1 //===---- ScheduleDAGInstrs.cpp - MachineInstr Rescheduling ---------------===//
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 the ScheduleDAGInstrs class, which implements re-scheduling
13 //===----------------------------------------------------------------------===//
15 #define DEBUG_TYPE "sched-instrs"
16 #include "ScheduleDAGInstrs.h"
17 #include "llvm/Operator.h"
18 #include "llvm/Analysis/AliasAnalysis.h"
19 #include "llvm/CodeGen/MachineFunctionPass.h"
20 #include "llvm/CodeGen/MachineMemOperand.h"
21 #include "llvm/CodeGen/MachineRegisterInfo.h"
22 #include "llvm/CodeGen/PseudoSourceValue.h"
23 #include "llvm/Target/TargetMachine.h"
24 #include "llvm/Target/TargetInstrInfo.h"
25 #include "llvm/Target/TargetRegisterInfo.h"
26 #include "llvm/Target/TargetSubtarget.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/raw_ostream.h"
29 #include "llvm/ADT/SmallSet.h"
32 ScheduleDAGInstrs::ScheduleDAGInstrs(MachineFunction &mf,
33 const MachineLoopInfo &mli,
34 const MachineDominatorTree &mdt)
35 : ScheduleDAG(mf), MLI(mli), MDT(mdt), LoopRegs(MLI, MDT) {
36 MFI = mf.getFrameInfo();
40 /// Run - perform scheduling.
42 void ScheduleDAGInstrs::Run(MachineBasicBlock *bb,
43 MachineBasicBlock::iterator begin,
44 MachineBasicBlock::iterator end,
48 InsertPosIndex = endcount;
50 ScheduleDAG::Run(bb, end);
53 /// getUnderlyingObjectFromInt - This is the function that does the work of
54 /// looking through basic ptrtoint+arithmetic+inttoptr sequences.
55 static const Value *getUnderlyingObjectFromInt(const Value *V) {
57 if (const Operator *U = dyn_cast<Operator>(V)) {
58 // If we find a ptrtoint, we can transfer control back to the
59 // regular getUnderlyingObjectFromInt.
60 if (U->getOpcode() == Instruction::PtrToInt)
61 return U->getOperand(0);
62 // If we find an add of a constant or a multiplied value, it's
63 // likely that the other operand will lead us to the base
64 // object. We don't have to worry about the case where the
65 // object address is somehow being computed by the multiply,
66 // because our callers only care when the result is an
67 // identifibale object.
68 if (U->getOpcode() != Instruction::Add ||
69 (!isa<ConstantInt>(U->getOperand(1)) &&
70 Operator::getOpcode(U->getOperand(1)) != Instruction::Mul))
76 assert(V->getType()->isIntegerTy() && "Unexpected operand type!");
80 /// getUnderlyingObject - This is a wrapper around Value::getUnderlyingObject
81 /// and adds support for basic ptrtoint+arithmetic+inttoptr sequences.
82 static const Value *getUnderlyingObject(const Value *V) {
83 // First just call Value::getUnderlyingObject to let it do what it does.
85 V = V->getUnderlyingObject();
86 // If it found an inttoptr, use special code to continue climing.
87 if (Operator::getOpcode(V) != Instruction::IntToPtr)
89 const Value *O = getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
90 // If that succeeded in finding a pointer, continue the search.
91 if (!O->getType()->isPointerTy())
98 /// getUnderlyingObjectForInstr - If this machine instr has memory reference
99 /// information and it can be tracked to a normal reference to a known
100 /// object, return the Value for that object. Otherwise return null.
101 static const Value *getUnderlyingObjectForInstr(const MachineInstr *MI,
102 const MachineFrameInfo *MFI,
105 if (!MI->hasOneMemOperand() ||
106 !(*MI->memoperands_begin())->getValue() ||
107 (*MI->memoperands_begin())->isVolatile())
110 const Value *V = (*MI->memoperands_begin())->getValue();
114 V = getUnderlyingObject(V);
115 if (const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(V)) {
116 // For now, ignore PseudoSourceValues which may alias LLVM IR values
117 // because the code that uses this function has no way to cope with
119 if (PSV->isAliased(MFI))
122 MayAlias = PSV->mayAlias(MFI);
126 if (isIdentifiedObject(V))
132 void ScheduleDAGInstrs::StartBlock(MachineBasicBlock *BB) {
133 if (MachineLoop *ML = MLI.getLoopFor(BB))
134 if (BB == ML->getLoopLatch()) {
135 MachineBasicBlock *Header = ML->getHeader();
136 for (MachineBasicBlock::livein_iterator I = Header->livein_begin(),
137 E = Header->livein_end(); I != E; ++I)
138 LoopLiveInRegs.insert(*I);
139 LoopRegs.VisitLoop(ML);
143 void ScheduleDAGInstrs::BuildSchedGraph(AliasAnalysis *AA) {
144 // We'll be allocating one SUnit for each instruction, plus one for
145 // the region exit node.
