diff options
Diffstat (limited to 'lib/CodeGen/MachineScheduler.cpp')
-rw-r--r-- | lib/CodeGen/MachineScheduler.cpp | 707 |
1 files changed, 531 insertions, 176 deletions
diff --git a/lib/CodeGen/MachineScheduler.cpp b/lib/CodeGen/MachineScheduler.cpp index a6c5a9f..e71c4df 100644 --- a/lib/CodeGen/MachineScheduler.cpp +++ b/lib/CodeGen/MachineScheduler.cpp @@ -53,6 +53,12 @@ static cl::opt<unsigned> MISchedCutoff("misched-cutoff", cl::Hidden, static bool ViewMISchedDAGs = false; #endif // NDEBUG +static cl::opt<bool> EnableRegPressure("misched-regpressure", cl::Hidden, + cl::desc("Enable register pressure scheduling."), cl::init(true)); + +static cl::opt<bool> EnableCyclicPath("misched-cyclicpath", cl::Hidden, + cl::desc("Enable cyclic critical path analysis."), cl::init(true)); + static cl::opt<bool> EnableLoadCluster("misched-cluster", cl::Hidden, cl::desc("Enable load clustering."), cl::init(true)); @@ -66,6 +72,10 @@ static cl::opt<bool> VerifyScheduling("verify-misched", cl::Hidden, // DAG subtrees must have at least this many nodes. static const unsigned MinSubtreeSize = 8; +// Pin the vtables to this file. +void MachineSchedStrategy::anchor() {} +void ScheduleDAGMutation::anchor() {} + //===----------------------------------------------------------------------===// // Machine Instruction Scheduling Pass and Registry //===----------------------------------------------------------------------===// @@ -95,6 +105,9 @@ public: virtual void print(raw_ostream &O, const Module* = 0) const; static char ID; // Class identification, replacement for typeinfo + +protected: + ScheduleDAGInstrs *createMachineScheduler(); }; } // namespace @@ -149,12 +162,13 @@ DefaultSchedRegistry("default", "Use the target's default scheduler choice.", /// Forward declare the standard machine scheduler. This will be used as the /// default scheduler if the target does not set a default. -static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C); +static ScheduleDAGInstrs *createGenericSched(MachineSchedContext *C); /// Decrement this iterator until reaching the top or a non-debug instr. -static MachineBasicBlock::iterator -priorNonDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator Beg) { +static MachineBasicBlock::const_iterator +priorNonDebug(MachineBasicBlock::const_iterator I, + MachineBasicBlock::const_iterator Beg) { assert(I != Beg && "reached the top of the region, cannot decrement"); while (--I != Beg) { if (!I->isDebugValue()) @@ -163,10 +177,19 @@ priorNonDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator Beg) { return I; } +/// Non-const version. +static MachineBasicBlock::iterator +priorNonDebug(MachineBasicBlock::iterator I, + MachineBasicBlock::const_iterator Beg) { + return const_cast<MachineInstr*>( + &*priorNonDebug(MachineBasicBlock::const_iterator(I), Beg)); +} + /// If this iterator is a debug value, increment until reaching the End or a /// non-debug instruction. -static MachineBasicBlock::iterator -nextIfDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator End) { +static MachineBasicBlock::const_iterator +nextIfDebug(MachineBasicBlock::const_iterator I, + MachineBasicBlock::const_iterator End) { for(; I != End; ++I) { if (!I->isDebugValue()) break; @@ -174,6 +197,34 @@ nextIfDebug(MachineBasicBlock::iterator I, MachineBasicBlock::iterator End) { return I; } +/// Non-const version. +static MachineBasicBlock::iterator +nextIfDebug(MachineBasicBlock::iterator I, + MachineBasicBlock::const_iterator End) { + // Cast the return value to nonconst MachineInstr, then cast to an + // instr_iterator, which does not check for null, finally return a + // bundle_iterator. + return MachineBasicBlock::instr_iterator( + const_cast<MachineInstr*>( + &*nextIfDebug(MachineBasicBlock::const_iterator(I), End))); +} + +/// Instantiate a ScheduleDAGInstrs that will be owned by the caller. +ScheduleDAGInstrs *MachineScheduler::createMachineScheduler() { + // Select the scheduler, or set the default. + MachineSchedRegistry::ScheduleDAGCtor Ctor = MachineSchedOpt; + if (Ctor != useDefaultMachineSched) + return Ctor(this); + + // Get the default scheduler set by the target for this function. + ScheduleDAGInstrs *Scheduler = PassConfig->createMachineScheduler(this); + if (Scheduler) + return Scheduler; + + // Default to GenericScheduler. + return createGenericSched(this); +} + /// Top-level MachineScheduler pass driver. /// /// Visit blocks in function order. Divide each block into scheduling regions @@ -209,18 +260,9 @@ bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) { } RegClassInfo->runOnMachineFunction(*MF); - // Select the scheduler, or set the default. - MachineSchedRegistry::ScheduleDAGCtor Ctor = MachineSchedOpt; - if (Ctor == useDefaultMachineSched) { - // Get the default scheduler set by the target. - Ctor = MachineSchedRegistry::getDefault(); - if (!Ctor) { - Ctor = createConvergingSched; - MachineSchedRegistry::setDefault(Ctor); - } - } - // Instantiate the selected scheduler. - OwningPtr<ScheduleDAGInstrs> Scheduler(Ctor(this)); + // Instantiate the selected scheduler for this target, function, and + // optimization level. + OwningPtr<ScheduleDAGInstrs> Scheduler(createMachineScheduler()); // Visit all machine basic blocks. // @@ -255,14 +297,15 @@ bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) { // The next region starts above the previous region. Look backward in the // instruction stream until we find the nearest boundary. + unsigned NumRegionInstrs = 0; MachineBasicBlock::iterator I = RegionEnd; - for(;I != MBB->begin(); --I, --RemainingInstrs) { + for(;I != MBB->begin(); --I, --RemainingInstrs, ++NumRegionInstrs) { if (TII->isSchedulingBoundary(llvm::prior(I), MBB, *MF)) break; } // Notify the scheduler of the region, even if we may skip scheduling // it. Perhaps it still needs to be bundled. - Scheduler->enterRegion(MBB, I, RegionEnd, RemainingInstrs); + Scheduler->enterRegion(MBB, I, RegionEnd, NumRegionInstrs); // Skip empty scheduling regions (0 or 1 schedulable instructions). if (I == RegionEnd || I == llvm::prior(RegionEnd)) { @@ -277,7 +320,8 @@ bool MachineScheduler::runOnMachineFunction(MachineFunction &mf) { << "\n From: " << *I << " To: "; if (RegionEnd != MBB->end()) dbgs() << *RegionEnd; else dbgs() << "End"; - dbgs() << " Remaining: " << RemainingInstrs << "\n"); + dbgs() << " RegionInstrs: " << NumRegionInstrs + << " Remaining: " << RemainingInstrs << "\n"); // Schedule a region: possibly reorder instructions. // This invalidates 'RegionEnd' and 'I'. @@ -446,13 +490,19 @@ bool ScheduleDAGMI::checkSchedLimit() { void ScheduleDAGMI::enterRegion(MachineBasicBlock *bb, MachineBasicBlock::iterator begin, MachineBasicBlock::iterator end, - unsigned endcount) + unsigned regioninstrs) { - ScheduleDAGInstrs::enterRegion(bb, begin, end, endcount); + ScheduleDAGInstrs::enterRegion(bb, begin, end, regioninstrs); // For convenience remember the end of the liveness region. LiveRegionEnd = (RegionEnd == bb->end()) ? RegionEnd : llvm::next(RegionEnd); + + SUPressureDiffs.clear(); + + SchedImpl->initPolicy(begin, end, regioninstrs); + + ShouldTrackPressure = SchedImpl->shouldTrackPressure(); } // Setup the register pressure trackers for the top scheduled top and bottom @@ -483,9 +533,16 @@ void ScheduleDAGMI::initRegPressure() { dumpRegSetPressure(BotRPTracker.getLiveThru(), TRI)); }; + // For each live out vreg reduce the pressure change associated with other + // uses of the same vreg below the live-out reaching def. + updatePressureDiffs(RPTracker.getPressure().LiveOutRegs); + // Account for liveness generated by the region boundary. - if (LiveRegionEnd != RegionEnd) - BotRPTracker.recede(); + if (LiveRegionEnd != RegionEnd) { + SmallVector<unsigned, 8> LiveUses; + BotRPTracker.recede(&LiveUses); + updatePressureDiffs(LiveUses); + } assert(BotRPTracker.getPos() == RegionEnd && "Can't find the region bottom"); @@ -500,34 +557,83 @@ void ScheduleDAGMI::initRegPressure() { DEBUG(dbgs() << TRI->getRegPressureSetName(i) << " Limit " << Limit << " Actual " << RegionPressure[i] << "\n"); - RegionCriticalPSets.push_back(PressureElement(i, 0)); + RegionCriticalPSets.push_back(PressureChange(i)); } } DEBUG(dbgs() << "Excess PSets: "; for (unsigned i = 0, e = RegionCriticalPSets.size(); i != e; ++i) dbgs() << TRI->getRegPressureSetName( - RegionCriticalPSets[i].PSetID) << " "; + RegionCriticalPSets[i].getPSet()) << " "; dbgs() << "\n"); } -// FIXME: When the pressure tracker deals in pressure differences then we won't -// iterate over all RegionCriticalPSets[i]. void ScheduleDAGMI:: -updateScheduledPressure(const std::vector<unsigned> &NewMaxPressure) { - for (unsigned i = 0, e = RegionCriticalPSets.size(); i < e; ++i) { - unsigned ID = RegionCriticalPSets[i].PSetID; - int &MaxUnits = RegionCriticalPSets[i].UnitIncrease; - if ((int)NewMaxPressure[ID] > MaxUnits) - MaxUnits = NewMaxPressure[ID]; +updateScheduledPressure(const SUnit *SU, + const std::vector<unsigned> &NewMaxPressure) { + const PressureDiff &PDiff = getPressureDiff(SU); + unsigned CritIdx = 0, CritEnd = RegionCriticalPSets.size(); + for (PressureDiff::const_iterator I = PDiff.begin(), E = PDiff.end(); + I != E; ++I) { + if (!I->isValid()) + break; + unsigned ID = I->getPSet(); + while (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() < ID) + ++CritIdx; + if (CritIdx != CritEnd && RegionCriticalPSets[CritIdx].getPSet() == ID) { + if ((int)NewMaxPressure[ID] > RegionCriticalPSets[CritIdx].getUnitInc() + && NewMaxPressure[ID] <= INT16_MAX) + RegionCriticalPSets[CritIdx].setUnitInc(NewMaxPressure[ID]); + } + unsigned Limit = RegClassInfo->getRegPressureSetLimit(ID); + if (NewMaxPressure[ID] >= Limit - 2) { + DEBUG(dbgs() << " " << TRI->getRegPressureSetName(ID) << ": " + << NewMaxPressure[ID] << " > " << Limit << "(+ " + << BotRPTracker.getLiveThru()[ID] << " livethru)\n"); + } } - DEBUG( - for (unsigned i = 0, e = NewMaxPressure.size(); i < e; ++i) { - unsigned Limit = RegClassInfo->getRegPressureSetLimit(i); - if (NewMaxPressure[i] > Limit ) { - dbgs() << " " << TRI->getRegPressureSetName(i) << ": " - << NewMaxPressure[i] << " > " << Limit << "\n"; +} + +/// Update the PressureDiff array for liveness after scheduling this +/// instruction. +void ScheduleDAGMI::updatePressureDiffs(ArrayRef<unsigned> LiveUses) { + for (unsigned LUIdx = 0, LUEnd = LiveUses.size(); LUIdx != LUEnd; ++LUIdx) { + /// FIXME: Currently assuming single-use physregs. + unsigned Reg = LiveUses[LUIdx]; + DEBUG(dbgs() << " LiveReg: " << PrintVRegOrUnit(Reg, TRI) << "\n"); + if (!TRI->isVirtualRegister(Reg)) + continue; + + // This may be called before CurrentBottom has been initialized. However, + // BotRPTracker must have a valid position. We want the value live into the + // instruction or live out of the block, so ask for the previous + // instruction's live-out. + const LiveInterval &LI = LIS->getInterval(Reg); + VNInfo *VNI; + MachineBasicBlock::const_iterator I = + nextIfDebug(BotRPTracker.getPos(), BB->end()); + if (I == BB->end()) + VNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB)); + else { + LiveQueryResult LRQ = LI.Query(LIS->getInstructionIndex(I)); + VNI = LRQ.valueIn(); + } + // RegisterPressureTracker guarantees that readsReg is true for LiveUses. + assert(VNI && "No live value at use."); + for (VReg2UseMap::iterator + UI = VRegUses.find(Reg); UI != VRegUses.end(); ++UI) { + SUnit *SU = UI->SU; + DEBUG(dbgs() << " UpdateRegP: SU(" << SU->NodeNum << ") " + << *SU->getInstr()); + // If this use comes before the reaching def, it cannot be a last use, so + // descrease its pressure change. + if (!SU->isScheduled && SU != &ExitSU) { + LiveQueryResult LRQ + = LI.Query(LIS->getInstructionIndex(SU->getInstr())); + if (LRQ.valueIn() == VNI) + getPressureDiff(SU).addPressureChange(Reg, true, &MRI); } - }); + } + } } /// schedule - Called back from MachineScheduler::runOnMachineFunction @@ -585,6 +691,13 @@ void ScheduleDAGMI::schedule() { /// Build the DAG and setup three register pressure trackers. void ScheduleDAGMI::buildDAGWithRegPressure() { + if (!ShouldTrackPressure) { + RPTracker.reset(); + RegionCriticalPSets.clear(); + buildSchedGraph(AA); + return; + } + // Initialize the register pressure tracker used by buildSchedGraph. RPTracker.init(&MF, RegClassInfo, LIS, BB, LiveRegionEnd, /*TrackUntiedDefs=*/true); @@ -594,7 +707,7 @@ void