diff --git a/lib/CodeGen/ModuloScheduling/MSSchedule.cpp b/lib/CodeGen/ModuloScheduling/MSSchedule.cpp deleted file mode 100644 index 4ce85e4949d..00000000000 --- a/lib/CodeGen/ModuloScheduling/MSSchedule.cpp +++ /dev/null @@ -1,196 +0,0 @@ -//===-- MSSchedule.cpp Schedule ---------------------------------*- 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. -// -//===----------------------------------------------------------------------===// -// -// -// -//===----------------------------------------------------------------------===// -#define DEBUG_TYPE "ModuloSched" - -#include "MSSchedule.h" -#include "llvm/Support/Debug.h" -#include "llvm/Target/TargetSchedInfo.h" -#include "../../Target/SparcV9/SparcV9Internals.h" - -using namespace llvm; - -bool MSSchedule::insert(MSchedGraphNode *node, int cycle) { - - //First, check if the cycle has a spot free to start - if(schedule.find(cycle) != schedule.end()) { - if (schedule[cycle].size() < numIssue) { - if(resourcesFree(node, cycle)) { - schedule[cycle].push_back(node); - DEBUG(std::cerr << "Found spot in map, and there is an issue slot\n"); - return false; - } - } - } - //Not in the map yet so put it in - else { - if(resourcesFree(node,cycle)) { - std::vector nodes; - nodes.push_back(node); - schedule[cycle] = nodes; - DEBUG(std::cerr << "Nothing in map yet so taking an issue slot\n"); - return false; - } - } - - DEBUG(std::cerr << "All issue slots taken\n"); - return true; - -} - -bool MSSchedule::resourcesFree(MSchedGraphNode *node, int cycle) { - - //Get Resource usage for this instruction - const TargetSchedInfo *msi = node->getParent()->getTarget()->getSchedInfo(); - int currentCycle = cycle; - bool success = true; - - //Get resource usage for this instruction - InstrRUsage rUsage = msi->getInstrRUsage(node->getInst()->getOpcode()); - std::vector > 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) { - int resourceNum = resources[i][j]; - - //Check if this resource is available for this cycle - std::map >::iterator resourcesForCycle = resourceNumPerCycle.find(currentCycle); - - //First check map of resources for this cycle - if(resourcesForCycle != resourceNumPerCycle.end()) { - //A map exists for this cycle, so lets check for the resource - std::map::iterator resourceUse = resourcesForCycle->second.find(resourceNum); - if(resourceUse != resourcesForCycle->second.end()) { - //Check if there are enough of this resource and if so, increase count and move on - if(resourceUse->second < CPUResource::getCPUResource(resourceNum)->maxNumUsers) - ++resourceUse->second; - else { - success = false; - } - } - //Not in the map yet, so put it - else - resourcesForCycle->second[resourceNum] = 1; - - } - else { - //Create a new map and put in our resource - std::map resourceMap; - resourceMap[resourceNum] = 1; - resourceNumPerCycle[cycle] = resourceMap; - } - if(!success) - break; - } - if(!success) - break; - //Increase cycle - currentCycle++; - } - - if(!success) { - int oldCycle = cycle; - DEBUG(std::cerr << "Backtrack\n"); - //Get resource usage for this instruction - InstrRUsage rUsage = msi->getInstrRUsage(node->getInst()->getOpcode()); - std::vector > resources = rUsage.resourcesByCycle; - - //Loop over resources in each cycle and increments their usage count - for(unsigned i=0; i < resources.size(); ++i) { - if(oldCycle < currentCycle) { - - //Check if this resource is available for this cycle - std::map >::iterator resourcesForCycle = resourceNumPerCycle.find(oldCycle); - - for(unsigned j=0; j < resources[i].size(); ++j) { - int resourceNum = resources[i][j]; - //remove from map - std::map::iterator resourceUse = resourcesForCycle->second.find(resourceNum); - //assert if not in the map.. since it should be! - //assert(resourceUse != resourcesForCycle.end() && "Resource should be in map!"); - --resourceUse->second; - } - } - else - break; - oldCycle++; - } - return false; - - } - - return true; - -} - -bool MSSchedule::constructKernel(int II) { - MSchedGraphNode *branchNode = 0; - MSchedGraphNode *branchANode = 0; - - int stageNum = (schedule.rbegin()->first)/ II; - DEBUG(std::cerr << "Number of Stages: " << stageNum << "\n"); - - for(int index = 0; index < II; ++index) { - int count = 0; - for(int i = index; i <= (schedule.rbegin()->first); i+=II) { - if(schedule.count(i)) { - for(std::vector::iterator I = schedule[i].begin(), - E = schedule[i].end(); I != E; ++I) { - //Check if its a branch - if((*I)->isBranch()) { - if((*I)->getInst()->getOpcode() == V9::BA) - branchANode = *I; - else - branchNode = *I; - assert(count == 0 && "Branch can not be from a previous iteration"); - } - else - //FIXME: Check if the instructions in the earlier stage conflict - kernel.push_back(std::make_pair(*I, count)); - } - } - ++count; - } - } - - //Add Branch to the end - kernel.push_back(std::make_pair(branchNode, 0)); - - //Add Branch Always to the end - kernel.push_back(std::make_pair(branchANode, 0)); - - - if(stageNum > 0) - maxStage = stageNum; - else - maxStage = 0; - - return true; -} - - -void MSSchedule::print(std::ostream &os) const { - os << "Schedule:\n"; - - for(schedule_const_iterator I = schedule.begin(), E = schedule.end(); I != E; ++I) { - os << "Cycle: " << I->first << "\n"; - for(std::vector::const_iterator node = I->second.begin(), nodeEnd = I->second.end(); node != nodeEnd; ++node) - os << **node << "\n"; - } - - os << "Kernel:\n"; - for(std::vector >::const_iterator I = kernel.begin(), - E = kernel.end(); I != E; ++I) - os << "Node: " << *(I->first) << " Stage: " << I->second << "\n"; -} - diff --git a/lib/CodeGen/ModuloScheduling/MSSchedule.h b/lib/CodeGen/ModuloScheduling/MSSchedule.h deleted file mode 100644 index 248a7485917..00000000000 --- a/lib/CodeGen/ModuloScheduling/MSSchedule.h +++ /dev/null @@ -1,66 +0,0 @@ -//===-- MSSchedule.h - Schedule ------- -------------------------*- 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. -// -//===----------------------------------------------------------------------===// -// -// The schedule generated by a scheduling algorithm -// -//===----------------------------------------------------------------------===// - -#ifndef LLVM_MSSCHEDULE_H -#define LLVM_MSSCHEDULE_H - -#include "MSchedGraph.h" -#include - -namespace llvm { - - class MSSchedule { - std::map > schedule; - unsigned numIssue; - - //Internal map to keep track of explicit resources - std::map > resourceNumPerCycle; - - //Check if all resources are free - bool resourcesFree(MSchedGraphNode*, int); - - //Resulting kernel - std::vector > kernel; - - //Max stage count - int maxStage; - - public: - MSSchedule(int num) : numIssue(num) {} - MSSchedule() : numIssue(4) {} - bool insert(MSchedGraphNode *node, int cycle); - int getStartCycle(MSchedGraphNode *node); - void clear() { schedule.clear(); resourceNumPerCycle.clear(); kernel.clear(); } - std::vector >* getKernel() { return &kernel; } - bool constructKernel(int II); - int getMaxStage() { return maxStage; } - - - //iterators - typedef std::map >::iterator schedule_iterator; - typedef std::map >::const_iterator schedule_const_iterator; - schedule_iterator begin() { return schedule.begin(); }; - schedule_iterator end() { return schedule.end(); }; - void print(std::ostream &os) const; - - typedef std::vector >::iterator kernel_iterator; - typedef std::vector >::const_iterator kernel_const_iterator; - kernel_iterator kernel_begin() { return kernel.begin(); } - kernel_iterator kernel_end() { return kernel.end(); } - - }; - -} - - -#endif diff --git a/lib/CodeGen/ModuloScheduling/MSchedGraph.cpp b/lib/CodeGen/ModuloScheduling/MSchedGraph.cpp deleted file mode 100644 index 2b246ca674e..00000000000 --- a/lib/CodeGen/ModuloScheduling/MSchedGraph.cpp +++ /dev/null @@ -1,429 +0,0 @@ -//===-- MSchedGraph.cpp - Scheduling Graph ----------------------*- 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. -// -//===----------------------------------------------------------------------===// -// -// A graph class for dependencies -// -//===----------------------------------------------------------------------===// -#define DEBUG_TYPE "ModuloSched" - -#include "MSchedGraph.h" -#include "../../Target/SparcV9/SparcV9RegisterInfo.h" -#include "llvm/CodeGen/MachineBasicBlock.h" -#include "llvm/Target/TargetInstrInfo.h" -#include "llvm/Support/Debug.h" -#include -#include -using namespace llvm; - -MSchedGraphNode::MSchedGraphNode(const MachineInstr* inst, - MSchedGraph *graph, - unsigned late, bool isBranch) - : Inst(inst), Parent(graph), latency(late), isBranchInstr(isBranch) { - - //Add to the graph - graph->addNode(inst, this); -} - -void MSchedGraphNode::print(std::ostream &os) const { - os << "MSchedGraphNode: Inst=" << *Inst << ", latency= " << latency << "\n"; -} - -MSchedGraphEdge MSchedGraphNode::getInEdge(MSchedGraphNode *pred) { - //Loop over all the successors of our predecessor - //return the edge the corresponds to this in edge - for (MSchedGraphNode::succ_iterator I = pred->succ_begin(), - E = pred->succ_end(); I != E; ++I) { - if (*I == this) - return I.getEdge(); - } - assert(0 && "Should have found edge between this node and its predecessor!"); - abort(); -} - -unsigned MSchedGraphNode::getInEdgeNum(MSchedGraphNode *pred) { - //Loop over all the successors of our predecessor - //return the edge the corresponds to this in edge - int count = 0; - for(MSchedGraphNode::succ_iterator I = pred->succ_begin(), E = pred->succ_end(); - I != E; ++I) { - if(*I == this) - return count; - count++; - } - assert(0 && "Should have found edge between this node and its predecessor!"); - abort(); -} -bool MSchedGraphNode::isSuccessor(MSchedGraphNode *succ) { - for(succ_iterator I = succ_begin(), E = succ_end(); I != E; ++I) - if(*I == succ) - return true; - return false; -} - - -bool MSchedGraphNode::isPredecessor(MSchedGraphNode *pred) { - if(std::find( Predecessors.begin(), Predecessors.end(), pred) != Predecessors.end()) - return true; - else - return false; -} - - -void MSchedGraph::addNode(const MachineInstr *MI, - MSchedGraphNode *node) { - - //Make sure node does not already exist - assert(GraphMap.find(MI) == GraphMap.end() - && "New MSchedGraphNode already exists for this instruction"); - - GraphMap[MI] = node; -} - -MSchedGraph::MSchedGraph(const MachineBasicBlock *bb, const TargetMachine &targ) - : BB(bb), Target(targ) { - - //Make sure BB is not null, - assert(BB != NULL && "Basic Block is null"); - - //DEBUG(std::cerr << "Constructing graph for " << bb << "\n"); - - //Create nodes and edges for this BB - buildNodesAndEdges(); -} - -MSchedGraph::~MSchedGraph () { - for(MSchedGraph::iterator I = GraphMap.begin(), E = GraphMap.end(); I != E; ++I) - delete I->second; -} - -void MSchedGraph::buildNodesAndEdges() { - - //Get Machine target information for calculating latency - const TargetInstrInfo *MTI = Target.getInstrInfo(); - - std::vector memInstructions; - std::map > regNumtoNodeMap; - std::map > valuetoNodeMap; - - //Save PHI instructions to deal with later - std::vector phiInstrs; - - //Loop over instructions in MBB and add nodes and edges - for (MachineBasicBlock::const_iterator MI = BB->begin(), e = BB->end(); MI != e; ++MI) { - //Get each instruction of machine basic block, get the delay - //using the op code, create a new node for it, and add to the - //graph. - - MachineOpCode opCode = MI->getOpcode(); - int delay; - -#if 0 // FIXME: LOOK INTO THIS - //Check if subsequent instructions can be issued before - //the result is ready, if so use min delay. - if(MTI->hasResultInterlock(MIopCode)) - delay = MTI->minLatency(MIopCode); - else -#endif - //Get delay - delay = MTI->maxLatency(opCode); - - //Create new node for this machine instruction and add to the graph. - //Create only if not a nop - if(MTI->isNop(opCode)) - continue; - - //Add PHI to phi