X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FPostRASchedulerList.cpp;h=27f56762de9be4e8947bc37bd06d86a3eaf26ed3;hb=228546b8cb53686f1cc8859ec39bbb11f5c15e90;hp=5917e76004ab1bebe21baeb4cae06c2e079cb90b;hpb=0855dee564f80160abf95497475306af38ab7f84;p=oota-llvm.git diff --git a/lib/CodeGen/PostRASchedulerList.cpp b/lib/CodeGen/PostRASchedulerList.cpp index 5917e76004a..27f56762de9 100644 --- a/lib/CodeGen/PostRASchedulerList.cpp +++ b/lib/CodeGen/PostRASchedulerList.cpp @@ -19,35 +19,33 @@ //===----------------------------------------------------------------------===// #define DEBUG_TYPE "post-RA-sched" -#include "AntiDepBreaker.h" +#include "llvm/CodeGen/Passes.h" #include "AggressiveAntiDepBreaker.h" +#include "AntiDepBreaker.h" #include "CriticalAntiDepBreaker.h" -#include "ExactHazardRecognizer.h" -#include "SimpleHazardRecognizer.h" -#include "ScheduleDAGInstrs.h" -#include "llvm/CodeGen/Passes.h" +#include "llvm/ADT/BitVector.h" +#include "llvm/ADT/Statistic.h" +#include "llvm/Analysis/AliasAnalysis.h" #include "llvm/CodeGen/LatencyPriorityQueue.h" -#include "llvm/CodeGen/SchedulerRegistry.h" #include "llvm/CodeGen/MachineDominators.h" #include "llvm/CodeGen/MachineFrameInfo.h" #include "llvm/CodeGen/MachineFunctionPass.h" +#include "llvm/CodeGen/MachineInstrBuilder.h" #include "llvm/CodeGen/MachineLoopInfo.h" #include "llvm/CodeGen/MachineRegisterInfo.h" +#include "llvm/CodeGen/RegisterClassInfo.h" +#include "llvm/CodeGen/ScheduleDAGInstrs.h" #include "llvm/CodeGen/ScheduleHazardRecognizer.h" -#include "llvm/Analysis/AliasAnalysis.h" -#include "llvm/Target/TargetLowering.h" -#include "llvm/Target/TargetMachine.h" -#include "llvm/Target/TargetInstrInfo.h" -#include "llvm/Target/TargetRegisterInfo.h" -#include "llvm/Target/TargetSubtarget.h" +#include "llvm/CodeGen/SchedulerRegistry.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/ErrorHandling.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/ADT/BitVector.h" -#include "llvm/ADT/Statistic.h" -#include -#include +#include "llvm/Target/TargetInstrInfo.h" +#include "llvm/Target/TargetLowering.h" +#include "llvm/Target/TargetMachine.h" +#include "llvm/Target/TargetRegisterInfo.h" +#include "llvm/Target/TargetSubtargetInfo.h" using namespace llvm; STATISTIC(NumNoops, "Number of noops inserted"); @@ -55,7 +53,7 @@ STATISTIC(NumStalls, "Number of pipeline stalls"); STATISTIC(NumFixedAnti, "Number of fixed anti-dependencies"); // Post-RA scheduling is enabled with -// TargetSubtarget.enablePostRAScheduler(). This flag can be used to +// TargetSubtargetInfo.enablePostRAScheduler(). This flag can be used to // override the target. static cl::opt EnablePostRAScheduler("post-RA-scheduler", @@ -66,10 +64,6 @@ EnableAntiDepBreaking("break-anti-dependencies", cl::desc("Break post-RA scheduling anti-dependencies: " "\"critical\", \"all\", or \"none\""), cl::init("none"), cl::Hidden); -static cl::opt -EnablePostRAHazardAvoidance("avoid-hazards", - cl::desc("Enable exact hazard avoidance"), - cl::init(true), cl::Hidden); // If DebugDiv > 0 then only schedule MBB with (ID % DebugDiv) == DebugMod static cl::opt @@ -85,17 +79,17 @@ AntiDepBreaker::~AntiDepBreaker() { } namespace { class PostRAScheduler : public MachineFunctionPass { - AliasAnalysis *AA; - CodeGenOpt::Level OptLevel; + const TargetInstrInfo *TII; + RegisterClassInfo RegClassInfo; public: static char ID; - PostRAScheduler(CodeGenOpt::Level ol) : - MachineFunctionPass(&ID), OptLevel(ol) {} + PostRAScheduler() : MachineFunctionPass(ID) {} void getAnalysisUsage(AnalysisUsage &AU) const { AU.setPreservesCFG(); AU.addRequired(); + AU.addRequired(); AU.addRequired(); AU.addPreserved(); AU.addRequired(); @@ -103,10 +97,6 @@ namespace { MachineFunctionPass::getAnalysisUsage(AU); } - const char *getPassName() const { - return "Post RA top-down list latency scheduler"; - } - bool runOnMachineFunction(MachineFunction &Fn); }; char PostRAScheduler::ID = 0; @@ -115,16 +105,13 @@ namespace { /// AvailableQueue - The priority queue to use for the available SUnits. /// LatencyPriorityQueue AvailableQueue; - + /// PendingQueue - This contains all of the instructions whose operands have /// been issued, but their results are not ready yet (due to the latency of /// the operation). Once the operands becomes available, the instruction is /// added to the AvailableQueue. std::vector PendingQueue; - /// Topo - A topological ordering for SUnits. - ScheduleDAGTopologicalSort Topo; - /// HazardRec - The hazard recognizer