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 #include "llvm/CodeGen/ScheduleDAGInstrs.h"
16 #include "llvm/ADT/MapVector.h"
17 #include "llvm/ADT/SmallPtrSet.h"
18 #include "llvm/ADT/SmallSet.h"
19 #include "llvm/Analysis/AliasAnalysis.h"
20 #include "llvm/Analysis/ValueTracking.h"
21 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
22 #include "llvm/CodeGen/MachineFunctionPass.h"
23 #include "llvm/CodeGen/MachineFrameInfo.h"
24 #include "llvm/CodeGen/MachineInstrBuilder.h"
25 #include "llvm/CodeGen/MachineMemOperand.h"
26 #include "llvm/CodeGen/MachineRegisterInfo.h"
27 #include "llvm/CodeGen/PseudoSourceValue.h"
28 #include "llvm/CodeGen/RegisterPressure.h"
29 #include "llvm/CodeGen/ScheduleDFS.h"
30 #include "llvm/IR/Operator.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/Debug.h"
33 #include "llvm/Support/Format.h"
34 #include "llvm/Support/raw_ostream.h"
35 #include "llvm/Target/TargetInstrInfo.h"
36 #include "llvm/Target/TargetMachine.h"
37 #include "llvm/Target/TargetRegisterInfo.h"
38 #include "llvm/Target/TargetSubtargetInfo.h"
43 #define DEBUG_TYPE "misched"
45 static cl::opt<bool> EnableAASchedMI("enable-aa-sched-mi", cl::Hidden,
46 cl::ZeroOrMore, cl::init(false),
47 cl::desc("Enable use of AA during MI DAG construction"));
49 static cl::opt<bool> UseTBAA("use-tbaa-in-sched-mi", cl::Hidden,
50 cl::init(true), cl::desc("Enable use of TBAA during MI DAG construction"));
52 ScheduleDAGInstrs::ScheduleDAGInstrs(MachineFunction &mf,
53 const MachineLoopInfo *mli,
56 : ScheduleDAG(mf), MLI(mli), MFI(mf.getFrameInfo()), LIS(LIS),
57 RemoveKillFlags(RemoveKillFlags), CanHandleTerminators(false),
58 FirstDbgValue(nullptr) {
61 const TargetSubtargetInfo &ST = mf.getSubtarget();
62 SchedModel.init(ST.getSchedModel(), &ST, TII);
65 /// getUnderlyingObjectFromInt - This is the function that does the work of
66 /// looking through basic ptrtoint+arithmetic+inttoptr sequences.
67 static const Value *getUnderlyingObjectFromInt(const Value *V) {
69 if (const Operator *U = dyn_cast<Operator>(V)) {
70 // If we find a ptrtoint, we can transfer control back to the
71 // regular getUnderlyingObjectFromInt.
72 if (U->getOpcode() == Instruction::PtrToInt)
73 return U->getOperand(0);
74 // If we find an add of a constant, a multiplied value, or a phi, it's
75 // likely that the other operand will lead us to the base
76 // object. We don't have to worry about the case where the
77 // object address is somehow being computed by the multiply,
78 // because our callers only care when the result is an
79 // identifiable object.
80 if (U->getOpcode() != Instruction::Add ||
81 (!isa<ConstantInt>(U->getOperand(1)) &&
82 Operator::getOpcode(U->getOperand(1)) != Instruction::Mul &&
83 !isa<PHINode>(U->getOperand(1))))
89 assert(V->getType()->isIntegerTy() && "Unexpected operand type!");
93 /// getUnderlyingObjects - This is a wrapper around GetUnderlyingObjects
94 /// and adds support for basic ptrtoint+arithmetic+inttoptr sequences.
95 static void getUnderlyingObjects(const Value *V,
96 SmallVectorImpl<Value *> &Objects,
97 const DataLayout &DL) {
98 SmallPtrSet<const Value *, 16> Visited;
99 SmallVector<const Value *, 4> Working(1, V);
101 V = Working.pop_back_val();
103 SmallVector<Value *, 4> Objs;
104 GetUnderlyingObjects(const_cast<Value *>(V), Objs, DL);
106 for (SmallVectorImpl<Value *>::iterator I = Objs.begin(), IE = Objs.end();
109 if (!Visited.insert(V).second)
111 if (Operator::getOpcode(V) == Instruction::IntToPtr) {
113 getUnderlyingObjectFromInt(cast<User>(V)->getOperand(0));
114 if (O->getType()->isPointerTy()) {
115 Working.push_back(O);
119 Objects.push_back(const_cast<Value *>(V));
121 } while (!Working.empty());
124 typedef PointerUnion<const Value *, const PseudoSourceValue *> ValueType;
125 typedef SmallVector<PointerIntPair<ValueType, 1, bool>, 4>
126 UnderlyingObjectsVector;
128 /// getUnderlyingObjectsForInstr - If this machine instr has memory reference
129 /// information and it can be tracked to a normal reference to a known
130 /// object, return the Value for that object.
131 static void getUnderlyingObjectsForInstr(const MachineInstr *MI,
132 const MachineFrameInfo *MFI,
133 UnderlyingObjectsVector &Objects,
134 const DataLayout &DL) {
135 if (!MI->hasOneMemOperand() ||
136 (!(*MI->memoperands_begin())->getValue() &&
137 !(*MI->memoperands_begin())->getPseudoValue()) ||
138 (*MI->memoperands_begin())->isVolatile())
141 if (const PseudoSourceValue *PSV =
142 (*MI->memoperands_begin())->getPseudoValue()) {
143 // Function that contain tail calls don't have unique PseudoSourceValue
144 // objects. Two PseudoSourceValues might refer to the same or overlapping
145 // locations. The client code calling this function assumes this is not the
146 // case. So return a conservative answer of no known object.
147 if (MFI->hasTailCall())
150 // For now, ignore PseudoSourceValues which may alias LLVM IR values
151 // because the code that uses this function has no way to cope with
153 if (!PSV->isAliased(MFI)) {
154 bool MayAlias = PSV->mayAlias(MFI);
155 Objects.push_back(UnderlyingObjectsVector::value_type(PSV, MayAlias));
160 const Value *V = (*MI->memoperands_begin())->getValue();
164 SmallVector<Value *, 4> Objs;
165 getUnderlyingObjects(V, Objs, DL);
167 for (Value *V : Objs) {
168 if (!isIdentifiedObject(V)) {
173 Objects.push_back(UnderlyingObjectsVector::value_type(V, true));
177 void ScheduleDAGInstrs::startBlock(MachineBasicBlock *bb) {
181 void ScheduleDAGInstrs::finishBlock() {
182 // Subclasses should no longer refer to the old block.
186 /// Initialize the DAG and common scheduler state for the current scheduling
187 /// region. This does not actually create the DAG, only clears it. The
188 /// scheduling driver may call BuildSchedGraph multiple times per scheduling
190 void ScheduleDAGInstrs::enterRegion(MachineBasicBlock *bb,
191 MachineBasicBlock::iterator begin,
192 MachineBasicBlock::iterator end,
193 unsigned regioninstrs) {
194 assert(bb == BB && "startBlock should set BB");
197 NumRegionInstrs = regioninstrs;
200 /// Close the current scheduling region. Don't clear any state in case the
201 /// driver wants to refer to the previous scheduling region.
202 void ScheduleDAGInstrs::exitRegion() {
206 /// addSchedBarrierDeps - Add dependencies from instructions in the current
207 /// list of instructions being scheduled to scheduling barrier by adding
208 /// the exit SU to the register defs and use list. This is because we want to
209 /// make sure instructions which define registers that are either used by
210 /// the terminator or are live-out are properly scheduled. This is
211 /// especially important when the definition latency of the return value(s)
212 /// are too high to be hidden by the branch or when the liveout registers
213 /// used by instructions in the fallthrough block.
