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 "ExactHazardRecognizer.h"
23 #include "SimpleHazardRecognizer.h"
24 #include "ScheduleDAGInstrs.h"
25 #include "llvm/CodeGen/Passes.h"
26 #include "llvm/CodeGen/LatencyPriorityQueue.h"
27 #include "llvm/CodeGen/SchedulerRegistry.h"
28 #include "llvm/CodeGen/MachineDominators.h"
29 #include "llvm/CodeGen/MachineFunctionPass.h"
30 #include "llvm/CodeGen/MachineLoopInfo.h"
31 #include "llvm/CodeGen/MachineRegisterInfo.h"
32 #include "llvm/CodeGen/ScheduleHazardRecognizer.h"
33 #include "llvm/Target/TargetLowering.h"
34 #include "llvm/Target/TargetMachine.h"
35 #include "llvm/Target/TargetInstrInfo.h"
36 #include "llvm/Target/TargetRegisterInfo.h"
37 #include "llvm/Target/TargetSubtarget.h"
38 #include "llvm/Support/Compiler.h"
39 #include "llvm/Support/Debug.h"
40 #include "llvm/Support/ErrorHandling.h"
41 #include "llvm/Support/raw_ostream.h"
42 #include "llvm/ADT/Statistic.h"
47 STATISTIC(NumNoops, "Number of noops inserted");
48 STATISTIC(NumStalls, "Number of pipeline stalls");
51 EnableAntiDepBreaking("break-anti-dependencies",
52 cl::desc("Break post-RA scheduling anti-dependencies"),
53 cl::init(true), cl::Hidden);
56 EnablePostRAHazardAvoidance("avoid-hazards",
57 cl::desc("Enable exact hazard avoidance"),
58 cl::init(true), cl::Hidden);
60 // If DebugDiv > 0 then only schedule MBB with (ID % DebugDiv) == DebugMod
62 DebugDiv("postra-sched-debugdiv",
63 cl::desc("Debug control MBBs that are scheduled"),
64 cl::init(0), cl::Hidden);
66 DebugMod("postra-sched-debugmod",
67 cl::desc("Debug control MBBs that are scheduled"),
68 cl::init(0), cl::Hidden);
71 class VISIBILITY_HIDDEN PostRAScheduler : public MachineFunctionPass {
74 PostRAScheduler() : MachineFunctionPass(&ID) {}
76 void getAnalysisUsage(AnalysisUsage &AU) const {
78 AU.addRequired<MachineDominatorTree>();
79 AU.addPreserved<MachineDominatorTree>();
80 AU.addRequired<MachineLoopInfo>();
81 AU.addPreserved<MachineLoopInfo>();
82 MachineFunctionPass::getAnalysisUsage(AU);
85 const char *getPassName() const {
86 return "Post RA top-down list latency scheduler";
89 bool runOnMachineFunction(MachineFunction &Fn);
91 char PostRAScheduler::ID = 0;
93 class VISIBILITY_HIDDEN SchedulePostRATDList : public ScheduleDAGInstrs {
94 /// AvailableQueue - The priority queue to use for the available SUnits.
96 LatencyPriorityQueue AvailableQueue;
98 /// PendingQueue - This contains all of the instructions whose operands have
99 /// been issued, but their results are not ready yet (due to the latency of
100 /// the operation). Once the operands becomes available, the instruction is
101 /// added to the AvailableQueue.
102 std::vector<SUnit*> PendingQueue;
104 /// Topo - A topological ordering for SUnits.
105 ScheduleDAGTopologicalSort Topo;
107 /// AllocatableSet - The set of allocatable registers.
108 /// We'll be ignoring anti-dependencies on non-allocatable registers,
109 /// because they may not be safe to break.
110 const BitVector AllocatableSet;
112 /// HazardRec - The hazard recognizer to use.
113 ScheduleHazardRecognizer *HazardRec;
115 /// Classes - For live regs that are only used in one register class in a
116 /// live range, the register class. If the register is not live, the
117 /// corresponding value is null. If the register is live but used in
118 /// multiple register classes, the corresponding value is -1 casted to a
120 const TargetRegisterClass *
121 Classes[TargetRegisterInfo::FirstVirtualRegister];
123 /// RegRegs - Map registers to all their references within a live range.
124 std::multimap<unsigned, MachineOperand *> RegRefs;
126 /// KillIndices - The index of the most recent kill (proceding bottom-up),
127 /// or ~0u if the register is not live.
128 unsigned KillIndices[TargetRegisterInfo::FirstVirtualRegister];
130 /// DefIndices - The index of the most recent complete def (proceding bottom
131 /// up), or ~0u if the register is live.
132 unsigned DefIndices[TargetRegisterInfo::FirstVirtualRegister];
134 /// KeepRegs - A set of registers which are live and cannot be changed to
135 /// break anti-dependencies.