to use. ScheduleHazardRecognizer *HazardRec; @@ -134,40 +121,49 @@ namespace { /// AA - AliasAnalysis for making memory reference queries. AliasAnalysis *AA; - /// KillIndices - The index of the most recent kill (proceding bottom-up), - /// or ~0u if the register is not live. - unsigned KillIndices[TargetRegisterInfo::FirstVirtualRegister]; + /// LiveRegs - true if the register is live. + BitVector LiveRegs; + + /// The schedule. Null SUnit*'s represent noop instructions. + std::vector Sequence; public: - SchedulePostRATDList(MachineFunction &MF, - const MachineLoopInfo &MLI, - const MachineDominatorTree &MDT, - ScheduleHazardRecognizer *HR, - AntiDepBreaker *ADB, - AliasAnalysis *aa) - : ScheduleDAGInstrs(MF, MLI, MDT), Topo(SUnits), - HazardRec(HR), AntiDepBreak(ADB), AA(aa) {} - - ~SchedulePostRATDList() { - } + SchedulePostRATDList( + MachineFunction &MF, MachineLoopInfo &MLI, MachineDominatorTree &MDT, + AliasAnalysis *AA, const RegisterClassInfo&, + TargetSubtargetInfo::AntiDepBreakMode AntiDepMode, + SmallVectorImpl &CriticalPathRCs); - /// StartBlock - Initialize register live-range state for scheduling in + ~SchedulePostRATDList(); + + /// startBlock - Initialize register live-range state for scheduling in /// this block. /// - void StartBlock(MachineBasicBlock *BB); + void startBlock(MachineBasicBlock *BB); + + /// Initialize the scheduler state for the next scheduling region. + virtual void enterRegion(MachineBasicBlock *bb, + MachineBasicBlock::iterator begin, + MachineBasicBlock::iterator end, + unsigned endcount); + + /// Notify that the scheduler has finished scheduling the current region. + virtual void exitRegion(); /// Schedule - Schedule the instruction range using list scheduling. /// - void Schedule(); - + void schedule(); + + void EmitSchedule(); + /// Observe - Update liveness information to account for the current /// instruction, which will not be scheduled. /// void Observe(MachineInstr *MI, unsigned Count); - /// FinishBlock - Clean up register live-range state. + /// finishBlock - Clean up register live-range state. /// - void FinishBlock(); + void finishBlock(); /// FixupKills - Fix register kill flags that have been made /// invalid due to scheduling @@ -175,80 +171,123 @@ namespace { void FixupKills(MachineBasicBlock *MBB); private: - void ReleaseSucc(SUnit *SU, SDep *SuccEdge, bool IgnoreAntiDep); - void ReleaseSuccessors(SUnit *SU, bool IgnoreAntiDep); - void ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle, bool IgnoreAntiDep); - void ListScheduleTopDown( - AntiDepBreaker::CandidateMap *AntiDepCandidates); + void ReleaseSucc(SUnit *SU, SDep *SuccEdge); + void ReleaseSuccessors(SUnit *SU); + void ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle); + void ListScheduleTopDown(); void StartBlockForKills(MachineBasicBlock *BB); - + // ToggleKillFlag - Toggle a register operand kill flag. Other // adjustments may be made to the instruction if necessary. Return // true if the operand has been deleted, false if not. bool ToggleKillFlag(MachineInstr *MI, MachineOperand &MO); + + void dumpSchedule() const; }; } -/// isSchedulingBoundary - Test if the given instruction should be -/// considered a scheduling boundary. This primarily includes labels -/// and terminators. -/// -static bool isSchedulingBoundary(const MachineInstr *MI, - const MachineFunction &MF) { - // Terminators and labels can't be scheduled around. - if (MI->getDesc().isTerminator() || MI->isLabel()) - return true; - - // Don't attempt to schedule around any instruction that modifies - // a stack-oriented pointer, as it's unlikely to be profitable. This - // saves compile time, because it doesn't require every single - // stack slot reference to depend on the instruction that does the - // modification. - const TargetLowering &TLI = *MF.getTarget().getTargetLowering(); - if (MI->modifiesRegister(TLI.getStackPointerRegisterToSaveRestore())) - return true; +char &llvm::PostRASchedulerID = PostRAScheduler::ID; + +INITIALIZE_PASS(PostRAScheduler, "post-RA-sched", + "Post RA top-down list latency scheduler", false, false) + +SchedulePostRATDList::SchedulePostRATDList( + MachineFunction &MF, MachineLoopInfo &MLI, MachineDominatorTree &MDT, + AliasAnalysis *AA, const RegisterClassInfo &RCI, + TargetSubtargetInfo::AntiDepBreakMode AntiDepMode, + SmallVectorImpl &CriticalPathRCs) + : ScheduleDAGInstrs(MF, MLI, MDT, /*IsPostRA=*/true), AA(AA), + LiveRegs(TRI->getNumRegs()) +{ + const TargetMachine &TM = MF.getTarget(); + const