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/TargetLowering.h"
32 #include "llvm/Target/TargetMachine.h"
33 #include "llvm/Target/TargetInstrInfo.h"
34 #include "llvm/Target/TargetRegisterInfo.h"
35 #include "llvm/Support/Compiler.h"
36 #include "llvm/Support/Debug.h"
37 #include "llvm/ADT/Statistic.h"
41 STATISTIC(NumNoops, "Number of noops inserted");
42 STATISTIC(NumStalls, "Number of pipeline stalls");
45 EnableAntiDepBreaking("break-anti-dependencies",
46 cl::desc("Break post-RA scheduling anti-dependencies"),
47 cl::init(true), cl::Hidden);
50 EnablePostRAHazardAvoidance("avoid-hazards",
51 cl::desc("Enable simple hazard-avoidance"),
52 cl::init(true), cl::Hidden);
55 class VISIBILITY_HIDDEN PostRAScheduler : public MachineFunctionPass {
58 PostRAScheduler() : MachineFunctionPass(&ID) {}
60 void getAnalysisUsage(AnalysisUsage &AU) const {
61 AU.addRequired<MachineDominatorTree>();
62 AU.addPreserved<MachineDominatorTree>();
63 AU.addRequired<MachineLoopInfo>();
64 AU.addPreserved<MachineLoopInfo>();
65 MachineFunctionPass::getAnalysisUsage(AU);
68 const char *getPassName() const {
69 return "Post RA top-down list latency scheduler";
72 bool runOnMachineFunction(MachineFunction &Fn);
74 char PostRAScheduler::ID = 0;
76 class VISIBILITY_HIDDEN SchedulePostRATDList : public ScheduleDAGInstrs {
77 /// AvailableQueue - The priority queue to use for the available SUnits.
79 LatencyPriorityQueue AvailableQueue;
81 /// PendingQueue - This contains all of the instructions whose operands have
82 /// been issued, but their results are not ready yet (due to the latency of
83 /// the operation). Once the operands becomes available, the instruction is
84 /// added to the AvailableQueue.
85 std::vector<SUnit*> PendingQueue;
87 /// Topo - A topological ordering for SUnits.
88 ScheduleDAGTopologicalSort Topo;
90 /// AllocatableSet - The set of allocatable registers.
91 /// We'll be ignoring anti-dependencies on non-allocatable registers,
92 /// because they may not be safe to break.
93 const BitVector AllocatableSet;
95 /// HazardRec - The hazard recognizer to use.
96 ScheduleHazardRecognizer *HazardRec;
98 /// Classes - For live regs that are only used in one register class in a
99 /// live range, the register class. If the register is not live, the
100 /// corresponding value is null. If the register is live but used in
101 /// multiple register classes, the corresponding value is -1 casted to a
103 const TargetRegisterClass *
104 Classes[TargetRegisterInfo::FirstVirtualRegister];
106 /// RegRegs - Map registers to all their references within a live range.
107 std::multimap<unsigned, MachineOperand *> RegRefs;
109 /// The index of the most recent kill (proceding bottom-up), or ~0u if
110 /// the register is not live.
111 unsigned KillIndices[TargetRegisterInfo::FirstVirtualRegister];
113 /// The index of the most recent complete def (proceding bottom up), or ~0u
114 /// if the register is live.
115 unsigned DefIndices[TargetRegisterInfo::FirstVirtualRegister];
118 SchedulePostRATDList(MachineFunction &MF,
119 const MachineLoopInfo &MLI,
120 const MachineDominatorTree &MDT,
121 ScheduleHazardRecognizer *HR)
122 : ScheduleDAGInstrs(MF, MLI, MDT), Topo(SUnits),
123 AllocatableSet(TRI->getAllocatableSet(MF)),
126 ~SchedulePostRATDList() {
130 /// StartBlock - Initialize register live-range state for scheduling in
133 void StartBlock(MachineBasicBlock *BB);
135 /// Schedule - Schedule the instruction range using list scheduling.
139 /// Observe - Update liveness information to account for the current
140 /// instruction, which will not be scheduled.
142 void Observe(MachineInstr *MI);
144 /// FinishBlock - Clean up register live-range state.