214 void ScheduleDAGInstrs::addSchedBarrierDeps() {
215 MachineInstr *ExitMI = RegionEnd != BB->end() ? &*RegionEnd : nullptr;
216 ExitSU.setInstr(ExitMI);
217 bool AllDepKnown = ExitMI &&
218 (ExitMI->isCall() || ExitMI->isBarrier());
219 if (ExitMI && AllDepKnown) {
220 // If it's a call or a barrier, add dependencies on the defs and uses of
222 for (unsigned i = 0, e = ExitMI->getNumOperands(); i != e; ++i) {
223 const MachineOperand &MO = ExitMI->getOperand(i);
224 if (!MO.isReg() || MO.isDef()) continue;
225 unsigned Reg = MO.getReg();
226 if (Reg == 0) continue;
228 if (TRI->isPhysicalRegister(Reg))
229 Uses.insert(PhysRegSUOper(&ExitSU, -1, Reg));
230 else if (MO.readsReg()) // ignore undef operands
231 addVRegUseDeps(&ExitSU, i);
234 // For others, e.g. fallthrough, conditional branch, assume the exit
235 // uses all the registers that are livein to the successor blocks.
236 assert(Uses.empty() && "Uses in set before adding deps?");
237 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
238 SE = BB->succ_end(); SI != SE; ++SI)
239 for (const auto &LI : (*SI)->liveins()) {
240 if (!Uses.contains(LI.PhysReg))
241 Uses.insert(PhysRegSUOper(&ExitSU, -1, LI.PhysReg));
246 /// MO is an operand of SU's instruction that defines a physical register. Add
247 /// data dependencies from SU to any uses of the physical register.
248 void ScheduleDAGInstrs::addPhysRegDataDeps(SUnit *SU, unsigned OperIdx) {
249 const MachineOperand &MO = SU->getInstr()->getOperand(OperIdx);
250 assert(MO.isDef() && "expect physreg def");
252 // Ask the target if address-backscheduling is desirable, and if so how much.
253 const TargetSubtargetInfo &ST = MF.getSubtarget();
255 for (MCRegAliasIterator Alias(MO.getReg(), TRI, true);
256 Alias.isValid(); ++Alias) {
257 if (!Uses.contains(*Alias))
259 for (Reg2SUnitsMap::iterator I = Uses.find(*Alias); I != Uses.end(); ++I) {
260 SUnit *UseSU = I->SU;
264 // Adjust the dependence latency using operand def/use information,
265 // then allow the target to perform its own adjustments.
266 int UseOp = I->OpIdx;
267 MachineInstr *RegUse = nullptr;
270 Dep = SDep(SU, SDep::Artificial);
272 // Set the hasPhysRegDefs only for physreg defs that have a use within
273 // the scheduling region.
274 SU->hasPhysRegDefs = true;
275 Dep = SDep(SU, SDep::Data, *Alias);
276 RegUse = UseSU->getInstr();
279 SchedModel.computeOperandLatency(SU->getInstr(), OperIdx, RegUse,
282 ST.adjustSchedDependency(SU, UseSU, Dep);
288 /// addPhysRegDeps - Add register dependencies (data, anti, and output) from
289 /// this SUnit to following instructions in the same scheduling region that
290 /// depend the physical register referenced at OperIdx.
291 void ScheduleDAGInstrs::addPhysRegDeps(SUnit *SU, unsigned OperIdx) {
292 MachineInstr *MI = SU->getInstr();
293 MachineOperand &MO = MI->getOperand(OperIdx);
295 // Optionally add output and anti dependencies. For anti
296 // dependencies we use a latency of 0 because for a multi-issue
297 // target we want to allow the defining instruction to issue
298 // in the same cycle as the using instruction.
299 // TODO: Using a latency of 1 here for output dependencies assumes
300 // there's no cost for reusing registers.
301 SDep::Kind Kind = MO.isUse() ? SDep::Anti : SDep::Output;
302 for (MCRegAliasIterator Alias(MO.getReg(), TRI, true);
303 Alias.isValid(); ++Alias) {
304 if (!Defs.contains(*Alias))
306 for (Reg2SUnitsMap::iterator I = Defs.find(*Alias); I != Defs.end(); ++I) {
307 SUnit *DefSU = I->SU;
308 if (DefSU == &ExitSU)
311 (Kind != SDep::Output || !MO.isDead() ||
312 !DefSU->getInstr()->registerDefIsDead(*Alias))) {
313 if (Kind == SDep::Anti)
314 DefSU->addPred(SDep(SU, Kind, /*Reg=*/*Alias));
316 SDep Dep(SU, Kind, /*Reg=*/*Alias);
318 SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
326 SU->hasPhysRegUses = true;
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.insert(PhysRegSUOper(SU, OperIdx, MO.getReg()));
335 addPhysRegDataDeps(SU, OperIdx);
336 unsigned Reg = MO.getReg();
338 // clear this register's use list
339 if (Uses.contains(Reg))
344 } else if (SU->isCall) {
345 // Calls will not be reordered because of chain dependencies (see
346 // below). Since call operands are dead, calls may continue to be added
347 // to the DefList making dependence checking quadratic in the size of
348 // the block. Instead, we leave only one call at the back of the
350 Reg2SUnitsMap::RangePair P = Defs.equal_range(Reg);
351 Reg2SUnitsMap::iterator B = P.first;
352 Reg2SUnitsMap::iterator I = P.second;
353 for (bool isBegin = I == B; !isBegin; /* empty */) {
354 isBegin = (--I) == B;
361 // Defs are pushed in the order they are visited and never reordered.
362 Defs.insert(PhysRegSUOper(SU, OperIdx, Reg));
366 /// addVRegDefDeps - Add register output and data dependencies from this SUnit
367 /// to instructions that occur later in the same scheduling region if they read
368 /// from or write to the virtual register defined at OperIdx.
370 /// TODO: Hoist loop induction variable increments. This has to be
371 /// reevaluated. Generally, IV scheduling should be done before coalescing.
372 void ScheduleDAGInstrs::addVRegDefDeps(SUnit *SU, unsigned OperIdx) {
373 const MachineInstr *MI = SU->getInstr();
374 unsigned Reg = MI->getOperand(OperIdx).getReg();
376 // Singly defined vregs do not have output/anti dependencies.
377 // The current operand is a def, so we have at least one.
378 // Check here if there are any others...
379 if (MRI.hasOneDef(Reg))
382 // Add output dependence to the next nearest def of this vreg.
384 // Unless this definition is dead, the output dependence should be
385 // transitively redundant with antidependencies from this definition's
386 // uses. We're conservative for now until we have a way to guarantee the uses
387 // are not eliminated sometime during scheduling. The output dependence edge
388 // is also useful if output latency exceeds def-use latency.
389 VReg2SUnitMap::iterator DefI = VRegDefs.find(Reg);
390 if (DefI == VRegDefs.end())
391 VRegDefs.insert(VReg2SUnit(Reg, SU));
393 SUnit *DefSU = DefI->SU;
394 if (DefSU != SU && DefSU != &ExitSU) {
395 SDep Dep(SU, SDep::Output, Reg);
397 SchedModel.computeOutputLatency(MI, OperIdx, DefSU->getInstr()));
404 /// addVRegUseDeps - Add a register data dependency if the instruction that
405 /// defines the virtual register used at OperIdx is mapped to an SUnit. Add a
406 /// register antidependency from this SUnit to instructions that occur later in
407 /// the same scheduling region if they write the virtual register.
409 /// TODO: Handle ExitSU "uses" properly.
