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), CanHandleTerminators(false), LoopRegs(MDT),
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 ScheduleDAG::clearDAG();
183 /// Close the current scheduling region. Don't clear any state in case the
184 /// driver wants to refer to the previous scheduling region.
185 void ScheduleDAGInstrs::exitRegion() {
189 /// addSchedBarrierDeps - Add dependencies from instructions in the current
190 /// list of instructions being scheduled to scheduling barrier by adding
191 /// the exit SU to the register defs and use list. This is because we want to
192 /// make sure instructions which define registers that are either used by
193 /// the terminator or are live-out are properly scheduled. This is
194 /// especially important when the definition latency of the return value(s)
195 /// are too high to be hidden by the branch or when the liveout registers
196 /// used by instructions in the fallthrough block.
197 void ScheduleDAGInstrs::addSchedBarrierDeps() {
198 MachineInstr *ExitMI = RegionEnd != BB->end() ? &*RegionEnd : 0;
199 ExitSU.setInstr(ExitMI);
200 bool AllDepKnown = ExitMI &&
201 (ExitMI->isCall() || ExitMI->isBarrier());
202 if (ExitMI && AllDepKnown) {
203 // If it's a call or a barrier, add dependencies on the defs and uses of
205 for (unsigned i = 0, e = ExitMI->getNumOperands(); i != e; ++i) {
206 const MachineOperand &MO = ExitMI->getOperand(i);
207 if (!MO.isReg() || MO.isDef()) continue;
208 unsigned Reg = MO.getReg();
209 if (Reg == 0) continue;
211 if (TRI->isPhysicalRegister(Reg))
212 Uses[Reg].push_back(PhysRegSUOper(&ExitSU, -1));
214 assert(!IsPostRA && "Virtual register encountered after regalloc.");
215 addVRegUseDeps(&ExitSU, i);
219 // For others, e.g. fallthrough, conditional branch, assume the exit
220 // uses all the registers that are livein to the successor blocks.
221 assert(Uses.empty() && "Uses in set before adding deps?");
222 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
223 SE = BB->succ_end(); SI != SE; ++SI)
224 for (MachineBasicBlock::livein_iterator I = (*SI)->livein_begin(),
225 E = (*SI)->livein_end(); I != E; ++I) {
227 if (!Uses.contains(Reg))
228 Uses[Reg].push_back(PhysRegSUOper(&ExitSU, -1));
233 /// MO is an operand of SU's instruction that defines a physical register. Add
234 /// data dependencies from SU to any uses of the physical register.
235 void ScheduleDAGInstrs::addPhysRegDataDeps(SUnit *SU, unsigned OperIdx) {
236 const MachineOperand &MO = SU->getInstr()->getOperand(OperIdx);
237 assert(MO.isDef() && "expect physreg def");
239 // Ask the target if address-backscheduling is desirable, and if so how much.
240 const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
241 unsigned SpecialAddressLatency = ST.getSpecialAddressLatency();
242 unsigned DataLatency = SU->Latency;
244 for (MCRegAliasIterator Alias(MO.getReg(), TRI, true);
245 Alias.isValid(); ++Alias) {
246 if (!Uses.contains(*Alias))
248 std::vector<PhysRegSUOper> &UseList = Uses[*Alias];
249 for (unsigned i = 0, e = UseList.size(); i != e; ++i) {
250 SUnit *UseSU = UseList[i].SU;
253 MachineInstr *UseMI = UseSU->getInstr();
254 int UseOp = UseList[i].OpIdx;
255 unsigned LDataLatency = DataLatency;
256 // Optionally add in a special extra latency for nodes that
258 // TODO: Perhaps we should get rid of
259 // SpecialAddressLatency and just move this into
260 // adjustSchedDependency for the targets that care about it.
261 if (SpecialAddressLatency != 0 && UseSU != &ExitSU) {
262 const MCInstrDesc &UseMCID = UseMI->getDesc();
263 int RegUseIndex = UseMI->findRegisterUseOperandIdx(*Alias);
264 assert(RegUseIndex >= 0 && "UseMI doesn't use register!");
265 if (RegUseIndex >= 0 &&
266 (UseMI->mayLoad() || UseMI->mayStore()) &&
267 (unsigned)RegUseIndex < UseMCID.getNumOperands() &&
268 UseMCID.OpInfo[RegUseIndex].isLookupPtrRegClass())
269 LDataLatency += SpecialAddressLatency;
271 // Adjust the dependence latency using operand def/use
272 // information (if any), and then allow the target to
273 // perform its own adjustments.
