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 "llvm/Operator.h"
17 #include "llvm/Analysis/AliasAnalysis.h"
18 #include "llvm/Analysis/ValueTracking.h"
19 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
20 #include "llvm/CodeGen/MachineFunctionPass.h"
21 #include "llvm/CodeGen/MachineMemOperand.h"
22 #include "llvm/CodeGen/MachineRegisterInfo.h"
23 #include "llvm/CodeGen/PseudoSourceValue.h"
24 #include "llvm/CodeGen/RegisterPressure.h"
25 #include "llvm/CodeGen/ScheduleDAGInstrs.h"
26 #include "llvm/MC/MCInstrItineraries.h"
27 #include "llvm/Target/TargetMachine.h"
28 #include "llvm/Target/TargetInstrInfo.h"
29 #include "llvm/Target/TargetRegisterInfo.h"
30 #include "llvm/Target/TargetSubtargetInfo.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/raw_ostream.h"
34 #include "llvm/ADT/SmallSet.h"
35 #include "llvm/ADT/SmallPtrSet.h"
38 static cl::opt<bool> EnableAASchedMI("enable-aa-sched-mi", cl::Hidden,
39 cl::ZeroOrMore, cl::init(false),
40 cl::desc("Enable use of AA during MI GAD construction"));
42 ScheduleDAGInstrs::ScheduleDAGInstrs(MachineFunction &mf,
43 const MachineLoopInfo &mli,
44 const MachineDominatorTree &mdt,
47 : ScheduleDAG(mf), MLI(mli), MDT(mdt), MFI(mf.getFrameInfo()),
48 InstrItins(mf.getTarget().getInstrItineraryData()), LIS(lis),
49 IsPostRA(IsPostRAFlag), UnitLatencies(false), CanHandleTerminators(false),
50 LoopRegs(MDT), FirstDbgValue(0) {
51 assert((IsPostRA || LIS) && "PreRA scheduling requires LiveIntervals");
53 assert(!(IsPostRA && MRI.getNumVirtRegs()) &&
54 "Virtual registers must be removed prior to PostRA scheduling");
56 const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
57 SchedModel.init(*ST.getSchedModel(), &ST, TII);
60 /// getUnderlyingObjectFromInt - This is the function that does the work of
61 /// looking through basic ptrtoint+arithmetic+inttoptr sequences.
62 static const Value *getUnderlyingObjectFromInt(const Value *V) {
64 if (const Operator *U = dyn_cast<Operator>(V)) {
65 // If we find a ptrtoint, we can transfer control back to the
66 // regular getUnderlyingObjectFromInt.
67 if (U->getOpcode() == Instruction::PtrToInt)
68 return U->getOperand(0);
69 // If we find an add of a constant or a multiplied value, it's
70 // likely that the other operand will lead us to the base
71 // object. We don't have to worry about the case where the
72 // object address is somehow being computed by the multiply,
73 // because our callers only care when the result is an
74 // identifibale object.
75 if (U->getOpcode() != Instruction::Add ||
76 (!isa<ConstantInt>(U->getOperand(1)) &&
77 Operator::getOpcode(U->getOperand(1)) != Instruction::Mul))
83 assert(V->getType()->isIntegerTy() && "Unexpected operand type!");
87 /// getUnderlyingObject - This is a wrapper around GetUnderlyingObject
88 /// and adds support for basic ptrtoint+arithmetic+inttoptr sequences.
89 static const Value *getUnderlyingObject(const Value *V) {
90 // First just call Value::getUnderlyingObject to let it do what it does.
92 V = GetUnderlyingObject(V);
93 // If it found an inttoptr, use special code to continue climing.
94 if (Operator::getOpcode(V) != Instruction::IntToPtr)
96 const Value *O = getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
97 // If that succeeded in finding a pointer, continue the search.
98 if (!O->getType()->isPointerTy())
105 /// getUnderlyingObjectForInstr - If this machine instr has memory reference
106 /// information and it can be tracked to a normal reference to a known
107 /// object, return the Value for that object. Otherwise return null.
108 static const Value *getUnderlyingObjectForInstr(const MachineInstr *MI,
109 const MachineFrameInfo *MFI,
112 if (!MI->hasOneMemOperand() ||
113 !(*MI->memoperands_begin())->getValue() ||
114 (*MI->memoperands_begin())->isVolatile())
117 const Value *V = (*MI->memoperands_begin())->getValue();
121 V = getUnderlyingObject(V);
122 if (const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(V)) {
123 // For now, ignore PseudoSourceValues which may alias LLVM IR values
124 // because the code that uses this function has no way to cope with
126 if (PSV->isAliased(MFI))
129 MayAlias = PSV->mayAlias(MFI);
133 if (isIdentifiedObject(V))
139 void ScheduleDAGInstrs::startBlock(MachineBasicBlock *bb) {
141 LoopRegs.Deps.clear();
142 if (MachineLoop *ML = MLI.getLoopFor(BB))
143 if (BB == ML->getLoopLatch())
144 LoopRegs.VisitLoop(ML);
147 void ScheduleDAGInstrs::finishBlock() {
148 // Subclasses should no longer refer to the old block.