410 void ScheduleDAGInstrs::addVRegUseDeps(SUnit *SU, unsigned OperIdx) {
411 MachineInstr *MI = SU->getInstr();
412 unsigned Reg = MI->getOperand(OperIdx).getReg();
414 // Record this local VReg use.
415 VReg2UseMap::iterator UI = VRegUses.find(Reg);
416 for (; UI != VRegUses.end(); ++UI) {
420 if (UI == VRegUses.end())
421 VRegUses.insert(VReg2SUnit(Reg, SU));
423 // Lookup this operand's reaching definition.
424 assert(LIS && "vreg dependencies requires LiveIntervals");
426 = LIS->getInterval(Reg).Query(LIS->getInstructionIndex(MI));
427 VNInfo *VNI = LRQ.valueIn();
429 // VNI will be valid because MachineOperand::readsReg() is checked by caller.
430 assert(VNI && "No value to read by operand");
431 MachineInstr *Def = LIS->getInstructionFromIndex(VNI->def);
432 // Phis and other noninstructions (after coalescing) have a NULL Def.
434 SUnit *DefSU = getSUnit(Def);
436 // The reaching Def lives within this scheduling region.
437 // Create a data dependence.
438 SDep dep(DefSU, SDep::Data, Reg);
439 // Adjust the dependence latency using operand def/use information, then
440 // allow the target to perform its own adjustments.
441 int DefOp = Def->findRegisterDefOperandIdx(Reg);
442 dep.setLatency(SchedModel.computeOperandLatency(Def, DefOp, MI, OperIdx));
444 const TargetSubtargetInfo &ST = MF.getSubtarget();
445 ST.adjustSchedDependency(DefSU, SU, const_cast<SDep &>(dep));
450 // Add antidependence to the following def of the vreg it uses.
451 VReg2SUnitMap::iterator DefI = VRegDefs.find(Reg);
452 if (DefI != VRegDefs.end() && DefI->SU != SU)
453 DefI->SU->addPred(SDep(SU, SDep::Anti, Reg));
456 /// Return true if MI is an instruction we are unable to reason about
457 /// (like a call or something with unmodeled side effects).
458 static inline bool isGlobalMemoryObject(AliasAnalysis *AA, MachineInstr *MI) {
459 return MI->isCall() || MI->hasUnmodeledSideEffects() ||
460 (MI->hasOrderedMemoryRef() &&
461 (!MI->mayLoad() || !MI->isInvariantLoad(AA)));
464 // This MI might have either incomplete info, or known to be unsafe
465 // to deal with (i.e. volatile object).
466 static inline bool isUnsafeMemoryObject(MachineInstr *MI,
467 const MachineFrameInfo *MFI,
468 const DataLayout &DL) {
469 if (!MI || MI->memoperands_empty())
471 // We purposefully do no check for hasOneMemOperand() here
472 // in hope to trigger an assert downstream in order to
473 // finish implementation.
474 if ((*MI->memoperands_begin())->isVolatile() ||
475 MI->hasUnmodeledSideEffects())
478 if ((*MI->memoperands_begin())->getPseudoValue()) {
479 // Similarly to getUnderlyingObjectForInstr:
480 // For now, ignore PseudoSourceValues which may alias LLVM IR values
481 // because the code that uses this function has no way to cope with
486 const Value *V = (*MI->memoperands_begin())->getValue();
490 SmallVector<Value *, 4> Objs;
491 getUnderlyingObjects(V, Objs, DL);
492 for (Value *V : Objs) {
493 // Does this pointer refer to a distinct and identifiable object?
494 if (!isIdentifiedObject(V))
501 /// This returns true if the two MIs need a chain edge between them.
502 /// If these are not even memory operations, we still may need
503 /// chain deps between them. The question really is - could
504 /// these two MIs be reordered during scheduling from memory dependency
506 static bool MIsNeedChainEdge(AliasAnalysis *AA, const MachineFrameInfo *MFI,
507 const DataLayout &DL, MachineInstr *MIa,
509 const MachineFunction *MF = MIa->getParent()->getParent();
510 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
512 // Cover a trivial case - no edge is need to itself.
516 // Let the target decide if memory accesses cannot possibly overlap.
517 if ((MIa->mayLoad() || MIa->mayStore()) &&
518 (MIb->mayLoad() || MIb->mayStore()))
519 if (TII->areMemAccessesTriviallyDisjoint(MIa, MIb, AA))
522 // FIXME: Need to handle multiple memory operands to support all targets.
523 if (!MIa->hasOneMemOperand() || !MIb->hasOneMemOperand())
526 if (isUnsafeMemoryObject(MIa, MFI, DL) || isUnsafeMemoryObject(MIb, MFI, DL))
529 // If we are dealing with two "normal" loads, we do not need an edge
530 // between them - they could be reordered.
531 if (!MIa->mayStore() && !MIb->mayStore())
534 // To this point analysis is generic. From here on we do need AA.
538 MachineMemOperand *MMOa = *MIa->memoperands_begin();
539 MachineMemOperand *MMOb = *MIb->memoperands_begin();
541 if (!MMOa->getValue() || !MMOb->getValue())
544 // The following interface to AA is fashioned after DAGCombiner::isAlias
545 // and operates with MachineMemOperand offset with some important
547 // - LLVM fundamentally assumes flat address spaces.
548 // - MachineOperand offset can *only* result from legalization and
549 // cannot affect queries other than the trivial case of overlap
551 // - These offsets never wrap and never step outside
552 // of allocated objects.
553 // - There should never be any negative offsets here.
555 // FIXME: Modify API to hide this math from "user"
556 // FIXME: Even before we go to AA we can reason locally about some
557 // memory objects. It can save compile time, and possibly catch some
558 // corner cases not currently covered.
560 assert ((MMOa->getOffset() >= 0) && "Negative MachineMemOperand offset");
561 assert ((MMOb->getOffset() >= 0) && "Negative MachineMemOperand offset");
563 int64_t MinOffset = std::min(MMOa->getOffset(), MMOb->getOffset());
564 int64_t Overlapa = MMOa->getSize() + MMOa->getOffset() - MinOffset;
565 int64_t Overlapb = MMOb->getSize() + MMOb->getOffset() - MinOffset;
567 AliasResult AAResult =
568 AA->alias(MemoryLocation(MMOa->getValue(), Overlapa,
569 UseTBAA ? MMOa->getAAInfo() : AAMDNodes()),
570 MemoryLocation(MMOb->getValue(), Overlapb,
571 UseTBAA ? MMOb->getAAInfo() : AAMDNodes()));
573 return (AAResult != NoAlias);
576 /// This recursive function iterates over chain deps of SUb looking for
577 /// "latest" node that needs a chain edge to SUa.
578 static unsigned iterateChainSucc(AliasAnalysis *AA, const MachineFrameInfo *MFI,
579 const DataLayout &DL, SUnit *SUa, SUnit *SUb,
580 SUnit *ExitSU, unsigned *Depth,
581 SmallPtrSetImpl<const SUnit *> &Visited) {
582 if (!SUa || !SUb || SUb == ExitSU)
585 // Remember visited nodes.
586 if (!Visited.insert(SUb).second)
588 // If there is _some_ dependency already in place, do not
589 // descend any further.
590 // TODO: Need to make sure that if that dependency got eliminated or ignored
591 // for any reason in the future, we would not violate DAG topology.
592 // Currently it does not happen, but makes an implicit assumption about
593 // future implementation.
595 // Independently, if we encounter node that is some sort of global
596 // object (like a call) we already have full set of dependencies to it
597 // and we can stop descending.