274 SDep dep(SU, SDep::Data, LDataLatency, *Alias);
275 MachineInstr *RegUse = UseOp < 0 ? 0 : UseMI;
277 SchedModel.computeOperandLatency(SU->getInstr(), OperIdx,
278 RegUse, UseOp, /*FindMin=*/false));
280 SchedModel.computeOperandLatency(SU->getInstr(), OperIdx,
281 RegUse, UseOp, /*FindMin=*/true));
283 ST.adjustSchedDependency(SU, UseSU, dep);
289 /// addPhysRegDeps - Add register dependencies (data, anti, and output) from
290 /// this SUnit to following instructions in the same scheduling region that
291 /// depend the physical register referenced at OperIdx.
292 void ScheduleDAGInstrs::addPhysRegDeps(SUnit *SU, unsigned OperIdx) {
293 const MachineInstr *MI = SU->getInstr();
294 const MachineOperand &MO = MI->getOperand(OperIdx);
296 // Optionally add output and anti dependencies. For anti
297 // dependencies we use a latency of 0 because for a multi-issue
298 // target we want to allow the defining instruction to issue
299 // in the same cycle as the using instruction.
300 // TODO: Using a latency of 1 here for output dependencies assumes
301 // there's no cost for reusing registers.
302 SDep::Kind Kind = MO.isUse() ? SDep::Anti : SDep::Output;
303 for (MCRegAliasIterator Alias(MO.getReg(), TRI, true);
304 Alias.isValid(); ++Alias) {
305 if (!Defs.contains(*Alias))
307 std::vector<PhysRegSUOper> &DefList = Defs[*Alias];
308 for (unsigned i = 0, e = DefList.size(); i != e; ++i) {
309 SUnit *DefSU = DefList[i].SU;
310 if (DefSU == &ExitSU)
313 (Kind != SDep::Output || !MO.isDead() ||
314 !DefSU->getInstr()->registerDefIsDead(*Alias))) {
315 if (Kind == SDep::Anti)
316 DefSU->addPred(SDep(SU, Kind, 0, /*Reg=*/*Alias));
318 unsigned AOLat = TII->getOutputLatency(InstrItins, MI, OperIdx,
320 DefSU->addPred(SDep(SU, Kind, AOLat, /*Reg=*/*Alias));
327 // Either insert a new Reg2SUnits entry with an empty SUnits list, or
328 // retrieve the existing SUnits list for this register's uses.
329 // Push this SUnit on the use list.
330 Uses[MO.getReg()].push_back(PhysRegSUOper(SU, OperIdx));
333 addPhysRegDataDeps(SU, OperIdx);
335 // Either insert a new Reg2SUnits entry with an empty SUnits list, or
336 // retrieve the existing SUnits list for this register's defs.
337 std::vector<PhysRegSUOper> &DefList = Defs[MO.getReg()];
339 // If a def is going to wrap back around to the top of the loop,
341 if (DefList.empty()) {
342 LoopDependencies::LoopDeps::iterator I = LoopRegs.Deps.find(MO.getReg());
343 if (I != LoopRegs.Deps.end()) {
344 const MachineOperand *UseMO = I->second.first;
345 unsigned Count = I->second.second;
346 const MachineInstr *UseMI = UseMO->getParent();
347 unsigned UseMOIdx = UseMO - &UseMI->getOperand(0);
348 const MCInstrDesc &UseMCID = UseMI->getDesc();
349 const TargetSubtargetInfo &ST =
350 TM.getSubtarget<TargetSubtargetInfo>();
351 unsigned SpecialAddressLatency = ST.getSpecialAddressLatency();
352 // TODO: If we knew the total depth of the region here, we could
353 // handle the case where the whole loop is inside the region but
354 // is large enough that the isScheduleHigh trick isn't needed.
355 if (UseMOIdx < UseMCID.getNumOperands()) {
356 // Currently, we only support scheduling regions consisting of
357 // single basic blocks. Check to see if the instruction is in
358 // the same region by checking to see if it has the same parent.