152 /// Initialize the map with the number of registers.
153 void Reg2SUnitsMap::setRegLimit(unsigned Limit) {
154 PhysRegSet.setUniverse(Limit);
155 SUnits.resize(Limit);
158 /// Clear the map without deallocating storage.
159 void Reg2SUnitsMap::clear() {
160 for (const_iterator I = reg_begin(), E = reg_end(); I != E; ++I) {
166 /// Initialize the DAG and common scheduler state for the current scheduling
167 /// region. This does not actually create the DAG, only clears it. The
168 /// scheduling driver may call BuildSchedGraph multiple times per scheduling
170 void ScheduleDAGInstrs::enterRegion(MachineBasicBlock *bb,
171 MachineBasicBlock::iterator begin,
172 MachineBasicBlock::iterator end,
174 assert(bb == BB && "startBlock should set BB");
180 // Check to see if the scheduler cares about latencies.
181 UnitLatencies = forceUnitLatencies();
183 ScheduleDAG::clearDAG();
186 /// Close the current scheduling region. Don't clear any state in case the
187 /// driver wants to refer to the previous scheduling region.
188 void ScheduleDAGInstrs::exitRegion() {
192 /// addSchedBarrierDeps - Add dependencies from instructions in the current
193 /// list of instructions being scheduled to scheduling barrier by adding
194 /// the exit SU to the register defs and use list. This is because we want to
195 /// make sure instructions which define registers that are either used by
196 /// the terminator or are live-out are properly scheduled. This is
197 /// especially important when the definition latency of the return value(s)
198 /// are too high to be hidden by the branch or when the liveout registers
199 /// used by instructions in the fallthrough block.
200 void ScheduleDAGInstrs::addSchedBarrierDeps() {
201 MachineInstr *ExitMI = RegionEnd != BB->end() ? &*RegionEnd : 0;
202 ExitSU.setInstr(ExitMI);
203 bool AllDepKnown = ExitMI &&
204 (ExitMI->isCall() || ExitMI->isBarrier());
205 if (ExitMI && AllDepKnown) {
206 // If it's a call or a barrier, add dependencies on the defs and uses of
208 for (unsigned i = 0, e = ExitMI->getNumOperands(); i != e; ++i) {
209 const MachineOperand &MO = ExitMI->getOperand(i);
210 if (!MO.isReg() || MO.isDef()) continue;
211 unsigned Reg = MO.getReg();
212 if (Reg == 0) continue;
214 if (TRI->isPhysicalRegister(Reg))
215 Uses[Reg].push_back(PhysRegSUOper(&ExitSU, -1));
217 assert(!IsPostRA && "Virtual register encountered after regalloc.");
218 addVRegUseDeps(&ExitSU, i);
222 // For others, e.g. fallthrough, conditional branch, assume the exit
223 // uses all the registers that are livein to the successor blocks.
224 assert(Uses.empty() && "Uses in set before adding deps?");
225 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
226 SE = BB->succ_end(); SI != SE; ++SI)
227 for (MachineBasicBlock::livein_iterator I = (*SI)->livein_begin(),
228 E = (*SI)->livein_end(); I != E; ++I) {
230 if (!Uses.contains(Reg))
231 Uses[Reg].push_back(PhysRegSUOper(&ExitSU, -1));
236 /// MO is an operand of SU's instruction that defines a physical register. Add
237 /// data dependencies from SU to any uses of the physical register.
238 void ScheduleDAGInstrs::addPhysRegDataDeps(SUnit *SU, unsigned OperIdx) {
239 const MachineOperand &MO = SU->getInstr()->getOperand(OperIdx);
240 assert(MO.isDef() && "expect physreg def");
242 // Ask the target if address-backscheduling is desirable, and if so how much.
243 const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
244 unsigned SpecialAddressLatency = ST.getSpecialAddressLatency();
245 unsigned DataLatency = SU->Latency;
247 for (MCRegAliasIterator Alias(MO.getReg(), TRI, true);
248 Alias.isValid(); ++Alias) {
249 if (!Uses.contains(*Alias))
251 std::vector<PhysRegSUOper> &UseList = Uses[*Alias];
252 for (unsigned i = 0, e = UseList.size(); i != e; ++i) {
253 SUnit *UseSU = UseList[i].SU;
256 MachineInstr *UseMI = UseSU->getInstr();
257 int UseOp = UseList[i].OpIdx;
258 unsigned LDataLatency = DataLatency;
259 // Optionally add in a special extra latency for nodes that
261 // TODO: Perhaps we should get rid of
262 // SpecialAddressLatency and just move this into
263 // adjustSchedDependency for the targets that care about it.