146 SUnits.reserve(BB->size());
148 // We build scheduling units by walking a block's instruction list from bottom
151 // Remember where a generic side-effecting instruction is as we procede.
152 SUnit *BarrierChain = 0, *AliasChain = 0;
154 // Memory references to specific known memory locations are tracked
155 // so that they can be given more precise dependencies. We track
156 // separately the known memory locations that may alias and those
157 // that are known not to alias
158 std::map<const Value *, SUnit *> AliasMemDefs, NonAliasMemDefs;
159 std::map<const Value *, std::vector<SUnit *> > AliasMemUses, NonAliasMemUses;
161 // Keep track of dangling debug references to registers.
162 std::pair<MachineInstr*, unsigned>
163 DanglingDebugValue[TargetRegisterInfo::FirstVirtualRegister];
165 // Check to see if the scheduler cares about latencies.
166 bool UnitLatencies = ForceUnitLatencies();
168 // Ask the target if address-backscheduling is desirable, and if so how much.
169 const TargetSubtarget &ST = TM.getSubtarget<TargetSubtarget>();
170 unsigned SpecialAddressLatency = ST.getSpecialAddressLatency();
172 // Remove any stale debug info; sometimes BuildSchedGraph is called again
173 // without emitting the info from the previous call.
175 std::memset(DanglingDebugValue, 0, sizeof(DanglingDebugValue));
177 // Walk the list of instructions, from bottom moving up.
178 for (MachineBasicBlock::iterator MII = InsertPos, MIE = Begin;
180 MachineInstr *MI = prior(MII);
181 // DBG_VALUE does not have SUnit's built, so just remember these for later
183 if (MI->isDebugValue()) {
184 if (MI->getNumOperands()==3 && MI->getOperand(0).isReg() &&
185 MI->getOperand(0).getReg())
186 DanglingDebugValue[MI->getOperand(0).getReg()] =
187 std::make_pair(MI, DbgValueVec.size());
188 DbgValueVec.push_back(MI);
191 const TargetInstrDesc &TID = MI->getDesc();
192 assert(!TID.isTerminator() && !MI->isLabel() &&
193 "Cannot schedule terminators or labels!");
194 // Create the SUnit for this MI.
195 SUnit *SU = NewSUnit(MI);
197 // Assign the Latency field of SU using target-provided information.
203 // Add register-based dependencies (data, anti, and output).
204 for (unsigned j = 0, n = MI->getNumOperands(); j != n; ++j) {
205 const MachineOperand &MO = MI->getOperand(j);
206 if (!MO.isReg()) continue;
207 unsigned Reg = MO.getReg();
208 if (Reg == 0) continue;
210 assert(TRI->isPhysicalRegister(Reg) && "Virtual register encountered!");
212 if (MO.isDef() && DanglingDebugValue[Reg].first!=0) {
213 SU->setDbgInstr(DanglingDebugValue[Reg].first);
214 DbgValueVec[DanglingDebugValue[Reg].second] = 0;
215 DanglingDebugValue[Reg] = std::make_pair((MachineInstr*)0, 0);
218 std::vector<SUnit *> &UseList = Uses[Reg];
219 std::vector<SUnit *> &DefList = Defs[Reg];
220 // Optionally add output and anti dependencies. For anti
221 // dependencies we use a latency of 0 because for a multi-issue
222 // target we want to allow the defining instruction to issue
223 // in the same cycle as the using instruction.
224 // TODO: Using a latency of 1 here for output dependencies assumes
225 // there's no cost for reusing registers.
226 SDep::Kind Kind = MO.isUse() ? SDep::Anti : SDep::Output;
227 unsigned AOLatency = (Kind == SDep::Anti) ? 0 : 1;
228 for (unsigned i = 0, e = DefList.size(); i != e; ++i) {
229 SUnit *DefSU = DefList[i];
231 (Kind != SDep::Output || !MO.isDead() ||
232 !DefSU->getInstr()->registerDefIsDead(Reg)))
233 DefSU->addPred(SDep(SU, Kind, AOLatency, /*Reg=*/Reg));
235 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
236 std::vector<SUnit *> &DefList = Defs[*Alias];
237 for (unsigned i = 0, e = DefList.size(); i != e; ++i) {
238 SUnit *DefSU = DefList[i];
240 (Kind != SDep::Output || !MO.isDead() ||
241 !DefSU->getInstr()->registerDefIsDead(*Alias)))
242 DefSU->addPred(SDep(SU, Kind, AOLatency, /*Reg=*/ *Alias));
247 // Add any data dependencies.