149 void PrescanInstruction(MachineInstr *MI);
150 void ScanInstruction(MachineInstr *MI, unsigned Count);
151 void ReleaseSucc(SUnit *SU, SDep *SuccEdge);
152 void ReleaseSuccessors(SUnit *SU);
153 void ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle);
154 void ListScheduleTopDown();
155 bool BreakAntiDependencies();
158 /// SimpleHazardRecognizer - A *very* simple hazard recognizer. It uses
159 /// a coarse classification and attempts to avoid that instructions of
160 /// a given class aren't grouped too densely together.
161 class SimpleHazardRecognizer : public ScheduleHazardRecognizer {
162 /// Class - A simple classification for SUnits.
167 /// Window - The Class values of the most recently issued
171 /// getClass - Classify the given SUnit.
172 Class getClass(const SUnit *SU) {
173 const MachineInstr *MI = SU->getInstr();
174 const TargetInstrDesc &TID = MI->getDesc();
182 /// Step - Rotate the existing entries in Window and insert the
183 /// given class value in position as the most recent.
185 std::copy(Window+1, array_endof(Window), Window);
186 Window[array_lengthof(Window)-1] = C;
190 SimpleHazardRecognizer() : Window() {}
192 virtual HazardType getHazardType(SUnit *SU) {
193 Class C = getClass(SU);
197 for (unsigned i = 0; i != array_lengthof(Window); ++i)
200 if (Score > array_lengthof(Window) * 2)
205 virtual void EmitInstruction(SUnit *SU) {
209 virtual void AdvanceCycle() {
215 /// isSchedulingBoundary - Test if the given instruction should be
216 /// considered a scheduling boundary. This primarily includes labels
219 static bool isSchedulingBoundary(const MachineInstr *MI,
220 const MachineFunction &MF) {
221 // Terminators and labels can't be scheduled around.
222 if (MI->getDesc().isTerminator() || MI->isLabel())
225 // Don't attempt to schedule around any instruction that modifies
226 // a stack-oriented pointer, as it's unlikely to be profitable. This
227 // saves compile time, because it doesn't require every single
228 // stack slot reference to depend on the instruction that does the
230 const TargetLowering &TLI = *MF.getTarget().getTargetLowering();
231 if (MI->modifiesRegister(TLI.getStackPointerRegisterToSaveRestore()))
237 bool PostRAScheduler::runOnMachineFunction(MachineFunction &Fn) {
238 DOUT << "PostRAScheduler\n";
240 const MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();
241 const MachineDominatorTree &MDT = getAnalysis<MachineDominatorTree>();
242 ScheduleHazardRecognizer *HR = EnablePostRAHazardAvoidance ?
243 new SimpleHazardRecognizer :
244 new ScheduleHazardRecognizer();
246 SchedulePostRATDList Scheduler(Fn, MLI, MDT, HR);
248 // Loop over all of the basic blocks
249 for (MachineFunction::iterator MBB = Fn.begin(), MBBe = Fn.end();
250 MBB != MBBe; ++MBB) {
251 // Initialize register live-range state for scheduling in this block.
252 Scheduler.StartBlock(MBB);
254 // Schedule each sequence of instructions not interrupted by a label
255 // or anything else that effectively needs to shut down scheduling.
256 MachineBasicBlock::iterator Current = MBB->end();
257 for (MachineBasicBlock::iterator I = Current; I != MBB->begin(); ) {
258 MachineInstr *MI = prior(I);
259 if (isSchedulingBoundary(MI, Fn)) {
261 Scheduler.Run(0, MBB, I, Current);
262 Scheduler.EmitSchedule();
264 Scheduler.Observe(MI);
269 Scheduler.Run(0, MBB, MBB->begin(), Current);
270 Scheduler.EmitSchedule();
272 // Clean up register live-range state.
273 Scheduler.FinishBlock();
279 /// StartBlock - Initialize register live-range state for scheduling in
282 void SchedulePostRATDList::StartBlock(MachineBasicBlock *BB) {
283 // Call the superclass.
284 ScheduleDAGInstrs::StartBlock(BB);
286 // Clear out the register class data.
287 std::fill(Classes, array_endof(Classes),
288 static_cast<const TargetRegisterClass *>(0));
290 // Initialize the indices to indicate that no registers are live.
291 std::fill(KillIndices, array_endof(KillIndices), ~0u);
292 std::fill(DefIndices, array_endof(DefIndices), BB->size());
294 // Determine the live-out physregs for this block.