264 if (SpecialAddressLatency != 0 && !UnitLatencies &&
266 const MCInstrDesc &UseMCID = UseMI->getDesc();
267 int RegUseIndex = UseMI->findRegisterUseOperandIdx(*Alias);
268 assert(RegUseIndex >= 0 && "UseMI doesn't use register!");
269 if (RegUseIndex >= 0 &&
270 (UseMI->mayLoad() || UseMI->mayStore()) &&
271 (unsigned)RegUseIndex < UseMCID.getNumOperands() &&
272 UseMCID.OpInfo[RegUseIndex].isLookupPtrRegClass())
273 LDataLatency += SpecialAddressLatency;
275 // Adjust the dependence latency using operand def/use
276 // information (if any), and then allow the target to
277 // perform its own adjustments.
278 SDep dep(SU, SDep::Data, LDataLatency, *Alias);
279 if (!UnitLatencies) {
280 MachineInstr *RegUse = UseOp < 0 ? 0 : UseMI;
282 SchedModel.computeOperandLatency(SU->getInstr(), OperIdx,
283 RegUse, UseOp, /*FindMin=*/false));
285 SchedModel.computeOperandLatency(SU->getInstr(), OperIdx,
286 RegUse, UseOp, /*FindMin=*/true));
288 ST.adjustSchedDependency(SU, UseSU, dep);
295 /// addPhysRegDeps - Add register dependencies (data, anti, and output) from
296 /// this SUnit to following instructions in the same scheduling region that
297 /// depend the physical register referenced at OperIdx.
298 void ScheduleDAGInstrs::addPhysRegDeps(SUnit *SU, unsigned OperIdx) {
299 const MachineInstr *MI = SU->getInstr();
300 const MachineOperand &MO = MI->getOperand(OperIdx);
302 // Optionally add output and anti dependencies. For anti
303 // dependencies we use a latency of 0 because for a multi-issue
304 // target we want to allow the defining instruction to issue
305 // in the same cycle as the using instruction.
306 // TODO: Using a latency of 1 here for output dependencies assumes
307 // there's no cost for reusing registers.
308 SDep::Kind Kind = MO.isUse() ? SDep::Anti : SDep::Output;
309 for (MCRegAliasIterator Alias(MO.getReg(), TRI, true);
310 Alias.isValid(); ++Alias) {
311 if (!Defs.contains(*Alias))
313 std::vector<PhysRegSUOper> &DefList = Defs[*Alias];
314 for (unsigned i = 0, e = DefList.size(); i != e; ++i) {
315 SUnit *DefSU = DefList[i].SU;
316 if (DefSU == &ExitSU)
319 (Kind != SDep::Output || !MO.isDead() ||
320 !DefSU->getInstr()->registerDefIsDead(*Alias))) {
321 if (Kind == SDep::Anti)
322 DefSU->addPred(SDep(SU, Kind, 0, /*Reg=*/*Alias));
324 unsigned AOLat = TII->getOutputLatency(InstrItins, MI, OperIdx,
326 DefSU->addPred(SDep(SU, Kind, AOLat, /*Reg=*/*Alias));
333 // Either insert a new Reg2SUnits entry with an empty SUnits list, or
334 // retrieve the existing SUnits list for this register's uses.
335 // Push this SUnit on the use list.
336 Uses[MO.getReg()].push_back(PhysRegSUOper(SU, OperIdx));
339 addPhysRegDataDeps(SU, OperIdx);
341 // Either insert a new Reg2SUnits entry with an empty SUnits list, or
342 // retrieve the existing SUnits list for this register's defs.
343 std::vector<PhysRegSUOper> &DefList = Defs[MO.getReg()];
345 // If a def is going to wrap back around to the top of the loop,
347 if (!UnitLatencies && DefList.empty()) {
348 LoopDependencies::LoopDeps::iterator I = LoopRegs.Deps.find(MO.getReg());
349 if (I != LoopRegs.Deps.end()) {
350 const MachineOperand *UseMO = I->second.first;
351 unsigned Count = I->second.second;
352 const MachineInstr *UseMI = UseMO->getParent();
353 unsigned UseMOIdx = UseMO - &UseMI->getOperand(0);
354 const MCInstrDesc &UseMCID = UseMI->getDesc();
355 const TargetSubtargetInfo &ST =
356 TM.getSubtarget<TargetSubtargetInfo>();
357 unsigned SpecialAddressLatency = ST.getSpecialAddressLatency();
358 // TODO: If we knew the total depth of the region here, we could
359 // handle the case where the whole loop is inside the region but
360 // is large enough that the isScheduleHigh trick isn't needed.
361 if (UseMOIdx < UseMCID.getNumOperands()) {
362 // Currently, we only support scheduling regions consisting of
363 // single basic blocks. Check to see if the instruction is in
364 // the same region by checking to see if it has the same parent.
365 if (UseMI->getParent() != MI->getParent()) {
366 unsigned Latency = SU->Latency;
367 if (UseMCID.OpInfo[UseMOIdx].isLookupPtrRegClass())
368 Latency += SpecialAddressLatency;
369 // This is a wild guess as to the portion of the latency which
370 // will be overlapped by work done outside the current
371 // scheduling region.
372 Latency -= std::min(Latency, Count);
373 // Add the artificial edge.
