X-Git-Url: http://plrg.eecs.uci.edu/git/?p=oota-llvm.git;a=blobdiff_plain;f=lib%2FTransforms%2FScalar%2FLoopUnswitch.cpp;h=8d39a4dece31e4374d8727bb897376e6ca3a0441;hp=a91d2955405e7ede896435cbdc4fda0ce59f4e3c;hb=4ee451de366474b9c228b4e5fa573795a715216d;hpb=3ed469ccd7b028a030b550d84b7336d146f5d8fa diff --git a/lib/Transforms/Scalar/LoopUnswitch.cpp b/lib/Transforms/Scalar/LoopUnswitch.cpp index a91d2955405..8d39a4dece3 100644 --- a/lib/Transforms/Scalar/LoopUnswitch.cpp +++ b/lib/Transforms/Scalar/LoopUnswitch.cpp @@ -2,8 +2,8 @@ // // 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 file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. // //===----------------------------------------------------------------------===// // @@ -29,41 +29,66 @@ #define DEBUG_TYPE "loop-unswitch" #include "llvm/Transforms/Scalar.h" #include "llvm/Constants.h" +#include "llvm/DerivedTypes.h" #include "llvm/Function.h" #include "llvm/Instructions.h" +#include "llvm/Analysis/ConstantFolding.h" #include "llvm/Analysis/LoopInfo.h" +#include "llvm/Analysis/LoopPass.h" +#include "llvm/Analysis/Dominators.h" #include "llvm/Transforms/Utils/Cloning.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" #include "llvm/ADT/Statistic.h" -#include "llvm/ADT/PostOrderIterator.h" -#include "llvm/Support/Debug.h" +#include "llvm/ADT/SmallPtrSet.h" #include "llvm/Support/CommandLine.h" +#include "llvm/Support/Compiler.h" +#include "llvm/Support/Debug.h" #include -#include #include using namespace llvm; +STATISTIC(NumBranches, "Number of branches unswitched"); +STATISTIC(NumSwitches, "Number of switches unswitched"); +STATISTIC(NumSelects , "Number of selects unswitched"); +STATISTIC(NumTrivial , "Number of unswitches that are trivial"); +STATISTIC(NumSimplify, "Number of simplifications of unswitched code"); + namespace { - Statistic<> NumBranches("loop-unswitch", "Number of branches unswitched"); - Statistic<> NumSwitches("loop-unswitch", "Number of switches unswitched"); - Statistic<> NumSelects ("loop-unswitch", "Number of selects unswitched"); - Statistic<> NumTrivial ("loop-unswitch", - "Number of unswitches that are trivial"); - Statistic<> NumSimplify("loop-unswitch", - "Number of simplifications of unswitched code"); cl::opt Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"), cl::init(10), cl::Hidden); - class LoopUnswitch : public FunctionPass { + class VISIBILITY_HIDDEN LoopUnswitch : public LoopPass { LoopInfo *LI; // Loop information + LPPassManager *LPM; - // LoopProcessWorklist - List of loops we need to process. + // LoopProcessWorklist - Used to check if second loop needs processing + // after RewriteLoopBodyWithConditionConstant rewrites first loop. std::vector LoopProcessWorklist; + SmallPtrSet UnswitchedVals; + + bool OptimizeForSize; + bool redoLoop; + + DominanceFrontier *DF; + DominatorTree *DT; + + /// LoopDF - Loop's dominance frontier. This set is a collection of + /// loop exiting blocks' DF member blocks. However this does set does not + /// includes basic blocks that are inside loop. + SmallPtrSet LoopDF; + + /// OrigLoopExitMap - This is used to map loop exiting block with + /// corresponding loop exit block, before updating CFG. + DenseMap OrigLoopExitMap; public: - virtual bool runOnFunction(Function &F); - bool visitLoop(Loop *L); + static char ID; // Pass ID, replacement for typeid + explicit LoopUnswitch(bool Os = false) : + LoopPass((intptr_t)&ID), OptimizeForSize(Os), redoLoop(false) {} + + bool runOnLoop(Loop *L, LPPassManager &LPM); + bool processLoop(Loop *L); /// This transformation requires natural loop information & requires that /// loop preheaders be inserted into the CFG... @@ -75,9 +100,12 @@ namespace { AU.addPreserved(); AU.addRequiredID(LCSSAID); AU.addPreservedID(LCSSAID); + AU.addPreserved(); + AU.addPreserved(); } private: + /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist, /// remove it. void RemoveLoopFromWorklist(Loop *L) { @@ -86,47 +114,42 @@ namespace { if (I != LoopProcessWorklist.end()) LoopProcessWorklist.erase(I); } + + /// Split all of the edges from inside the loop to their exit blocks. Update + /// the appropriate Phi nodes as we do so. + void SplitExitEdges(Loop *L, const SmallVector &ExitBlocks, + SmallVector &MiddleBlocks); + + /// If BB's dominance frontier has a member that is not part of loop L then + /// remove it. Add NewDFMember in BB's dominance frontier. + void ReplaceLoopExternalDFMember(Loop *L, BasicBlock *BB, + BasicBlock *NewDFMember); bool UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L); unsigned getLoopUnswitchCost(Loop *L, Value *LIC); void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val, BasicBlock *ExitBlock); void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L); - BasicBlock *SplitEdge(BasicBlock *From, BasicBlock *To); - BasicBlock *SplitBlock(BasicBlock *Old, Instruction *SplitPt); void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, Constant *Val, bool isEqual); - - void SimplifyCode(std::vector &Worklist); + + void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, + BasicBlock *TrueDest, + BasicBlock *FalseDest, + Instruction *InsertPt); + + void