ScheduleDAGMI::buildDAGWithRegPressure() { RPTracker.recede(); // Build the DAG, and compute current register pressure. - buildSchedGraph(AA, &RPTracker); + buildSchedGraph(AA, &RPTracker, &SUPressureDiffs); // Initialize top/bottom trackers after computing region pressure. initRegPressure(); @@ -637,6 +750,91 @@ void ScheduleDAGMI::findRootsAndBiasEdges(SmallVectorImpl<SUnit*> &TopRoots, ExitSU.biasCriticalPath(); } +/// Compute the max cyclic critical path through the DAG. The scheduling DAG +/// only provides the critical path for single block loops. To handle loops that +/// span blocks, we could use the vreg path latencies provided by +/// MachineTraceMetrics instead. However, MachineTraceMetrics is not currently +/// available for use in the scheduler. +/// +/// The cyclic path estimation identifies a def-use pair that crosses the back +/// edge and considers the depth and height of the nodes. For example, consider +/// the following instruction sequence where each instruction has unit latency +/// and defines an epomymous virtual register: +/// +/// a->b(a,c)->c(b)->d(c)->exit +/// +/// The cyclic critical path is a two cycles: b->c->b +/// The acyclic critical path is four cycles: a->b->c->d->exit +/// LiveOutHeight = height(c) = len(c->d->exit) = 2 +/// LiveOutDepth = depth(c) + 1 = len(a->b->c) + 1 = 3 +/// LiveInHeight = height(b) + 1 = len(b->c->d->exit) + 1 = 4 +/// LiveInDepth = depth(b) = len(a->b) = 1 +/// +/// LiveOutDepth - LiveInDepth = 3 - 1 = 2 +/// LiveInHeight - LiveOutHeight = 4 - 2 = 2 +/// CyclicCriticalPath = min(2, 2) = 2 +unsigned ScheduleDAGMI::computeCyclicCriticalPath() { + // This only applies to single block loop. + if (!BB->isSuccessor(BB)) + return 0; + + unsigned MaxCyclicLatency = 0; + // Visit each live out vreg def to find def/use pairs that cross iterations. + ArrayRef<unsigned> LiveOuts = RPTracker.getPressure().LiveOutRegs; + for (ArrayRef<unsigned>::iterator RI = LiveOuts.begin(), RE = LiveOuts.end(); + RI != RE; ++RI) { + unsigned Reg = *RI; + if (!TRI->isVirtualRegister(Reg)) + continue; + const LiveInterval &LI = LIS->getInterval(Reg); + const VNInfo *DefVNI = LI.getVNInfoBefore(LIS->getMBBEndIdx(BB)); + if (!DefVNI) + continue; + + MachineInstr *DefMI = LIS->getInstructionFromIndex(DefVNI->def); + const SUnit *DefSU = getSUnit(DefMI); + if (!DefSU) + continue; + + unsigned LiveOutHeight = DefSU->getHeight(); + unsigned LiveOutDepth = DefSU->getDepth() + DefSU->Latency; + // Visit all local users of the vreg def. + for (VReg2UseMap::iterator + UI = VRegUses.find(Reg); UI != VRegUses.end(); ++UI) { + if (UI->SU == &ExitSU) + continue; + + // Only consider uses of the phi. + LiveQueryResult LRQ = + LI.Query(LIS->getInstructionIndex(UI->SU->getInstr())); + if (!LRQ.valueIn()->isPHIDef()) + continue; + + // Assume that a path spanning two iterations is a cycle, which could + // overestimate in strange cases. This allows cyclic latency to be + // estimated as the minimum slack of the vreg's depth or height. + unsigned CyclicLatency = 0; + if (LiveOutDepth > UI->SU->getDepth()) + CyclicLatency = LiveOutDepth - UI->SU->getDepth(); + + unsigned LiveInHeight = UI->SU->getHeight() + DefSU->Latency; + if (LiveInHeight > LiveOutHeight) { + if (LiveInHeight - LiveOutHeight < CyclicLatency) + CyclicLatency = LiveInHeight - LiveOutHeight; + } + else + CyclicLatency = 0; + + DEBUG(dbgs() << "Cyclic Path: SU(" << DefSU->NodeNum << ") -> SU(" + << UI->SU->NodeNum << ") = " << CyclicLatency << "c\n"); + if (CyclicLatency > MaxCyclicLatency) + MaxCyclicLatency = CyclicLatency; + } + } + DEBUG(dbgs() << "Cyclic Critical Path: " << MaxCyclicLatency << "c\n"); + return MaxCyclicLatency; +} + /// Identify DAG roots and setup scheduler queues. void ScheduleDAGMI::initQueues(ArrayRef<SUnit*> TopRoots, ArrayRef<SUnit*> BotRoots) { @@ -664,11 +862,13 @@ void ScheduleDAGMI::initQueues(ArrayRef<SUnit*> TopRoots, SchedImpl->registerRoots(); // Advance past initial DebugValues. - assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker"); CurrentTop = nextIfDebug(RegionBegin, RegionEnd); - TopRPTracker.setPos(CurrentTop); - CurrentBottom = RegionEnd; + + if (ShouldTrackPressure) { + assert(TopRPTracker.getPos() == RegionBegin && "bad initial Top tracker"); + TopRPTracker.setPos(CurrentTop); + } } /// Move an instruction and update register pressure. @@ -685,10 +885,12 @@ void ScheduleDAGMI::scheduleMI(SUnit *SU, bool IsTopNode) { TopRPTracker.setPos(MI); } - // Update top scheduled pressure. - TopRPTracker.advance(); - assert(TopRPTracker.getPos() == CurrentTop && "out of sync"); - updateScheduledPressure(TopRPTracker.getPressure().MaxSetPressure); + if (ShouldTrackPressure) { + // Update top scheduled pressure. + TopRPTracker.advance(); + assert(TopRPTracker.getPos() == CurrentTop && "out of sync"); + updateScheduledPressure(SU, TopRPTracker.getPressure().MaxSetPressure); + } } else { assert(SU->isBottomReady() && "node still has unscheduled dependencies"); @@ -704,10 +906,14 @@ void ScheduleDAGMI::scheduleMI(SUnit *SU, bool IsTopNode) { moveInstruction(MI, CurrentBottom); CurrentBottom = MI; } - // Update bottom scheduled pressure. - BotRPTracker.recede(); - assert(BotRPTracker.getPos() == CurrentBottom && "out of sync"); - updateScheduledPressure(BotRPTracker.getPressure().MaxSetPressure); + if (ShouldTrackPressure) { + // Update bottom scheduled pressure. + SmallVector<unsigned, 8> LiveUses; + BotRPTracker.recede(&LiveUses); + assert(BotRPTracker.getPos() == CurrentBottom && "out of sync"); + updateScheduledPressure(SU, BotRPTracker.getPressure().MaxSetPressure); + updatePressureDiffs(LiveUses); + } } } @@ -1113,13 +1319,13 @@ void CopyConstrain::apply(ScheduleDAGMI *DAG) { } //===----------------------------------------------------------------------===// -// ConvergingScheduler - Implementation of the generic MachineSchedStrategy. +// GenericScheduler - Implementation of the generic MachineSchedStrategy. //===----------------------------------------------------------------------===// namespace { -/// ConvergingScheduler shrinks the unscheduled zone using heuristics to balance +/// GenericScheduler shrinks the unscheduled zone using heuristics to balance /// the schedule. -class