374 ExitSU.addPred(SDep(SU, SDep::Order, Latency,
375 /*Reg=*/0, /*isNormalMemory=*/false,
376 /*isMustAlias=*/false,
377 /*isArtificial=*/true));
378 } else if (SpecialAddressLatency > 0 &&
379 UseMCID.OpInfo[UseMOIdx].isLookupPtrRegClass()) {
380 // The entire loop body is within the current scheduling region
381 // and the latency of this operation is assumed to be greater
382 // than the latency of the loop.
383 // TODO: Recursively mark data-edge predecessors as
384 // isScheduleHigh too.
385 SU->isScheduleHigh = true;
388 LoopRegs.Deps.erase(I);
392 // clear this register's use list
393 if (Uses.contains(MO.getReg()))
394 Uses[MO.getReg()].clear();
399 // Calls will not be reordered because of chain dependencies (see
400 // below). Since call operands are dead, calls may continue to be added
401 // to the DefList making dependence checking quadratic in the size of
402 // the block. Instead, we leave only one call at the back of the
405 while (!DefList.empty() && DefList.back().SU->isCall)
408 // Defs are pushed in the order they are visited and never reordered.
409 DefList.push_back(PhysRegSUOper(SU, OperIdx));
413 /// addVRegDefDeps - Add register output and data dependencies from this SUnit
414 /// to instructions that occur later in the same scheduling region if they read
415 /// from or write to the virtual register defined at OperIdx.
417 /// TODO: Hoist loop induction variable increments. This has to be
418 /// reevaluated. Generally, IV scheduling should be done before coalescing.
419 void ScheduleDAGInstrs::addVRegDefDeps(SUnit *SU, unsigned OperIdx) {
420 const MachineInstr *MI = SU->getInstr();
421 unsigned Reg = MI->getOperand(OperIdx).getReg();
423 // Singly defined vregs do not have output/anti dependencies.
424 // The current operand is a def, so we have at least one.
425 // Check here if there are any others...
426 if (MRI.hasOneDef(Reg))
429 // Add output dependence to the next nearest def of this vreg.
431 // Unless this definition is dead, the output dependence should be
432 // transitively redundant with antidependencies from this definition's
433 // uses. We're conservative for now until we have a way to guarantee the uses
434 // are not eliminated sometime during scheduling. The output dependence edge
435 // is also useful if output latency exceeds def-use latency.
436 VReg2SUnitMap::iterator DefI = VRegDefs.find(Reg);
437 if (DefI == VRegDefs.end())
438 VRegDefs.insert(VReg2SUnit(Reg, SU));
440 SUnit *DefSU = DefI->SU;
441 if (DefSU != SU && DefSU != &ExitSU) {
442 unsigned OutLatency = TII->getOutputLatency(InstrItins, MI, OperIdx,
444 DefSU->addPred(SDep(SU, SDep::Output, OutLatency, Reg));
450 /// addVRegUseDeps - Add a register data dependency if the instruction that
451 /// defines the virtual register used at OperIdx is mapped to an SUnit. Add a
452 /// register antidependency from this SUnit to instructions that occur later in
453 /// the same scheduling region if they write the virtual register.
455 /// TODO: Handle ExitSU "uses" properly.
456 void ScheduleDAGInstrs::addVRegUseDeps(SUnit *SU, unsigned OperIdx) {
457 MachineInstr *MI = SU->getInstr();
458 unsigned Reg = MI->getOperand(OperIdx).getReg();
460 // Lookup this operand's reaching definition.
461 assert(LIS && "vreg dependencies requires LiveIntervals");
462 LiveRangeQuery LRQ(LIS->getInterval(Reg), LIS->getInstructionIndex(MI));
463 VNInfo *VNI = LRQ.valueIn();
465 // VNI will be valid because MachineOperand::readsReg() is checked by caller.
466 assert(VNI && "No value to read by operand");
467 MachineInstr *Def = LIS->getInstructionFromIndex(VNI->def);
468 // Phis and other noninstructions (after coalescing) have a NULL Def.
470 SUnit *DefSU = getSUnit(Def);
472 // The reaching Def lives within this scheduling region.
473 // Create a data dependence.
475 // TODO: Handle "special" address latencies cleanly.
476 SDep dep(DefSU, SDep::Data, DefSU->Latency, Reg);
477 if (!UnitLatencies) {
478 // Adjust the dependence latency using operand def/use information, then
479 // allow the target to perform its own adjustments.
480 int DefOp = Def->findRegisterDefOperandIdx(Reg);
482 SchedModel.computeOperandLatency(Def, DefOp, MI, OperIdx, false));
484 SchedModel.computeOperandLatency(Def, DefOp, MI, OperIdx, true));
486 const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
487 ST.adjustSchedDependency(DefSU, SU, const_cast<SDep &>(dep));
493 // Add antidependence to the following def of the vreg it uses.
