//===-- AMDGPUStructurizeCFG.cpp - ------------------===// // // The LLVM Compiler Infrastructure // // This file is distributed under the University of Illinois Open Source // License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // /// \file /// The pass implemented in this file transforms the programs control flow /// graph into a form that's suitable for code generation on hardware that /// implements control flow by execution masking. This currently includes all /// AMD GPUs but may as well be useful for other types of hardware. // //===----------------------------------------------------------------------===// #include "AMDGPU.h" #include "llvm/ADT/SCCIterator.h" #include "llvm/Analysis/RegionInfo.h" #include "llvm/Analysis/RegionIterator.h" #include "llvm/Analysis/RegionPass.h" #include "llvm/Module.h" #include "llvm/Transforms/Utils/SSAUpdater.h" using namespace llvm; namespace { // Definition of the complex types used in this pass. typedef std::pair BBValuePair; typedef ArrayRef BBVecRef; typedef SmallVector RNVector; typedef SmallVector BBVector; typedef SmallVector BBValueVector; typedef DenseMap PhiMap; typedef DenseMap BBPhiMap; typedef DenseMap BBPredicates; typedef DenseMap PredMap; typedef DenseMap VisitedMap; // The name for newly created blocks. static const char *FlowBlockName = "Flow"; /// @brief Transforms the control flow graph on one single entry/exit region /// at a time. /// /// After the transform all "If"/"Then"/"Else" style control flow looks like /// this: /// /// \verbatim /// 1 /// || /// | | /// 2 | /// | / /// |/ /// 3 /// || Where: /// | | 1 = "If" block, calculates the condition /// 4 | 2 = "Then" subregion, runs if the condition is true /// | / 3 = "Flow" blocks, newly inserted flow blocks, rejoins the flow /// |/ 4 = "Else" optional subregion, runs if the condition is false /// 5 5 = "End" block, also rejoins the control flow /// \endverbatim /// /// Control flow is expressed as a branch where the true exit goes into the /// "Then"/"Else" region, while the false exit skips the region /// The condition for the optional "Else" region is expressed as a PHI node. /// The incomming values of the PHI node are true for the "If" edge and false /// for the "Then" edge. /// /// Additionally to that even complicated loops look like this: /// /// \verbatim /// 1 /// || /// | | /// 2 ^ Where: /// | / 1 = "Entry" block /// |/ 2 = "Loop" optional subregion, with all exits at "Flow" block /// 3 3 = "Flow" block, with back edge to entry block /// | /// \endverbatim /// /// The back edge of the "Flow" block is always on the false side of the branch /// while the true side continues the general flow. So the loop condition /// consist of a network of PHI nodes where the true incoming values expresses /// breaks and the false values expresses continue states. class AMDGPUStructurizeCFG : public RegionPass { static char ID; Type *Boolean; ConstantInt *BoolTrue; ConstantInt *BoolFalse; UndefValue *BoolUndef; Function *Func; Region *ParentRegion; DominatorTree *DT; RNVector Order; VisitedMap Visited; PredMap Predicates; BBPhiMap DeletedPhis; BBVector FlowsInserted; BasicBlock *LoopStart; BasicBlock *LoopEnd; BBPredicates LoopPred; void orderNodes(); void buildPredicate(BranchInst *Term, unsigned Idx, BBPredicates &Pred, bool Invert); void analyzeBlock(BasicBlock *BB); void analyzeLoop(BasicBlock *BB, unsigned &LoopIdx); void collectInfos(); bool dominatesPredicates(BasicBlock *A, BasicBlock *B); void killTerminator(BasicBlock *BB); RegionNode *skipChained(RegionNode *Node); void delPhiValues(BasicBlock *From, BasicBlock *To); void addPhiValues(BasicBlock *From, BasicBlock *To); BasicBlock *getNextFlow(BasicBlock *Prev); bool