ConvergingScheduler : public MachineSchedStrategy { +class GenericScheduler : public MachineSchedStrategy { public: /// Represent the type of SchedCandidate found within a single queue. /// pickNodeBidirectional depends on these listed by decreasing priority. @@ -1129,7 +1335,7 @@ public: TopDepthReduce, TopPathReduce, NextDefUse, NodeOrder}; #ifndef NDEBUG - static const char *getReasonStr(ConvergingScheduler::CandReason Reason); + static const char *getReasonStr(GenericScheduler::CandReason Reason); #endif /// Policy for scheduling the next instruction in the candidate's zone. @@ -1160,7 +1366,7 @@ public: } }; - /// Store the state used by ConvergingScheduler heuristics, required for the + /// Store the state used by GenericScheduler heuristics, required for the /// lifetime of one invocation of pickNode(). struct SchedCandidate { CandPolicy Policy; @@ -1205,16 +1411,21 @@ public: struct SchedRemainder { // Critical path through the DAG in expected latency. unsigned CriticalPath; + unsigned CyclicCritPath; // Scaled count of micro-ops left to schedule. unsigned RemIssueCount; + bool IsAcyclicLatencyLimited; + // Unscheduled resources SmallVector<unsigned, 16> RemainingCounts; void reset() { CriticalPath = 0; + CyclicCritPath = 0; RemIssueCount = 0; + IsAcyclicLatencyLimited = false; RemainingCounts.clear(); } @@ -1288,13 +1499,16 @@ public: void reset() { // A new HazardRec is created for each DAG and owned by SchedBoundary. - delete HazardRec; - + // Destroying and reconstructing it is very expensive though. So keep + // invalid, placeholder HazardRecs. + if (HazardRec && HazardRec->isEnabled()) { + delete HazardRec; + HazardRec = 0; + } Available.clear(); Pending.clear(); CheckPending = false; NextSUs.clear(); - HazardRec = 0; CurrCycle = 0; CurrMOps = 0; MinReadyCycle = UINT_MAX; @@ -1316,7 +1530,7 @@ public: /// PendingFlag set. SchedBoundary(unsigned ID, const Twine &Name): DAG(0), SchedModel(0), Rem(0), Available(ID, Name+".A"), - Pending(ID << ConvergingScheduler::LogMaxQID, Name+".P"), + Pending(ID << GenericScheduler::LogMaxQID, Name+".P"), HazardRec(0) { reset(); } @@ -1327,7 +1541,7 @@ public: SchedRemainder *rem); bool isTop() const { - return Available.getID() == ConvergingScheduler::TopQID; + return Available.getID() == GenericScheduler::TopQID; } #ifndef NDEBUG @@ -1399,6 +1613,7 @@ public: }; private: + const MachineSchedContext *Context; ScheduleDAGMI *DAG; const TargetSchedModel *SchedModel; const TargetRegisterInfo *TRI; @@ -1408,6 +1623,7 @@ private: SchedBoundary Top; SchedBoundary Bot; + MachineSchedPolicy RegionPolicy; public: /// SUnit::NodeQueueId: 0 (none), 1 (top), 2 (bot), 3 (both) enum { @@ -1416,8 +1632,15 @@ public: LogMaxQID = 2 }; - ConvergingScheduler(): - DAG(0), SchedModel(0), TRI(0), Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {} + GenericScheduler(const MachineSchedContext *C): + Context(C), DAG(0), SchedModel(0), TRI(0), + Top(TopQID, "TopQ"), Bot(BotQID, "BotQ") {} + + virtual void initPolicy(MachineBasicBlock::iterator Begin, + MachineBasicBlock::iterator End, + unsigned NumRegionInstrs); + + bool shouldTrackPressure() const { return RegionPolicy.ShouldTrackPressure; } virtual void initialize(ScheduleDAGMI *dag); @@ -1432,6 +1655,8 @@ public: virtual void registerRoots(); protected: + void checkAcyclicLatency(); + void tryCandidate(SchedCandidate &Cand, SchedCandidate &TryCand, SchedBoundary &Zone, @@ -1452,7 +1677,7 @@ protected: }; } // namespace -void ConvergingScheduler::SchedRemainder:: +void GenericScheduler::SchedRemainder:: init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) { reset(); if (!SchedModel->hasInstrSchedModel()) @@ -1473,7 +1698,7 @@ init(ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) { } } -void ConvergingScheduler::SchedBoundary:: +void GenericScheduler::SchedBoundary:: init(ScheduleDAGMI *dag, const TargetSchedModel *smodel, SchedRemainder *rem) { reset(); DAG = dag; @@ -1483,7 +1708,49 @@ init(ScheduleDAGMI *dag, const TargetSchedModel *smodel, SchedRemainder *rem) { ExecutedResCounts.resize(SchedModel->getNumProcResourceKinds()); } -void ConvergingScheduler::initialize(ScheduleDAGMI *dag) { +/// Initialize the per-region scheduling policy. +void GenericScheduler::initPolicy(MachineBasicBlock::iterator Begin, + MachineBasicBlock::iterator End, + unsigned NumRegionInstrs) { + const TargetMachine &TM = Context->MF->getTarget(); + + // Avoid setting up the register pressure tracker for small regions to save + // compile time. As a rough heuristic, only track pressure when the number of + // schedulable instructions exceeds half the integer register file. + unsigned NIntRegs = Context->RegClassInfo->getNumAllocatableRegs( + TM.getTargetLowering()->getRegClassFor(MVT::i32)); + + RegionPolicy.ShouldTrackPressure = NumRegionInstrs > (NIntRegs / 2); + + // For generic targets, we default to bottom-up, because it's simpler and more + // compile-time optimizations have been implemented in that direction. + RegionPolicy.OnlyBottomUp = true; + + // Allow the subtarget to override default policy. + const TargetSubtargetInfo &ST = TM.getSubtarget<TargetSubtargetInfo>(); + ST.overrideSchedPolicy(RegionPolicy, Begin, End, NumRegionInstrs); + + // After subtarget overrides, apply command line options. + if (!EnableRegPressure) + RegionPolicy.ShouldTrackPressure = false; + + // Check -misched-topdown/bottomup can force or unforce scheduling direction. + // e.g. -misched-bottomup=false allows scheduling in both directions. + assert((!ForceTopDown || !ForceBottomUp) && + "-misched-topdown incompatible with -misched-bottomup"); + if (ForceBottomUp.getNumOccurrences() > 0) { + RegionPolicy.OnlyBottomUp = ForceBottomUp; + if (RegionPolicy.OnlyBottomUp) + RegionPolicy.OnlyTopDown = false; + } + if (ForceTopDown.getNumOccurrences() > 0) { + RegionPolicy.OnlyTopDown = ForceTopDown; + if (RegionPolicy.OnlyTopDown) + RegionPolicy.OnlyBottomUp = false; + } +} + +void GenericScheduler::initialize(ScheduleDAGMI *dag) { DAG = dag; SchedModel = DAG->getSchedModel(); TRI = DAG->TRI; @@ -1498,14 +1765,17 @@ void ConvergingScheduler::initialize(ScheduleDAGMI *dag) { // are disabled, then these HazardRecs will be disabled. const InstrItineraryData *Itin = SchedModel->getInstrItineraries(); const TargetMachine &TM = DAG->MF.getTarget(); - Top.