494 VReg2SUnitMap::iterator DefI = VRegDefs.find(Reg);
495 if (DefI != VRegDefs.end() && DefI->SU != SU)
496 DefI->SU->addPred(SDep(SU, SDep::Anti, 0, Reg));
499 /// Return true if MI is an instruction we are unable to reason about
500 /// (like a call or something with unmodeled side effects).
501 static inline bool isGlobalMemoryObject(AliasAnalysis *AA, MachineInstr *MI) {
502 if (MI->isCall() || MI->hasUnmodeledSideEffects() ||
503 (MI->hasOrderedMemoryRef() &&
504 (!MI->mayLoad() || !MI->isInvariantLoad(AA))))
509 // This MI might have either incomplete info, or known to be unsafe
510 // to deal with (i.e. volatile object).
511 static inline bool isUnsafeMemoryObject(MachineInstr *MI,
512 const MachineFrameInfo *MFI) {
513 if (!MI || MI->memoperands_empty())
515 // We purposefully do no check for hasOneMemOperand() here
516 // in hope to trigger an assert downstream in order to
517 // finish implementation.
518 if ((*MI->memoperands_begin())->isVolatile() ||
519 MI->hasUnmodeledSideEffects())
522 const Value *V = (*MI->memoperands_begin())->getValue();
526 V = getUnderlyingObject(V);
527 if (const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(V)) {
528 // Similarly to getUnderlyingObjectForInstr:
529 // For now, ignore PseudoSourceValues which may alias LLVM IR values
530 // because the code that uses this function has no way to cope with
532 if (PSV->isAliased(MFI))
535 // Does this pointer refer to a distinct and identifiable object?
536 if (!isIdentifiedObject(V))
542 /// This returns true if the two MIs need a chain edge betwee them.
543 /// If these are not even memory operations, we still may need
544 /// chain deps between them. The question really is - could
545 /// these two MIs be reordered during scheduling from memory dependency
547 static bool MIsNeedChainEdge(AliasAnalysis *AA, const MachineFrameInfo *MFI,
550 // Cover a trivial case - no edge is need to itself.
554 if (isUnsafeMemoryObject(MIa, MFI) || isUnsafeMemoryObject(MIb, MFI))
557 // If we are dealing with two "normal" loads, we do not need an edge
558 // between them - they could be reordered.
559 if (!MIa->mayStore() && !MIb->mayStore())
562 // To this point analysis is generic. From here on we do need AA.
566 MachineMemOperand *MMOa = *MIa->memoperands_begin();
567 MachineMemOperand *MMOb = *MIb->memoperands_begin();
569 // FIXME: Need to handle multiple memory operands to support all targets.
570 if (!MIa->hasOneMemOperand() || !MIb->hasOneMemOperand())
571 llvm_unreachable("Multiple memory operands.");
573 // The following interface to AA is fashioned after DAGCombiner::isAlias
574 // and operates with MachineMemOperand offset with some important
576 // - LLVM fundamentally assumes flat address spaces.
577 // - MachineOperand offset can *only* result from legalization and
578 // cannot affect queries other than the trivial case of overlap
580 // - These offsets never wrap and never step outside
581 // of allocated objects.
582 // - There should never be any negative offsets here.
584 // FIXME: Modify API to hide this math from "user"
585 // FIXME: Even before we go to AA we can reason locally about some
586 // memory objects. It can save compile time, and possibly catch some
587 // corner cases not currently covered.
589 assert ((MMOa->getOffset() >= 0) && "Negative MachineMemOperand offset");
590 assert ((MMOb->getOffset() >= 0) && "Negative MachineMemOperand offset");
592 int64_t MinOffset = std::min(MMOa->getOffset(), MMOb->getOffset());
593 int64_t Overlapa = MMOa->getSize() + MMOa->getOffset() - MinOffset;
594 int64_t Overlapb = MMOb->getSize() + MMOb->getOffset() - MinOffset;
596 AliasAnalysis::AliasResult AAResult = AA->alias(
597 AliasAnalysis::Location(MMOa->getValue(), Overlapa,
598 MMOa->getTBAAInfo()),
599 AliasAnalysis::Location(MMOb->getValue(), Overlapb,
600 MMOb->getTBAAInfo()));
602 return (AAResult != AliasAnalysis::NoAlias);
605 /// This recursive function iterates over chain deps of SUb looking for
606 /// "latest" node that needs a chain edge to SUa.
608 iterateChainSucc(AliasAnalysis *AA, const MachineFrameInfo *MFI,
609 SUnit *SUa, SUnit *SUb, SUnit *ExitSU, unsigned *Depth,
610 SmallPtrSet<const SUnit*, 16> &Visited) {
611 if (!SUa || !SUb || SUb == ExitSU)
614 // Remember visited nodes.
615 if (!Visited.insert(SUb))
617 // If there is _some_ dependency already in place, do not
618 // descend any further.
619 // TODO: Need to make sure that if that dependency got eliminated or ignored
620 // for any reason in the future, we would not violate DAG topology.
621 // Currently it does not happen, but makes an implicit assumption about
622 // future implementation.