InstrItineraryData *InstrItins = TM.getInstrItineraryData(); + HazardRec = + TM.getInstrInfo()->CreateTargetPostRAHazardRecognizer(InstrItins, this); + + assert((AntiDepMode == TargetSubtargetInfo::ANTIDEP_NONE || + MRI.tracksLiveness()) && + "Live-ins must be accurate for anti-dependency breaking"); + AntiDepBreak = + ((AntiDepMode == TargetSubtargetInfo::ANTIDEP_ALL) ? + (AntiDepBreaker *)new AggressiveAntiDepBreaker(MF, RCI, CriticalPathRCs) : + ((AntiDepMode == TargetSubtargetInfo::ANTIDEP_CRITICAL) ? + (AntiDepBreaker *)new CriticalAntiDepBreaker(MF, RCI) : NULL)); +} - return false; +SchedulePostRATDList::~SchedulePostRATDList() { + delete HazardRec; + delete AntiDepBreak; +} + +/// Initialize state associated with the next scheduling region. +void SchedulePostRATDList::enterRegion(MachineBasicBlock *bb, + MachineBasicBlock::iterator begin, + MachineBasicBlock::iterator end, + unsigned endcount) { + ScheduleDAGInstrs::enterRegion(bb, begin, end, endcount); + Sequence.clear(); +} + +/// Print the schedule before exiting the region. +void SchedulePostRATDList::exitRegion() { + DEBUG({ + dbgs() << "*** Final schedule ***\n"; + dumpSchedule(); + dbgs() << '\n'; + }); + ScheduleDAGInstrs::exitRegion(); +} + +#if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) +/// dumpSchedule - dump the scheduled Sequence. +void SchedulePostRATDList::dumpSchedule() const { + for (unsigned i = 0, e = Sequence.size(); i != e; i++) { + if (SUnit *SU = Sequence[i]) + SU->dump(this); + else + dbgs() << "**** NOOP ****\n"; + } } +#endif bool PostRAScheduler::runOnMachineFunction(MachineFunction &Fn) { - AA = &getAnalysis(); + TII = Fn.getTarget().getInstrInfo(); + MachineLoopInfo &MLI = getAnalysis(); + MachineDominatorTree &MDT = getAnalysis(); + AliasAnalysis *AA = &getAnalysis(); + TargetPassConfig *PassConfig = &getAnalysis(); + + RegClassInfo.runOnMachineFunction(Fn); // Check for explicit enable/disable of post-ra scheduling. - TargetSubtarget::AntiDepBreakMode AntiDepMode = TargetSubtarget::ANTIDEP_NONE; - TargetSubtarget::ExcludedRCVector ExcludedRCs; + TargetSubtargetInfo::AntiDepBreakMode AntiDepMode = + TargetSubtargetInfo::ANTIDEP_NONE; + SmallVector CriticalPathRCs; if (EnablePostRAScheduler.getPosition() > 0) { if (!EnablePostRAScheduler) return false; } else { // Check that post-RA scheduling is enabled for this target. - const TargetSubtarget &ST = Fn.getTarget().getSubtarget(); - if (!ST.enablePostRAScheduler(OptLevel, AntiDepMode, ExcludedRCs)) + // This may upgrade the AntiDepMode. + const TargetSubtargetInfo &ST = Fn.getTarget().getSubtarget(); + if (!ST.enablePostRAScheduler(PassConfig->getOptLevel(), AntiDepMode, + CriticalPathRCs)) return false; } // Check for antidep breaking override... if (EnableAntiDepBreaking.getPosition() > 0) { - AntiDepMode = (EnableAntiDepBreaking == "all") ? TargetSubtarget::ANTIDEP_ALL : - (EnableAntiDepBreaking == "critical") ? TargetSubtarget::ANTIDEP_CRITICAL : - TargetSubtarget::ANTIDEP_NONE; + AntiDepMode = (EnableAntiDepBreaking == "all") + ? TargetSubtargetInfo::ANTIDEP_ALL + : ((EnableAntiDepBreaking == "critical") + ? TargetSubtargetInfo::ANTIDEP_CRITICAL + : TargetSubtargetInfo::ANTIDEP_NONE); } - DEBUG(errs() << "PostRAScheduler\n"); + DEBUG(dbgs() << "PostRAScheduler\n"); - const MachineLoopInfo &MLI = getAnalysis(); - const MachineDominatorTree &MDT = getAnalysis(); - const InstrItineraryData &InstrItins = Fn.getTarget().getInstrItineraryData(); - ScheduleHazardRecognizer *HR = EnablePostRAHazardAvoidance ? - (ScheduleHazardRecognizer *)new ExactHazardRecognizer(InstrItins) : - (ScheduleHazardRecognizer *)new SimpleHazardRecognizer(); - AntiDepBreaker *ADB = - ((AntiDepMode == TargetSubtarget::ANTIDEP_ALL) ? - (AntiDepBreaker *)new AggressiveAntiDepBreaker(Fn, ExcludedRCs) : - ((AntiDepMode == TargetSubtarget::ANTIDEP_CRITICAL) ? - (AntiDepBreaker *)new CriticalAntiDepBreaker(Fn) : NULL)); - - SchedulePostRATDList Scheduler(Fn, MLI, MDT, HR, ADB, AA); + SchedulePostRATDList Scheduler(Fn, MLI, MDT, AA, RegClassInfo, AntiDepMode, + CriticalPathRCs); // Loop over all of the basic blocks for (MachineFunction::iterator MBB = Fn.begin(), MBBe = Fn.end(); @@ -259,55 +298,61 @@ bool PostRAScheduler::runOnMachineFunction(MachineFunction &Fn) { static int bbcnt = 0; if (bbcnt++ % DebugDiv != DebugMod) continue; - errs() << "*** DEBUG scheduling " << Fn.getFunction()->getNameStr() << - ":BB#" << MBB->getNumber() << " ***\n"; + dbgs() << "*** DEBUG scheduling " << Fn.getName() + << ":BB#" << MBB->getNumber() << " ***\n"; } #endif // Initialize register live-range state for scheduling in this block. - Scheduler.StartBlock(MBB); + Scheduler.startBlock(MBB); // Schedule each sequence of instructions not interrupted by a label // or anything else that effectively needs to shut down scheduling. MachineBasicBlock::iterator Current = MBB->end(); unsigned Count = MBB->size(), CurrentCount = Count; for (MachineBasicBlock::iterator I = Current; I != MBB->begin(); ) { - MachineInstr *MI = prior(I); - if (isSchedulingBoundary(MI, Fn)) { - Scheduler.Run(MBB, I, Current, CurrentCount); - Scheduler.EmitSchedule(0); + MachineInstr *MI = llvm::prior(I); + // Calls are not scheduling boundaries before register allocation, but + // post-ra we don't gain anything by scheduling across calls since we + // don't need to worry about register pressure. + if (MI->isCall() || TII->isSchedulingBoundary(MI, MBB, Fn)) { + Scheduler.enterRegion(MBB, I, Current, CurrentCount); + Scheduler.schedule(); + Scheduler.exitRegion(); + Scheduler.EmitSchedule(); Current = MI; CurrentCount = Count - 1; Scheduler.Observe(MI, CurrentCount); } I = MI; --Count; + if (MI->isBundle()) + Count -= MI->getBundleSize(); } assert(Count == 0 && "Instruction count mismatch!"); assert((MBB->begin() == Current || CurrentCount != 0) && "Instruction count mismatch!"); - Scheduler.Run(MBB, MBB->begin(), Current, CurrentCount); - Scheduler.EmitSchedule(0); + Scheduler.enterRegion(MBB, MBB->begin(), Current, CurrentCount); + Scheduler.schedule(); + Scheduler.exitRegion(); + Scheduler.EmitSchedule(); // Clean up register live-range state. - Scheduler.FinishBlock(); + Scheduler.finishBlock(); // Update register kills Scheduler.FixupKills(MBB); } - delete HR; - delete ADB; - return true; } - + /// StartBlock - Initialize register live-range state for scheduling in /// this block. /// -void SchedulePostRATDList::StartBlock(MachineBasicBlock *BB) { +void SchedulePostRATDList::startBlock(MachineBasicBlock *BB) { // Call the superclass. - ScheduleDAGInstrs::StartBlock(BB); + ScheduleDAGInstrs::startBlock(BB); // Reset the hazard recognizer and anti-dep breaker. HazardRec->Reset(); @@ -317,64 +362,35 @@ void SchedulePostRATDList::StartBlock(MachineBasicBlock *BB) { /// Schedule - Schedule the instruction range using list scheduling. /// -void SchedulePostRATDList::Schedule() { +void SchedulePostRATDList::schedule() { // Build the scheduling graph. - BuildSchedGraph(AA); + buildSchedGraph(AA); if (AntiDepBreak != NULL) { - AntiDepBreaker::CandidateMap AntiDepCandidates; - const bool NeedCandidates = AntiDepBreak->NeedCandidates(); - - for (unsigned i = 0, Trials = AntiDepBreak->GetMaxTrials(); - i < Trials; ++i) { - DEBUG(errs() << "\n********** Break Anti-Deps, Trial " << - i << " **********\n"); - - // If candidates are required, then schedule forward ignoring - // anti-dependencies to collect the candidate operands for - // anti-dependence breaking. The candidates will be the def - // operands for the anti-dependencies that if broken would allow - // an improved schedule - if (NeedCandidates) { - DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su) - SUnits[su].dumpAll(this)); - - AntiDepCandidates.clear(); - AvailableQueue.initNodes(SUnits); - ListScheduleTopDown(&AntiDepCandidates); - AvailableQueue.releaseState(); - } - - unsigned Broken = - AntiDepBreak->BreakAntiDependencies(SUnits, AntiDepCandidates, - Begin, InsertPos, InsertPosIndex); + unsigned Broken = + AntiDepBreak->BreakAntiDependencies(SUnits, RegionBegin, RegionEnd, + EndIndex, DbgValues); + if (Broken != 0) { // We made changes. Update the dependency graph. // Theoretically we could update the graph in place: // When a live range is changed to use a different register, remove // the def's anti-dependence *and* output-dependence edges due to // that register, and add new anti-dependence and output-dependence // edges based on the next live range of the register. - if ((Broken != 0) || NeedCandidates) { - SUnits.clear(); - Sequence.clear(); - EntrySU = SUnit(); - ExitSU = SUnit(); - BuildSchedGraph(AA); - } + ScheduleDAG::clearDAG(); + buildSchedGraph(AA); NumFixedAnti += Broken; - if (Broken == 0) - break; } } - DEBUG(errs() << "********** List Scheduling **********\n"); + DEBUG(dbgs() << "********** List Scheduling **********\n"); DEBUG(for (unsigned su = 0, e = SUnits.size(); su != e; ++su) SUnits[su].dumpAll(this)); AvailableQueue.initNodes(SUnits); - ListScheduleTopDown(NULL); + ListScheduleTopDown(); AvailableQueue.releaseState(); } @@ -383,53 +399,35 @@ void SchedulePostRATDList::Schedule() { /// void SchedulePostRATDList::Observe(MachineInstr *MI, unsigned Count) { if (AntiDepBreak != NULL) - AntiDepBreak->Observe(MI, Count, InsertPosIndex); + AntiDepBreak->Observe(MI, Count, EndIndex); } /// FinishBlock - Clean up register live-range state. /// -void SchedulePostRATDList::FinishBlock() { +void SchedulePostRATDList::finishBlock() { if (AntiDepBreak != NULL) AntiDepBreak->FinishBlock(); // Call the superclass. - ScheduleDAGInstrs::FinishBlock(); + ScheduleDAGInstrs::finishBlock(); } /// StartBlockForKills - Initialize register live-range state for updating kills /// void SchedulePostRATDList::StartBlockForKills(MachineBasicBlock *BB) { - // Initialize the indices to indicate that no registers are live. - std::fill(KillIndices, array_endof(KillIndices), ~0u); - - // Determine the live-out physregs for this block. - if (!BB->empty() && BB->back().getDesc().isReturn()) { - // In a return block, examine the function live-out regs. - for (MachineRegisterInfo::liveout_iterator I = MRI.liveout_begin(), - E = MRI.liveout_end(); I != E; ++I) { + // Start with no live registers. + LiveRegs.reset(); + + // Examine the live-in regs of all successors. + for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), + SE = BB->succ_end(); SI != SE; ++SI) { + for (MachineBasicBlock::livein_iterator I = (*SI)->livein_begin(), + E = (*SI)->livein_end(); I != E; ++I) { unsigned Reg = *I; - KillIndices[Reg] = BB->size(); - // Repeat, for all subregs. - for (const unsigned *Subreg = TRI->getSubRegisters(Reg); - *Subreg; ++Subreg) { - KillIndices[*Subreg] = BB->size(); - } - } - } - else { - // In a non-return block, examine the live-in regs of all successors. - for (MachineBasicBlock::succ_iterator SI = BB->succ_begin(), - SE = BB->succ_end(); SI != SE; ++SI) { - for (MachineBasicBlock::livein_iterator I = (*SI)->livein_begin(), - E = (*SI)->livein_end(); I != E; ++I) { - unsigned Reg = *I; - KillIndices[Reg] = BB->size(); - // Repeat, for all subregs. - for (const unsigned *Subreg = TRI->getSubRegisters(Reg); - *Subreg; ++Subreg) { - KillIndices[*Subreg] = BB->size(); - } - } + // Repeat, for reg and all subregs. + for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true); + SubRegs.isValid(); ++SubRegs) + LiveRegs.set(*SubRegs); } } } @@ -441,9 +439,9 @@ bool SchedulePostRATDList::ToggleKillFlag(MachineInstr *MI, MO.setIsKill(true); return false; } - + // If MO itself is live, clear the kill flag... - if (KillIndices[MO.getReg()] != ~0u) { + if (LiveRegs.test(MO.getReg())) { MO.setIsKill(false); return false; } @@ -453,14 +451,10 @@ bool SchedulePostRATDList::ToggleKillFlag(MachineInstr *MI, MO.setIsKill(false); bool AllDead = true; const unsigned SuperReg = MO.getReg(); - for (const unsigned *Subreg = TRI->getSubRegisters(SuperReg); - *Subreg; ++Subreg) { - if (KillIndices[*Subreg] != ~0u) { - MI->addOperand(MachineOperand::CreateReg(*Subreg, - true /*IsDef*/, - true /*IsImp*/, - false /*IsKill*/, - false /*IsDead*/)); + MachineInstrBuilder MIB(MF, MI); + for (MCSubRegIterator SubRegs(SuperReg, TRI); SubRegs.isValid(); ++SubRegs) { + if (LiveRegs.test(*SubRegs)) { + MIB.addReg(*SubRegs, RegState::ImplicitDefine); AllDead = false; } } @@ -474,57 +468,56 @@ bool SchedulePostRATDList::ToggleKillFlag(MachineInstr *MI, /// incorrect by instruction reordering. /// void SchedulePostRATDList::FixupKills(MachineBasicBlock *MBB) { - DEBUG(errs() << "Fixup kills for BB#" << MBB->getNumber() << '\n'); + DEBUG(dbgs() << "Fixup kills for BB#" << MBB->getNumber() << '\n'); - std::set killedRegs; - BitVector ReservedRegs = TRI->getReservedRegs(MF); + BitVector killedRegs(TRI->getNumRegs()); StartBlockForKills(MBB); - + // Examine block from end to start... unsigned Count = MBB->size(); for (MachineBasicBlock::iterator I = MBB->end(), E = MBB->begin(); I != E; --Count) { MachineInstr *MI = --I; + if (MI->isDebugValue()) + continue; // Update liveness. Registers that are defed but not used in this // instruction are now dead. Mark register and all subregs as they // are completely defined. for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { MachineOperand &MO = MI->getOperand(i); + if (MO.isRegMask()) + LiveRegs.clearBitsNotInMask(MO.getRegMask()); if (!MO.isReg()) continue; unsigned Reg = MO.getReg(); if (Reg == 0) continue; if (!MO.isDef()) continue; // Ignore two-addr defs. if (MI->isRegTiedToUseOperand(i)) continue; - - KillIndices[Reg] = ~0u; - - // Repeat for all subregs. - for (const unsigned *Subreg = TRI->getSubRegisters(Reg); - *Subreg; ++Subreg) { - KillIndices[*Subreg] = ~0u; - } + + // Repeat for reg and all subregs. + for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true); + SubRegs.isValid(); ++SubRegs) + LiveRegs.reset(*SubRegs); } // Examine all used registers and set/clear kill flag. When a // register is used multiple times we only set the kill flag on // the first use. - killedRegs.clear(); + killedRegs.reset(); for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { MachineOperand &MO = MI->getOperand(i); if (!MO.isReg() || !MO.isUse()) continue; unsigned Reg = MO.getReg(); - if ((Reg == 0) || ReservedRegs.test(Reg)) continue; + if ((Reg == 0) || MRI.isReserved(Reg)) continue; bool kill = false; - if (killedRegs.find(Reg) == killedRegs.end()) { + if (!killedRegs.test(Reg)) { kill = true; // A register is not killed if any subregs are live... - for (const unsigned *Subreg = TRI->getSubRegisters(Reg); - *Subreg; ++Subreg) { - if (KillIndices[*Subreg] != ~0u) { + for (MCSubRegIterator SubRegs(Reg, TRI); SubRegs.isValid(); ++SubRegs) { + if (LiveRegs.test(*SubRegs)) { kill = false; break; } @@ -533,36 +526,30 @@ void SchedulePostRATDList::FixupKills(MachineBasicBlock *MBB) { // If subreg is not live, then register is killed if it became // live in this instruction if (kill) - kill = (KillIndices[Reg] == ~0u); + kill = !LiveRegs.test(Reg); } - + if (MO.isKill() != kill) { - bool removed = ToggleKillFlag(MI, MO); - if (removed) { - DEBUG(errs() << "Fixed in "); - } else { - DEBUG(errs() << "Fixed " << MO << " in "); - } + DEBUG(dbgs() << "Fixing " << MO << " in "); + // Warning: ToggleKillFlag may invalidate MO. + ToggleKillFlag(MI, MO); DEBUG(MI->dump()); } - - killedRegs.insert(Reg); + + killedRegs.set(Reg); } - + // Mark any used register (that is not using undef) and subregs as // now live... for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) { MachineOperand &MO = MI->getOperand(i); if (!MO.isReg() || !MO.isUse() || MO.isUndef()) continue; unsigned Reg = MO.getReg(); - if ((Reg == 0) || ReservedRegs.test(Reg)) continue; + if ((Reg == 0) || MRI.isReserved(Reg)) continue; - KillIndices[Reg] = Count; - - for (const unsigned *Subreg = TRI->getSubRegisters(Reg); - *Subreg; ++Subreg) { - KillIndices[*Subreg] = Count; - } + for (MCSubRegIterator SubRegs(Reg, TRI, /*IncludeSelf=*/true); + SubRegs.isValid(); ++SubRegs) + LiveRegs.set(*SubRegs); } } } @@ -572,27 +559,35 @@ void SchedulePostRATDList::FixupKills(MachineBasicBlock *MBB) { //===----------------------------------------------------------------------===// /// ReleaseSucc - Decrement the NumPredsLeft count of a successor. Add it to -/// the PendingQueue if the count reaches zero. Also update its cycle bound. -void SchedulePostRATDList::ReleaseSucc(SUnit *SU, SDep *SuccEdge, - bool IgnoreAntiDep) { +/// the PendingQueue if the count reaches zero. +void SchedulePostRATDList::ReleaseSucc(SUnit *SU, SDep *SuccEdge) { SUnit *SuccSU = SuccEdge->getSUnit(); + if (SuccEdge->isWeak()) { + --SuccSU->WeakPredsLeft; + return; + } #ifndef NDEBUG if (SuccSU->NumPredsLeft == 0) { - errs() << "*** Scheduling failed! ***\n"; + dbgs() << "*** Scheduling failed! ***\n"; SuccSU->dump(this); - errs() << " has been released too many times!\n"; + dbgs() << " has been released too many times!\n"; llvm_unreachable(0); } #endif --SuccSU->NumPredsLeft; - // Compute how many cycles it will be before this actually becomes - // available. This is the max of the start time of all predecessors plus - // their latencies. - SuccSU->setDepthToAtLeast(SU->getDepth(IgnoreAntiDep) + - SuccEdge->getLatency(), IgnoreAntiDep); - + // Standard scheduler algorithms will recompute the depth of the successor + // here as such: + // SuccSU->setDepthToAtLeast(SU->getDepth() + SuccEdge->getLatency()); + // + // However, we lazily compute node depth instead. Note that + // ScheduleNodeTopDown has already updated the depth of this node which causes + // all descendents to be marked dirty. Setting the successor depth explicitly + // here would cause depth to be recomputed for all its ancestors. If the + // successor is not yet ready (because of a transitively redundant edge) then + // this causes depth computation to be quadratic in the size of the DAG. + // If all the node's predecessors are scheduled, this node is ready // to be scheduled. Ignore the special ExitSU