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) {}
37 /// Run - perform scheduling.
39 void ScheduleDAGInstrs::Run(MachineBasicBlock *bb,
40 MachineBasicBlock::iterator begin,
41 MachineBasicBlock::iterator end,
45 InsertPosIndex = endcount;
47 ScheduleDAG::Run(bb, end);
50 /// getUnderlyingObjectFromInt - This is the function that does the work of
51 /// looking through basic ptrtoint+arithmetic+inttoptr sequences.
52 static const Value *getUnderlyingObjectFromInt(const Value *V) {
54 if (const Operator *U = dyn_cast<Operator>(V)) {
55 // If we find a ptrtoint, we can transfer control back to the
56 // regular getUnderlyingObjectFromInt.
57 if (U->getOpcode() == Instruction::PtrToInt)
58 return U->getOperand(0);
59 // If we find an add of a constant or a multiplied value, it's
60 // likely that the other operand will lead us to the base
61 // object. We don't have to worry about the case where the
62 // object address is somehow being computed by the multiply,
63 // because our callers only care when the result is an
64 // identifibale object.
65 if (U->getOpcode() != Instruction::Add ||
66 (!isa<ConstantInt>(U->getOperand(1)) &&
67 Operator::getOpcode(U->getOperand(1)) != Instruction::Mul))
73 assert(isa<IntegerType>(V->getType()) && "Unexpected operand type!");
77 /// getUnderlyingObject - This is a wrapper around Value::getUnderlyingObject
78 /// and adds support for basic ptrtoint+arithmetic+inttoptr sequences.
79 static const Value *getUnderlyingObject(const Value *V) {
80 // First just call Value::getUnderlyingObject to let it do what it does.
82 V = V->getUnderlyingObject();
83 // If it found an inttoptr, use special code to continue climing.
84 if (Operator::getOpcode(V) != Instruction::IntToPtr)
86 const Value *O = getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
87 // If that succeeded in finding a pointer, continue the search.
88 if (!isa<PointerType>(O->getType()))
95 /// getUnderlyingObjectForInstr - If this machine instr has memory reference
96 /// information and it can be tracked to a normal reference to a known
97 /// object, return the Value for that object. Otherwise return null.
98 static const Value *getUnderlyingObjectForInstr(const MachineInstr *MI) {
99 if (!MI->hasOneMemOperand() ||
100 !(*MI->memoperands_begin())->getValue() ||
101 (*MI->memoperands_begin())->isVolatile())
104 const Value *V = (*MI->memoperands_begin())->getValue();
108 V = getUnderlyingObject(V);
109 if (const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(V)) {
110 // For now, ignore PseudoSourceValues which may alias LLVM IR values
111 // because the code that uses this function has no way to cope with
113 if (PSV->isAliased())
118 if (isIdentifiedObject(V))
124 void ScheduleDAGInstrs::StartBlock(MachineBasicBlock *BB) {
125 if (MachineLoop *ML = MLI.getLoopFor(BB))
126 if (BB == ML->getLoopLatch()) {
127 MachineBasicBlock *Header = ML->getHeader();
128 for (MachineBasicBlock::livein_iterator I = Header->livein_begin(),
129 E = Header->livein_end(); I != E; ++I)
130 LoopLiveInRegs.insert(*I);
131 LoopRegs.VisitLoop(ML);
135 void ScheduleDAGInstrs::BuildSchedGraph(AliasAnalysis *AA) {
136 // We'll be allocating one SUnit for each instruction, plus one for
137 // the region exit node.
138 SUnits.reserve(BB->size());
140 // We build scheduling units by walking a block's instruction list from bottom
143 // Remember where a generic side-effecting instruction is as we procede. If
144 // ChainMMO is null, this is assumed to have arbitrary side-effects. If
145 // ChainMMO is non-null, then Chain makes only a single memory reference.
147 MachineMemOperand *ChainMMO = 0;
149 // Memory references to specific known memory locations are tracked so that
150 // they can be given more precise dependencies.
151 std::map<const Value *, SUnit *> MemDefs;
152 std::map<const Value *, std::vector<SUnit *> > MemUses;
154 // Check to see if the scheduler cares about latencies.
155 bool UnitLatencies = ForceUnitLatencies();
157 // Ask the target if address-backscheduling is desirable, and if so how much.
158 const TargetSubtarget &ST = TM.getSubtarget<TargetSubtarget>();
159 unsigned SpecialAddressLatency = ST.getSpecialAddressLatency();
161 // Walk the list of instructions, from bottom moving up.
162 for (MachineBasicBlock::iterator MII = InsertPos, MIE = Begin;
164 MachineInstr *MI = prior(MII);
165 const TargetInstrDesc &TID = MI->getDesc();
166 assert(!TID.isTerminator() && !MI->isLabel() &&
167 "Cannot schedule terminators or labels!");
168 // Create the SUnit for this MI.
