X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FCodeGen%2FMachineTraceMetrics.cpp;h=ff7f94ea9423613be26fcd42254736104ba57ad0;hb=eb3430cfbde7c28586a0d55cff8a88dbb3aa348f;hp=1baab9b1edaea36198a4508fbaa3c9c6c9e6e263;hpb=5f8e8bd656bb174b3e22c0e56ce3d1eb958ac2e2;p=oota-llvm.git diff --git a/lib/CodeGen/MachineTraceMetrics.cpp b/lib/CodeGen/MachineTraceMetrics.cpp index 1baab9b1eda..ff7f94ea942 100644 --- a/lib/CodeGen/MachineTraceMetrics.cpp +++ b/lib/CodeGen/MachineTraceMetrics.cpp @@ -7,22 +7,26 @@ // //===----------------------------------------------------------------------===// -#define DEBUG_TYPE "early-ifcvt" -#include "MachineTraceMetrics.h" +#include "llvm/CodeGen/MachineTraceMetrics.h" +#include "llvm/ADT/PostOrderIterator.h" +#include "llvm/ADT/SparseSet.h" #include "llvm/CodeGen/MachineBasicBlock.h" #include "llvm/CodeGen/MachineBranchProbabilityInfo.h" #include "llvm/CodeGen/MachineLoopInfo.h" #include "llvm/CodeGen/MachineRegisterInfo.h" #include "llvm/CodeGen/Passes.h" -#include "llvm/Target/TargetInstrInfo.h" -#include "llvm/Target/TargetRegisterInfo.h" +#include "llvm/MC/MCSubtargetInfo.h" #include "llvm/Support/Debug.h" +#include "llvm/Support/Format.h" #include "llvm/Support/raw_ostream.h" -#include "llvm/ADT/PostOrderIterator.h" -#include "llvm/ADT/SparseSet.h" +#include "llvm/Target/TargetInstrInfo.h" +#include "llvm/Target/TargetRegisterInfo.h" +#include "llvm/Target/TargetSubtargetInfo.h" using namespace llvm; +#define DEBUG_TYPE "machine-trace-metrics" + char MachineTraceMetrics::ID = 0; char &llvm::MachineTraceMetricsID = MachineTraceMetrics::ID; @@ -34,8 +38,9 @@ INITIALIZE_PASS_END(MachineTraceMetrics, "machine-trace-metrics", "Machine Trace Metrics", false, true) MachineTraceMetrics::MachineTraceMetrics() - : MachineFunctionPass(ID), MF(0), TII(0), TRI(0), MRI(0), Loops(0) { - std::fill(Ensembles, array_endof(Ensembles), (Ensemble*)0); + : MachineFunctionPass(ID), MF(nullptr), TII(nullptr), TRI(nullptr), + MRI(nullptr), Loops(nullptr) { + std::fill(std::begin(Ensembles), std::end(Ensembles), nullptr); } void MachineTraceMetrics::getAnalysisUsage(AnalysisUsage &AU) const { @@ -49,19 +54,23 @@ bool MachineTraceMetrics::runOnMachineFunction(MachineFunction &Func) { MF = &Func; TII = MF->getTarget().getInstrInfo(); TRI = MF->getTarget().getRegisterInfo(); - ItinData = MF->getTarget().getInstrItineraryData(); MRI = &MF->getRegInfo(); Loops = &getAnalysis(); + const TargetSubtargetInfo &ST = + MF->getTarget().getSubtarget(); + SchedModel.init(*ST.getSchedModel(), &ST, TII); BlockInfo.resize(MF->getNumBlockIDs()); + ProcResourceCycles.resize(MF->getNumBlockIDs() * + SchedModel.getNumProcResourceKinds()); return false; } void MachineTraceMetrics::releaseMemory() { - MF = 0; + MF = nullptr; BlockInfo.clear(); for (unsigned i = 0; i != TS_NumStrategies; ++i) { delete Ensembles[i]; - Ensembles[i] = 0; + Ensembles[i] = nullptr; } } @@ -81,9 +90,13 @@ MachineTraceMetrics::getResources(const MachineBasicBlock *MBB) { return FBI; // Compute resource usage in the block. - // FIXME: Compute per-functional unit counts. FBI->HasCalls = false; unsigned InstrCount = 0; + + // Add up per-processor resource cycles as well. + unsigned PRKinds = SchedModel.getNumProcResourceKinds(); + SmallVector PRCycles(PRKinds); + for (MachineBasicBlock::const_iterator I = MBB->begin(), E = MBB->end(); I != E; ++I) { const MachineInstr *MI = I; @@ -92,11 +105,43 @@ MachineTraceMetrics::getResources(const MachineBasicBlock *MBB) { ++InstrCount; if (MI->isCall()) FBI->HasCalls = true; + + // Count processor resources used. + if (!SchedModel.hasInstrSchedModel()) + continue; + const MCSchedClassDesc *SC = SchedModel.resolveSchedClass(MI); + if (!SC->isValid()) + continue; + + for (TargetSchedModel::ProcResIter + PI = SchedModel.getWriteProcResBegin(SC), + PE = SchedModel.getWriteProcResEnd(SC); PI != PE; ++PI) { + assert(PI->ProcResourceIdx < PRKinds && "Bad processor resource kind"); + PRCycles[PI->ProcResourceIdx] += PI->Cycles; + } } FBI->InstrCount = InstrCount; + + // Scale the resource cycles so they are comparable. + unsigned PROffset = MBB->getNumber() * PRKinds; + for (unsigned K = 0; K != PRKinds; ++K) + ProcResourceCycles[PROffset + K] = + PRCycles[K] * SchedModel.getResourceFactor(K); + return FBI; } +ArrayRef +MachineTraceMetrics::getProcResourceCycles(unsigned MBBNum) const { + assert(BlockInfo[MBBNum].hasResources() && + "getResources() must be called before