instruction list to be processed later - if (opCode == TargetInstrInfo::PHI) - phiInstrs.push_back(MI); - - bool isBranch = false; - - //We want to flag the branch node to treat it special - if(MTI->isBranch(opCode)) - isBranch = true; - - //Node is created and added to the graph automatically - MSchedGraphNode *node = new MSchedGraphNode(MI, this, delay, isBranch); - - DEBUG(std::cerr << "Created Node: " << *node << "\n"); - - //Check OpCode to keep track of memory operations to add memory dependencies later. - if(MTI->isLoad(opCode) || MTI->isStore(opCode)) - memInstructions.push_back(node); - - //Loop over all operands, and put them into the register number to - //graph node map for determining dependencies - //If an operands is a use/def, we have an anti dependence to itself - for(unsigned i=0; i < MI->getNumOperands(); ++i) { - //Get Operand - const MachineOperand &mOp = MI->getOperand(i); - - //Check if it has an allocated register - if(mOp.hasAllocatedReg()) { - int regNum = mOp.getReg(); - - if(regNum != SparcV9::g0) { - //Put into our map - regNumtoNodeMap[regNum].push_back(std::make_pair(i, node)); - } - continue; - } - - - //Add virtual registers dependencies - //Check if any exist in the value map already and create dependencies - //between them. - if(mOp.getType() == MachineOperand::MO_VirtualRegister || mOp.getType() == MachineOperand::MO_CCRegister) { - - //Make sure virtual register value is not null - assert((mOp.getVRegValue() != NULL) && "Null value is defined"); - - //Check if this is a read operation in a phi node, if so DO NOT PROCESS - if(mOp.isUse() && (opCode == TargetInstrInfo::PHI)) - continue; - - - if (const Value* srcI = mOp.getVRegValue()) { - - //Find value in the map - std::map >::iterator V - = valuetoNodeMap.find(srcI); - - //If there is something in the map already, add edges from - //those instructions - //to this one we are processing - if(V != valuetoNodeMap.end()) { - addValueEdges(V->second, node, mOp.isUse(), mOp.isDef()); - - //Add to value map - V->second.push_back(std::make_pair(i,node)); - } - //Otherwise put it in the map - else - //Put into value map - valuetoNodeMap[mOp.getVRegValue()].push_back(std::make_pair(i, node)); - } - } - } - } - addMemEdges(memInstructions); - addMachRegEdges(regNumtoNodeMap); - - //Finally deal with PHI Nodes and Value* - for(std::vector::iterator I = phiInstrs.begin(), E = phiInstrs.end(); I != E; ++I) { - //Get Node for this instruction - MSchedGraphNode *node = find(*I)->second; - - //Loop over operands for this instruction and add value edges - for(unsigned i=0; i < (*I)->getNumOperands(); ++i) { - //Get Operand - const MachineOperand &mOp = (*I)->getOperand(i); - if((mOp.getType() == MachineOperand::MO_VirtualRegister || mOp.getType() == MachineOperand::MO_CCRegister) && mOp.isUse()) { - //find the value in the map - if (const Value* srcI = mOp.getVRegValue()) { - - //Find value in the map - std::map >::iterator V - = valuetoNodeMap.find(srcI); - - //If there is something in the map already, add edges from - //those instructions - //to this one we are processing - if(V != valuetoNodeMap.end()) { - addValueEdges(V->second, node, mOp.isUse(), mOp.isDef(), 1); - } - } - } - } - } -} - -void MSchedGraph::addValueEdges(std::vector &NodesInMap, - MSchedGraphNode *destNode, bool nodeIsUse, - bool nodeIsDef, int diff) { - - for(std::vector::iterator I = NodesInMap.begin(), - E = NodesInMap.end(); I != E; ++I) { - - //Get node in vectors machine operand that is the same value as node - MSchedGraphNode *srcNode = I->second; - MachineOperand mOp = srcNode->getInst()->getOperand(I->first); - - //Node is a Def, so add output dep. - if(nodeIsDef) { - if(mOp.isUse()) - srcNode->addOutEdge(destNode, MSchedGraphEdge::ValueDep, - MSchedGraphEdge::AntiDep, diff); - if(mOp.isDef()) - srcNode->addOutEdge(destNode, MSchedGraphEdge::ValueDep, - MSchedGraphEdge::OutputDep, diff); - - } - if(nodeIsUse) { - if(mOp.isDef()) - srcNode->addOutEdge(destNode, MSchedGraphEdge::ValueDep, - MSchedGraphEdge::TrueDep, diff); - } - } -} - - -void MSchedGraph::addMachRegEdges(std::map >& regNumtoNodeMap) { - //Loop over all machine registers in the map, and add dependencies - //between the instructions that use it - typedef std::map > regNodeMap; - for(regNodeMap::iterator I = regNumtoNodeMap.begin(); I != regNumtoNodeMap.end(); ++I) { - //Get the register number - int regNum = (*I).first; - - //Get Vector of nodes that use this register - std::vector Nodes = (*I).second; - - //Loop over nodes and determine the dependence between the other - //nodes in the vector - for(unsigned i =0; i < Nodes.size(); ++i) { - - //Get src node operator index that uses this machine register - int srcOpIndex = Nodes[i].first; - - //Get the actual src Node - MSchedGraphNode *srcNode = Nodes[i].second; - - //Get Operand - const MachineOperand &srcMOp = srcNode->getInst()->getOperand(srcOpIndex); - - bool srcIsUseandDef = srcMOp.isDef() && srcMOp.isUse(); - bool srcIsUse = srcMOp.isUse() && !srcMOp.isDef(); - - - //Look at all instructions after this in execution order - for(unsigned j=i+1; j < Nodes.size(); ++j) { - - //Sink node is a write - if(Nodes[j].second->getInst()->getOperand(Nodes[j].first).isDef()) { - //Src only uses the register (read) - if(srcIsUse) - srcNode->addOutEdge(Nodes[j].second, MSchedGraphEdge::MachineRegister, - MSchedGraphEdge::AntiDep); - - else if(srcIsUseandDef) { - srcNode->addOutEdge(Nodes[j].second, MSchedGraphEdge::MachineRegister, - MSchedGraphEdge::AntiDep); - - srcNode->addOutEdge(Nodes[j].second, MSchedGraphEdge::MachineRegister, - MSchedGraphEdge::OutputDep); - } - else - srcNode->addOutEdge(Nodes[j].second, MSchedGraphEdge::MachineRegister, - MSchedGraphEdge::OutputDep); - } - //Dest node is a read - else { - if(!srcIsUse || srcIsUseandDef) - srcNode->addOutEdge(Nodes[j].second, MSchedGraphEdge::MachineRegister, - MSchedGraphEdge::TrueDep); - } - - } - - //Look at all the instructions before this one since machine registers - //could live across iterations. - for(unsigned j = 0; j < i; ++j) { - //Sink node is a write - if(Nodes[j].second->getInst()->getOperand(Nodes[j].first).isDef()) { - //Src only uses the register (read) - if(srcIsUse) - srcNode->addOutEdge(Nodes[j].second, MSchedGraphEdge::MachineRegister, - MSchedGraphEdge::AntiDep, 1); - - else if(srcIsUseandDef) { - srcNode->addOutEdge(Nodes[j].second, MSchedGraphEdge::MachineRegister, - MSchedGraphEdge::AntiDep, 1); - - srcNode->addOutEdge(Nodes[j].second, MSchedGraphEdge::MachineRegister, - MSchedGraphEdge::OutputDep, 1); - } - else - srcNode->addOutEdge(Nodes[j].second, MSchedGraphEdge::MachineRegister, - MSchedGraphEdge::OutputDep, 1); - } - //Dest node is a read - else { - if(!srcIsUse || srcIsUseandDef) - srcNode->addOutEdge(Nodes[j].second, MSchedGraphEdge::MachineRegister, - MSchedGraphEdge::TrueDep,1 ); - } - - - } - - } - - } - -} - -void MSchedGraph::addMemEdges(const std::vector& memInst) { - - //Get Target machine instruction info - const TargetInstrInfo *TMI = Target.getInstrInfo(); - - //Loop over all memory instructions in the vector - //Knowing that they are in execution, add true, anti, and output dependencies - for (unsigned srcIndex = 0; srcIndex < memInst.size(); ++srcIndex) { - - //Get the machine opCode to determine type of memory instruction - MachineOpCode srcNodeOpCode = memInst[srcIndex]->getInst()->getOpcode(); - - //All instructions after this one in execution order have an iteration delay of 0 - for(unsigned destIndex = srcIndex + 1; destIndex < memInst.size(); ++destIndex) { - - //source is a Load, so add anti-dependencies (store after load) - if(TMI->isLoad(srcNodeOpCode)) - if(TMI->isStore(memInst[destIndex]->getInst()->getOpcode())) - memInst[srcIndex]->addOutEdge(memInst[destIndex], - MSchedGraphEdge::MemoryDep, - MSchedGraphEdge::AntiDep); - - //If source is a store, add output and true dependencies - if(TMI->isStore(srcNodeOpCode)) { - if(TMI->isStore(memInst[destIndex]->getInst()->getOpcode())) - memInst[srcIndex]->addOutEdge(memInst[destIndex], - MSchedGraphEdge::MemoryDep, - MSchedGraphEdge::OutputDep); - else - memInst[srcIndex]->addOutEdge(memInst[destIndex], - MSchedGraphEdge::MemoryDep, - MSchedGraphEdge::TrueDep); - } - } - - //All instructions before the src in execution order have an iteration delay of 1 - for(unsigned destIndex = 0; destIndex < srcIndex; ++destIndex) { - //source is a Load, so add anti-dependencies (store after load) - if(TMI->isLoad(srcNodeOpCode)) - if(TMI->isStore(memInst[destIndex]->getInst()->getOpcode())) - memInst[srcIndex]->addOutEdge(memInst[destIndex], - MSchedGraphEdge::MemoryDep, - MSchedGraphEdge::AntiDep, 1); - if(TMI->isStore(srcNodeOpCode)) { - if(TMI->isStore(memInst[destIndex]->getInst()->getOpcode())) - memInst[srcIndex]->addOutEdge(memInst[destIndex], - MSchedGraphEdge::MemoryDep, - MSchedGraphEdge::OutputDep, 1); - else - memInst[srcIndex]->addOutEdge(memInst[destIndex], - MSchedGraphEdge::MemoryDep, - MSchedGraphEdge::TrueDep, 1); - } - - } - - } -} diff --git a/lib/CodeGen/ModuloScheduling/MSchedGraph.h b/lib/CodeGen/ModuloScheduling/MSchedGraph.h deleted file mode 100644 index 4ea572a380b..00000000000 --- a/lib/CodeGen/ModuloScheduling/MSchedGraph.h +++ /dev/null @@ -1,316 +0,0 @@ -//===-- MSchedGraph.h - Scheduling Graph ------------------------*- 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. -// -//===----------------------------------------------------------------------===// -// -// A graph class for dependencies -// -//===----------------------------------------------------------------------===// - -#ifndef LLVM_MSCHEDGRAPH_H -#define LLVM_MSCHEDGRAPH_H - -#include "llvm/CodeGen/MachineInstr.h" -#include "llvm/Target/TargetMachine.h" -#include "llvm/ADT/GraphTraits.h" -#include "llvm/ADT/STLExtras.h" -#include "llvm/ADT/iterator" -#include - -namespace llvm { - class MSchedGraph; - class MSchedGraphNode; - template - class MSchedGraphNodeIterator; - - - struct MSchedGraphEdge { - enum DataDepOrderType { - TrueDep, AntiDep, OutputDep, NonDataDep - }; - - enum MSchedGraphEdgeType { - MemoryDep, ValueDep, MachineRegister - }; - - MSchedGraphNode *getDest() const { return dest; } - unsigned getIteDiff() { return iteDiff; } - unsigned getDepOrderType() { return depOrderType; } - - private: - friend class MSchedGraphNode; - MSchedGraphEdge(MSchedGraphNode *destination, MSchedGraphEdgeType type, - unsigned deptype, unsigned diff) - : dest(destination), depType(type), depOrderType(deptype), iteDiff(diff) {} - - MSchedGraphNode *dest; - MSchedGraphEdgeType depType; - unsigned depOrderType; - unsigned iteDiff; - }; - - class MSchedGraphNode { - - const MachineInstr* Inst; //Machine Instruction - MSchedGraph* Parent; //Graph this node belongs to - unsigned latency; //Latency of Instruction - bool isBranchInstr; //Is this node the branch instr or not - - std::vector Predecessors; //Predecessor Nodes - std::vector Successors; - - public: - MSchedGraphNode(const MachineInstr *inst, MSchedGraph *graph, - unsigned late=0, bool isBranch=false); - - //Iterators - typedef std::vector::iterator pred_iterator; - pred_iterator pred_begin() { return Predecessors.begin(); } - pred_iterator pred_end() { return Predecessors.end(); } - - typedef std::vector::const_iterator pred_const_iterator; - pred_const_iterator pred_begin() const { return Predecessors.begin(); } - pred_const_iterator pred_end() const { return Predecessors.end(); } - - // Successor iterators. - typedef MSchedGraphNodeIterator::const_iterator, - const MSchedGraphNode> succ_const_iterator; - succ_const_iterator succ_begin() const; - succ_const_iterator succ_end() const; - - typedef MSchedGraphNodeIterator::iterator, - MSchedGraphNode> succ_iterator; - succ_iterator succ_begin(); - succ_iterator succ_end(); - - - - void addOutEdge(MSchedGraphNode *destination, - MSchedGraphEdge::MSchedGraphEdgeType type, - unsigned deptype, unsigned diff=0) { - Successors.push_back(MSchedGraphEdge(destination, type, deptype,diff)); - destination->Predecessors.push_back(this); - } - const MachineInstr* getInst() { return Inst; } - MSchedGraph* getParent() { return Parent; } - bool hasPredecessors() { return (Predecessors.size() > 0); } - bool hasSuccessors() { return (Successors.size() > 0); } - int getLatency() { return latency; } - MSchedGraphEdge getInEdge(MSchedGraphNode *pred); - unsigned getInEdgeNum(MSchedGraphNode *pred); - - bool isSuccessor(MSchedGraphNode *); - bool isPredecessor(MSchedGraphNode *); - bool isBranch() { return isBranchInstr; } - //Debug support - void print(std::ostream &os) const; - - }; - - template - class MSchedGraphNodeIterator : public forward_iterator { - IteratorType I; // std::vector::iterator or const_iterator - public: - MSchedGraphNodeIterator(IteratorType i) : I(i) {} - - bool operator==(const MSchedGraphNodeIterator RHS) const { return I == RHS.I; } - bool operator!