598 if (SUa->isSucc(SUb) ||
599 isGlobalMemoryObject(AA, SUb->getInstr()))
602 // If we do need an edge, or we have exceeded depth budget,
603 // add that edge to the predecessors chain of SUb,
604 // and stop descending.
606 MIsNeedChainEdge(AA, MFI, DL, SUa->getInstr(), SUb->getInstr())) {
607 SUb->addPred(SDep(SUa, SDep::MayAliasMem));
610 // Track current depth.
612 // Iterate over memory dependencies only.
613 for (SUnit::const_succ_iterator I = SUb->Succs.begin(), E = SUb->Succs.end();
615 if (I->isNormalMemoryOrBarrier())
616 iterateChainSucc(AA, MFI, DL, SUa, I->getSUnit(), ExitSU, Depth, Visited);
620 /// This function assumes that "downward" from SU there exist
621 /// tail/leaf of already constructed DAG. It iterates downward and
622 /// checks whether SU can be aliasing any node dominated
624 static void adjustChainDeps(AliasAnalysis *AA, const MachineFrameInfo *MFI,
625 const DataLayout &DL, SUnit *SU, SUnit *ExitSU,
626 std::set<SUnit *> &CheckList,
627 unsigned LatencyToLoad) {
631 SmallPtrSet<const SUnit*, 16> Visited;
634 for (std::set<SUnit *>::iterator I = CheckList.begin(), IE = CheckList.end();
638 if (MIsNeedChainEdge(AA, MFI, DL, SU->getInstr(), (*I)->getInstr())) {
639 SDep Dep(SU, SDep::MayAliasMem);
640 Dep.setLatency(((*I)->getInstr()->mayLoad()) ? LatencyToLoad : 0);
644 // Iterate recursively over all previously added memory chain
645 // successors. Keep track of visited nodes.
646 for (SUnit::const_succ_iterator J = (*I)->Succs.begin(),
647 JE = (*I)->Succs.end(); J != JE; ++J)
648 if (J->isNormalMemoryOrBarrier())
649 iterateChainSucc(AA, MFI, DL, SU, J->getSUnit(), ExitSU, &Depth,
654 /// Check whether two objects need a chain edge, if so, add it
655 /// otherwise remember the rejected SU.
656 static inline void addChainDependency(AliasAnalysis *AA,
657 const MachineFrameInfo *MFI,
658 const DataLayout &DL, SUnit *SUa,
659 SUnit *SUb, std::set<SUnit *> &RejectList,
660 unsigned TrueMemOrderLatency = 0,
661 bool isNormalMemory = false) {
662 // If this is a false dependency,
663 // do not add the edge, but remember the rejected node.
664 if (MIsNeedChainEdge(AA, MFI, DL, SUa->getInstr(), SUb->getInstr())) {
665 SDep Dep(SUa, isNormalMemory ? SDep::MayAliasMem : SDep::Barrier);
666 Dep.setLatency(TrueMemOrderLatency);
670 // Duplicate entries should be ignored.
671 RejectList.insert(SUb);
672 DEBUG(dbgs() << "\tReject chain dep between SU("
673 << SUa->NodeNum << ") and SU("
674 << SUb->NodeNum << ")\n");
678 /// Create an SUnit for each real instruction, numbered in top-down topological
679 /// order. The instruction order A < B, implies that no edge exists from B to A.
681 /// Map each real instruction to its SUnit.
683 /// After initSUnits, the SUnits vector cannot be resized and the scheduler may
684 /// hang onto SUnit pointers. We may relax this in the future by using SUnit IDs
685 /// instead of pointers.
687 /// MachineScheduler relies on initSUnits numbering the nodes by their order in
688 /// the original instruction list.
689 void ScheduleDAGInstrs::initSUnits() {
690 // We'll be allocating one SUnit for each real instruction in the region,
691 // which is contained within a basic block.
692 SUnits.reserve(NumRegionInstrs);
694 for (MachineBasicBlock::iterator I = RegionBegin; I != RegionEnd; ++I) {
695 MachineInstr *MI = I;
696 if (MI->isDebugValue())
699 SUnit *SU = newSUnit(MI);
702 SU->isCall = MI->isCall();
703 SU->isCommutable = MI->isCommutable();
705 // Assign the Latency field of SU using target-provided information.
706 SU->Latency = SchedModel.computeInstrLatency(SU->getInstr());
708 // If this SUnit uses a reserved or unbuffered resource, mark it as such.
710 // Reserved resources block an instruction from issuing and stall the
711 // entire pipeline. These are identified by BufferSize=0.
713 // Unbuffered resources prevent execution of subsequent instructions that
714 // require the same resources. This is used for in-order execution pipelines
715 // within an out-of-order core. These are identified by BufferSize=1.
716 if (SchedModel.hasInstrSchedModel()) {
717 const MCSchedClassDesc *SC = getSchedClass(SU);
718 for (TargetSchedModel::ProcResIter
719 PI = SchedModel.getWriteProcResBegin(SC),
720 PE = SchedModel.getWriteProcResEnd(SC); PI != PE; ++PI) {
721 switch (SchedModel.getProcResource(PI->ProcResourceIdx)->BufferSize) {
723 SU->hasReservedResource = true;
726 SU->isUnbuffered = true;
736 /// If RegPressure is non-null, compute register pressure as a side effect. The
737 /// DAG builder is an efficient place to do it because it already visits
739 void ScheduleDAGInstrs::buildSchedGraph(AliasAnalysis *AA,
740 RegPressureTracker *RPTracker,
741 PressureDiffs *PDiffs) {
742 const TargetSubtargetInfo &ST = MF.getSubtarget();
743 bool UseAA = EnableAASchedMI.getNumOccurrences() > 0 ? EnableAASchedMI
745 AliasAnalysis *AAForDep = UseAA ? AA : nullptr;
748 ScheduleDAG::clearDAG();
750 // Create an SUnit for each real instruction.
754 PDiffs->init(SUnits.size());
756 // We build scheduling units by walking a block's instruction list from bottom
759 // Remember where a generic side-effecting instruction is as we proceed.
760 SUnit *BarrierChain = nullptr, *AliasChain = nullptr;
762 // Memory references to specific known memory locations are tracked
763 // so that they can be given more precise dependencies. We track
764 // separately the known memory locations that may alias and those
765 // that are known not to alias
766 MapVector<ValueType, std::vector<SUnit *> > AliasMemDefs, NonAliasMemDefs;
767 MapVector<ValueType, std::vector<SUnit *> > AliasMemUses, NonAliasMemUses;
768 std::set<SUnit*> RejectMemNodes;
770 // Remove any stale debug info; sometimes BuildSchedGraph is called again
771 // without emitting the info from the previous call.
773 FirstDbgValue = nullptr;
775 assert(Defs.empty() && Uses.empty() &&
776 "Only BuildGraph should update Defs/Uses");
777 Defs.setUniverse(TRI->getNumRegs());
778 Uses.setUniverse(TRI->getNumRegs());
780 assert(VRegDefs.empty() && "Only BuildSchedGraph may access VRegDefs");
782 VRegDefs.setUniverse(MRI.getNumVirtRegs());
783 VRegUses.setUniverse(MRI.getNumVirtRegs());
785 // Model data dependencies between instructions being scheduled and the
787 addSchedBarrierDeps();
789 // Walk the list of instructions, from bottom moving up.