getProcResourceCycles()"); + unsigned PRKinds = SchedModel.getNumProcResourceKinds(); + assert((MBBNum+1) * PRKinds <= ProcResourceCycles.size()); + return ArrayRef(ProcResourceCycles.data() + MBBNum * PRKinds, + PRKinds); +} + + //===----------------------------------------------------------------------===// // Ensemble utility functions //===----------------------------------------------------------------------===// @@ -104,6 +149,9 @@ MachineTraceMetrics::getResources(const MachineBasicBlock *MBB) { MachineTraceMetrics::Ensemble::Ensemble(MachineTraceMetrics *ct) : MTM(*ct) { BlockInfo.resize(MTM.BlockInfo.size()); + unsigned PRKinds = MTM.SchedModel.getNumProcResourceKinds(); + ProcResourceDepths.resize(MTM.BlockInfo.size() * PRKinds); + ProcResourceHeights.resize(MTM.BlockInfo.size() * PRKinds); } // Virtual destructor serves as an anchor. @@ -119,21 +167,32 @@ MachineTraceMetrics::Ensemble::getLoopFor(const MachineBasicBlock *MBB) const { void MachineTraceMetrics::Ensemble:: computeDepthResources(const MachineBasicBlock *MBB) { TraceBlockInfo *TBI = &BlockInfo[MBB->getNumber()]; + unsigned PRKinds = MTM.SchedModel.getNumProcResourceKinds(); + unsigned PROffset = MBB->getNumber() * PRKinds; // Compute resources from trace above. The top block is simple. if (!TBI->Pred) { TBI->InstrDepth = 0; TBI->Head = MBB->getNumber(); + std::fill(ProcResourceDepths.begin() + PROffset, + ProcResourceDepths.begin() + PROffset + PRKinds, 0); return; } // Compute from the block above. A post-order traversal ensures the // predecessor is always computed first. - TraceBlockInfo *PredTBI = &BlockInfo[TBI->Pred->getNumber()]; + unsigned PredNum = TBI->Pred->getNumber(); + TraceBlockInfo *PredTBI = &BlockInfo[PredNum]; assert(PredTBI->hasValidDepth() && "Trace above has not been computed yet"); const FixedBlockInfo *PredFBI = MTM.getResources(TBI->Pred); TBI->InstrDepth = PredTBI->InstrDepth + PredFBI->InstrCount; TBI->Head = PredTBI->Head; + + // Compute per-resource depths. + ArrayRef PredPRDepths = getProcResourceDepths(PredNum); + ArrayRef PredPRCycles = MTM.getProcResourceCycles(PredNum); + for (unsigned K = 0; K != PRKinds; ++K) + ProcResourceDepths[PROffset + K] = PredPRDepths[K] + PredPRCycles[K]; } // Update resource-related information in the TraceBlockInfo for MBB. @@ -141,22 +200,33 @@ computeDepthResources(const MachineBasicBlock *MBB) { void MachineTraceMetrics::Ensemble:: computeHeightResources(const MachineBasicBlock *MBB) { TraceBlockInfo *TBI = &BlockInfo[MBB->getNumber()]; + unsigned PRKinds = MTM.SchedModel.getNumProcResourceKinds(); + unsigned PROffset = MBB->getNumber() * PRKinds; // Compute resources for the current block. TBI->InstrHeight = MTM.getResources(MBB)->InstrCount; + ArrayRef PRCycles = MTM.getProcResourceCycles(MBB->getNumber()); // The trace tail is done. if (!TBI->Succ) { TBI->Tail = MBB->getNumber(); + std::copy(PRCycles.begin(), PRCycles.end(), + ProcResourceHeights.begin() + PROffset); return; } // Compute from the block below. A post-order traversal ensures the // predecessor is always computed first. - TraceBlockInfo *SuccTBI = &BlockInfo[TBI->Succ->getNumber()]; + unsigned SuccNum = TBI->Succ->getNumber(); + TraceBlockInfo *SuccTBI = &BlockInfo[SuccNum]; assert(SuccTBI->hasValidHeight() && "Trace below has not been computed yet"); TBI->InstrHeight += SuccTBI->InstrHeight; TBI->Tail = SuccTBI->Tail; + + // Compute per-resource heights. + ArrayRef SuccPRHeights = getProcResourceHeights(SuccNum); + for (unsigned K = 0; K != PRKinds; ++K) + ProcResourceHeights[PROffset + K] = SuccPRHeights[K] + PRCycles[K]; } // Check if depth resources for MBB are valid and return the TBI. @@ -165,7 +235,7 @@ const MachineTraceMetrics::TraceBlockInfo* MachineTraceMetrics::Ensemble:: getDepthResources(const MachineBasicBlock *MBB) const { const TraceBlockInfo *TBI = &BlockInfo[MBB->getNumber()]; - return TBI->hasValidDepth() ? TBI : 0; + return TBI->hasValidDepth() ? TBI : nullptr; } // Check if height resources for MBB are valid and return the TBI. @@ -174,7 +244,36 @@ const MachineTraceMetrics::TraceBlockInfo* MachineTraceMetrics::Ensemble:: getHeightResources(const MachineBasicBlock *MBB) const { const TraceBlockInfo *TBI = &BlockInfo[MBB->getNumber()]; - return TBI->hasValidHeight() ? TBI : 0; + return TBI->hasValidHeight() ? TBI : nullptr; +} + +/// Get an array of processor resource depths for MBB. Indexed by processor +/// resource kind, this