diff options
Diffstat (limited to 'lib/Target/SparcV9/ModuloScheduling/ModuloSchedulingSuperBlock.cpp')
| -rw-r--r-- | lib/Target/SparcV9/ModuloScheduling/ModuloSchedulingSuperBlock.cpp | 2968 |
1 files changed, 2922 insertions, 46 deletions
diff --git a/lib/Target/SparcV9/ModuloScheduling/ModuloSchedulingSuperBlock.cpp b/lib/Target/SparcV9/ModuloScheduling/ModuloSchedulingSuperBlock.cpp index bc18027ff4..80594c3451 100644 --- a/lib/Target/SparcV9/ModuloScheduling/ModuloSchedulingSuperBlock.cpp +++ b/lib/Target/SparcV9/ModuloScheduling/ModuloSchedulingSuperBlock.cpp @@ -17,12 +17,16 @@ #include "DependenceAnalyzer.h" #include "ModuloSchedulingSuperBlock.h" +#include "llvm/Constants.h" #include "llvm/ADT/Statistic.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/Passes.h" #include "llvm/Support/CFG.h" #include "llvm/Support/Debug.h" #include "llvm/Support/GraphWriter.h" +#include "llvm/Support/Timer.h" +#include "llvm/ADT/StringExtras.h" +#include "llvm/ADT/SCCIterator.h" #include "llvm/Instructions.h" #include "../MachineCodeForInstruction.h" #include "../SparcV9RegisterInfo.h" @@ -30,6 +34,8 @@ #include "../SparcV9TmpInstr.h" #include <fstream> #include <sstream> +#include <cmath> +#include <utility> using namespace llvm; /// Create ModuloSchedulingSBPass @@ -39,9 +45,18 @@ FunctionPass *llvm::createModuloSchedulingSBPass(TargetMachine & targ) { return new ModuloSchedulingSBPass(targ); } + +#if 1 +#define TIME_REGION(VARNAME, DESC) \ + NamedRegionTimer VARNAME(DESC) +#else +#define TIME_REGION(VARNAME, DESC) +#endif + + //Graph Traits for printing out the dependence graph template<typename GraphType> -static void WriteGraphToFile(std::ostream &O, const std::string &GraphName, +static void WriteGraphToFileSB(std::ostream &O, const std::string &GraphName, const GraphType >) { std::string Filename = GraphName + ".dot"; O << "Writing '" << Filename << "'..."; @@ -58,8 +73,82 @@ namespace llvm { Statistic<> NumLoops("moduloschedSB-numLoops", "Total Number of Loops"); Statistic<> NumSB("moduloschedSB-numSuperBlocks", "Total Number of SuperBlocks"); Statistic<> BBWithCalls("modulosched-BBCalls", "Basic Blocks rejected due to calls"); - Statistic<> BBWithCondMov("modulosched-loopCondMov", "Basic Blocks rejected due to conditional moves"); + Statistic<> BBWithCondMov("modulosched-loopCondMov", + "Basic Blocks rejected due to conditional moves"); + Statistic<> SBResourceConstraint("modulosched-resourceConstraint", + "Loops constrained by resources"); + Statistic<> SBRecurrenceConstraint("modulosched-recurrenceConstraint", + "Loops constrained by recurrences"); + Statistic<> SBFinalIISum("modulosched-finalIISum", "Sum of all final II"); + Statistic<> SBIISum("modulosched-IISum", "Sum of all theoretical II"); + Statistic<> SBMSLoops("modulosched-schedLoops", "Number of loops successfully modulo-scheduled"); + Statistic<> SBNoSched("modulosched-noSched", "No schedule"); + Statistic<> SBSameStage("modulosched-sameStage", "Max stage is 0"); + Statistic<> SBBLoops("modulosched-SBBLoops", "Number single basic block loops"); + Statistic<> SBInvalid("modulosched-SBInvalid", "Number invalid superblock loops"); + Statistic<> SBValid("modulosched-SBValid", "Number valid superblock loops"); + Statistic<> SBSize("modulosched-SBSize", "Total size of all valid superblocks"); + + template<> + struct DOTGraphTraits<MSchedGraphSB*> : public DefaultDOTGraphTraits { + static std::string getGraphName(MSchedGraphSB *F) { + return "Dependence Graph"; + } + + static std::string getNodeLabel(MSchedGraphSBNode *Node, MSchedGraphSB *Graph) { + if(!Node->isPredicate()) { + if (Node->getInst()) { + std::stringstream ss; + ss << *(Node->getInst()); + return ss.str(); //((MachineInstr*)Node->getInst()); + } + else + return "No Inst"; + } + else + return "Pred Node"; + } + static std::string getEdgeSourceLabel(MSchedGraphSBNode *Node, + MSchedGraphSBNode::succ_iterator