X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=blobdiff_plain;f=lib%2FTransforms%2FScalar%2FLoopRotation.cpp;h=7eeb1527ad401c9b187832224a18f944d3061890;hp=1118ed55918751c3a2ca93b48ebcfe3c9a5fb82d;hb=a94d6e87c4c49f2e81b01d66d8bfb591277f8f96;hpb=f6784a326293941cd4ff6d25da64732760e993f3 diff --git a/lib/Transforms/Scalar/LoopRotation.cpp b/lib/Transforms/Scalar/LoopRotation.cpp index 1118ed55918..7eeb1527ad4 100644 --- a/lib/Transforms/Scalar/LoopRotation.cpp +++ b/lib/Transforms/Scalar/LoopRotation.cpp @@ -12,21 +12,20 @@ //===----------------------------------------------------------------------===// #define DEBUG_TYPE "loop-rotate" - #include "llvm/Transforms/Scalar.h" #include "llvm/Function.h" -#include "llvm/Instructions.h" -#include "llvm/Analysis/LoopInfo.h" +#include "llvm/IntrinsicInst.h" +#include "llvm/Analysis/CodeMetrics.h" #include "llvm/Analysis/LoopPass.h" -#include "llvm/Analysis/Dominators.h" +#include "llvm/Analysis/InstructionSimplify.h" #include "llvm/Analysis/ScalarEvolution.h" +#include "llvm/Analysis/ValueTracking.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" -#include "llvm/Support/CommandLine.h" +#include "llvm/Transforms/Utils/SSAUpdater.h" +#include "llvm/Transforms/Utils/ValueMapper.h" #include "llvm/Support/Debug.h" #include "llvm/ADT/Statistic.h" -#include "llvm/ADT/SmallVector.h" - using namespace llvm; #define MAX_HEADER_SIZE 16 @@ -34,523 +33,426 @@ using namespace llvm; STATISTIC(NumRotated, "Number of loops rotated"); namespace { - class VISIBILITY_HIDDEN RenameData { - public: - RenameData(Instruction *O, Value *P, Instruction *H) - : Original(O), PreHeader(P), Header(H) { } - public: - Instruction *Original; // Original instruction - Value *PreHeader; // Original pre-header replacement - Instruction *Header; // New header replacement - }; - - class VISIBILITY_HIDDEN LoopRotate : public LoopPass { - + class LoopRotate : public LoopPass { public: static char ID; // Pass ID, replacement for typeid - LoopRotate() : LoopPass(&ID) {} - - // Rotate Loop L as many times as possible. Return true if - // loop is rotated at least once. - bool runOnLoop(Loop *L, LPPassManager &LPM); + LoopRotate() : LoopPass(ID) { + initializeLoopRotatePass(*PassRegistry::getPassRegistry()); + } // LCSSA form makes instruction renaming easier. virtual void getAnalysisUsage(AnalysisUsage &AU) const { + AU.addPreserved(); + AU.addRequired(); + AU.addPreserved(); AU.addRequiredID(LoopSimplifyID); AU.addPreservedID(LoopSimplifyID); AU.addRequiredID(LCSSAID); AU.addPreservedID(LCSSAID); AU.addPreserved(); - AU.addPreserved(); - AU.addPreserved(); - AU.addPreserved(); } - // Helper functions - - /// Do actual work - bool rotateLoop(Loop *L, LPPassManager &LPM); - - /// Initialize local data - void initialize(); - - /// Make sure all Exit block PHINodes have required incoming values. - /// If incoming value is constant or defined outside the loop then - /// PHINode may not have an entry for original pre-header. - void updateExitBlock(); - - /// Return true if this instruction is used outside original header. - bool usedOutsideOriginalHeader(Instruction *In); - - /// Find Replacement information for instruction. Return NULL if it is - /// not available. - const RenameData *findReplacementData(Instruction *I); - - /// After loop rotation, loop pre-header has multiple sucessors. - /// Insert one forwarding basic block to ensure that loop pre-header - /// has only one successor. - void preserveCanonicalLoopForm(LPPassManager &LPM); + bool runOnLoop(Loop *L, LPPassManager &LPM); + void simplifyLoopLatch(Loop *L); + bool rotateLoop(Loop *L); private: - - Loop *L; - BasicBlock *OrigHeader; - BasicBlock *OrigPreHeader; - BasicBlock *OrigLatch; - BasicBlock *NewHeader; - BasicBlock *Exit; - LPPassManager *LPM_Ptr; - SmallVector LoopHeaderInfo; + LoopInfo *LI; }; } - + char LoopRotate::ID = 0; -static RegisterPass X("loop-rotate", "Rotate Loops"); +INITIALIZE_PASS_BEGIN(LoopRotate, "loop-rotate", "Rotate Loops", false, false) +INITIALIZE_PASS_DEPENDENCY(LoopInfo) +INITIALIZE_PASS_DEPENDENCY(LoopSimplify) +INITIALIZE_PASS_DEPENDENCY(LCSSA) +INITIALIZE_PASS_END(LoopRotate, "loop-rotate", "Rotate Loops", false, false) Pass *llvm::createLoopRotatePass() { return new LoopRotate(); } /// Rotate Loop L as many times as possible. Return true if -/// loop is rotated at least once. -bool LoopRotate::runOnLoop(Loop *Lp, LPPassManager &LPM) { +/// the loop is rotated at least once. +bool LoopRotate::runOnLoop(Loop *L, LPPassManager &LPM) { + LI = &getAnalysis(); - bool RotatedOneLoop = false; - initialize(); - LPM_Ptr = &LPM; + // Simplify the loop latch before attempting to rotate the header + // upward. Rotation may not be needed if the loop tail can be folded into the + // loop exit. + simplifyLoopLatch(L); // One loop can be rotated multiple times. - while (rotateLoop(Lp,LPM)) { - RotatedOneLoop = true; - initialize(); - } + bool MadeChange = false; + while (rotateLoop(L)) + MadeChange = true; - return RotatedOneLoop; + return MadeChange; } -/// Rotate loop LP. Return true if the loop is rotated. -bool LoopRotate::rotateLoop(Loop *Lp, LPPassManager &LPM) { - L = Lp; +/// RewriteUsesOfClonedInstructions - We just cloned the instructions from the +/// old header into the preheader. If there were uses of the values produced by +/// these instruction that were outside of the loop, we have to insert PHI nodes +/// to merge the two values. Do this now. +static void RewriteUsesOfClonedInstructions(BasicBlock *OrigHeader, + BasicBlock *OrigPreheader, + ValueToValueMapTy &ValueMap) { + // Remove PHI node entries that are no longer live. + BasicBlock::iterator I, E = OrigHeader->end(); + for (I = OrigHeader->begin(); PHINode *PN = dyn_cast(I); ++I) + PN->removeIncomingValue(PN->getBasicBlockIndex(OrigPreheader)); + + // Now fix up users of the instructions in OrigHeader, inserting PHI nodes + // as necessary. + SSAUpdater SSA; + for (I = OrigHeader->begin(); I != E; ++I) { + Value *OrigHeaderVal = I; + + // If there are no uses of the value (e.g. because it returns void), there + // is nothing to rewrite. + if (OrigHeaderVal->use_empty()) + continue; - OrigHeader = L->getHeader(); - OrigPreHeader = L->getLoopPreheader(); - OrigLatch = L->getLoopLatch(); + Value *OrigPreHeaderVal = ValueMap[OrigHeaderVal]; - // If loop has only one block then there is not much to rotate. - if (L->getBlocks().size() == 1) - return false; + // The value now exits in two versions: the initial value in the preheader + // and the loop "next" value in the original header. + SSA.Initialize(OrigHeaderVal->getType(), OrigHeaderVal->getName()); + SSA.AddAvailableValue(OrigHeader, OrigHeaderVal); + SSA.AddAvailableValue(OrigPreheader, OrigPreHeaderVal); - assert(OrigHeader && OrigLatch && OrigPreHeader && - "Loop is not in canonical form"); + // Visit each use of the OrigHeader instruction. + for (Value::use_iterator UI = OrigHeaderVal->use_begin(), + UE = OrigHeaderVal->use_end(); UI != UE; ) { + // Grab the use before incrementing the iterator. + Use &U = UI.getUse(); - // If loop header is not one of the loop exit block then - // either this loop is already rotated or it is not - // suitable for loop rotation transformations. - if (!L->isLoopExit(OrigHeader)) - return false; + // Increment the iterator before removing the use from the list. + ++UI; - BranchInst *BI = dyn_cast(OrigHeader->getTerminator()); - if (!BI) - return false; - assert(BI->isConditional() && "Branch Instruction is not conditional"); + // SSAUpdater can't handle a non-PHI use in the same block as an + // earlier def. We can easily handle those cases manually. + Instruction *UserInst = cast(U.getUser()); + if (!isa(UserInst)) { + BasicBlock *UserBB = UserInst->getParent(); - // Updating PHInodes in loops with multiple exits adds complexity. - // Keep it simple, and restrict loop rotation to loops with one exit only. - // In future, lift this restriction and support for multiple exits if - // required. - SmallVector ExitBlocks; - L->getExitBlocks(ExitBlocks); - if (ExitBlocks.size() > 1) - return false; - - // Check size of original header and reject - // loop if it is very big. - if (OrigHeader->size() > MAX_HEADER_SIZE) - return false; - - // Now, this loop is suitable for rotation. - - // Find new Loop header. NewHeader is a Header's one and only successor - // that is inside loop. Header's other successor is outside the - // loop. Otherwise loop is not suitable for rotation. - Exit = BI->getSuccessor(0); - NewHeader = BI->getSuccessor(1); - if (L->contains(Exit)) - std::swap(Exit, NewHeader); - assert(NewHeader && "Unable to determine new loop header"); - assert(L->contains(NewHeader) && !L->contains(Exit) && - "Unable to determine loop header and exit blocks"); - - // Copy PHI nodes and other instructions from original header - // into original pre-header. Unlike original header, original pre-header is - // not a member of loop. - // - // New loop header is one and only successor of original header that - // is inside the loop. All other original header successors are outside - // the loop. Copy PHI Nodes from original header into new loop header. - // Add second incoming value, from original loop pre-header into these phi - // nodes. If a value defined in original header is used outside original - // header then new loop header will need new phi nodes with two incoming - // values, one definition from original header and second definition is - // from original loop pre-header. - - // Remove terminator from Original pre-header. Original pre-header will - // receive a clone of original header terminator as a new terminator. - OrigPreHeader->getInstList().pop_back(); - BasicBlock::iterator I = OrigHeader->begin(), E = OrigHeader->end(); - PHINode *PN = 0; - for (; (PN = dyn_cast(I)); ++I) { - // PHI nodes are not copied into original pre-header. Instead their values - // are directly propagated. - Value *NPV = PN->getIncomingValueForBlock(OrigPreHeader); - - // Create new PHI node with two incoming values for NewHeader. - // One incoming value is from OrigLatch (through OrigHeader) and - // second incoming value is from original pre-header. - PHINode *NH = PHINode::Create(PN->getType(), PN->getName(), - NewHeader->begin()); - NH->addIncoming(PN->getIncomingValueForBlock(OrigLatch), OrigHeader); - NH->addIncoming(NPV, OrigPreHeader); - - // "In" can be replaced by NH at various places. - LoopHeaderInfo.push_back(RenameData(PN, NPV, NH)); - } + // The original users in the OrigHeader are already using the + // original definitions. + if (UserBB == OrigHeader) + continue; - // Now, handle non-phi instructions. - for (; I != E; ++I) { - Instruction *In = I; - assert(!isa(In) && "PHINode is not expected here"); - - // This is not a PHI instruction. Insert its clone into original pre-header. - // If this instruction is using a value from same basic block then - // update it to use value from cloned instruction. - Instruction *C = In->clone(); - C->setName(In->getName()); - OrigPreHeader->getInstList().push_back(C); - - for (unsigned opi = 0, e = In->getNumOperands(); opi != e; ++opi) { - Instruction *OpInsn = dyn_cast(In->getOperand(opi)); - if (!OpInsn) continue; // Ignore non-instruction values. - if (const RenameData *D = findReplacementData(OpInsn)) - C->setOperand(opi, D->PreHeader); - } + // Users in the OrigPreHeader need to use the value to which the + // original definitions are mapped. + if (UserBB == OrigPreheader) { + U = OrigPreHeaderVal; + continue; + } + } - // If this instruction is used outside this basic block then - // create new PHINode for this instruction. - Instruction *NewHeaderReplacement = NULL; - if (usedOutsideOriginalHeader(In)) { - PHINode *PN = PHINode::Create(In->getType(), In->getName(), - NewHeader->begin()); - PN->addIncoming(In, OrigHeader); - PN->addIncoming(C, OrigPreHeader); - NewHeaderReplacement = PN; + // Anything else can be handled by SSAUpdater. + SSA.RewriteUse(U); } - LoopHeaderInfo.push_back(RenameData(In, C, NewHeaderReplacement)); } +} - // Rename uses of original header instructions to reflect their new - // definitions (either from original pre-header node or from newly created - // new header PHINodes. - // - // Original header instructions are used in - // 1) Original header: - // - // If instruction is used in non-phi instructions then it is using - // defintion from original heder iteself. Do not replace this use - // with definition from new header or original pre-header. - // - // If instruction is used in phi node then it is an incoming - // value. Rename its use to reflect new definition from new-preheader - // or new header. - // - // 2) Inside loop but not in original header - // - // Replace this use to reflect definition from new header. - for (unsigned LHI = 0, LHI_E = LoopHeaderInfo.size(); LHI != LHI_E; ++LHI) { - const RenameData &ILoopHeaderInfo = LoopHeaderInfo[LHI]; - - if (!ILoopHeaderInfo.Header) - continue; +/// Determine whether the instructions in this range my be safely and cheaply +/// speculated. This is not an important enough situation to develop complex +/// heuristics. We handle a single arithmetic instruction along with any type +/// conversions. +static bool shouldSpeculateInstrs(BasicBlock::iterator Begin, + BasicBlock::iterator End) { + bool seenIncrement = false; + for (BasicBlock::iterator I = Begin; I != End; ++I) { - Instruction *OldPhi = ILoopHeaderInfo.Original; - Instruction *NewPhi = ILoopHeaderInfo.Header; - - // Before replacing uses, collect them first, so that iterator is - // not invalidated. - SmallVector AllUses; - for (Value::use_iterator UI = OldPhi->use_begin(), UE = OldPhi->use_end(); - UI != UE; ++UI) - AllUses.push_back(cast(UI)); - - for (SmallVector::iterator UI = AllUses.begin(), - UE = AllUses.end(); UI != UE; ++UI) { - Instruction *U = *UI; - BasicBlock *Parent = U->getParent(); - - // Used inside original header - if (Parent == OrigHeader) { - // Do not rename uses inside original header non-phi instructions. - PHINode *PU = dyn_cast(U); - if (!PU) - continue; + if (!isSafeToSpeculativelyExecute(I)) + return false; - // Do not rename uses inside original header phi nodes, if the - // incoming value is for new header. - if (PU->getBasicBlockIndex(NewHeader) != -1 - && PU->getIncomingValueForBlock(NewHeader) == U) - continue; - - U->replaceUsesOfWith(OldPhi, NewPhi); - continue; - } + if (isa(I)) + continue; - // Used inside loop, but not in original header. - if (L->contains(U->getParent())) { - if (U != NewPhi) - U->replaceUsesOfWith(OldPhi, NewPhi); - continue; - } - - // Used inside Exit Block. Since we are in LCSSA form, U must be PHINode. - if (U->getParent() == Exit) { - assert(isa(U) && "Use in Exit Block that is not PHINode"); - - PHINode *UPhi = cast(U); - // UPhi already has one incoming argument from original header. - // Add second incoming argument from new Pre header. - UPhi->addIncoming(ILoopHeaderInfo.PreHeader, OrigPreHeader); - } else { - // Used outside Exit block. Create a new PHI node from exit block - // to receive value from ne new header ane pre header. - PHINode *PN = PHINode::Create(U->getType(), U->getName(), - Exit->begin()); - PN->addIncoming(ILoopHeaderInfo.PreHeader, OrigPreHeader); - PN->addIncoming(OldPhi, OrigHeader); - U->replaceUsesOfWith(OldPhi, PN); - } + switch (I->getOpcode()) { + default: + return false; + case Instruction::GetElementPtr: + // GEPs are cheap if all indices are constant. + if (!cast(I)->hasAllConstantIndices()) + return false; + // fall-thru to increment case + case Instruction::Add: + case Instruction::Sub: + case Instruction::And: + case Instruction::Or: + case Instruction::Xor: + case Instruction::Shl: + case Instruction::LShr: + case Instruction::AShr: + if (seenIncrement) + return false; + seenIncrement = true; + break; + case Instruction::Trunc: + case Instruction::ZExt: + case Instruction::SExt: + // ignore type conversions + break; } } - - /// Make sure all Exit block PHINodes have required incoming values. - updateExitBlock(); + return true; +} - // Update CFG +/// Fold the loop tail into the loop exit by speculating the loop tail +/// instructions. Typically, this is a single post-increment. In the case of a +/// simple 2-block loop, hoisting the increment can be much better than +/// duplicating the entire loop header. In the cast of loops with early exits, +/// rotation will not work anyway, but simplifyLoopLatch will put the loop in +/// canonical form so downstream passes can handle it. +/// +/// I don't believe this invalidates SCEV. +void LoopRotate::simplifyLoopLatch(Loop *L) { + BasicBlock *Latch = L->getLoopLatch(); + if (!Latch || Latch->hasAddressTaken()) + return; + + BranchInst *Jmp = dyn_cast(Latch->getTerminator()); + if (!Jmp || !Jmp->isUnconditional()) + return; + + BasicBlock *LastExit = Latch->getSinglePredecessor(); + if (!LastExit || !L->isLoopExiting(LastExit)) + return; + + BranchInst *BI = dyn_cast(LastExit->getTerminator()); + if (!BI) + return; - // Removing incoming branch from loop preheader to original header. - // Now original header is inside the loop. - for (BasicBlock::iterator I = OrigHeader->begin(), E = OrigHeader->end(); - I != E; ++I) - if (PHINode *PN = dyn_cast(I)) - PN->removeIncomingValue(OrigPreHeader); + if (!shouldSpeculateInstrs(Latch->begin(), Jmp)) + return; - // Make NewHeader as the new header for the loop. - L->moveToHeader(NewHeader); + DEBUG(dbgs() << "Folding loop latch " << Latch->getName() << " into " + << LastExit->getName() << "\n"); - preserveCanonicalLoopForm(LPM); + // Hoist the instructions from Latch into LastExit. + LastExit->getInstList().splice(BI, Latch->getInstList(), Latch->begin(), Jmp); - NumRotated++; - return true; -} + unsigned FallThruPath = BI->getSuccessor(0) == Latch ? 0 : 1; + BasicBlock *Header = Jmp->getSuccessor(0); + assert(Header == L->getHeader() && "expected a backward branch"); -/// Make sure all Exit block PHINodes have required incoming values. -/// If incoming value is constant or defined outside the loop then -/// PHINode may not have an entry for original pre-header. -void LoopRotate::updateExitBlock() { + // Remove Latch from the CFG so that LastExit becomes the new Latch. + BI->setSuccessor(FallThruPath, Header); + Latch->replaceSuccessorsPhiUsesWith(LastExit); + Jmp->eraseFromParent(); - for (BasicBlock::iterator I = Exit->begin(), E = Exit->end(); - I != E; ++I) { + // Nuke the Latch block. + assert(Latch->empty() && "unable to evacuate Latch"); + LI->removeBlock(Latch); + if (DominatorTree *DT = getAnalysisIfAvailable()) + DT->eraseNode(Latch); + Latch->eraseFromParent(); +} - PHINode *PN = dyn_cast(I); - if (!PN) - break; +/// Rotate loop LP. Return true if the loop is rotated. +bool LoopRotate::rotateLoop(Loop *L) { + // If the loop has only one block then there is not much to rotate. + if (L->getBlocks().size() == 1) + return false; - // There is already one incoming value from original pre-header block. - if (PN->getBasicBlockIndex(OrigPreHeader) != -1) - continue; + BasicBlock *OrigHeader = L->getHeader(); - const RenameData *ILoopHeaderInfo; - Value *V = PN->getIncomingValueForBlock(OrigHeader); - if (isa(V) && - (ILoopHeaderInfo = findReplacementData(cast(V)))) { - assert(ILoopHeaderInfo->PreHeader && "Missing New Preheader Instruction"); - PN->addIncoming(ILoopHeaderInfo->PreHeader, OrigPreHeader); - } else { - PN->addIncoming(V, OrigPreHeader); - } - } -} + BranchInst *BI = dyn_cast(OrigHeader->getTerminator()); + if (BI == 0 || BI->isUnconditional()) + return false; -/// Initialize local data -void LoopRotate::initialize() { - L = NULL; - OrigHeader = NULL; - OrigPreHeader = NULL; - NewHeader = NULL; - Exit = NULL; + // If the loop header is not one of the loop exiting blocks then + // either this loop is already rotated or it is not + // suitable for loop rotation transformations. + if (!L->isLoopExiting(OrigHeader)) + return false; - LoopHeaderInfo.clear(); -} + // Updating PHInodes in loops with multiple exits adds complexity. + // Keep it simple, and restrict loop rotation to loops with one exit only. + // In future, lift this restriction and support for multiple exits if + // required. + SmallVector ExitBlocks; + L->getExitBlocks(ExitBlocks); + if (ExitBlocks.size() > 1) + return false; -/// Return true if this instruction is used by any instructions in the loop that -/// aren't in original header. -bool LoopRotate::usedOutsideOriginalHeader(Instruction *In) { - for (Value::use_iterator UI = In->use_begin(), UE = In->use_end(); - UI != UE; ++UI) { - BasicBlock *UserBB = cast(UI)->getParent(); - if (UserBB != OrigHeader && L->contains(UserBB)) - return true; + // Check size of original header and reject loop if it is very big. + { + CodeMetrics Metrics; + Metrics.analyzeBasicBlock(OrigHeader); + if (Metrics.NumInsts > MAX_HEADER_SIZE) + return false; } - return false; -} + // Now, this loop is suitable for rotation. + BasicBlock *OrigPreheader = L->getLoopPreheader(); + BasicBlock *OrigLatch = L->getLoopLatch(); -/// Find Replacement