From: Chris Lattner Date: Sun, 22 Jun 2003 20:10:28 +0000 (+0000) Subject: Initial checkin of Tail duplication pass. X-Git-Url: http://demsky.eecs.uci.edu/git/?a=commitdiff_plain;h=7a7bef43554e0166de2b2570e3d68d2be813d219;p=oota-llvm.git Initial checkin of Tail duplication pass. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@6846 91177308-0d34-0410-b5e6-96231b3b80d8 --- diff --git a/lib/Transforms/Scalar/TailDuplication.cpp b/lib/Transforms/Scalar/TailDuplication.cpp new file mode 100644 index 00000000000..b947c3238b8 --- /dev/null +++ b/lib/Transforms/Scalar/TailDuplication.cpp @@ -0,0 +1,324 @@ +//===- TailDuplication.cpp - Simplify CFG through tail duplication --------===// +// +// This pass performs a limited form of tail duplication, intended to simplify +// CFGs by removing some unconditional branches. This pass is necessary to +// straighten out loops created by the C front-end, but also is capable of +// making other code nicer. After this pass is run, the CFG simplify pass +// should be run to clean up the mess. +// +// This pass could be enhanced in the future to use profile information to be +// more aggressive. +// +//===----------------------------------------------------------------------===// + +#include "llvm/Transforms/Scalar.h" +#include "llvm/Function.h" +#include "llvm/iPHINode.h" +#include "llvm/iTerminators.h" +#include "llvm/Pass.h" +#include "llvm/Type.h" +#include "llvm/Support/CFG.h" +#include "llvm/Transforms/Utils/Local.h" +#include "Support/Statistic.h" + +namespace { + Statistic<> NumEliminated("tailduplicate", + "Number of unconditional branches eliminated"); + Statistic<> NumPHINodes("tailduplicate", "Number of phi nodes inserted"); + + class TailDup : public FunctionPass { + bool runOnFunction(Function &F); + private: + inline bool shouldEliminateUnconditionalBranch(TerminatorInst *TI); + inline void eliminateUnconditionalBranch(BranchInst *BI); + inline void InsertPHINodesIfNecessary(Instruction *OrigInst, Value *NewInst, + BasicBlock *NewBlock); + inline Value *GetValueInBlock(BasicBlock *BB, Value *OrigVal, + std::map &ValueMap, + std::map &OutValueMap); + inline Value *GetValueOutBlock(BasicBlock *BB, Value *OrigVal, + std::map &ValueMap, + std::map &OutValueMap); + }; + RegisterOpt X("tailduplicate", "Tail Duplication"); +} + +Pass *createTailDuplicationPass() { return new TailDup(); } + +/// runOnFunction - Top level algorithm - Loop over each unconditional branch in +/// the function, eliminating it if it looks attractive enough. +/// +bool TailDup::runOnFunction(Function &F) { + bool Changed = false; + for (Function::iterator I = F.begin(), E = F.end(); I != E; ) + if (shouldEliminateUnconditionalBranch(I->getTerminator())) { + eliminateUnconditionalBranch(cast(I->getTerminator())); + Changed = true; + } else { + ++I; + } + return Changed; +} + +/// shouldEliminateUnconditionalBranch - Return true if this branch looks +/// attractive to eliminate. We eliminate the branch if the destination basic +/// block has <= 5 instructions in it, not counting PHI nodes. In practice, +/// since one of these is a terminator instruction, this means that we will add +/// up to 4 instructions to the new block. +/// +/// We don't count PHI nodes in the count since they will be removed when the +/// contents of the block are copied over. +/// +bool TailDup::shouldEliminateUnconditionalBranch(TerminatorInst *TI) { + BranchInst *BI = dyn_cast(TI); + if (!BI || !BI->isUnconditional()) return false; // Not an uncond branch! + + BasicBlock *Dest = BI->getSuccessor(0); + if (Dest == BI->getParent()) return false; // Do not loop infinitely! + + // Do not bother working on dead blocks... + pred_iterator PI = pred_begin(Dest), PE = pred_end(Dest); + if (PI == PE && Dest != Dest->getParent()->begin()) + return false; // It's just a dead block, ignore it... + + // Also, do not bother with blocks with only a single predecessor: simplify + // CFG will fold these two blocks together! + ++PI; + if (PI == PE) return false; // Exactly one predecessor! + + BasicBlock::iterator I = Dest->begin(); + while (isa(*I)) ++I; + + for (unsigned Size = 0; I != Dest->end(); ++Size, ++I) + if (Size == 6) return false; // The block is too large... + return true; +} + + +/// eliminateUnconditionalBranch - Clone the instructions from the destination +/// block into the source block, eliminating the specified unconditional branch. +/// If the destination block defines values used by successors of the dest +/// block, we may need to insert PHI nodes. +/// +void