I) { + //Label each edge with the type of dependence + std::string edgelabel = ""; + switch (I.getEdge().getDepOrderType()) { + + case MSchedGraphSBEdge::TrueDep: + edgelabel = "True"; + break; + + case MSchedGraphSBEdge::AntiDep: + edgelabel = "Anti"; + break; + + case MSchedGraphSBEdge::OutputDep: + edgelabel = "Output"; + break; + + case MSchedGraphSBEdge::NonDataDep: + edgelabel = "Pred"; + break; + + default: + edgelabel = "Unknown"; + break; + } + + //FIXME + int iteDiff = I.getEdge().getIteDiff(); + std::string intStr = "(IteDiff: "; + intStr += itostr(iteDiff); + + intStr += ")"; + edgelabel += intStr; + + return edgelabel; + } + }; + bool ModuloSchedulingSBPass::runOnFunction(Function &F) { + alarm(100); bool Changed = false; //Get MachineFunction @@ -91,14 +180,97 @@ namespace llvm { continue; } - MSchedGraph *MSG = new MSchedGraph(*SB, target, indVarInstrs[*SB], DA, + MSchedGraphSB *MSG = new MSchedGraphSB(*SB, target, indVarInstrs[*SB], DA, machineTollvm[*SB]); //Write Graph out to file - DEBUG(WriteGraphToFile(std::cerr, F.getName(), MSG)); - DEBUG(MSG->print(std::cerr)); + DEBUG(WriteGraphToFileSB(std::cerr, F.getName(), MSG)); + + //Calculate Resource II + int ResMII = calculateResMII(*SB); + //Calculate Recurrence II + int RecMII = calculateRecMII(MSG, ResMII); + + DEBUG(std::cerr << "Number of reccurrences found: " << recurrenceList.size() << "\n"); + + //Our starting initiation interval is the maximum of RecMII and ResMII + if(RecMII < ResMII) + ++SBRecurrenceConstraint; + else + ++SBResourceConstraint; + + II = std::max(RecMII, ResMII); + int mII = II; + + + //Print out II, RecMII, and ResMII + DEBUG(std::cerr << "II starts out as " << II << " ( RecMII=" << RecMII << " and ResMII=" << ResMII << ")\n"); + + //Calculate Node Properties + calculateNodeAttributes(MSG, ResMII); + + //Dump node properties if in debug mode + DEBUG(for(std::map<MSchedGraphSBNode*, MSNodeSBAttributes>::iterator I = nodeToAttributesMap.begin(), + E = nodeToAttributesMap.end(); I !=E; ++I) { + std::cerr << "Node: " << *(I->first) << " ASAP: " << I->second.ASAP << " ALAP: " + << I->second.ALAP << " MOB: " << I->second.MOB << " Depth: " << I->second.depth + << " Height: " << I->second.height << "\n"; + }); + + + //Put nodes in order to schedule them + computePartialOrder(); + + //Dump out partial order + DEBUG(for(std::vector<std::set<MSchedGraphSBNode*> >::iterator I = partialOrder.begin(), + E = partialOrder.end(); I !=E; ++I) { + std::cerr << "Start set in PO\n"; + for(std::set<MSchedGraphSBNode*>::iterator J = I->begin(), JE = I->end(); J != JE; ++J) + std::cerr << "PO:" << **J << "\n"; + }); + + //Place nodes in final order + orderNodes(); + + //Dump out order of nodes + DEBUG(for(std::vector<MSchedGraphSBNode*>::iterator I = FinalNodeOrder.begin(), E = FinalNodeOrder.end(); I != E; ++I) { + std::cerr << "FO:" << **I << "\n"; + }); + + + //Finally schedule nodes + bool haveSched = computeSchedule(*SB, MSG); + + //Print out final schedule + DEBUG(schedule.print(std::cerr)); + + //Final scheduling step is to reconstruct the loop only if we actual have + //stage > 0 + if(haveSched) { + //schedule.printSchedule(std::cerr); + reconstructLoop(*SB); + ++SBMSLoops; + //Changed = true; + SBIISum += mII; + SBFinalIISum += II; + + if(schedule.getMaxStage() == 0) + ++SBSameStage; + } + else + ++SBNoSched; + + //Clear out our maps for the next basic block that is processed + nodeToAttributesMap.clear(); + partialOrder.clear(); + recurrenceList.clear(); + FinalNodeOrder.clear(); + schedule.clear(); + defMap.clear(); + } + alarm(0); return Changed; } @@ -145,63 +317,211 @@ namespace llvm { MachineBasicBlock *currentMBB = bbMap[header]; bool done = false; bool success = true; + unsigned offset = 0; + std::map<const MachineInstr*, unsigned> indexMap; while(!done) { - - if(MachineBBisValid(currentMBB)) { - - //Loop over successors of this BB, they