359 if (UseMI->getParent() != MI->getParent()) {
360 unsigned Latency = SU->Latency;
361 if (UseMCID.OpInfo[UseMOIdx].isLookupPtrRegClass())
362 Latency += SpecialAddressLatency;
363 // This is a wild guess as to the portion of the latency which
364 // will be overlapped by work done outside the current
365 // scheduling region.
366 Latency -= std::min(Latency, Count);
367 // Add the artificial edge.
368 ExitSU.addPred(SDep(SU, SDep::Order, Latency,
369 /*Reg=*/0, /*isNormalMemory=*/false,
370 /*isMustAlias=*/false,
371 /*isArtificial=*/true));
372 } else if (SpecialAddressLatency > 0 &&
373 UseMCID.OpInfo[UseMOIdx].isLookupPtrRegClass()) {
374 // The entire loop body is within the current scheduling region
375 // and the latency of this operation is assumed to be greater
376 // than the latency of the loop.
377 // TODO: Recursively mark data-edge predecessors as
378 // isScheduleHigh too.
379 SU->isScheduleHigh = true;
382 LoopRegs.Deps.erase(I);
386 // clear this register's use list
387 if (Uses.contains(MO.getReg()))
388 Uses[MO.getReg()].clear();
393 // Calls will not be reordered because of chain dependencies (see
394 // below). Since call operands are dead, calls may continue to be added
395 // to the DefList making dependence checking quadratic in the size of
396 // the block. Instead, we leave only one call at the back of the
399 while (!DefList.empty() && DefList.back().SU->isCall)
402 // Defs are pushed in the order they are visited and never reordered.
403 DefList.push_back(PhysRegSUOper(SU, OperIdx));
407 /// addVRegDefDeps - Add register output and data dependencies from this SUnit
408 /// to instructions that occur later in the same scheduling region if they read
409 /// from or write to the virtual register defined at OperIdx.
411 /// TODO: Hoist loop induction variable increments. This has to be
412 /// reevaluated. Generally, IV scheduling should be done before coalescing.
413 void ScheduleDAGInstrs::addVRegDefDeps(SUnit *SU, unsigned OperIdx) {
414 const MachineInstr *MI = SU->getInstr();
415 unsigned Reg = MI->getOperand(OperIdx).getReg();
417 // Singly defined vregs do not have output/anti dependencies.
418 // The current operand is a def, so we have at least one.
419 // Check here if there are any others...
420 if (MRI.hasOneDef(Reg))
423 // Add output dependence to the next nearest def of this vreg.
425 // Unless this definition is dead, the output dependence should be
426 // transitively redundant with antidependencies from this definition's
427 // uses. We're conservative for now until we have a way to guarantee the uses
428 // are not eliminated sometime during scheduling. The output dependence edge
429 // is also useful if output latency exceeds def-use latency.
430 VReg2SUnitMap::iterator DefI = VRegDefs.find(Reg);
431 if (DefI == VRegDefs.end())
432 VRegDefs.insert(VReg2SUnit(Reg, SU));
434 SUnit *DefSU = DefI->SU;
435 if (DefSU != SU && DefSU != &ExitSU) {
436 unsigned OutLatency = TII->getOutputLatency(InstrItins, MI, OperIdx,
438 DefSU->addPred(SDep(SU, SDep::Output, OutLatency, Reg));
444 /// addVRegUseDeps - Add a register data dependency if the instruction that
445 /// defines the virtual register used at OperIdx is mapped to an SUnit. Add a
446 /// register antidependency from this SUnit to instructions that occur later in
447 /// the same scheduling region if they write the virtual register.
449 /// TODO: Handle ExitSU "uses" properly.
450 void ScheduleDAGInstrs::addVRegUseDeps(SUnit *SU, unsigned OperIdx) {
451 MachineInstr *MI = SU->getInstr();
452 unsigned Reg = MI->getOperand(OperIdx).getReg();
454 // Lookup this operand's reaching definition.
455 assert(LIS && "vreg dependencies requires LiveIntervals");
456 LiveRangeQuery LRQ(LIS->getInterval(Reg), LIS->getInstructionIndex(MI));
457 VNInfo *VNI = LRQ.valueIn();
459 // VNI will be valid because MachineOperand::readsReg() is checked by caller.