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG); - Bot.HazardRec = TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG); - - assert((!ForceTopDown || !ForceBottomUp) && - "-misched-topdown incompatible with -misched-bottomup"); + if (!Top.HazardRec) { + Top.HazardRec = + TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG); + } + if (!Bot.HazardRec) { + Bot.HazardRec = + TM.getInstrInfo()->CreateTargetMIHazardRecognizer(Itin, DAG); + } } -void ConvergingScheduler::releaseTopNode(SUnit *SU) { +void GenericScheduler::releaseTopNode(SUnit *SU) { if (SU->isScheduled) return; @@ -1524,7 +1794,7 @@ void ConvergingScheduler::releaseTopNode(SUnit *SU) { Top.releaseNode(SU, SU->TopReadyCycle); } -void ConvergingScheduler::releaseBottomNode(SUnit *SU) { +void GenericScheduler::releaseBottomNode(SUnit *SU) { if (SU->isScheduled) return; @@ -1545,8 +1815,46 @@ void ConvergingScheduler::releaseBottomNode(SUnit *SU) { Bot.releaseNode(SU, SU->BotReadyCycle); } -void ConvergingScheduler::registerRoots() { +/// Set IsAcyclicLatencyLimited if the acyclic path is longer than the cyclic +/// critical path by more cycles than it takes to drain the instruction buffer. +/// We estimate an upper bounds on in-flight instructions as: +/// +/// CyclesPerIteration = max( CyclicPath, Loop-Resource-Height ) +/// InFlightIterations = AcyclicPath / CyclesPerIteration +/// InFlightResources = InFlightIterations * LoopResources +/// +/// TODO: Check execution resources in addition to IssueCount. +void GenericScheduler::checkAcyclicLatency() { + if (Rem.CyclicCritPath == 0 || Rem.CyclicCritPath >= Rem.CriticalPath) + return; + + // Scaled number of cycles per loop iteration. + unsigned IterCount = + std::max(Rem.CyclicCritPath * SchedModel->getLatencyFactor(), + Rem.RemIssueCount); + // Scaled acyclic critical path. + unsigned AcyclicCount = Rem.CriticalPath * SchedModel->getLatencyFactor(); + // InFlightCount = (AcyclicPath / IterCycles) * InstrPerLoop + unsigned InFlightCount = + (AcyclicCount * Rem.RemIssueCount + IterCount-1) / IterCount; + unsigned BufferLimit = + SchedModel->getMicroOpBufferSize() * SchedModel->getMicroOpFactor(); + + Rem.IsAcyclicLatencyLimited = InFlightCount > BufferLimit; + + DEBUG(dbgs() << "IssueCycles=" + << Rem.RemIssueCount / SchedModel->getLatencyFactor() << "c " + << "IterCycles=" << IterCount / SchedModel->getLatencyFactor() + << "c NumIters=" << (AcyclicCount + IterCount-1) / IterCount + << " InFlight=" << InFlightCount / SchedModel->getMicroOpFactor() + << "m BufferLim=" << SchedModel->getMicroOpBufferSize() << "m\n"; + if (Rem.IsAcyclicLatencyLimited) + dbgs() << " ACYCLIC LATENCY LIMIT\n"); +} + +void GenericScheduler::registerRoots() { Rem.CriticalPath = DAG->ExitSU.getDepth(); + // Some roots may not feed into ExitSU. Check all of them in case. for (std::vector<SUnit*>::const_iterator I = Bot.Available.begin(), E = Bot.Available.end(); I != E; ++I) { @@ -1554,6 +1862,11 @@ void ConvergingScheduler::registerRoots() { Rem.CriticalPath = (*I)->getDepth(); } DEBUG(dbgs() << "Critical Path: " << Rem.CriticalPath << '\n'); + + if (EnableCyclicPath) { + Rem.CyclicCritPath = DAG->computeCyclicCriticalPath(); + checkAcyclicLatency(); + } } /// Does this SU have a hazard within the current instruction group. @@ -1569,7 +1882,7 @@ void ConvergingScheduler::registerRoots() { /// can dispatch per cycle. /// /// TODO: Also check whether the SU must start a new group. -bool ConvergingScheduler::SchedBoundary::checkHazard(SUnit *SU) { +bool GenericScheduler::SchedBoundary::checkHazard(SUnit *SU) { if (HazardRec->isEnabled()) return HazardRec->getHazardType(SU) != ScheduleHazardRecognizer::NoHazard; @@ -1583,7 +1896,7 @@ bool ConvergingScheduler::SchedBoundary::checkHazard(SUnit *SU) { } // Find the unscheduled node in ReadySUs with the highest latency. -unsigned ConvergingScheduler::SchedBoundary:: +unsigned GenericScheduler::SchedBoundary:: findMaxLatency(ArrayRef<SUnit*> ReadySUs) { SUnit *LateSU = 0; unsigned RemLatency = 0; @@ -1605,7 +1918,7 @@ findMaxLatency(ArrayRef<SUnit*> ReadySUs) { // Count resources in this zone and the remaining unscheduled // instruction. Return the max count, scaled. Set OtherCritIdx to the critical // resource index, or zero if the zone is issue limited. -unsigned ConvergingScheduler::SchedBoundary:: +unsigned GenericScheduler::SchedBoundary:: getOtherResourceCount(unsigned &OtherCritIdx) { OtherCritIdx = 0; if (!SchedModel->hasInstrSchedModel()) @@ -1633,7 +1946,7 @@ getOtherResourceCount(unsigned &OtherCritIdx) { /// Set the CandPolicy for this zone given the current resources and latencies /// inside and outside the zone. -void ConvergingScheduler::SchedBoundary::setPolicy(CandPolicy &Policy, +void GenericScheduler::SchedBoundary::setPolicy(CandPolicy &Policy, SchedBoundary &OtherZone) { // Now that potential stalls have been considered, apply preemptive heuristics // based on the the total latency and resources inside and outside this @@ -1692,7 +2005,7 @@ void ConvergingScheduler::SchedBoundary::setPolicy(CandPolicy &Policy, Policy.DemandResIdx = OtherCritIdx; } -void ConvergingScheduler::SchedBoundary::releaseNode(SUnit *SU, +void GenericScheduler::SchedBoundary::releaseNode(SUnit *SU, unsigned ReadyCycle) { if (ReadyCycle < MinReadyCycle) MinReadyCycle = ReadyCycle; @@ -1710,7 +2023,7 @@ void ConvergingScheduler::SchedBoundary::releaseNode(SUnit *SU, } /// Move the boundary of scheduled code by one cycle. -void ConvergingScheduler::SchedBoundary::bumpCycle(unsigned NextCycle) { +void GenericScheduler::SchedBoundary::bumpCycle(unsigned NextCycle) { if (SchedModel->getMicroOpBufferSize() == 0) { assert(MinReadyCycle < UINT_MAX && "MinReadyCycle uninitialized"); if (MinReadyCycle > NextCycle) @@ -1748,7 +2061,7 @@ void ConvergingScheduler::SchedBoundary::bumpCycle(unsigned NextCycle) { DEBUG(dbgs() << "Cycle: " << CurrCycle << ' ' << Available.getName() << '\n'); } -void ConvergingScheduler::SchedBoundary::incExecutedResources(unsigned PIdx, +void GenericScheduler::SchedBoundary::incExecutedResources(unsigned PIdx, unsigned