information for instruction. Return NULL if it is -/// not available. -const RenameData *LoopRotate::findReplacementData(Instruction *In) { + // If the loop could not be converted to canonical form, it must have an + // indirectbr in it, just give up. + if (OrigPreheader == 0 || OrigLatch == 0) + return false; - // Since LoopHeaderInfo is small, linear walk is OK. - for (unsigned LHI = 0, LHI_E = LoopHeaderInfo.size(); LHI != LHI_E; ++LHI) { - const RenameData &ILoopHeaderInfo = LoopHeaderInfo[LHI]; - if (ILoopHeaderInfo.Original == In) - return &ILoopHeaderInfo; - } - return NULL; -} + // Anything ScalarEvolution may know about this loop or the PHI nodes + // in its header will soon be invalidated. + if (ScalarEvolution *SE = getAnalysisIfAvailable()) + SE->forgetLoop(L); -/// After loop rotation, loop pre-header has multiple sucessors. -/// Insert one forwarding basic block to ensure that loop pre-header -/// has only one successor. -void LoopRotate::preserveCanonicalLoopForm(LPPassManager &LPM) { - - // Right now original pre-header has two successors, new header and - // exit block. Insert new block between original pre-header and - // new header such that loop's new pre-header has only one successor. - BasicBlock *NewPreHeader = BasicBlock::Create("bb.nph", - OrigHeader->getParent(), - NewHeader); - LoopInfo &LI = LPM.getAnalysis(); - if (Loop *PL = LI.getLoopFor(OrigPreHeader)) - PL->addBasicBlockToLoop(NewPreHeader, LI.getBase()); - BranchInst::Create(NewHeader, NewPreHeader); - - BranchInst *OrigPH_BI = cast(OrigPreHeader->getTerminator()); - if (OrigPH_BI->getSuccessor(0) == NewHeader) - OrigPH_BI->setSuccessor(0, NewPreHeader); - else { - assert(OrigPH_BI->getSuccessor(1) == NewHeader && - "Unexpected original pre-header terminator"); - OrigPH_BI->setSuccessor(1, NewPreHeader); - } - - for (BasicBlock::iterator I = NewHeader->begin(), E = NewHeader->end(); - I != E; ++I) { - PHINode *PN = dyn_cast(I); - if (!PN) - break; + // Find new Loop header. NewHeader is a Header's one and only successor + // that is inside loop. Header's other successor is outside the + // loop. Otherwise loop is not suitable for rotation. + BasicBlock *Exit = BI->getSuccessor(0); + BasicBlock *NewHeader = BI->getSuccessor(1); + if (L->contains(Exit)) + std::swap(Exit, NewHeader); + assert(NewHeader && "Unable to determine new loop header"); + assert(L->contains(NewHeader) && !L->contains(Exit) && + "Unable to determine loop header and exit blocks"); - int index = PN->getBasicBlockIndex(OrigPreHeader); - assert(index != -1 && "Expected incoming value from Original PreHeader"); - PN->setIncomingBlock(index, NewPreHeader); - assert(PN->getBasicBlockIndex(OrigPreHeader) == -1 && - "Expected only one incoming value from Original PreHeader"); - } + // This code assumes that the new header has exactly one predecessor. + // Remove any single-entry PHI nodes in it. + assert(NewHeader->getSinglePredecessor() && + "New header doesn't have one pred!"); + FoldSingleEntryPHINodes(NewHeader); - if (DominatorTree *DT = getAnalysisToUpdate()) { - DT->addNewBlock(NewPreHeader, OrigPreHeader); - DT->changeImmediateDominator(L->getHeader(), NewPreHeader); - DT->changeImmediateDominator(Exit, OrigPreHeader); - for (Loop::block_iterator BI = L->block_begin(), BE = L->block_end(); - BI != BE; ++BI) { - BasicBlock *B = *BI; - if (L->getHeader() != B) { - DomTreeNode *Node = DT->getNode(B); - if (Node && Node->getBlock() == OrigHeader) - DT->changeImmediateDominator(*BI, L->getHeader()); - } + // Begin by walking OrigHeader and populating ValueMap with an entry for + // each Instruction. + BasicBlock::iterator I = OrigHeader->begin(), E = OrigHeader->end(); + ValueToValueMapTy ValueMap; + + // For PHI nodes, the value available in OldPreHeader is just the + // incoming value from OldPreHeader. + for (; PHINode *PN = dyn_cast(I); ++I) + ValueMap[PN] = PN->getIncomingValueForBlock(OrigPreheader); + + // For the rest of the instructions, either hoist to the OrigPreheader if + // possible or create a clone in the OldPreHeader if not. + TerminatorInst *LoopEntryBranch = OrigPreheader->getTerminator(); + while (I != E) { + Instruction *Inst = I++; + + // If the instruction's