248 unsigned DataLatency = SU->Latency;
249 for (unsigned i = 0, e = UseList.size(); i != e; ++i) {
250 SUnit *UseSU = UseList[i];
253 unsigned LDataLatency = DataLatency;
254 // Optionally add in a special extra latency for nodes that
256 // TODO: Do this for register aliases too.
257 // TODO: Perhaps we should get rid of
258 // SpecialAddressLatency and just move this into
259 // adjustSchedDependency for the targets that care about it.
260 if (SpecialAddressLatency != 0 && !UnitLatencies) {
261 MachineInstr *UseMI = UseSU->getInstr();
262 const TargetInstrDesc &UseTID = UseMI->getDesc();
263 int RegUseIndex = UseMI->findRegisterUseOperandIdx(Reg);
264 assert(RegUseIndex >= 0 && "UseMI doesn's use register!");
265 if ((UseTID.mayLoad() || UseTID.mayStore()) &&
266 (unsigned)RegUseIndex < UseTID.getNumOperands() &&
267 UseTID.OpInfo[RegUseIndex].isLookupPtrRegClass())
268 LDataLatency += SpecialAddressLatency;
270 // Adjust the dependence latency using operand def/use
271 // information (if any), and then allow the target to
272 // perform its own adjustments.
273 const SDep& dep = SDep(SU, SDep::Data, LDataLatency, Reg);
274 if (!UnitLatencies) {
275 ComputeOperandLatency(SU, UseSU, (SDep &)dep);
276 ST.adjustSchedDependency(SU, UseSU, (SDep &)dep);
280 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
281 std::vector<SUnit *> &UseList = Uses[*Alias];
282 for (unsigned i = 0, e = UseList.size(); i != e; ++i) {
283 SUnit *UseSU = UseList[i];
286 const SDep& dep = SDep(SU, SDep::Data, DataLatency, *Alias);
287 if (!UnitLatencies) {
288 ComputeOperandLatency(SU, UseSU, (SDep &)dep);
289 ST.adjustSchedDependency(SU, UseSU, (SDep &)dep);
295 // If a def is going to wrap back around to the top of the loop,
297 if (!UnitLatencies && DefList.empty()) {
298 LoopDependencies::LoopDeps::iterator I = LoopRegs.Deps.find(Reg);
299 if (I != LoopRegs.Deps.end()) {
300 const MachineOperand *UseMO = I->second.first;
301 unsigned Count = I->second.second;
302 const MachineInstr *UseMI = UseMO->getParent();
303 unsigned UseMOIdx = UseMO - &UseMI->getOperand(0);
304 const TargetInstrDesc &UseTID = UseMI->getDesc();
305 // TODO: If we knew the total depth of the region here, we could
306 // handle the case where the whole loop is inside the region but
307 // is large enough that the isScheduleHigh trick isn't needed.
308 if (UseMOIdx < UseTID.getNumOperands()) {
309 // Currently, we only support scheduling regions consisting of
310 // single basic blocks. Check to see if the instruction is in
311 // the same region by checking to see if it has the same parent.
312 if (UseMI->getParent() != MI->getParent()) {
313 unsigned Latency = SU->Latency;
314 if (UseTID.OpInfo[UseMOIdx].isLookupPtrRegClass())
315 Latency += SpecialAddressLatency;
316 // This is a wild guess as to the portion of the latency which
317 // will be overlapped by work done outside the current
318 // scheduling region.
319 Latency -= std::min(Latency, Count);
320 // Add the artifical edge.
321 ExitSU.addPred(SDep(SU, SDep::Order, Latency,
322 /*Reg=*/0, /*isNormalMemory=*/false,
323 /*isMustAlias=*/false,
324 /*isArtificial=*/true));
325 } else if (SpecialAddressLatency > 0 &&
326 UseTID.OpInfo[UseMOIdx].isLookupPtrRegClass()) {
327 // The entire loop body is within the current scheduling region
328 // and the latency of this operation is assumed to be greater
329 // than the latency of the loop.