136 SmallSet<unsigned, 4> KeepRegs;
139 SchedulePostRATDList(MachineFunction &MF,
140 const MachineLoopInfo &MLI,
141 const MachineDominatorTree &MDT,
142 ScheduleHazardRecognizer *HR)
143 : ScheduleDAGInstrs(MF, MLI, MDT), Topo(SUnits),
144 AllocatableSet(TRI->getAllocatableSet(MF)),
147 ~SchedulePostRATDList() {
151 /// StartBlock - Initialize register live-range state for scheduling in
154 void StartBlock(MachineBasicBlock *BB);
156 /// Schedule - Schedule the instruction range using list scheduling.
160 /// FixupKills - Fix register kill flags that have been made
161 /// invalid due to scheduling
163 void FixupKills(MachineBasicBlock *MBB);
165 /// Observe - Update liveness information to account for the current
166 /// instruction, which will not be scheduled.
168 void Observe(MachineInstr *MI, unsigned Count);
170 /// FinishBlock - Clean up register live-range state.
175 void PrescanInstruction(MachineInstr *MI);
176 void ScanInstruction(MachineInstr *MI, unsigned Count);
177 void ReleaseSucc(SUnit *SU, SDep *SuccEdge);
178 void ReleaseSuccessors(SUnit *SU);
179 void ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle);
180 void ListScheduleTopDown();
181 bool BreakAntiDependencies();
182 unsigned findSuitableFreeRegister(unsigned AntiDepReg,
184 const TargetRegisterClass *);
185 void StartBlockForKills(MachineBasicBlock *BB);
187 // ToggleKillFlag - Toggle a register operand kill flag. Other
188 // adjustments may be made to the instruction if necessary. Return
189 // true if the operand has been deleted, false if not.
190 bool ToggleKillFlag(MachineInstr *MI, MachineOperand &MO);
194 /// isSchedulingBoundary - Test if the given instruction should be
195 /// considered a scheduling boundary. This primarily includes labels
198 static bool isSchedulingBoundary(const MachineInstr *MI,
199 const MachineFunction &MF) {
200 // Terminators and labels can't be scheduled around.
201 if (MI->getDesc().isTerminator() || MI->isLabel())
204 // Don't attempt to schedule around any instruction that modifies
205 // a stack-oriented pointer, as it's unlikely to be profitable. This
206 // saves compile time, because it doesn't require every single
207 // stack slot reference to depend on the instruction that does the
209 const TargetLowering &TLI = *MF.getTarget().getTargetLowering();
210 if (MI->modifiesRegister(TLI.getStackPointerRegisterToSaveRestore()))
216 bool PostRAScheduler::runOnMachineFunction(MachineFunction &Fn) {
217 // Check that post-RA scheduling is enabled for this function
218 const TargetSubtarget &ST = Fn.getTarget().getSubtarget<TargetSubtarget>();
219 if (!ST.enablePostRAScheduler())
222 DEBUG(errs() << "PostRAScheduler\n");
224 const MachineLoopInfo &MLI = getAnalysis<MachineLoopInfo>();
225 const MachineDominatorTree &MDT = getAnalysis<MachineDominatorTree>();
226 const InstrItineraryData &InstrItins = Fn.getTarget().getInstrItineraryData();
227 ScheduleHazardRecognizer *HR = EnablePostRAHazardAvoidance ?
228 (ScheduleHazardRecognizer *)new ExactHazardRecognizer(InstrItins) :
229 (ScheduleHazardRecognizer *)new SimpleHazardRecognizer();
231 SchedulePostRATDList Scheduler(Fn, MLI, MDT, HR);
233 // Loop over all of the basic blocks
234 for (MachineFunction::iterator MBB = Fn.begin(), MBBe = Fn.end();
235 MBB != MBBe; ++MBB) {
237 // If DebugDiv > 0 then only schedule MBB with (ID % DebugDiv) == DebugMod
239 static int bbcnt = 0;
240 if (bbcnt++ % DebugDiv != DebugMod)
242 errs() << "*** DEBUG scheduling " << Fn.getFunction()->getNameStr() <<
243 ":MBB ID#" << MBB->getNumber() << " ***\n";
247 // Initialize register live-range state for scheduling in this block.
248 Scheduler.StartBlock(MBB);
250 // Schedule each sequence of instructions not interrupted by a label
251 // or anything else that effectively needs to shut down scheduling.
252 MachineBasicBlock::iterator Current = MBB->end();
253 unsigned Count = MBB->size(), CurrentCount = Count;
254 for (MachineBasicBlock::iterator I = Current; I != MBB->begin(); ) {
255 MachineInstr *MI = prior(I);
256 if (isSchedulingBoundary(MI, Fn)) {
257 Scheduler.Run(MBB, I, Current, CurrentCount);
258 Scheduler.EmitSchedule(0);
260 CurrentCount = Count - 1;
261 Scheduler.Observe(MI, CurrentCount);
266 assert(Count == 0 && "Instruction count mismatch!");
267 assert((MBB->begin() == Current || CurrentCount != 0) &&
268 "Instruction count mismatch!");
269 Scheduler.Run(MBB, MBB->begin(), Current, CurrentCount);
270 Scheduler.EmitSchedule(0);
272 // Clean up register live-range state.