SimplifyCode(std::vector &Worklist, Loop *L); void RemoveBlockIfDead(BasicBlock *BB, - std::vector &Worklist); + std::vector &Worklist, Loop *l); void RemoveLoopFromHierarchy(Loop *L); }; + char LoopUnswitch::ID = 0; RegisterPass X("loop-unswitch", "Unswitch loops"); } -FunctionPass *llvm::createLoopUnswitchPass() { return new LoopUnswitch(); } - -bool LoopUnswitch::runOnFunction(Function &F) { - bool Changed = false; - LI = &getAnalysis(); - - // Populate the worklist of loops to process in post-order. - for (LoopInfo::iterator I = LI->begin(), E = LI->end(); I != E; ++I) - for (po_iterator LI = po_begin(*I), E = po_end(*I); LI != E; ++LI) - LoopProcessWorklist.push_back(*LI); - - // Process the loops in worklist order, this is a post-order visitation of - // the loops. We use a worklist of loops so that loops can be removed at any - // time if they are deleted (e.g. the backedge of a loop is removed). - while (!LoopProcessWorklist.empty()) { - Loop *L = LoopProcessWorklist.back(); - LoopProcessWorklist.pop_back(); - Changed |= visitLoop(L); - } - - return Changed; +LoopPass *llvm::createLoopUnswitchPass(bool Os) { + return new LoopUnswitch(Os); } /// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is @@ -135,7 +158,7 @@ bool LoopUnswitch::runOnFunction(Function &F) { static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { // Constants should be folded, not unswitched on! if (isa(Cond)) return false; - + // TODO: Handle: br (VARIANT|INVARIANT). // TODO: Hoist simple expressions out of loops. if (L->isLoopInvariant(Cond)) return Cond; @@ -151,15 +174,31 @@ static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed)) return RHS; } - - return 0; + + return 0; } -bool LoopUnswitch::visitLoop(Loop *L) { +bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { + LI = &getAnalysis(); + LPM = &LPM_Ref; + DF = getAnalysisToUpdate(); + DT = getAnalysisToUpdate(); + + bool Changed = false; + + do { + redoLoop = false; + Changed |= processLoop(L); + } while(redoLoop); + + return Changed; +} + +/// processLoop - Do actual work and unswitch loop if possible and profitable. +bool LoopUnswitch::processLoop(Loop *L) { assert(L->isLCSSAForm()); - bool Changed = false; - + // Loop over all of the basic blocks in the loop. If we find an interior // block that is branching on a loop-invariant condition, we can unswitch this // loop. @@ -173,7 +212,7 @@ bool LoopUnswitch::visitLoop(Loop *L) { // See if this, or some part of it, is loop invariant. If so, we can // unswitch on it if we desire. Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), L, Changed); - if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantBool::getTrue(), + if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(), L)) { ++NumBranches; return true; @@ -185,6 +224,10 @@ bool LoopUnswitch::visitLoop(Loop *L) { // Find a value to unswitch on: // FIXME: this should chose the most expensive case! Constant *UnswitchVal = SI->getCaseValue(1); + // Do not process same value again and again. + if (!UnswitchedVals.insert(UnswitchVal)) + continue; + if (UnswitchIfProfitable(LoopCond, UnswitchVal, L)) { ++NumSwitches; return true; @@ -197,7 +240,7 @@ bool LoopUnswitch::visitLoop(Loop *L) { BBI != E; ++BBI) if (SelectInst *SI = dyn_cast(BBI)) { Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), L, Changed); - if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantBool::getTrue(), + if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(), L)) { ++NumSelects; return true; @@ -288,9 +331,9 @@ static bool IsTrivialUnswitchCondition(Loop *L, Value *Cond, Constant **Val = 0, // side-effects. If so, determine the value of Cond that causes it to do // this. if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(0)))) { - if (Val) *Val = ConstantBool::getTrue(); + if (Val) *Val = ConstantInt::getTrue(); } else if ((LoopExitBB = isTrivialLoopExitBlock(L, BI->getSuccessor(1)))) { - if (Val) *Val = ConstantBool::getFalse(); + if (Val) *Val = ConstantInt::getFalse(); } } else if (SwitchInst *SI = dyn_cast(HeaderTerm)) { // If this isn't a switch on Cond, we can't handle it. @@ -365,13 +408,18 @@ unsigned LoopUnswitch::getLoopUnswitchCost(Loop *L, Value *LIC) { bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L){ // Check to see if it would be profitable to unswitch this loop. unsigned Cost = getLoopUnswitchCost(L, LoopCond); + + // Do not do non-trivial unswitch while optimizing for size. + if (Cost && OptimizeForSize) + return false; + if (Cost > Threshold) { // FIXME: this should estimate growth by the amount of code shared by the // resultant unswitched loops. // - DEBUG(std::cerr << "NOT unswitching loop %" - << L->getHeader()->getName() << ", cost too high: " - << L->getBlocks().size() << "\n"); + DOUT << "NOT unswitching loop %" + << L->getHeader()->getName() << ", cost too high: " + << L->getBlocks().size() << "\n"; return false; } @@ -388,92 +436,117 @@ bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val,Loop *L){ return true; } -/// SplitBlock - Split the specified block at the specified instruction - every -/// thing before SplitPt stays in Old and everything starting with SplitPt moves -/// to