460 assert(VNI && "No value to read by operand");
461 MachineInstr *Def = LIS->getInstructionFromIndex(VNI->def);
462 // Phis and other noninstructions (after coalescing) have a NULL Def.
464 SUnit *DefSU = getSUnit(Def);
466 // The reaching Def lives within this scheduling region.
467 // Create a data dependence.
469 // TODO: Handle "special" address latencies cleanly.
470 SDep dep(DefSU, SDep::Data, DefSU->Latency, Reg);
471 // Adjust the dependence latency using operand def/use information, then
472 // allow the target to perform its own adjustments.
473 int DefOp = Def->findRegisterDefOperandIdx(Reg);
475 SchedModel.computeOperandLatency(Def, DefOp, MI, OperIdx, false));
477 SchedModel.computeOperandLatency(Def, DefOp, MI, OperIdx, true));
479 const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>();
480 ST.adjustSchedDependency(DefSU, SU, const_cast<SDep &>(dep));
485 // Add antidependence to the following def of the vreg it uses.
486 VReg2SUnitMap::iterator DefI = VRegDefs.find(Reg);
487 if (DefI != VRegDefs.end() && DefI->SU != SU)
488 DefI->SU->addPred(SDep(SU, SDep::Anti, 0, Reg));
491 /// Return true if MI is an instruction we are unable to reason about
492 /// (like a call or something with unmodeled side effects).
493 static inline bool isGlobalMemoryObject(AliasAnalysis *AA, MachineInstr *MI) {
494 if (MI->isCall() || MI->hasUnmodeledSideEffects() ||
495 (MI->hasOrderedMemoryRef() &&
496 (!MI->mayLoad() || !MI->isInvariantLoad(AA))))
501 // This MI might have either incomplete info, or known to be unsafe
502 // to deal with (i.e. volatile object).
503 static inline bool isUnsafeMemoryObject(MachineInstr *MI,
504 const MachineFrameInfo *MFI) {
505 if (!MI || MI->memoperands_empty())
507 // We purposefully do no check for hasOneMemOperand() here
508 // in hope to trigger an assert downstream in order to
509 // finish implementation.
510 if ((*MI->memoperands_begin())->isVolatile() ||
511 MI->hasUnmodeledSideEffects())
514 const Value *V = (*MI->memoperands_begin())->getValue();
518 V = getUnderlyingObject(V);
519 if (const PseudoSourceValue *PSV = dyn_cast<PseudoSourceValue>(V)) {
520 // Similarly to getUnderlyingObjectForInstr:
521 // For now, ignore PseudoSourceValues which may alias LLVM IR values
522 // because the code that uses this function has no way to cope with
524 if (PSV->isAliased(MFI))
527 // Does this pointer refer to a distinct and identifiable object?
528 if (!isIdentifiedObject(V))
534 /// This returns true if the two MIs need a chain edge betwee them.
535 /// If these are not even memory operations, we still may need
536 /// chain deps between them. The question really is - could
537 /// these two MIs be reordered during scheduling from memory dependency
539 static bool MIsNeedChainEdge(AliasAnalysis *AA, const MachineFrameInfo *MFI,
542 // Cover a trivial case - no edge is need to itself.
546 if (isUnsafeMemoryObject(MIa, MFI) || isUnsafeMemoryObject(MIb, MFI))
549 // If we are dealing with two "normal" loads, we do not need an edge
550 // between them - they could be reordered.
551 if (!MIa->mayStore() && !MIb->mayStore())
554 // To this point analysis is generic. From here on we do need AA.
558 MachineMemOperand *MMOa = *MIa->memoperands_begin();
559 MachineMemOperand *MMOb = *MIb->memoperands_begin();
561 // FIXME: Need to handle multiple memory operands to support all targets.
562 if (!MIa->hasOneMemOperand() || !MIb->hasOneMemOperand())
563 llvm_unreachable("Multiple memory operands.");
565 // The following interface to AA is fashioned after DAGCombiner::isAlias
566 // and operates with MachineMemOperand offset with some important
568 // - LLVM fundamentally assumes flat address spaces.
569 // - MachineOperand offset can *only* result from legalization and
570 // cannot affect queries other than the trivial case of overlap
572 // - These offsets never wrap and never step outside
573 // of allocated objects.
574 // - There should never be any negative offsets here.