273 Scheduler.FinishBlock();
275 // Update register kills
276 Scheduler.FixupKills(MBB);
282 /// StartBlock - Initialize register live-range state for scheduling in
285 void SchedulePostRATDList::StartBlock(MachineBasicBlock *BB) {
286 // Call the superclass.
287 ScheduleDAGInstrs::StartBlock(BB);
289 // Reset the hazard recognizer.
292 // Clear out the register class data.
293 std::fill(Classes, array_endof(Classes),
294 static_cast<const TargetRegisterClass *>(0));
296 // Initialize the indices to indicate that no registers are live.
297 std::fill(KillIndices, array_endof(KillIndices), ~0u);
298 std::fill(DefIndices, array_endof(DefIndices), BB->size());
300 // Clear "do not change" set.
303 // Determine the live-out physregs for this block.
304 if (!BB->empty() && BB->back().getDesc().isReturn())
305 // In a return block, examine the function live-out regs.
306 for (MachineRegisterInfo::liveout_iterator I = MRI.liveout_begin(),
307 E = MRI.liveout_end(); I != E; ++I) {
309 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
310 KillIndices[Reg] = BB->size();
311 DefIndices[Reg] = ~0u;
312 // Repeat, for all aliases.
313 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
314 unsigned AliasReg = *Alias;
315 Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
316 KillIndices[AliasReg] = BB->size();
317 DefIndices[AliasReg] = ~0u;
321 // In a non-return block, examine the live-in regs of all successors.
322 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
323 SE = BB->succ_end(); SI != SE; ++SI)
324 for (MachineBasicBlock::livein_iterator I = (*SI)->livein_begin(),
325 E = (*SI)->livein_end(); I != E; ++I) {
327 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
328 KillIndices[Reg] = BB->size();
329 DefIndices[Reg] = ~0u;
330 // Repeat, for all aliases.
331 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
332 unsigned AliasReg = *Alias;
333 Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
334 KillIndices[AliasReg] = BB->size();
335 DefIndices[AliasReg] = ~0u;
339 // Consider callee-saved registers as live-out, since we're running after
340 // prologue/epilogue insertion so there's no way to add additional
343 // TODO: there is a new method
344 // MachineFrameInfo::getPristineRegs(MBB). It gives you a list of
345 // CSRs that have not been saved when entering the MBB. The
346 // remaining CSRs have been saved and can be treated like call
347 // clobbered registers.
348 for (const unsigned *I = TRI->getCalleeSavedRegs(); *I; ++I) {
350 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
351 KillIndices[Reg] = BB->size();
352 DefIndices[Reg] = ~0u;
353 // Repeat, for all aliases.
354 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
355 unsigned AliasReg = *Alias;
356 Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
357 KillIndices[AliasReg] = BB->size();
358 DefIndices[AliasReg] = ~0u;
363 /// Schedule - Schedule the instruction range using list scheduling.
365 void SchedulePostRATDList::Schedule() {
366 DEBUG(errs() << "********** List Scheduling **********\n");
368 // Build the scheduling graph.
371 if (EnableAntiDepBreaking) {
372 if (BreakAntiDependencies()) {
373 // We made changes. Update the dependency graph.
374 // Theoretically we could update the graph in place:
375 // When a live range is changed to use a different register, remove
376 // the def's anti-dependence *and* output-dependence edges due to
377 // that register, and add new anti-dependence and output-dependence
378 // edges based on the next live range of the register.
386 DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su)
387 SUnits[su].dumpAll(this));
389 AvailableQueue.initNodes(SUnits);
391 ListScheduleTopDown();
393 AvailableQueue.releaseState();
396 /// Observe - Update liveness information to account for the current
397 /// instruction, which will not be scheduled.
399 void SchedulePostRATDList::Observe(MachineInstr *MI, unsigned Count) {
400 assert(Count < InsertPosIndex && "Instruction index out of expected range!");
402 // Any register which was defined within the previous scheduling region
403 // may have been rescheduled and its lifetime may overlap with registers
404 // in ways not reflected in our current liveness state. For each such
405 // register, adjust the liveness state to be conservatively correct.
406 for (unsigned Reg = 0; Reg != TargetRegisterInfo::FirstVirtualRegister; ++Reg)
407 if (DefIndices[Reg] < InsertPosIndex && DefIndices[Reg] >= Count) {
408 assert(KillIndices[Reg] == ~0u && "Clobbered register is live!");
409 // Mark this register to be non-renamable.
410 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
411 // Move the def index to the end of the previous region, to reflect
412 // that the def could theoretically have been scheduled at the end.
413 DefIndices[Reg] = InsertPosIndex;
416 PrescanInstruction(MI);
417 ScanInstruction(MI, Count);
420 /// FinishBlock - Clean up register live-range state.
422 void SchedulePostRATDList::FinishBlock() {
425 // Call the superclass.
426 ScheduleDAGInstrs::FinishBlock();
429 /// CriticalPathStep - Return the next SUnit after SU on the bottom-up
431 static SDep *CriticalPathStep(SUnit *SU) {
433 unsigned NextDepth = 0;
434 // Find the predecessor edge with the greatest depth.