should be in the loop block - //and be valid - BasicBlock *next = 0; - for(succ_iterator I = succ_begin(current), E = succ_end(current); - I != E; ++I) { - if(L->contains(*I)) { - if(!next) - next = *I; - else { - done = true; - success = false; - break; - } - } - } - if(success) { - superBlock.push_back(currentMBB); - if(next == header) - done = true; - else if(!next->getSinglePredecessor()) { + //Loop over successors of this BB, they should be in the + //loop block and be valid + BasicBlock *next = 0; + for(succ_iterator I = succ_begin(current), E = succ_end(current); + I != E; ++I) { + if(L->contains(*I)) { + if(!next) + next = *I; + else { done = true; success = false; - } - else { - //Check that the next BB only has one entry - current = next; - assert(bbMap.count(current) && "LLVM BB must have corresponding Machine BB"); - currentMBB = bbMap[current]; + break; } } } - else { - done = true; - success = false; + + if(success) { + superBlock.push_back(currentMBB); + if(next == header) + done = true; + else if(!next->getSinglePredecessor()) { + done = true; + success = false; + } + else { + //Check that the next BB only has one entry + current = next; + assert(bbMap.count(current) && "LLVM BB must have corresponding Machine BB"); + currentMBB = bbMap[current]; + } } } + + + + if(success) { ++NumSB; - Worklist.push_back(superBlock); + + //Loop over all the blocks in the superblock + for(std::vector<const MachineBasicBlock*>::iterator currentMBB = superBlock.begin(), MBBEnd = superBlock.end(); currentMBB != MBBEnd; ++currentMBB) { + if(!MachineBBisValid(*currentMBB, indexMap, offset)) { + success = false; + break; + } + } + } + + if(success) { + if(getIndVar(superBlock, bbMap, indexMap)) { + ++SBValid; + Worklist.push_back(superBlock); + SBSize += superBlock.size(); + } + else + ++SBInvalid; } + } + } + } + + + bool ModuloSchedulingSBPass::getIndVar(std::vector<const MachineBasicBlock*> &superBlock, std::map<BasicBlock*, MachineBasicBlock*> &bbMap, + std::map<const MachineInstr*, unsigned> &indexMap) { + //See if we can get induction var instructions + std::set<const BasicBlock*> llvmSuperBlock; + + for(unsigned i =0; i < superBlock.size(); ++i) + llvmSuperBlock.insert(superBlock[i]->getBasicBlock()); + + //Get Target machine instruction info + const TargetInstrInfo *TMI = target.getInstrInfo(); + + //Get the loop back branch + BranchInst *b = dyn_cast<BranchInst>(((BasicBlock*) (superBlock[superBlock.size()-1])->getBasicBlock())->getTerminator()); + std::set<Instruction*> indVar; + + if(b->isConditional()) { + //Get the condition for the branch + Value *cond = b->getCondition(); + + DEBUG(std::cerr << "Condition: " << *cond << "\n"); + + //List of instructions associated with induction variable + std::vector<Instruction*> stack; + + //Add branch + indVar.insert(b); + + if(Instruction *I = dyn_cast<Instruction>(cond)) + if(bbMap.count(I->getParent())) { + if (!assocIndVar(I, indVar, stack, bbMap, superBlock[(superBlock.size()-1)]->getBasicBlock(), llvmSuperBlock)) + return false; + } + else + return false; + else + return false; + } + else { + indVar.insert(b); + } + + //Dump out instructions associate with indvar for debug reasons + DEBUG(for(std::set<Instruction*>::iterator N = indVar.begin(), NE = indVar.end(); + N != NE; ++N) { + std::cerr << **N << "\n"; + }); + + //Create map of machine instr to llvm instr + std::map<MachineInstr*, Instruction*> mllvm; + for(std::vector<const MachineBasicBlock*>::iterator MBB = superBlock.begin(), MBE = superBlock.end(); MBB != MBE; ++MBB) { + BasicBlock *BB = (BasicBlock*) (*MBB)->getBasicBlock(); + for(BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { + MachineCodeForInstruction & tempMvec = MachineCodeForInstruction::get(I); + for (unsigned j = 0; j < tempMvec.size(); j++) { + mllvm[tempMvec[j]] = I; + } + } + } + //Convert list of LLVM Instructions to list of Machine