169 SUnit *SU = NewSUnit(MI);
171 // Assign the Latency field of SU using target-provided information.
177 // Add register-based dependencies (data, anti, and output).
178 for (unsigned j = 0, n = MI->getNumOperands(); j != n; ++j) {
179 const MachineOperand &MO = MI->getOperand(j);
180 if (!MO.isReg()) continue;
181 unsigned Reg = MO.getReg();
182 if (Reg == 0) continue;
184 assert(TRI->isPhysicalRegister(Reg) && "Virtual register encountered!");
185 std::vector<SUnit *> &UseList = Uses[Reg];
186 std::vector<SUnit *> &DefList = Defs[Reg];
187 // Optionally add output and anti dependencies. For anti
188 // dependencies we use a latency of 0 because for a multi-issue
189 // target we want to allow the defining instruction to issue
190 // in the same cycle as the using instruction.
191 // TODO: Using a latency of 1 here for output dependencies assumes
192 // there's no cost for reusing registers.
193 SDep::Kind Kind = MO.isUse() ? SDep::Anti : SDep::Output;
194 unsigned AOLatency = (Kind == SDep::Anti) ? 0 : 1;
195 for (unsigned i = 0, e = DefList.size(); i != e; ++i) {
196 SUnit *DefSU = DefList[i];
198 (Kind != SDep::Output || !MO.isDead() ||
199 !DefSU->getInstr()->registerDefIsDead(Reg)))
200 DefSU->addPred(SDep(SU, Kind, AOLatency, /*Reg=*/Reg));
202 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
203 std::vector<SUnit *> &DefList = Defs[*Alias];
204 for (unsigned i = 0, e = DefList.size(); i != e; ++i) {
205 SUnit *DefSU = DefList[i];
207 (Kind != SDep::Output || !MO.isDead() ||
208 !DefSU->getInstr()->registerDefIsDead(Reg)))
209 DefSU->addPred(SDep(SU, Kind, AOLatency, /*Reg=*/ *Alias));
214 // Add any data dependencies.
215 unsigned DataLatency = SU->Latency;
216 for (unsigned i = 0, e = UseList.size(); i != e; ++i) {
217 SUnit *UseSU = UseList[i];
219 unsigned LDataLatency = DataLatency;
220 // Optionally add in a special extra latency for nodes that
222 // TODO: Do this for register aliases too.
223 // TODO: Perhaps we should get rid of
224 // SpecialAddressLatency and just move this into
225 // adjustSchedDependency for the targets that care about
227 if (SpecialAddressLatency != 0 && !UnitLatencies) {
228 MachineInstr *UseMI = UseSU->getInstr();
229 const TargetInstrDesc &UseTID = UseMI->getDesc();
230 int RegUseIndex = UseMI->findRegisterUseOperandIdx(Reg);
231 assert(RegUseIndex >= 0 && "UseMI doesn's use register!");
232 if ((UseTID.mayLoad() || UseTID.mayStore()) &&
233 (unsigned)RegUseIndex < UseTID.getNumOperands() &&
234 UseTID.OpInfo[RegUseIndex].isLookupPtrRegClass())
235 LDataLatency += SpecialAddressLatency;
237 // Adjust the dependence latency using operand def/use
238 // information (if any), and then allow the target to
239 // perform its own adjustments.
240 const SDep& dep = SDep(SU, SDep::Data, LDataLatency, Reg);
241 if (!UnitLatencies) {
242 ComputeOperandLatency(SU, UseSU, (SDep &)dep);
243 ST.adjustSchedDependency(SU, UseSU, (SDep &)dep);
248 for (const unsigned *Alias = TRI->getAliasSet(Reg); *Alias; ++Alias) {
249 std::vector<SUnit *> &UseList = Uses[*Alias];
250 for (unsigned i = 0, e = UseList.size(); i != e; ++i) {
251 SUnit *UseSU = UseList[i];
253 const SDep& dep = SDep(SU, SDep::Data, DataLatency, *Alias);
254 if (!UnitLatencies) {
255 ComputeOperandLatency(SU, UseSU, (SDep &)dep);
256 ST.adjustSchedDependency(SU, UseSU, (SDep &)dep);
263 // If a def is going to wrap back around to the top of the loop,
265 if (!UnitLatencies && DefList.empty()) {
266 LoopDependencies::LoopDeps::iterator I = LoopRegs.Deps.find(Reg);
267 if (I != LoopRegs.Deps.end()) {
268 const MachineOperand *UseMO = I->second.first;
269 unsigned Count = I->second.second;
270 const MachineInstr *UseMI = UseMO->getParent();
271 unsigned UseMOIdx = UseMO - &UseMI->getOperand(0);
272 const TargetInstrDesc &UseTID = UseMI->getDesc();
273 // TODO: If we knew the total depth of the region here, we could
274 // handle the case where the whole loop is inside the region but
275 // is large enough that the isScheduleHigh trick isn't needed.