295 if (!BB->empty() && BB->back().getDesc().isReturn())
296 // In a return block, examine the function live-out regs.
297 for (MachineRegisterInfo::liveout_iterator I = MRI.liveout_begin(),
298 E = MRI.liveout_end(); I != E; ++I) {
300 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
301 KillIndices[Reg] = BB->size();
302 DefIndices[Reg] = ~0u;
303 // Repeat, for all aliases.
304 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
305 unsigned AliasReg = *Alias;
306 Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
307 KillIndices[AliasReg] = BB->size();
308 DefIndices[AliasReg] = ~0u;
312 // In a non-return block, examine the live-in regs of all successors.
313 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
314 SE = BB->succ_end(); SI != SE; ++SI)
315 for (MachineBasicBlock::livein_iterator I = (*SI)->livein_begin(),
316 E = (*SI)->livein_end(); I != E; ++I) {
318 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
319 KillIndices[Reg] = BB->size();
320 DefIndices[Reg] = ~0u;
321 // Repeat, for all aliases.
322 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
323 unsigned AliasReg = *Alias;
324 Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
325 KillIndices[AliasReg] = BB->size();
326 DefIndices[AliasReg] = ~0u;
330 // Consider callee-saved registers as live-out, since we're running after
331 // prologue/epilogue insertion so there's no way to add additional
334 // TODO: If the callee saves and restores these, then we can potentially
335 // use them between the save and the restore. To do that, we could scan
336 // the exit blocks to see which of these registers are defined.
337 // Alternatively, callee-saved registers that aren't saved and restored
338 // could be marked live-in in every block.
339 for (const unsigned *I = TRI->getCalleeSavedRegs(); *I; ++I) {
341 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
342 KillIndices[Reg] = BB->size();
343 DefIndices[Reg] = ~0u;
344 // Repeat, for all aliases.
345 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
346 unsigned AliasReg = *Alias;
347 Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
348 KillIndices[AliasReg] = BB->size();
349 DefIndices[AliasReg] = ~0u;
354 /// Schedule - Schedule the instruction range using list scheduling.
356 void SchedulePostRATDList::Schedule() {
357 DOUT << "********** List Scheduling **********\n";
359 // Build the scheduling graph.
362 if (EnableAntiDepBreaking) {
363 if (BreakAntiDependencies()) {
364 // We made changes. Update the dependency graph.
365 // Theoretically we could update the graph in place:
366 // When a live range is changed to use a different register, remove
367 // the def's anti-dependence *and* output-dependence edges due to
368 // that register, and add new anti-dependence and output-dependence
369 // edges based on the next live range of the register.
377 AvailableQueue.initNodes(SUnits);
379 ListScheduleTopDown();
381 AvailableQueue.releaseState();
384 /// Observe - Update liveness information to account for the current
385 /// instruction, which will not be scheduled.
387 void SchedulePostRATDList::Observe(MachineInstr *MI) {
388 PrescanInstruction(MI);
389 ScanInstruction(MI, 0);
392 /// FinishBlock - Clean up register live-range state.
394 void SchedulePostRATDList::FinishBlock() {
397 // Call the superclass.
398 ScheduleDAGInstrs::FinishBlock();
401 /// getInstrOperandRegClass - Return register class of the operand of an
402 /// instruction of the specified TargetInstrDesc.
403 static const TargetRegisterClass*
404 getInstrOperandRegClass(const TargetRegisterInfo *TRI,
405 const TargetInstrDesc &II, unsigned Op) {
406 if (Op >= II.getNumOperands())
408 if (II.OpInfo[Op].isLookupPtrRegClass())
409 return TRI->getPointerRegClass();
410 return TRI->getRegClass(II.OpInfo[Op].RegClass);
413 /// CriticalPathStep - Return the next SUnit after SU on the bottom-up
415 static SDep *CriticalPathStep(SUnit *SU) {
417 unsigned NextDepth = 0;
418 // Find the predecessor edge with the greatest depth.
419 for (SUnit::pred_iterator P = SU->Preds.begin(), PE = SU->Preds.end();
421 SUnit *PredSU = P->getSUnit();
422 unsigned PredLatency = P->getLatency();
423 unsigned PredTotalLatency = PredSU->getDepth() + PredLatency;
424 // In the case of a latency tie, prefer an anti-dependency edge over
425 // other types of edges.