array contains the scaled processor resources consumed +/// by all blocks preceding MBB in its trace. It does not include instructions +/// in MBB. +/// +/// Compare TraceBlockInfo::InstrDepth. +ArrayRef +MachineTraceMetrics::Ensemble:: +getProcResourceDepths(unsigned MBBNum) const { + unsigned PRKinds = MTM.SchedModel.getNumProcResourceKinds(); + assert((MBBNum+1) * PRKinds <= ProcResourceDepths.size()); + return ArrayRef(ProcResourceDepths.data() + MBBNum * PRKinds, + PRKinds); +} + +/// Get an array of processor resource heights for MBB. Indexed by processor +/// resource kind, this array contains the scaled processor resources consumed +/// by this block and all blocks following it in its trace. +/// +/// Compare TraceBlockInfo::InstrHeight. +ArrayRef +MachineTraceMetrics::Ensemble:: +getProcResourceHeights(unsigned MBBNum) const { + unsigned PRKinds = MTM.SchedModel.getNumProcResourceKinds(); + assert((MBBNum+1) * PRKinds <= ProcResourceHeights.size()); + return ArrayRef(ProcResourceHeights.data() + MBBNum * PRKinds, + PRKinds); } //===----------------------------------------------------------------------===// @@ -205,9 +304,9 @@ static bool isExitingLoop(const MachineLoop *From, const MachineLoop *To) { // instructions. namespace { class MinInstrCountEnsemble : public MachineTraceMetrics::Ensemble { - const char *getName() const { return "MinInstr"; } - const MachineBasicBlock *pickTracePred(const MachineBasicBlock*); - const MachineBasicBlock *pickTraceSucc(const MachineBasicBlock*); + const char *getName() const override { return "MinInstr"; } + const MachineBasicBlock *pickTracePred(const MachineBasicBlock*) override; + const MachineBasicBlock *pickTraceSucc(const MachineBasicBlock*) override; public: MinInstrCountEnsemble(MachineTraceMetrics *mtm) @@ -219,20 +318,22 @@ public: const MachineBasicBlock* MinInstrCountEnsemble::pickTracePred(const MachineBasicBlock *MBB) { if (MBB->pred_empty()) - return 0; + return nullptr; const MachineLoop *CurLoop = getLoopFor(MBB); // Don't leave loops, and never follow back-edges. if (CurLoop && MBB == CurLoop->getHeader()) - return 0; + return nullptr; unsigned CurCount = MTM.getResources(MBB)->InstrCount; - const MachineBasicBlock *Best = 0; + const MachineBasicBlock *Best = nullptr; unsigned BestDepth = 0; for (MachineBasicBlock::const_pred_iterator I = MBB->pred_begin(), E = MBB->pred_end(); I != E; ++I) { const MachineBasicBlock *Pred = *I; const MachineTraceMetrics::TraceBlockInfo *PredTBI = getDepthResources(Pred); - assert(PredTBI && "Predecessor must be visited first"); + // Ignore cycles that aren't natural loops. + if (!PredTBI) + continue; // Pick the predecessor that would give this block the smallest InstrDepth. unsigned Depth = PredTBI->InstrDepth + CurCount; if (!Best || Depth < BestDepth) @@ -245,9 +346,9 @@ MinInstrCountEnsemble::pickTracePred(const MachineBasicBlock *MBB) { const MachineBasicBlock* MinInstrCountEnsemble::pickTraceSucc(const MachineBasicBlock *MBB) { if (MBB->pred_empty()) - return 0; + return nullptr; const MachineLoop *CurLoop = getLoopFor(MBB); - const MachineBasicBlock *Best = 0; + const MachineBasicBlock *Best = nullptr; unsigned BestHeight = 0; for (MachineBasicBlock::const_succ_iterator I = MBB->succ_begin(), E = MBB->succ_end(); I != E; ++I) { @@ -260,7 +361,9 @@ MinInstrCountEnsemble::pickTraceSucc(const MachineBasicBlock *MBB) { continue; const MachineTraceMetrics::TraceBlockInfo *SuccTBI = getHeightResources(Succ); - assert(SuccTBI && "Successor must be visited first"); + // Ignore cycles that aren't natural loops. + if (!SuccTBI) + continue; // Pick the successor that would give this block the smallest InstrHeight. unsigned Height = SuccTBI->InstrHeight; if (!Best || Height < BestHeight) @@ -314,6 +417,7 @@ void MachineTraceMetrics::verifyAnalysis() const { namespace { struct LoopBounds { MutableArrayRef Blocks; + SmallPtrSet Visited; const MachineLoopInfo *Loops; bool Downward; LoopBounds(MutableArrayRef blocks, @@ -338,21 +442,19 @@ public: if (LB.Downward ? TBI.hasValidHeight() : TBI.hasValidDepth()) return false; // From is null once when To is the trace center block. - if (!From) - return true; - const MachineLoop *FromLoop = LB.Loops->getLoopFor(From); - if (!FromLoop) - return true; - // Don't follow backedges, don't leave FromLoop when going upwards. - if ((LB.Downward ? To : From) == FromLoop->getHeader()) - return false; - // Don't leave FromLoop. - if (isExitingLoop(FromLoop, LB.Loops->getLoopFor(To))) - return false; - // This is a new block. The PO traversal will compute height/depth - // resources, causing us to reject new edges to To. This only works because - // we reject back-edges, so the CFG is cycle-free. - return true; + if (From) { + if (const MachineLoop *FromLoop = LB.Loops->getLoopFor(From)) { + // Don't follow backedges, don't leave FromLoop when going upwards. + if ((LB.Downward ? To : From) == FromLoop->getHeader()) + return false; + // Don't leave FromLoop. + if (isExitingLoop(FromLoop, LB.Loops->getLoopFor(To))) + return false; + } + } + // To is a new block. Mark the block as visited in case the CFG has cycles + // that MachineLoopInfo didn't recognize as a natural loop. + return LB.Visited.insert(To); } }; } @@ -366,6 +468,7 @@ void MachineTraceMetrics::Ensemble::computeTrace(const MachineBasicBlock *MBB) { // Run an upwards post-order search for the trace start. Bounds.Downward = false; + Bounds.Visited.clear(); typedef ipo_ext_iterator UpwardPO; for (UpwardPO I = ipo_ext_begin(MBB, Bounds), E = ipo_ext_end(MBB, Bounds); I != E; ++I) { @@ -385,6 +488,7 @@ void MachineTraceMetrics::Ensemble::computeTrace(const MachineBasicBlock *MBB) { // Run a downwards post-order search for the trace end. Bounds.Downward = true; + Bounds.Visited.clear(); typedef po_ext_iterator DownwardPO; for (DownwardPO I = po_ext_begin(MBB, Bounds), E = po_ext_end(MBB, Bounds); I != E; ++I) { @@ -515,6 +619,20 @@ struct DataDep { const MachineInstr *DefMI; unsigned DefOp; unsigned UseOp; + + DataDep(const MachineInstr *DefMI, unsigned DefOp, unsigned UseOp) + : DefMI(DefMI), DefOp(DefOp), UseOp(UseOp) {} + + /// Create a DataDep from an SSA form virtual register. + DataDep(const MachineRegisterInfo *MRI, unsigned VirtReg, unsigned UseOp) + : UseOp(UseOp) { + assert(TargetRegisterInfo::isVirtualRegister(VirtReg)); + MachineRegisterInfo::def_iterator DefI = MRI->def_begin(VirtReg); + assert(!DefI.atEnd() && "Register has no defs"); + DefMI = DefI->getParent(); + DefOp = DefI.getOperandNo(); + assert((++DefI).atEnd() && "Register has multiple defs"); + } }; } @@ -535,14 +653,8 @@ static bool getDataDeps(const MachineInstr *UseMI, continue; } // Collect virtual register reads. - if (!MO->readsReg()) - continue; - MachineRegisterInfo::def_iterator DefI = MRI->def_begin(Reg); - DataDep Dep; - Dep.DefMI = &*DefI; - Dep.DefOp = DefI.getOperandNo(); - Dep.UseOp = MO.getOperandNo(); - Deps.push_back(Dep); + if (MO->readsReg()) + Deps.push_back(DataDep(MRI, Reg, MO.getOperandNo())); } return HasPhysRegs; } @@ -561,29 +673,26 @@ static void getPHIDeps(const MachineInstr *UseMI, for (unsigned i = 1; i != UseMI->getNumOperands(); i += 2) { if (UseMI->getOperand(i + 1).getMBB() == Pred) { unsigned Reg = UseMI->getOperand(i).getReg(); - assert(TargetRegisterInfo::isVirtualRegister(Reg) && "Bad PHI op"); - MachineRegisterInfo::def_iterator DefI = MRI->def_begin(Reg); - DataDep Dep; - Dep.DefMI = &*DefI; - Dep.DefOp = DefI.getOperandNo(); - Dep.UseOp = i; - Deps.push_back(Dep); + Deps.push_back(DataDep(MRI, Reg, i)); return; } } } // Keep track of physreg data dependencies by recording each live register unit. +// Associate each regunit with an instruction operand. Depending on the +// direction instructions are scanned, it could be the operand that defined the +// regunit, or the highest operand to read the regunit. namespace { struct LiveRegUnit { unsigned RegUnit; - unsigned DefOp; - const MachineInstr *DefMI; + unsigned Cycle; + const MachineInstr *MI; + unsigned Op; unsigned getSparseSetIndex() const { return RegUnit; } - LiveRegUnit(unsigned RU, const MachineInstr *MI = 0, unsigned OpNo = 0) - : RegUnit(RU), DefOp(OpNo), DefMI(MI) {} + LiveRegUnit(unsigned RU) : RegUnit(RU), Cycle(0), MI(nullptr), Op(0) {} }; } @@ -617,11 +726,7 @@ static void updatePhysDepsDownwards(const MachineInstr *UseMI, SparseSet::iterator I = RegUnits.find(*Units); if (I == RegUnits.end()) continue; - DataDep Dep; - Dep.DefMI = I->DefMI; - Dep.DefOp = I->DefOp; - Dep.UseOp = MO.getOperandNo(); - Deps.push_back(Dep); + Deps.push_back(DataDep(I->MI, I->Op, MO.getOperandNo())); break; } } @@ -638,14 +743,42 @@ static void updatePhysDepsDownwards(const MachineInstr *UseMI, for (MCRegUnitIterator Units(UseMI->getOperand(DefOp).getReg(), TRI); Units.isValid(); ++Units) { LiveRegUnit &LRU = RegUnits[*Units]; - LRU.DefMI = UseMI; - LRU.DefOp = DefOp; + LRU.MI = UseMI; + LRU.Op = DefOp; } } } - - +/// The length of the critical path through a trace is the maximum of two path +/// lengths: +/// +/// 1. The maximum height+depth over all instructions in the trace center block. +/// +/// 2. The longest cross-block dependency chain. For small blocks, it is +/// possible that the critical path through the trace doesn't include any +/// instructions in the block. +/// +/// This function computes the second number from the live-in list of the +/// center block. +unsigned MachineTraceMetrics::Ensemble:: +computeCrossBlockCriticalPath(const TraceBlockInfo &TBI) { + assert(TBI.HasValidInstrDepths && "Missing depth info"); + assert(TBI.HasValidInstrHeights && "Missing height info"); + unsigned MaxLen = 0; + for (unsigned i = 0, e = TBI.LiveIns.size(); i != e; ++i) { + const LiveInReg &LIR = TBI.LiveIns[i]; + if (!TargetRegisterInfo::isVirtualRegister(LIR.Reg)) + continue; + const MachineInstr *DefMI = MTM.MRI->getVRegDef(LIR.Reg); + // Ignore dependencies outside the current trace. + const TraceBlockInfo &DefTBI = BlockInfo[DefMI->getParent()->getNumber()]; + if (!DefTBI.isUsefulDominator(TBI)) + continue; + unsigned Len = LIR.Height + Cycles[DefMI].Depth; + MaxLen = std::max(MaxLen, Len); + } + return MaxLen; +} /// Compute instruction depths for all instructions above or in MBB in its /// trace. This assumes that the trace through MBB has already been computed. @@ -675,9 +808,28 @@ computeInstrDepths(const MachineBasicBlock *MBB) { SmallVector Deps; while (!Stack.empty()) { MBB = Stack.pop_back_val(); - DEBUG(dbgs() << "Depths for BB#" << MBB->getNumber() << ":\n"); + DEBUG(dbgs() << "\nDepths for BB#" << MBB->getNumber() << ":\n"); TraceBlockInfo &TBI = BlockInfo[MBB->getNumber()]; TBI.HasValidInstrDepths = true; + TBI.CriticalPath = 0; + + // Print out resource depths here as well. + DEBUG({ + dbgs() << format("%7u Instructions\n", TBI.InstrDepth); + ArrayRef PRDepths = getProcResourceDepths(MBB->getNumber()); + for (unsigned K = 0; K != PRDepths.size(); ++K) + if (PRDepths[K]) { + unsigned Factor = MTM.SchedModel.getResourceFactor(K); + dbgs() << format("%6uc @ ", MTM.getCycles(PRDepths[K])) + << MTM.SchedModel.getProcResource(K)->Name << " (" + << PRDepths[K]/Factor << " ops x" << Factor << ")\n"; + } + }); + + // Also compute the critical path length through MBB when possible. + if (TBI.HasValidInstrHeights) + TBI.CriticalPath = computeCrossBlockCriticalPath(TBI); + for (MachineBasicBlock::const_iterator I = MBB->begin(), E = MBB->end(); I != E; ++I) { const MachineInstr *UseMI = I; @@ -696,33 +848,396 @@ computeInstrDepths(const MachineBasicBlock *MBB) { const TraceBlockInfo&DepTBI = BlockInfo[Dep.DefMI->getParent()->getNumber()]; // Ignore dependencies from outside the current trace. - if (!DepTBI.hasValidDepth() || DepTBI.Head != TBI.Head) + if (!DepTBI.isUsefulDominator(TBI)) continue; assert(DepTBI.HasValidInstrDepths && "Inconsistent dependency"); unsigned DepCycle = Cycles.lookup(Dep.DefMI).Depth; // Add latency if DefMI is a real instruction. Transients get latency 0. if (!Dep.DefMI->isTransient()) - DepCycle += MTM.TII->computeOperandLatency(MTM.ItinData, - Dep.DefMI, Dep.DefOp, - UseMI, Dep.UseOp, - /* FindMin = */ false); + DepCycle += MTM.SchedModel + .computeOperandLatency(Dep.DefMI, Dep.DefOp, UseMI, Dep.UseOp); Cycle = std::max(Cycle, DepCycle); } // Remember the instruction depth. - Cycles[UseMI].Depth = Cycle; - DEBUG(dbgs() << Cycle << '\t' << *UseMI); + InstrCycles &MICycles = Cycles[UseMI]; + MICycles.Depth = Cycle; + + if (!TBI.HasValidInstrHeights) { + DEBUG(dbgs() << Cycle << '\t' << *UseMI); + continue; + } + // Update critical path length. + TBI.CriticalPath = std::max(TBI.CriticalPath, Cycle + MICycles.Height); + DEBUG(dbgs() << TBI.CriticalPath << '\t' << Cycle << '\t' << *UseMI); } } } +// Identify physreg dependencies for MI when scanning instructions upwards. +// Return the issue height of MI after considering any live regunits. +// Height is the issue height computed from virtual register dependencies alone. +static unsigned updatePhysDepsUpwards(const MachineInstr *MI, unsigned Height, + SparseSet &RegUnits, + const TargetSchedModel &SchedModel, + const TargetInstrInfo *TII, + const TargetRegisterInfo *TRI) { + SmallVector ReadOps; + for (ConstMIOperands MO(MI); MO.isValid(); ++MO) { + if (!MO->isReg()) + continue; + unsigned Reg = MO->getReg(); + if (!TargetRegisterInfo::isPhysicalRegister(Reg)) + continue; + if (MO->readsReg()) + ReadOps.push_back(MO.getOperandNo()); + if (!MO->isDef()) + continue; + // This is a def of Reg. Remove corresponding entries from RegUnits, and + // update MI Height to consider the physreg dependencies. + for (MCRegUnitIterator Units(Reg, TRI); Units.isValid(); ++Units) { + SparseSet::iterator I = RegUnits.find(*Units); + if (I == RegUnits.end()) + continue; + unsigned DepHeight = I->Cycle; + if (!MI->isTransient()) { + // We may not know the UseMI of this dependency, if it came from the + // live-in list. SchedModel can handle a NULL UseMI. + DepHeight += SchedModel + .computeOperandLatency(MI, MO.getOperandNo(), I->MI, I->Op); + } + Height = std::max(Height, DepHeight); + // This regunit is dead above MI. + RegUnits.erase(I); + } + } + + // Now we know the height of MI. Update any regunits read. + for (unsigned i = 0, e = ReadOps.size(); i != e; ++i) { + unsigned Reg = MI->getOperand(ReadOps[i]).getReg(); + for (MCRegUnitIterator Units(Reg, TRI); Units.isValid(); ++Units) { + LiveRegUnit &LRU = RegUnits[*Units]; + // Set the height to the highest reader of the unit. + if (LRU.Cycle <= Height && LRU.MI != MI) { + LRU.Cycle = Height; + LRU.MI = MI; + LRU.Op = ReadOps[i]; + } + } + } + + return Height; +} + + +typedef DenseMap MIHeightMap; + +// Push the height of DefMI upwards if required to match UseMI. +// Return true if this is the first time DefMI was seen. +static bool pushDepHeight(const DataDep &Dep, + const MachineInstr *UseMI, unsigned UseHeight, + MIHeightMap &Heights, + const TargetSchedModel &SchedModel, + const TargetInstrInfo *TII) { + // Adjust height by Dep.DefMI latency. + if (!Dep.DefMI->isTransient()) + UseHeight += SchedModel.computeOperandLatency(Dep.DefMI, Dep.DefOp, + UseMI, Dep.UseOp); + + // Update Heights[DefMI] to be the maximum height seen. + MIHeightMap::iterator I; + bool New; + std::tie(I, New) = Heights.insert(std::make_pair(Dep.DefMI, UseHeight)); + if (New) + return true; + + // DefMI has been pushed before. Give it the max height. + if (I->second < UseHeight) + I->second = UseHeight; + return false; +} + +/// Assuming that the virtual register defined by DefMI:DefOp was used by +/// Trace.back(), add it to the live-in lists of all the blocks in Trace. Stop +/// when reaching the block that contains DefMI. +void MachineTraceMetrics::Ensemble:: +addLiveIns(const MachineInstr *DefMI, unsigned DefOp, + ArrayRef Trace) { + assert(!Trace.empty() && "Trace should contain at least one block"); + unsigned Reg = DefMI->getOperand(DefOp).getReg(); + assert(TargetRegisterInfo::isVirtualRegister(Reg)); + const MachineBasicBlock *DefMBB = DefMI->getParent(); + + // Reg is live-in to all blocks in Trace that follow DefMBB. + for (unsigned i = Trace.size(); i; --i) { + const MachineBasicBlock *MBB = Trace[i-1]; + if (MBB == DefMBB) + return; + TraceBlockInfo &TBI = BlockInfo[MBB->getNumber()]; + // Just add the register. The height will be updated later. + TBI.LiveIns.push_back(Reg); + } +} + +/// Compute instruction heights in the trace through MBB. This updates MBB and +/// the blocks below it in the trace. It is assumed that the trace has already +/// been computed. +void MachineTraceMetrics::Ensemble:: +computeInstrHeights(const MachineBasicBlock *MBB) { + // The bottom of the trace may already be computed. + // Find the blocks that need updating. + SmallVector Stack; + do { + TraceBlockInfo &TBI = BlockInfo[MBB->getNumber()]; + assert(TBI.hasValidHeight() && "Incomplete trace"); + if (TBI.HasValidInstrHeights) + break; + Stack.push_back(MBB); + TBI.LiveIns.clear(); + MBB = TBI.Succ; + } while (MBB); + + // As we move upwards in the trace, keep track of instructions that are + // required by deeper trace instructions. Map MI -> height required so far. + MIHeightMap Heights; + + // For physregs, the def isn't known when we see the use. + // Instead, keep track of the highest use of each regunit. + SparseSet RegUnits; + RegUnits.setUniverse(MTM.TRI->getNumRegUnits()); + + // If the bottom of the trace was already precomputed, initialize heights + // from its live-in list. + // MBB is the highest precomputed block in the trace. + if (MBB) { + TraceBlockInfo &TBI = BlockInfo[MBB->getNumber()]; + for (unsigned i = 0, e = TBI.LiveIns.size(); i != e; ++i) { + LiveInReg LI = TBI.LiveIns[i]; + if (TargetRegisterInfo::isVirtualRegister(LI.Reg)) { + // For virtual registers, the def latency is included. + unsigned &Height = Heights[MTM.MRI->getVRegDef(LI.Reg)]; + if (Height < LI.Height) + Height = LI.Height; + } else { + // For register units, the def latency is not included because we don't + // know the def yet. + RegUnits[LI.Reg].Cycle = LI.Height; + } + } + } + + // Go through the trace blocks in bottom-up order. + SmallVector Deps; + for (;!Stack.empty(); Stack.pop_back()) { + MBB = Stack.back(); + DEBUG(dbgs() << "Heights for BB#" << MBB->getNumber() << ":\n"); + TraceBlockInfo &TBI = BlockInfo[MBB->getNumber()]; + TBI.HasValidInstrHeights = true; + TBI.CriticalPath = 0; + + DEBUG({ + dbgs() << format("%7u Instructions\n", TBI.InstrHeight); + ArrayRef PRHeights = getProcResourceHeights(MBB->getNumber()); + for (unsigned K = 0; K != PRHeights.size(); ++K) + if (PRHeights[K]) { + unsigned Factor = MTM.SchedModel.getResourceFactor(K); + dbgs() << format("%6uc @ ", MTM.getCycles(PRHeights[K])) + << MTM.SchedModel.getProcResource(K)->Name << " (" + << PRHeights[K]/Factor << " ops x" << Factor << ")\n"; + } + }); + + // Get dependencies from PHIs in the trace successor. + const MachineBasicBlock *Succ = TBI.Succ; + // If MBB is the last block in the trace, and it has a back-edge to the + // loop header, get loop-carried dependencies from PHIs in the header. For + // that purpose, pretend that all the loop header PHIs have height 0. + if (!Succ) + if (const MachineLoop *Loop = getLoopFor(MBB)) + if (MBB->isSuccessor(Loop->getHeader())) + Succ = Loop->getHeader(); + + if (Succ) { + for (MachineBasicBlock::const_iterator I = Succ->begin(), E = Succ->end(); + I != E && I->isPHI(); ++I) { + const MachineInstr *PHI = I; + Deps.clear(); + getPHIDeps(PHI, Deps, MBB, MTM.MRI); + if (!Deps.empty()) { + // Loop header PHI heights are all 0. + unsigned Height = TBI.Succ ? Cycles.lookup(PHI).Height : 0; + DEBUG(dbgs() << "pred\t" << Height << '\t' << *PHI); + if (pushDepHeight(Deps.front(), PHI, Height, + Heights, MTM.SchedModel, MTM.TII)) + addLiveIns(Deps.front().DefMI, Deps.front().DefOp, Stack); + } + } + } + + // Go through the block backwards. + for (MachineBasicBlock::const_iterator BI = MBB->end(), BB = MBB->begin(); + BI != BB;) { + const MachineInstr *MI = --BI; + + // Find the MI height as determined by virtual register uses in the + // trace below. + unsigned Cycle = 0; + MIHeightMap::iterator HeightI = Heights.find(MI); + if (HeightI != Heights.end()) { + Cycle = HeightI->second; + // We won't be seeing any more MI uses. + Heights.erase(HeightI); + } + + // Don't process PHI deps. They depend on the specific predecessor, and + // we'll get them when visiting the predecessor. + Deps.clear(); + bool HasPhysRegs = !MI->isPHI() && getDataDeps(MI, Deps, MTM.MRI); + + // There may also be regunit dependencies to include in the height. + if (HasPhysRegs) + Cycle = updatePhysDepsUpwards(MI, Cycle, RegUnits, + MTM.SchedModel, MTM.TII, MTM.TRI); + + // Update the required height of any virtual registers read by MI. + for (unsigned i = 0, e = Deps.size(); i != e; ++i) + if (pushDepHeight(Deps[i], MI, Cycle, Heights, MTM.SchedModel, MTM.TII)) + addLiveIns(Deps[i].DefMI, Deps[i].DefOp, Stack); + + InstrCycles &MICycles = Cycles[MI]; + MICycles.Height = Cycle; + if (!TBI.HasValidInstrDepths) { + DEBUG(dbgs() << Cycle << '\t' << *MI); + continue; + } + // Update critical path length. + TBI.CriticalPath = std::max(TBI.CriticalPath, Cycle + MICycles.Depth); + DEBUG(dbgs() << TBI.CriticalPath << '\t' << Cycle << '\t' << *MI); + } + + // Update virtual live-in heights. They were added by addLiveIns() with a 0 + // height because the final height isn't known until now. + DEBUG(dbgs() << "BB#" << MBB->getNumber() << " Live-ins:"); + for (unsigned i = 0, e = TBI.LiveIns.size(); i != e; ++i) { + LiveInReg &LIR = TBI.LiveIns[i]; + const MachineInstr *DefMI = MTM.MRI->getVRegDef(LIR.Reg); + LIR.Height = Heights.lookup(DefMI); + DEBUG(dbgs() << ' ' << PrintReg(LIR.Reg) << '@' << LIR.Height); + } + + // Transfer the live regunits to the live-in list. + for (SparseSet::const_iterator + RI = RegUnits.begin(), RE = RegUnits.end(); RI != RE; ++RI) { + TBI.LiveIns.push_back(LiveInReg(RI->RegUnit, RI->Cycle)); + DEBUG(dbgs() << ' ' << PrintRegUnit(RI->RegUnit, MTM.TRI) + << '@' << RI->Cycle); + } + DEBUG(dbgs() << '\n'); + + if (!TBI.HasValidInstrDepths) + continue; + // Add live-ins to the critical path length. + TBI.CriticalPath = std::max(TBI.CriticalPath, + computeCrossBlockCriticalPath(TBI)); + DEBUG(dbgs() << "Critical