435 for (SUnit::pred_iterator P = SU->Preds.begin(), PE = SU->Preds.end();
437 SUnit *PredSU = P->getSUnit();
438 unsigned PredLatency = P->getLatency();
439 unsigned PredTotalLatency = PredSU->getDepth() + PredLatency;
440 // In the case of a latency tie, prefer an anti-dependency edge over
441 // other types of edges.
442 if (NextDepth < PredTotalLatency ||
443 (NextDepth == PredTotalLatency && P->getKind() == SDep::Anti)) {
444 NextDepth = PredTotalLatency;
451 void SchedulePostRATDList::PrescanInstruction(MachineInstr *MI) {
452 // Scan the register operands for this instruction and update
453 // Classes and RegRefs.
454 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
455 MachineOperand &MO = MI->getOperand(i);
456 if (!MO.isReg()) continue;
457 unsigned Reg = MO.getReg();
458 if (Reg == 0) continue;
459 const TargetRegisterClass *NewRC = 0;
461 if (i < MI->getDesc().getNumOperands())
462 NewRC = MI->getDesc().OpInfo[i].getRegClass(TRI);
464 // For now, only allow the register to be changed if its register
465 // class is consistent across all uses.
466 if (!Classes[Reg] && NewRC)
467 Classes[Reg] = NewRC;
468 else if (!NewRC || Classes[Reg] != NewRC)
469 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
471 // Now check for aliases.
472 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
473 // If an alias of the reg is used during the live range, give up.
474 // Note that this allows us to skip checking if AntiDepReg
475 // overlaps with any of the aliases, among other things.
476 unsigned AliasReg = *Alias;
477 if (Classes[AliasReg]) {
478 Classes[AliasReg] = reinterpret_cast<TargetRegisterClass *>(-1);
479 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
483 // If we're still willing to consider this register, note the reference.
484 if (Classes[Reg] != reinterpret_cast<TargetRegisterClass *>(-1))
485 RegRefs.insert(std::make_pair(Reg, &MO));
489 void SchedulePostRATDList::ScanInstruction(MachineInstr *MI,
492 // Proceding upwards, registers that are defed but not used in this
493 // instruction are now dead.
494 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
495 MachineOperand &MO = MI->getOperand(i);
496 if (!MO.isReg()) continue;
497 unsigned Reg = MO.getReg();
498 if (Reg == 0) continue;
499 if (!MO.isDef()) continue;
500 // Ignore two-addr defs.
501 if (MI->isRegTiedToUseOperand(i)) continue;
503 DefIndices[Reg] = Count;
504 KillIndices[Reg] = ~0u;
505 assert(((KillIndices[Reg] == ~0u) !=
506 (DefIndices[Reg] == ~0u)) &&
507 "Kill and Def maps aren't consistent for Reg!");
511 // Repeat, for all subregs.
512 for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
514 unsigned SubregReg = *Subreg;
515 DefIndices[SubregReg] = Count;
516 KillIndices[SubregReg] = ~0u;
517 KeepRegs.erase(SubregReg);
518 Classes[SubregReg] = 0;
519 RegRefs.erase(SubregReg);
521 // Conservatively mark super-registers as unusable.
522 for (const unsigned *Super = TRI->getSuperRegisters(Reg);
524 unsigned SuperReg = *Super;
525 Classes[SuperReg] = reinterpret_cast<TargetRegisterClass *>(-1);
528 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
529 MachineOperand &MO = MI->getOperand(i);
530 if (!MO.isReg()) continue;
531 unsigned Reg = MO.getReg();
532 if (Reg == 0) continue;
533 if (!MO.isUse()) continue;
535 const TargetRegisterClass *NewRC = 0;
536 if (i < MI->getDesc().getNumOperands())
537 NewRC = MI->getDesc().OpInfo[i].getRegClass(TRI);
539 // For now, only allow the register to be changed if its register
540 // class is consistent across all uses.
541 if (!Classes[Reg] && NewRC)
542 Classes[Reg] = NewRC;
543 else if (!NewRC || Classes[Reg] != NewRC)
544 Classes[Reg] = reinterpret_cast<TargetRegisterClass *>(-1);
546 RegRefs.insert(std::make_pair(Reg, &MO));
548 // It wasn't previously live but now it is, this is a kill.
549 if (KillIndices[Reg] == ~0u) {
550 KillIndices[Reg] = Count;
551 DefIndices[Reg] = ~0u;
552 assert(((KillIndices[Reg] == ~0u) !=
553 (DefIndices[Reg] == ~0u)) &&
554 "Kill and Def maps aren't consistent for Reg!");
556 // Repeat, for all aliases.
557 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
558 unsigned AliasReg = *Alias;
559 if (KillIndices[AliasReg] == ~0u) {
560 KillIndices[AliasReg] = Count;
561 DefIndices[AliasReg] = ~0u;
568 SchedulePostRATDList::findSuitableFreeRegister(unsigned AntiDepReg,
570 const TargetRegisterClass *RC) {
571 for (TargetRegisterClass::iterator R = RC->allocation_order_begin(MF),
572 RE = RC->allocation_order_end(MF); R != RE; ++R) {
573 unsigned NewReg = *R;
574 // Don't replace a register with itself.