790 MachineInstr *DbgMI = nullptr;
791 for (MachineBasicBlock::iterator MII = RegionEnd, MIE = RegionBegin;
793 MachineInstr *MI = std::prev(MII);
795 DbgValues.push_back(std::make_pair(DbgMI, MI));
799 if (MI->isDebugValue()) {
803 SUnit *SU = MISUnitMap[MI];
804 assert(SU && "No SUnit mapped to this MI");
807 PressureDiff *PDiff = PDiffs ? &(*PDiffs)[SU->NodeNum] : nullptr;
808 RPTracker->recede(/*LiveUses=*/nullptr, PDiff);
809 assert(RPTracker->getPos() == std::prev(MII) &&
810 "RPTracker can't find MI");
814 (CanHandleTerminators || (!MI->isTerminator() && !MI->isPosition())) &&
815 "Cannot schedule terminators or labels!");
817 // Add register-based dependencies (data, anti, and output).
818 bool HasVRegDef = false;
819 for (unsigned j = 0, n = MI->getNumOperands(); j != n; ++j) {
820 const MachineOperand &MO = MI->getOperand(j);
821 if (!MO.isReg()) continue;
822 unsigned Reg = MO.getReg();
823 if (Reg == 0) continue;
825 if (TRI->isPhysicalRegister(Reg))
826 addPhysRegDeps(SU, j);
830 addVRegDefDeps(SU, j);
832 else if (MO.readsReg()) // ignore undef operands
833 addVRegUseDeps(SU, j);
836 // If we haven't seen any uses in this scheduling region, create a
837 // dependence edge to ExitSU to model the live-out latency. This is required
838 // for vreg defs with no in-region use, and prefetches with no vreg def.
840 // FIXME: NumDataSuccs would be more precise than NumSuccs here. This
841 // check currently relies on being called before adding chain deps.
842 if (SU->NumSuccs == 0 && SU->Latency > 1
843 && (HasVRegDef || MI->mayLoad())) {
844 SDep Dep(SU, SDep::Artificial);
845 Dep.setLatency(SU->Latency - 1);
849 // Add chain dependencies.
850 // Chain dependencies used to enforce memory order should have
851 // latency of 0 (except for true dependency of Store followed by
852 // aliased Load... we estimate that with a single cycle of latency
853 // assuming the hardware will bypass)
854 // Note that isStoreToStackSlot and isLoadFromStackSLot are not usable
855 // after stack slots are lowered to actual addresses.
856 // TODO: Use an AliasAnalysis and do real alias-analysis queries, and
857 // produce more precise dependence information.
858 unsigned TrueMemOrderLatency = MI->mayStore() ? 1 : 0;
859 if (isGlobalMemoryObject(AA, MI)) {
860 // Be conservative with these and add dependencies on all memory
861 // references, even those that are known to not alias.
862 for (MapVector<ValueType, std::vector<SUnit *> >::iterator I =
863 NonAliasMemDefs.begin(), E = NonAliasMemDefs.end(); I != E; ++I) {
864 for (unsigned i = 0, e = I->second.size(); i != e; ++i) {
865 I->second[i]->addPred(SDep(SU, SDep::Barrier));
868 for (MapVector<ValueType, std::vector<SUnit *> >::iterator I =
869 NonAliasMemUses.begin(), E = NonAliasMemUses.end(); I != E; ++I) {
870 for (unsigned i = 0, e = I->second.size(); i != e; ++i) {
871 SDep Dep(SU, SDep::Barrier);
872 Dep.setLatency(TrueMemOrderLatency);
873 I->second[i]->addPred(Dep);
876 // Add SU to the barrier chain.
878 BarrierChain->addPred(SDep(SU, SDep::Barrier));
880 // This is a barrier event that acts as a pivotal node in the DAG,
881 // so it is safe to clear list of exposed nodes.
882 adjustChainDeps(AA, MFI, MF.getDataLayout(), SU, &ExitSU, RejectMemNodes,
883 TrueMemOrderLatency);
884 RejectMemNodes.clear();
885 NonAliasMemDefs.clear();
886 NonAliasMemUses.clear();
890 // Chain all possibly aliasing memory references through SU.
892 unsigned ChainLatency = 0;
893 if (AliasChain->getInstr()->mayLoad())
894 ChainLatency = TrueMemOrderLatency;
895 addChainDependency(AAForDep, MFI, MF.getDataLayout(), SU, AliasChain,
896 RejectMemNodes, ChainLatency);
899 for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
900 addChainDependency(AAForDep, MFI, MF.getDataLayout(), SU,
901 PendingLoads[k], RejectMemNodes,
902 TrueMemOrderLatency);
903 for (MapVector<ValueType, std::vector<SUnit *> >::iterator I =
904 AliasMemDefs.begin(), E = AliasMemDefs.end(); I != E; ++I) {
905 for (unsigned i = 0, e = I->second.size(); i != e; ++i)
906 addChainDependency(AAForDep, MFI, MF.getDataLayout(), SU,
907 I->second[i], RejectMemNodes);
909 for (MapVector<ValueType, std::vector<SUnit *> >::iterator I =
910 AliasMemUses.begin(), E = AliasMemUses.end(); I != E; ++I) {
911 for (unsigned i = 0, e = I->second.size(); i != e; ++i)
912 addChainDependency(AAForDep, MFI, MF.getDataLayout(), SU,
913 I->second[i], RejectMemNodes, TrueMemOrderLatency);
915 adjustChainDeps(AA, MFI, MF.getDataLayout(), SU, &ExitSU, RejectMemNodes,
916 TrueMemOrderLatency);
917 PendingLoads.clear();
918 AliasMemDefs.clear();
919 AliasMemUses.clear();
920 } else if (MI->mayStore()) {
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::Barrier));
928 UnderlyingObjectsVector Objs;
929 getUnderlyingObjectsForInstr(MI, MFI, Objs, MF.getDataLayout());
932 // Treat all other stores conservatively.
933 goto new_alias_chain;
936 bool MayAlias = false;
937 for (UnderlyingObjectsVector::iterator K = Objs.begin(), KE = Objs.end();
939 ValueType V = K->getPointer();
940 bool ThisMayAlias = K->getInt();
944 // A store to a specific PseudoSourceValue. Add precise dependencies.
945 // Record the def in MemDefs, first adding a dep if there is
947 MapVector<ValueType, std::vector<SUnit *> >::iterator I =
948 ((ThisMayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V));
949 MapVector<ValueType, std::vector<SUnit *> >::iterator IE =
950 ((ThisMayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
952 for (unsigned i = 0, e = I->second.size(); i != e; ++i)
953 addChainDependency(AAForDep, MFI, MF.getDataLayout(), SU,
954 I->second[i], RejectMemNodes, 0, true);
956 // If we're not using AA, then we only need one store per object.
959 I->second.push_back(SU);
963 AliasMemDefs[V].clear();
964 AliasMemDefs[V].push_back(SU);
967 NonAliasMemDefs[V].clear();
968 NonAliasMemDefs[V].push_back(SU);
971 // Handle the uses in MemUses, if there are any.
972 MapVector<ValueType, std::vector<SUnit *> >::iterator J =
973 ((ThisMayAlias) ? AliasMemUses.find(V) : NonAliasMemUses.find(V));
974 MapVector<ValueType, std::vector<SUnit *> >::iterator JE =
975 ((ThisMayAlias) ? AliasMemUses.end() : NonAliasMemUses.end());
977 for (unsigned i = 0, e = J->second.size(); i != e; ++i)
978 addChainDependency(AAForDep, MFI, MF.getDataLayout(), SU,
979 J->second[i], RejectMemNodes,
980 TrueMemOrderLatency, true);
985 // Add dependencies from all the PendingLoads, i.e. loads
986 // with no underlying object.
987 for (unsigned k = 0, m = PendingLoads.size(); k != m; ++k)
988 addChainDependency(AAForDep, MFI, MF.getDataLayout(), SU,
989 PendingLoads[k], RejectMemNodes,
990 TrueMemOrderLatency);
991 // Add dependence on alias chain, if needed.
