//===-- 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. // //===----------------------------------------------------------------------===// // // // //===----------------------------------------------------------------------===// #define DEBUG_TYPE "ModuloSched" #include "ModuloScheduling.h" #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/Passes.h" #include "llvm/Support/CFG.h" #include "llvm/Target/TargetSchedInfo.h" #include "Support/Debug.h" #include "Support/GraphWriter.h" #include #include #include #include #include using namespace llvm; /// Create ModuloSchedulingPass /// FunctionPass *llvm::createModuloSchedulingPass(TargetMachine & targ) { DEBUG(std::cerr << "Created ModuloSchedulingPass\n"); return new ModuloSchedulingPass(targ); } 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"; }; 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; } if(I.getEdge().getIteDiff() > 0) edgelabel += I.getEdge().getIteDiff(); return edgelabel; } }; } /// ModuloScheduling::runOnFunction - main transformation entry point bool ModuloSchedulingPass::runOnFunction(Function &F) { bool Changed = false; DEBUG(std::cerr << "Creating ModuloSchedGraph for each BasicBlock in" + F.getName() + "\n"); //Get MachineFunction MachineFunction &MF = MachineFunction::get(&F); //Iterate over BasicBlocks and do ModuloScheduling if they are valid for (MachineFunction::const_iterator BI = MF.begin(); BI != MF.end(); ++BI) { if(MachineBBisValid(BI)) { MSchedGraph *MSG = new MSchedGraph(BI, target); //Write Graph out to file DEBUG(WriteGraphToFile(std::cerr, "dependgraph", MSG)); //Print out BB for debugging DEBUG(BI->print(std::cerr)); //Calculate Resource II int ResMII = calculateResMII(BI); calculateNodeAttributes(MSG, ResMII); } } return Changed; } bool ModuloSchedulingPass::MachineBBisValid(const MachineBasicBlock *BI) { //Valid basic blocks must be loops and can not have if/else statements or calls. 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) { DEBUG(std::cerr << "Basic Block is not a loop\n"); return false; } else DEBUG(std::cerr << "Basic Block is a loop\n"); //Get Target machine instruction info /*const TargetInstrInfo& TMI = targ.getInstrInfo(); //Check each instruction and look for calls or if/else statements unsigned count = 0; for(MachineBasicBlock::const_iterator I = BI->begin(), E = BI->end(); I != E; ++I) { //Get opcode to check instruction type MachineOpCode OC = I->getOpcode(); if(TMI.isControlFlow(OC) && (count+1 < BI->size())) return false; count++; }*/ 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. 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 = msi.getCPUResourceNum(RB->first); //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); DEBUG(std::cerr << "Resource ID: " << RB->first << " (usage=" << usageCount << ", resourceNum=X" << ", issueSlots=" << issueSlots << ", finalUsage=" << finalUsageCount << ")\n"); //Only keep track of the max ResMII = std::max( (int) finalUsageCount, ResMII); } DEBUG(std::cerr << "Final Resource MII: " << ResMII << "\n"); return ResMII; } 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; std::vector vNodes; //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, (I->second), MII, visitedNodes); findAllReccurrences(I->first, vNodes); vNodes.clear(); visitedNodes.clear(); } //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, (I->second), MII, MII, visitedNodes); 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 = (I->second).ALAP - (I->second).ASAP; DEBUG(std::cerr << "MOB: " << (I->second).MOB << " (" << *(I->first) << ")\n"); calculateDepth(I->first, (I->second), visitedNodes); visitedNodes.clear(); calculateHeight(I->first, (I->second), visitedNodes); visitedNodes.clear(); }*/ } void ModuloSchedulingPass::calculateASAP(MSchedGraphNode *node, MSNodeAttributes &attributes, int MII, std::set &visitedNodes) { DEBUG(std::cerr << "Calculating ASAP for " << *node << "\n"); if(attributes.ASAP != -1 || (visitedNodes.find(node) != visitedNodes.end())) { visitedNodes.erase(node); return; } if(node->hasPredecessors()) { int maxPredValue = 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) { //Get that nodes ASAP MSNodeAttributes predAttributes = nodeToAttributesMap.find(*P)->second; if(predAttributes.ASAP == -1) { //Put into set before you recurse visitedNodes.insert(node); calculateASAP(*P, predAttributes, MII, visitedNodes); predAttributes = nodeToAttributesMap.find(*P)->second; } int iteDiff = node->getInEdge(*P).getIteDiff(); int currentPredValue = predAttributes.ASAP + node->getLatency() - iteDiff * MII; DEBUG(std::cerr << "Current ASAP pred: " << currentPredValue << "\n"); maxPredValue = std::max(maxPredValue, currentPredValue); } visitedNodes.erase(node); attributes.ASAP = maxPredValue; } else { visitedNodes.erase(node); attributes.ASAP = 0; } DEBUG(std::cerr << "ASAP: " << attributes.ASAP << " (" << *node << ")\n"); } void ModuloSchedulingPass::calculateALAP(MSchedGraphNode *node, MSNodeAttributes &attributes, int MII, int maxASAP, std::set &visitedNodes) { DEBUG(std::cerr << "Calculating AlAP for " << *node << "\n"); if(attributes.ALAP != -1|| (visitedNodes.find(node) != visitedNodes.end())) { visitedNodes.erase(node); return; } if(node->hasSuccessors()) { int minSuccValue = 0; //Iterate over all of the predecessors and fine max for(MSchedGraphNode::succ_iterator P = node->succ_begin(), E = node->succ_end(); P != E; ++P) { MSNodeAttributes succAttributes = nodeToAttributesMap.find(*P)->second; if(succAttributes.ASAP == -1) { //Put into set before recursing visitedNodes.insert(node); calculateALAP(*P, succAttributes, MII, maxASAP, visitedNodes); succAttributes = nodeToAttributesMap.find(*P)->second; assert(succAttributes.ASAP == -1 && "Successors ALAP should have been caclulated"); } int iteDiff = P.getEdge().getIteDiff(); int currentSuccValue = succAttributes.ALAP + node->getLatency() + iteDiff * MII; minSuccValue = std::min(minSuccValue, currentSuccValue); } visitedNodes.erase(node); attributes.ALAP = minSuccValue; } else { visitedNodes.erase(node); attributes.ALAP = maxASAP; } DEBUG(std::cerr << "ALAP: " << attributes.ALAP << " (" << *node << ")\n"); } 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; } void ModuloSchedulingPass::calculateHeight(MSchedGraphNode *node, MSNodeAttributes &attributes, std::set &visitedNodes) { if(attributes.depth != -1 || (visitedNodes.find(node) != visitedNodes.end())) { //Remove from map before returning visitedNodes.erase(node); return; } if(node->hasSuccessors()) { int maxHeight = 0; //Iterate over all of the predecessors and fine max for(MSchedGraphNode::succ_iterator P = node->succ_begin(), E = node->succ_end(); P != E; ++P) { MSNodeAttributes succAttributes = nodeToAttributesMap.find(*P)->second; if(succAttributes.height == -1) { //Put into map before recursing visitedNodes.insert(node); calculateHeight(*P, succAttributes, visitedNodes); succAttributes = nodeToAttributesMap.find(*P)->second; assert(succAttributes.height == -1 && "Successors Height should have been caclulated"); } int currentHeight = succAttributes.height + node->getLatency(); maxHeight = std::max(maxHeight, currentHeight); } visitedNodes.erase(node); attributes.height = maxHeight; } else { visitedNodes.erase(node); attributes.height = 0; } DEBUG(std::cerr << "Height: " << attributes.height << " (" << *node << ")\n"); } void ModuloSchedulingPass::calculateDepth(MSchedGraphNode *node, MSNodeAttributes &attributes, std::set &visitedNodes) { if(attributes.depth != -1 || (visitedNodes.find(node) != visitedNodes.end())) { //Remove from map before returning visitedNodes.erase(node); return; } if(node->hasPredecessors()) { 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) { //Get that nodes depth MSNodeAttributes predAttributes = nodeToAttributesMap.find(*P)->second; if(predAttributes.depth == -1) { //Put into set before recursing visitedNodes.insert(node); calculateDepth(*P, predAttributes, visitedNodes); predAttributes = nodeToAttributesMap.find(*P)->second; assert(predAttributes.depth == -1 && "Predecessors ASAP should have been caclulated"); } int currentDepth = predAttributes.depth + node->getLatency(); maxDepth = std::max(maxDepth, currentDepth); } //Remove from map before returning visitedNodes.erase(node); attributes.height = maxDepth; } else { //Remove from map before returning visitedNodes.erase(node); attributes.depth = 0; } DEBUG(std::cerr << "Depth: " << attributes.depth << " (" << *node << "*)\n"); } void ModuloSchedulingPass::findAllReccurrences(MSchedGraphNode *node, std::vector &visitedNodes) { if(find(visitedNodes.begin(), visitedNodes.end(), node) != visitedNodes.end()) { //DUMP out recurrence DEBUG(std::cerr << "Reccurrence:\n"); bool first = true; for(std::vector::iterator I = visitedNodes.begin(), E = visitedNodes.end(); I !