575 if (NewReg == AntiDepReg) continue;
576 // Don't replace a register with one that was recently used to repair
577 // an anti-dependence with this AntiDepReg, because that would
578 // re-introduce that anti-dependence.
579 if (NewReg == LastNewReg) continue;
580 // If NewReg is dead and NewReg's most recent def is not before
581 // AntiDepReg's kill, it's safe to replace AntiDepReg with NewReg.
582 assert(((KillIndices[AntiDepReg] == ~0u) != (DefIndices[AntiDepReg] == ~0u)) &&
583 "Kill and Def maps aren't consistent for AntiDepReg!");
584 assert(((KillIndices[NewReg] == ~0u) != (DefIndices[NewReg] == ~0u)) &&
585 "Kill and Def maps aren't consistent for NewReg!");
586 if (KillIndices[NewReg] != ~0u ||
587 Classes[NewReg] == reinterpret_cast<TargetRegisterClass *>(-1) ||
588 KillIndices[AntiDepReg] > DefIndices[NewReg])
593 // No registers are free and available!
597 /// BreakAntiDependencies - Identifiy anti-dependencies along the critical path
598 /// of the ScheduleDAG and break them by renaming registers.
600 bool SchedulePostRATDList::BreakAntiDependencies() {
601 // The code below assumes that there is at least one instruction,
602 // so just duck out immediately if the block is empty.
603 if (SUnits.empty()) return false;
605 // Find the node at the bottom of the critical path.
607 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
608 SUnit *SU = &SUnits[i];
609 if (!Max || SU->getDepth() + SU->Latency > Max->getDepth() + Max->Latency)
613 DEBUG(errs() << "Critical path has total latency "
614 << (Max->getDepth() + Max->Latency) << "\n");
616 // Track progress along the critical path through the SUnit graph as we walk
618 SUnit *CriticalPathSU = Max;
619 MachineInstr *CriticalPathMI = CriticalPathSU->getInstr();
621 // Consider this pattern:
630 // There are three anti-dependencies here, and without special care,
631 // we'd break all of them using the same register:
640 // because at each anti-dependence, B is the first register that
641 // isn't A which is free. This re-introduces anti-dependencies
642 // at all but one of the original anti-dependencies that we were
643 // trying to break. To avoid this, keep track of the most recent
644 // register that each register was replaced with, avoid
645 // using it to repair an anti-dependence on the same register.
646 // This lets us produce this:
655 // This still has an anti-dependence on B, but at least it isn't on the
656 // original critical path.
658 // TODO: If we tracked more than one register here, we could potentially
659 // fix that remaining critical edge too. This is a little more involved,
660 // because unlike the most recent register, less recent registers should
661 // still be considered, though only if no other registers are available.
662 unsigned LastNewReg[TargetRegisterInfo::FirstVirtualRegister] = {};
664 // Attempt to break anti-dependence edges on the critical path. Walk the
665 // instructions from the bottom up, tracking information about liveness
666 // as we go to help determine which registers are available.
667 bool Changed = false;
668 unsigned Count = InsertPosIndex - 1;
669 for (MachineBasicBlock::iterator I = InsertPos, E = Begin;
671 MachineInstr *MI = --I;
673 // After regalloc, KILL instructions aren't safe to treat as
674 // dependence-breaking. In the case of an INSERT_SUBREG, the KILL
675 // is left behind appearing to clobber the super-register, while the
676 // subregister needs to remain live. So we just ignore them.
677 if (MI->getOpcode() == TargetInstrInfo::KILL)
680 // Check if this instruction has a dependence on the critical path that
681 // is an anti-dependence that we may be able to break. If it is, set
682 // AntiDepReg to the non-zero register associated with the anti-dependence.
684 // We limit our attention to the critical path as a heuristic to avoid
685 // breaking anti-dependence edges that aren't going to significantly
686 // impact the overall schedule. There are a limited number of registers
687 // and we want to save them for the important edges.
689 // TODO: Instructions with multiple defs could have multiple
690 // anti-dependencies. The current code here only knows how to break one
691 // edge per instruction. Note that we'd have to be able to break all of
692 // the anti-dependencies in an instruction in order to be effective.
693 unsigned AntiDepReg = 0;
694 if (MI == CriticalPathMI) {
695 if (SDep *Edge = CriticalPathStep(CriticalPathSU)) {
696 SUnit *NextSU = Edge->getSUnit();
698 // Only consider anti-dependence edges.
699 if (Edge->getKind() == SDep::Anti) {
700 AntiDepReg = Edge->getReg();
701 assert(AntiDepReg != 0 && "Anti-dependence on reg0?");
702 if (!AllocatableSet.test(AntiDepReg))
703 // Don't break anti-dependencies on non-allocatable registers.
705 else if (KeepRegs.count(AntiDepReg))
706 // Don't break anti-dependencies if an use down below requires
707 // this exact register.
710 // If the SUnit has other dependencies on the SUnit that it
711 // anti-depends on, don't bother breaking the anti-dependency
712 // since those edges would prevent such units from being
713 // scheduled past each other regardless.