426 if (NextDepth < PredTotalLatency ||
427 (NextDepth == PredTotalLatency && P->getKind() == SDep::Anti)) {
428 NextDepth = PredTotalLatency;
435 void SchedulePostRATDList::PrescanInstruction(MachineInstr *MI) {
436 // Scan the register operands for this instruction and update
437 // Classes and RegRefs.
438 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
439 MachineOperand &MO = MI->getOperand(i);
440 if (!MO.isReg()) continue;
441 unsigned Reg = MO.getReg();
442 if (Reg == 0) continue;
443 const TargetRegisterClass *NewRC =
444 getInstrOperandRegClass(TRI, MI->getDesc(), i);
446 // For now, only allow the register to be changed if its register
447 // class is consistent across all uses.
448 if (!Classes[Reg] && NewRC)
449 Classes[Reg] = NewRC;
450 else if (!NewRC || Classes[Reg] != NewRC)
451 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
453 // Now check for aliases.
454 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
455 // If an alias of the reg is used during the live range, give up.
456 // Note that this allows us to skip checking if AntiDepReg
457 // overlaps with any of the aliases, among other things.
458 unsigned AliasReg = *Alias;
459 if (Classes[AliasReg]) {
460 Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
461 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
465 // If we're still willing to consider this register, note the reference.
466 if (Classes[Reg] != reinterpret_cast<TargetRegisterClass *>(-1))
467 RegRefs.insert(std::make_pair(Reg, &MO));
471 void SchedulePostRATDList::ScanInstruction(MachineInstr *MI,
474 // Proceding upwards, registers that are defed but not used in this
475 // instruction are now dead.
476 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
477 MachineOperand &MO = MI->getOperand(i);
478 if (!MO.isReg()) continue;
479 unsigned Reg = MO.getReg();
480 if (Reg == 0) continue;
481 if (!MO.isDef()) continue;
482 // Ignore two-addr defs.
483 if (MI->isRegReDefinedByTwoAddr(i)) continue;
485 DefIndices[Reg] = Count;
486 KillIndices[Reg] = ~0u;
489 // Repeat, for all subregs.
490 for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
492 unsigned SubregReg = *Subreg;
493 DefIndices[SubregReg] = Count;
494 KillIndices[SubregReg] = ~0u;
495 Classes[SubregReg] = 0;
496 RegRefs.erase(SubregReg);
498 // Conservatively mark super-registers as unusable.
499 for (const unsigned *Super = TRI->getSuperRegisters(Reg);
501 unsigned SuperReg = *Super;
502 Classes[SuperReg] = reinterpret_cast<TargetRegisterClass *>(-1);
505 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
506 MachineOperand &MO = MI->getOperand(i);
507 if (!MO.isReg()) continue;
508 unsigned Reg = MO.getReg();
509 if (Reg == 0) continue;
510 if (!MO.isUse()) continue;
512 const TargetRegisterClass *NewRC =
513 getInstrOperandRegClass(TRI, MI->getDesc(), i);
515 // For now, only allow the register to be changed if its register
516 // class is consistent across all uses.
517 if (!Classes[Reg] && NewRC)
518 Classes[Reg] = NewRC;
519 else if (!NewRC || Classes[Reg] != NewRC)
520 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
522 RegRefs.insert(std::make_pair(Reg, &MO));
524 // It wasn't previously live but now it is, this is a kill.
525 if (KillIndices[Reg] == ~0u) {
526 KillIndices[Reg] = Count;
527 DefIndices[Reg] = ~0u;
529 // Repeat, for all aliases.
530 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
531 unsigned AliasReg = *Alias;
532 if (KillIndices[AliasReg] == ~0u) {
533 KillIndices[AliasReg] = Count;
534 DefIndices[AliasReg] = ~0u;
540 /// BreakAntiDependencies - Identifiy anti-dependencies along the critical path
541 /// of the ScheduleDAG and break them by renaming registers.
543 bool SchedulePostRATDList::BreakAntiDependencies() {
544 // The code below assumes that there is at least one instruction,
545 // so just duck out immediately if the block is empty.
546 if (SUnits.empty()) return false;
548 // Find the node at the bottom of the critical path.