624 // Independently, if we encounter node that is some sort of global
625 // object (like a call) we already have full set of dependencies to it
626 // and we can stop descending.
627 if (SUa->isSucc(SUb) ||
628 isGlobalMemoryObject(AA, SUb->getInstr()))
631 // If we do need an edge, or we have exceeded depth budget,
632 // add that edge to the predecessors chain of SUb,
633 // and stop descending.
635 MIsNeedChainEdge(AA, MFI, SUa->getInstr(), SUb->getInstr())) {
636 SUb->addPred(SDep(SUa, SDep::Order, /*Latency=*/0, /*Reg=*/0,
637 /*isNormalMemory=*/true));
640 // Track current depth.
642 // Iterate over chain dependencies only.
643 for (SUnit::const_succ_iterator I = SUb->Succs.begin(), E = SUb->Succs.end();
646 iterateChainSucc (AA, MFI, SUa, I->getSUnit(), ExitSU, Depth, Visited);
650 /// This function assumes that "downward" from SU there exist
651 /// tail/leaf of already constructed DAG. It iterates downward and
652 /// checks whether SU can be aliasing any node dominated
654 static void adjustChainDeps(AliasAnalysis *AA, const MachineFrameInfo *MFI,
655 SUnit *SU, SUnit *ExitSU, std::set<SUnit *> &CheckList,
656 unsigned LatencyToLoad) {
660 SmallPtrSet<const SUnit*, 16> Visited;
663 for (std::set<SUnit *>::iterator I = CheckList.begin(), IE = CheckList.end();
667 if (MIsNeedChainEdge(AA, MFI, SU->getInstr(), (*I)->getInstr())) {
668 unsigned Latency = ((*I)->getInstr()->mayLoad()) ? LatencyToLoad : 0;
669 (*I)->addPred(SDep(SU, SDep::Order, Latency, /*Reg=*/0,
670 /*isNormalMemory=*/true));
672 // Now go through all the chain successors and iterate from them.
673 // Keep track of visited nodes.
674 for (SUnit::const_succ_iterator J = (*I)->Succs.begin(),
675 JE = (*I)->Succs.end(); J != JE; ++J)
677 iterateChainSucc (AA, MFI, SU, J->getSUnit(),
678 ExitSU, &Depth, Visited);
682 /// Check whether two objects need a chain edge, if so, add it
683 /// otherwise remember the rejected SU.
685 void addChainDependency (AliasAnalysis *AA, const MachineFrameInfo *MFI,
686 SUnit *SUa, SUnit *SUb,
687 std::set<SUnit *> &RejectList,
688 unsigned TrueMemOrderLatency = 0,
689 bool isNormalMemory = false) {
690 // If this is a false dependency,
691 // do not add the edge, but rememeber the rejected node.
692 if (!EnableAASchedMI ||
693 MIsNeedChainEdge(AA, MFI, SUa->getInstr(), SUb->getInstr()))
694 SUb->addPred(SDep(SUa, SDep::Order, TrueMemOrderLatency, /*Reg=*/0,
697 // Duplicate entries should be ignored.
698 RejectList.insert(SUb);
699 DEBUG(dbgs() << "\tReject chain dep between SU("
700 << SUa->NodeNum << ") and SU("
701 << SUb->NodeNum << ")\n");
705 /// Create an SUnit for each real instruction, numbered in top-down toplological
706 /// order. The instruction order A < B, implies that no edge exists from B to A.
708 /// Map each real instruction to its SUnit.
710 /// After initSUnits, the SUnits vector cannot be resized and the scheduler may
711 /// hang onto SUnit pointers. We may relax this in the future by using SUnit IDs
712 /// instead of pointers.
714 /// MachineScheduler relies on initSUnits numbering the nodes by their order in
715 /// the original instruction list.
716 void ScheduleDAGInstrs::initSUnits() {
717 // We'll be allocating one SUnit for each real instruction in the region,
718 // which is contained within a basic block.
719 SUnits.reserve(BB->size());
721 for (MachineBasicBlock::iterator I = RegionBegin; I != RegionEnd; ++I) {
722 MachineInstr *MI = I;
723 if (MI->isDebugValue())
726 SUnit *SU = newSUnit(MI);
729 SU->isCall = MI->isCall();
730 SU->isCommutable = MI->isCommutable();
732 // Assign the Latency field of SU using target-provided information.
740 /// If RegPressure is non null, compute register pressure as a side effect. The
741 /// DAG builder is an efficient place to do it because it already visits
743 void ScheduleDAGInstrs::buildSchedGraph(AliasAnalysis *AA,
744 RegPressureTracker *RPTracker) {
745 // Create an SUnit for each real instruction.
748 // We build scheduling units by walking a block's instruction list from bottom
751 // Remember where a generic side-effecting instruction is as we procede.
