X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FScalar%2FLoopUnswitch.cpp;h=e7c91420b17540bec8416b3f279f57c940fedec5;hb=a5a1e46d0004bc4a2de5d4726e3af8301fab1532;hp=de124d10aaae08acec35c6afd02ee4344a5959f7;hpb=a78130c3207e59bfb2ec0b77ca3f48677685d9ae;p=oota-llvm.git diff --git a/lib/Transforms/Scalar/LoopUnswitch.cpp b/lib/Transforms/Scalar/LoopUnswitch.cpp index de124d10aaa..e7c91420b17 100644 --- a/lib/Transforms/Scalar/LoopUnswitch.cpp +++ b/lib/Transforms/Scalar/LoopUnswitch.cpp @@ -32,11 +32,12 @@ #include "llvm/DerivedTypes.h" #include "llvm/Function.h" #include "llvm/Instructions.h" -#include "llvm/Analysis/ConstantFolding.h" #include "llvm/Analysis/InlineCost.h" +#include "llvm/Analysis/InstructionSimplify.h" #include "llvm/Analysis/LoopInfo.h" #include "llvm/Analysis/LoopPass.h" #include "llvm/Analysis/Dominators.h" +#include "llvm/Analysis/ScalarEvolution.h" #include "llvm/Transforms/Utils/Cloning.h" #include "llvm/Transforms/Utils/Local.h" #include "llvm/Transforms/Utils/BasicBlockUtils.h" @@ -55,14 +56,67 @@ 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"); +STATISTIC(TotalInsts, "Total number of instructions analyzed"); // The specific value of 50 here was chosen based only on intuition and a // few specific examples. static cl::opt Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"), - cl::init(50), cl::Hidden); + cl::init(100), cl::Hidden); namespace { + + class LUAnalysisCache { + + typedef DenseMap > + UnswitchedValsMap; + + typedef UnswitchedValsMap::iterator UnswitchedValsIt; + + struct LoopProperties { + unsigned CanBeUnswitchedCount; + unsigned SizeEstimation; + UnswitchedValsMap UnswitchedVals; + }; + + typedef DenseMap LoopPropsMap; + typedef LoopPropsMap::iterator LoopPropsMapIt; + + LoopPropsMap LoopsProperties; + UnswitchedValsMap* CurLoopInstructions; + LoopProperties* CurrentLoopProperties; + + unsigned MaxSize; + + public: + + LUAnalysisCache() : + CurLoopInstructions(NULL), CurrentLoopProperties(NULL), + MaxSize(Threshold) + {} + + // Analyze loop. Check its size, calculate is it possible to unswitch + // it. Returns true if we can unswitch this loop. + bool countLoop(const Loop* L); + + // Clean all data related to given loop. + void forgetLoop(const Loop* L); + + // Mark case value as unswitched. + // Since SI instruction can be partly unswitched, in order to avoid + // extra unswitching in cloned loops keep track all unswitched values. + void setUnswitched(const SwitchInst* SI, const Value* V); + + // Check was this case value unswitched before or not. + bool isUnswitched(const SwitchInst* SI, const Value* V); + + // Clone all loop-unswitch related loop properties. + // Redistribute unswitching quotas. + // Note, that new loop data is stored inside the VMap. + void cloneData(const Loop* NewLoop, const Loop* OldLoop, + const ValueToValueMapTy& VMap); + }; + class LoopUnswitch : public LoopPass { LoopInfo *LI; // Loop information LPPassManager *LPM; @@ -70,13 +124,26 @@ namespace { // LoopProcessWorklist - Used to check if second loop needs processing // after RewriteLoopBodyWithConditionConstant rewrites first loop. std::vector LoopProcessWorklist; - SmallPtrSet UnswitchedVals; + + // TODO: This few lines are here for cosmetic purposes only. + // Will be removed with the next commit. + struct LoopProperties { + unsigned CanBeUnswitchedCount; + unsigned SizeEstimation; + }; + + // TODO: This few lines are here for cosmetic purposes only. + // Will be removed with the next commit. + typedef DenseMap LoopPropsMap; + typedef LoopPropsMap::iterator LoopPropsMapIt; + LoopPropsMap LoopsProperties; + + LUAnalysisCache BranchesInfo; bool OptimizeForSize; bool redoLoop; Loop *currentLoop; - DominanceFrontier *DF; DominatorTree *DT; BasicBlock *loopHeader; BasicBlock *loopPreheader; @@ -91,15 +158,17 @@ namespace { public: static char ID; // Pass ID, replacement for typeid explicit LoopUnswitch(bool Os = false) : - LoopPass(&ID), OptimizeForSize(Os), redoLoop(false), - currentLoop(NULL), DF(NULL), DT(NULL), loopHeader(NULL), - loopPreheader(NULL) {} + LoopPass(ID), OptimizeForSize(Os), redoLoop(false), + currentLoop(NULL), DT(NULL), loopHeader(NULL), + loopPreheader(NULL) { + initializeLoopUnswitchPass(*PassRegistry::getPassRegistry()); + } bool runOnLoop(Loop *L, LPPassManager &LPM); bool processCurrentLoop(); /// This transformation requires natural loop information & requires that - /// loop preheaders be inserted into the CFG... + /// loop preheaders be inserted into the CFG. /// virtual void getAnalysisUsage(AnalysisUsage &AU) const { AU.addRequiredID(LoopSimplifyID); @@ -109,13 +178,13 @@ namespace { AU.addRequiredID(LCSSAID); AU.addPreservedID(LCSSAID); AU.addPreserved(); - AU.addPreserved(); + AU.addPreserved(); } private: virtual void releaseMemory() { - UnswitchedVals.clear(); + BranchesInfo.forgetLoop(currentLoop); } /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist, @@ -126,7 +195,7 @@ namespace { if (I != LoopProcessWorklist.end()) LoopProcessWorklist.erase(I); } - + void initLoopData() { loopHeader = currentLoop->getHeader(); loopPreheader = currentLoop->getLoopPreheader(); @@ -158,8 +227,121 @@ namespace { }; } + +// Analyze loop. Check its size, calculate is it possible to unswitch +// it. Returns true if we can unswitch this loop. +bool LUAnalysisCache::countLoop(const Loop* L) { + + std::pair InsertRes = + LoopsProperties.insert(std::make_pair(L, LoopProperties())); + + LoopProperties& Props = InsertRes.first->second; + + if (InsertRes.second) { + // New loop. + + // Limit the number of instructions to avoid causing significant code + // expansion, and the number of basic blocks, to avoid loops with + // large numbers of branches which cause loop unswitching to go crazy. + // This is a very ad-hoc heuristic. + + // FIXME: This is overly conservative because it does not take into + // consideration code simplification opportunities and code that can + // be shared by the resultant unswitched loops. + CodeMetrics Metrics; + for (Loop::block_iterator I = L->block_begin(), + E = L->block_end(); + I != E; ++I) + Metrics.analyzeBasicBlock(*I); + + Props.SizeEstimation = std::min(Metrics.NumInsts, Metrics.NumBlocks * 5); + Props.CanBeUnswitchedCount = MaxSize / (Props.SizeEstimation); + MaxSize -= Props.SizeEstimation * Props.CanBeUnswitchedCount; + } + + if (!Props.CanBeUnswitchedCount) { + DEBUG(dbgs() << "NOT unswitching loop %" + << L->getHeader()->getName() << ", cost too high: " + << L->getBlocks().size() << "\n"); + + return false; + } + + // Be careful. This links are good only before new loop addition. + CurrentLoopProperties = &Props; + CurLoopInstructions = &Props.UnswitchedVals; + + return true; +} + +// Clean all data related to given loop. +void LUAnalysisCache::forgetLoop(const Loop* L) { + + LoopPropsMapIt LIt = LoopsProperties.find(L); + + if (LIt != LoopsProperties.end()) { + LoopProperties& Props = LIt->second; + MaxSize += Props.CanBeUnswitchedCount * Props.SizeEstimation; + LoopsProperties.erase(LIt); + } + + CurrentLoopProperties = NULL; + CurLoopInstructions = NULL; +} + +// Mark case value as unswitched. +// Since SI instruction can be partly unswitched, in order to avoid +// extra unswitching in cloned loops keep track all unswitched values. +void LUAnalysisCache::setUnswitched(const SwitchInst* SI, const Value* V) { + + (*CurLoopInstructions)[SI].insert(V); +} + +// Check was this case value unswitched before or not. +bool LUAnalysisCache::isUnswitched(const