550 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
551 SUnit *SU = &SUnits[i];
552 if (!Max || SU->getDepth() + SU->Latency > Max->getDepth() + Max->Latency)
556 DOUT << "Critical path has total latency "
557 << (Max->getDepth() + Max->Latency) << "\n";
559 // Track progress along the critical path through the SUnit graph as we walk
561 SUnit *CriticalPathSU = Max;
562 MachineInstr *CriticalPathMI = CriticalPathSU->getInstr();
564 // Consider this pattern:
573 // There are three anti-dependencies here, and without special care,
574 // we'd break all of them using the same register:
583 // because at each anti-dependence, B is the first register that
584 // isn't A which is free. This re-introduces anti-dependencies
585 // at all but one of the original anti-dependencies that we were
586 // trying to break. To avoid this, keep track of the most recent
587 // register that each register was replaced with, avoid avoid
588 // using it to repair an anti-dependence on the same register.
589 // This lets us produce this:
598 // This still has an anti-dependence on B, but at least it isn't on the
599 // original critical path.
601 // TODO: If we tracked more than one register here, we could potentially
602 // fix that remaining critical edge too. This is a little more involved,
603 // because unlike the most recent register, less recent registers should
604 // still be considered, though only if no other registers are available.
605 unsigned LastNewReg[TargetRegisterInfo::FirstVirtualRegister] = {};
607 // Attempt to break anti-dependence edges on the critical path. Walk the
608 // instructions from the bottom up, tracking information about liveness
609 // as we go to help determine which registers are available.
610 bool Changed = false;
611 unsigned Count = SUnits.size() - 1;
612 for (MachineBasicBlock::iterator I = End, E = Begin;
614 MachineInstr *MI = --I;
616 // After regalloc, IMPLICIT_DEF instructions aren't safe to treat as
617 // dependence-breaking. In the case of an INSERT_SUBREG, the IMPLICIT_DEF
618 // is left behind appearing to clobber the super-register, while the
619 // subregister needs to remain live. So we just ignore them.
620 if (MI->getOpcode() == TargetInstrInfo::IMPLICIT_DEF)
623 // Check if this instruction has a dependence on the critical path that
624 // is an anti-dependence that we may be able to break. If it is, set
625 // AntiDepReg to the non-zero register associated with the anti-dependence.
627 // We limit our attention to the critical path as a heuristic to avoid
628 // breaking anti-dependence edges that aren't going to significantly
629 // impact the overall schedule. There are a limited number of registers
630 // and we want to save them for the important edges.
632 // TODO: Instructions with multiple defs could have multiple
633 // anti-dependencies. The current code here only knows how to break one
634 // edge per instruction. Note that we'd have to be able to break all of
635 // the anti-dependencies in an instruction in order to be effective.
636 unsigned AntiDepReg = 0;
637 if (MI == CriticalPathMI) {
638 if (SDep *Edge = CriticalPathStep(CriticalPathSU)) {
639 SUnit *NextSU = Edge->getSUnit();
641 // Only consider anti-dependence edges.
642 if (Edge->getKind() == SDep::Anti) {
643 AntiDepReg = Edge->getReg();
644 assert(AntiDepReg != 0 && "Anti-dependence on reg0?");
645 // Don't break anti-dependencies on non-allocatable registers.
646 if (!AllocatableSet.test(AntiDepReg))
649 // If the SUnit has other dependencies on the SUnit that it
650 // anti-depends on, don't bother breaking the anti-dependency
651 // since those edges would prevent such units from being
652 // scheduled past each other regardless.
654 // Also, if there are dependencies on other SUnits with the
655 // same register as the anti-dependency, don't attempt to
657 for (SUnit::pred_iterator P = CriticalPathSU->Preds.begin(),
658 PE = CriticalPathSU->Preds.end(); P != PE; ++P)
659 if (P->getSUnit() == NextSU ?
660 (P->getKind() != SDep::Anti || P->getReg() != AntiDepReg) :
661 (P->getKind() == SDep::Data && P->getReg() == AntiDepReg)) {
667 CriticalPathSU = NextSU;
668 CriticalPathMI = CriticalPathSU->getInstr();
670 // We've reached the end of the critical path.
676 PrescanInstruction(MI);
678 // If this instruction has a use of AntiDepReg, breaking it
680 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
681 MachineOperand &MO = MI->getOperand(i);
682 if (!MO.isReg()) continue;
683 unsigned Reg = MO.getReg();
684 if (Reg == 0) continue;
685 if (MO.isUse() && AntiDepReg == Reg) {
691 // Determine AntiDepReg's register class, if it is live and is
692 // consistently used within a single class.