576 // FIXME: Modify API to hide this math from "user"
577 // FIXME: Even before we go to AA we can reason locally about some
578 // memory objects. It can save compile time, and possibly catch some
579 // corner cases not currently covered.
581 assert ((MMOa->getOffset() >= 0) && "Negative MachineMemOperand offset");
582 assert ((MMOb->getOffset() >= 0) && "Negative MachineMemOperand offset");
584 int64_t MinOffset = std::min(MMOa->getOffset(), MMOb->getOffset());
585 int64_t Overlapa = MMOa->getSize() + MMOa->getOffset() - MinOffset;
586 int64_t Overlapb = MMOb->getSize() + MMOb->getOffset() - MinOffset;
588 AliasAnalysis::AliasResult AAResult = AA->alias(
589 AliasAnalysis::Location(MMOa->getValue(), Overlapa,
590 MMOa->getTBAAInfo()),
591 AliasAnalysis::Location(MMOb->getValue(), Overlapb,
592 MMOb->getTBAAInfo()));
594 return (AAResult != AliasAnalysis::NoAlias);
597 /// This recursive function iterates over chain deps of SUb looking for
598 /// "latest" node that needs a chain edge to SUa.
600 iterateChainSucc(AliasAnalysis *AA, const MachineFrameInfo *MFI,
601 SUnit *SUa, SUnit *SUb, SUnit *ExitSU, unsigned *Depth,
602 SmallPtrSet<const SUnit*, 16> &Visited) {
603 if (!SUa || !SUb || SUb == ExitSU)
606 // Remember visited nodes.
607 if (!Visited.insert(SUb))
609 // If there is _some_ dependency already in place, do not
610 // descend any further.
611 // TODO: Need to make sure that if that dependency got eliminated or ignored
612 // for any reason in the future, we would not violate DAG topology.
613 // Currently it does not happen, but makes an implicit assumption about
614 // future implementation.
616 // Independently, if we encounter node that is some sort of global
617 // object (like a call) we already have full set of dependencies to it
618 // and we can stop descending.
619 if (SUa->isSucc(SUb) ||
620 isGlobalMemoryObject(AA, SUb->getInstr()))
623 // If we do need an edge, or we have exceeded depth budget,
624 // add that edge to the predecessors chain of SUb,
625 // and stop descending.
627 MIsNeedChainEdge(AA, MFI, SUa->getInstr(), SUb->getInstr())) {
628 SUb->addPred(SDep(SUa, SDep::Order, /*Latency=*/0, /*Reg=*/0,
629 /*isNormalMemory=*/true));
632 // Track current depth.
634 // Iterate over chain dependencies only.
635 for (SUnit::const_succ_iterator I = SUb->Succs.begin(), E = SUb->Succs.end();
638 iterateChainSucc (AA, MFI, SUa, I->getSUnit(), ExitSU, Depth, Visited);
642 /// This function assumes that "downward" from SU there exist
643 /// tail/leaf of already constructed DAG. It iterates downward and
644 /// checks whether SU can be aliasing any node dominated
646 static void adjustChainDeps(AliasAnalysis *AA, const MachineFrameInfo *MFI,
647 SUnit *SU, SUnit *ExitSU, std::set<SUnit *> &CheckList,
648 unsigned LatencyToLoad) {
652 SmallPtrSet<const SUnit*, 16> Visited;
655 for (std::set<SUnit *>::iterator I = CheckList.begin(), IE = CheckList.end();
659 if (MIsNeedChainEdge(AA, MFI, SU->getInstr(), (*I)->getInstr())) {
660 unsigned Latency = ((*I)->getInstr()->mayLoad()) ? LatencyToLoad : 0;
661 (*I)->addPred(SDep(SU, SDep::Order, Latency, /*Reg=*/0,
662 /*isNormalMemory=*/true));
664 // Now go through all the chain successors and iterate from them.
665 // Keep track of visited nodes.
666 for (SUnit::const_succ_iterator J = (*I)->Succs.begin(),
667 JE = (*I)->Succs.end(); J != JE; ++J)
669 iterateChainSucc (AA, MFI, SU, J->getSUnit(),
670 ExitSU, &Depth, Visited);
674 /// Check whether two objects need a chain edge, if so, add it
675 /// otherwise remember the rejected SU.