Count) { ExecutedResCounts[PIdx] += Count; if (ExecutedResCounts[PIdx] > MaxExecutedResCount) @@ -1762,7 +2075,7 @@ void ConvergingScheduler::SchedBoundary::incExecutedResources(unsigned PIdx, /// /// \return the next cycle at which the instruction may execute without /// oversubscribing resources. -unsigned ConvergingScheduler::SchedBoundary:: +unsigned GenericScheduler::SchedBoundary:: countResource(unsigned PIdx, unsigned Cycles, unsigned ReadyCycle) { unsigned Factor = SchedModel->getResourceFactor(PIdx); unsigned Count = Factor * Cycles; @@ -1787,7 +2100,7 @@ countResource(unsigned PIdx, unsigned Cycles, unsigned ReadyCycle) { } /// Move the boundary of scheduled code by one SUnit. -void ConvergingScheduler::SchedBoundary::bumpNode(SUnit *SU) { +void GenericScheduler::SchedBoundary::bumpNode(SUnit *SU) { // Update the reservation table. if (HazardRec->isEnabled()) { if (!isTop() && SU->isCall) { @@ -1891,7 +2204,7 @@ void ConvergingScheduler::SchedBoundary::bumpNode(SUnit *SU) { /// Release pending ready nodes in to the available queue. This makes them /// visible to heuristics. -void ConvergingScheduler::SchedBoundary::releasePending() { +void GenericScheduler::SchedBoundary::releasePending() { // If the available queue is empty, it is safe to reset MinReadyCycle. if (Available.empty()) MinReadyCycle = UINT_MAX; @@ -1921,7 +2234,7 @@ void ConvergingScheduler::SchedBoundary::releasePending() { } /// Remove SU from the ready set for this boundary. -void ConvergingScheduler::SchedBoundary::removeReady(SUnit *SU) { +void GenericScheduler::SchedBoundary::removeReady(SUnit *SU) { if (Available.isInQueue(SU)) Available.remove(Available.find(SU)); else { @@ -1933,7 +2246,7 @@ void ConvergingScheduler::SchedBoundary::removeReady(SUnit *SU) { /// If this queue only has one ready candidate, return it. As a side effect, /// defer any nodes that now hit a hazard, and advance the cycle until at least /// one node is ready. If multiple instructions are ready, return NULL. -SUnit *ConvergingScheduler::SchedBoundary::pickOnlyChoice() { +SUnit *GenericScheduler::SchedBoundary::pickOnlyChoice() { if (CheckPending) releasePending(); @@ -1962,7 +2275,7 @@ SUnit *ConvergingScheduler::SchedBoundary::pickOnlyChoice() { #ifndef NDEBUG // This is useful information to dump after bumpNode. // Note that the Queue contents are more useful before pickNodeFromQueue. -void ConvergingScheduler::SchedBoundary::dumpScheduledState() { +void GenericScheduler::SchedBoundary::dumpScheduledState() { unsigned ResFactor; unsigned ResCount; if (ZoneCritResIdx) { @@ -1985,7 +2298,7 @@ void ConvergingScheduler::SchedBoundary::dumpScheduledState() { } #endif -void ConvergingScheduler::SchedCandidate:: +void GenericScheduler::SchedCandidate:: initResourceDelta(const ScheduleDAGMI *DAG, const TargetSchedModel *SchedModel) { if (!Policy.ReduceResIdx && !Policy.DemandResIdx) @@ -2005,9 +2318,9 @@ initResourceDelta(const ScheduleDAGMI *DAG, /// Return true if this heuristic determines order. static bool tryLess(int TryVal, int CandVal, - ConvergingScheduler::SchedCandidate &TryCand, - ConvergingScheduler::SchedCandidate &Cand, - ConvergingScheduler::CandReason Reason) { + GenericScheduler::SchedCandidate &TryCand, + GenericScheduler::SchedCandidate &Cand, + GenericScheduler::CandReason Reason) { if (TryVal < CandVal) { TryCand.Reason = Reason; return true; @@ -2022,9 +2335,9 @@ static bool tryLess(int TryVal, int CandVal, } static bool tryGreater(int TryVal, int CandVal, - ConvergingScheduler::SchedCandidate &TryCand, - ConvergingScheduler::SchedCandidate &Cand, - ConvergingScheduler::CandReason Reason) { + GenericScheduler::SchedCandidate &TryCand, + GenericScheduler::SchedCandidate &Cand, + GenericScheduler::CandReason Reason) { if (TryVal > CandVal) { TryCand.Reason = Reason; return true; @@ -2038,26 +2351,26 @@ static bool tryGreater(int TryVal, int CandVal, return false; } -static bool tryPressure(const PressureElement &TryP, - const PressureElement &CandP, - ConvergingScheduler::SchedCandidate &TryCand, - ConvergingScheduler::SchedCandidate &Cand, - ConvergingScheduler::CandReason Reason) { +static bool tryPressure(const PressureChange &TryP, + const PressureChange &CandP, + GenericScheduler::SchedCandidate &TryCand, + GenericScheduler::SchedCandidate &Cand, + GenericScheduler::CandReason Reason) { + int TryRank = TryP.getPSetOrMax(); + int CandRank = CandP.getPSetOrMax(); // If both candidates affect the same set, go with the smallest increase. - if (TryP.PSetID == CandP.PSetID) { - return tryLess(TryP.UnitIncrease, CandP.UnitIncrease, TryCand, Cand, + if (TryRank == CandRank) { + return tryLess(TryP.getUnitInc(), CandP.getUnitInc(), TryCand, Cand, Reason); } // If one candidate decreases and the other increases, go with it. - if (tryLess(TryP.UnitIncrease < 0, CandP.UnitIncrease < 0, TryCand, Cand, + // Invalid candidates have UnitInc==0. + if (tryLess(TryP.getUnitInc() < 0, CandP.getUnitInc() < 0, TryCand, Cand, Reason)) { return true; } - // If TryP has lower Rank, it has a higher priority. - int TryRank = TryP.PSetRank(); - int CandRank = CandP.PSetRank(); // If the candidates are decreasing pressure, reverse priority. - if (TryP.UnitIncrease < 0) + if (TryP.getUnitInc() < 0) std::swap(TryRank, CandRank); return tryGreater(TryRank, CandRank, TryCand, Cand, Reason); } @@ -2094,6 +2407,32 @@ static int biasPhysRegCopy(const SUnit *SU, bool isTop) { return 0; } +static bool tryLatency(GenericScheduler::SchedCandidate &TryCand, + GenericScheduler::SchedCandidate &Cand, + GenericScheduler::SchedBoundary &Zone) { + if (Zone.isTop()) { + if (Cand.SU->getDepth() > Zone.getScheduledLatency()) { + if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(), + TryCand, Cand, GenericScheduler::TopDepthReduce)) + return true; + } + if (tryGreater(TryCand.SU->getHeight(), Cand.SU->getHeight(), + TryCand, Cand, GenericScheduler::TopPathReduce)) + return true; + } + else { + if (Cand.SU->getHeight() > Zone.getScheduledLatency()) { + if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(), + TryCand, Cand, GenericScheduler::BotHeightReduce)) + return true; + } + if (tryGreater(TryCand.SU->getDepth(), Cand.SU->getDepth(), + TryCand, Cand, GenericScheduler::BotPathReduce)) + return true; + } + return false; +} + /// Apply a set of heursitics to a new candidate. Heuristics are currently /// hierarchical. This may be more efficient than a graduated cost model because /// we don't need to evaluate all aspects of the model for each node in the @@ -2105,16 +2444,44 @@ static int biasPhysRegCopy(const SUnit *SU, bool isTop) { /// \param Zone describes the scheduled zone that we are extending. /// \param RPTracker describes reg pressure within the scheduled zone. /// \param TempTracker is a scratch pressure tracker to reuse in queries. -void ConvergingScheduler::tryCandidate(SchedCandidate &Cand, +void GenericScheduler::tryCandidate(SchedCandidate &Cand, SchedCandidate &TryCand, SchedBoundary &Zone, const RegPressureTracker &RPTracker, RegPressureTracker &TempTracker) { - // Always initialize TryCand's RPDelta. - TempTracker.getMaxPressureDelta(TryCand.SU->getInstr(), TryCand.RPDelta, - DAG->getRegionCriticalPSets(), - DAG->getRegPressure().MaxSetPressure); + if (DAG->isTrackingPressure()) { + // Always initialize TryCand's RPDelta. + if (Zone.isTop()) { + TempTracker.getMaxDownwardPressureDelta( + TryCand.SU->getInstr(), + TryCand.RPDelta, + DAG->getRegionCriticalPSets(), + DAG->getRegPressure().MaxSetPressure); + } + else { + if (VerifyScheduling) { + TempTracker.getMaxUpwardPressureDelta( + TryCand.SU->getInstr(), + &DAG->getPressureDiff(TryCand.SU), + TryCand.RPDelta, + DAG->getRegionCriticalPSets(), + DAG->getRegPressure().MaxSetPressure); + } + else { + RPTracker.getUpwardPressureDelta( + TryCand.SU->getInstr(), + DAG->getPressureDiff(TryCand.SU), + TryCand.RPDelta, + DAG->getRegionCriticalPSets(), + DAG->getRegPressure().MaxSetPressure); + } + } + } + DEBUG(if (TryCand.RPDelta.Excess.isValid()) + dbgs() << " SU(" << TryCand.SU->NodeNum << ") " + << TRI->getRegPressureSetName(TryCand.RPDelta.Excess.getPSet()) + << ":" << TryCand.RPDelta.Excess.getUnitInc() << "\n"); // Initialize the candidate if needed. if (!Cand.isValid()) { @@ -2129,13 +2496,22 @@ void ConvergingScheduler::tryCandidate(SchedCandidate &Cand, // Avoid exceeding the target's limit. If signed PSetID is negative, it is // invalid; convert it to INT_MAX to give it lowest priority. - if (tryPressure(TryCand.RPDelta.Excess, Cand.RPDelta.Excess, TryCand, Cand, - RegExcess)) + if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.Excess, + Cand.RPDelta.Excess, + TryCand, Cand, RegExcess)) return; // Avoid increasing the max critical pressure in the scheduled region. - if (tryPressure(TryCand.RPDelta.CriticalMax, Cand.RPDelta.CriticalMax, - TryCand, Cand, RegCritical)) + if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CriticalMax, + Cand.RPDelta.CriticalMax, + TryCand, Cand, RegCritical)) + return; + + // For loops that are acyclic path limited, aggressively schedule for latency. + // This can result in very long dependence chains scheduled in sequence, so + // once every cycle (when CurrMOps == 0), switch to normal heuristics. + if (Rem.IsAcyclicLatencyLimited && !Zone.CurrMOps + && tryLatency(TryCand, Cand, Zone)) return; // Keep clustered nodes together to encourage downstream peephole @@ -2157,8 +2533,9 @@ void ConvergingScheduler::tryCandidate(SchedCandidate &Cand, return; } // Avoid increasing the max pressure of the entire region. - if (tryPressure(TryCand.RPDelta.CurrentMax, Cand.RPDelta.CurrentMax, - TryCand, Cand, RegMax)) + if (DAG->isTrackingPressure() && tryPressure(TryCand.RPDelta.CurrentMax, + Cand.RPDelta.CurrentMax, + TryCand, Cand, RegMax)) return; // Avoid critical resource consumption and balance the schedule. @@ -2172,27 +2549,10 @@ void ConvergingScheduler::tryCandidate(SchedCandidate &Cand, return; // Avoid serializing long latency dependence chains. - if (Cand.Policy.ReduceLatency) { - if (Zone.isTop()) { - if (Cand.SU->getDepth() > Zone.getScheduledLatency()) { - if (tryLess(TryCand.SU->getDepth(), Cand.SU->getDepth(), - TryCand, Cand, TopDepthReduce)) - return; - } - if (tryGreater(TryCand.SU->getHeight(), Cand.SU->getHeight(), - TryCand, Cand, TopPathReduce)) - return; - } - else { - if (Cand.SU->getHeight() > Zone.getScheduledLatency()) { - if (tryLess(TryCand.SU->getHeight(), Cand.SU->getHeight(), - TryCand, Cand, BotHeightReduce)) - return; - } - if (tryGreater(TryCand.SU->getDepth(), Cand.SU->getDepth(), - TryCand, Cand, BotPathReduce)) - return; - } + // For acyclic path limited loops, latency was already checked above. + if (Cand.Policy.ReduceLatency && !Rem.IsAcyclicLatencyLimited + && tryLatency(TryCand, Cand, Zone)) { + return; } // Prefer immediate defs/users of the last scheduled instruction. This is a @@ -2210,8 +2570,8 @@ void ConvergingScheduler::tryCandidate(SchedCandidate &Cand, } #ifndef NDEBUG -const char *ConvergingScheduler::getReasonStr( - ConvergingScheduler::CandReason Reason) { +const char *GenericScheduler::getReasonStr( + GenericScheduler::CandReason Reason) { switch (Reason) { case NoCand: return "NOCAND "; case PhysRegCopy: return "PREG-COPY"; @@ -2232,8 +2592,8 @@ const char *ConvergingScheduler::getReasonStr( llvm_unreachable("Unknown reason!"); } -void ConvergingScheduler::traceCandidate(const SchedCandidate &Cand) { - PressureElement P; +void GenericScheduler::traceCandidate(const SchedCandidate &Cand) { + PressureChange P; unsigned ResIdx = 0; unsigned Latency = 0; switch (Cand.Reason) { @@ -2269,8 +2629,8 @@ void ConvergingScheduler::traceCandidate(const SchedCandidate &Cand) { } dbgs() << " SU(" << Cand.SU->NodeNum << ") " << getReasonStr(Cand.Reason); if (P.isValid()) - dbgs() << " " << TRI->getRegPressureSetName(P.PSetID) - << ":" << P.UnitIncrease << " "; + dbgs() << " " << TRI->getRegPressureSetName(P.getPSet()) + << ":" << P.getUnitInc() << " "; else dbgs() << " "; if (ResIdx) @@ -2285,12 +2645,12 @@ void ConvergingScheduler::traceCandidate(const SchedCandidate &Cand) { } #endif -/// Pick the best candidate from the top queue. +/// Pick the best candidate from the queue. /// /// TODO: getMaxPressureDelta results can be mostly cached for each SUnit during /// DAG building. To adjust for the current scheduling location we need to /// maintain the number of vreg