=E; ++I) { if(*I == node) first = false; if(first) continue; DEBUG(std::cerr << **I << "\n"); } DEBUG(std::cerr << "End Reccurrence:\n"); return; } for(MSchedGraphNode::succ_iterator I = node->succ_begin(), E = node->succ_end(); I != E; ++I) { visitedNodes.push_back(node); findAllReccurrences(*I, visitedNodes); visitedNodes.pop_back(); } } void ModuloSchedulingPass::orderNodes() { int BOTTOM_UP = 0; int TOP_DOWN = 1; //FIXME: Group nodes into sets and order all the sets based on RecMII typedef std::vector NodeVector; typedef std::pair NodeSet; std::vector NodeSetsToOrder; //Order the resulting sets NodeVector FinalNodeOrder; //Loop over all the sets and place them in the final node order for(unsigned i=0; i < NodeSetsToOrder.size(); ++i) { //Set default order int order = BOTTOM_UP; //Get Nodes in Current set NodeVector CurrentSet = NodeSetsToOrder[i].second; //Loop through the predecessors for each node in the final order //and only keeps nodes both in the pred_set and currentset NodeVector IntersectCurrent; //Sort CurrentSet so we can use lowerbound 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) { if(lower_bound(CurrentSet.begin(), CurrentSet.end(), *P) != CurrentSet.end()) IntersectCurrent.push_back(*P); } } //If the intersection of predecessor and current set is not empty //sort nodes bottom up if(IntersectCurrent.size() != 0) order = BOTTOM_UP; //If empty, use successors else { 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) { if(lower_bound(CurrentSet.begin(), CurrentSet.end(), *P) != CurrentSet.end()) IntersectCurrent.push_back(*P); } } //sort top-down if(IntersectCurrent.size() != 0) order = TOP_DOWN; else { //Find node with max ASAP in current Set MSchedGraphNode *node; int maxASAP = 0; 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!"); if(maxASAP < nodeAttr.ASAP) { maxASAP = nodeAttr.ASAP; node = CurrentSet[j]; } } order = BOTTOM_UP; } } //Repeat until all nodes are put into the final order from current set /*while(IntersectCurrent.size() > 0) { if(order == TOP_DOWN) { while(IntersectCurrent.size() > 0) { //FIXME //Get node attributes MSNodeAttributes nodeAttr= nodeToAttributesMap.find(IntersectCurrent[0])->second; assert(nodeAttr != nodeToAttributesMap.end() && "Node not in attributes map!"); //Get node with highest height, if a tie, use one with lowest //MOB int MOB = nodeAttr.MBO; int height = nodeAttr.height; ModuloSchedGraphNode *V = IntersectCurrent[0]; for(unsigned j=0; j < IntersectCurrent.size(); ++j) { int temp = IntersectCurrent[j]->getHeight(); if(height < temp) { V = IntersectCurrent[j]; height = temp; MOB = V->getMobility(); } else if(height == temp) { if(MOB > IntersectCurrent[j]->getMobility()) { V = IntersectCurrent[j]; height = temp; MOB = V->getMobility(); } } } //Append V to the NodeOrder NodeOrder.push_back(V); //Remove V from IntersectOrder IntersectCurrent.erase(find(IntersectCurrent.begin(), IntersectCurrent.end(), V)); //Intersect V's successors with CurrentSet for(mod_succ_iterator P = succ_begin(V), E = succ_end(V); P != E; ++P) { if(lower_bound(CurrentSet.begin(), CurrentSet.end(), *P) != CurrentSet.end()) { //If not already in Intersect, add if(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(); for(unsigned j=0; j < NodeOrder.size(); ++j) { for(mod_pred_iterator P = pred_begin(NodeOrder[j]), E = pred_end(NodeOrder[j]); P != E; ++P) { if(lower_bound(CurrentSet.begin(), CurrentSet.end(), *P) != CurrentSet.end()) IntersectCurrent.push_back(*P); } } } //End If TOP_DOWN //Begin if BOTTOM_UP else { while(IntersectCurrent.size() > 0) { //Get node with highest depth, if a tie, use one with lowest //MOB int MOB = IntersectCurrent[0]->getMobility(); int depth = IntersectCurrent[0]->getDepth(); ModuloSchedGraphNode *V = IntersectCurrent[0]; for(unsigned j=0; j < IntersectCurrent.size(); ++j) { int temp = IntersectCurrent[j]->getDepth(); if(depth < temp) { V = IntersectCurrent[j]; depth = temp; MOB = V->getMobility(); } else if(depth == temp) { if(MOB > IntersectCurrent[j]->getMobility()) { V = IntersectCurrent[j]; depth = temp; MOB = V->getMobility(); } } } //Append V to the NodeOrder NodeOrder.push_back(V); //Remove V from IntersectOrder IntersectCurrent.erase(find(IntersectCurrent.begin(), IntersectCurrent.end(),V)); //Intersect V's pred with CurrentSet for(mod_pred_iterator P = pred_begin(V), E = pred_end(V); P != E; ++P) { if(lower_bound(CurrentSet.begin(), CurrentSet.end(), *P) != CurrentSet.end()) { //If not already in Intersect, add if(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(); for(unsigned j=0; j < NodeOrder.size(); ++j) { for(mod_succ_iterator P = succ_begin(NodeOrder[j]), E = succ_end(NodeOrder[j]); P != E; ++P) { if(lower_bound(CurrentSet.begin(), CurrentSet.end(), *P) != CurrentSet.end()) IntersectCurrent.push_back(*P); } } } //End if BOTTOM_DOWN }*/ //End Wrapping while loop }//End for over all sets of nodes //Return final Order //return FinalNodeOrder; }