677 void addChainDependency (AliasAnalysis *AA, const MachineFrameInfo *MFI,
678 SUnit *SUa, SUnit *SUb,
679 std::set<SUnit *> &RejectList,
680 unsigned TrueMemOrderLatency = 0,
681 bool isNormalMemory = false) {
682 // If this is a false dependency,
683 // do not add the edge, but rememeber the rejected node.
684 if (!EnableAASchedMI ||
685 MIsNeedChainEdge(AA, MFI, SUa->getInstr(), SUb->getInstr()))
686 SUb->addPred(SDep(SUa, SDep::Order, TrueMemOrderLatency, /*Reg=*/0,
689 // Duplicate entries should be ignored.
690 RejectList.insert(SUb);
691 DEBUG(dbgs() << "\tReject chain dep between SU("
692 << SUa->NodeNum << ") and SU("
693 << SUb->NodeNum << ")\n");
697 /// Create an SUnit for each real instruction, numbered in top-down toplological
698 /// order. The instruction order A < B, implies that no edge exists from B to A.
700 /// Map each real instruction to its SUnit.
702 /// After initSUnits, the SUnits vector cannot be resized and the scheduler may
703 /// hang onto SUnit pointers. We may relax this in the future by using SUnit IDs
704 /// instead of pointers.
706 /// MachineScheduler relies on initSUnits numbering the nodes by their order in
707 /// the original instruction list.
708 void ScheduleDAGInstrs::initSUnits() {
709 // We'll be allocating one SUnit for each real instruction in the region,
710 // which is contained within a basic block.
711 SUnits.reserve(BB->size());
713 for (MachineBasicBlock::iterator I = RegionBegin; I != RegionEnd; ++I) {
714 MachineInstr *MI = I;
715 if (MI->isDebugValue())
718 SUnit *SU = newSUnit(MI);
721 SU->isCall = MI->isCall();
722 SU->isCommutable = MI->isCommutable();
724 // Assign the Latency field of SU using target-provided information.
729 /// If RegPressure is non null, compute register pressure as a side effect. The
730 /// DAG builder is an efficient place to do it because it already visits
732 void ScheduleDAGInstrs::buildSchedGraph(AliasAnalysis *AA,
733 RegPressureTracker *RPTracker) {
734 // Create an SUnit for each real instruction.
737 // We build scheduling units by walking a block's instruction list from bottom
740 // Remember where a generic side-effecting instruction is as we procede.
741 SUnit *BarrierChain = 0, *AliasChain = 0;
743 // Memory references to specific known memory locations are tracked
744 // so that they can be given more precise dependencies. We track
745 // separately the known memory locations that may alias and those
746 // that are known not to alias
747 std::map<const Value *, SUnit *> AliasMemDefs, NonAliasMemDefs;
748 std::map<const Value *, std::vector<SUnit *> > AliasMemUses, NonAliasMemUses;
749 std::set<SUnit*> RejectMemNodes;
751 // Remove any stale debug info; sometimes BuildSchedGraph is called again
752 // without emitting the info from the previous call.
754 FirstDbgValue = NULL;
756 assert(Defs.empty() && Uses.empty() &&
757 "Only BuildGraph should update Defs/Uses");
758 Defs.setRegLimit(TRI->getNumRegs());
759 Uses.setRegLimit(TRI->getNumRegs());
761 assert(VRegDefs.empty() && "Only BuildSchedGraph may access VRegDefs");
762 // FIXME: Allow SparseSet to reserve space for the creation of virtual
763 // registers during scheduling. Don't artificially inflate the Universe
764 // because we want to assert that vregs are not created during DAG building.
765 VRegDefs.setUniverse(MRI.getNumVirtRegs());
767 // Model data dependencies between instructions being scheduled and the
769 addSchedBarrierDeps();
771 // Walk the list of instructions, from bottom moving up.
772 MachineInstr *PrevMI = NULL;
773 for (MachineBasicBlock::iterator MII = RegionEnd, MIE = RegionBegin;
775 MachineInstr *MI = prior(MII);
777 DbgValues.push_back(std::make_pair(PrevMI, MI));
781 if (MI->isDebugValue()) {
787 assert(RPTracker->getPos() == prior(MII) && "RPTracker can't find MI");
790 assert((!MI->isTerminator() || CanHandleTerminators) && !MI->isLabel() &&
791 "Cannot schedule terminators or labels!");
793 SUnit *SU = MISUnitMap[MI];
794 assert(SU && "No SUnit mapped to this MI");
796 // Add register-based dependencies (data, anti, and output).