TailDup::eliminateUnconditionalBranch(BranchInst *Branch) { + BasicBlock *SourceBlock = Branch->getParent(); + BasicBlock *DestBlock = Branch->getSuccessor(0); + assert(SourceBlock != DestBlock && "Our predicate is broken!"); + + DEBUG(std::cerr << "TailDuplication[" << SourceBlock->getParent()->getName() + << "]: Eliminating branch: " << *Branch); + + // We are going to have to map operands from the original block B to the new + // copy of the block B'. If there are PHI nodes in the DestBlock, these PHI + // nodes also define part of this mapping. Loop over these PHI nodes, adding + // them to our mapping. + std::map ValueMapping; + + BasicBlock::iterator BI = DestBlock->begin(); + bool HadPHINodes = isa(BI); + for (; PHINode *PN = dyn_cast(BI); ++BI) + ValueMapping[PN] = PN->getIncomingValueForBlock(SourceBlock); + + // Clone the non-phi instructions of the dest block into the source block, + // keeping track of the mapping... + // + for (; BI != DestBlock->end(); ++BI) { + Instruction *New = BI->clone(); + New->setName(BI->getName()); + SourceBlock->getInstList().push_back(New); + ValueMapping[BI] = New; + } + + // Now that we have built the mapping information and cloned all of the + // instructions (giving us a new terminator, among other things), walk the new + // instructions, rewriting references of old instructions to use new + // instructions. + // + BI = Branch; ++BI; // Get an iterator to the first new instruction + for (; BI != SourceBlock->end(); ++BI) + for (unsigned i = 0, e = BI->getNumOperands(); i != e; ++i) + if (Value *Remapped = ValueMapping[BI->getOperand(i)]) + BI->setOperand(i, Remapped); + + // Next we check to see if any of the successors of DestBlock had PHI nodes. + // If so, we need to add entries to the PHI nodes for SourceBlock now. + for (succ_iterator SI = succ_begin(DestBlock), SE = succ_end(DestBlock); + SI != SE; ++SI) { + BasicBlock *Succ = *SI; + for (BasicBlock::iterator PNI = Succ->begin(); + PHINode *PN = dyn_cast(PNI); ++PNI) { + // Ok, we have a PHI node. Figure out what the incoming value was for the + // DestBlock. + Value *IV = PN->getIncomingValueForBlock(DestBlock); + + // Remap the value if necessary... + if (Value *MappedIV = ValueMapping[IV]) + IV = MappedIV; + PN->addIncoming(IV, SourceBlock); + } + } + + // Now that all of the instructions are correctly copied into the SourceBlock, + // we have one more minor problem: the successors of the original DestBB may + // use the values computed in DestBB either directly (if DestBB dominated the + // block), or through a PHI node. In either case, we need to insert PHI nodes + // into any successors of DestBB (which are now our successors) for each value + // that is computed in DestBB, but is used outside of it. All of these uses + // we have to rewrite with the new PHI node. + // + if (succ_begin(SourceBlock) != succ_end(SourceBlock)) // Avoid wasting time... + for (BI = DestBlock->begin(); BI != DestBlock->end(); ++BI) + if (BI->getType() != Type::VoidTy) + InsertPHINodesIfNecessary(BI, ValueMapping[BI], SourceBlock); + + // Final step: now that we have finished everything up, walk the cloned + // instructions one last time, constant propagating and DCE'ing them, because + // they may not be needed anymore. + // + BI = Branch; ++BI; // Get an iterator to the first new instruction + if (HadPHINodes) + while (BI != SourceBlock->end()) + if (!dceInstruction(BI) && !doConstantPropagation(BI)) + ++BI; + + DestBlock->removePredecessor(SourceBlock); // Remove entries in PHI nodes... + SourceBlock->getInstList().erase(Branch); // Destroy the uncond branch... + + ++NumEliminated; // We just killed a branch! +} + +/// InsertPHINodesIfNecessary - So at this point, we cloned the OrigInst +/// instruction into the NewBlock with the value of NewInst. If OrigInst was +/// used outside of its defining basic block, we need to insert a PHI nodes into +/// the successors. +/// +void TailDup::InsertPHINodesIfNecessary(Instruction *OrigInst, Value *NewInst, + BasicBlock *NewBlock) { + // Loop over all of the uses of OrigInst, rewriting them to be newly inserted + // PHI nodes, unless they are in the same basic block as OrigInst. + BasicBlock *OrigBlock = OrigInst->getParent(); + std::vector Users; + Users.reserve(OrigInst->use_size()); + for (Value::use_iterator I = OrigInst->use_begin(), E = OrigInst->use_end(); + I != E; ++I) { + Instruction *In = cast(*I); + if (In->getParent() != OrigBlock) // Don't modify uses in the orig block! + Users.push_back(In); + } + + // The