diff options
Diffstat (limited to 'lib/Target/SparcV9/ModuloScheduling/ModuloSchedulingSuperBlock.cpp')
-rw-r--r-- | lib/Target/SparcV9/ModuloScheduling/ModuloSchedulingSuperBlock.cpp | 3155 |
1 files changed, 0 insertions, 3155 deletions
diff --git a/lib/Target/SparcV9/ModuloScheduling/ModuloSchedulingSuperBlock.cpp b/lib/Target/SparcV9/ModuloScheduling/ModuloSchedulingSuperBlock.cpp deleted file mode 100644 index 8b3185155e..0000000000 --- a/lib/Target/SparcV9/ModuloScheduling/ModuloSchedulingSuperBlock.cpp +++ /dev/null @@ -1,3155 +0,0 @@ -//===-- ModuloSchedulingSuperBlock.cpp - ModuloScheduling--------*- C++ -*-===// -// -// The LLVM Compiler Infrastructure -// -// This file was developed by the LLVM research group and is distributed under -// the University of Illinois Open Source License. See LICENSE.TXT for details. -// -//===----------------------------------------------------------------------===// -// -// This ModuloScheduling pass is based on the Swing Modulo Scheduling -// algorithm, but has been extended to support SuperBlocks (multiple -// basic block, single entry, multipl exit loops). -// -//===----------------------------------------------------------------------===// - -#define DEBUG_TYPE "ModuloSchedSB" - -#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" -#include "../SparcV9Internals.h" -#include "../SparcV9TmpInstr.h" -#include <fstream> -#include <sstream> -#include <cmath> -#include <utility> - -using namespace llvm; -/// Create ModuloSchedulingSBPass -/// -FunctionPass *llvm::createModuloSchedulingSBPass(TargetMachine & targ) { - DEBUG(std::cerr << "Created ModuloSchedulingSBPass\n"); - 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 WriteGraphToFileSB(std::ostream &O, const std::string &GraphName, - const GraphType >) { - std::string Filename = GraphName + ".dot"; - O << "Writing '" << Filename << "'..."; - std::ofstream F(Filename.c_str()); - - if (F.good()) - WriteGraph(F, GT); - else - O << " error opening file for writing!"; - O << "\n"; -}; - -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<> 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) { - bool Changed = false; - - //Get MachineFunction - MachineFunction &MF = MachineFunction::get(&F); - - //Get Loop Info & Dependence Anaysis info - LoopInfo &LI = getAnalysis<LoopInfo>(); - DependenceAnalyzer &DA = getAnalysis<DependenceAnalyzer>(); - - //Worklist of superblocks - std::vector<std::vector<const MachineBasicBlock*> > Worklist; - FindSuperBlocks(F, LI, Worklist); - - DEBUG(if(Worklist.size() == 0) std::cerr << "No superblocks in function to ModuloSchedule\n"); - - //Loop over worklist and ModuloSchedule each SuperBlock - for(std::vector<std::vector<const MachineBasicBlock*> >::iterator SB = Worklist.begin(), - SBE = Worklist.end(); SB != SBE; ++SB) { - - //Print out Superblock - DEBUG(std::cerr << "ModuloScheduling SB: \n"; - for(std::vector<const MachineBasicBlock*>::const_iterator BI = SB->begin(), - BE = SB->end(); BI != BE; ++BI) { - (*BI)->print(std::cerr);}); - - if(!CreateDefMap(*SB)) { - defaultInst = 0; - defMap.clear(); - continue; - } - - MSchedGraphSB *MSG = new MSchedGraphSB(*SB, target, indVarInstrs[*SB], DA, - machineTollvm[*SB]); - - //Write Graph out to file - 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(); - - } - return Changed; - } - - void ModuloSchedulingSBPass::FindSuperBlocks(Function &F, LoopInfo &LI, - std::vector<std::vector<const