node. if (SuccSU->NumPredsLeft == 0 && SuccSU != &ExitSU) @@ -600,73 +595,48 @@ void SchedulePostRATDList::ReleaseSucc(SUnit *SU, SDep *SuccEdge, } /// ReleaseSuccessors - Call ReleaseSucc on each of SU's successors. -void SchedulePostRATDList::ReleaseSuccessors(SUnit *SU, bool IgnoreAntiDep) { +void SchedulePostRATDList::ReleaseSuccessors(SUnit *SU) { for (SUnit::succ_iterator I = SU->Succs.begin(), E = SU->Succs.end(); I != E; ++I) { - if (IgnoreAntiDep && (I->getKind() == SDep::Anti)) continue; - ReleaseSucc(SU, &*I, IgnoreAntiDep); + ReleaseSucc(SU, &*I); } } /// ScheduleNodeTopDown - Add the node to the schedule. Decrement the pending /// count of its successors. If a successor pending count is zero, add it to /// the Available queue. -void SchedulePostRATDList::ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle, - bool IgnoreAntiDep) { - DEBUG(errs() << "*** Scheduling [" << CurCycle << "]: "); +void SchedulePostRATDList::ScheduleNodeTopDown(SUnit *SU, unsigned CurCycle) { + DEBUG(dbgs() << "*** Scheduling [" << CurCycle << "]: "); DEBUG(SU->dump(this)); - + Sequence.push_back(SU); - assert(CurCycle >= SU->getDepth(IgnoreAntiDep) && + assert(CurCycle >= SU->getDepth() && "Node scheduled above its depth!"); - SU->setDepthToAtLeast(CurCycle, IgnoreAntiDep); + SU->setDepthToAtLeast(CurCycle); - ReleaseSuccessors(SU, IgnoreAntiDep); + ReleaseSuccessors(SU); SU->isScheduled = true; - AvailableQueue.ScheduledNode(SU); + AvailableQueue.scheduledNode(SU); } /// ListScheduleTopDown - The main loop of list scheduling for top-down /// schedulers. -void SchedulePostRATDList::ListScheduleTopDown( - AntiDepBreaker::CandidateMap *AntiDepCandidates) { +void SchedulePostRATDList::ListScheduleTopDown() { unsigned CurCycle = 0; - const bool IgnoreAntiDep = (AntiDepCandidates != NULL); - + // We're scheduling top-down but we're visiting the regions in // bottom-up order, so we don't know the hazards at the start of a // region. So assume no hazards (this should usually be ok as most // blocks are a single region). HazardRec->Reset(); - // If ignoring anti-dependencies, the Schedule DAG still has Anti - // dep edges, but we ignore them for scheduling purposes - AvailableQueue.setIgnoreAntiDep(IgnoreAntiDep); - // Release any successors of the special Entry node. - ReleaseSuccessors(&EntrySU, IgnoreAntiDep); + ReleaseSuccessors(&EntrySU); - // Add all leaves to Available queue. If ignoring antideps we also - // adjust the predecessor count for each node to not include antidep - // edges. + // Add all leaves to Available queue. for (unsigned i = 0, e = SUnits.size(); i != e; ++i) { // It is available if it has no predecessors. - bool available = SUnits[i].Preds.empty(); - // If we are ignoring anti-dependencies then a node that has only - // anti-dep predecessors is available. - if (!available && IgnoreAntiDep) { - available = true; - for (SUnit::const_pred_iterator I = SUnits[i].Preds.begin(), - E = SUnits[i].Preds.end(); I != E; ++I) { - if (I->getKind() != SDep::Anti) { - available = false; - } else { - SUnits[i].NumPredsLeft -= 1; - } - } - } - - if (available) { + if (!SUnits[i].NumPredsLeft && !SUnits[i].isAvailable) { AvailableQueue.push(&SUnits[i]); SUnits[i].isAvailable = true; } @@ -685,23 +655,17 @@ void SchedulePostRATDList::ListScheduleTopDown( // so, add them to the available queue. unsigned MinDepth = ~0u; for (unsigned i = 0, e = PendingQueue.size(); i != e; ++i) { - if (PendingQueue[i]->getDepth(IgnoreAntiDep) <= CurCycle) { + if (PendingQueue[i]->getDepth() <= CurCycle) { AvailableQueue.push(PendingQueue[i]); PendingQueue[i]->isAvailable = true; PendingQueue[i] = PendingQueue.back(); PendingQueue.pop_back(); --i; --e; - } else if (PendingQueue[i]->getDepth(IgnoreAntiDep) < MinDepth) - MinDepth = PendingQueue[i]->getDepth(IgnoreAntiDep); + } else if (PendingQueue[i]->getDepth() < MinDepth) + MinDepth = PendingQueue[i]->getDepth(); } - DEBUG(errs() << "\n*** Examining Available\n"; - LatencyPriorityQueue q = AvailableQueue; - while (!q.empty()) { - SUnit *su = q.pop(); - errs() << "Height " << su->getHeight(IgnoreAntiDep) << ": "; - su->dump(this); - }); + DEBUG(dbgs() << "\n*** Examining Available\n"; AvailableQueue.dump(this)); SUnit *FoundSUnit = 0; bool HasNoopHazards = false; @@ -709,7 +673,7 @@ void SchedulePostRATDList::ListScheduleTopDown( SUnit *CurSUnit = AvailableQueue.pop(); ScheduleHazardRecognizer::HazardType HT = - HazardRec->getHazardType(CurSUnit); + HazardRec->getHazardType(CurSUnit, 0/*no