a new block. The two blocks are joined by an unconditional branch and -/// the loop info is updated. -/// -BasicBlock *LoopUnswitch::SplitBlock(BasicBlock *Old, Instruction *SplitPt) { - BasicBlock::iterator SplitIt = SplitPt; - while (isa(SplitIt)) - ++SplitIt; - BasicBlock *New = Old->splitBasicBlock(SplitIt, Old->getName()+".split"); - - // The new block lives in whichever loop the old one did. - if (Loop *L = LI->getLoopFor(Old)) - L->addBasicBlockToLoop(New, *LI); - - return New; -} - - -BasicBlock *LoopUnswitch::SplitEdge(BasicBlock *BB, BasicBlock *Succ) { - TerminatorInst *LatchTerm = BB->getTerminator(); - unsigned SuccNum = 0; - for (unsigned i = 0, e = LatchTerm->getNumSuccessors(); ; ++i) { - assert(i != e && "Didn't find edge?"); - if (LatchTerm->getSuccessor(i) == Succ) { - SuccNum = i; - break; - } - } - - // If this is a critical edge, let SplitCriticalEdge do it. - if (SplitCriticalEdge(BB->getTerminator(), SuccNum, this)) - return LatchTerm->getSuccessor(SuccNum); - - // If the edge isn't critical, then BB has a single successor or Succ has a - // single pred. Split the block. - BasicBlock::iterator SplitPoint; - if (BasicBlock *SP = Succ->getSinglePredecessor()) { - // If the successor only has a single pred, split the top of the successor - // block. - assert(SP == BB && "CFG broken"); - return SplitBlock(Succ, Succ->begin()); - } else { - // Otherwise, if BB has a single successor, split it at the bottom of the - // block. - assert(BB->getTerminator()->getNumSuccessors() == 1 && - "Should have a single succ!"); - return SplitBlock(BB, BB->getTerminator()); - } -} - - - // RemapInstruction - Convert the instruction operands from referencing the // current values into those specified by ValueMap. // static inline void RemapInstruction(Instruction *I, - std::map &ValueMap) { + DenseMap &ValueMap) { for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) { Value *Op = I->getOperand(op); - std::map::iterator It = ValueMap.find(Op); + DenseMap::iterator It = ValueMap.find(Op); if (It != ValueMap.end()) Op = It->second; I->setOperand(op, Op); } } +// CloneDomInfo - NewBB is cloned from Orig basic block. Now clone Dominator +// Info. +// +// If Orig block's immediate dominator is mapped in VM then use corresponding +// immediate dominator from the map. Otherwise Orig block's dominator is also +// NewBB's dominator. +// +// OrigPreheader is loop pre-header before this pass started +// updating CFG. NewPrehader is loops new pre-header. However, after CFG +// manipulation, loop L may not exist. So rely on input parameter NewPreheader. +void CloneDomInfo(BasicBlock *NewBB, BasicBlock *Orig, + BasicBlock *NewPreheader, BasicBlock *OrigPreheader, + BasicBlock *OrigHeader, + DominatorTree *DT, DominanceFrontier *DF, + DenseMap &VM) { + + // If NewBB alreay has found its place in domiantor tree then no need to do + // anything. + if (DT->getNode(NewBB)) + return; + + // If Orig does not have any immediate domiantor then its clone, NewBB, does + // not need any immediate dominator. + DomTreeNode *OrigNode = DT->getNode(Orig); + if (!OrigNode) + return; + DomTreeNode *OrigIDomNode = OrigNode->getIDom(); + if (!OrigIDomNode) + return; + + BasicBlock *OrigIDom = NULL; + + // If Orig is original loop header then its immediate dominator is + // NewPreheader. + if (Orig == OrigHeader) + OrigIDom = NewPreheader; + + // If Orig is new pre-header then its immediate dominator is + // original pre-header. + else if (Orig == NewPreheader) + OrigIDom = OrigPreheader; + + // Other as DT to find Orig's immediate dominator. + else + OrigIDom = OrigIDomNode->getBlock(); + + // Initially use Orig's immediate dominator as NewBB's immediate dominator. + BasicBlock *NewIDom = OrigIDom; + DenseMap::iterator I = VM.find(OrigIDom); + if (I != VM.end()) { + NewIDom = cast(I->second); + + // If NewIDom does not have corresponding dominatore tree node then + // get one. + if (!DT->getNode(NewIDom)) + CloneDomInfo(NewIDom, OrigIDom, NewPreheader, OrigPreheader, + OrigHeader, DT, DF, VM); + } + + DT->addNewBlock(NewBB, NewIDom); + + // Copy cloned dominance frontiner set + DominanceFrontier::DomSetType NewDFSet; + if (DF) { + DominanceFrontier::iterator DFI = DF->find(Orig); + if ( DFI != DF->end()) { + DominanceFrontier::DomSetType S = DFI->second; + for (DominanceFrontier::DomSetType::iterator I = S.begin(), E = S.end(); + I != E; ++I) { + BasicBlock *BB = *I; + DenseMap::iterator IDM = VM.find(BB); + if (IDM != VM.end()) + NewDFSet.insert(cast(IDM->second)); + else + NewDFSet.insert(BB); + } + } + DF->addBasicBlock(NewBB, NewDFSet); + } +} + /// CloneLoop - Recursively clone the specified loop and all of its children, /// mapping the blocks with the specified map. -static Loop *CloneLoop(Loop *L, Loop *PL, std::map &VM, - LoopInfo *LI) { +static Loop *CloneLoop(Loop *L, Loop *PL, DenseMap &VM, + LoopInfo *LI, LPPassManager *LPM) { Loop *New = new Loop(); - if (PL) - PL->addChildLoop(New); - else - LI->addTopLevelLoop(New); + LPM->insertLoop(New, PL); // Add all of the blocks in L to the new loop. for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); I != E; ++I) if (LI->getLoopFor(*I) == L) - New->addBasicBlockToLoop(cast(VM[*I]), *LI); + New->addBasicBlockToLoop(cast(VM[*I]), LI->getBase()); // Add all of the subloops to the new loop. for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I) - CloneLoop(*I, New, VM, LI); + CloneLoop(*I, New, VM, LI, LPM); return New; } @@ -481,22 +554,22 @@ static Loop *CloneLoop(Loop *L, Loop *PL, std::map &VM, /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest. Insert the /// code immediately before InsertPt. -static void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, - BasicBlock *TrueDest, - BasicBlock *FalseDest, - Instruction *InsertPt) { +void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, + BasicBlock *TrueDest, + BasicBlock *FalseDest, + Instruction *InsertPt) { // Insert a conditional branch on LIC to the two preheaders. The original // code is the true version and the new code is the false version. Value *BranchVal = LIC; - if (!isa(Val)) { - BranchVal = BinaryOperator::createSetEQ(LIC, Val, "tmp", InsertPt); - } else if (Val != ConstantBool::getTrue()) { + if (!isa(Val) || Val->getType() != Type::Int1Ty) + BranchVal = new ICmpInst(ICmpInst::ICMP_EQ, LIC, Val, "tmp", InsertPt); + else if (Val != ConstantInt::getTrue()) // We want to enter the new loop when the condition is true. std::swap(TrueDest, FalseDest); - } // Insert the new branch. new BranchInst(TrueDest, FalseDest, BranchVal, InsertPt); + } @@ -508,16 +581,16 @@ static void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val, BasicBlock *ExitBlock) { - DEBUG(std::cerr << "loop-unswitch: Trivial-Unswitch loop %" - << L->getHeader()->getName() << " [" << L->getBlocks().size() - << " blocks] in Function " << L->getHeader()->getParent()->getName() - << " on cond: " << *Val << " == " << *Cond << "\n"); + DOUT << "loop-unswitch: Trivial-Unswitch loop %" + << L->getHeader()->getName() << " [" << L->getBlocks().size() + << " blocks] in Function " << L->getHeader()->getParent()->getName() + << " on cond: " << *Val << " == " << *Cond << "\n"; // First step, split the preheader, so that we know that there is a safe place // to insert the conditional branch. We will change 'OrigPH' to have a // conditional branch on Cond. BasicBlock *OrigPH = L->getLoopPreheader(); - BasicBlock *NewPH = SplitEdge(OrigPH, L->getHeader()); + BasicBlock *NewPH = SplitEdge(OrigPH, L->getHeader(), this); // Now that we have a place to insert the conditional branch, create a place // to branch to: this is the exit block out of the loop that we should @@ -528,16 +601,17 @@ void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, // without actually branching to it (the exit block should be dominated by the // loop header, not the preheader). assert(!L->contains(ExitBlock) && "Exit block is in the loop?"); - BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin()); + BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this); // Okay, now we have a position to branch from and a position to branch to, // insert the new conditional branch. EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, OrigPH->getTerminator()); + LPM->deleteSimpleAnalysisValue(OrigPH->getTerminator(), L); OrigPH->getTerminator()->eraseFromParent(); // We need to reprocess this loop, it could be unswitched again. - LoopProcessWorklist.push_back(L); + redoLoop = true; // Now that we know that the loop is never entered when this condition is a // particular value, rewrite the loop with this info. We know that this will @@ -546,44 +620,43 @@ void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, ++NumTrivial; } +/// ReplaceLoopExternalDFMember - +/// If BB's dominance frontier has a member that is not part of loop L then +/// remove it. Add NewDFMember in BB's dominance frontier. +void LoopUnswitch::ReplaceLoopExternalDFMember(Loop *L, BasicBlock *BB, + BasicBlock *NewDFMember) { + + DominanceFrontier::iterator DFI = DF->find(BB); + if (DFI == DF->end()) + return; + + DominanceFrontier::DomSetType &DFSet = DFI->second; + for (DominanceFrontier::DomSetType::iterator DI = DFSet.begin(), + DE = DFSet.end(); DI != DE;) { + BasicBlock *B = *DI++; + if (L->contains(B)) + continue; -/// VersionLoop - We determined that the loop is profitable to unswitch when LIC -/// equal Val. Split it into loop versions and test the condition outside of -/// either loop. Return the loops created as Out1/Out2. -void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, - Loop *L) { - Function *F = L->getHeader()->getParent(); - DEBUG(std::cerr << "loop-unswitch: Unswitching loop %" - << L->getHeader()->getName() << " [" << L->getBlocks().size() - << " blocks] in Function " << F->getName() - << " when '" << *Val << "' == " << *LIC << "\n"); - - // LoopBlocks contains all of the basic blocks of the loop, including the - // preheader of the loop, the body of the loop, and the exit blocks of the - // loop, in that order. - std::vector LoopBlocks; - - // First step, split the preheader and exit blocks, and add these blocks to - // the LoopBlocks list. - BasicBlock *OrigPreheader = L->getLoopPreheader(); - LoopBlocks.push_back(SplitEdge(OrigPreheader, L->getHeader())); + DF->removeFromFrontier(DFI, B); + LoopDF.insert(B); + } - // We