715 // Also, if there are dependencies on other SUnits with the
716 // same register as the anti-dependency, don't attempt to
718 for (SUnit::pred_iterator P = CriticalPathSU->Preds.begin(),
719 PE = CriticalPathSU->Preds.end(); P != PE; ++P)
720 if (P->getSUnit() == NextSU ?
721 (P->getKind() != SDep::Anti || P->getReg() != AntiDepReg) :
722 (P->getKind() == SDep::Data && P->getReg() == AntiDepReg)) {
728 CriticalPathSU = NextSU;
729 CriticalPathMI = CriticalPathSU->getInstr();
731 // We've reached the end of the critical path.
737 PrescanInstruction(MI);
739 if (MI->getDesc().hasExtraSrcRegAllocReq()) {
740 // It's not safe to change register allocation for source operands of
741 // that have special allocation requirements.
742 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
743 MachineOperand &MO = MI->getOperand(i);
744 if (!MO.isReg()) continue;
745 unsigned Reg = MO.getReg();
746 if (Reg == 0) continue;
748 if (KeepRegs.insert(Reg)) {
749 for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
751 KeepRegs.insert(*Subreg);
757 if (MI->getDesc().hasExtraDefRegAllocReq())
758 // If this instruction's defs have special allocation requirement, don't
759 // break this anti-dependency.
761 else if (AntiDepReg) {
762 // If this instruction has a use of AntiDepReg, breaking it
764 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
765 MachineOperand &MO = MI->getOperand(i);
766 if (!MO.isReg()) continue;
767 unsigned Reg = MO.getReg();
768 if (Reg == 0) continue;
769 if (MO.isUse() && AntiDepReg == Reg) {
776 // Determine AntiDepReg's register class, if it is live and is
777 // consistently used within a single class.
778 const TargetRegisterClass *RC = AntiDepReg != 0 ? Classes[AntiDepReg] : 0;
779 assert((AntiDepReg == 0 || RC != NULL) &&
780 "Register should be live if it's causing an anti-dependence!");
781 if (RC == reinterpret_cast<TargetRegisterClass *>(-1))
784 // Look for a suitable register to use to break the anti-depenence.
786 // TODO: Instead of picking the first free register, consider which might
788 if (AntiDepReg != 0) {
789 if (unsigned NewReg = findSuitableFreeRegister(AntiDepReg,
790 LastNewReg[AntiDepReg],
792 DEBUG(errs() << "Breaking anti-dependence edge on "
793 << TRI->getName(AntiDepReg)
794 << " with " << RegRefs.count(AntiDepReg) << " references"
795 << " using " << TRI->getName(NewReg) << "!\n");
797 // Update the references to the old register to refer to the new
799 std::pair<std::multimap<unsigned, MachineOperand *>::iterator,
800 std::multimap<unsigned, MachineOperand *>::iterator>
801 Range = RegRefs.equal_range(AntiDepReg);
802 for (std::multimap<unsigned, MachineOperand *>::iterator
803 Q = Range.first, QE = Range.second; Q != QE; ++Q)
804 Q->second->setReg(NewReg);
806 // We just went back in time and modified history; the
807 // liveness information for the anti-depenence reg is now
808 // inconsistent. Set the state as if it were dead.
809 Classes[NewReg] = Classes[AntiDepReg];
810 DefIndices[NewReg] = DefIndices[AntiDepReg];
811 KillIndices[NewReg] = KillIndices[AntiDepReg];
812 assert(((KillIndices[NewReg] == ~0u) !=
813 (DefIndices[NewReg] == ~0u)) &&
814 "Kill and Def maps aren't consistent for NewReg!");
816 Classes[AntiDepReg] = 0;
817 DefIndices[AntiDepReg] = KillIndices[AntiDepReg];
818 KillIndices[AntiDepReg] = ~0u;
819 assert(((KillIndices[AntiDepReg] == ~0u) !=
820 (DefIndices[AntiDepReg] == ~0u)) &&
821 "Kill and Def maps aren't consistent for AntiDepReg!");
823 RegRefs.erase(AntiDepReg);
825 LastNewReg[AntiDepReg] = NewReg;
829 ScanInstruction(MI, Count);
835 /// StartBlockForKills - Initialize register live-range state for updating kills
837 void SchedulePostRATDList::StartBlockForKills(MachineBasicBlock *BB) {
838 // Initialize the indices to indicate that no registers are live.
839 std::fill(KillIndices, array_endof(KillIndices), ~0u);
841 // Determine the live-out physregs for this block.
842 if (!BB->empty() && BB->back().getDesc().isReturn()) {
843 // In a return block, examine the function live-out regs.
844 for (MachineRegisterInfo::liveout_iterator I = MRI.liveout_begin(),
845 E = MRI.liveout_end(); I != E; ++I) {
847 KillIndices[Reg] = BB->size();
848 // Repeat, for all subregs.
849 for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
851 KillIndices[*Subreg] = BB->size();
856 // In a non-return block, examine the live-in regs of all successors.