stalls*/); if (HT == ScheduleHazardRecognizer::NoHazard) { FoundSUnit = CurSUnit; break; @@ -729,59 +693,34 @@ void SchedulePostRATDList::ListScheduleTopDown( // If we found a node to schedule... if (FoundSUnit) { - // If we are ignoring anti-dependencies and the SUnit we are - // scheduling has an antidep predecessor that has not been - // scheduled, then we will need to break that antidep if we want - // to get this schedule when not ignoring anti-dependencies. - if (IgnoreAntiDep) { - AntiDepBreaker::AntiDepRegVector AntiDepRegs; - for (SUnit::const_pred_iterator I = FoundSUnit->Preds.begin(), - E = FoundSUnit->Preds.end(); I != E; ++I) { - if ((I->getKind() == SDep::Anti) && !I->getSUnit()->isScheduled) - AntiDepRegs.push_back(I->getReg()); - } - - if (AntiDepRegs.size() > 0) { - DEBUG(errs() << "*** AntiDep Candidate: "); - DEBUG(FoundSUnit->dump(this)); - AntiDepCandidates->insert( - AntiDepBreaker::CandidateMap::value_type(FoundSUnit, AntiDepRegs)); - } - } - // ... schedule the node... - ScheduleNodeTopDown(FoundSUnit, CurCycle, IgnoreAntiDep); + ScheduleNodeTopDown(FoundSUnit, CurCycle); HazardRec->EmitInstruction(FoundSUnit); CycleHasInsts = true; - - // If we are using the target-specific hazards, then don't - // advance the cycle time just because we schedule a node. If - // the target allows it we can schedule multiple nodes in the - // same cycle. - if (!EnablePostRAHazardAvoidance) { - if (FoundSUnit->Latency) // Don't increment CurCycle for pseudo-ops! - ++CurCycle; + if (HazardRec->atIssueLimit()) { + DEBUG(dbgs() << "*** Max instructions per cycle " << CurCycle << '\n'); + HazardRec->AdvanceCycle(); + ++CurCycle; + CycleHasInsts = false; } } else { if (CycleHasInsts) { - DEBUG(errs() << "*** Finished cycle " << CurCycle << '\n'); + DEBUG(dbgs() << "*** Finished cycle " << CurCycle << '\n'); HazardRec->AdvanceCycle(); } else if (!HasNoopHazards) { // Otherwise, we have a pipeline stall, but no other problem, // just advance the current cycle and try again. - DEBUG(errs() << "*** Stall in cycle " << CurCycle << '\n'); + DEBUG(dbgs() << "*** Stall in cycle " << CurCycle << '\n'); HazardRec->AdvanceCycle(); - if (!IgnoreAntiDep) - ++NumStalls; + ++NumStalls; } else { // Otherwise, we have no instructions to issue and we have instructions // that will fault if we don't do this right. This is the case for // processors without pipeline interlocks and other cases. - DEBUG(errs() << "*** Emitting noop in cycle " << CurCycle << '\n'); + DEBUG(dbgs() << "*** Emitting noop in cycle " << CurCycle << '\n'); HazardRec->EmitNoop(); Sequence.push_back(0); // NULL here means noop - if (!IgnoreAntiDep) - ++NumNoops; + ++NumNoops; } ++CurCycle; @@ -790,14 +729,46 @@ void SchedulePostRATDList::ListScheduleTopDown( } #ifndef NDEBUG - VerifySchedule(/*isBottomUp=*/false); -#endif + unsigned ScheduledNodes = VerifyScheduledDAG(/*isBottomUp=*/false); + unsigned Noops = 0; + for (unsigned i = 0, e = Sequence.size(); i != e; ++i) + if (!Sequence[i]) + ++Noops; + assert(Sequence.size() - Noops == ScheduledNodes && + "The number of nodes scheduled doesn't match the expected number!"); +#endif // NDEBUG } -//===----------------------------------------------------------------------===// -// Public Constructor Functions -//===----------------------------------------------------------------------===// +// EmitSchedule - Emit the machine code in scheduled order. +void SchedulePostRATDList::EmitSchedule() { + RegionBegin = RegionEnd; + + // If first instruction was a DBG_VALUE then put it back. + if (FirstDbgValue) + BB->splice(RegionEnd, BB, FirstDbgValue); + + // Then re-insert them according to the given schedule. + for (unsigned i = 0, e = Sequence.size(); i != e; i++) { + if (SUnit *SU = Sequence[i]) + BB->splice(RegionEnd, BB, SU->getInstr()); + else + // Null SUnit* is a noop. + TII->insertNoop(*BB, RegionEnd); + + // Update the Begin iterator, as the first instruction in the block + // may have been scheduled later. + if (i == 0) + RegionBegin = prior(RegionEnd); + } -FunctionPass *llvm::createPostRAScheduler(CodeGenOpt::Level OptLevel) { - return new PostRAScheduler(OptLevel); + // Reinsert any remaining debug_values. + for (std::vector >::iterator + DI = DbgValues.end(), DE = DbgValues.begin(); DI != DE; --DI) { + std::pair P = *prior(DI); + MachineInstr *DbgValue = P.first; + MachineBasicBlock::iterator OrigPrivMI = P.second; + BB->splice(++OrigPrivMI, BB, DbgValue); + } + DbgValues.clear(); + FirstDbgValue = NULL; }