276 if (UseMOIdx < UseTID.getNumOperands()) {
277 // Currently, we only support scheduling regions consisting of
278 // single basic blocks. Check to see if the instruction is in
279 // the same region by checking to see if it has the same parent.
280 if (UseMI->getParent() != MI->getParent()) {
281 unsigned Latency = SU->Latency;
282 if (UseTID.OpInfo[UseMOIdx].isLookupPtrRegClass())
283 Latency += SpecialAddressLatency;
284 // This is a wild guess as to the portion of the latency which
285 // will be overlapped by work done outside the current
286 // scheduling region.
287 Latency -= std::min(Latency, Count);
288 // Add the artifical edge.
289 ExitSU.addPred(SDep(SU, SDep::Order, Latency,
290 /*Reg=*/0, /*isNormalMemory=*/false,
291 /*isMustAlias=*/false,
292 /*isArtificial=*/true));
293 } else if (SpecialAddressLatency > 0 &&
294 UseTID.OpInfo[UseMOIdx].isLookupPtrRegClass()) {
295 // The entire loop body is within the current scheduling region
296 // and the latency of this operation is assumed to be greater
297 // than the latency of the loop.
298 // TODO: Recursively mark data-edge predecessors as
299 // isScheduleHigh too.
300 SU->isScheduleHigh = true;
303 LoopRegs.Deps.erase(I);
310 DefList.push_back(SU);
312 UseList.push_back(SU);
316 // Add chain dependencies.
317 // Note that isStoreToStackSlot and isLoadFromStackSLot are not usable
318 // after stack slots are lowered to actual addresses.
319 // TODO: Use an AliasAnalysis and do real alias-analysis queries, and
320 // produce more precise dependence information.
321 if (TID.isCall() || TID.hasUnmodeledSideEffects()) {
323 // This is the conservative case. Add dependencies on all memory
326 Chain->addPred(SDep(SU, SDep::Order, SU->Latency));
328 for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
329 PendingLoads[k]->addPred(SDep(SU, SDep::Order, SU->Latency));
330 PendingLoads.clear();
331 for (std::map<const Value *, SUnit *>::iterator I = MemDefs.begin(),
332 E = MemDefs.end(); I != E; ++I) {
333 I->second->addPred(SDep(SU, SDep::Order, SU->Latency));
336 for (std::map<const Value *, std::vector<SUnit *> >::iterator I =
337 MemUses.begin(), E = MemUses.end(); I != E; ++I) {
338 for (unsigned i = 0, e = I->second.size(); i != e; ++i)
339 I->second[i]->addPred(SDep(SU, SDep::Order, SU->Latency));
342 // See if it is known to just have a single memory reference.
343 MachineInstr *ChainMI = Chain->getInstr();
344 const TargetInstrDesc &ChainTID = ChainMI->getDesc();
345 if (!ChainTID.isCall() &&
346 !ChainTID.hasUnmodeledSideEffects() &&
347 ChainMI->hasOneMemOperand() &&
348 !(*ChainMI->memoperands_begin())->isVolatile() &&
349 (*ChainMI->memoperands_begin())->getValue())
350 // We know that the Chain accesses one specific memory location.
351 ChainMMO = *ChainMI->memoperands_begin();
353 // Unknown memory accesses. Assume the worst.
355 } else if (TID.mayStore()) {
356 if (const Value *V = getUnderlyingObjectForInstr(MI)) {
357 // A store to a specific PseudoSourceValue. Add precise dependencies.
358 // Handle the def in MemDefs, if there is one.
359 std::map<const Value *, SUnit *>::iterator I = MemDefs.find(V);
360 if (I != MemDefs.end()) {
361 I->second->addPred(SDep(SU, SDep::Order, SU->Latency, /*Reg=*/0,
362 /*isNormalMemory=*/true));
367 // Handle the uses in MemUses, if there are any.
368 std::map<const Value *, std::vector<SUnit *> >::iterator J =
370 if (J != MemUses.end()) {
371 for (unsigned i = 0, e = J->second.size(); i != e; ++i)
372 J->second[i]->addPred(SDep(SU, SDep::Order, SU->Latency, /*Reg=*/0,
373 /*isNormalMemory=*/true));
376 // Add dependencies from all the PendingLoads, since without
377 // memoperands we must assume they alias anything.
378 for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
379 PendingLoads[k]->addPred(SDep(SU, SDep::Order, SU->Latency));
380 // Add a general dependence too, if needed.
382 Chain->addPred(SDep(SU, SDep::Order, SU->Latency));
384 // Treat all other stores conservatively.