=(const MSchedGraphNodeIterator RHS) const { return I != RHS.I; } - - const MSchedGraphNodeIterator &operator=(const MSchedGraphNodeIterator &RHS) { - I = RHS.I; - return *this; - } - - NodeType* operator*() const { - return I->getDest(); - } - NodeType* operator->() const { return operator*(); } - - MSchedGraphNodeIterator& operator++() { // Preincrement - ++I; - return *this; - } - MSchedGraphNodeIterator operator++(int) { // Postincrement - MSchedGraphNodeIterator tmp = *this; ++*this; return tmp; - } - - MSchedGraphEdge &getEdge() { - return *I; - } - const MSchedGraphEdge &getEdge() const { - return *I; - } - }; - - inline MSchedGraphNode::succ_const_iterator MSchedGraphNode::succ_begin() const { - return succ_const_iterator(Successors.begin()); - } - inline MSchedGraphNode::succ_const_iterator MSchedGraphNode::succ_end() const { - return succ_const_iterator(Successors.end()); - } - inline MSchedGraphNode::succ_iterator MSchedGraphNode::succ_begin() { - return succ_iterator(Successors.begin()); - } - inline MSchedGraphNode::succ_iterator MSchedGraphNode::succ_end() { - return succ_iterator(Successors.end()); - } - - // ostream << operator for MSGraphNode class - inline std::ostream &operator<<(std::ostream &os, - const MSchedGraphNode &node) { - node.print(os); - return os; - } - - - - class MSchedGraph { - - const MachineBasicBlock *BB; //Machine basic block - const TargetMachine &Target; //Target Machine - - //Nodes - std::map GraphMap; - - //Add Nodes and Edges to this graph for our BB - typedef std::pair OpIndexNodePair; - void buildNodesAndEdges(); - void addValueEdges(std::vector &NodesInMap, - MSchedGraphNode *node, - bool nodeIsUse, bool nodeIsDef, int diff=0); - void addMachRegEdges(std::map >& regNumtoNodeMap); - void addMemEdges(const std::vector& memInst); - - public: - MSchedGraph(const MachineBasicBlock *bb, const TargetMachine &targ); - ~MSchedGraph(); - - //Add Nodes to the Graph - void addNode(const MachineInstr* MI, MSchedGraphNode *node); - - //iterators - typedef std::map::iterator iterator; - typedef std::map::const_iterator const_iterator; - typedef std::map::reverse_iterator reverse_iterator; - iterator find(const MachineInstr* I) { return GraphMap.find(I); } - iterator end() { return GraphMap.end(); } - iterator begin() { return GraphMap.begin(); } - reverse_iterator rbegin() { return GraphMap.rbegin(); } - reverse_iterator rend() { return GraphMap.rend(); } - const TargetMachine* getTarget() { return &Target; } - }; - - - static MSchedGraphNode& getSecond(std::pair &Pair) { - return *Pair.second; - } - - - - // Provide specializations of GraphTraits to be able to use graph - // iterators on the scheduling graph! - // - template <> struct GraphTraits { - typedef MSchedGraphNode NodeType; - typedef MSchedGraphNode::succ_iterator ChildIteratorType; - - static inline ChildIteratorType child_begin(NodeType *N) { - return N->succ_begin(); - } - static inline ChildIteratorType child_end(NodeType *N) { - return N->succ_end(); - } - - typedef std::pointer_to_unary_function&, MSchedGraphNode&> DerefFun; - - typedef mapped_iterator nodes_iterator; - static nodes_iterator nodes_begin(MSchedGraph *G) { - return map_iterator(((MSchedGraph*)G)->begin(), DerefFun(getSecond)); - } - static nodes_iterator nodes_end(MSchedGraph *G) { - return map_iterator(((MSchedGraph*)G)->end(), DerefFun(getSecond)); - } - - - }; - - template <> struct GraphTraits { - typedef const MSchedGraphNode NodeType; - typedef MSchedGraphNode::succ_const_iterator ChildIteratorType; - - static inline ChildIteratorType child_begin(NodeType *N) { - return N->succ_begin(); - } - static inline ChildIteratorType child_end(NodeType *N) { - return N->succ_end(); - } - typedef std::pointer_to_unary_function&, MSchedGraphNode&> DerefFun; - - typedef mapped_iterator nodes_iterator; - static nodes_iterator nodes_begin(MSchedGraph *G) { - return map_iterator(((MSchedGraph*)G)->begin(), DerefFun(getSecond)); - } - static nodes_iterator nodes_end(MSchedGraph *G) { - return map_iterator(((MSchedGraph*)G)->end(), DerefFun(getSecond)); - } - }; - - template <> struct GraphTraits > { - typedef MSchedGraphNode NodeType; - typedef MSchedGraphNode::pred_iterator ChildIteratorType; - - static inline ChildIteratorType child_begin(NodeType *N) { - return N->pred_begin(); - } - static inline ChildIteratorType child_end(NodeType *N) { - return N->pred_end(); - } - typedef std::pointer_to_unary_function&, MSchedGraphNode&> DerefFun; - - typedef mapped_iterator nodes_iterator; - static nodes_iterator nodes_begin(MSchedGraph *G) { - return map_iterator(((MSchedGraph*)G)->begin(), DerefFun(getSecond)); - } - static nodes_iterator nodes_end(MSchedGraph *G) { - return map_iterator(((MSchedGraph*)G)->end(), DerefFun(getSecond)); - } - }; - - template <> struct GraphTraits > { - typedef const MSchedGraphNode NodeType; - typedef MSchedGraphNode::pred_const_iterator ChildIteratorType; - - static inline ChildIteratorType child_begin(NodeType *N) { - return N->pred_begin(); - } - static inline ChildIteratorType child_end(NodeType *N) { - return N->pred_end(); - } - - typedef std::pointer_to_unary_function&, MSchedGraphNode&> DerefFun; - - typedef mapped_iterator nodes_iterator; - static nodes_iterator nodes_begin(MSchedGraph *G) { - return map_iterator(((MSchedGraph*)G)->begin(), DerefFun(getSecond)); - } - static nodes_iterator nodes_end(MSchedGraph *G) { - return map_iterator(((MSchedGraph*)G)->end(), DerefFun(getSecond)); - } - }; - - - - -} - -#endif diff --git a/lib/CodeGen/ModuloScheduling/Makefile b/lib/CodeGen/ModuloScheduling/Makefile deleted file mode 100644 index febdbf8d18d..00000000000 --- a/lib/CodeGen/ModuloScheduling/Makefile +++ /dev/null @@ -1,15 +0,0 @@ -##===- lib/CodeGen/ModuloScheduling/Makefile ---------------*- Makefile -*-===## -# -# 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. -# -##===----------------------------------------------------------------------===## -LEVEL = ../../.. - -DIRS = - -LIBRARYNAME = modulosched - -include $(LEVEL)/Makefile.common diff --git a/lib/CodeGen/ModuloScheduling/Makefile.am b/lib/CodeGen/ModuloScheduling/Makefile.am deleted file mode 100644 index e1fd097d381..00000000000 --- a/lib/CodeGen/ModuloScheduling/Makefile.am +++ /dev/null @@ -1,19 +0,0 @@ -#===-- lib/CodeGen/ModuloScheduling/Makefile.am ------------*- Makefile -*--===# -# -# The LLVM Compiler Infrastructure -# -# This file was developed by Reid Spencer and is distributed under the -# University of Illinois Open Source License. See LICENSE.TXT for details. -# -#===------------------------------------------------------------------------===# - -include $(top_srcdir)/Makefile.rules.am - -libexec_PROGRAMS = LLVMModuloScheduling.o - -LLVMModuloScheduling_o_SOURCES = \ - ModuloScheduling.cpp \ - MSchedGraph.cpp \ - MSSchedule.cpp - -LIBS= diff --git a/lib/CodeGen/ModuloScheduling/ModuloScheduling.cpp b/lib/CodeGen/ModuloScheduling/ModuloScheduling.cpp deleted file mode 100644 index ffb3404ff8e..00000000000 --- a/lib/CodeGen/ModuloScheduling/ModuloScheduling.cpp +++ /dev/null @@ -1,1967 +0,0 @@ -//===-- ModuloScheduling.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. -// -//===----------------------------------------------------------------------===// - -#define DEBUG_TYPE "ModuloSched" - -#include "ModuloScheduling.h" -#include "llvm/Instructions.h" -#include "llvm/Function.h" -#include "llvm/CodeGen/MachineFunction.h" -#include "llvm/CodeGen/Passes.h" -#include "llvm/Support/CFG.h" -#include "llvm/Target/TargetSchedInfo.h" -#include "llvm/Support/Debug.h" -#include "llvm/Support/GraphWriter.h" -#include "llvm/ADT/StringExtras.h" -#include -#include -#include -#include -#include -#include -#include "../../Target/SparcV9/MachineCodeForInstruction.h" -#include "../../Target/SparcV9/SparcV9TmpInstr.h" -#include "../../Target/SparcV9/SparcV9Internals.h" -#include "../../Target/SparcV9/SparcV9RegisterInfo.h" -using namespace llvm; - -/// Create ModuloSchedulingPass -/// -FunctionPass *llvm::createModuloSchedulingPass(TargetMachine & targ) { - DEBUG(std::cerr << "Created ModuloSchedulingPass\n"); - return new ModuloSchedulingPass(targ); -} - - -//Graph Traits for printing out the dependence graph -template -static void WriteGraphToFile(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"; -}; - -//Graph Traits for printing out the dependence graph -namespace llvm { - - template<> - struct DOTGraphTraits : public DefaultDOTGraphTraits { - static std::string getGraphName(MSchedGraph *F) { - return "Dependence Graph"; - } - - static std::string getNodeLabel(MSchedGraphNode *Node, MSchedGraph *Graph) { - if (Node->getInst()) { - std::stringstream ss; - ss << *(Node->getInst()); - return ss.str(); //((MachineInstr*)Node->getInst()); - } - else - return "No Inst"; - } - static std::string getEdgeSourceLabel(MSchedGraphNode *Node, - MSchedGraphNode::succ_iterator I) { - //Label each edge with the type of dependence - std::string edgelabel = ""; - switch (I.getEdge().getDepOrderType()) { - - case MSchedGraphEdge::TrueDep: - edgelabel = "True"; - break; - - case MSchedGraphEdge::AntiDep: - edgelabel = "Anti"; - break; - - case MSchedGraphEdge::OutputDep: - edgelabel = "Output"; - break; - - default: - edgelabel = "Unknown"; - break; - } - - //FIXME - int iteDiff = I.getEdge().getIteDiff(); - std::string intStr = "(IteDiff: "; - intStr += itostr(iteDiff); - - intStr += ")"; - edgelabel += intStr; - - return edgelabel; - } - }; -} - -/// ModuloScheduling::runOnFunction - main transformation entry point -/// The Swing Modulo Schedule algorithm has three basic steps: -/// 1) Computation and Analysis of the dependence graph -/// 2) Ordering of the nodes -/// 3) Scheduling -/// -bool ModuloSchedulingPass::runOnFunction(Function &F) { - - bool Changed = false; - - DEBUG(std::cerr << "Creating ModuloSchedGraph for each valid BasicBlock in " + F.getName() + "\n"); - - //Get MachineFunction - MachineFunction &MF = MachineFunction::get(&F); - - //Worklist - std::vector Worklist; - - //Iterate over BasicBlocks and put them into our worklist if they are valid - for (MachineFunction::iterator BI = MF.begin(); BI != MF.end(); ++BI) - if(MachineBBisValid(BI)) - Worklist.push_back(&*BI); - - DEBUG(if(Worklist.size() == 0) std::cerr << "No single basic block loops in function to ModuloSchedule\n"); - - //Iterate over the worklist and perform scheduling - for(std::vector::iterator BI = Worklist.begin(), - BE = Worklist.end(); BI != BE; ++BI) { - - MSchedGraph *MSG = new MSchedGraph(*BI, target); - - //Write Graph out to file - DEBUG(WriteGraphToFile(std::cerr, F.getName(), MSG)); - - //Print out BB for debugging - DEBUG(std::cerr << "ModuloScheduling BB: \n"; (*BI)->print(std::cerr)); - - //Calculate Resource II - int ResMII = calculateResMII(*BI); - - //Calculate Recurrence II - int RecMII = calculateRecMII(MSG, ResMII); - - //Our starting initiation interval is the maximum of RecMII and ResMII - II = std::max(RecMII, ResMII); - - //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::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 >::iterator I = partialOrder.begin(), - E = partialOrder.end(); I !