330 // TODO: Recursively mark data-edge predecessors as
331 // isScheduleHigh too.
332 SU->isScheduleHigh = true;
335 LoopRegs.Deps.erase(I);
342 DefList.push_back(SU);
344 UseList.push_back(SU);
348 // Add chain dependencies.
349 // Chain dependencies used to enforce memory order should have
350 // latency of 0 (except for true dependency of Store followed by
351 // aliased Load... we estimate that with a single cycle of latency
352 // assuming the hardware will bypass)
353 // Note that isStoreToStackSlot and isLoadFromStackSLot are not usable
354 // after stack slots are lowered to actual addresses.
355 // TODO: Use an AliasAnalysis and do real alias-analysis queries, and
356 // produce more precise dependence information.
357 #define STORE_LOAD_LATENCY 1
358 unsigned TrueMemOrderLatency = 0;
359 if (TID.isCall() || TID.hasUnmodeledSideEffects() ||
360 (MI->hasVolatileMemoryRef() &&
361 (!TID.mayLoad() || !MI->isInvariantLoad(AA)))) {
362 // Be conservative with these and add dependencies on all memory
363 // references, even those that are known to not alias.
364 for (std::map<const Value *, SUnit *>::iterator I =
365 NonAliasMemDefs.begin(), E = NonAliasMemDefs.end(); I != E; ++I) {
366 I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
368 for (std::map<const Value *, std::vector<SUnit *> >::iterator I =
369 NonAliasMemUses.begin(), E = NonAliasMemUses.end(); I != E; ++I) {
370 for (unsigned i = 0, e = I->second.size(); i != e; ++i)
371 I->second[i]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency));
373 NonAliasMemDefs.clear();
374 NonAliasMemUses.clear();
375 // Add SU to the barrier chain.
377 BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
382 // Chain all possibly aliasing memory references though SU.
384 AliasChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
386 for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
387 PendingLoads[k]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency));
388 for (std::map<const Value *, SUnit *>::iterator I = AliasMemDefs.begin(),
389 E = AliasMemDefs.end(); I != E; ++I) {
390 I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
392 for (std::map<const Value *, std::vector<SUnit *> >::iterator I =
393 AliasMemUses.begin(), E = AliasMemUses.end(); I != E; ++I) {
394 for (unsigned i = 0, e = I->second.size(); i != e; ++i)
395 I->second[i]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency));
397 PendingLoads.clear();
398 AliasMemDefs.clear();
399 AliasMemUses.clear();
400 } else if (TID.mayStore()) {
401 bool MayAlias = true;
402 TrueMemOrderLatency = STORE_LOAD_LATENCY;
403 if (const Value *V = getUnderlyingObjectForInstr(MI, MFI, MayAlias)) {
404 // A store to a specific PseudoSourceValue. Add precise dependencies.
405 // Record the def in MemDefs, first adding a dep if there is
407 std::map<const Value *, SUnit *>::iterator I =
408 ((MayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V));
409 std::map<const Value *, SUnit *>::iterator IE =
410 ((MayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
412 I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0, /*Reg=*/0,
413 /*isNormalMemory=*/true));
417 AliasMemDefs[V] = SU;
419 NonAliasMemDefs[V] = SU;
421 // Handle the uses in MemUses, if there are any.
422 std::map<const Value *, std::vector<SUnit *> >::iterator J =
423 ((MayAlias) ? AliasMemUses.find(V) : NonAliasMemUses.find(V));
424 std::map<const Value *, std::vector<SUnit *> >::iterator JE =
425 ((MayAlias) ? AliasMemUses.end() : NonAliasMemUses.end());
427 for (unsigned i = 0, e = J->second.size(); i != e; ++i)
428 J->second[i]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency,
429 /*Reg=*/0, /*isNormalMemory=*/true));
433 // Add dependencies from all the PendingLoads, i.e. loads
434 // with no underlying object.
435 for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
436 PendingLoads[k]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency));
437 // Add dependence on alias chain, if needed.
439 AliasChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
441 // Add dependence on barrier chain, if needed.
443 BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
445 // Treat all other stores conservatively.
446 goto new_alias_chain;
448 } else if (TID.mayLoad()) {
449 bool MayAlias = true;
450 TrueMemOrderLatency = 0;
451 if (MI->isInvariantLoad(AA)) {
452 // Invariant load, no chain dependencies needed!
455 getUnderlyingObjectForInstr(MI, MFI, MayAlias)) {
456 // A load from a specific PseudoSourceValue. Add precise dependencies.