752 SUnit *BarrierChain = 0, *AliasChain = 0;
754 // Memory references to specific known memory locations are tracked
755 // so that they can be given more precise dependencies. We track
756 // separately the known memory locations that may alias and those
757 // that are known not to alias
758 std::map<const Value *, SUnit *> AliasMemDefs, NonAliasMemDefs;
759 std::map<const Value *, std::vector<SUnit *> > AliasMemUses, NonAliasMemUses;
760 std::set<SUnit*> RejectMemNodes;
762 // Remove any stale debug info; sometimes BuildSchedGraph is called again
763 // without emitting the info from the previous call.
765 FirstDbgValue = NULL;
767 assert(Defs.empty() && Uses.empty() &&
768 "Only BuildGraph should update Defs/Uses");
769 Defs.setRegLimit(TRI->getNumRegs());
770 Uses.setRegLimit(TRI->getNumRegs());
772 assert(VRegDefs.empty() && "Only BuildSchedGraph may access VRegDefs");
773 // FIXME: Allow SparseSet to reserve space for the creation of virtual
774 // registers during scheduling. Don't artificially inflate the Universe
775 // because we want to assert that vregs are not created during DAG building.
776 VRegDefs.setUniverse(MRI.getNumVirtRegs());
778 // Model data dependencies between instructions being scheduled and the
780 addSchedBarrierDeps();
782 // Walk the list of instructions, from bottom moving up.
783 MachineInstr *PrevMI = NULL;
784 for (MachineBasicBlock::iterator MII = RegionEnd, MIE = RegionBegin;
786 MachineInstr *MI = prior(MII);
788 DbgValues.push_back(std::make_pair(PrevMI, MI));
792 if (MI->isDebugValue()) {
798 assert(RPTracker->getPos() == prior(MII) && "RPTracker can't find MI");
801 assert((!MI->isTerminator() || CanHandleTerminators) && !MI->isLabel() &&
802 "Cannot schedule terminators or labels!");
804 SUnit *SU = MISUnitMap[MI];
805 assert(SU && "No SUnit mapped to this MI");
807 // Add register-based dependencies (data, anti, and output).
808 for (unsigned j = 0, n = MI->getNumOperands(); j != n; ++j) {
809 const MachineOperand &MO = MI->getOperand(j);
810 if (!MO.isReg()) continue;
811 unsigned Reg = MO.getReg();
812 if (Reg == 0) continue;
814 if (TRI->isPhysicalRegister(Reg))
815 addPhysRegDeps(SU, j);
817 assert(!IsPostRA && "Virtual register encountered!");
819 addVRegDefDeps(SU, j);
820 else if (MO.readsReg()) // ignore undef operands
821 addVRegUseDeps(SU, j);
825 // Add chain dependencies.
826 // Chain dependencies used to enforce memory order should have
827 // latency of 0 (except for true dependency of Store followed by
828 // aliased Load... we estimate that with a single cycle of latency
829 // assuming the hardware will bypass)
830 // Note that isStoreToStackSlot and isLoadFromStackSLot are not usable
831 // after stack slots are lowered to actual addresses.
832 // TODO: Use an AliasAnalysis and do real alias-analysis queries, and
833 // produce more precise dependence information.
834 unsigned TrueMemOrderLatency = MI->mayStore() ? 1 : 0;
835 if (isGlobalMemoryObject(AA, MI)) {
836 // Be conservative with these and add dependencies on all memory
837 // references, even those that are known to not alias.
838 for (std::map<const Value *, SUnit *>::iterator I =
839 NonAliasMemDefs.begin(), E = NonAliasMemDefs.end(); I != E; ++I) {
840 I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
842 for (std::map<const Value *, std::vector<SUnit *> >::iterator I =
843 NonAliasMemUses.begin(), E = NonAliasMemUses.end(); I != E; ++I) {
844 for (unsigned i = 0, e = I->second.size(); i != e; ++i)
845 I->second[i]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency));
847 // Add SU to the barrier chain.
849 BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
851 // This is a barrier event that acts as a pivotal node in the DAG,
852 // so it is safe to clear list of exposed nodes.
853 adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
854 TrueMemOrderLatency);
855 RejectMemNodes.clear();
856 NonAliasMemDefs.clear();
857 NonAliasMemUses.clear();
861 // Chain all possibly aliasing memory references though SU.
863 unsigned ChainLatency = 0;
864 if (AliasChain->getInstr()->mayLoad())
865 ChainLatency = TrueMemOrderLatency;
866 addChainDependency(AA, MFI, SU, AliasChain, RejectMemNodes,
870 for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
871 addChainDependency(AA, MFI, SU, PendingLoads[k], RejectMemNodes,
872 TrueMemOrderLatency);
873 for (std::map<const Value *, SUnit *>::iterator I = AliasMemDefs.begin(),
874 E = AliasMemDefs.end(); I != E; ++I)
875 addChainDependency(AA, MFI, SU, I->second, RejectMemNodes);
876 for (std::map<const Value *, std::vector<SUnit *> >::iterator I =
877 AliasMemUses.begin(), E = AliasMemUses.end(); I != E; ++I) {
878 for (unsigned i = 0, e = I->second.size(); i != e; ++i)
879 addChainDependency(AA, MFI, SU, I->second[i], RejectMemNodes,
880 TrueMemOrderLatency);
882 adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
883 TrueMemOrderLatency);
884 PendingLoads.clear();
885 AliasMemDefs.clear();
886 AliasMemUses.clear();
887 } else if (MI->mayStore()) {
888 bool MayAlias = true;
889 if (const Value *V = getUnderlyingObjectForInstr(MI, MFI, MayAlias)) {
890 // A store to a specific PseudoSourceValue. Add precise dependencies.