=E; ++I) { - std::cerr << "Start set in PO\n"; - for(std::vector::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::iterator I = FinalNodeOrder.begin(), E = FinalNodeOrder.end(); I != E; ++I) { - std::cerr << "FO:" << **I << "\n"; - }); - - //Finally schedule nodes - computeSchedule(); - - //Print out final schedule - DEBUG(schedule.print(std::cerr)); - - - //Final scheduling step is to reconstruct the loop - reconstructLoop(*BI); - - //Print out new loop - - - //Clear out our maps for the next basic block that is processed - nodeToAttributesMap.clear(); - partialOrder.clear(); - recurrenceList.clear(); - FinalNodeOrder.clear(); - schedule.clear(); - - //Clean up. Nuke old MachineBB and llvmBB - //BasicBlock *llvmBB = (BasicBlock*) (*BI)->getBasicBlock(); - //Function *parent = (Function*) llvmBB->getParent(); - //Should't std::find work?? - //parent->getBasicBlockList().erase(std::find(parent->getBasicBlockList().begin(), parent->getBasicBlockList().end(), *llvmBB)); - //parent->getBasicBlockList().erase(llvmBB); - - //delete(llvmBB); - //delete(*BI); - } - - - return Changed; -} - - -/// 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 ModuloSchedulingPass::MachineBBisValid(const MachineBasicBlock *BI) { - - bool isLoop = false; - - //Check first if its a valid loop - for(succ_const_iterator I = succ_begin(BI->getBasicBlock()), - E = succ_end(BI->getBasicBlock()); I != E; ++I) { - if (*I == BI->getBasicBlock()) // has single block loop - isLoop = true; - } - - if(!isLoop) - return false; - - //Get Target machine instruction info - const TargetInstrInfo *TMI = target.getInstrInfo(); - - //Check each instruction and look for calls - for(MachineBasicBlock::const_iterator I = BI->begin(), E = BI->end(); I != E; ++I) { - //Get opcode to check instruction type - MachineOpCode OC = I->getOpcode(); - if(TMI->isCall(OC)) - 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 ModuloSchedulingPass::calculateResMII(const MachineBasicBlock *BI) { - - 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 resourceUsageCount; - - 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 > 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.find(resources[i][j]) == resourceUsageCount.end()) { - 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::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; - 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 -int ModuloSchedulingPass::calculateRecMII(MSchedGraph *graph, int MII) { - std::vector vNodes; - //Loop over all nodes in the graph - for(MSchedGraph::iterator I = graph->begin(), E = graph->end(); I != E; ++I) { - findAllReccurrences(I->second, vNodes, MII); - vNodes.clear(); - } - - int RecMII = 0; - - for(std::set > >::iterator I = recurrenceList.begin(), E=recurrenceList.end(); I !=E; ++I) { - DEBUG(for(std::vector::const_iterator N = I->second.begin(), NE = I->second.end(); N != NE; ++N) { - std::cerr << **N << "\n"; - }); - RecMII = std::max(RecMII, I->first); - } - - return MII; -} - -/// calculateNodeAttributes - The following properties are calculated for -/// each node in the dependence graph: ASAP, ALAP, Depth, Height, and -/// MOB. -void ModuloSchedulingPass::calculateNodeAttributes(MSchedGraph *graph, int MII) { - - //Loop over the nodes and add them to the map - for(MSchedGraph::iterator I = graph->begin(), E = graph->end(); I != E; ++I) { - //Assert if its already in the map - assert(nodeToAttributesMap.find(I->second) == nodeToAttributesMap.end() && "Node attributes are already in the map"); - - //Put into the map with default attribute values - nodeToAttributesMap[I->second] = MSNodeAttributes(); - } - - //Create set to deal with reccurrences - std::set visitedNodes; - - //Now Loop over map and calculate the node attributes - for(std::map::iterator I = nodeToAttributesMap.begin(), E = nodeToAttributesMap.end(); I != E; ++I) { - calculateASAP(I->first, MII, (MSchedGraphNode*) 0); - visitedNodes.clear(); - } - - int maxASAP = findMaxASAP(); - //Calculate ALAP which depends on ASAP being totally calculated - for(std::map::iterator I = nodeToAttributesMap.begin(), E = nodeToAttributesMap.end(); I != E; ++I) { - calculateALAP(I->first, MII, maxASAP, (MSchedGraphNode*) 0); - visitedNodes.clear(); - } - - //Calculate MOB which depends on ASAP being totally calculated, also do depth and height - for(std::map::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, (MSchedGraphNode*) 0); - calculateHeight(I->first, (MSchedGraphNode*) 0); - } - - -} - -/// ignoreEdge - Checks to see if this edge of a recurrence should be ignored or not -bool ModuloSchedulingPass::ignoreEdge(MSchedGraphNode *srcNode, MSchedGraphNode *destNode) { - if(destNode == 0 || srcNode ==0) - return false; - - bool findEdge = edgesToIgnore.count(std::make_pair(srcNode, destNode->getInEdgeNum(srcNode))); - - return findEdge; -} - - -/// calculateASAP - Calculates the -int ModuloSchedulingPass::calculateASAP(MSchedGraphNode *node, int MII, MSchedGraphNode *destNode) { - - DEBUG(std::cerr << "Calculating ASAP for " << *node << "\n"); - - //Get current node attributes - MSNodeAttributes &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(MSchedGraphNode::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 << ")\n"); - - return maxPredValue; -} - - -int ModuloSchedulingPass::calculateALAP(MSchedGraphNode *node, int MII, - int maxASAP, MSchedGraphNode *srcNode) { - - DEBUG(std::cerr << "Calculating ALAP for " << *node << "\n"); - - MSNodeAttributes &attributes = nodeToAttributesMap.find(node)->second; - - if(attributes.ALAP != -1) - return attributes.ALAP; - - if(node->hasSuccessors()) { - - //Trying to deal with the issue where the node has successors, but - //we are ignoring all of the edges to them. So this is my hack for - //now.. there is probably a more elegant way of doing this (FIXME) - bool processedOneEdge = false; - - //FIXME, set to something high to start - int minSuccValue = 9999999; - - //Iterate over all of the predecessors and fine max - for(MSchedGraphNode::succ_iterator P = node->succ_begin(), - E = node->succ_end(); P != E; ++P) { - - //Only process if we are not ignoring the edge - if(!ignoreEdge(node, *P)) { - processedOneEdge = true; - int succALAP = -1; - succALAP = calculateALAP(*P, MII, maxASAP, node); - - assert(succALAP != -1 && "Successors ALAP should have been caclulated"); - - int iteDiff = P.getEdge().getIteDiff(); - - int currentSuccValue = succALAP - node->getLatency() + iteDiff * MII; - - DEBUG(std::cerr << "succ ALAP: " << succALAP << ", iteDiff: " << iteDiff << ", SuccLatency: " << (*P)->getLatency() << ", Current ALAP succ: " << currentSuccValue << "\n"); - - minSuccValue = std::min(minSuccValue, currentSuccValue); - } - } - - if(processedOneEdge) - attributes.ALAP = minSuccValue; - - else - attributes.ALAP = maxASAP; - } - else - attributes.ALAP = maxASAP; - - DEBUG(std::cerr << "ALAP: " << attributes.ALAP << " (" << *node << ")\n"); - - if(attributes.ALAP < 0) - attributes.ALAP = 0; - - return attributes.ALAP; -} - -int ModuloSchedulingPass::findMaxASAP() { - int maxASAP = 0; - - for(std::map::iterator I = nodeToAttributesMap.begin(), - E = nodeToAttributesMap.end(); I != E; ++I) - maxASAP = std::max(maxASAP, I->second.ASAP); - return maxASAP; -} - - -int ModuloSchedulingPass::calculateHeight(MSchedGraphNode *node,MSchedGraphNode *srcNode) { - - MSNodeAttributes &attributes = nodeToAttributesMap.find(node)->second; - - if(attributes.height != -1) - return attributes.height; - - int maxHeight = 0; - - //Iterate over all of the predecessors and find max - for(MSchedGraphNode::succ_iterator P = node->succ_begin(), - E = node->succ_end(); P != E; ++P) { - - - if(!ignoreEdge(node, *P)) { - int succHeight = calculateHeight(*P, node); - - assert(succHeight != -1 && "Successors Height should have been caclulated"); - - int currentHeight = succHeight + node->getLatency(); - maxHeight = std::max(maxHeight, currentHeight); - } - } - attributes.height = maxHeight; - DEBUG(std::cerr << "Height: " << attributes.height << " (" << *node << ")\n"); - return maxHeight; -} - - -int ModuloSchedulingPass::calculateDepth(MSchedGraphNode *node, - MSchedGraphNode *destNode) { - - MSNodeAttributes &attributes = nodeToAttributesMap.find(node)->second; - - if(attributes.depth != -1) - return attributes.depth; - - int maxDepth = 0; - - //Iterate over all of the predecessors and fine max - for(MSchedGraphNode::pred_iterator P = node->pred_begin(), E = node->pred_end(); P != E; ++P) { - - if(!ignoreEdge(*P, node)) { - int predDepth = -1; - predDepth = calculateDepth(*P, node); - - assert(predDepth != -1 && "Predecessors ASAP should have been caclulated"); - - int currentDepth = predDepth + (*P)->getLatency(); - maxDepth = std::max(maxDepth, currentDepth); - } - } - attributes.depth = maxDepth; - - DEBUG(std::cerr << "Depth: " << attributes.depth << " (" << *node << "*)\n"); - return maxDepth; -} - - - -void ModuloSchedulingPass::addReccurrence(std::vector &recurrence, int II, MSchedGraphNode *srcBENode, MSchedGraphNode *destBENode) { - //Check to make sure that this recurrence is unique - bool same = false; - - - //Loop over all recurrences already in our list - for(std::set > >::iterator R = recurrenceList.begin(), RE = recurrenceList.end(); R != RE; ++R) { - - bool all_same = true; - //First compare size - if(R->second.size() == recurrence.size()) { - - for(std::vector::const_iterator node = R->second.begin(), end = R->second.end(); node != end; ++node) { - if(std::find(recurrence.begin(), recurrence.end(), *node) == recurrence.end()) { - all_same = all_same && false; - break; - } - else - all_same = all_same && true; - } - if(all_same) { - same = true; - break; - } - } - } - - if(!same) { - srcBENode = recurrence.back(); - destBENode = recurrence.front(); - - //FIXME - if(destBENode->getInEdge(srcBENode).getIteDiff() == 0) { - //DEBUG(std::cerr << "NOT A BACKEDGE\n"); - //find actual backedge HACK HACK - for(unsigned i=0; i< recurrence.size()-1; ++i) { - if(recurrence[i+1]->getInEdge(recurrence[i]).getIteDiff() == 1) { - srcBENode = recurrence[i]; - destBENode = recurrence[i+1]; - break; - } - - } - - } - DEBUG(std::cerr << "Back Edge to Remove: " << *srcBENode << " to " << *destBENode << "\n"); - edgesToIgnore.insert(std::make_pair(srcBENode, destBENode->getInEdgeNum(srcBENode))); - recurrenceList.insert(std::make_pair(II, recurrence)); - } - -} - -void ModuloSchedulingPass::findAllReccurrences(MSchedGraphNode *node, - std::vector &visitedNodes, - int II) { - - if(std::find(visitedNodes.begin(), visitedNodes.end(), node) != visitedNodes.end()) { - std::vector recurrence; - bool first = true; - int delay = 0; - int distance = 0; - int RecMII = II; //Starting value - MSchedGraphNode *last = node; - MSchedGraphNode *srcBackEdge = 0; - MSchedGraphNode *destBackEdge = 0; - - - - for(std::vector::iterator I = visitedNodes.begin(), E = visitedNodes.end(); - I !