want the loop to come after the preheader, but before the exit blocks. - LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end()); + DF->addToFrontier(DFI, NewDFMember); +} - std::vector ExitBlocks; - L->getUniqueExitBlocks(ExitBlocks); +/// SplitExitEdges - +/// Split all of the edges from inside the loop to their exit blocks. Update +/// the appropriate Phi nodes as we do so. +void LoopUnswitch::SplitExitEdges(Loop *L, const SmallVector &ExitBlocks, + SmallVector &MiddleBlocks) { - // Split all of the edges from inside the loop to their exit blocks. Update - // the appropriate Phi nodes as we do so. for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { BasicBlock *ExitBlock = ExitBlocks[i]; std::vector Preds(pred_begin(ExitBlock), pred_end(ExitBlock)); for (unsigned j = 0, e = Preds.size(); j != e; ++j) { - assert(L->contains(Preds[j]) && - "All preds of loop exit blocks must be the same loop!"); - BasicBlock* MiddleBlock = SplitEdge(Preds[j], ExitBlock); + BasicBlock* MiddleBlock = SplitEdge(Preds[j], ExitBlock, this); + MiddleBlocks.push_back(MiddleBlock); BasicBlock* StartBlock = Preds[j]; BasicBlock* EndBlock; if (MiddleBlock->getSinglePredecessor() == ExitBlock) { @@ -593,6 +666,8 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, EndBlock = ExitBlock; } + OrigLoopExitMap[StartBlock] = EndBlock; + std::set InsertedPHIs; PHINode* OldLCSSA = 0; for (BasicBlock::iterator I = EndBlock->begin(); @@ -618,9 +693,60 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, OldLCSSA->replaceAllUsesWith(NewLCSSA); NewLCSSA->addIncoming(OldLCSSA, MiddleBlock); } + + if (DF && DT) { + // StartBlock -- > MiddleBlock -- > EndBlock + // StartBlock is loop exiting block. EndBlock will become merge point + // of two loop exits after loop unswitch. + + // If StartBlock's DF member includes a block that is not loop member + // then replace that DF member with EndBlock. + + // If MiddleBlock's DF member includes a block that is not loop member + // tnen replace that DF member with EndBlock. + + ReplaceLoopExternalDFMember(L, StartBlock, EndBlock); + ReplaceLoopExternalDFMember(L, MiddleBlock, EndBlock); + } } } - + +} + +/// UnswitchNontrivialCondition - We determined that the loop is profitable +/// to unswitch when LIC equal Val. Split it into loop versions and test the +/// condition outside of either loop. Return the loops created as Out1/Out2. +void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, + Loop *L) { + Function *F = L->getHeader()->getParent(); + DOUT << "loop-unswitch: Unswitching loop %" + << L->getHeader()->getName() << " [" << L->getBlocks().size() + << " blocks] in Function " << F->getName() + << " when '" << *Val << "' == " << *LIC << "\n"; + + // LoopBlocks contains all of the basic blocks of the loop, including the + // preheader of the loop, the body of the loop, and the exit blocks of the + // loop, in that order. + std::vector LoopBlocks; + + // First step, split the preheader and exit blocks, and add these blocks to + // the LoopBlocks list. + BasicBlock *OrigHeader = L->getHeader(); + BasicBlock *OrigPreheader = L->getLoopPreheader(); + BasicBlock *NewPreheader = SplitEdge(OrigPreheader, L->getHeader(), this); + LoopBlocks.push_back(NewPreheader); + + // We want the loop to come after the preheader, but before the exit blocks. + LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end()); + + SmallVector ExitBlocks; + L->getUniqueExitBlocks(ExitBlocks); + + // Split all of the edges from inside the loop to their exit blocks. Update + // the appropriate Phi nodes as we do so. + SmallVector MiddleBlocks; + SplitExitEdges(L, ExitBlocks, MiddleBlocks); + // The exit blocks may have been changed due to edge splitting, recompute. ExitBlocks.clear(); L->getUniqueExitBlocks(ExitBlocks); @@ -633,11 +759,27 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, // the instructions and blocks. std::vector NewBlocks; NewBlocks.reserve(LoopBlocks.size()); - std::map ValueMap; + DenseMap ValueMap; for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { BasicBlock *New = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F); NewBlocks.push_back(New); ValueMap[LoopBlocks[i]] = New; // Keep the BB mapping. + LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], New, L); + } + + // OutSiders are basic block that are dominated by original header and + // at the same time they are not part of loop. + SmallPtrSet OutSiders; + if (DT) { + DomTreeNode *OrigHeaderNode = DT->getNode(OrigHeader); + for(std::vector::iterator DI = OrigHeaderNode->begin(), + DE = OrigHeaderNode->end(); DI != DE; ++DI) { + BasicBlock *B = (*DI)->getBlock(); + + DenseMap::iterator VI = ValueMap.find(B); + if (VI == ValueMap.end()) + OutSiders.insert(B); + } } // Splice the newly inserted blocks into the function right before the @@ -646,19 +788,19 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, NewBlocks[0], F->end()); // Now we create the new Loop object for the versioned loop. - Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI); + Loop *NewLoop = CloneLoop(L, L->getParentLoop(), ValueMap, LI, LPM); Loop *ParentLoop = L->getParentLoop(); if (ParentLoop) { // Make sure to add the cloned preheader and exit blocks to the parent loop // as well. - ParentLoop->addBasicBlockToLoop(NewBlocks[0], *LI); + ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase()); } for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { BasicBlock *NewExit = cast(ValueMap[ExitBlocks[i]]); // The new exit block should be in the same loop as the old one. if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i])) - ExitBBLoop->addBasicBlockToLoop(NewExit, *LI); + ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase()); assert(NewExit->getTerminator()->getNumSuccessors() == 1 && "Exit block should have been split to have one successor!"); @@ -670,7 +812,7 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, for (BasicBlock::iterator I = ExitSucc->begin(); (PN = dyn_cast(I)); ++I) { Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); - std::map::iterator It = ValueMap.find(V); + DenseMap::iterator It = ValueMap.find(V); if (It != ValueMap.end()) V = It->second; PN->addIncoming(V, NewExit); } @@ -689,10 +831,97 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, // Emit the new branch that selects between the two versions of this loop. EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR); + LPM->deleteSimpleAnalysisValue(OldBR, L); OldBR->eraseFromParent(); - - LoopProcessWorklist.push_back(L); + + // Update dominator info + if (DF && DT) { + + SmallVector ExitingBlocks; + L->getExitingBlocks(ExitingBlocks); + + // Clone dominator info for all cloned basic block. + for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { + BasicBlock *LBB = LoopBlocks[i]; + BasicBlock *NBB = NewBlocks[i]; + CloneDomInfo(NBB, LBB, NewPreheader, OrigPreheader, + OrigHeader, DT, DF, ValueMap); + + // If LBB's dominance frontier includes DFMember + // such that DFMember is also a member of LoopDF then + // - Remove DFMember from LBB's dominance frontier + // - Copy loop exiting blocks', that are dominated by BB, dominance frontier + // member in BB's dominance frontier + + DominanceFrontier::iterator LBBI = DF->find(LBB); + DominanceFrontier::iterator NBBI = DF->find(NBB); + if (LBBI == DF->end()) + continue; + + DominanceFrontier::DomSetType &LBSet = LBBI->second; + for (DominanceFrontier::DomSetType::iterator LI = LBSet.begin(), + LE = LBSet.end(); LI != LE; /* NULL */) { + BasicBlock *B = *LI++; + if (B == LBB && B == L->getHeader()) + continue; + bool removeB = false; + if (!LoopDF.count(B)) + continue; + + // If LBB dominates loop exits then insert loop exit block's DF + // into B's DF. + for(SmallVector::iterator LExitI = ExitingBlocks.begin(), + LExitE = ExitingBlocks.end(); LExitI != LExitE; ++LExitI) { + BasicBlock *E = *LExitI; + + if (!DT->dominates(LBB,E)) + continue; + + DenseMap::iterator DFBI = + OrigLoopExitMap.find(E); + if (DFBI == OrigLoopExitMap.end()) + continue; + + BasicBlock *DFB = DFBI->second; + DF->addToFrontier(LBBI, DFB); + DF->addToFrontier(NBBI, DFB); + removeB = true; + } + + // If B's replacement is inserted in DF then now is the time to remove B. + if (removeB) { + DF->removeFromFrontier(LBBI, B); + if (L->contains(B)) + DF->removeFromFrontier(NBBI, cast(ValueMap[B])); + else + DF->removeFromFrontier(NBBI, B); + } + } + + } + + // MiddleBlocks are dominated by original pre header. SplitEdge updated + // MiddleBlocks' dominance frontier appropriately. + for (unsigned i = 0, e = MiddleBlocks.size(); i != e; ++i) { + BasicBlock *MBB = MiddleBlocks[i]; + if (!MBB->getSinglePredecessor()) + DT->changeImmediateDominator(MBB, OrigPreheader); + } + + // All Outsiders are now dominated by original pre header. + for (SmallPtrSet::iterator OI = OutSiders.begin(), + OE = OutSiders.end(); OI != OE; ++OI) { + BasicBlock *OB = *OI; + DT->changeImmediateDominator(OB, OrigPreheader); + } + + // New loop headers are dominated by original preheader + DT->changeImmediateDominator(NewBlocks[0], OrigPreheader); + DT->changeImmediateDominator(LoopBlocks[0], OrigPreheader); + } + LoopProcessWorklist.push_back(NewLoop); + redoLoop = true; // Now we rewrite the original code to know that the condition is true and the // new code to know that the condition is false. @@ -720,8 +949,9 @@ static void RemoveFromWorklist(Instruction *I, /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the /// program, replacing all uses with V and update the worklist. static void ReplaceUsesOfWith(Instruction *I, Value *V, - std::vector &Worklist) { - DEBUG(std::cerr << "Replace with '" << *V << "': " << *I); + std::vector &Worklist, + Loop *L, LPPassManager *LPM) { + DOUT << "Replace with '" << *V << "': " << *I; // Add uses to the worklist, which may be dead now. for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) @@ -732,9 +962,10 @@ static void ReplaceUsesOfWith(Instruction *I, Value *V, for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI) Worklist.push_back(cast(*UI)); + LPM->deleteSimpleAnalysisValue(I, L); + RemoveFromWorklist(I, Worklist); I->replaceAllUsesWith(V); I->eraseFromParent(); - RemoveFromWorklist(I, Worklist); ++NumSimplify; } @@ -742,7 +973,8 @@ static void