797 for (unsigned j = 0, n = MI->getNumOperands(); j != n; ++j) {
798 const MachineOperand &MO = MI->getOperand(j);
799 if (!MO.isReg()) continue;
800 unsigned Reg = MO.getReg();
801 if (Reg == 0) continue;
803 if (TRI->isPhysicalRegister(Reg))
804 addPhysRegDeps(SU, j);
806 assert(!IsPostRA && "Virtual register encountered!");
808 addVRegDefDeps(SU, j);
809 else if (MO.readsReg()) // ignore undef operands
810 addVRegUseDeps(SU, j);
814 // Add chain dependencies.
815 // Chain dependencies used to enforce memory order should have
816 // latency of 0 (except for true dependency of Store followed by
817 // aliased Load... we estimate that with a single cycle of latency
818 // assuming the hardware will bypass)
819 // Note that isStoreToStackSlot and isLoadFromStackSLot are not usable
820 // after stack slots are lowered to actual addresses.
821 // TODO: Use an AliasAnalysis and do real alias-analysis queries, and
822 // produce more precise dependence information.
823 unsigned TrueMemOrderLatency = MI->mayStore() ? 1 : 0;
824 if (isGlobalMemoryObject(AA, MI)) {
825 // Be conservative with these and add dependencies on all memory
826 // references, even those that are known to not alias.
827 for (std::map<const Value *, SUnit *>::iterator I =
828 NonAliasMemDefs.begin(), E = NonAliasMemDefs.end(); I != E; ++I) {
829 I->second->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
831 for (std::map<const Value *, std::vector<SUnit *> >::iterator I =
832 NonAliasMemUses.begin(), E = NonAliasMemUses.end(); I != E; ++I) {
833 for (unsigned i = 0, e = I->second.size(); i != e; ++i)
834 I->second[i]->addPred(SDep(SU, SDep::Order, TrueMemOrderLatency));
836 // Add SU to the barrier chain.
838 BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
840 // This is a barrier event that acts as a pivotal node in the DAG,
841 // so it is safe to clear list of exposed nodes.
842 adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
843 TrueMemOrderLatency);
844 RejectMemNodes.clear();
845 NonAliasMemDefs.clear();
846 NonAliasMemUses.clear();
850 // Chain all possibly aliasing memory references though SU.
852 unsigned ChainLatency = 0;
853 if (AliasChain->getInstr()->mayLoad())
854 ChainLatency = TrueMemOrderLatency;
855 addChainDependency(AA, MFI, SU, AliasChain, RejectMemNodes,
859 for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
860 addChainDependency(AA, MFI, SU, PendingLoads[k], RejectMemNodes,
861 TrueMemOrderLatency);
862 for (std::map<const Value *, SUnit *>::iterator I = AliasMemDefs.begin(),
863 E = AliasMemDefs.end(); I != E; ++I)
864 addChainDependency(AA, MFI, SU, I->second, RejectMemNodes);
865 for (std::map<const Value *, std::vector<SUnit *> >::iterator I =
866 AliasMemUses.begin(), E = AliasMemUses.end(); I != E; ++I) {
867 for (unsigned i = 0, e = I->second.size(); i != e; ++i)
868 addChainDependency(AA, MFI, SU, I->second[i], RejectMemNodes,
869 TrueMemOrderLatency);
871 adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
872 TrueMemOrderLatency);
873 PendingLoads.clear();
874 AliasMemDefs.clear();
875 AliasMemUses.clear();
876 } else if (MI->mayStore()) {
877 bool MayAlias = true;
878 if (const Value *V = getUnderlyingObjectForInstr(MI, MFI, MayAlias)) {
879 // A store to a specific PseudoSourceValue. Add precise dependencies.
880 // Record the def in MemDefs, first adding a dep if there is
882 std::map<const Value *, SUnit *>::iterator I =
883 ((MayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V));
884 std::map<const Value *, SUnit *>::iterator IE =
885 ((MayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
887 addChainDependency(AA, MFI, SU, I->second, RejectMemNodes,
892 AliasMemDefs[V] = SU;
894 NonAliasMemDefs[V] = SU;
896 // Handle the uses in MemUses, if there are any.