=E; ++I) { - - if(*I == node) - first = false; - if(first) - continue; - - delay = delay + (*I)->getLatency(); - - if(*I != node) { - int diff = (*I)->getInEdge(last).getIteDiff(); - distance += diff; - if(diff > 0) { - srcBackEdge = last; - destBackEdge = *I; - } - } - - recurrence.push_back(*I); - last = *I; - } - - - - //Get final distance calc - distance += node->getInEdge(last).getIteDiff(); - - - //Adjust II until we get close to the inequality delay - II*distance <= 0 - - int value = delay-(RecMII * distance); - int lastII = II; - while(value <= 0) { - - lastII = RecMII; - RecMII--; - value = delay-(RecMII * distance); - } - - - DEBUG(std::cerr << "Final II for this recurrence: " << lastII << "\n"); - addReccurrence(recurrence, lastII, srcBackEdge, destBackEdge); - assert(distance != 0 && "Recurrence distance should not be zero"); - return; - } - - for(MSchedGraphNode::succ_iterator I = node->succ_begin(), E = node->succ_end(); I != E; ++I) { - visitedNodes.push_back(node); - findAllReccurrences(*I, visitedNodes, II); - visitedNodes.pop_back(); - } -} - - - - - -void ModuloSchedulingPass::computePartialOrder() { - - - //Loop over all recurrences and add to our partial order - //be sure to remove nodes that are already in the partial order in - //a different recurrence and don't add empty recurrences. - for(std::set > >::reverse_iterator I = recurrenceList.rbegin(), E=recurrenceList.rend(); I !=E; ++I) { - - //Add nodes that connect this recurrence to the previous recurrence - - //If this is the first recurrence in the partial order, add all predecessors - for(std::vector::const_iterator N = I->second.begin(), NE = I->second.end(); N != NE; ++N) { - - } - - - std::vector new_recurrence; - //Loop through recurrence and remove any nodes already in the partial order - for(std::vector::const_iterator N = I->second.begin(), NE = I->second.end(); N != NE; ++N) { - bool found = false; - for(std::vector >::iterator PO = partialOrder.begin(), PE = partialOrder.end(); PO != PE; ++PO) { - if(std::find(PO->begin(), PO->end(), *N) != PO->end()) - found = true; - } - if(!found) { - new_recurrence.push_back(*N); - - if(partialOrder.size() == 0) - //For each predecessors, add it to this recurrence ONLY if it is not already in it - for(MSchedGraphNode::pred_iterator P = (*N)->pred_begin(), - PE = (*N)->pred_end(); P != PE; ++P) { - - //Check if we are supposed to ignore this edge or not - if(!ignoreEdge(*P, *N)) - //Check if already in this recurrence - if(std::find(I->second.begin(), I->second.end(), *P) == I->second.end()) { - //Also need to check if in partial order - bool predFound = false; - for(std::vector >::iterator PO = partialOrder.begin(), PEND = partialOrder.end(); PO != PEND; ++PO) { - if(std::find(PO->begin(), PO->end(), *P) != PO->end()) - predFound = true; - } - - if(!predFound) - if(std::find(new_recurrence.begin(), new_recurrence.end(), *P) == new_recurrence.end()) - new_recurrence.push_back(*P); - - } - } - } - } - - - if(new_recurrence.size() > 0) - partialOrder.push_back(new_recurrence); - } - - //Add any nodes that are not already in the partial order - std::vector lastNodes; - for(std::map::iterator I = nodeToAttributesMap.begin(), E = nodeToAttributesMap.end(); I != E; ++I) { - bool found = false; - //Check if its already in our partial order, if not add it to the final vector - for(std::vector >::iterator PO = partialOrder.begin(), PE = partialOrder.end(); PO != PE; ++PO) { - if(std::find(PO->begin(), PO->end(), I->first) != PO->end()) - found = true; - } - if(!found) - lastNodes.push_back(I->first); - } - - if(lastNodes.size() > 0) - partialOrder.push_back(lastNodes); - -} - - -void ModuloSchedulingPass::predIntersect(std::vector &CurrentSet, std::vector &IntersectResult) { - - //Sort CurrentSet so we can use lowerbound - std::sort(CurrentSet.begin(), CurrentSet.end()); - - for(unsigned j=0; j < FinalNodeOrder.size(); ++j) { - for(MSchedGraphNode::pred_iterator P = FinalNodeOrder[j]->pred_begin(), - E = FinalNodeOrder[j]->pred_end(); P != E; ++P) { - - //Check if we are supposed to ignore this edge or not - if(ignoreEdge(*P,FinalNodeOrder[j])) - continue; - - if(std::find(CurrentSet.begin(), - CurrentSet.end(), *P) != CurrentSet.end()) - if(std::find(FinalNodeOrder.begin(), FinalNodeOrder.end(), *P) == FinalNodeOrder.end()) - IntersectResult.push_back(*P); - } - } -} - -void ModuloSchedulingPass::succIntersect(std::vector &CurrentSet, std::vector &IntersectResult) { - - //Sort CurrentSet so we can use lowerbound - std::sort(CurrentSet.begin(), CurrentSet.end()); - - for(unsigned j=0; j < FinalNodeOrder.size(); ++j) { - for(MSchedGraphNode::succ_iterator P = FinalNodeOrder[j]->succ_begin(), - E = FinalNodeOrder[j]->succ_end(); P != E; ++P) { - - //Check if we are supposed to ignore this edge or not - if(ignoreEdge(FinalNodeOrder[j],*P)) - continue; - - if(std::find(CurrentSet.begin(), - CurrentSet.end(), *P) != CurrentSet.end()) - if(std::find(FinalNodeOrder.begin(), FinalNodeOrder.end(), *P) == FinalNodeOrder.end()) - IntersectResult.push_back(*P); - } - } -} - -void dumpIntersection(std::vector &IntersectCurrent) { - std::cerr << "Intersection ("; - for(std::vector::iterator I = IntersectCurrent.begin(), E = IntersectCurrent.end(); I != E; ++I) - std::cerr << **I << ", "; - std::cerr << ")\n"; -} - - - -void ModuloSchedulingPass::orderNodes() { - - int BOTTOM_UP = 0; - int TOP_DOWN = 1; - - //Set default order - int order = BOTTOM_UP; - - - //Loop over all the sets and place them in the final node order - for(std::vector >::iterator CurrentSet = partialOrder.begin(), E= partialOrder.end(); CurrentSet != E; ++CurrentSet) { - - DEBUG(std::cerr << "Processing set in S\n"); - DEBUG(dumpIntersection(*CurrentSet)); - - //Result of intersection - std::vector IntersectCurrent; - - predIntersect(*CurrentSet, IntersectCurrent); - - //If the intersection of predecessor and current set is not empty - //sort nodes bottom up - if(IntersectCurrent.size() != 0) { - DEBUG(std::cerr << "Final Node Order Predecessors and Current Set interesection is NOT empty\n"); - order = BOTTOM_UP; - } - //If empty, use successors - else { - DEBUG(std::cerr << "Final Node Order Predecessors and Current Set interesection is empty\n"); - - succIntersect(*CurrentSet, IntersectCurrent); - - //sort top-down - if(IntersectCurrent.size() != 0) { - DEBUG(std::cerr << "Final Node Order Successors and Current Set interesection is NOT empty\n"); - order = TOP_DOWN; - } - else { - DEBUG(std::cerr << "Final Node Order Successors and Current Set interesection is empty\n"); - //Find node with max ASAP in current Set - MSchedGraphNode *node; - int maxASAP = 0; - DEBUG(std::cerr << "Using current set of size " << CurrentSet->size() << "to find max ASAP\n"); - for(unsigned j=0; j < CurrentSet->size(); ++j) { - //Get node attributes - MSNodeAttributes nodeAttr= nodeToAttributesMap.find((*CurrentSet)[j])->second; - //assert(nodeAttr != nodeToAttributesMap.end() && "Node not in attributes map!"); - DEBUG(std::cerr << "CurrentSet index " << j << "has ASAP: " << nodeAttr.ASAP << "\n"); - if(maxASAP < nodeAttr.ASAP) { - maxASAP = nodeAttr.ASAP; - node = (*CurrentSet)[j]; - } - } - assert(node != 0 && "In node ordering node should not be null"); - IntersectCurrent.push_back(node); - order = BOTTOM_UP; - } - } - - //Repeat until all nodes are put into the final order from current set - while(IntersectCurrent.size() > 0) { - - if(order == TOP_DOWN) { - DEBUG(std::cerr << "Order is TOP DOWN\n"); - - while(IntersectCurrent.size() > 0) { - DEBUG(std::cerr << "Intersection is not empty, so find heighest height\n"); - - int MOB = 0; - int height = 0; - MSchedGraphNode *highestHeightNode = IntersectCurrent[0]; - - //Find node in intersection with highest heigh and lowest MOB - for(std::vector::iterator I = IntersectCurrent.begin(), - E = IntersectCurrent.end(); I != E; ++I) { - - //Get current nodes properties - MSNodeAttributes nodeAttr= nodeToAttributesMap.find(*I)->second; - - if(height < nodeAttr.height) { - highestHeightNode = *I; - height = nodeAttr.height; - MOB = nodeAttr.MOB; - } - else if(height == nodeAttr.height) { - if(MOB > nodeAttr.height) { - highestHeightNode = *I; - height = nodeAttr.height; - MOB = nodeAttr.MOB; - } - } - } - - //Append our node with greatest height to the NodeOrder - if(std::find(FinalNodeOrder.begin(), FinalNodeOrder.end(), highestHeightNode) == FinalNodeOrder.end()) { - DEBUG(std::cerr << "Adding node to Final Order: " << *highestHeightNode << "\n"); - FinalNodeOrder.push_back(highestHeightNode); - } - - //Remove V from IntersectOrder - IntersectCurrent.erase(std::find(IntersectCurrent.begin(), - IntersectCurrent.end(), highestHeightNode)); - - - //Intersect V's successors with CurrentSet - for(MSchedGraphNode::succ_iterator P = highestHeightNode->succ_begin(), - E = highestHeightNode->succ_end(); P != E; ++P) { - //if(lower_bound(CurrentSet->begin(), - // CurrentSet->end(), *P) != CurrentSet->end()) { - if(std::find(CurrentSet->begin(), CurrentSet->end(), *P) != CurrentSet->end()) { - if(ignoreEdge(highestHeightNode, *P)) - continue; - //If not already in Intersect, add - if(std::find(IntersectCurrent.begin(), IntersectCurrent.end(), *P) == IntersectCurrent.end()) - IntersectCurrent.push_back(*P); - } - } - } //End while loop over Intersect Size - - //Change direction - order = BOTTOM_UP; - - //Reset Intersect to reflect changes in OrderNodes - IntersectCurrent.clear(); - predIntersect(*CurrentSet, IntersectCurrent); - - } //End If TOP_DOWN - - //Begin if BOTTOM_UP - else { - DEBUG(std::cerr << "Order is BOTTOM UP\n"); - while(IntersectCurrent.size() > 0) { - DEBUG(std::cerr << "Intersection of size " << IntersectCurrent.size() << ", finding highest depth\n"); - - //dump intersection - DEBUG(dumpIntersection(IntersectCurrent)); - //Get node with highest depth, if a tie, use one with lowest - //MOB - int MOB = 0; - int depth = 0; - MSchedGraphNode *highestDepthNode = IntersectCurrent[0]; - - for(std::vector::iterator I = IntersectCurrent.begin(), - E = IntersectCurrent.end(); I != E; ++I) { - //Find node attribute in graph - MSNodeAttributes nodeAttr= nodeToAttributesMap.find(*I)->second; - - if(depth < nodeAttr.depth) { - highestDepthNode = *I; - depth = nodeAttr.depth; - MOB = nodeAttr.MOB; - } - else if(depth == nodeAttr.depth) { - if(MOB > nodeAttr.MOB) { - highestDepthNode = *I; - depth = nodeAttr.depth; - MOB = nodeAttr.MOB; - } - } - } - - - - //Append highest depth node to the NodeOrder - if(std::find(FinalNodeOrder.begin(), FinalNodeOrder.end(), highestDepthNode) == FinalNodeOrder.end()) { - DEBUG(std::cerr << "Adding node to Final Order: " << *highestDepthNode << "\n"); - FinalNodeOrder.push_back(highestDepthNode); - } - //Remove heightestDepthNode from IntersectOrder - IntersectCurrent.erase(std::find(IntersectCurrent.begin(), - IntersectCurrent.end(),highestDepthNode)); - - - //Intersect heightDepthNode's pred with CurrentSet - for(MSchedGraphNode::pred_iterator P = highestDepthNode->pred_begin(), - E = highestDepthNode->pred_end(); P != E; ++P) { - //if(lower_bound(CurrentSet->begin(), - // CurrentSet->end(), *P) != CurrentSet->end()) { - if(std::find(CurrentSet->begin(), CurrentSet->end(), *P) != CurrentSet->end()) { - - if(ignoreEdge(*P, highestDepthNode)) - continue; - - //If not already in Intersect, add - if(std::find(IntersectCurrent.begin(), - IntersectCurrent.end(), *P) == IntersectCurrent.end()) - IntersectCurrent.push_back(*P); - } - } - - } //End while loop over Intersect Size - - //Change order - order = TOP_DOWN; - - //Reset IntersectCurrent to reflect changes in OrderNodes - IntersectCurrent.clear(); - succIntersect(*CurrentSet, IntersectCurrent); - } //End if BOTTOM_DOWN - - DEBUG(std::cerr << "Current Intersection Size: " << IntersectCurrent.size() << "\n"); - } - //End Wrapping while loop - DEBUG(std::cerr << "Ending Size of Current Set: " << CurrentSet->size() << "\n"); - }//End for over all sets of nodes - - //FIXME: As the algorithm stands it will NEVER add an instruction such as ba (with no - //data dependencies) to the final order. We add this manually. It will always be - //in the last set of S since its not part of a recurrence - //Loop over all the sets and place them in the final node order - std::vector > ::reverse_iterator LastSet = partialOrder.rbegin(); - for(std::vector::iterator CurrentNode = LastSet->begin(), LastNode = LastSet->end(); - CurrentNode != LastNode; ++CurrentNode) { - if((*CurrentNode)->getInst()->getOpcode() == V9::BA) - FinalNodeOrder.push_back(*CurrentNode); - } - //Return final Order - //return FinalNodeOrder; -} - -void ModuloSchedulingPass::computeSchedule() { - - bool success = false; - - while(!success) { - - //Loop over the final node order and process each node - for(std::vector::iterator I = FinalNodeOrder.begin(), - E = FinalNodeOrder.end(); I != E; ++I) { - - //CalculateEarly and Late start - int EarlyStart = -1; - int LateStart = 99999; //Set to something higher then we would ever expect (FIXME) - bool hasSucc = false; - bool hasPred = false; - - if(!