operands are invariant and it doesn't read or write + // memory, then it is safe to hoist. Doing this doesn't change the order of + // execution in the preheader, but does prevent the instruction from + // executing in each iteration of the loop. This means it is safe to hoist + // something that might trap, but isn't safe to hoist something that reads + // memory (without proving that the loop doesn't write). + if (L->hasLoopInvariantOperands(Inst) && + !Inst->mayReadFromMemory() && !Inst->mayWriteToMemory() && + !isa(Inst) && !isa(Inst) && + !isa(Inst)) { + Inst->moveBefore(LoopEntryBranch); + continue; } - DT->changeImmediateDominator(OrigHeader, OrigLatch); - } - if (DominanceFrontier *DF = getAnalysisToUpdate()) { - // New Preheader's dominance frontier is Exit block. - DominanceFrontier::DomSetType NewPHSet; - NewPHSet.insert(Exit); - DF->addBasicBlock(NewPreHeader, NewPHSet); - - // New Header's dominance frontier now includes itself and Exit block - DominanceFrontier::iterator HeadI = DF->find(L->getHeader()); - if (HeadI != DF->end()) { - DominanceFrontier::DomSetType & HeaderSet = HeadI->second; - HeaderSet.clear(); - HeaderSet.insert(L->getHeader()); - HeaderSet.insert(Exit); + // Otherwise, create a duplicate of the instruction. + Instruction *C = Inst->clone(); + + // Eagerly remap the operands of the instruction. + RemapInstruction(C, ValueMap, + RF_NoModuleLevelChanges|RF_IgnoreMissingEntries); + + // With the operands remapped, see if the instruction constant folds or is + // otherwise simplifyable. This commonly occurs because the entry from PHI + // nodes allows icmps and other instructions to fold. + Value *V = SimplifyInstruction(C); + if (V && LI->replacementPreservesLCSSAForm(C, V)) { + // If so, then delete the temporary instruction and stick the folded value + // in the map. + delete C; + ValueMap[Inst] = V; } else { - DominanceFrontier::DomSetType HeaderSet; - HeaderSet.insert(L->getHeader()); - HeaderSet.insert(Exit); - DF->addBasicBlock(L->getHeader(), HeaderSet); + // Otherwise, stick the new instruction into the new block! + C->setName(Inst->getName()); + C->insertBefore(LoopEntryBranch); + ValueMap[Inst] = C; } + } - // Original header (new Loop Latch)'s dominance frontier is Exit. - DominanceFrontier::iterator LatchI = DF->find(L->getLoopLatch()); - if (LatchI != DF->end()) { - DominanceFrontier::DomSetType &LatchSet = LatchI->second; - LatchSet = LatchI->second; - LatchSet.clear(); - LatchSet.insert(Exit); - } else { - DominanceFrontier::DomSetType LatchSet; - LatchSet.insert(Exit); - DF->addBasicBlock(L->getHeader(), LatchSet); + // Along with all the other instructions, we just cloned OrigHeader's + // terminator into OrigPreHeader. Fix up the PHI nodes in each of OrigHeader's + // successors by duplicating their incoming values for OrigHeader. + TerminatorInst *TI = OrigHeader->getTerminator(); + for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) + for (BasicBlock::iterator BI = TI->getSuccessor(i)->begin(); + PHINode *PN = dyn_cast(BI); ++BI) + PN->addIncoming(PN->getIncomingValueForBlock(OrigHeader), OrigPreheader); + + // Now that OrigPreHeader has a clone of OrigHeader's terminator, remove + // OrigPreHeader's old terminator (the original branch into the loop), and + // remove the corresponding incoming values from the PHI nodes in OrigHeader. + LoopEntryBranch->eraseFromParent(); + + // If there were any uses of instructions in the duplicated block outside the + // loop, update them, inserting PHI nodes as required + RewriteUsesOfClonedInstructions(OrigHeader, OrigPreheader, ValueMap); + + // NewHeader is now the header of the loop. + L->moveToHeader(NewHeader); + assert(L->getHeader() == NewHeader && "Latch block is our new header"); + + + // At this point, we've finished our major CFG changes. As part of cloning + // the loop into the preheader we've simplified instructions and the + // duplicated conditional branch may now be branching on a constant. If it is + // branching on a constant and if that constant means that we enter the loop, + // then we fold away the cond branch to an uncond branch. This simplifies the + // loop in cases important for nested loops, and it also means we