SwitchInst* SI, const Value* V) { + return (*CurLoopInstructions)[SI].count(V); +} + +// Clone all loop-unswitch related loop properties. +// Redistribute unswitching quotas. +// Note, that new loop data is stored inside the VMap. +void LUAnalysisCache::cloneData(const Loop* NewLoop, const Loop* OldLoop, + const ValueToValueMapTy& VMap) { + + LoopProperties& NewLoopProps = LoopsProperties[NewLoop]; + LoopProperties& OldLoopProps = *CurrentLoopProperties; + UnswitchedValsMap& Insts = OldLoopProps.UnswitchedVals; + + // Reallocate "can-be-unswitched quota" + + --OldLoopProps.CanBeUnswitchedCount; + unsigned Quota = OldLoopProps.CanBeUnswitchedCount; + NewLoopProps.CanBeUnswitchedCount = Quota / 2; + OldLoopProps.CanBeUnswitchedCount = Quota - Quota / 2; + + NewLoopProps.SizeEstimation = OldLoopProps.SizeEstimation; + + // Clone unswitched values info: + // for new loop switches we clone info about values that was + // already unswitched and has redundant successors. + for (UnswitchedValsIt I = Insts.begin(); I != Insts.end(); ++I) { + const SwitchInst* OldInst = I->first; + Value* NewI = VMap.lookup(OldInst); + const SwitchInst* NewInst = cast_or_null(NewI); + assert(NewInst && "All instructions that are in SrcBB must be in VMap."); + + NewLoopProps.UnswitchedVals[NewInst] = OldLoopProps.UnswitchedVals[OldInst]; + } +} + char LoopUnswitch::ID = 0; -static RegisterPass X("loop-unswitch", "Unswitch loops"); +INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops", + false, false) +INITIALIZE_PASS_DEPENDENCY(LoopSimplify) +INITIALIZE_PASS_DEPENDENCY(LoopInfo) +INITIALIZE_PASS_DEPENDENCY(LCSSA) +INITIALIZE_PASS_END(LoopUnswitch, "loop-unswitch", "Unswitch loops", + false, false) Pass *llvm::createLoopUnswitchPass(bool Os) { return new LoopUnswitch(Os); @@ -169,6 +351,10 @@ Pass *llvm::createLoopUnswitchPass(bool Os) { /// invariant in the loop, or has an invariant piece, return the invariant. /// Otherwise, return null. static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { + + // We started analyze new instruction, increment scanned instructions counter. + ++TotalInsts; + // We can never unswitch on vector conditions. if (Cond->getType()->isVectorTy()) return 0; @@ -200,7 +386,6 @@ static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { LI = &getAnalysis(); LPM = &LPM_Ref; - DF = getAnalysisIfAvailable(); DT = getAnalysisIfAvailable(); currentLoop = L; Function *F = currentLoop->getHeader()->getParent(); @@ -215,8 +400,6 @@ bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { // FIXME: Reconstruct dom info, because it is not preserved properly. if (DT) DT->runOnFunction(*F); - if (DF) - DF->runOnFunction(*F); } return Changed; } @@ -225,7 +408,19 @@ bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { /// and profitable. bool LoopUnswitch::processCurrentLoop() { bool Changed = false; - LLVMContext &Context = currentLoop->getHeader()->getContext(); + + initLoopData(); + + // If LoopSimplify was unable to form a preheader, don't do any unswitching. + if (!loopPreheader) + return false; + + LLVMContext &Context = loopHeader->getContext(); + + // Probably we reach the quota of branches for this loop. If so + // stop unswitching. + if (!BranchesInfo.countLoop(currentLoop)) + return 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 @@ -250,12 +445,25 @@ bool LoopUnswitch::processCurrentLoop() { } else if (SwitchInst *SI = dyn_cast(TI)) { Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), currentLoop, Changed); - if (LoopCond && SI->getNumCases() > 1) { + unsigned NumCases = SI->getNumCases(); + if (LoopCond && NumCases > 1) { // Find a value to unswitch on: // FIXME: this should chose the most expensive case! - Constant *UnswitchVal = SI->getCaseValue(1); + // FIXME: scan for a case with a non-critical edge? + Constant *UnswitchVal = NULL; + // Do not process same value again and again. - if (!UnswitchedVals.insert(UnswitchVal)) + // At this point we have some cases already unswitched and + // some not yet unswitched. Let's find the first not yet unswitched one. + for (unsigned i = 1; i < NumCases; ++i) { + Constant* UnswitchValCandidate = SI->getCaseValue(i); + if (!BranchesInfo.isUnswitched(SI, UnswitchValCandidate)) { + UnswitchVal = UnswitchValCandidate; + break; + } + } + + if (!UnswitchVal) continue; if (UnswitchIfProfitable(LoopCond, UnswitchVal)) { @@ -281,19 +489,19 @@ bool LoopUnswitch::processCurrentLoop() { return Changed; } -/// isTrivialLoopExitBlock - Check to see if all paths from BB either: -/// 1. Exit the loop with no side effects. -/// 2. Branch to the latch block with no side-effects. +/// isTrivialLoopExitBlock - Check to see if all paths from BB exit the +/// loop with no side effects (including infinite loops). /// -/// If these conditions are true, we return true and set ExitBB to the block we +/// If true, we return true and set ExitBB to the block we /// exit through. /// static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB, BasicBlock *&ExitBB, std::set &Visited) { if (!Visited.insert(BB).second) { - // Already visited and Ok, end of recursion. - return true; + // Already visited. Without more analysis, this could indicate an infinite + // loop. + return false; } else if (!L->contains(BB)) { // Otherwise, this is a loop exit, this is fine so long as this is the // first exit. @@ -323,7 +531,7 @@ static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB, /// process. If so, return the block that is exited to, otherwise return null. static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) { std::set Visited; - Visited.insert(L->getHeader()); // Branches to header are ok. + Visited.insert(L->getHeader()); // Branches to header make infinite loops. BasicBlock *ExitBB = 0; if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited)) return ExitBB; @@ -355,8 +563,8 @@ bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val, if (!BI->isConditional() || BI->getCondition() != Cond) return false; - // Check to see if a successor of the branch is guaranteed to go to the - // latch block or exit through a one exit block without having any + // Check to see if a successor of the branch is guaranteed to + // exit through a unique exit block without having any // side-effects. If so, determine the value of Cond that causes it to do // this. if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, @@ -373,14 +581,25 @@ bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val, // Check to see if a successor of the switch is guaranteed to go to the // latch block or exit through a one exit block without having any // side-effects. If so, determine the value of Cond that causes it to do - // this. Note that we can't trivially unswitch on the default case. - for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) - if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, + // this. + // Note that we can't trivially unswitch on the default case or + // on already unswitched cases. + for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i) { + BasicBlock* LoopExitCandidate; + if ((LoopExitCandidate = isTrivialLoopExitBlock(currentLoop, SI->getSuccessor(i)))) { // Okay, we found a trivial case, remember the value that is trivial. - if (Val) *Val = SI->getCaseValue(i); + ConstantInt* CaseVal = SI->getCaseValue(i); + + // Check that it was not unswitched before, since already unswitched + // trivial vals are looks trivial too. + if (BranchesInfo.isUnswitched(SI, CaseVal)) + continue; + LoopExitBB = LoopExitCandidate; + if (Val) *Val = CaseVal; break; } + } } // If we didn't find a single unique LoopExit block, or if the loop exit block @@ -406,12 +625,6 @@ bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val, /// unswitch the loop, reprocess the pieces, then return true. bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) { - initLoopData(); - - // If LoopSimplify was unable to form a preheader, don't do any unswitching. - if (!loopPreheader) - return false; - Function *F = loopHeader->getParent(); Constant *CondVal = 0; @@ -429,48 +642,13 @@ bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) { if (OptimizeForSize || F->hasFnAttr(Attribute::OptimizeForSize)) return false; - // FIXME: This is overly conservative because it does not take into - // consideration code simplification opportunities and code that can - // be shared by the resultant unswitched loops. - CodeMetrics Metrics; - for (Loop::block_iterator I = currentLoop->block_begin(), - E = currentLoop->block_end(); - I != E; ++I) - Metrics.analyzeBasicBlock(*I); - - // Limit the number of instructions to avoid causing significant code - // expansion, and the number of basic blocks, to avoid loops with - // large numbers of branches which cause loop unswitching to go crazy. - // This is a very ad-hoc heuristic. - if (Metrics.NumInsts > Threshold || - Metrics.NumBlocks * 5 > Threshold || - Metrics.NeverInline) { - DEBUG(dbgs() << "NOT unswitching loop %" - << currentLoop->getHeader()->getName() << ", cost too high: " - << currentLoop->getBlocks().size() << "\n"); - return false; - } - UnswitchNontrivialCondition(LoopCond, Val, currentLoop); return true; } -// RemapInstruction - Convert the instruction operands from referencing the -// current values into those specified by ValueMap. -// -static inline void RemapInstruction(Instruction *I, - DenseMap &ValueMap) { - for (unsigned op = 0, E = I->getNumOperands(); op != E; ++op) { - Value *Op = I->getOperand(op); - DenseMap::iterator It = ValueMap.find(Op); - if (It != ValueMap.end()) Op = It->second; - I->setOperand(op, Op); - } -} - /// 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, DenseMap &VM, +static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM, LoopInfo *LI, LPPassManager *LPM) { Loop *New = new Loop(); LPM->insertLoop(New, PL); @@ -500,7 +678,7 @@ void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, Value *BranchVal = LIC; if (!isa(Val) || Val->getType() != Type::getInt1Ty(LIC->getContext())) - BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val, "tmp"); + BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val); else if (Val != ConstantInt::getTrue(Val->getContext())) // We want to enter the new loop when the condition is true. std::swap(TrueDest, FalseDest); @@ -569,8 +747,16 @@ void LoopUnswitch::SplitExitEdges(Loop *L, BasicBlock *ExitBlock = ExitBlocks[i]; SmallVector Preds(pred_begin(ExitBlock), pred_end(ExitBlock)); - SplitBlockPredecessors(ExitBlock, Preds.data(), Preds.size(), - ".us-lcssa", this); + + // Although SplitBlockPredecessors doesn't preserve loop-simplify in + // general, if we call it on all predecessors of all exits then it does. + if (!ExitBlock->isLandingPad()) { + SplitBlockPredecessors(ExitBlock, Preds, ".us-lcssa", this); + } else { + SmallVector NewBBs; + SplitLandingPadPredecessors(ExitBlock, Preds, ".us-lcssa", ".us-lcssa", + this, NewBBs); + } } } @@ -585,6 +771,9 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, << " blocks] in Function " << F->getName() << " when '" << *Val << "' == " << *LIC << "\n"); + if (ScalarEvolution *SE = getAnalysisIfAvailable()) + SE->forgetLoop(L); + LoopBlocks.clear(); NewBlocks.clear(); @@ -614,11 +803,12 