(*I)->isBranch()) { - //Loop over nodes in the schedule and determine if they are predecessors - //or successors of the node we are trying to schedule - for(MSSchedule::schedule_iterator nodesByCycle = schedule.begin(), nodesByCycleEnd = schedule.end(); - nodesByCycle != nodesByCycleEnd; ++nodesByCycle) { - - //For this cycle, get the vector of nodes schedule and loop over it - for(std::vector::iterator schedNode = nodesByCycle->second.begin(), SNE = nodesByCycle->second.end(); schedNode != SNE; ++schedNode) { - - if((*I)->isPredecessor(*schedNode)) { - if(!ignoreEdge(*schedNode, *I)) { - int diff = (*I)->getInEdge(*schedNode).getIteDiff(); - int ES_Temp = nodesByCycle->first + (*schedNode)->getLatency() - diff * II; - DEBUG(std::cerr << "Diff: " << diff << " Cycle: " << nodesByCycle->first << "\n"); - DEBUG(std::cerr << "Temp EarlyStart: " << ES_Temp << " Prev EarlyStart: " << EarlyStart << "\n"); - EarlyStart = std::max(EarlyStart, ES_Temp); - hasPred = true; - } - } - if((*I)->isSuccessor(*schedNode)) { - if(!ignoreEdge(*I,*schedNode)) { - int diff = (*schedNode)->getInEdge(*I).getIteDiff(); - int LS_Temp = nodesByCycle->first - (*I)->getLatency() + diff * II; - DEBUG(std::cerr << "Diff: " << diff << " Cycle: " << nodesByCycle->first << "\n"); - DEBUG(std::cerr << "Temp LateStart: " << LS_Temp << " Prev LateStart: " << LateStart << "\n"); - LateStart = std::min(LateStart, LS_Temp); - hasSucc = true; - } - } - } - } - } - else { - //WARNING: HACK! FIXME!!!! - if((*I)->getInst()->getOpcode() == V9::BA) { - EarlyStart = II-1; - LateStart = II-1; - } - else { - EarlyStart = II-1; - LateStart = II-1; - assert( (EarlyStart >= 0) && (LateStart >=0) && "EarlyStart and LateStart must be greater then 0"); - } - hasPred = 1; - hasSucc = 1; - } - - - DEBUG(std::cerr << "Has Successors: " << hasSucc << ", Has Pred: " << hasPred << "\n"); - DEBUG(std::cerr << "EarlyStart: " << EarlyStart << ", LateStart: " << LateStart << "\n"); - - //Check if the node has no pred or successors and set Early Start to its ASAP - if(!hasSucc && !hasPred) - EarlyStart = nodeToAttributesMap.find(*I)->second.ASAP; - - //Now, try to schedule this node depending upon its pred and successor in the schedule - //already - if(!hasSucc && hasPred) - success = scheduleNode(*I, EarlyStart, (EarlyStart + II -1)); - else if(!hasPred && hasSucc) - success = scheduleNode(*I, LateStart, (LateStart - II +1)); - else if(hasPred && hasSucc) - success = scheduleNode(*I, EarlyStart, std::min(LateStart, (EarlyStart + II -1))); - else - success = scheduleNode(*I, EarlyStart, EarlyStart + II - 1); - - if(!success) { - ++II; - schedule.clear(); - break; - } - - } - - DEBUG(std::cerr << "Constructing Kernel\n"); - success = schedule.constructKernel(II); - if(!success) { - ++II; - schedule.clear(); - } - } -} - - -bool ModuloSchedulingPass::scheduleNode(MSchedGraphNode *node, - int start, int end) { - bool success = false; - - DEBUG(std::cerr << *node << " (Start Cycle: " << start << ", End Cycle: " << end << ")\n"); - - //Make sure start and end are not negative - if(start < 0) - start = 0; - if(end < 0) - end = 0; - - bool forward = true; - if(start > end) - forward = false; - - bool increaseSC = true; - int cycle = start ; - - - while(increaseSC) { - - increaseSC = false; - - increaseSC = schedule.insert(node, cycle); - - if(!increaseSC) - return true; - - //Increment cycle to try again - if(forward) { - ++cycle; - DEBUG(std::cerr << "Increase cycle: " << cycle << "\n"); - if(cycle > end) - return false; - } - else { - --cycle; - DEBUG(std::cerr << "Decrease cycle: " << cycle << "\n"); - if(cycle < end) - return false; - } - } - - return success; -} - -void ModuloSchedulingPass::writePrologues(std::vector &prologues, MachineBasicBlock *origBB, std::vector &llvm_prologues, std::map > &valuesToSave, std::map > &newValues, std::map &newValLocation) { - - //Keep a map to easily know whats in the kernel - std::map > inKernel; - int maxStageCount = 0; - - MSchedGraphNode *branch = 0; - MSchedGraphNode *BAbranch = 0; - - for(MSSchedule::kernel_iterator I = schedule.kernel_begin(), E = schedule.kernel_end(); I != E; ++I) { - maxStageCount = std::max(maxStageCount, I->second); - - //Ignore the branch, we will handle this separately - if(I->first->isBranch()) { - if (I->first->getInst()->getOpcode() == V9::BA) - BAbranch = I->first; - else - branch = I->first; - continue; - } - - //Put int the map so we know what instructions in each stage are in the kernel - DEBUG(std::cerr << "Inserting instruction " << *(I->first->getInst()) << " into map at stage " << I->second << "\n"); - inKernel[I->second].insert(I->first->getInst()); - } - - //Get target information to look at machine operands - const TargetInstrInfo *mii = target.getInstrInfo(); - - //Now write the prologues - for(int i = 0; i < maxStageCount; ++i) { - BasicBlock *llvmBB = new BasicBlock("PROLOGUE", (Function*) (origBB->getBasicBlock()->getParent())); - MachineBasicBlock *machineBB = new MachineBasicBlock(llvmBB); - - DEBUG(std::cerr << "i=" << i << "\n"); - for(int j = 0; j <= i; ++j) { - for(MachineBasicBlock::const_iterator MI = origBB->begin(), ME = origBB->end(); ME != MI; ++MI) { - if(inKernel[j].count(&*MI)) { - MachineInstr *instClone = MI->clone(); - machineBB->push_back(instClone); - - DEBUG(std::cerr << "Cloning: " << *MI << "\n"); - - Instruction *tmp; - - //After cloning, we may need to save the value that this instruction defines - for(unsigned opNum=0; opNum < MI->getNumOperands(); ++opNum) { - //get machine operand - const MachineOperand &mOp = instClone->getOperand(opNum); - if(mOp.getType() == MachineOperand::MO_VirtualRegister && mOp.isDef()) { - - //Check if this is a value we should save - if(valuesToSave.count(mOp.getVRegValue())) { - //Save copy in tmpInstruction - tmp = new TmpInstruction(mOp.getVRegValue()); - - DEBUG(std::cerr << "Value: " << *(mOp.getVRegValue()) << " New Value: " << *tmp << " Stage: " << i << "\n"); - - newValues[mOp.getVRegValue()][i]= tmp; - newValLocation[tmp] = machineBB; - - DEBUG(std::cerr << "Machine Instr Operands: " << *(mOp.getVRegValue()) << ", 0, " << *tmp << "\n"); - - //Create machine instruction and put int machineBB - MachineInstr *saveValue = BuildMI(machineBB, V9::ORr, 3).addReg(mOp.getVRegValue()).addImm(0).addRegDef(tmp); - - DEBUG(std::cerr << "Created new machine instr: " << *saveValue << "\n"); - } - } - - //We may also need to update the value that we use if its from an earlier prologue - if(j != 0) { - if(mOp.getType() == MachineOperand::MO_VirtualRegister && mOp.isUse()) { - if(newValues.count(mOp.getVRegValue())) - if(newValues[mOp.getVRegValue()].count(j-1)) { - DEBUG(std::cerr << "Replaced this value: " << mOp.getVRegValue() << " With:" << (newValues[mOp.getVRegValue()][i-1]) << "\n"); - //Update the operand with the right value - instClone->getOperand(opNum).setValueReg(newValues[mOp.getVRegValue()][i-1]); - } - } - } - } - } - } - } - - - //Stick in branch at the end - machineBB->push_back(branch->getInst()->clone()); - - //Stick in BA branch at the end - machineBB->push_back(BAbranch->getInst()->clone()); - - (((MachineBasicBlock*)origBB)->getParent())->getBasicBlockList().push_back(machineBB); - prologues.push_back(machineBB); - llvm_prologues.push_back(llvmBB); - } -} - -void ModuloSchedulingPass::writeEpilogues(std::vector &epilogues, const MachineBasicBlock *origBB, std::vector &llvm_epilogues, std::map > &valuesToSave, std::map > &newValues,std::map &newValLocation, std::map > &kernelPHIs ) { - - std::map > inKernel; - - for(MSSchedule::kernel_iterator I = schedule.kernel_begin(), E = schedule.kernel_end(); I != E; ++I) { - - //Ignore the branch, we will handle this separately - if(I->first->isBranch()) - continue; - - //Put int the map so we know what instructions in each stage are in the kernel - inKernel[I->second].insert(I->first->getInst()); - } - - std::map valPHIs; - - //some debug stuff, will remove later - DEBUG(for(std::map >::iterator V = newValues.begin(), E = newValues.end(); V !=E; ++V) { - std::cerr << "Old Value: " << *(V->first) << "\n"; - for(std::map::iterator I = V->second.begin(), IE = V->second.end(); I != IE; ++I) - std::cerr << "Stage: " << I->first << " Value: " << *(I->second) << "\n"; - }); - - //some debug stuff, will remove later - DEBUG(for(std::map >::iterator V = kernelPHIs.begin(), E = kernelPHIs.end(); V !=E; ++V) { - std::cerr << "Old Value: " << *(V->first) << "\n"; - for(std::map::iterator I = V->second.begin(), IE = V->second.end(); I != IE; ++I) - std::cerr << "Stage: " << I->first << " Value: " << *(I->second) << "\n"; - }); - - //Now write the epilogues - for(int i = schedule.getMaxStage()-1; i >= 0; --i) { - BasicBlock *llvmBB = new BasicBlock("EPILOGUE", (Function*) (origBB->getBasicBlock()->getParent())); - MachineBasicBlock *machineBB = new MachineBasicBlock(llvmBB); - - DEBUG(std::cerr << " Epilogue #: " << i << "\n"); - - - - - for(MachineBasicBlock::const_iterator MI = origBB->begin(), ME = origBB->end(); ME != MI; ++MI) { - for(int j=schedule.getMaxStage(); j > i; --j) { - if(inKernel[j].count(&*MI)) { - DEBUG(std::cerr << "Cloning instruction " << *MI << "\n"); - MachineInstr *clone = MI->clone(); - - //Update operands that need to use the result from the phi - for(unsigned opNum=0; opNum < clone->getNumOperands(); ++opNum) { - //get machine operand - const MachineOperand &mOp = clone->getOperand(opNum); - - //If this is the last instructions for the max iterations ago, don't update operands - if(j == schedule.getMaxStage() && (i == 0)) - continue; - - if((mOp.getType() == MachineOperand::MO_VirtualRegister && mOp.isUse())) { - - DEBUG(std::cerr << "Writing PHI for " << *(mOp.getVRegValue()) << "\n"); - - //Quickly write appropriate phis for this operand - if(newValues.count(mOp.getVRegValue())) { - if(newValues[mOp.getVRegValue()].count(i)) { - Instruction *tmp = new TmpInstruction(newValues[mOp.getVRegValue()][i]); - MachineInstr *saveValue = BuildMI(machineBB, V9::PHI, 3).addReg(newValues[mOp.getVRegValue()][i]).addReg(kernelPHIs[mOp.getVRegValue()][i]).addRegDef(tmp); - DEBUG(std::cerr << "Resulting PHI: " << *saveValue << "\n"); - valPHIs[mOp.getVRegValue()] = tmp; - } - } - - if(valPHIs.count(mOp.getVRegValue())) { - //Update the operand in the cloned instruction - clone->getOperand(opNum).setValueReg(valPHIs[mOp.getVRegValue()]); - } - } - } - machineBB->push_back(clone); - } - } - } - - (((MachineBasicBlock*)origBB)->getParent())->getBasicBlockList().push_back(machineBB); - epilogues.push_back(machineBB); - llvm_epilogues.push_back(llvmBB); - - DEBUG(std::cerr << "EPILOGUE #" << i << "\n"); - DEBUG(machineBB->print(std::cerr)); - } -} - -void ModuloSchedulingPass::writeKernel(BasicBlock *llvmBB, MachineBasicBlock *machineBB, std::map > &valuesToSave, std::map > &newValues, std::map &newValLocation, std::map > &kernelPHIs) { - - //Keep track of operands that are read and saved from a previous iteration. The new clone - //instruction will use the result of the phi instead. - std::map finalPHIValue; - std::map kernelValue; - - //Create TmpInstructions for the final phis - for(MSSchedule::kernel_iterator I = schedule.kernel_begin(), E = schedule.kernel_end(); I != E; ++I) { - - DEBUG(std::cerr << "Stage: " << I->second << " Inst: " << *(I->first->getInst()) << "\n";); - - //Clone instruction - const MachineInstr *inst = I->first->getInst(); - MachineInstr *instClone = inst->clone(); - - //Insert into machine basic block - machineBB->push_back(instClone); - - - //Loop over Machine Operands - for(unsigned i=0; i < inst->getNumOperands(); ++i) { - //get machine operand - const MachineOperand &mOp = inst->getOperand(i); - - if(I->second != 0) { - if(mOp.getType() == MachineOperand::MO_VirtualRegister && mOp.isUse()) { - - //Check to see where this operand is defined if this instruction is from max stage - if(I->second == schedule.getMaxStage()) { - DEBUG(std::cerr << "VREG: " << *(mOp.getVRegValue()) << "\n"); - } - - //If its in the value saved, we need to create a temp instruction and use that instead - if(valuesToSave.count(mOp.getVRegValue())) { - TmpInstruction *tmp = new TmpInstruction(mOp.getVRegValue()); - - //Update the operand in the cloned instruction - instClone->getOperand(i).setValueReg(tmp); - - //save this as our final phi - finalPHIValue[mOp.getVRegValue()] = tmp; - newValLocation[tmp] = machineBB; - } - } - } - if(I->second != schedule.getMaxStage()) { - if(mOp.getType() == MachineOperand::MO_VirtualRegister && mOp.isDef()) { - if(valuesToSave.count(mOp.getVRegValue())) { - - TmpInstruction *tmp = new TmpInstruction(mOp.getVRegValue()); - - //Create new machine instr and put in MBB - MachineInstr *saveValue = BuildMI(machineBB, V9::ORr, 3).addReg(mOp.getVRegValue()).addImm(0).addRegDef(tmp); - - //Save for future cleanup - kernelValue[mOp.getVRegValue()] = tmp; - newValLocation[tmp] = machineBB; - kernelPHIs[mOp.getVRegValue()][schedule.getMaxStage()-1] = tmp; - } - } - } - } - - } - - DEBUG(std::cerr << "KERNEL before PHIs\n"); - DEBUG(machineBB->print(std::cerr)); - - - //Loop over each value we need to generate phis for - for(std::map >::iterator V = newValues.begin(), - E = newValues.end(); V != E; ++V) { - - - DEBUG(std::cerr << "Writing phi for" << *(V->first)); - DEBUG(std::cerr << "\nMap of Value* for this phi\n"); - DEBUG(for(std::map::iterator I = V->second.begin(), - IE = V->second.end(); I != IE; ++I) { - std::cerr << "Stage: " << I->first; - std::cerr << " Value: " << *(I->second) << "\n"; - }); - - //If we only have one current iteration live, its safe to set lastPhi = to kernel value - if(V->second.size() == 1) { - assert(kernelValue[V->first] != 0 && "Kernel value* must exist to create phi"); - MachineInstr *saveValue = BuildMI(*machineBB, machineBB->begin(),V9::PHI, 3).addReg(V->second.begin()->second).addReg(kernelValue[V->first]).addRegDef(finalPHIValue[V->first]); - DEBUG(std::cerr << "Resulting PHI: " << *saveValue << "\n"); - kernelPHIs[V->first][schedule.getMaxStage()-1] = kernelValue[V->first]; - } - else { - - //Keep track of last phi created. - Instruction *lastPhi = 0; - - unsigned count = 1; - //Loop over the the map backwards to generate phis - for(std::map::reverse_iterator I = V->second.rbegin(), IE = V->second.rend(); - I != IE; ++I) { - - if(count < (V->second).size()) { - if(lastPhi == 0) { - lastPhi = new TmpInstruction(I->second); - MachineInstr *saveValue = BuildMI(*machineBB, machineBB->begin(), V9::PHI, 3).addReg(kernelValue[V->first]).addReg(I->second).addRegDef(lastPhi); - DEBUG(std::cerr << "Resulting PHI: " << *saveValue << "\n"); - newValLocation[lastPhi] = machineBB; - } - else { - Instruction *tmp = new TmpInstruction(I->second); - MachineInstr *saveValue = BuildMI(*machineBB, machineBB->begin(), V9::PHI, 3).addReg(lastPhi).addReg(I->second).addRegDef(tmp); - DEBUG(std::cerr << "Resulting PHI: " << *saveValue << "\n"); - lastPhi = tmp; - kernelPHIs[V->first][I->first] = lastPhi; - newValLocation[lastPhi] = machineBB; - } - } - //Final phi value - else { - //The resulting value must be the Value* we created earlier - assert(lastPhi != 0 && "Last phi is NULL!\n"); - MachineInstr *saveValue = BuildMI(*machineBB, machineBB->begin(), V9::PHI, 3).addReg(lastPhi).addReg(I->second).addRegDef(finalPHIValue[V->first]); - DEBUG(std::cerr << "Resulting PHI: " << *saveValue << "\n"); - kernelPHIs[V->first][I->first] = finalPHIValue[V->first]; - } - - ++count; - } - - } - } - - DEBUG(std::cerr << "KERNEL after PHIs\n"); - DEBUG(machineBB->print(std::cerr)); -} - - -void ModuloSchedulingPass::removePHIs(const MachineBasicBlock *origBB, std::vector &prologues, std::vector &epilogues, MachineBasicBlock *kernelBB, std::map &newValLocation) { - - //Worklist to delete things - std::vector > worklist; - - const TargetInstrInfo *TMI = target.getInstrInfo(); - - //Start with the kernel and for each phi insert a copy for the phi def and for each arg - for(MachineBasicBlock::iterator I = kernelBB->begin(), E = kernelBB->end(); I != E; ++I) { - //Get op code and check if its a phi - if(I->getOpcode() == V9::PHI) { - Instruction *tmp = 0; - for(unsigned i = 0; i < I->getNumOperands(); ++i) { - //Get Operand - const MachineOperand &mOp = I->getOperand(i); - assert(mOp.getType() == MachineOperand::MO_VirtualRegister && "Should be a Value*\n"); - - if(!tmp) { - tmp = new TmpInstruction(mOp.getVRegValue()); - } - - //Now for all our arguments we read, OR to the new TmpInstruction that we created - if(mOp.isUse()) { - DEBUG(std::cerr << "Use: " << mOp << "\n"); - //Place a copy at the end of its BB but before the branches - assert(newValLocation.count(mOp.getVRegValue()) && "We must know where this value is located\n"); - //Reverse iterate to find the branches, we can safely assume no instructions have been - //put in the nop positions - for(MachineBasicBlock::iterator inst = --(newValLocation[mOp.getVRegValue()])->end(), endBB = (newValLocation[mOp.getVRegValue()])->begin(); inst != endBB; --inst) { - MachineOpCode opc = inst->getOpcode(); - if(TMI->isBranch(opc) || TMI->isNop(opc)) - continue; - else { - BuildMI(*(newValLocation[mOp.getVRegValue()]), ++inst, V9::ORr, 3).addReg(mOp.getVRegValue()).addImm(0).addRegDef(tmp); - break; - } - - } - - } - else { - //Remove the phi and replace it with an OR - DEBUG(std::cerr << "Def: " << mOp << "\n"); - BuildMI(*kernelBB, I, V9::ORr, 3).addReg(tmp).addImm(0).addRegDef(mOp.getVRegValue()); - worklist.push_back(std::make_pair(kernelBB, I)); - } - - } - } - - } - - //Remove phis from epilogue - for(std::vector::iterator MB = epilogues.begin(), ME = epilogues.end(); MB != ME; ++MB) { - for(MachineBasicBlock::iterator I = (*MB)->begin(), E = (*MB)->end(); I != E; ++I) { - //Get op code and check if its a phi - if(I->getOpcode() == V9::PHI) { - Instruction *tmp = 0; - for(unsigned i = 0; i < I->getNumOperands(); ++i) { - //Get Operand - const MachineOperand &mOp = I->getOperand(i); - assert(mOp.getType() == MachineOperand::MO_VirtualRegister && "Should be a Value*\n"); - - if(!tmp) { - tmp = new TmpInstruction(mOp.getVRegValue()); - } - - //Now for all our arguments we read, OR to the new TmpInstruction that we created - if(mOp.isUse()) { - DEBUG(std::cerr << "Use: " << mOp << "\n"); - //Place a copy at the end of its BB but before the branches - assert(newValLocation.count(mOp.getVRegValue()) && "We must know where this value is located\n"); - //Reverse iterate to find the branches, we can safely assume no instructions have been - //put in the nop positions - for(MachineBasicBlock::iterator inst = --(newValLocation[mOp.getVRegValue()])->end(), endBB = (newValLocation[mOp.getVRegValue()])->begin(); inst != endBB; --inst) { - MachineOpCode opc = inst->getOpcode(); - if(TMI->isBranch(opc) || TMI->isNop(opc)) - continue; - else { - BuildMI(*(newValLocation[mOp.getVRegValue()]), ++inst, V9::ORr, 3).addReg(mOp.getVRegValue()).addImm(0).addRegDef(tmp); - break; - } - - } - - } - else { - //Remove the phi and replace it with an OR - DEBUG(std::cerr << "Def: " << mOp << "\n"); - BuildMI(**MB, I, V9::ORr, 3).addReg(tmp).addImm(0).addRegDef(mOp.getVRegValue()); - worklist.push_back(std::make_pair(*MB,I)); - } - - } - } - } - } - - //Delete the phis - for(std::vector >::iterator I = worklist.begin(), E = worklist.end(); I != E; ++I) { - DEBUG(std::cerr << "Deleting PHI " << I->second << "\n"); - I->first->erase(I->second); - - } - -} - - -void ModuloSchedulingPass::reconstructLoop(MachineBasicBlock *BB) { - - DEBUG(std::cerr << "Reconstructing Loop\n"); - - //First find the value *'s that we need to "save" - std::map > valuesToSave; - - //Keep track of instructions we have already seen and their stage because - //we don't want to "save" values if they are used in the kernel immediately - std::map lastInstrs; - - //Loop over kernel and only look at instructions from a stage > 0 - //Look at its operands and save values *'s that are read - for(MSSchedule::kernel_iterator I = schedule.kernel_begin(), E = schedule.kernel_end(); I != E; ++I) { - - if(I->second !=0) { - //For this instruction, get the Value*'s that it reads and put them into the set. - //Assert if there is an operand of another type that we need to save - const MachineInstr *inst = I->first->getInst(); - lastInstrs[inst] = I->second; - - for(unsigned i=0; i < inst->getNumOperands(); ++i) { - //get machine operand - const MachineOperand &mOp = inst->getOperand(i); - - if(mOp.getType() == MachineOperand::MO_VirtualRegister && mOp.isUse()) { - //find the value in the map - if (const Value* srcI = mOp.getVRegValue()) { - - //Before we declare this Value* one that we should save - //make sure its def is not of the same stage as this instruction - //because it will be consumed before its used - Instruction *defInst = (Instruction*) srcI; - - //Should we save this value? - bool save = true; - - //Get Machine code for this instruction, and loop backwards over the array - //to find the def - MachineCodeForInstruction & tempMvec = MachineCodeForInstruction::get(defInst); - for (int j = tempMvec.size()-1; j >= 0; j--) { - MachineInstr *temp = tempMvec[j]; - - //Loop over instructions - for(unsigned opNum = 0; opNum < temp->getNumOperands(); ++opNum) { - MachineOperand &mDefOp = temp->getOperand(opNum); - - if (mDefOp.getType() == MachineOperand::MO_VirtualRegister && mDefOp.isDef()) { - const Value* defVReg = mDefOp.getVRegValue(); - if(defVReg == srcI) { - //Check if instruction has been seen already and is of same stage - if(lastInstrs.count(temp)) { - if(lastInstrs[temp] == I->second) - save = false; - } - } - } - } - } - if(save) - valuesToSave[srcI] = std::make_pair(I->first, i); - } - } - - if(mOp.getType() != MachineOperand::MO_VirtualRegister && mOp.isUse()) { - assert("Our assumption is wrong. We have another type of register that needs to be saved\n"); - } - } - } - } - - //The new loop will consist of one or more prologues, the kernel, and one or more epilogues. - - //Map to keep track of old to new values - std::map > newValues; - - //Map to keep track of old to new values in kernel - std::map > kernelPHIs; - - //Another map to keep track of what machine basic blocks these new value*s are in since - //they have no llvm instruction equivalent - std::map newValLocation; - - std::vector prologues; - std::vector llvm_prologues; - - - //Write prologue - writePrologues(prologues, BB, llvm_prologues, valuesToSave, newValues, newValLocation); - - //Print out epilogues and prologue - DEBUG(for(std::vector::iterator I = prologues.begin(), E = prologues.end(); - I != E; ++I) { - std::cerr << "PROLOGUE\n"; - (*I)->print(std::cerr); - }); - - BasicBlock *llvmKernelBB = new BasicBlock("Kernel", (Function*) (BB->getBasicBlock()->getParent())); - MachineBasicBlock *machineKernelBB = new