857 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
858 SE = BB->succ_end(); SI != SE; ++SI) {
859 for (MachineBasicBlock::livein_iterator I = (*SI)->livein_begin(),
860 E = (*SI)->livein_end(); I != E; ++I) {
862 KillIndices[Reg] = BB->size();
863 // Repeat, for all subregs.
864 for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
866 KillIndices[*Subreg] = BB->size();
873 bool SchedulePostRATDList::ToggleKillFlag(MachineInstr *MI,
874 MachineOperand &MO) {
875 // Setting kill flag...
881 // If MO itself is live, clear the kill flag...
882 if (KillIndices[MO.getReg()] != ~0u) {
887 // If any subreg of MO is live, then create an imp-def for that
888 // subreg and keep MO marked as killed.
890 const unsigned SuperReg = MO.getReg();
891 for (const unsigned *Subreg = TRI->getSubRegisters(SuperReg);
893 if (KillIndices[*Subreg] != ~0u) {
894 MI->addOperand(MachineOperand::CreateReg(*Subreg,
903 MO.setIsKill(AllDead);
907 /// FixupKills - Fix the register kill flags, they may have been made
908 /// incorrect by instruction reordering.
910 void SchedulePostRATDList::FixupKills(MachineBasicBlock *MBB) {
911 DEBUG(errs() << "Fixup kills for BB ID#" << MBB->getNumber() << '\n');
913 std::set<unsigned> killedRegs;
914 BitVector ReservedRegs = TRI->getReservedRegs(MF);
916 StartBlockForKills(MBB);
918 // Examine block from end to start...
919 unsigned Count = MBB->size();
920 for (MachineBasicBlock::iterator I = MBB->end(), E = MBB->begin();
922 MachineInstr *MI = --I;
924 // Update liveness. Registers that are defed but not used in this
925 // instruction are now dead. Mark register and all subregs as they
926 // are completely defined.
927 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
928 MachineOperand &MO = MI->getOperand(i);
929 if (!MO.isReg()) continue;
930 unsigned Reg = MO.getReg();
931 if (Reg == 0) continue;
932 if (!MO.isDef()) continue;
933 // Ignore two-addr defs.
934 if (MI->isRegTiedToUseOperand(i)) continue;
936 KillIndices[Reg] = ~0u;
938 // Repeat for all subregs.
939 for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
941 KillIndices[*Subreg] = ~0u;
945 // Examine all used registers and set/clear kill flag. When a
946 // register is used multiple times we only set the kill flag on
949 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
950 MachineOperand &MO = MI->getOperand(i);
951 if (!MO.isReg() || !MO.isUse()) continue;
952 unsigned Reg = MO.getReg();
953 if ((Reg == 0) || ReservedRegs.test(Reg)) continue;
956 if (killedRegs.find(Reg) == killedRegs.end()) {
958 // A register is not killed if any subregs are live...
959 for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
961 if (KillIndices[*Subreg] != ~0u) {
967 // If subreg is not live, then register is killed if it became
968 // live in this instruction
970 kill = (KillIndices[Reg] == ~0u);
973 if (MO.isKill() != kill) {
974 bool removed = ToggleKillFlag(MI, MO);
976 DEBUG(errs() << "Fixed <removed> in ");
978 DEBUG(errs() << "Fixed " << MO << " in ");
983 killedRegs.insert(Reg);
986 // Mark any used register (that is not using undef) and subregs as
988 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
989 MachineOperand &MO = MI->getOperand(i);
990 if (!MO.isReg() || !MO.isUse() || MO.isUndef()) continue;
991 unsigned Reg = MO.getReg();
992 if ((Reg == 0) || ReservedRegs.test(Reg)) continue;
994 KillIndices[Reg] = Count;
996 for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
998 KillIndices[*Subreg] = Count;
1004 //===----------------------------------------------------------------------===//
1005 // Top-Down Scheduling
1006 //===----------------------------------------------------------------------===//
1008 /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to
1009 /// the PendingQueue if the count reaches zero. Also update its cycle bound.
1010 void SchedulePostRATDList::ReleaseSucc(SUnit *SU, SDep *SuccEdge) {
1011 SUnit *SuccSU = SuccEdge->getSUnit();
1014 if (SuccSU->NumPredsLeft == 0) {
1015 errs() << "*** Scheduling failed! ***\n";
1017 errs() << " has been released too many times!\n";
1018 llvm_unreachable(0);
1021 --SuccSU->NumPredsLeft;
1023 // Compute how many cycles it will be before this actually becomes
1024 // available. This is the max of the start time of all predecessors plus
1026 SuccSU->setDepthToAtLeast(SU->getDepth() + SuccEdge->getLatency());
1028 // If all the node's predecessors are scheduled, this node is ready
1029 // to be scheduled. Ignore the special ExitSU node.
1030 if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU)
1031 PendingQueue.push_back(SuccSU);
1034 /// ReleaseSuccessors - Call ReleaseSucc on each of SU's successors.
1035 void SchedulePostRATDList::ReleaseSuccessors(SUnit *SU) {
1036 for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end();
1038 ReleaseSucc(SU, &*I);
1041 /// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending
1042 /// count of its successors. If a successor pending count is zero, add it to
1043 /// the Available queue.