993 addChainDependency(AAForDep, MFI, MF.getDataLayout(), SU, AliasChain,
996 adjustChainDeps(AA, MFI, MF.getDataLayout(), SU, &ExitSU, RejectMemNodes,
997 TrueMemOrderLatency);
998 } else if (MI->mayLoad()) {
999 bool MayAlias = true;
1000 if (MI->isInvariantLoad(AA)) {
1001 // Invariant load, no chain dependencies needed!
1003 UnderlyingObjectsVector Objs;
1004 getUnderlyingObjectsForInstr(MI, MFI, Objs, MF.getDataLayout());
1007 // A load with no underlying object. Depend on all
1008 // potentially aliasing stores.
1009 for (MapVector<ValueType, std::vector<SUnit *> >::iterator I =
1010 AliasMemDefs.begin(), E = AliasMemDefs.end(); I != E; ++I)
1011 for (unsigned i = 0, e = I->second.size(); i != e; ++i)
1012 addChainDependency(AAForDep, MFI, MF.getDataLayout(), SU,
1013 I->second[i], RejectMemNodes);
1015 PendingLoads.push_back(SU);
1021 for (UnderlyingObjectsVector::iterator
1022 J = Objs.begin(), JE = Objs.end(); J != JE; ++J) {
1023 ValueType V = J->getPointer();
1024 bool ThisMayAlias = J->getInt();
1029 // A load from a specific PseudoSourceValue. Add precise dependencies.
1030 MapVector<ValueType, std::vector<SUnit *> >::iterator I =
1031 ((ThisMayAlias) ? AliasMemDefs.find(V) : NonAliasMemDefs.find(V));
1032 MapVector<ValueType, std::vector<SUnit *> >::iterator IE =
1033 ((ThisMayAlias) ? AliasMemDefs.end() : NonAliasMemDefs.end());
1035 for (unsigned i = 0, e = I->second.size(); i != e; ++i)
1036 addChainDependency(AAForDep, MFI, MF.getDataLayout(), SU,
1037 I->second[i], RejectMemNodes, 0, true);
1039 AliasMemUses[V].push_back(SU);
1041 NonAliasMemUses[V].push_back(SU);
1044 adjustChainDeps(AA, MFI, MF.getDataLayout(), SU, &ExitSU,
1045 RejectMemNodes, /*Latency=*/0);
1046 // Add dependencies on alias and barrier chains, if needed.
1047 if (MayAlias && AliasChain)
1048 addChainDependency(AAForDep, MFI, MF.getDataLayout(), SU, AliasChain,
1051 BarrierChain->addPred(SDep(SU, SDep::Barrier));
1056 FirstDbgValue = DbgMI;
1061 PendingLoads.clear();
1064 /// \brief Initialize register live-range state for updating kills.
1065 void ScheduleDAGInstrs::startBlockForKills(MachineBasicBlock *BB) {
1066 // Start with no live registers.
1069 // Examine the live-in regs of all successors.
1070 for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(),
1071 SE = BB->succ_end(); SI != SE; ++SI) {
1072 for (const auto &LI : (*SI)->liveins()) {
1073 // Repeat, for reg and all subregs.
1074 for (MCSubRegIterator SubRegs(LI.PhysReg, TRI, /*IncludeSelf=*/true);
1075 SubRegs.isValid(); ++SubRegs)
1076 LiveRegs.set(*SubRegs);
1081 /// \brief If we change a kill flag on the bundle instruction implicit register
1082 /// operands, then we also need to propagate that to any instructions inside
1083 /// the bundle which had the same kill state.
1084 static void toggleBundleKillFlag(MachineInstr *MI, unsigned Reg,
1085 bool NewKillState) {
1086 if (MI->getOpcode() != TargetOpcode::BUNDLE)
1089 // Walk backwards from the last instruction in the bundle to the first.
1090 // Once we set a kill flag on an instruction, we bail out, as otherwise we
1091 // might set it on too many operands. We will clear as many flags as we
1093 MachineBasicBlock::instr_iterator Begin = MI->getIterator();
1094 MachineBasicBlock::instr_iterator End = getBundleEnd(MI);
1095 while (Begin != End) {
1096 for (MachineOperand &MO : (--End)->operands()) {
1097 if (!MO.isReg() || MO.isDef() || Reg != MO.getReg())
1100 // DEBUG_VALUE nodes do not contribute to code generation and should
1101 // always be ignored. Failure to do so may result in trying to modify
1102 // KILL flags on DEBUG_VALUE nodes, which is distressing.
1106 // If the register has the internal flag then it could be killing an
1107 // internal def of the register. In this case, just skip. We only want
1108 // to toggle the flag on operands visible outside the bundle.
1109 if (MO.isInternalRead())
1112 if (MO.isKill() == NewKillState)
1114 MO.setIsKill(NewKillState);
1121 bool ScheduleDAGInstrs::toggleKillFlag(MachineInstr *MI, MachineOperand &MO) {
1122 // Setting kill flag...
1125 toggleBundleKillFlag(MI, MO.getReg(), true);
1129 // If MO itself is live, clear the kill flag...
1130 if (LiveRegs.test(MO.getReg())) {
1131 MO.setIsKill(false);
1132 toggleBundleKillFlag(MI, MO.getReg(), false);
1136 // If any subreg of MO is live, then create an imp-def for that
1137 // subreg and keep MO marked as killed.
1138 MO.setIsKill(false);
1139 toggleBundleKillFlag(MI, MO.getReg(), false);
1140 bool AllDead = true;
1141 const unsigned SuperReg = MO.getReg();
1142 MachineInstrBuilder MIB(MF, MI);
1143 for (MCSubRegIterator SubRegs(SuperReg, TRI); SubRegs.isValid(); ++SubRegs) {
1144 if (LiveRegs.test(*SubRegs)) {
1145 MIB.addReg(*SubRegs, RegState::ImplicitDefine);
1152 toggleBundleKillFlag(MI, MO.getReg(), true);
1157 // FIXME: Reuse the LivePhysRegs utility for this.
1158 void ScheduleDAGInstrs::fixupKills(MachineBasicBlock *MBB) {
1159 DEBUG(dbgs() << "Fixup kills for BB#" << MBB->getNumber() << '\n');
1161 LiveRegs.resize(TRI->getNumRegs());
1162 BitVector killedRegs(TRI->getNumRegs());
1164 startBlockForKills(MBB);
1166 // Examine block from end to start...
1167 unsigned Count = MBB->size();
1168 for (MachineBasicBlock::iterator I = MBB->end(), E = MBB->begin();
1170 MachineInstr *MI = --I;
1171 if (MI->isDebugValue())
1174 // Update liveness. Registers that are defed but not used in this
1175 // instruction are now dead. Mark register and all subregs as they
1176 // are completely defined.
1177 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1178 MachineOperand &MO = MI->getOperand(i);
1180 LiveRegs.clearBitsNotInMask(MO.getRegMask());
1181 if (!MO.isReg()) continue;
1182 unsigned Reg = MO.getReg();
1183 if (Reg == 0) continue;
1184 if (!MO.isDef()) continue;
1185 // Ignore two-addr defs.
1186 if (MI->isRegTiedToUseOperand(i)) continue;
1188 // Repeat for reg and all subregs.
1189 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
1190 SubRegs.isValid(); ++SubRegs)
1191 LiveRegs.reset(*SubRegs);
1194 // Examine all used registers and set/clear kill flag. When a
1195 // register is used multiple times we only set the kill flag on
1196 // the first use. Don't set kill flags on undef operands.