MachineBasicBlock(llvmKernelBB); - (((MachineBasicBlock*)BB)->getParent())->getBasicBlockList().push_back(machineKernelBB); - writeKernel(llvmKernelBB, machineKernelBB, valuesToSave, newValues, newValLocation, kernelPHIs); - - - std::vector epilogues; - std::vector llvm_epilogues; - - //Write epilogues - writeEpilogues(epilogues, BB, llvm_epilogues, valuesToSave, newValues, newValLocation, kernelPHIs); - - - const TargetInstrInfo *TMI = target.getInstrInfo(); - - //Fix up machineBB and llvmBB branches - for(unsigned I = 0; I < prologues.size(); ++I) { - - MachineInstr *branch = 0; - - //Find terminator since getFirstTerminator does not work! - for(MachineBasicBlock::reverse_iterator mInst = prologues[I]->rbegin(), mInstEnd = prologues[I]->rend(); mInst != mInstEnd; ++mInst) { - MachineOpCode OC = mInst->getOpcode(); - if(TMI->isBranch(OC)) { - branch = &*mInst; - DEBUG(std::cerr << *mInst << "\n"); - break; - } - } - - - - //Update branch - for(unsigned opNum = 0; opNum < branch->getNumOperands(); ++opNum) { - MachineOperand &mOp = branch->getOperand(opNum); - if (mOp.getType() == MachineOperand::MO_PCRelativeDisp) { - mOp.setValueReg(llvm_epilogues[(llvm_epilogues.size()-1-I)]); - } - } - - //Update llvm basic block with our new branch instr - DEBUG(std::cerr << BB->getBasicBlock()->getTerminator() << "\n"); - const BranchInst *branchVal = dyn_cast(BB->getBasicBlock()->getTerminator()); - TmpInstruction *tmp = new TmpInstruction(branchVal->getCondition()); - if(I == prologues.size()-1) { - TerminatorInst *newBranch = new BranchInst(llvmKernelBB, - llvm_epilogues[(llvm_epilogues.size()-1-I)], - tmp, - llvm_prologues[I]); - } - else - TerminatorInst *newBranch = new BranchInst(llvm_prologues[I+1], - llvm_epilogues[(llvm_epilogues.size()-1-I)], - tmp, - llvm_prologues[I]); - - assert(branch != 0 && "There must be a terminator for this machine basic block!\n"); - - //Push nop onto end of machine basic block - BuildMI(prologues[I], V9::NOP, 0); - - //Add a unconditional branch to the next prologue - if(I != prologues.size()-1) - BuildMI(prologues[I], V9::BA, 1).addPCDisp(llvm_prologues[I+1]); - else - BuildMI(prologues[I], V9::BA, 1).addPCDisp(llvmKernelBB); - - //Add one more nop! - BuildMI(prologues[I], V9::NOP, 0); - } - - //Fix up kernel machine branches - MachineInstr *branch = 0; - for(MachineBasicBlock::reverse_iterator mInst = machineKernelBB->rbegin(), mInstEnd = machineKernelBB->rend(); mInst != mInstEnd; ++mInst) { - MachineOpCode OC = mInst->getOpcode(); - if(TMI->isBranch(OC)) { - branch = &*mInst; - DEBUG(std::cerr << *mInst << "\n"); - break; - } - } - - assert(branch != 0 && "There must be a terminator for the kernel machine basic block!\n"); - - //Update kernel self loop branch - for(unsigned opNum = 0; opNum < branch->getNumOperands(); ++opNum) { - MachineOperand &mOp = branch->getOperand(opNum); - - if (mOp.getType() == MachineOperand::MO_PCRelativeDisp) { - mOp.setValueReg(llvmKernelBB); - } - } - - //Update kernelLLVM branches - const BranchInst *branchVal = dyn_cast(BB->getBasicBlock()->getTerminator()); - TerminatorInst *newBranch = new BranchInst(llvmKernelBB, - llvm_epilogues[0], - new TmpInstruction(branchVal->getCondition()), - llvmKernelBB); - - //Add kernel noop - BuildMI(machineKernelBB, V9::NOP, 0); - - //Add unconditional branch to first epilogue - BuildMI(machineKernelBB, V9::BA, 1).addPCDisp(llvm_epilogues[0]); - - - //Add kernel noop - BuildMI(machineKernelBB, V9::NOP, 0); - - //Lastly add unconditional branches for the epilogues - for(unsigned I = 0; I < epilogues.size(); ++I) { - - //Now since I don't trust fall throughs, add a unconditional branch to the next prologue - if(I != epilogues.size()-1) { - BuildMI(epilogues[I], V9::BA, 1).addPCDisp(llvm_epilogues[I+1]); - //Add unconditional branch to end of epilogue - TerminatorInst *newBranch = new BranchInst(llvm_epilogues[I+1], - llvm_epilogues[I]); - - } - else { - MachineBasicBlock *origBlock = (MachineBasicBlock*) BB; - for(MachineBasicBlock::reverse_iterator inst = origBlock->rbegin(), instEnd = origBlock->rend(); inst != instEnd; ++inst) { - MachineOpCode OC = inst->getOpcode(); - if(TMI->isBranch(OC)) { - branch = &*inst; - DEBUG(std::cerr << "Exit branch from loop" << *inst << "\n"); - break; - - } - - for(unsigned opNum = 0; opNum < branch->getNumOperands(); ++opNum) { - MachineOperand &mOp = branch->getOperand(opNum); - - if (mOp.getType() == MachineOperand::MO_PCRelativeDisp) { - BuildMI(epilogues[I], V9::BA, 1).addPCDisp(mOp.getVRegValue()); - break; - } - } - - } - - //Update last epilogue exit branch - BranchInst *branchVal = (BranchInst*) dyn_cast(BB->getBasicBlock()->getTerminator()); - //Find where we are supposed to branch to - BasicBlock *nextBlock = 0; - for(unsigned j=0; j getNumSuccessors(); ++j) { - if(branchVal->getSuccessor(j) != BB->getBasicBlock()) - nextBlock = branchVal->getSuccessor(j); - } - TerminatorInst *newBranch = new BranchInst(nextBlock, llvm_epilogues[I]); - } - //Add one more nop! - BuildMI(epilogues[I], V9::NOP, 0); - - } - - //FIX UP Machine BB entry!! - //We are looking at the predecesor of our loop basic block and we want to change its ba instruction - - - //Find all llvm basic blocks that branch to the loop entry and change to our first prologue. - const BasicBlock *llvmBB = BB->getBasicBlock(); - - for(pred_const_iterator P = pred_begin(llvmBB), PE = pred_end(llvmBB); P != PE; ++PE) { - if(*P == llvmBB) - continue; - else { - DEBUG(std::cerr << "Found our entry BB\n"); - //Get the Terminator instruction for this basic block and print it out - DEBUG(std::cerr << *((*P)->getTerminator()) << "\n"); - //Update the terminator - TerminatorInst *term = ((BasicBlock*)*P)->getTerminator(); - for(unsigned i=0; i < term->getNumSuccessors(); ++i) { - if(term->getSuccessor(i) == llvmBB) { - DEBUG(std::cerr << "Replacing successor bb\n"); - if(llvm_prologues.size() > 0) { - term->setSuccessor(i, llvm_prologues[0]); - //Also update its corresponding machine instruction - MachineCodeForInstruction & tempMvec = - MachineCodeForInstruction::get(term); - for (unsigned j = 0; j < tempMvec.size(); j++) { - MachineInstr *temp = tempMvec[j]; - MachineOpCode opc = temp->getOpcode(); - if(TMI->isBranch(opc)) { - DEBUG(std::cerr << *temp << "\n"); - //Update branch - for(unsigned opNum = 0; opNum < temp->getNumOperands(); ++opNum) { - MachineOperand &mOp = temp->getOperand(opNum); - if (mOp.getType() == MachineOperand::MO_PCRelativeDisp) { - mOp.setValueReg(llvm_prologues[0]); - } - } - } - } - } - else { - term->setSuccessor(i, llvmKernelBB); - //Also update its corresponding machine instruction - MachineCodeForInstruction & tempMvec = - MachineCodeForInstruction::get(term); - for (unsigned j = 0; j < tempMvec.size(); j++) { - MachineInstr *temp = tempMvec[j]; - MachineOpCode opc = temp->getOpcode(); - if(TMI->isBranch(opc)) { - DEBUG(std::cerr << *temp << "\n"); - //Update branch - for(unsigned opNum = 0; opNum < temp->getNumOperands(); ++opNum) { - MachineOperand &mOp = temp->getOperand(opNum); - if (mOp.getType() == MachineOperand::MO_PCRelativeDisp) { - mOp.setValueReg(llvmKernelBB); - } - } - } - } - } - } - } - break; - } - } - - removePHIs(BB, prologues, epilogues, machineKernelBB, newValLocation); - - - - //Print out epilogues and prologue - DEBUG(for(std::vector::iterator I = prologues.begin(), E = prologues.end(); - I != E; ++I) { - std::cerr << "PROLOGUE\n"; - (*I)->print(std::cerr); - }); - - DEBUG(std::cerr << "KERNEL\n"); - DEBUG(machineKernelBB->print(std::cerr)); - - DEBUG(for(std::vector::iterator I = epilogues.begin(), E = epilogues.end(); - I != E; ++I) { - std::cerr << "EPILOGUE\n"; - (*I)->print(std::cerr); - }); - - - DEBUG(std::cerr << "New Machine Function" << "\n"); - DEBUG(std::cerr << BB->getParent() << "\n"); - - //BB->getParent()->getBasicBlockList().erase(BB); - -} - diff --git a/lib/CodeGen/ModuloScheduling/ModuloScheduling.h b/lib/CodeGen/ModuloScheduling/ModuloScheduling.h deleted file mode 100644 index d1376b702d7..00000000000 --- a/lib/CodeGen/ModuloScheduling/ModuloScheduling.h +++ /dev/null @@ -1,112 +0,0 @@ -//===-- ModuloScheduling.h - Swing Modulo Scheduling------------*- 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. -// -//===----------------------------------------------------------------------===// -// -// -//===----------------------------------------------------------------------===// - -#ifndef LLVM_MODULOSCHEDULING_H -#define LLVM_MODULOSCHEDULING_H - -#include "MSchedGraph.h" -#include "MSSchedule.h" -#include "llvm/Function.h" -#include "llvm/Pass.h" -#include - -namespace llvm { - - - //Struct to contain ModuloScheduling Specific Information for each node - struct MSNodeAttributes { - int ASAP; //Earliest time at which the opreation can be scheduled - int ALAP; //Latest time at which the operation can be scheduled. - int MOB; - int depth; - int height; - MSNodeAttributes(int asap=-1, int alap=-1, int mob=-1, - int d=-1, int h=-1) : ASAP(asap), ALAP(alap), - MOB(mob), depth(d), - height(h) {} - }; - - - class ModuloSchedulingPass : public FunctionPass { - const TargetMachine ⌖ - - //Map that holds node to node attribute information - std::map nodeToAttributesMap; - - //Map to hold all reccurrences - std::set > > recurrenceList; - - //Set of edges to ignore, stored as src node and index into vector of successors - std::set > edgesToIgnore; - - //Vector containing the partial order - std::vector > partialOrder; - - //Vector containing the final node order - std::vector FinalNodeOrder; - - //Schedule table, key is the cycle number and the vector is resource, node pairs - MSSchedule schedule; - - //Current initiation interval - int II; - - //Internal functions - bool MachineBBisValid(const MachineBasicBlock *BI); - int calculateResMII(const MachineBasicBlock *BI); - int calculateRecMII(MSchedGraph *graph, int MII); - void calculateNodeAttributes(MSchedGraph *graph, int MII); - - bool ignoreEdge(MSchedGraphNode *srcNode, MSchedGraphNode *destNode); - - int calculateASAP(MSchedGraphNode *node, int MII,MSchedGraphNode *destNode); - int calculateALAP(MSchedGraphNode *node, int MII, int maxASAP, MSchedGraphNode *srcNode); - - int calculateHeight(MSchedGraphNode *node,MSchedGraphNode *srcNode); - int calculateDepth(MSchedGraphNode *node, MSchedGraphNode *destNode); - - int findMaxASAP(); - void orderNodes(); - void findAllReccurrences(MSchedGraphNode *node, - std::vector &visitedNodes, int II); - void addReccurrence(std::vector &recurrence, int II, MSchedGraphNode*, MSchedGraphNode*); - - void computePartialOrder(); - void computeSchedule(); - bool scheduleNode(MSchedGraphNode *node, - int start, int end); - - void predIntersect(std::vector &CurrentSet, std::vector &IntersectResult); - void succIntersect(std::vector &CurrentSet, std::vector &IntersectResult); - - void reconstructLoop(MachineBasicBlock*); - - //void saveValue(const MachineInstr*, const std::set&, std::vector*); - - void writePrologues(std::vector &prologues, MachineBasicBlock *origBB, std::vector &llvm_prologues, std::map > &valuesToSave, std::map > &newValues, std::map &newValLocation); - - void writeEpilogues(std::vector &epilogues, const MachineBasicBlock *origBB, std::vector &llvm_epilogues, std::map > &valuesToSave,std::map > &newValues, std::map &newValLocation, std::map > &kernelPHIs); - - - void writeKernel(BasicBlock *llvmBB, MachineBasicBlock *machineBB, std::map > &valuesToSave, std::map > &newValues, std::map &newValLocation, std::map > &kernelPHIs); - - void removePHIs(const MachineBasicBlock *origBB, std::vector &prologues, std::vector &epilogues, MachineBasicBlock *kernelBB, std::map &newValLocation); - - public: - ModuloSchedulingPass(TargetMachine &targ) : target(targ) {} - virtual bool runOnFunction(Function &F); - }; - -} - - -#endif