instructions + std::map<const MachineInstr*, unsigned> mIndVar; + for(std::set<Instruction*>::iterator N = indVar.begin(), + NE = indVar.end(); N != NE; ++N) { + + //If we have a load, we can't handle this loop because + //there is no way to preserve dependences between loads + //and stores + if(isa<LoadInst>(*N)) + return false; + + MachineCodeForInstruction & tempMvec = MachineCodeForInstruction::get(*N); + for (unsigned j = 0; j < tempMvec.size(); j++) { + MachineOpCode OC = (tempMvec[j])->getOpcode(); + if(TMI->isNop(OC)) + continue; + if(!indexMap.count(tempMvec[j])) + continue; + mIndVar[(MachineInstr*) tempMvec[j]] = indexMap[(MachineInstr*) tempMvec[j]]; + DEBUG(std::cerr << *(tempMvec[j]) << " at index " << indexMap[(MachineInstr*) tempMvec[j]] << "\n"); + } } + + //Put into a map for future access + indVarInstrs[superBlock] = mIndVar; + machineTollvm[superBlock] = mllvm; + + return true; + + } + bool ModuloSchedulingSBPass::assocIndVar(Instruction *I, + std::set<Instruction*> &indVar, + std::vector<Instruction*> &stack, + std::map<BasicBlock*, MachineBasicBlock*> &bbMap, + const BasicBlock *last, std::set<const BasicBlock*> &llvmSuperBlock) { + + stack.push_back(I); + + //If this is a phi node, check if its the canonical indvar + if(PHINode *PN = dyn_cast<PHINode>(I)) { + if(llvmSuperBlock.count(PN->getParent())) { + if (Instruction *Inc = + dyn_cast<Instruction>(PN->getIncomingValueForBlock(last))) + if (Inc->getOpcode() == Instruction::Add && Inc->getOperand(0) == PN) + if (ConstantInt *CI = dyn_cast<ConstantInt>(Inc->getOperand(1))) + if (CI->equalsInt(1)) { + //We have found the indvar, so add the stack, and inc instruction to the set + indVar.insert(stack.begin(), stack.end()); + indVar.insert(Inc); + stack.pop_back(); + return true; + } + return false; + } } + else { + //Loop over each of the instructions operands, check if they are an instruction and in this BB + for(unsigned i = 0; i < I->getNumOperands(); ++i) { + if(Instruction *N = dyn_cast<Instruction>(I->getOperand(i))) { + if(bbMap.count(N->getParent())) + if(!assocIndVar(N, indVar, stack, bbMap, last, llvmSuperBlock)) + return false; + } + } + } + + stack.pop_back(); + return true; } + /// This function checks if a Machine Basic Block is valid for modulo /// scheduling. This means that it has no control flow (if/else or /// calls) in the block. Currently ModuloScheduling only works on /// single basic block loops. - bool ModuloSchedulingSBPass::MachineBBisValid(const MachineBasicBlock *BI) { + bool ModuloSchedulingSBPass::MachineBBisValid(const MachineBasicBlock *BI, + std::map<const MachineInstr*, unsigned> &indexMap, + unsigned &offset) { //Check size of our basic block.. make sure we have more then just the terminator in it if(BI->getBasicBlock()->size() == 1) @@ -210,9 +530,6 @@ namespace llvm { //Get Target machine instruction info const TargetInstrInfo *TMI = target.getInstrInfo(); - //Check each instruction and look for calls, keep map to get index later - std::map<const MachineInstr*, unsigned> indexMap; - unsigned count = 0; for(MachineBasicBlock::const_iterator I = BI->begin(), E = BI->end(); I != E; ++I) { //Get opcode to check instruction type @@ -236,7 +553,7 @@ namespace llvm { return false; } - indexMap[I] = count; + indexMap[I] = count + offset; if(TMI->isNop(OC)) continue; @@ -244,6 +561,8 @@ namespace llvm { ++count; } + offset += count; + return true; } } @@ -258,10 +577,16 @@ bool ModuloSchedulingSBPass::CreateDefMap(std::vector<const MachineBasicBlock*> for(unsigned opNum = 0; opNum < I->getNumOperands(); ++opNum) { const MachineOperand &mOp = I->getOperand(opNum); if(mOp.getType() == MachineOperand::MO_VirtualRegister && mOp.isDef()) { - + Value *V = mOp.getVRegValue(); //assert if this is the second def we have seen - assert(!defMap.count(mOp.getVRegValue()) && "Def already in the map"); - defMap[mOp.getVRegValue()] = (MachineInstr*) &*I; + if(defMap.count(V) && isa<PHINode>(V)) + DEBUG(std::cerr << "FIXME: Dup def for phi!