ReplaceUsesOfWith(Instruction *I, Value *V, /// information, and remove any dead successors it has. /// void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, - std::vector &Worklist) { + std::vector &Worklist, + Loop *L) { if (pred_begin(BB) != pred_end(BB)) { // This block isn't dead, since an edge to BB was just removed, see if there // are any easy simplifications we can do now. @@ -751,7 +983,7 @@ void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, while (isa(BB->begin())) ReplaceUsesOfWith(BB->begin(), cast(BB->begin())->getIncomingValue(0), - Worklist); + Worklist, L, LPM); // If this is the header of a loop and the only pred is the latch, we now // have an unreachable loop. @@ -760,6 +992,7 @@ void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, // Remove the branch from the latch to the header block, this makes // the header dead, which will make the latch dead (because the header // dominates the latch). + LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L); Pred->getTerminator()->eraseFromParent(); new UnreachableInst(Pred); @@ -767,7 +1000,7 @@ void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, RemoveLoopFromHierarchy(L); // Reprocess the header, which now IS dead. - RemoveBlockIfDead(BB, Worklist); + RemoveBlockIfDead(BB, Worklist, L); return; } @@ -780,7 +1013,7 @@ void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, return; } - DEBUG(std::cerr << "Nuking dead block: " << *BB); + DOUT << "Nuking dead block: " << *BB; // Remove the instructions in the basic block from the worklist. for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) { @@ -817,8 +1050,8 @@ void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end()); // Remove the basic block, including all of the instructions contained in it. + LPM->deleteSimpleAnalysisValue(BB, L); BB->eraseFromParent(); - // Remove successor blocks here that are not dead, so that we know we only // have dead blocks in this list. Nondead blocks have a way of becoming dead, // then getting removed before we revisit them, which is badness. @@ -837,7 +1070,7 @@ void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, } for (unsigned i = 0, e = Succs.size(); i != e; ++i) - RemoveBlockIfDead(Succs[i], Worklist); + RemoveBlockIfDead(Succs[i], Worklist, L); } /// RemoveLoopFromHierarchy - We have discovered that the specified loop has @@ -847,54 +1080,7 @@ void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, /// so they just reparent loops. If the loops are actually dead, they will be /// removed later. void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) { - if (Loop *ParentLoop = L->getParentLoop()) { // Not a top-level loop. - // Reparent all of the blocks in this loop. Since BBLoop had a parent, - // they are now all in it. - for (Loop::block_iterator I = L->block_begin(), E = L->block_end(); - I != E; ++I) - if (LI->getLoopFor(*I) == L) // Don't change blocks in subloops. - LI->changeLoopFor(*I, ParentLoop); - - // Remove the loop from its parent loop. - for (Loop::iterator I = ParentLoop->begin(), E = ParentLoop->end();; - ++I) { - assert(I != E && "Couldn't find loop"); - if (*I == L) { - ParentLoop->removeChildLoop(I); - break; - } - } - - // Move all subloops into the parent loop. - while (L->begin() != L->end()) - ParentLoop->addChildLoop(L->removeChildLoop(L->end()-1)); - } else { - // Reparent all of the blocks in this loop. Since BBLoop had no parent, - // they no longer in a loop at all. - - for (unsigned i = 0; i != L->getBlocks().size(); ++i) { - // Don't change blocks in subloops. - if (LI->getLoopFor(L->getBlocks()[i]) == L) { - LI->removeBlock(L->getBlocks()[i]); - --i; - } - } - - // Remove the loop from the top-level LoopInfo object. - for (LoopInfo::iterator I = LI->begin(), E = LI->end();; ++I) { - assert(I != E && "Couldn't find loop"); - if (*I == L) { - LI->removeLoop(I); - break; - } - } - - // Move all of the subloops to the top-level. - while (L->begin() != L->end()) - LI->addTopLevelLoop(L->removeChildLoop(L->end()-1)); - } - - delete L; + LPM->deleteLoopFromQueue(L); RemoveLoopFromWorklist(L); } @@ -922,12 +1108,13 @@ void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC // in the loop with the appropriate one directly. - if (IsEqual || isa(Val)) { + if (IsEqual || (isa(Val) && Val->getType() == Type::Int1Ty)) { Value *Replacement; if (IsEqual) Replacement = Val; else - Replacement = ConstantBool::get(!cast(Val)->getValue()); + Replacement = ConstantInt::get(Type::Int1Ty, + !cast(Val)->getZExtValue()); for (unsigned i = 0, e = Users.size(); i != e; ++i) if (Instruction *U = cast(Users[i])) { @@ -961,15 +1148,15 @@ void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, // loop structure and put the block on an dead code path. BasicBlock* Old = SI->getParent(); - BasicBlock* Split = SplitBlock(Old, SI); + BasicBlock* Split = SplitBlock(Old, SI, this); Instruction* OldTerm = Old->getTerminator(); new BranchInst(Split, SI->getSuccessor(i), - ConstantBool::getTrue(), OldTerm); - + ConstantInt::getTrue(), OldTerm); + + LPM->deleteSimpleAnalysisValue(Old->getTerminator(), L); Old->getTerminator()->eraseFromParent(); - PHINode *PN; for (BasicBlock::iterator II = SI->getSuccessor(i)->begin(); (PN = dyn_cast(II)); ++II) { @@ -988,7 +1175,7 @@ void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, } } - SimplifyCode(Worklist); + SimplifyCode(Worklist, L); } /// SimplifyCode - Okay, now that we have simplified some instructions in the @@ -1000,27 +1187,28 @@ void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, /// FIXME: When the loop optimizer is more mature, separate this out to a new /// pass. /// -void LoopUnswitch::SimplifyCode(std::vector &Worklist) { +void LoopUnswitch::SimplifyCode(std::vector &Worklist, Loop *L) { while (!Worklist.empty()) { Instruction *I = Worklist.back(); Worklist.pop_back(); // Simple constant folding. if (Constant *C = ConstantFoldInstruction(I)) { - ReplaceUsesOfWith(I, C, Worklist); + ReplaceUsesOfWith(I, C, Worklist, L, LPM); continue; } // Simple DCE. if (isInstructionTriviallyDead(I)) { - DEBUG(std::cerr << "Remove dead instruction '" << *I); + DOUT << "Remove dead instruction '" << *I; // Add uses to the worklist, which may be dead now. for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) if (Instruction *Use = dyn_cast(I->getOperand(i))) Worklist.push_back(Use); - I->eraseFromParent(); + LPM->deleteSimpleAnalysisValue(I, L); RemoveFromWorklist(I, Worklist); + I->eraseFromParent(); ++NumSimplify; continue; } @@ -1028,32 +1216,37 @@ void LoopUnswitch::SimplifyCode(std::vector &Worklist) { // Special case hacks that appear commonly in unswitched code. switch (I->getOpcode()) { case Instruction::Select: - if (ConstantBool *CB = dyn_cast(I->getOperand(0))) { - ReplaceUsesOfWith(I, I->getOperand(!CB->getValue()+1), Worklist); + if (ConstantInt *CB = dyn_cast(I->getOperand(0))) { + ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L, + LPM); continue; } break; case Instruction::And: - if (isa(I->getOperand(0))) // constant -> RHS + if (isa(I->getOperand(0)) && + I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS cast(I)->swapOperands(); - if (ConstantBool *CB = dyn_cast(I->getOperand(1))) { - if (CB->getValue()) // X & 1 -> X - ReplaceUsesOfWith(I, I->getOperand(0), Worklist); - else // X & 0 -> 0 - ReplaceUsesOfWith(I, I->getOperand(1), Worklist); - continue; - } + if (ConstantInt *CB = dyn_cast(I->getOperand(1))) + if (CB->getType() == Type::Int1Ty) { + if (CB->isOne()) // X & 1 -> X + ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM); + else // X & 0 -> 0 + ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM); + continue; + } break; case Instruction::Or: - if (isa(I->getOperand(0))) // constant -> RHS + if (isa(I->getOperand(0)) && + I->getOperand(0)->getType() == Type::Int1Ty) // constant -> RHS cast(I)->swapOperands(); - if (ConstantBool *CB = dyn_cast(I->getOperand(1))) { - if (CB->getValue()) // X | 1 -> 1 - ReplaceUsesOfWith(I, I->getOperand(1), Worklist); - else // X | 0 -> X - ReplaceUsesOfWith(I, I->getOperand(0), Worklist); - continue; - } + if (ConstantInt *CB = dyn_cast(I->getOperand(1))) + if (CB->getType() == Type::Int1Ty) { + if (CB->isOne()) // X | 1 -> 1 + ReplaceUsesOfWith(I, I->getOperand(1), Worklist, L, LPM); + else // X | 0 -> X + ReplaceUsesOfWith(I, I->getOperand(0), Worklist, L, LPM); + continue; + } break; case Instruction::Br: { BranchInst *BI = cast(I); @@ -1066,16 +1259,17 @@ void LoopUnswitch::SimplifyCode(std::vector &Worklist) { if (!SinglePred) continue; // Nothing to do. assert(SinglePred == Pred && "CFG broken"); - DEBUG(std::cerr << "Merging blocks: " << Pred->getName() << " <- " - << Succ->getName() << "\n"); + DOUT << "Merging blocks: " << Pred->getName() << " <- " + << Succ->getName() << "\n"; // Resolve any single entry PHI nodes in Succ. while (PHINode *PN = dyn_cast(Succ->begin())) - ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist); + ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM); // Move all of the successor contents from Succ to Pred. Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(), Succ->end()); + LPM->deleteSimpleAnalysisValue(BI, L); BI->eraseFromParent(); RemoveFromWorklist(BI, Worklist); @@ -1085,23 +1279,25 @@ void LoopUnswitch::SimplifyCode(std::vector &Worklist) { // Remove Succ from the loop tree. LI->removeBlock(Succ); + LPM->deleteSimpleAnalysisValue(Succ, L); Succ->eraseFromParent(); ++NumSimplify; - } else if (ConstantBool *CB = dyn_cast(BI->getCondition())){ + } else if (ConstantInt *CB = dyn_cast(BI->getCondition())){ // Conditional branch. Turn it into an unconditional branch, then // remove dead blocks. break; // FIXME: Enable. - DEBUG(std::cerr << "Folded branch: " << *BI); - BasicBlock *DeadSucc = BI->getSuccessor(CB->getValue()); - BasicBlock *LiveSucc = BI->getSuccessor(!CB->getValue()); + DOUT << "Folded branch: " << *BI; + BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue()); + BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue()); DeadSucc->removePredecessor(BI->getParent(), true); Worklist.push_back(new BranchInst(LiveSucc, BI)); + LPM->deleteSimpleAnalysisValue(BI, L); BI->eraseFromParent(); RemoveFromWorklist(BI, Worklist); ++NumSimplify; - RemoveBlockIfDead(DeadSucc, Worklist); + RemoveBlockIfDead(DeadSucc, Worklist, L); } break; }