don't have + // to split as many edges. + BranchInst *PHBI = cast(OrigPreheader->getTerminator()); + assert(PHBI->isConditional() && "Should be clone of BI condbr!"); + if (!isa(PHBI->getCondition()) || + PHBI->getSuccessor(cast(PHBI->getCondition())->isZero()) + != NewHeader) { + // The conditional branch can't be folded, handle the general case. + // Update DominatorTree to reflect the CFG change we just made. Then split + // edges as necessary to preserve LoopSimplify form. + if (DominatorTree *DT = getAnalysisIfAvailable()) { + // Since OrigPreheader now has the conditional branch to Exit block, it is + // the dominator of Exit. + DT->changeImmediateDominator(Exit, OrigPreheader); + DT->changeImmediateDominator(NewHeader, OrigPreheader); + + // Update OrigHeader to be dominated by the new header block. + DT->changeImmediateDominator(OrigHeader, OrigLatch); } - // If a loop block dominates new loop latch then its frontier is - // new header and Exit. - BasicBlock *NewLatch = L->getLoopLatch(); - DominatorTree *DT = getAnalysisToUpdate(); - for (Loop::block_iterator BI = L->block_begin(), BE = L->block_end(); - BI != BE; ++BI) { - BasicBlock *B = *BI; - if (DT->dominates(B, NewLatch)) { - DominanceFrontier::iterator BDFI = DF->find(B); - if (BDFI != DF->end()) { - DominanceFrontier::DomSetType &BSet = BDFI->second; - BSet = BDFI->second; - BSet.clear(); - BSet.insert(L->getHeader()); - BSet.insert(Exit); - } else { - DominanceFrontier::DomSetType BSet; - BSet.insert(L->getHeader()); - BSet.insert(Exit); - DF->addBasicBlock(B, BSet); - } - } + // Right now OrigPreHeader has two successors, NewHeader and ExitBlock, and + // thus is not a preheader anymore. + // Split the edge to form a real preheader. + BasicBlock *NewPH = SplitCriticalEdge(OrigPreheader, NewHeader, this); + NewPH->setName(NewHeader->getName() + ".lr.ph"); + + // Preserve canonical loop form, which means that 'Exit' should have only + // one predecessor. + BasicBlock *ExitSplit = SplitCriticalEdge(L->getLoopLatch(), Exit, this); + ExitSplit->moveBefore(Exit); + } else { + // We can fold the conditional branch in the preheader, this makes things + // simpler. The first step is to remove the extra edge to the Exit block. + Exit->removePredecessor(OrigPreheader, true /*preserve LCSSA*/); + BranchInst *NewBI = BranchInst::Create(NewHeader, PHBI); + NewBI->setDebugLoc(PHBI->getDebugLoc()); + PHBI->eraseFromParent(); + + // With our CFG finalized, update DomTree if it is available. + if (DominatorTree *DT = getAnalysisIfAvailable()) { + // Update OrigHeader to be dominated by the new header block. + DT->changeImmediateDominator(NewHeader, OrigPreheader); + DT->changeImmediateDominator(OrigHeader, OrigLatch); } } - // Preserve canonical loop form, which means Exit block should - // have only one predecessor. - BasicBlock *NExit = SplitEdge(L->getLoopLatch(), Exit, this); - - // Preserve LCSSA. - BasicBlock::iterator I = Exit->begin(), E = Exit->end(); - PHINode *PN = NULL; - for (; (PN = dyn_cast(I)); ++I) { - unsigned N = PN->getNumIncomingValues(); - for (unsigned index = 0; index < N; ++index) - if (PN->getIncomingBlock(index) == NExit) { - PHINode *NewPN = PHINode::Create(PN->getType(), PN->getName(), - NExit->begin()); - NewPN->addIncoming(PN->getIncomingValue(index), L->getLoopLatch()); - PN->setIncomingValue(index, NewPN); - PN->setIncomingBlock(index, NExit); - break; - } - } + assert(L->getLoopPreheader() && "Invalid loop preheader after loop rotation"); + assert(L->getLoopLatch() && "Invalid loop latch after loop rotation"); - assert(NewHeader && L->getHeader() == NewHeader && - "Invalid loop header after loop rotation"); - assert(NewPreHeader && L->getLoopPreheader() == NewPreHeader && - "Invalid loop preheader after loop rotation"); - assert(L->getLoopLatch() && - "Invalid loop latch after loop rotation"); + // Now that the CFG and DomTree are in a consistent state again, try to merge + // the OrigHeader block into OrigLatch. This will succeed if they are + // connected by an unconditional branch. This is just a cleanup so the + // emitted code isn't too gross in this common case. + MergeBlockIntoPredecessor(OrigHeader, this); + + ++NumRotated; + return true; } +