1198 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1199 MachineOperand &MO = MI->getOperand(i);
1200 if (!MO.isReg() || !MO.isUse() || MO.isUndef()) continue;
1201 unsigned Reg = MO.getReg();
1202 if ((Reg == 0) || MRI.isReserved(Reg)) continue;
1205 if (!killedRegs.test(Reg)) {
1207 // A register is not killed if any subregs are live...
1208 for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) {
1209 if (LiveRegs.test(*SubRegs)) {
1215 // If subreg is not live, then register is killed if it became
1216 // live in this instruction
1218 kill = !LiveRegs.test(Reg);
1221 if (MO.isKill() != kill) {
1222 DEBUG(dbgs() << "Fixing " << MO << " in ");
1223 // Warning: toggleKillFlag may invalidate MO.
1224 toggleKillFlag(MI, MO);
1226 DEBUG(if (MI->getOpcode() == TargetOpcode::BUNDLE) {
1227 MachineBasicBlock::instr_iterator Begin = MI->getIterator();
1228 MachineBasicBlock::instr_iterator End = getBundleEnd(MI);
1229 while (++Begin != End)
1230 DEBUG(Begin->dump());
1234 killedRegs.set(Reg);
1237 // Mark any used register (that is not using undef) and subregs as
1239 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
1240 MachineOperand &MO = MI->getOperand(i);
1241 if (!MO.isReg() || !MO.isUse() || MO.isUndef()) continue;
1242 unsigned Reg = MO.getReg();
1243 if ((Reg == 0) || MRI.isReserved(Reg)) continue;
1245 for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true);
1246 SubRegs.isValid(); ++SubRegs)
1247 LiveRegs.set(*SubRegs);
1252 void ScheduleDAGInstrs::dumpNode(const SUnit *SU) const {
1253 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
1254 SU->getInstr()->dump();
1258 std::string ScheduleDAGInstrs::getGraphNodeLabel(const SUnit *SU) const {
1260 raw_string_ostream oss(s);
1263 else if (SU == &ExitSU)
1266 SU->getInstr()->print(oss, /*SkipOpers=*/true);
1270 /// Return the basic block label. It is not necessarilly unique because a block
1271 /// contains multiple scheduling regions. But it is fine for visualization.
1272 std::string ScheduleDAGInstrs::getDAGName() const {
1273 return "dag." + BB->getFullName();
1276 //===----------------------------------------------------------------------===//
1277 // SchedDFSResult Implementation
1278 //===----------------------------------------------------------------------===//
1281 /// \brief Internal state used to compute SchedDFSResult.
1282 class SchedDFSImpl {
1285 /// Join DAG nodes into equivalence classes by their subtree.
1286 IntEqClasses SubtreeClasses;
1287 /// List PredSU, SuccSU pairs that represent data edges between subtrees.
1288 std::vector<std::pair<const SUnit*, const SUnit*> > ConnectionPairs;
1292 unsigned ParentNodeID; // Parent node (member of the parent subtree).
1293 unsigned SubInstrCount; // Instr count in this tree only, not children.
1295 RootData(unsigned id): NodeID(id),
1296 ParentNodeID(SchedDFSResult::InvalidSubtreeID),
1299 unsigned getSparseSetIndex() const { return NodeID; }
1302 SparseSet<RootData> RootSet;
1305 SchedDFSImpl(SchedDFSResult &r): R(r), SubtreeClasses(R.DFSNodeData.size()) {
1306 RootSet.setUniverse(R.DFSNodeData.size());
1309 /// Return true if this node been visited by the DFS traversal.
1311 /// During visitPostorderNode the Node's SubtreeID is assigned to the Node
1312 /// ID. Later, SubtreeID is updated but remains valid.
1313 bool isVisited(const SUnit *SU) const {
1314 return R.DFSNodeData[SU->NodeNum].SubtreeID
1315 != SchedDFSResult::InvalidSubtreeID;
1318 /// Initialize this node's instruction count. We don't need to flag the node
1319 /// visited until visitPostorder because the DAG cannot have cycles.
1320 void visitPreorder(const SUnit *SU) {
1321 R.DFSNodeData[SU->NodeNum].InstrCount =
1322 SU->getInstr()->isTransient() ? 0 : 1;
1325 /// Called once for each node after all predecessors are visited. Revisit this
1326 /// node's predecessors and potentially join them now that we know the ILP of
1327 /// the other predecessors.
1328 void visitPostorderNode(const SUnit *SU) {
1329 // Mark this node as the root of a subtree. It may be joined with its
1330 // successors later.
1331 R.DFSNodeData[SU->NodeNum].SubtreeID = SU->NodeNum;
1332 RootData RData(SU->NodeNum);
1333 RData.SubInstrCount = SU->getInstr()->isTransient() ? 0 : 1;
1335 // If any predecessors are still in their own subtree, they either cannot be
1336 // joined or are large enough to remain separate. If this parent node's
1337 // total instruction count is not greater than a child subtree by at least
1338 // the subtree limit, then try to join it now since splitting subtrees is
1339 // only useful if multiple high-pressure paths are possible.
1340 unsigned InstrCount = R.DFSNodeData[SU->NodeNum].InstrCount;
1341 for (SUnit::const_pred_iterator
1342 PI = SU->Preds.begin(), PE = SU->Preds.end(); PI != PE; ++PI) {
1343 if (PI->getKind() != SDep::Data)
1345 unsigned PredNum = PI->getSUnit()->NodeNum;
1346 if ((InstrCount - R.DFSNodeData[PredNum].InstrCount) < R.SubtreeLimit)
1347 joinPredSubtree(*PI, SU, /*CheckLimit=*/false);
1349 // Either link or merge the TreeData entry from the child to the parent.
1350 if (R.DFSNodeData[PredNum].SubtreeID == PredNum) {
1351 // If the predecessor's parent is invalid, this is a tree edge and the
1352 // current node is the parent.
1353 if (RootSet[PredNum].ParentNodeID == SchedDFSResult::InvalidSubtreeID)
1354 RootSet[PredNum].ParentNodeID = SU->NodeNum;
1356 else if (RootSet.count(PredNum)) {
1357 // The predecessor is not a root, but is still in the root set. This
1358 // must be the new parent that it was just joined to. Note that
1359 // RootSet[PredNum].ParentNodeID may either be invalid or may still be
1360 // set to the original parent.
1361 RData.SubInstrCount += RootSet[PredNum].SubInstrCount;
1362 RootSet.erase(PredNum);
1365 RootSet[SU->NodeNum] = RData;
1368 /// Called once for each tree edge after calling visitPostOrderNode on the
1369 /// predecessor. Increment the parent node's instruction count and
1370 /// preemptively join this subtree to its parent's if it is small enough.
1371 void visitPostorderEdge(const SDep &PredDep, const SUnit *Succ) {
1372 R.DFSNodeData[Succ->NodeNum].InstrCount
1373 += R.DFSNodeData[PredDep.getSUnit()->NodeNum].InstrCount;
1374 joinPredSubtree(PredDep, Succ);
1377 /// Add a connection for cross edges.