path: " << TBI.CriticalPath << '\n'); + } +} + MachineTraceMetrics::Trace MachineTraceMetrics::Ensemble::getTrace(const MachineBasicBlock *MBB) { // FIXME: Check cache tags, recompute as needed. computeTrace(MBB); computeInstrDepths(MBB); + computeInstrHeights(MBB); return Trace(*this, BlockInfo[MBB->getNumber()]); } +unsigned +MachineTraceMetrics::Trace::getInstrSlack(const MachineInstr *MI) const { + assert(MI && "Not an instruction."); + assert(getBlockNum() == unsigned(MI->getParent()->getNumber()) && + "MI must be in the trace center block"); + InstrCycles Cyc = getInstrCycles(MI); + return getCriticalPath() - (Cyc.Depth + Cyc.Height); +} + +unsigned +MachineTraceMetrics::Trace::getPHIDepth(const MachineInstr *PHI) const { + const MachineBasicBlock *MBB = TE.MTM.MF->getBlockNumbered(getBlockNum()); + SmallVector Deps; + getPHIDeps(PHI, Deps, MBB, TE.MTM.MRI); + assert(Deps.size() == 1 && "PHI doesn't have MBB as a predecessor"); + DataDep &Dep = Deps.front(); + unsigned DepCycle = getInstrCycles(Dep.DefMI).Depth; + // Add latency if DefMI is a real instruction. Transients get latency 0. + if (!Dep.DefMI->isTransient()) + DepCycle += TE.MTM.SchedModel + .computeOperandLatency(Dep.DefMI, Dep.DefOp, PHI, Dep.UseOp); + return DepCycle; +} + +unsigned MachineTraceMetrics::Trace::getResourceDepth(bool Bottom) const { + // Find the limiting processor resource. + // Numbers have been pre-scaled to be comparable. + unsigned PRMax = 0; + ArrayRef PRDepths = TE.getProcResourceDepths(getBlockNum()); + if (Bottom) { + ArrayRef PRCycles = TE.MTM.getProcResourceCycles(getBlockNum()); + for (unsigned K = 0; K != PRDepths.size(); ++K) + PRMax = std::max(PRMax, PRDepths[K] + PRCycles[K]); + } else { + for (unsigned K = 0; K != PRDepths.size(); ++K) + PRMax = std::max(PRMax, PRDepths[K]); + } + // Convert to cycle count. + PRMax = TE.MTM.getCycles(PRMax); + + unsigned Instrs = TBI.InstrDepth; + if (Bottom) + Instrs += TE.MTM.BlockInfo[getBlockNum()].InstrCount; + if (unsigned IW = TE.MTM.SchedModel.getIssueWidth()) + Instrs /= IW; + // Assume issue width 1 without a schedule model. + return std::max(Instrs, PRMax); +} + + +unsigned MachineTraceMetrics::Trace:: +getResourceLength(ArrayRef Extrablocks, + ArrayRef ExtraInstrs) const { + // Add up resources above and below the center block. + ArrayRef PRDepths = TE.getProcResourceDepths(getBlockNum()); + ArrayRef PRHeights = TE.getProcResourceHeights(getBlockNum()); + unsigned PRMax = 0; + for (unsigned K = 0; K != PRDepths.size(); ++K) { + unsigned PRCycles = PRDepths[K] + PRHeights[K]; + for (unsigned I = 0; I != Extrablocks.size(); ++I) + PRCycles += TE.MTM.getProcResourceCycles(Extrablocks[I]->getNumber())[K]; + for (unsigned I = 0; I != ExtraInstrs.size(); ++I) { + const MCSchedClassDesc* SC = ExtraInstrs[I]; + if (!SC->isValid()) + continue; + for (TargetSchedModel::ProcResIter + PI = TE.MTM.SchedModel.getWriteProcResBegin(SC), + PE = TE.MTM.SchedModel.getWriteProcResEnd(SC); PI != PE; ++PI) { + if (PI->ProcResourceIdx != K) + continue; + PRCycles += (PI->Cycles * TE.MTM.SchedModel.getResourceFactor(K)); + } + } + PRMax = std::max(PRMax, PRCycles); + } + // Convert to cycle count. + PRMax = TE.MTM.getCycles(PRMax); + + unsigned Instrs = TBI.InstrDepth + TBI.InstrHeight; + for (unsigned i = 0, e = Extrablocks.size(); i != e; ++i) + Instrs += TE.MTM.getResources(Extrablocks[i])->InstrCount; + if (unsigned IW = TE.MTM.SchedModel.getIssueWidth()) + Instrs /= IW; + // Assume issue width 1 without a schedule model. + return std::max(Instrs, PRMax); +} + void MachineTraceMetrics::Ensemble::print(raw_ostream &OS) const { OS << getName() << " ensemble:\n"; for (unsigned i = 0, e = BlockInfo.size(); i != e; ++i) { @@ -756,6 +1271,8 @@ void MachineTraceMetrics::TraceBlockInfo::print(raw_ostream &OS) const { OS << " +instrs"; } else OS << "height invalid"; + if (HasValidInstrDepths && HasValidInstrHeights) + OS << ", crit=" << CriticalPath; } void MachineTraceMetrics::Trace::print(raw_ostream &OS) const { @@ -765,6 +1282,8 @@ void MachineTraceMetrics::Trace::print(raw_ostream &OS) const { << " --> BB#" << TBI.Tail << ':'; if (TBI.hasValidHeight() && TBI.hasValidDepth()) OS << ' ' << getInstrCount() << " instrs."; + if (TBI.HasValidInstrDepths && TBI.HasValidInstrHeights) + OS << ' ' << TBI.CriticalPath << " cycles."; const MachineTraceMetrics::TraceBlockInfo *Block = &TBI; OS << "\nBB#" << MBBNum;