MachineBasicBlock*> > &Worklist) { - - //Get MachineFunction - MachineFunction &MF = MachineFunction::get(&F); - - //Map of LLVM BB to machine BB - std::map<BasicBlock*, MachineBasicBlock*> bbMap; - - for (MachineFunction::iterator BI = MF.begin(); BI != MF.end(); ++BI) { - BasicBlock *llvmBB = (BasicBlock*) BI->getBasicBlock(); - assert(!bbMap.count(llvmBB) && "LLVM BB already in map!"); - bbMap[llvmBB] = &*BI; - } - - //Iterate over the loops, and find super blocks - for(LoopInfo::iterator LB = LI.begin(), LE = LI.end(); LB != LE; ++LB) { - Loop *L = *LB; - ++NumLoops; - - //If loop is not single entry, try the next one - if(!L->getLoopPreheader()) - continue; - - //Check size of this loop, we don't want SBB loops - if(L->getBlocks().size() == 1) - continue; - - //Check if this loop contains no sub loops - if(L->getSubLoops().size() == 0) { - - std::vector<const MachineBasicBlock*> superBlock; - - //Get Loop Headers - BasicBlock *header = L->getHeader(); - - //Follow the header and make sure each BB only has one entry and is valid - BasicBlock *current = header; - assert(bbMap.count(current) && "LLVM BB must have corresponding Machine BB\n"); - MachineBasicBlock *currentMBB = bbMap[header]; - bool done = false; - bool success = true; - unsigned offset = 0; - std::map<const MachineInstr*, unsigned> indexMap; - - while(!done) { - //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()) { - 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; - - //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, - 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) - return false; - - //Get Target machine instruction info - const TargetInstrInfo *TMI = target.getInstrInfo(); - - unsigned count = 0; - for(MachineBasicBlock::const_iterator I = BI->begin(), E = BI->end(); I != E; ++I) { - //Get opcode to check instruction type - MachineOpCode OC = I->getOpcode(); - - //Look for calls - if(TMI->isCall(OC)) { - ++BBWithCalls; - return false; - } - - //Look for conditional move - if(OC == V9::MOVRZr || OC == V9::MOVRZi || OC == V9::MOVRLEZr || OC == V9::MOVRLEZi - || OC == V9::MOVRLZr || OC == V9::MOVRLZi || OC == V9::MOVRNZr || OC == V9::MOVRNZi - || OC == V9::MOVRGZr || OC == V9::MOVRGZi || OC == V9::MOVRGEZr - || OC == V9::MOVRGEZi || OC == V9::MOVLEr || OC == V9::MOVLEi || OC == V9::MOVLEUr - || OC == V9::MOVLEUi || OC == V9::MOVFLEr || OC == V9::MOVFLEi - || OC == V9::MOVNEr || OC == V9::MOVNEi || OC == V9::MOVNEGr || OC == V9::MOVNEGi - || OC == V9::MOVFNEr || OC == V9::MOVFNEi) { - ++BBWithCondMov; - return false; - } - - indexMap[I] = count + offset; - - if(TMI->isNop(OC)) - continue; - - ++count; - } - - offset += count; - - return true; - } -} - -bool ModuloSchedulingSBPass::CreateDefMap(std::vector<const MachineBasicBlock*> &SB) { - defaultInst = 0; - - for(std::vector<const MachineBasicBlock*>::iterator BI = SB.begin(), - BE = SB.end(); BI != BE; ++BI) { - - for(MachineBasicBlock::const_iterator I = (*BI)->begin(), E = (*BI)->end(); I != E; ++I) { - 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 - 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 - if(!defaultInst && mOp.getType() == MachineOperand::MO_VirtualRegister) { - Value *V = mOp.getVRegValue(); - if(!isa<TmpInstruction>(V) && !isa<Argument>(V) && !isa<Constant>(V) && !isa<PHINode>(V)) - defaultInst = (Instruction*) V; - } - } - } - } - - if(!defaultInst) - return false; - - 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 != |