1044 void SchedulePostRATDList::ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle) {
1045 DEBUG(errs() << "*** Scheduling [" << CurCycle << "]: ");
1046 DEBUG(SU->dump(this));
1048 Sequence.push_back(SU);
1049 assert(CurCycle >= SU->getDepth() && "Node scheduled above its depth!");
1050 SU->setDepthToAtLeast(CurCycle);
1052 ReleaseSuccessors(SU);
1053 SU->isScheduled = true;
1054 AvailableQueue.ScheduledNode(SU);
1057 /// ListScheduleTopDown - The main loop of list scheduling for top-down
1059 void SchedulePostRATDList::ListScheduleTopDown() {
1060 unsigned CurCycle = 0;
1062 // Release any successors of the special Entry node.
1063 ReleaseSuccessors(&EntrySU);
1065 // All leaves to Available queue.
1066 for (unsigned i = 0, e = SUnits.size(); i != e; ++i) {
1067 // It is available if it has no predecessors.
1068 if (SUnits[i].Preds.empty()) {
1069 AvailableQueue.push(&SUnits[i]);
1070 SUnits[i].isAvailable = true;
1074 // In any cycle where we can't schedule any instructions, we must
1075 // stall or emit a noop, depending on the target.
1076 bool CycleHasInsts = false;
1078 // While Available queue is not empty, grab the node with the highest
1079 // priority. If it is not ready put it back. Schedule the node.
1080 std::vector<SUnit*> NotReady;
1081 Sequence.reserve(SUnits.size());
1082 while (!AvailableQueue.empty() || !PendingQueue.empty()) {
1083 // Check to see if any of the pending instructions are ready to issue. If
1084 // so, add them to the available queue.
1085 unsigned MinDepth = ~0u;
1086 for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) {
1087 if (PendingQueue[i]->getDepth() <= CurCycle) {
1088 AvailableQueue.push(PendingQueue[i]);
1089 PendingQueue[i]->isAvailable = true;
1090 PendingQueue[i] = PendingQueue.back();
1091 PendingQueue.pop_back();
1093 } else if (PendingQueue[i]->getDepth() < MinDepth)
1094 MinDepth = PendingQueue[i]->getDepth();
1097 DEBUG(errs() << "\n*** Examining Available\n";
1098 LatencyPriorityQueue q = AvailableQueue;
1099 while (!q.empty()) {
1100 SUnit *su = q.pop();
1101 errs() << "Height " << su->getHeight() << ": ";
1105 SUnit *FoundSUnit = 0;
1107 bool HasNoopHazards = false;
1108 while (!AvailableQueue.empty()) {
1109 SUnit *CurSUnit = AvailableQueue.pop();
1111 ScheduleHazardRecognizer::HazardType HT =
1112 HazardRec->getHazardType(CurSUnit);
1113 if (HT == ScheduleHazardRecognizer::NoHazard) {
1114 FoundSUnit = CurSUnit;
1118 // Remember if this is a noop hazard.
1119 HasNoopHazards |= HT == ScheduleHazardRecognizer::NoopHazard;
1121 NotReady.push_back(CurSUnit);
1124 // Add the nodes that aren't ready back onto the available list.
1125 if (!NotReady.empty()) {
1126 AvailableQueue.push_all(NotReady);
1130 // If we found a node to schedule, do it now.
1132 ScheduleNodeTopDown(FoundSUnit, CurCycle);
1133 HazardRec->EmitInstruction(FoundSUnit);
1134 CycleHasInsts = true;
1136 // If we are using the target-specific hazards, then don't
1137 // advance the cycle time just because we schedule a node. If
1138 // the target allows it we can schedule multiple nodes in the
1140 if (!EnablePostRAHazardAvoidance) {
1141 if (FoundSUnit->Latency) // Don't increment CurCycle for pseudo-ops!
1145 if (CycleHasInsts) {
1146 DEBUG(errs() << "*** Finished cycle " << CurCycle << '\n');
1147 HazardRec->AdvanceCycle();
1148 } else if (!HasNoopHazards) {
1149 // Otherwise, we have a pipeline stall, but no other problem,
1150 // just advance the current cycle and try again.
1151 DEBUG(errs() << "*** Stall in cycle " << CurCycle << '\n');
1152 HazardRec->AdvanceCycle();
1155 // Otherwise, we have no instructions to issue and we have instructions
1156 // that will fault if we don't do this right. This is the case for
1157 // processors without pipeline interlocks and other cases.
1158 DEBUG(errs() << "*** Emitting noop in cycle " << CurCycle << '\n');
1159 HazardRec->EmitNoop();
1160 Sequence.push_back(0); // NULL here means noop
1165 CycleHasInsts = false;
1170 VerifySchedule(/*isBottomUp=*/false);
1174 //===----------------------------------------------------------------------===//
1175 // Public Constructor Functions
1176 //===----------------------------------------------------------------------===//
1178 FunctionPass *llvm::createPostRAScheduler() {
1179 return new PostRAScheduler();