386 } else if (TID.mayLoad()) {
387 if (MI->isInvariantLoad(AA)) {
388 // Invariant load, no chain dependencies needed!
389 } else if (const Value *V = getUnderlyingObjectForInstr(MI)) {
390 // A load from a specific PseudoSourceValue. Add precise dependencies.
391 std::map<const Value *, SUnit *>::iterator I = MemDefs.find(V);
392 if (I != MemDefs.end())
393 I->second->addPred(SDep(SU, SDep::Order, SU->Latency, /*Reg=*/0,
394 /*isNormalMemory=*/true));
395 MemUses[V].push_back(SU);
397 // Add a general dependence too, if needed.
398 if (Chain && (!ChainMMO ||
399 (ChainMMO->isStore() || ChainMMO->isVolatile())))
400 Chain->addPred(SDep(SU, SDep::Order, SU->Latency));
401 } else if (MI->hasVolatileMemoryRef()) {
402 // Treat volatile loads conservatively. Note that this includes
403 // cases where memoperand information is unavailable.
406 // A normal load. Depend on the general chain, as well as on
407 // all stores. In the absense of MachineMemOperand information,
408 // we can't even assume that the load doesn't alias well-behaved
411 Chain->addPred(SDep(SU, SDep::Order, SU->Latency));
412 for (std::map<const Value *, SUnit *>::iterator I = MemDefs.begin(),
413 E = MemDefs.end(); I != E; ++I)
414 I->second->addPred(SDep(SU, SDep::Order, SU->Latency));
415 PendingLoads.push_back(SU);
420 for (int i = 0, e = TRI->getNumRegs(); i != e; ++i) {
424 PendingLoads.clear();
427 void ScheduleDAGInstrs::FinishBlock() {
431 void ScheduleDAGInstrs::ComputeLatency(SUnit *SU) {
432 const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
434 // Compute the latency for the node.
436 InstrItins.getStageLatency(SU->getInstr()->getDesc().getSchedClass());
438 // Simplistic target-independent heuristic: assume that loads take
440 if (InstrItins.isEmpty())
441 if (SU->getInstr()->getDesc().mayLoad())
445 void ScheduleDAGInstrs::ComputeOperandLatency(SUnit *Def, SUnit *Use,
447 const InstrItineraryData &InstrItins = TM.getInstrItineraryData();
448 if (InstrItins.isEmpty())
451 // For a data dependency with a known register...
452 if ((dep.getKind() != SDep::Data) || (dep.getReg() == 0))
455 const unsigned Reg = dep.getReg();
457 // ... find the definition of the register in the defining
459 MachineInstr *DefMI = Def->getInstr();
460 int DefIdx = DefMI->findRegisterDefOperandIdx(Reg);
462 int DefCycle = InstrItins.getOperandCycle(DefMI->getDesc().getSchedClass(), DefIdx);
464 MachineInstr *UseMI = Use->getInstr();
465 const unsigned UseClass = UseMI->getDesc().getSchedClass();
467 // For all uses of the register, calculate the maxmimum latency
469 for (unsigned i = 0, e = UseMI->getNumOperands(); i != e; ++i) {
470 const MachineOperand &MO = UseMI->getOperand(i);
471 if (!MO.isReg() || !MO.isUse())
473 unsigned MOReg = MO.getReg();
477 int UseCycle = InstrItins.getOperandCycle(UseClass, i);
479 Latency = std::max(Latency, DefCycle - UseCycle + 1);
482 // If we found a latency, then replace the existing dependence latency.
484 dep.setLatency(Latency);
489 void ScheduleDAGInstrs::dumpNode(const SUnit *SU) const {
490 SU->getInstr()->dump();
493 std::string ScheduleDAGInstrs::getGraphNodeLabel(const SUnit *SU) const {
495 raw_string_ostream oss(s);
498 else if (SU == &ExitSU)
501 SU->getInstr()->print(oss);
505 // EmitSchedule - Emit the machine code in scheduled order.
506 MachineBasicBlock *ScheduleDAGInstrs::
507 EmitSchedule(DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) {
508 // For MachineInstr-based scheduling, we're rescheduling the instructions in
509 // the block, so start by removing them from the block.
510 while (Begin != InsertPos) {
511 MachineBasicBlock::iterator I = Begin;
516 // Then re-insert them according to the given schedule.
517 for (unsigned i = 0, e = Sequence.size(); i != e; i++) {
518 SUnit *SU = Sequence[i];
520 // Null SUnit* is a noop.
525 BB->insert(InsertPos, SU->getInstr());
528 // Update the Begin iterator, as the first instruction in the block
529 // may have been scheduled later.
530 if (!Sequence.empty())
531 Begin = Sequence[0]->getInstr();