457 std::map<const Value *, SUnit *>::iterator I =
458 ((MayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V));
459 std::map<const Value *, SUnit *>::iterator IE =
460 ((MayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
462 I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0, /*Reg=*/0,
463 /*isNormalMemory=*/true));
465 AliasMemUses[V].push_back(SU);
467 NonAliasMemUses[V].push_back(SU);
469 // A load with no underlying object. Depend on all
470 // potentially aliasing stores.
471 for (std::map<const Value *, SUnit *>::iterator I =
472 AliasMemDefs.begin(), E = AliasMemDefs.end(); I != E; ++I)
473 I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
475 PendingLoads.push_back(SU);
479 // Add dependencies on alias and barrier chains, if needed.
480 if (MayAlias && AliasChain)
481 AliasChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
483 BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
488 for (int i = 0, e = TRI->getNumRegs(); i != e; ++i) {
492 PendingLoads.clear();
495 void ScheduleDAGInstrs::FinishBlock() {
499 void ScheduleDAGInstrs::ComputeLatency(SUnit *SU) {
500 const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
502 // Compute the latency for the node.
504 InstrItins.getStageLatency(SU->getInstr()->getDesc().getSchedClass());
506 // Simplistic target-independent heuristic: assume that loads take
508 if (InstrItins.isEmpty())
509 if (SU->getInstr()->getDesc().mayLoad())
513 void ScheduleDAGInstrs::ComputeOperandLatency(SUnit *Def, SUnit *Use,
515 const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
516 if (InstrItins.isEmpty())
519 // For a data dependency with a known register...
520 if ((dep.getKind() != SDep::Data) || (dep.getReg() == 0))
523 const unsigned Reg = dep.getReg();
525 // ... find the definition of the register in the defining
527 MachineInstr *DefMI = Def->getInstr();
528 int DefIdx = DefMI->findRegisterDefOperandIdx(Reg);
530 int DefCycle = InstrItins.getOperandCycle(DefMI->getDesc().getSchedClass(),
533 MachineInstr *UseMI = Use->getInstr();
534 const unsigned UseClass = UseMI->getDesc().getSchedClass();
536 // For all uses of the register, calculate the maxmimum latency
538 for (unsigned i = 0, e = UseMI->getNumOperands(); i != e; ++i) {
539 const MachineOperand &MO = UseMI->getOperand(i);
540 if (!MO.isReg() || !MO.isUse())
542 unsigned MOReg = MO.getReg();
546 int UseCycle = InstrItins.getOperandCycle(UseClass, i);
548 Latency = std::max(Latency, DefCycle - UseCycle + 1);
551 // If we found a latency, then replace the existing dependence latency.
553 dep.setLatency(Latency);
558 void ScheduleDAGInstrs::dumpNode(const SUnit *SU) const {
559 SU->getInstr()->dump();
562 std::string ScheduleDAGInstrs::getGraphNodeLabel(const SUnit *SU) const {
564 raw_string_ostream oss(s);
567 else if (SU == &ExitSU)
570 SU->getInstr()->print(oss);
574 // EmitSchedule - Emit the machine code in scheduled order.
575 MachineBasicBlock *ScheduleDAGInstrs::
576 EmitSchedule(DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) {
577 // For MachineInstr-based scheduling, we're rescheduling the instructions in
578 // the block, so start by removing them from the block.
579 while (Begin != InsertPos) {
580 MachineBasicBlock::iterator I = Begin;
585 // First reinsert any remaining debug_values; these are either constants,
586 // or refer to live-in registers. The beginning of the block is the right
587 // place for the latter. The former might reasonably be placed elsewhere
588 // using some kind of ordering algorithm, but right now it doesn't matter.
589 for (int i = DbgValueVec.size()-1; i>=0; --i)
591 BB->insert(InsertPos, DbgValueVec[i]);
593 // Then re-insert them according to the given schedule.
594 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
595 SUnit *SU = Sequence[i];
597 // Null SUnit* is a noop.
602 BB->insert(InsertPos, SU->getInstr());
603 if (SU->getDbgInstr())
604 BB->insert(InsertPos, SU->getDbgInstr());
607 // Update the Begin iterator, as the first instruction in the block
608 // may have been scheduled later.
609 if (!DbgValueVec.empty()) {
610 for (int i = DbgValueVec.size()-1; i>=0; --i)
611 if (DbgValueVec[i]!=0) {
612 Begin = DbgValueVec[DbgValueVec.size()-1];
615 } else if (!Sequence.empty())
616 Begin = Sequence[0]->getInstr();