\n"); + else { + //assert(!defMap.count(V) && "Def already in the map"); + if(defMap.count(V)) + return false; + defMap[V] = (MachineInstr*) &*I; + } } //See if we can use this Value* as our defaultInst @@ -280,3 +605,2554 @@ bool ModuloSchedulingSBPass::CreateDefMap(std::vector<const MachineBasicBlock*> return true; } + + +//ResMII is calculated by determining the usage count for each resource +//and using the maximum. +//FIXME: In future there should be a way to get alternative resources +//for each instruction +int ModuloSchedulingSBPass::calculateResMII(std::vector<const MachineBasicBlock*> &superBlock) { + + TIME_REGION(X, "calculateResMII"); + + const TargetInstrInfo *mii = target.getInstrInfo(); + const TargetSchedInfo *msi = target.getSchedInfo(); + + int ResMII = 0; + + //Map to keep track of usage count of each resource + std::map<unsigned, unsigned> resourceUsageCount; + + for(std::vector<const MachineBasicBlock*>::iterator BI = superBlock.begin(), BE = superBlock.end(); BI != BE; ++BI) { + for(MachineBasicBlock::const_iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I) { + + //Get resource usage for this instruction + InstrRUsage rUsage = msi->getInstrRUsage(I->getOpcode()); + std::vector<std::vector<resourceId_t> > resources = rUsage.resourcesByCycle; + + //Loop over resources in each cycle and increments their usage count + for(unsigned i=0; i < resources.size(); ++i) + for(unsigned j=0; j < resources[i].size(); ++j) { + if(!resourceUsageCount.count(resources[i][j])) { + resourceUsageCount[resources[i][j]] = 1; + } + else { + resourceUsageCount[resources[i][j]] = resourceUsageCount[resources[i][j]] + 1; + } + } + } + } + + //Find maximum usage count + + //Get max number of instructions that can be issued at once. (FIXME) + int issueSlots = msi->maxNumIssueTotal; + + for(std::map<unsigned,unsigned>::iterator RB = resourceUsageCount.begin(), RE = resourceUsageCount.end(); RB != RE; ++RB) { + + //Get the total number of the resources in our cpu + int resourceNum = CPUResource::getCPUResource(RB->first)->maxNumUsers; + + //Get total usage count for this resources + unsigned usageCount = RB->second; + + //Divide the usage count by either the max number we can issue or the number of + //resources (whichever is its upper bound) + double finalUsageCount; + DEBUG(std::cerr << "Resource Num: " << RB->first << " Usage: " << usageCount << " TotalNum: " << resourceNum << "\n"); + + if( resourceNum <= issueSlots) + finalUsageCount = ceil(1.0 * usageCount / resourceNum); + else + finalUsageCount = ceil(1.0 * usageCount / issueSlots); + + + //Only keep track of the max + ResMII = std::max( (int) finalUsageCount, ResMII); + + } + + return ResMII; + +} + +/// calculateRecMII - Calculates the value of the highest recurrence +/// By value we mean the total latency/distance +int ModuloSchedulingSBPass::calculateRecMII(MSchedGraphSB *graph, int MII) { + + TIME_REGION(X, "calculateRecMII"); + + findAllCircuits(graph, MII); + int RecMII = 0; + + for(std::set<std::pair<int, std::vector<MSchedGraphSBNode*> > >::iterator I = recurrenceList.begin(), E=recurrenceList.end(); I !=E; ++I) { + RecMII = std::max(RecMII, I->first); + } + + return MII; +} + +int CircCountSB; + +void ModuloSchedulingSBPass::unblock(MSchedGraphSBNode *u, std::set<MSchedGraphSBNode*> &blocked, + std::map<MSchedGraphSBNode*, std::set<MSchedGraphSBNode*> > &B) { + + //Unblock u + DEBUG(std::cerr << "Unblocking: " << *u << "\n"); + blocked.erase(u); + + //std::set<MSchedGraphSBNode*> toErase; + while (!B[u].empty()) { + MSchedGraphSBNode *W = *B[u].begin(); + B[u].erase(W); + //toErase.insert(*W); + DEBUG(std::cerr << "Removed: " << *W << "from B-List\n"); + if(blocked.count(W)) + unblock(W, blocked, B); + } + +} + +void