isPredictableTrue(BasicBlock *Prev, BasicBlock *Node); BasicBlock *wireFlowBlock(BasicBlock *Prev, RegionNode *Node); void createFlow(); void insertConditions(); void rebuildSSA(); public: AMDGPUStructurizeCFG(): RegionPass(ID) { initializeRegionInfoPass(*PassRegistry::getPassRegistry()); } virtual bool doInitialization(Region *R, RGPassManager &RGM); virtual bool runOnRegion(Region *R, RGPassManager &RGM); virtual const char *getPassName() const { return "AMDGPU simplify control flow"; } void getAnalysisUsage(AnalysisUsage &AU) const { AU.addRequired(); AU.addPreserved(); RegionPass::getAnalysisUsage(AU); } }; } // end anonymous namespace char AMDGPUStructurizeCFG::ID = 0; /// \brief Initialize the types and constants used in the pass bool AMDGPUStructurizeCFG::doInitialization(Region *R, RGPassManager &RGM) { LLVMContext &Context = R->getEntry()->getContext(); Boolean = Type::getInt1Ty(Context); BoolTrue = ConstantInt::getTrue(Context); BoolFalse = ConstantInt::getFalse(Context); BoolUndef = UndefValue::get(Boolean); return false; } /// \brief Build up the general order of nodes void AMDGPUStructurizeCFG::orderNodes() { scc_iterator I = scc_begin(ParentRegion), E = scc_end(ParentRegion); for (Order.clear(); I != E; ++I) { std::vector &Nodes = *I; Order.append(Nodes.begin(), Nodes.end()); } } /// \brief Build blocks and loop predicates void AMDGPUStructurizeCFG::buildPredicate(BranchInst *Term, unsigned Idx, BBPredicates &Pred, bool Invert) { Value *True = Invert ? BoolFalse : BoolTrue; Value *False = Invert ? BoolTrue : BoolFalse; RegionInfo *RI = ParentRegion->getRegionInfo(); BasicBlock *BB = Term->getParent(); // Handle the case where multiple regions start at the same block Region *R = BB != ParentRegion->getEntry() ? RI->getRegionFor(BB) : ParentRegion; if (R == ParentRegion) { // It's a top level block in our region Value *Cond = True; if (Term->isConditional()) { BasicBlock *Other = Term->getSuccessor(!Idx); if (Visited.count(Other)) { if (!Pred.count(Other)) Pred[Other] = False; if (!Pred.count(BB)) Pred[BB] = True; return; } Cond = Term->getCondition(); if (Idx != Invert) Cond = BinaryOperator::CreateNot(Cond, "", Term); } Pred[BB] = Cond; } else if (ParentRegion->contains(R)) { // It's a block in a sub region while(R->getParent() != ParentRegion) R = R->getParent(); Pred[R->getEntry()] = True; } else { // It's a branch from outside into our parent region Pred[BB] = True; } } /// \brief Analyze the successors of each block and build up predicates void AMDGPUStructurizeCFG::analyzeBlock(BasicBlock *BB) { pred_iterator PI = pred_begin(BB), PE = pred_end(BB); BBPredicates &Pred = Predicates[BB]; for (; PI != PE; ++PI) { BranchInst *Term = cast((*PI)->getTerminator()); for (unsigned i = 0, e = Term->getNumSuccessors(); i != e; ++i) { BasicBlock *Succ = Term->getSuccessor(i); if (Succ != BB) continue; buildPredicate(Term, i, Pred, false); } } } /// \brief Analyze the conditions leading to loop to a previous block void AMDGPUStructurizeCFG::analyzeLoop(BasicBlock *BB, unsigned &LoopIdx) { BranchInst *Term = cast(BB->getTerminator()); for (unsigned i = 0, e = Term->getNumSuccessors(); i != e; ++i) { BasicBlock *Succ = Term->getSuccessor(i); // Ignore it if it's not a back edge if (!Visited.count(Succ)) continue; buildPredicate(Term, i, LoopPred, true); LoopEnd = BB; if (Visited[Succ] < LoopIdx) { LoopIdx = Visited[Succ]; LoopStart = Succ; } } } /// \brief Collect various loop and predicate infos void AMDGPUStructurizeCFG::collectInfos() { unsigned Number = 0, LoopIdx = ~0; // Reset predicate Predicates.clear(); // and loop infos LoopStart = LoopEnd = 0; LoopPred.clear(); RNVector::reverse_iterator OI = Order.rbegin(), OE = Order.rend(); for (Visited.clear(); OI != OE; Visited[(*OI++)->getEntry()] = ++Number) { // Analyze all the conditions leading to a node analyzeBlock((*OI)->getEntry()); if ((*OI)->isSubRegion()) continue; // Find the first/last loop nodes and loop predicates analyzeLoop((*OI)->getNodeAs(), LoopIdx); } } /// \brief Does A dominate all the predicates of B ? bool AMDGPUStructurizeCFG::dominatesPredicates(BasicBlock *A, BasicBlock *B) { BBPredicates &Preds = Predicates[B]; for (BBPredicates::iterator PI = Preds.begin(), PE = Preds.end(); PI != PE; ++PI) { if (!DT->dominates(A, PI->first)) return false; } return true; } /// \brief Remove phi values from all successors and the remove the terminator. void AMDGPUStructurizeCFG::killTerminator(BasicBlock *BB) { TerminatorInst *Term = BB->getTerminator(); if (!Term) return; for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE; ++SI) { delPhiValues(BB, *SI); } Term->eraseFromParent(); } /// First: Skip forward to the first region node that either isn't a subregion or not /// dominating it's exit, remove all the skipped nodes from the node order. /// /// Second: Handle the first successor directly if the resulting nodes successor /// predicates are still dominated by the original entry RegionNode *AMDGPUStructurizeCFG::skipChained(RegionNode *Node) { BasicBlock *Entry = Node->getEntry(); // Skip forward as long as it is just a linear flow while (true) { BasicBlock *Entry = Node->getEntry(); BasicBlock *Exit; if (Node->isSubRegion()) { Exit = Node->getNodeAs()->getExit(); } else { TerminatorInst *Term = Entry->getTerminator(); if (Term->getNumSuccessors() != 1) break; Exit = Term->getSuccessor(0); } // It's a back edge, break here so we can insert a loop node if (!Visited.count(Exit)) return Node; // More than node edges are pointing to exit if (!DT->dominates(Entry, Exit)) return Node; RegionNode *Next = ParentRegion->getNode(Exit); RNVector::iterator I = std::find(Order.begin(), Order.end(), Next); assert(I != Order.end()); Visited.erase(Next->getEntry()); Order.erase(I); Node = Next; } BasicBlock *BB = Node->getEntry(); TerminatorInst *Term = BB->getTerminator(); if (Term->getNumSuccessors() != 2) return Node; // Our node has exactly two succesors, check if we can handle // any of them directly BasicBlock *Succ = Term->getSuccessor(0); if (!Visited.count(Succ) || !dominatesPredicates(Entry, Succ)) { Succ = Term->getSuccessor(1); if (!Visited.count(Succ) || !dominatesPredicates(Entry, Succ)) return Node; } else { BasicBlock *Succ2 = Term->getSuccessor(1); if (Visited.count(Succ2) && Visited[Succ] > Visited[Succ2] && dominatesPredicates(Entry, Succ2)) Succ = Succ2; } RegionNode *Next = ParentRegion->getNode(Succ); RNVector::iterator E = Order.end(); RNVector::iterator I = std::find(Order.begin(), E, Next); assert(I != E); killTerminator(BB); FlowsInserted.push_back(BB); Visited.erase(Succ); Order.erase(I); return ParentRegion->getNode(wireFlowBlock(BB, Next)); } /// \brief Remove all PHI values coming from "From" into "To" and remember /// them in DeletedPhis void AMDGPUStructurizeCFG::delPhiValues(BasicBlock *From, BasicBlock *To) { PhiMap &Map = DeletedPhis[To]; for (BasicBlock::iterator I = To->begin(), E = To->end(); I != E && isa(*I);) { PHINode &Phi = cast(*I++); while (Phi.getBasicBlockIndex(From) != -1) { Value *Deleted = Phi.removeIncomingValue(From, false); Map[&Phi].push_back(std::make_pair(From, Deleted)); } } } /// \brief Add the PHI values back once we knew the new predecessor void AMDGPUStructurizeCFG::addPhiValues(BasicBlock *From, BasicBlock *To) { if (!DeletedPhis.count(To)) return; PhiMap &Map = DeletedPhis[To]; SSAUpdater Updater; for (PhiMap::iterator I = Map.begin(), E = Map.end(); I != E; ++I) { PHINode *Phi = I->first; Updater.Initialize(Phi->getType(), ""); BasicBlock *Fallback = To; bool HaveFallback = false; for (BBValueVector::iterator VI = I->second.begin(), VE = I->second.end(); VI != VE; ++VI) { Updater.AddAvailableValue(VI->first, VI->second); BasicBlock *Dom = DT->findNearestCommonDominator(Fallback, VI->first); if (Dom == VI->first) HaveFallback = true; else if (Dom != Fallback) HaveFallback = false; Fallback = Dom; } if (!HaveFallback) { Value *Undef = UndefValue::get(Phi->getType()); Updater.AddAvailableValue(Fallback, Undef); } Phi->addIncoming(Updater.GetValueAtEndOfBlock(From), From); } DeletedPhis.erase(To); } /// \brief Create a new flow node and update dominator tree and region info BasicBlock *AMDGPUStructurizeCFG::getNextFlow(BasicBlock *Prev) { LLVMContext &Context = Func->getContext(); BasicBlock *Insert = Order.empty() ? ParentRegion->getExit() : Order.back()->getEntry(); BasicBlock *Flow = BasicBlock::Create(Context, FlowBlockName, Func, Insert); DT->addNewBlock(Flow, Prev); ParentRegion->getRegionInfo()->setRegionFor(Flow, ParentRegion); FlowsInserted.push_back(Flow); return Flow; } /// \brief Can we predict that this node will always be called? bool AMDGPUStructurizeCFG::isPredictableTrue(BasicBlock *Prev, BasicBlock *Node) { BBPredicates &Preds = Predicates[Node]; bool Dominated = false; for (BBPredicates::iterator I = Preds.begin(), E = Preds.end(); I != E; ++I) { if (I->second != BoolTrue) return false; if (!Dominated && DT->dominates(I->first, Prev)) Dominated = true; } return Dominated; } /// \brief Wire up the new control flow by inserting or updating the branch /// instructions at node exits BasicBlock *AMDGPUStructurizeCFG::wireFlowBlock(BasicBlock *Prev, RegionNode *Node) { BasicBlock *Entry = Node->getEntry(); if (LoopStart == Entry) { LoopStart = Prev; LoopPred[Prev] = BoolTrue; } // Wire it up temporary, skipChained may recurse into us BranchInst::Create(Entry, Prev); DT->changeImmediateDominator(Entry, Prev); addPhiValues(Prev, Entry); Node = skipChained(Node); BasicBlock *Next = getNextFlow(Prev); if (!isPredictableTrue(Prev, Entry)) { // Let Prev point to entry and next block Prev->getTerminator()->eraseFromParent(); BranchInst::Create(Entry, Next, BoolUndef, Prev); } else { DT->changeImmediateDominator(Next, Entry); } // Let node exit(s) point to next block if (Node->isSubRegion()) { Region *SubRegion = Node->getNodeAs(); BasicBlock *Exit = SubRegion->getExit(); // Find all the edges from the sub region to the exit BBVector ToDo; for (pred_iterator I = pred_begin(Exit), E = pred_end(Exit); I != E; ++I) { if (SubRegion->contains(*I)) ToDo.push_back(*I); } // Modify the edges to point to the new flow block for (BBVector::iterator I = ToDo.begin(), E = ToDo.end(); I != E; ++I) { delPhiValues(*I, Exit); TerminatorInst *Term = (*I)->getTerminator(); Term->replaceUsesOfWith(Exit, Next); } // Update the region info SubRegion->replaceExit(Next); } else { BasicBlock *BB = Node->getNodeAs(); killTerminator(BB); BranchInst::Create(Next, BB); if (BB == LoopEnd) LoopEnd = 0; } return Next; } /// Destroy node order and visited map, build up flow order instead. /// After this function control flow looks like it should be, but /// branches only have undefined conditions. void AMDGPUStructurizeCFG::createFlow() { DeletedPhis.clear(); BasicBlock *Prev = Order.pop_back_val()->getEntry(); assert(Prev == ParentRegion->getEntry() && "Incorrect node order!"); Visited.erase(Prev); if (LoopStart == Prev) { // Loop starts at entry, split entry so that we can predicate it BasicBlock::iterator Insert = Prev->getFirstInsertionPt(); BasicBlock *Split = Prev->splitBasicBlock(Insert, FlowBlockName); DT->addNewBlock(Split, Prev); ParentRegion->getRegionInfo()->setRegionFor(Split, ParentRegion); Predicates[Split] = Predicates[Prev]; Order.push_back(ParentRegion->getBBNode(Split)); LoopPred[Prev] = BoolTrue; } else if (LoopStart == Order.back()->getEntry()) { // Loop starts behind entry, split entry so that we can jump to it Instruction *Term = Prev->getTerminator(); BasicBlock *Split = Prev->splitBasicBlock(Term, FlowBlockName); DT->addNewBlock(Split, Prev); ParentRegion->getRegionInfo()->setRegionFor(Split, ParentRegion); Prev = Split; } killTerminator(Prev); FlowsInserted.clear(); FlowsInserted.push_back(Prev); while (!Order.empty()) { RegionNode *Node = Order.pop_back_val(); Visited.erase(Node->getEntry()); Prev = wireFlowBlock(Prev, Node); if (LoopStart && !LoopEnd) { // Create an extra loop end node LoopEnd = Prev; Prev = getNextFlow(LoopEnd); BranchInst::Create(Prev, LoopStart, BoolUndef, LoopEnd); addPhiValues(LoopEnd, LoopStart); } } BasicBlock *Exit = ParentRegion->getExit(); BranchInst::Create(Exit, Prev); addPhiValues(Prev, Exit); if (DT->dominates(ParentRegion->getEntry(), Exit)) DT->changeImmediateDominator(Exit, Prev); if (LoopStart && LoopEnd) { BBVector::iterator FI = std::find(FlowsInserted.begin(), FlowsInserted.end(), LoopStart); for (; *FI != LoopEnd; ++FI) { addPhiValues(*FI, (*FI)->getTerminator()->getSuccessor(0)); } } assert(Order.empty()); assert(Visited.empty()); assert(DeletedPhis.empty()); } /// \brief Insert the missing branch conditions void AMDGPUStructurizeCFG::insertConditions() { SSAUpdater PhiInserter; for (BBVector::iterator FI = FlowsInserted.begin(), FE = FlowsInserted.end(); FI != FE; ++FI) { BranchInst *Term = cast((*FI)->getTerminator()); if (Term->isUnconditional()) continue; PhiInserter.Initialize(Boolean, ""); PhiInserter.AddAvailableValue(&Func->getEntryBlock(), BoolFalse); BasicBlock *Succ = Term->getSuccessor(0); BBPredicates &Preds = (*FI == LoopEnd) ? LoopPred : Predicates[Succ]; for (BBPredicates::iterator PI = Preds.begin(), PE = Preds.end(); PI != PE; ++PI) { PhiInserter.AddAvailableValue(PI->first, PI->second); } Term->setCondition(PhiInserter.GetValueAtEndOfBlock(*FI)); } } /// Handle a rare case where the disintegrated nodes instructions /// no longer dominate all their uses. Not sure if this is really nessasary void AMDGPUStructurizeCFG::rebuildSSA() { SSAUpdater Updater; for (Region::block_iterator I = ParentRegion->block_begin(), E = ParentRegion->block_end(); I != E; ++I) { BasicBlock *BB = *I; for (BasicBlock::iterator II = BB->begin(), IE = BB->end(); II != IE; ++II) { bool Initialized = false; for (Use *I = &II->use_begin().getUse(), *Next; I; I = Next) { Next = I->getNext(); Instruction *User = cast(I->getUser()); if (User->getParent() == BB) { continue; } else if (PHINode *UserPN = dyn_cast(User)) { if (UserPN->getIncomingBlock(*I) == BB) continue; } if (DT->dominates(II, User)) continue; if (!Initialized) { Value *Undef = UndefValue::get(II->getType()); Updater.Initialize(II->getType(), ""); Updater.AddAvailableValue(&Func->getEntryBlock(), Undef); Updater.AddAvailableValue(BB, II); Initialized = true; } Updater.RewriteUseAfterInsertions(*I); } } } } /// \brief Run the transformation for each region found bool AMDGPUStructurizeCFG::runOnRegion(Region *R, RGPassManager &RGM) { if (R->isTopLevelRegion()) return false; Func = R->getEntry()->getParent(); ParentRegion = R; DT = &getAnalysis(); orderNodes(); collectInfos(); createFlow(); insertConditions(); rebuildSSA(); Order.clear(); Visited.clear(); Predicates.clear(); DeletedPhis.clear(); FlowsInserted.clear(); return true; } /// \brief Create the pass Pass *llvm::createAMDGPUStructurizeCFGPass() { return new AMDGPUStructurizeCFG(); }