common case is that the instruction is only used within the block that + // defines it. If we have this case, quick exit. + // + if (Users.empty()) return; + + // Otherwise, we have a more complex case, handle it now. This requires the + // construction of a mapping between a basic block and the value to use when + // in the scope of that basic block. This map will map to the original and + // new values when in the original or new block, but will map to inserted PHI + // nodes when in other blocks. + // + std::map ValueMap; + std::map OutValueMap; // The outgoing value map + OutValueMap[OrigBlock] = OrigInst; + OutValueMap[NewBlock ] = NewInst; // Seed the initial values... + + DEBUG(std::cerr << " ** Inserting PHI nodes for " << OrigInst); + while (!Users.empty()) { + Instruction *User = Users.back(); Users.pop_back(); + + if (PHINode *PN = dyn_cast(User)) { + // PHI nodes must be handled specially here, because their operands are + // actually defined in predecessor basic blocks, NOT in the block that the + // PHI node lives in. Note that we have already added entries to PHI nods + // which are in blocks that are immediate successors of OrigBlock, so + // don't modify them again. + for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) + if (PN->getIncomingValue(i) == OrigInst && + PN->getIncomingBlock(i) != OrigBlock) { + Value *V = GetValueOutBlock(PN->getIncomingBlock(i), OrigInst, + ValueMap, OutValueMap); + PN->setIncomingValue(i, V); + } + + } else { + // Any other user of the instruction can just replace any uses with the + // new value defined in the block it resides in. + Value *V = GetValueInBlock(User->getParent(), OrigInst, ValueMap, + OutValueMap); + User->replaceUsesOfWith(OrigInst, V); + } + } +} + +/// GetValueInBlock - This is a recursive method which inserts PHI nodes into +/// the function until there is a value available in basic block BB. +/// +Value *TailDup::GetValueInBlock(BasicBlock *BB, Value *OrigVal, + std::map &ValueMap, + std::map &OutValueMap) { + Value*& BBVal = ValueMap[BB]; + if (BBVal) return BBVal; // Value already computed for this block? + + assert(pred_begin(BB) != pred_end(BB) && + "Propagating PHI nodes to unreachable blocks?"); + + // If there is no value already available in this basic block, we need to + // either reuse a value from an incoming, dominating, basic block, or we need + // to create a new PHI node to merge in different incoming values. Because we + // don't know if we're part of a loop at this point or not, we create a PHI + // node, even if we will ultimately eliminate it. + PHINode *PN = new PHINode(OrigVal->getType(), OrigVal->getName()+".pn", + BB->begin()); + BBVal = PN; // Insert this into the BBVal slot in case of cycles... + + Value*& BBOutVal = OutValueMap[BB]; + if (BBOutVal == 0) BBOutVal = PN; + + // Now that we have created the PHI node, loop over all of the predecessors of + // this block, computing an incoming value for the predecessor. + std::vector Preds(pred_begin(BB), pred_end(BB)); + for (unsigned i = 0, e = Preds.size(); i != e; ++i) + PN->addIncoming(GetValueOutBlock(Preds[i], OrigVal, ValueMap, OutValueMap), + Preds[i]); + + // The PHI node is complete. In many cases, however the PHI node was + // ultimately unnecessary: we could have just reused a dominating incoming + // value. If this is the case, nuke the PHI node and replace the map entry + // with the dominating value. + // + assert(PN->getNumIncomingValues() > 0 && "No predecessors?"); + + // Check to see if all of the elements in the PHI node are either the PHI node + // itself or ONE particular value. + unsigned i = 0; + Value *ReplVal = PN->getIncomingValue(i); + for (; ReplVal == PN && i != PN->getNumIncomingValues(); ++i) + ReplVal = PN->getIncomingValue(i); // Skip values equal to the PN + + for (; i != PN->getNumIncomingValues(); ++i) + if (PN->getIncomingValue(i) != PN && PN->getIncomingValue(i) != ReplVal) { + ReplVal = 0; + break; + } + + // Found a value to replace the PHI node with? + if (ReplVal) { + PN->replaceAllUsesWith(ReplVal); + BBVal = ReplVal; + if (BBOutVal == PN) BBOutVal = ReplVal; + BB->getInstList().erase(PN); // Erase the PHI node... + } else { + ++NumPHINodes; + } + + return BBVal; +} + +Value *TailDup::GetValueOutBlock(BasicBlock *BB, Value *OrigVal, + std::map &ValueMap, + std::map &OutValueMap) { + Value*& BBVal = OutValueMap[BB]; + if (BBVal) return BBVal; // Value already computed for this block? + + return BBVal = GetValueInBlock(BB, OrigVal, ValueMap, OutValueMap); +}