ModuloSchedulingSBPass::addSCC(std::vector<MSchedGraphSBNode*> &SCC, std::map<MSchedGraphSBNode*, MSchedGraphSBNode*> &newNodes) { + + int totalDelay = 0; + int totalDistance = 0; + std::vector<MSchedGraphSBNode*> recc; + MSchedGraphSBNode *start = 0; + MSchedGraphSBNode *end = 0; + + //Loop over recurrence, get delay and distance + for(std::vector<MSchedGraphSBNode*>::iterator N = SCC.begin(), NE = SCC.end(); N != NE; ++N) { + DEBUG(std::cerr << **N << "\n"); + totalDelay += (*N)->getLatency(); + + for(unsigned i = 0; i < (*N)->succ_size(); ++i) { + MSchedGraphSBEdge *edge = (*N)->getSuccessor(i); + if(find(SCC.begin(), SCC.end(), edge->getDest()) != SCC.end()) { + totalDistance += edge->getIteDiff(); + if(edge->getIteDiff() > 0) + if(!start && !end) { + start = *N; + end = edge->getDest(); + } + + } + } + + + //Get the original node + recc.push_back(newNodes[*N]); + + + } + + DEBUG(std::cerr << "End Recc\n"); + + + assert( (start && end) && "Must have start and end node to ignore edge for SCC"); + + if(start && end) { + //Insert reccurrence into the list + DEBUG(std::cerr << "Ignore Edge from!!: " << *start << " to " << *end << "\n"); + edgesToIgnore.insert(std::make_pair(newNodes[start], (newNodes[end])->getInEdgeNum(newNodes[start]))); + } + + int lastII = totalDelay / totalDistance; + + + recurrenceList.insert(std::make_pair(lastII, recc)); + +} + +bool ModuloSchedulingSBPass::circuit(MSchedGraphSBNode *v, std::vector<MSchedGraphSBNode*> &stack, + std::set<MSchedGraphSBNode*> &blocked, std::vector<MSchedGraphSBNode*> &SCC, + MSchedGraphSBNode *s, std::map<MSchedGraphSBNode*, std::set<MSchedGraphSBNode*> > &B, + int II, std::map<MSchedGraphSBNode*, MSchedGraphSBNode*> &newNodes) { + bool f = false; + + DEBUG(std::cerr << "Finding Circuits Starting with: ( " << v << ")"<< *v << "\n"); + + //Push node onto the stack + stack.push_back(v); + + //block this node + blocked.insert(v); + + //Loop over all successors of node v that are in the scc, create Adjaceny list + std::set<MSchedGraphSBNode*> AkV; + for(MSchedGraphSBNode::succ_iterator I = v->succ_begin(), E = v->succ_end(); I != E; ++I) { + if((std::find(SCC.begin(), SCC.end(), *I) != SCC.end())) { + AkV.insert(*I); + } + } + + for(std::set<MSchedGraphSBNode*>::iterator I = AkV.begin(), E = AkV.end(); I != E; ++I) { + if(*I == s) { + //We have a circuit, so add it to our list + addRecc(stack, newNodes); + f = true; + } + else if(!blocked.count(*I)) { + if(circuit(*I, stack, blocked, SCC, s, B, II, newNodes)) + f = true; + } + else + DEBUG(std::cerr << "Blocked: " << **I << "\n"); + } + + + if(f) { + unblock(v, blocked, B); + } + else { + for(std::set<MSchedGraphSBNode*>::iterator I = AkV.begin(), E = AkV.end(); I != E; ++I) + B[*I].insert(v); + + } + + //Pop v + stack.pop_back(); + + return f; + +} + +void ModuloSchedulingSBPass::addRecc(std::vector<MSchedGraphSBNode*> &stack, std::map<MSchedGraphSBNode*, MSchedGraphSBNode*> &newNodes) { + std::vector<MSchedGraphSBNode*> recc; + //Dump recurrence for now + DEBUG(std::cerr << "Starting Recc\n"); + + int totalDelay = 0; + int totalDistance = 0; + MSchedGraphSBNode *lastN = 0; + MSchedGraphSBNode *start = 0; + MSchedGraphSBNode *end = 0; + + //Loop over recurrence, get delay and distance + for(std::vector<MSchedGraphSBNode*>::iterator N = stack.begin(), NE = stack.end(); N != NE; ++N) { + DEBUG(std::cerr << **N << "\n"); + totalDelay += (*N)->getLatency(); + if(lastN) { + int iteDiff = (*N)->getInEdge(lastN).getIteDiff(); + totalDistance += iteDiff; + + if(iteDiff > 0) { + start = lastN; + end = *N; + } + } + //Get the original node + lastN = *N; + recc.push_back(newNodes[*N]); + + + } + + //Get the loop edge + totalDistance += lastN->getIteDiff(*stack.begin()); + + DEBUG(std::cerr << "End Recc\n"); + CircCountSB++; + + if(start && end) { + //Insert reccurrence into the list + DEBUG(std::cerr << "Ignore Edge from!!: " << *start << " to " << *end << "\n"); + edgesToIgnore.insert(std::make_pair(newNodes[start], (newNodes[end])->getInEdgeNum(newNodes[start]))); + } + else { + //Insert reccurrence into the list + DEBUG(std::cerr << "Ignore Edge from: " << *lastN << " to " << **stack.begin() << "\n"); + edgesToIgnore.insert(std::make_pair(newNodes[lastN], newNodes[(*stack.begin())]->getInEdgeNum(newNodes[lastN]))); + + } + //Adjust II until we get close to the inequality delay - II*distance <= 0 + int RecMII = II; //Starting value + int value = totalDelay-(RecMII * totalDistance); + int lastII = II; + while(value < 0) { + + lastII = RecMII; + RecMII--; + value = totalDelay-(RecMII * totalDistance); + } + + recurrenceList.insert(std::make_pair(lastII, recc)); + +} + + +void ModuloSchedulingSBPass::findAllCircuits(MSchedGraphSB *g, int II) { + + CircCountSB = 0; + + //Keep old to new node mapping information + std::map<MSchedGraphSBNode*, MSchedGraphSBNode*> newNodes; + + //copy the graph + MSchedGraphSB *MSG = new MSchedGraphSB(*g, newNodes); + + DEBUG(std::cerr << "Finding All Circuits\n"); + + //Set of blocked nodes + std::set<MSchedGraphSBNode*> blocked; + + //Stack holding current circuit + std::vector<MSchedGraphSBNode*> stack; + + //Map for B Lists + std::map<MSchedGraphSBNode*, std::set<MSchedGraphSBNode*> > B; + + //current node + MSchedGraphSBNode *s; + + + //Iterate over the graph until its down to one node or empty + while(MSG->size() > 1) { + + //Write Graph out to file + //WriteGraphToFile(std::cerr, "Graph" + utostr(MSG->size()), MSG); + + DEBUG(std::cerr << "Graph Size: " << MSG->size() << "\n"); + DEBUG(std::cerr << "Finding strong component Vk with least vertex\n"); + + //Iterate over all the SCCs in the graph + std::set<MSchedGraphSBNode*> Visited; + std::vector<MSchedGraphSBNode*> Vk; + MSchedGraphSBNode* s = 0; + int numEdges = 0; + + //Find scc with the least vertex + for (MSchedGraphSB::iterator GI = MSG->begin(), E = MSG->end(); GI != E; ++GI) + if (Visited.insert(GI->second).second) { + for (scc_iterator<MSchedGraphSBNode*> SCCI = scc_begin(GI->second), + E = scc_end(GI->second); SCCI != E; ++SCCI) { + std::vector<MSchedGraphSBNode*> &nextSCC = *SCCI; + + if (Visited.insert(nextSCC[0]).second) { + Visited.insert(nextSCC.begin()+1, nextSCC.end()); + + if(nextSCC.size() > 1) { + DEBUG(std::cerr << "SCC size: " << nextSCC.size() << "\n"); + + for(unsigned i = 0; i < nextSCC.size(); ++i) { + //Loop over successor and see if in scc, then count edge + MSchedGraphSBNode *node = nextSCC[i]; + for(MSchedGraphSBNode::succ_iterator S = node->succ_begin(), SE = node->succ_end(); S != SE; ++S) { + if(find(nextSCC.begin(), nextSCC.end(), *S) != nextSCC.end()) + numEdges++; + } + } + DEBUG(std::cerr << "Num Edges: " << numEdges << "\n"); + } + + //Ignore self loops + if(nextSCC.size() > 1) { + + //Get least vertex in Vk + if(!s) { + s = nextSCC[0]; + Vk = nextSCC; + } + + for(unsigned i = 0; i < nextSCC.size(); ++i) { + if(nextSCC[i] < s) { + s = nextSCC[i]; + Vk = nextSCC; + } + } + } + } + } + } + + + + //Process SCC + DEBUG(for(std::vector<MSchedGraphSBNode*>::iterator N = Vk.begin(), NE = Vk.end(); + N != NE; ++N) { std::cerr << *((*N)->getInst()); }); + + //Iterate over all nodes in this scc + for(std::vector<MSchedGraphSBNode*>::iterator N = Vk.begin(), NE = Vk.end(); + N != NE; ++N) { + blocked.erase(*N); + B[*N].clear(); + } + if(Vk.size() > 1) { + if(numEdges < 98) + circuit(s, stack, blocked, Vk, s, B, II, newNodes); + else + addSCC(Vk, newNodes); + + + //Delete nodes from the graph + //Find all nodes up to s and delete them + std::vector<MSchedGraphSBNode*> nodesToRemove; + nodesToRemove.push_back(s); + for(MSchedGraphSB::iterator N = MSG->begin(), NE = MSG->end(); N != NE; ++N) { + if(N->second < s ) + nodesToRemove.push_back(N->second); + } + for(std::vector<MSchedGraphSBNode*>::iterator N = nodesToRemove.begin(), NE = nodesToRemove.end(); N != NE; ++N) { + DEBUG(std::cerr << "Deleting Node: " << **N << "\n"); + MSG->deleteNode(*N); + } + } + else + break; + } + DEBUG(std::cerr << "Num Circuits found: " << CircCountSB << "\n"); +} +/// calculateNodeAttributes - The following properties are calculated for +/// each node in the dependence graph: ASAP, ALAP, Depth, Height, and +/// MOB. +void ModuloSchedulingSBPass::calculateNodeAttributes(MSchedGraphSB *graph, int MII) { + + TIME_REGION(X, "calculateNodeAttributes"); + + assert(nodeToAttributesMap.empty() && "Node attribute map was not cleared"); + + //Loop over the nodes and add them to the map + for(MSchedGraphSB::iterator I = graph->begin(), E = graph->end(); I != E; ++I) { + + DEBUG(std::cerr << "Inserting node into attribute map: " << *I->second << "\n"); + + //Assert if its already in the map + assert(nodeToAttributesMap.count(I->second) == 0 && + "Node attributes are already in the map"); + + //Put into the map with default attribute values + nodeToAttributesMap[I->second] = MSNodeSBAttributes(); + } + + //Create set to deal with reccurrences + std::set<MSchedGraphSBNode*> visitedNodes; + + //Now Loop over map and calculate the node attributes + for(std::map<MSchedGraphSBNode*, MSNodeSBAttributes>::iterator I = nodeToAttributesMap.begin(), E = nodeToAttributesMap.end(); I != E; ++I) { + calculateASAP(I->first, MII, (MSchedGraphSBNode*) 0); + visitedNodes.clear(); + } + + int maxASAP = findMaxASAP(); + //Calculate ALAP which depends on ASAP being totally calculated + for(std::map<MSchedGraphSBNode*, MSNodeSBAttributes>::iterator I = nodeToAttributesMap.begin(), E = nodeToAttributesMap.end(); I != E; ++I) { + calculateALAP(I->first, MII, maxASAP, (MSchedGraphSBNode*) 0); + visitedNodes.clear(); + } + + //Calculate MOB which depends on ASAP being totally calculated, also do depth and height + for(std::map<MSchedGraphSBNode*, MSNodeSBAttributes>::iterator I = nodeToAttributesMap.begin(), E = nodeToAttributesMap.end(); I != E; ++I) { + (I->second).MOB = std::max(0,(I->second).ALAP - (I->second).ASAP); + + DEBUG(std::cerr << "MOB: " << (I->second).MOB << " (" << *(I->first) << ")\n"); + calculateDepth(I->first, (MSchedGraphSBNode*) 0); + calculateHeight(I->first, (MSchedGraphSBNode*) 0); + } + + +} + +/// ignoreEdge - Checks to see if this edge of a recurrence should be ignored or not +bool ModuloSchedulingSBPass::ignoreEdge(MSchedGraphSBNode *srcNode, MSchedGraphSBNode *destNode) { + if(destNode == 0 || srcNode ==0) + return false; + + bool findEdge = edgesToIgnore.count(std::make_pair(srcNode, destNode->getInEdgeNum(srcNode))); + + DEBUG(std::cerr << "Ignoring edge? from: " << *srcNode << " to " << *destNode << "\n"); + + return findEdge; +} + + +/// calculateASAP - Calculates the +int ModuloSchedulingSBPass::calculateASAP(MSchedGraphSBNode *node, int MII, MSchedGraphSBNode *destNode) { + + DEBUG(std::cerr << "Calculating ASAP for " << *node << "\n"); + + //Get current node attributes + MSNodeSBAttributes &attributes = nodeToAttributesMap.find(node)->second; + + if(attributes.ASAP != -1) + return attributes.ASAP; + + int maxPredValue = 0; + + //Iterate over all of the predecessors and find max + for(MSchedGraphSBNode::pred_iterator P = node->pred_begin(), E = node->pred_end(); P != E; ++P) { + + //Only process if we are not ignoring the edge + if(!ignoreEdge(*P, node)) { + int predASAP = -1; + predASAP = calculateASAP(*P, MII, node); + + assert(predASAP != -1 && "ASAP has not been calculated"); + int iteDiff = node->getInEdge(*P).getIteDiff(); + + int currentPredValue = predASAP + (*P)->getLatency() - (iteDiff * MII); + DEBUG(std::cerr << "pred ASAP: " << predASAP << ", iteDiff: " << iteDiff << ", PredLatency: " << (*P)->getLatency() << ", Current ASAP pred: " << currentPredValue << "\n"); + maxPredValue = std::max(maxPredValue, currentPredValue); + } + } + + attributes.ASAP = maxPredValue; + + DEBUG(std::cerr << "ASAP: " << attributes.ASAP << " (" << *node <&l |