693 const TargetRegisterClass *RC = AntiDepReg != 0 ? Classes[AntiDepReg] : 0;
694 assert((AntiDepReg == 0 || RC != NULL) &&
695 "Register should be live if it's causing an anti-dependence!");
696 if (RC == reinterpret_cast<TargetRegisterClass *>(-1))
699 // Look for a suitable register to use to break the anti-depenence.
701 // TODO: Instead of picking the first free register, consider which might
703 if (AntiDepReg != 0) {
704 for (TargetRegisterClass::iterator R = RC->allocation_order_begin(MF),
705 RE = RC->allocation_order_end(MF); R != RE; ++R) {
706 unsigned NewReg = *R;
707 // Don't replace a register with itself.
708 if (NewReg == AntiDepReg) continue;
709 // Don't replace a register with one that was recently used to repair
710 // an anti-dependence with this AntiDepReg, because that would
711 // re-introduce that anti-dependence.
712 if (NewReg == LastNewReg[AntiDepReg]) continue;
713 // If NewReg is dead and NewReg's most recent def is not before
714 // AntiDepReg's kill, it's safe to replace AntiDepReg with NewReg.
715 assert(((KillIndices[AntiDepReg] == ~0u) != (DefIndices[AntiDepReg] == ~0u)) &&
716 "Kill and Def maps aren't consistent for AntiDepReg!");
717 assert(((KillIndices[NewReg] == ~0u) != (DefIndices[NewReg] == ~0u)) &&
718 "Kill and Def maps aren't consistent for NewReg!");
719 if (KillIndices[NewReg] == ~0u &&
720 Classes[NewReg] != reinterpret_cast<TargetRegisterClass *>(-1) &&
721 KillIndices[AntiDepReg] <= DefIndices[NewReg]) {
722 DOUT << "Breaking anti-dependence edge on "
723 << TRI->getName(AntiDepReg)
724 << " with " << RegRefs.count(AntiDepReg) << " references"
725 << " using " << TRI->getName(NewReg) << "!\n";
727 // Update the references to the old register to refer to the new
729 std::pair<std::multimap<unsigned, MachineOperand *>::iterator,
730 std::multimap<unsigned, MachineOperand *>::iterator>
731 Range = RegRefs.equal_range(AntiDepReg);
732 for (std::multimap<unsigned, MachineOperand *>::iterator
733 Q = Range.first, QE = Range.second; Q != QE; ++Q)
734 Q->second->setReg(NewReg);
736 // We just went back in time and modified history; the
737 // liveness information for the anti-depenence reg is now
738 // inconsistent. Set the state as if it were dead.
739 Classes[NewReg] = Classes[AntiDepReg];
740 DefIndices[NewReg] = DefIndices[AntiDepReg];
741 KillIndices[NewReg] = KillIndices[AntiDepReg];
743 Classes[AntiDepReg] = 0;
744 DefIndices[AntiDepReg] = KillIndices[AntiDepReg];
745 KillIndices[AntiDepReg] = ~0u;
747 RegRefs.erase(AntiDepReg);
749 LastNewReg[AntiDepReg] = NewReg;
755 ScanInstruction(MI, Count);
757 assert(Count == ~0u && "Count mismatch!");
762 //===----------------------------------------------------------------------===//
763 // Top-Down Scheduling
764 //===----------------------------------------------------------------------===//
766 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to
767 /// the PendingQueue if the count reaches zero. Also update its cycle bound.
768 void SchedulePostRATDList::ReleaseSucc(SUnit *SU, SDep *SuccEdge) {
769 SUnit *SuccSU = SuccEdge->getSUnit();
770 --SuccSU->NumPredsLeft;
773 if (SuccSU->NumPredsLeft < 0) {
774 cerr << "*** Scheduling failed! ***\n";
776 cerr << " has been released too many times!\n";
781 // Compute how many cycles it will be before this actually becomes
782 // available. This is the max of the start time of all predecessors plus
784 SuccSU->setDepthToAtLeast(SU->getDepth() + SuccEdge->getLatency());
786 // If all the node's predecessors are scheduled, this node is ready
787 // to be scheduled. Ignore the special ExitSU node.
788 if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
789 PendingQueue.push_back(SuccSU);
792 /// ReleaseSuccessors - Call ReleaseSucc on each of SU's successors.