1378 void visitCrossEdge(const SDep &PredDep, const SUnit *Succ) {
1379 ConnectionPairs.push_back(std::make_pair(PredDep.getSUnit(), Succ));
1382 /// Set each node's subtree ID to the representative ID and record connections
1385 SubtreeClasses.compress();
1386 R.DFSTreeData.resize(SubtreeClasses.getNumClasses());
1387 assert(SubtreeClasses.getNumClasses() == RootSet.size()
1388 && "number of roots should match trees");
1389 for (SparseSet<RootData>::const_iterator
1390 RI = RootSet.begin(), RE = RootSet.end(); RI != RE; ++RI) {
1391 unsigned TreeID = SubtreeClasses[RI->NodeID];
1392 if (RI->ParentNodeID != SchedDFSResult::InvalidSubtreeID)
1393 R.DFSTreeData[TreeID].ParentTreeID = SubtreeClasses[RI->ParentNodeID];
1394 R.DFSTreeData[TreeID].SubInstrCount = RI->SubInstrCount;
1395 // Note that SubInstrCount may be greater than InstrCount if we joined
1396 // subtrees across a cross edge. InstrCount will be attributed to the
1397 // original parent, while SubInstrCount will be attributed to the joined
1400 R.SubtreeConnections.resize(SubtreeClasses.getNumClasses());
1401 R.SubtreeConnectLevels.resize(SubtreeClasses.getNumClasses());
1402 DEBUG(dbgs() << R.getNumSubtrees() << " subtrees:\n");
1403 for (unsigned Idx = 0, End = R.DFSNodeData.size(); Idx != End; ++Idx) {
1404 R.DFSNodeData[Idx].SubtreeID = SubtreeClasses[Idx];
1405 DEBUG(dbgs() << " SU(" << Idx << ") in tree "
1406 << R.DFSNodeData[Idx].SubtreeID << '\n');
1408 for (std::vector<std::pair<const SUnit*, const SUnit*> >::const_iterator
1409 I = ConnectionPairs.begin(), E = ConnectionPairs.end();
1411 unsigned PredTree = SubtreeClasses[I->first->NodeNum];
1412 unsigned SuccTree = SubtreeClasses[I->second->NodeNum];
1413 if (PredTree == SuccTree)
1415 unsigned Depth = I->first->getDepth();
1416 addConnection(PredTree, SuccTree, Depth);
1417 addConnection(SuccTree, PredTree, Depth);
1422 /// Join the predecessor subtree with the successor that is its DFS
1423 /// parent. Apply some heuristics before joining.
1424 bool joinPredSubtree(const SDep &PredDep, const SUnit *Succ,
1425 bool CheckLimit = true) {
1426 assert(PredDep.getKind() == SDep::Data && "Subtrees are for data edges");
1428 // Check if the predecessor is already joined.
1429 const SUnit *PredSU = PredDep.getSUnit();
1430 unsigned PredNum = PredSU->NodeNum;
1431 if (R.DFSNodeData[PredNum].SubtreeID != PredNum)
1434 // Four is the magic number of successors before a node is considered a
1436 unsigned NumDataSucs = 0;
1437 for (SUnit::const_succ_iterator SI = PredSU->Succs.begin(),
1438 SE = PredSU->Succs.end(); SI != SE; ++SI) {
1439 if (SI->getKind() == SDep::Data) {
1440 if (++NumDataSucs >= 4)
1444 if (CheckLimit && R.DFSNodeData[PredNum].InstrCount > R.SubtreeLimit)
1446 R.DFSNodeData[PredNum].SubtreeID = Succ->NodeNum;
1447 SubtreeClasses.join(Succ->NodeNum, PredNum);
1451 /// Called by finalize() to record a connection between trees.
1452 void addConnection(unsigned FromTree, unsigned ToTree, unsigned Depth) {
1457 SmallVectorImpl<SchedDFSResult::Connection> &Connections =
1458 R.SubtreeConnections[FromTree];
1459 for (SmallVectorImpl<SchedDFSResult::Connection>::iterator
1460 I = Connections.begin(), E = Connections.end(); I != E; ++I) {
1461 if (I->TreeID == ToTree) {
1462 I->Level = std::max(I->Level, Depth);
1466 Connections.push_back(SchedDFSResult::Connection(ToTree, Depth));
1467 FromTree = R.DFSTreeData[FromTree].ParentTreeID;
1468 } while (FromTree != SchedDFSResult::InvalidSubtreeID);
1474 /// \brief Manage the stack used by a reverse depth-first search over the DAG.
1475 class SchedDAGReverseDFS {
1476 std::vector<std::pair<const SUnit*, SUnit::const_pred_iterator> > DFSStack;
1478 bool isComplete() const { return DFSStack.empty(); }
1480 void follow(const SUnit *SU) {
1481 DFSStack.push_back(std::make_pair(SU, SU->Preds.begin()));
1483 void advance() { ++DFSStack.back().second; }
1485 const SDep *backtrack() {
1486 DFSStack.pop_back();
1487 return DFSStack.empty() ? nullptr : std::prev(DFSStack.back().second);
1490 const SUnit *getCurr() const { return DFSStack.back().first; }
1492 SUnit::const_pred_iterator getPred() const { return DFSStack.back().second; }
1494 SUnit::const_pred_iterator getPredEnd() const {
1495 return getCurr()->Preds.end();
1500 static bool hasDataSucc(const SUnit *SU) {
1501 for (SUnit::const_succ_iterator
1502 SI = SU->Succs.begin(), SE = SU->Succs.end(); SI != SE; ++SI) {
1503 if (SI->getKind() == SDep::Data && !SI->getSUnit()->isBoundaryNode())
1509 /// Compute an ILP metric for all nodes in the subDAG reachable via depth-first
1510 /// search from this root.
1511 void SchedDFSResult::compute(ArrayRef<SUnit> SUnits) {
1513 llvm_unreachable("Top-down ILP metric is unimplemnted");
1515 SchedDFSImpl Impl(*this);
1516 for (ArrayRef<SUnit>::const_iterator
1517 SI = SUnits.begin(), SE = SUnits.end(); SI != SE; ++SI) {
1518 const SUnit *SU = &*SI;
1519 if (Impl.isVisited(SU) || hasDataSucc(SU))
1522 SchedDAGReverseDFS DFS;
1523 Impl.visitPreorder(SU);
1526 // Traverse the leftmost path as far as possible.
1527 while (DFS.getPred() != DFS.getPredEnd()) {
1528 const SDep &PredDep = *DFS.getPred();
1530 // Ignore non-data edges.
1531 if (PredDep.getKind() != SDep::Data
1532 || PredDep.getSUnit()->isBoundaryNode()) {
1535 // An already visited edge is a cross edge, assuming an acyclic DAG.
1536 if (Impl.isVisited(PredDep.getSUnit())) {
1537 Impl.visitCrossEdge(PredDep, DFS.getCurr());
1540 Impl.visitPreorder(PredDep.getSUnit());
1541 DFS.follow(PredDep.getSUnit());
1543 // Visit the top of the stack in postorder and backtrack.
1544 const SUnit *Child = DFS.getCurr();
1545 const SDep *PredDep = DFS.backtrack();
1546 Impl.visitPostorderNode(Child);
1548 Impl.visitPostorderEdge(*PredDep, DFS.getCurr());
1549 if (DFS.isComplete())
1556 /// The root of the given SubtreeID was just scheduled. For all subtrees
1557 /// connected to this tree, record the depth of the connection so that the
1558 /// nearest connected subtrees can be prioritized.
1559 void SchedDFSResult::scheduleTree(unsigned SubtreeID) {
1560 for (SmallVectorImpl<Connection>::const_iterator
1561 I = SubtreeConnections[SubtreeID].begin(),
1562 E = SubtreeConnections[SubtreeID].end(); I != E; ++I) {
1563 SubtreeConnectLevels[I->TreeID] =
1564 std::max(SubtreeConnectLevels[I->TreeID], I->Level);
1565 DEBUG(dbgs() << " Tree: " << I->TreeID
1566 << " @" << SubtreeConnectLevels[I->TreeID] << '\n');
1571 void ILPValue::print(raw_ostream &OS) const {
1572 OS << InstrCount << " / " << Length << " = ";
1576 OS << format("%g", ((double)InstrCount / Length));
1580 void ILPValue::dump() const {
1581 dbgs() << *this << '\n';
1587 raw_ostream &operator<<(raw_ostream &OS, const ILPValue &Val) {