897 std::map<const Value *, std::vector<SUnit *> >::iterator J =
898 ((MayAlias) ? AliasMemUses.find(V) : NonAliasMemUses.find(V));
899 std::map<const Value *, std::vector<SUnit *> >::iterator JE =
900 ((MayAlias) ? AliasMemUses.end() : NonAliasMemUses.end());
902 for (unsigned i = 0, e = J->second.size(); i != e; ++i)
903 addChainDependency(AA, MFI, SU, J->second[i], RejectMemNodes,
904 TrueMemOrderLatency, true);
908 // Add dependencies from all the PendingLoads, i.e. loads
909 // with no underlying object.
910 for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
911 addChainDependency(AA, MFI, SU, PendingLoads[k], RejectMemNodes,
912 TrueMemOrderLatency);
913 // Add dependence on alias chain, if needed.
915 addChainDependency(AA, MFI, SU, AliasChain, RejectMemNodes);
916 // But we also should check dependent instructions for the
918 adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes,
919 TrueMemOrderLatency);
921 // Add dependence on barrier chain, if needed.
922 // There is no point to check aliasing on barrier event. Even if
923 // SU and barrier _could_ be reordered, they should not. In addition,
924 // we have lost all RejectMemNodes below barrier.
926 BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
928 // Treat all other stores conservatively.
929 goto new_alias_chain;
932 if (!ExitSU.isPred(SU))
933 // Push store's up a bit to avoid them getting in between cmp
935 ExitSU.addPred(SDep(SU, SDep::Order, 0,
936 /*Reg=*/0, /*isNormalMemory=*/false,
937 /*isMustAlias=*/false,
938 /*isArtificial=*/true));
939 } else if (MI->mayLoad()) {
940 bool MayAlias = true;
941 if (MI->isInvariantLoad(AA)) {
942 // Invariant load, no chain dependencies needed!
945 getUnderlyingObjectForInstr(MI, MFI, MayAlias)) {
946 // A load from a specific PseudoSourceValue. Add precise dependencies.
947 std::map<const Value *, SUnit *>::iterator I =
948 ((MayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V));
949 std::map<const Value *, SUnit *>::iterator IE =
950 ((MayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
952 addChainDependency(AA, MFI, SU, I->second, RejectMemNodes, 0, true);
954 AliasMemUses[V].push_back(SU);
956 NonAliasMemUses[V].push_back(SU);
958 // A load with no underlying object. Depend on all
959 // potentially aliasing stores.
960 for (std::map<const Value *, SUnit *>::iterator I =
961 AliasMemDefs.begin(), E = AliasMemDefs.end(); I != E; ++I)
962 addChainDependency(AA, MFI, SU, I->second, RejectMemNodes);
964 PendingLoads.push_back(SU);
968 adjustChainDeps(AA, MFI, SU, &ExitSU, RejectMemNodes, /*Latency=*/0);
969 // Add dependencies on alias and barrier chains, if needed.
970 if (MayAlias && AliasChain)
971 addChainDependency(AA, MFI, SU, AliasChain, RejectMemNodes);
973 BarrierChain->addPred(SDep(SU, SDep::Order, /*Latency=*/0));
978 FirstDbgValue = PrevMI;
983 PendingLoads.clear();
986 void ScheduleDAGInstrs::computeLatency(SUnit *SU) {
987 // Compute the latency for the node. We only provide a default for missing
988 // itineraries. Empty itineraries still have latency properties.
992 // Simplistic target-independent heuristic: assume that loads take
994 if (SU->getInstr()->mayLoad())
997 SU->Latency = TII->getInstrLatency(InstrItins, SU->getInstr());
1001 void ScheduleDAGInstrs::dumpNode(const SUnit *SU) const {
1002 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1003 SU->getInstr()->dump();
1007 std::string ScheduleDAGInstrs::getGraphNodeLabel(const SUnit *SU) const {
1009 raw_string_ostream oss(s);
1012 else if (SU == &ExitSU)
1015 SU->getInstr()->print(oss);
1019 /// Return the basic block label. It is not necessarilly unique because a block
1020 /// contains multiple scheduling regions. But it is fine for visualization.
1021 std::string ScheduleDAGInstrs::getDAGName() const {
1022 return "dag." + BB->getFullName();