793 void SchedulePostRATDList::ReleaseSuccessors(SUnit *SU) {
794 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
796 ReleaseSucc(SU, &*I);
799 /// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending
800 /// count of its successors. If a successor pending count is zero, add it to
801 /// the Available queue.
802 void SchedulePostRATDList::ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle) {
803 DOUT << "*** Scheduling [" << CurCycle << "]: ";
804 DEBUG(SU->dump(this));
806 Sequence.push_back(SU);
807 assert(CurCycle >= SU->getDepth() && "Node scheduled above its depth!");
808 SU->setDepthToAtLeast(CurCycle);
810 ReleaseSuccessors(SU);
811 SU->isScheduled = true;
812 AvailableQueue.ScheduledNode(SU);
815 /// ListScheduleTopDown - The main loop of list scheduling for top-down
817 void SchedulePostRATDList::ListScheduleTopDown() {
818 unsigned CurCycle = 0;
820 // Release any successors of the special Entry node.
821 ReleaseSuccessors(&EntrySU);
823 // All leaves to Available queue.
824 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
825 // It is available if it has no predecessors.
826 if (SUnits[i].Preds.empty()) {
827 AvailableQueue.push(&SUnits[i]);
828 SUnits[i].isAvailable = true;
832 // While Available queue is not empty, grab the node with the highest
833 // priority. If it is not ready put it back. Schedule the node.
834 std::vector<SUnit*> NotReady;
835 Sequence.reserve(SUnits.size());
836 while (!AvailableQueue.empty() || !PendingQueue.empty()) {
837 // Check to see if any of the pending instructions are ready to issue. If
838 // so, add them to the available queue.
839 unsigned MinDepth = ~0u;
840 for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
841 if (PendingQueue[i]->getDepth() <= CurCycle) {
842 AvailableQueue.push(PendingQueue[i]);
843 PendingQueue[i]->isAvailable = true;
844 PendingQueue[i] = PendingQueue.back();
845 PendingQueue.pop_back();
847 } else if (PendingQueue[i]->getDepth() < MinDepth)
848 MinDepth = PendingQueue[i]->getDepth();
851 // If there are no instructions available, don't try to issue anything, and
852 // don't advance the hazard recognizer.
853 if (AvailableQueue.empty()) {
854 CurCycle = MinDepth != ~0u ? MinDepth : CurCycle + 1;
858 SUnit *FoundSUnit = 0;
860 bool HasNoopHazards = false;
861 while (!AvailableQueue.empty()) {
862 SUnit *CurSUnit = AvailableQueue.pop();
864 ScheduleHazardRecognizer::HazardType HT =
865 HazardRec->getHazardType(CurSUnit);
866 if (HT == ScheduleHazardRecognizer::NoHazard) {
867 FoundSUnit = CurSUnit;
871 // Remember if this is a noop hazard.
872 HasNoopHazards |= HT == ScheduleHazardRecognizer::NoopHazard;
874 NotReady.push_back(CurSUnit);
877 // Add the nodes that aren't ready back onto the available list.
878 if (!NotReady.empty()) {
879 AvailableQueue.push_all(NotReady);
883 // If we found a node to schedule, do it now.
885 ScheduleNodeTopDown(FoundSUnit, CurCycle);
886 HazardRec->EmitInstruction(FoundSUnit);
888 // If this is a pseudo-op node, we don't want to increment the current
890 if (FoundSUnit->Latency) // Don't increment CurCycle for pseudo-ops!
892 } else if (!HasNoopHazards) {
893 // Otherwise, we have a pipeline stall, but no other problem, just advance
894 // the current cycle and try again.
895 DOUT << "*** Advancing cycle, no work to do\n";
896 HazardRec->AdvanceCycle();
900 // Otherwise, we have no instructions to issue and we have instructions
901 // that will fault if we don't do this right. This is the case for
902 // processors without pipeline interlocks and other cases.
903 DOUT << "*** Emitting noop\n";
904 HazardRec->EmitNoop();
905 Sequence.push_back(0); // NULL here means noop
912 VerifySchedule(/*isBottomUp=*/false);
916 //===----------------------------------------------------------------------===//
917 // Public Constructor Functions
918 //===----------------------------------------------------------------------===//
920 FunctionPass *llvm::createPostRAScheduler() {
921 return new PostRAScheduler();