uses remaining to be top-scheduled. -void ConvergingScheduler::pickNodeFromQueue(SchedBoundary &Zone, +void GenericScheduler::pickNodeFromQueue(SchedBoundary &Zone, const RegPressureTracker &RPTracker, SchedCandidate &Cand) { ReadyQueue &Q = Zone.Available; @@ -2315,14 +2675,14 @@ void ConvergingScheduler::pickNodeFromQueue(SchedBoundary &Zone, } } -static void tracePick(const ConvergingScheduler::SchedCandidate &Cand, +static void tracePick(const GenericScheduler::SchedCandidate &Cand, bool IsTop) { DEBUG(dbgs() << "Pick " << (IsTop ? "Top " : "Bot ") - << ConvergingScheduler::getReasonStr(Cand.Reason) << '\n'); + << GenericScheduler::getReasonStr(Cand.Reason) << '\n'); } /// Pick the best candidate node from either the top or bottom queue. -SUnit *ConvergingScheduler::pickNodeBidirectional(bool &IsTopNode) { +SUnit *GenericScheduler::pickNodeBidirectional(bool &IsTopNode) { // Schedule as far as possible in the direction of no choice. This is most // efficient, but also provides the best heuristics for CriticalPSets. if (SUnit *SU = Bot.pickOnlyChoice()) { @@ -2377,7 +2737,7 @@ SUnit *ConvergingScheduler::pickNodeBidirectional(bool &IsTopNode) { } /// Pick the best node to balance the schedule. Implements MachineSchedStrategy. -SUnit *ConvergingScheduler::pickNode(bool &IsTopNode) { +SUnit *GenericScheduler::pickNode(bool &IsTopNode) { if (DAG->top() == DAG->bottom()) { assert(Top.Available.empty() && Top.Pending.empty() && Bot.Available.empty() && Bot.Pending.empty() && "ReadyQ garbage"); @@ -2385,24 +2745,26 @@ SUnit *ConvergingScheduler::pickNode(bool &IsTopNode) { } SUnit *SU; do { - if (ForceTopDown) { + if (RegionPolicy.OnlyTopDown) { SU = Top.pickOnlyChoice(); if (!SU) { CandPolicy NoPolicy; SchedCandidate TopCand(NoPolicy); pickNodeFromQueue(Top, DAG->getTopRPTracker(), TopCand); - assert(TopCand.Reason != NoCand && "failed to find the first candidate"); + assert(TopCand.Reason != NoCand && "failed to find a candidate"); + tracePick(TopCand, true); SU = TopCand.SU; } IsTopNode = true; } - else if (ForceBottomUp) { + else if (RegionPolicy.OnlyBottomUp) { SU = Bot.pickOnlyChoice(); if (!SU) { CandPolicy NoPolicy; SchedCandidate BotCand(NoPolicy); pickNodeFromQueue(Bot, DAG->getBotRPTracker(), BotCand); - assert(BotCand.Reason != NoCand && "failed to find the first candidate"); + assert(BotCand.Reason != NoCand && "failed to find a candidate"); + tracePick(BotCand, false); SU = BotCand.SU; } IsTopNode = false; @@ -2421,7 +2783,7 @@ SUnit *ConvergingScheduler::pickNode(bool &IsTopNode) { return SU; } -void ConvergingScheduler::reschedulePhysRegCopies(SUnit *SU, bool isTop) { +void GenericScheduler::reschedulePhysRegCopies(SUnit *SU, bool isTop) { MachineBasicBlock::iterator InsertPos = SU->getInstr(); if (!isTop) @@ -2452,7 +2814,7 @@ void ConvergingScheduler::reschedulePhysRegCopies(SUnit *SU, bool isTop) { /// /// FIXME: Eventually, we may bundle physreg copies rather than rescheduling /// them here. See comments in biasPhysRegCopy. -void ConvergingScheduler::schedNode(SUnit *SU, bool IsTopNode) { +void GenericScheduler::schedNode(SUnit *SU, bool IsTopNode) { if (IsTopNode) { SU->TopReadyCycle = std::max(SU->TopReadyCycle, Top.CurrCycle); Top.bumpNode(SU); @@ -2469,25 +2831,23 @@ void ConvergingScheduler::schedNode(SUnit *SU, bool IsTopNode) { /// Create the standard converging machine scheduler. This will be used as the /// default scheduler if the target does not set a default. -static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C) { - assert((!ForceTopDown || !ForceBottomUp) && - "-misched-topdown incompatible with -misched-bottomup"); - ScheduleDAGMI *DAG = new ScheduleDAGMI(C, new ConvergingScheduler()); +static ScheduleDAGInstrs *createGenericSched(MachineSchedContext *C) { + ScheduleDAGMI *DAG = new ScheduleDAGMI(C, new GenericScheduler(C)); // Register DAG post-processors. // // FIXME: extend the mutation API to allow earlier mutations to instantiate // data and pass it to later mutations. Have a single mutation that gathers // the interesting nodes in one pass. DAG->addMutation(new CopyConstrain(DAG->TII, DAG->TRI)); - if (EnableLoadCluster) + if (EnableLoadCluster && DAG->TII->enableClusterLoads()) DAG->addMutation(new LoadClusterMutation(DAG->TII, DAG->TRI)); if (EnableMacroFusion) DAG->addMutation(new MacroFusion(DAG->TII)); return DAG; } static MachineSchedRegistry -ConvergingSchedRegistry("converge", "Standard converging scheduler.", - createConvergingSched); +GenericSchedRegistry("converge", "Standard converging scheduler.", + createGenericSched); //===----------------------------------------------------------------------===// // ILP Scheduler. Currently for experimental analysis of heuristics. @@ -2529,15 +2889,6 @@ struct ILPOrder { /// \brief Schedule based on the ILP metric. class ILPScheduler : public MachineSchedStrategy { - /// In case all subtrees are eventually connected to a common root through - /// data dependence (e.g. reduction), place an upper limit on their size. - /// - /// FIXME: A subtree limit is generally good, but in the situation commented - /// above, where multiple similar subtrees feed a common root, we should - /// only split at a point where the resulting subtrees will be balanced. - /// (a motivating test case must be found). - static const unsigned SubtreeLimit = 16; - ScheduleDAGMI *DAG; ILPOrder Cmp; @@ -2721,7 +3072,7 @@ struct DOTGraphTraits<ScheduleDAGMI*> : public DefaultDOTGraphTraits { } static bool isNodeHidden(const SUnit *Node) { - return (Node->NumPreds > 10 || Node->NumSuccs > 10); + return (Node->Preds.size() > 10 || Node->Succs.size() > 10); } static bool hasNodeAddressLabel(const SUnit *Node, @@ -2744,7 +3095,11 @@ struct DOTGraphTraits<ScheduleDAGMI*> : public DefaultDOTGraphTraits { static std::string getNodeLabel(const SUnit *SU, const ScheduleDAG *G) { std::string Str; raw_string_ostream SS(Str); - SS << "SU(" << SU->NodeNum << ')'; + const SchedDFSResult *DFS = + static_cast<const ScheduleDAGMI*>(G)->getDFSResult(); + SS << "SU:" << SU->NodeNum; + if (DFS) + SS << " I:" << DFS->getNumInstrs(SU); return SS.str(); } static std::string getNodeDescription(const SUnit *SU, const ScheduleDAG *G) { |