
620 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
621 MachineOperand &MO = MI->getOperand(i);
622 if (!MO.isReg()) continue;
623 unsigned Reg = MO.getReg();
624 if (Reg == 0) continue;
625 if (!MO.isUse()) continue;
627 const TargetRegisterClass *NewRC =
628 getInstrOperandRegClass(TRI, TII, MI->getDesc(), i);
630 // For now, only allow the register to be changed if its register
631 // class is consistent across all uses.
632 if (!Classes[Reg] && NewRC)
633 Classes[Reg] = NewRC;
634 else if (!NewRC || Classes[Reg] != NewRC)
635 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
637 RegRefs.insert(std::make_pair(Reg, &MO));
639 // It wasn't previously live but now it is, this is a kill.
640 if (KillIndices[Reg] == ~0u) {
641 KillIndices[Reg] = Count;
642 DefIndices[Reg] = ~0u;
644 // Repeat, for all aliases.
645 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
646 unsigned AliasReg = *Alias;
647 if (KillIndices[AliasReg] == ~0u) {
648 KillIndices[AliasReg] = Count;
649 DefIndices[AliasReg] = ~0u;
654 assert(Count == ~0u && "Count mismatch!");
659 //===----------------------------------------------------------------------===//
660 // Top-Down Scheduling
661 //===----------------------------------------------------------------------===//
663 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to
664 /// the PendingQueue if the count reaches zero. Also update its cycle bound.
665 void SchedulePostRATDList::ReleaseSucc(SUnit *SU, SDep *SuccEdge) {
666 SUnit *SuccSU = SuccEdge->getSUnit();
667 --SuccSU->NumPredsLeft;
670 if (SuccSU->NumPredsLeft < 0) {
671 cerr << "*** Scheduling failed! ***\n";
673 cerr << " has been released too many times!\n";
678 // Compute how many cycles it will be before this actually becomes
679 // available. This is the max of the start time of all predecessors plus
681 SuccSU->setDepthToAtLeast(SU->getDepth() + SuccEdge->getLatency());
683 if (SuccSU->NumPredsLeft == 0) {
684 PendingQueue.push_back(SuccSU);
688 /// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending
689 /// count of its successors. If a successor pending count is zero, add it to
690 /// the Available queue.
691 void SchedulePostRATDList::ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle) {
692 DOUT << "*** Scheduling [" << CurCycle << "]: ";
693 DEBUG(SU->dump(this));
695 Sequence.push_back(SU);
696 assert(CurCycle >= SU->getDepth() && "Node scheduled above its depth!");
697 SU->setDepthToAtLeast(CurCycle);
699 // Top down: release successors.
700 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
702 ReleaseSucc(SU, &*I);
704 SU->isScheduled = true;
705 AvailableQueue.ScheduledNode(SU);
708 /// ListScheduleTopDown - The main loop of list scheduling for top-down
710 void SchedulePostRATDList::ListScheduleTopDown() {
711 unsigned CurCycle = 0;
713 // All leaves to Available queue.
714 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
715 // It is available if it has no predecessors.
716 if (SUnits[i].Preds.empty()) {
717 AvailableQueue.push(&SUnits[i]);
718 SUnits[i].isAvailable = true;
722 // While Available queue is not empty, grab the node with the highest
723 // priority. If it is not ready put it back. Schedule the node.
724 std::vector<SUnit*> NotReady;
725 Sequence.reserve(SUnits.size());
726 while (!AvailableQueue.empty() || !PendingQueue.empty()) {
727 // Check to see if any of the pending instructions are ready to issue. If
728 // so, add them to the available queue.
729 unsigned MinDepth = ~0u;
730 for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
731 if (PendingQueue[i]->getDepth() <= CurCycle) {
732 AvailableQueue.push(PendingQueue[i]);
733 PendingQueue[i]->isAvailable = true;
734 PendingQueue[i] = PendingQueue.back();
735 PendingQueue.pop_back();
737 } else if (PendingQueue[i]->getDepth() < MinDepth)
738 MinDepth = PendingQueue[i]->getDepth();
741 // If there are no instructions available, don't try to issue anything, and
742 // don't advance the hazard recognizer.
743 if (AvailableQueue.empty()) {
744 CurCycle = MinDepth != ~0u ? MinDepth : CurCycle + 1;
748 SUnit *FoundSUnit = 0;
750 bool HasNoopHazards = false;
751 while (!AvailableQueue.empty()) {
752 SUnit *CurSUnit = AvailableQueue.pop();
754 ScheduleHazardRecognizer::HazardType HT =
755 HazardRec->getHazardType(CurSUnit);
756 if (HT == ScheduleHazardRecognizer::NoHazard) {
757 FoundSUnit = CurSUnit;
761 // Remember if this is a noop hazard.
762 HasNoopHazards |= HT == ScheduleHazardRecognizer::NoopHazard;
764 NotReady.push_back(CurSUnit);
767 // Add the nodes that aren't ready back onto the available list.
768 if (!NotReady.empty()) {
769 AvailableQueue.push_all(NotReady);
773 // If we found a node to schedule, do it now.
775 ScheduleNodeTopDown(FoundSUnit, CurCycle);
776 HazardRec->EmitInstruction(FoundSUnit);
778 // If this is a pseudo-op node, we don't want to increment the current
780 if (FoundSUnit->Latency) // Don't increment CurCycle for pseudo-ops!
782 } else if (!HasNoopHazards) {
783 // Otherwise, we have a pipeline stall, but no other problem, just advance
784 // the current cycle and try again.
785 DOUT << "*** Advancing cycle, no work to do\n";
786 HazardRec->AdvanceCycle();
790 // Otherwise, we have no instructions to issue and we have instructions
791 // that will fault if we don't do this right. This is the case for
792 // processors without pipeline interlocks and other cases.
793 DOUT << "*** Emitting noop\n";
794 HazardRec->EmitNoop();
795 Sequence.push_back(0); // NULL here means noop
802 VerifySchedule(/*isBottomUp=*/false);
806 //===----------------------------------------------------------------------===//
807 // Public Constructor Functions
808 //===----------------------------------------------------------------------===//
810 FunctionPass *llvm::createPostRAScheduler() {
811 return new PostRAScheduler();