891 // Record the def in MemDefs, first adding a dep if there is
893 std::map<const Value *, SUnit *>::iterator I =
894 ((MayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V));
895 std::map<const Value *, SUnit *>::iterator IE =
896 ((MayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
898 addChainDependency(AA, MFI, SU, I->second, RejectMemNodes,
903 AliasMemDefs[V] = SU;
905 NonAliasMemDefs[V] = SU;
907 // Handle the uses in MemUses, if there are any.
908 std::map<const Value *, std::vector<SUnit *> >::iterator J =
909 ((MayAlias) ? AliasMemUses.find(V) : NonAliasMemUses.find(V));
910 std::map<const Value *, std::vector<SUnit *> >::iterator JE =
911 ((MayAlias) ? AliasMemUses.end() : NonAliasMemUses.end());
913 for (unsigned i = 0, e = J->second.size(); i != e; ++i)
914 addChainDependency(AA, MFI, SU, J->second[i], RejectMemNodes,
915 TrueMemOrderLatency, true);
919 // Add dependencies from all the PendingLoads, i.e. loads
920 // with no underlying object.
921 for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
922 addChainDependency(AA, MFI, SU, PendingLoads[k], RejectMemNodes,
923 TrueMemOrderLatency);
924 // Add dependence on alias chain, if needed.
926 addChainDependency(AA, MFI, SU, AliasChain, RejectMemNodes);
927 // But we also should check dependent instructions for the
929 adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
930 TrueMemOrderLatency);
932 // Add dependence on barrier chain, if needed.
933 // There is no point to check aliasing on barrier event. Even if
934 // SU and barrier _could_ be reordered, they should not. In addition,
935 // we have lost all RejectMemNodes below barrier.
937 BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
939 // Treat all other stores conservatively.
940 goto new_alias_chain;
943 if (!ExitSU.isPred(SU))
944 // Push store's up a bit to avoid them getting in between cmp
946 ExitSU.addPred(SDep(SU, SDep::Order, 0,
947 /*Reg=*/0, /*isNormalMemory=*/false,
948 /*isMustAlias=*/false,
949 /*isArtificial=*/true));
950 } else if (MI->mayLoad()) {
951 bool MayAlias = true;
952 if (MI->isInvariantLoad(AA)) {
953 // Invariant load, no chain dependencies needed!
956 getUnderlyingObjectForInstr(MI, MFI, MayAlias)) {
957 // A load from a specific PseudoSourceValue. Add precise dependencies.
958 std::map<const Value *, SUnit *>::iterator I =
959 ((MayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V));
960 std::map<const Value *, SUnit *>::iterator IE =
961 ((MayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
963 addChainDependency(AA, MFI, SU, I->second, RejectMemNodes, 0, true);
965 AliasMemUses[V].push_back(SU);
967 NonAliasMemUses[V].push_back(SU);
969 // A load with no underlying object. Depend on all
970 // potentially aliasing stores.
971 for (std::map<const Value *, SUnit *>::iterator I =
972 AliasMemDefs.begin(), E = AliasMemDefs.end(); I != E; ++I)
973 addChainDependency(AA, MFI, SU, I->second, RejectMemNodes);
975 PendingLoads.push_back(SU);
979 adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes, /*Latency=*/0);
980 // Add dependencies on alias and barrier chains, if needed.
981 if (MayAlias && AliasChain)
982 addChainDependency(AA, MFI, SU, AliasChain, RejectMemNodes);
984 BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
989 FirstDbgValue = PrevMI;
994 PendingLoads.clear();
997 void ScheduleDAGInstrs::computeLatency(SUnit *SU) {
998 // Compute the latency for the node. We only provide a default for missing
999 // itineraries. Empty itineraries still have latency properties.
1003 // Simplistic target-independent heuristic: assume that loads take
1005 if (SU->getInstr()->mayLoad())
1008 SU->Latency = TII->getInstrLatency(InstrItins, SU->getInstr());
1012 void ScheduleDAGInstrs::dumpNode(const SUnit *SU) const {
1013 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1014 SU->getInstr()->dump();
1018 std::string ScheduleDAGInstrs::getGraphNodeLabel(const SUnit *SU) const {
1020 raw_string_ostream oss(s);
1023 else if (SU == &ExitSU)
1026 SU->getInstr()->print(oss);
1030 /// Return the basic block label. It is not necessarilly unique because a block
1031 /// contains multiple scheduling regions. But it is fine for visualization.
1032 std::string ScheduleDAGInstrs::getDAGName() const {
1033 return "dag." + BB->getFullName();