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, // the loop preheader and exit blocks), keeping track of the mapping between // the instructions and blocks. NewBlocks.reserve(LoopBlocks.size()); - DenseMap ValueMap; + ValueToValueMapTy VMap; for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { - BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], ValueMap, ".us", F); + BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F); + NewBlocks.push_back(NewBB); - ValueMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping. + VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping. LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L); } @@ -628,16 +818,21 @@ 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, LPM); + Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM); + + // Recalculate unswitching quota, inherit simplified switches info for NewBB, + // Probably clone more loop-unswitch related loop properties. + BranchesInfo.cloneData(NewLoop, L, VMap); + 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->getBase()); } - + for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { - BasicBlock *NewExit = cast(ValueMap[ExitBlocks[i]]); + BasicBlock *NewExit = cast(VMap[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->getBase()); @@ -652,17 +847,30 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, for (BasicBlock::iterator I = ExitSucc->begin(); isa(I); ++I) { PN = cast(I); Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); - DenseMap::iterator It = ValueMap.find(V); - if (It != ValueMap.end()) V = It->second; + ValueToValueMapTy::iterator It = VMap.find(V); + if (It != VMap.end()) V = It->second; PN->addIncoming(V, NewExit); } + + if (LandingPadInst *LPad = NewExit->getLandingPadInst()) { + PN = PHINode::Create(LPad->getType(), 0, "", + ExitSucc->getFirstInsertionPt()); + + for (pred_iterator I = pred_begin(ExitSucc), E = pred_end(ExitSucc); + I != E; ++I) { + BasicBlock *BB = *I; + LandingPadInst *LPI = BB->getLandingPadInst(); + LPI->replaceAllUsesWith(PN); + PN->addIncoming(LPI, BB); + } + } } // Rewrite the code to refer to itself. for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i) for (BasicBlock::iterator I = NewBlocks[i]->begin(), E = NewBlocks[i]->end(); I != E; ++I) - RemapInstruction(I, ValueMap); + RemapInstruction(I, VMap,RF_NoModuleLevelChanges|RF_IgnoreMissingEntries); // Rewrite the original preheader to select between versions of the loop. BranchInst *OldBR = cast(loopPreheader->getTerminator()); @@ -792,8 +1000,13 @@ void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, // If this is the edge to the header block for a loop, remove the loop and // promote all subloops. if (Loop *BBLoop = LI->getLoopFor(BB)) { - if (BBLoop->getLoopLatch() == BB) + if (BBLoop->getLoopLatch() == BB) { RemoveLoopFromHierarchy(BBLoop); + if (currentLoop == BBLoop) { + currentLoop = 0; + redoLoop = false; + } + } } // Remove the block from the loop info, which removes it from any loops it @@ -865,7 +1078,6 @@ void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches, // selects, switches. - std::vector Users(LIC->use_begin(), LIC->use_end()); std::vector Worklist; LLVMContext &Context = Val->getContext(); @@ -881,13 +1093,18 @@ void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()), !cast(Val)->getZExtValue()); - for (unsigned i = 0, e = Users.size(); i != e; ++i) - if (Instruction *U = cast(Users[i])) { - if (!L->contains(U)) - continue; - U->replaceUsesOfWith(LIC, Replacement); - Worklist.push_back(U); - } + for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end(); + UI != E; ++UI) { + Instruction *U = dyn_cast(*UI); + if (!U || !L->contains(U)) + continue; + Worklist.push_back(U); + } + + for (std::vector::iterator UI = Worklist.begin(); + UI != Worklist.end(); ++UI) + (*UI)->replaceUsesOfWith(LIC, Replacement); + SimplifyCode(Worklist, L); return; } @@ -895,9 +1112,10 @@ void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, // Otherwise, we don't know the precise value of LIC, but we do know that it // is certainly NOT "Val". As such, simplify any uses in the loop that we // can. This case occurs when we unswitch switch statements. - for (unsigned i = 0, e = Users.size(); i != e; ++i) { - Instruction *U = cast(Users[i]); - if (!L->contains(U)) + for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end(); + UI != E; ++UI) { + Instruction *U = dyn_cast(*UI); + if (!U || !L->contains(U)) continue; Worklist.push_back(U); @@ -915,13 +1133,25 @@ void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, // Found a dead case value. Don't remove PHI nodes in the // successor if they become single-entry, those PHI nodes may // be in the Users list. - + + BasicBlock *Switch = SI->getParent(); + BasicBlock *SISucc = SI->getSuccessor(DeadCase); + BasicBlock *Latch = L->getLoopLatch(); + + BranchesInfo.setUnswitched(SI, Val); + + if (!SI->findCaseDest(SISucc)) continue; // Edge is critical. + // If the DeadCase successor dominates the loop latch, then the + // transformation isn't safe since it will delete the sole predecessor edge + // to the latch. + if (Latch && DT->dominates(SISucc, Latch)) + continue; + // FIXME: This is a hack. We need to keep the successor around // and hooked up so as to preserve the loop structure, because // trying to update it is complicated. So instead we preserve the // loop structure and put the block on a dead code path. - BasicBlock *Switch = SI->getParent(); - SplitEdge(Switch, SI->getSuccessor(DeadCase), this); + SplitEdge(Switch, SISucc, this); // Compute the successors instead of relying on the return value // of SplitEdge, since it may have split the switch successor // after PHI nodes. @@ -966,13 +1196,7 @@ 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, L, LPM); - continue; - } - + // Simple DCE. if (isInstructionTriviallyDead(I)) { DEBUG(dbgs() << "Remove dead instruction '" << *I); @@ -987,48 +1211,18 @@ void LoopUnswitch::SimplifyCode(std::vector &Worklist, Loop *L) { ++NumSimplify; continue; } - - // FIXME: Change this to use instruction simplify interfaces! - - // Special case hacks that appear commonly in unswitched code. - switch (I->getOpcode()) { - case Instruction::Select: - if (ConstantInt *CB = dyn_cast(I->getOperand(0))) { - ReplaceUsesOfWith(I, I->getOperand(!CB->getZExtValue()+1), Worklist, L, - LPM); + + // See if instruction simplification can hack this up. This is common for + // things like "select false, X, Y" after unswitching made the condition be + // 'false'. + if (Value *V = SimplifyInstruction(I, 0, 0, DT)) + if (LI->replacementPreservesLCSSAForm(I, V)) { + ReplaceUsesOfWith(I, V, Worklist, L, LPM); continue; } - break; - case Instruction::And: - if (isa(I->getOperand(0)) && - // constant -> RHS - I->getOperand(0)->getType()->isIntegerTy(1)) - cast(I)->swapOperands(); - if (ConstantInt *CB = dyn_cast(I->getOperand(1))) - if (CB->getType()->isIntegerTy(1)) { - 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 - I->getOperand(0)->getType()->isIntegerTy(1)) - cast(I)->swapOperands(); - if (ConstantInt *CB = dyn_cast(I->getOperand(1))) - if (CB->getType()->isIntegerTy(1)) { - 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); + + // Special case hacks that appear commonly in unswitched code. + if (BranchInst *BI = dyn_cast(I)) { if (BI->isUnconditional()) { // If BI's parent is the only pred of the successor, fold the two blocks // together. @@ -1045,6 +1239,10 @@ void LoopUnswitch::SimplifyCode(std::vector &Worklist, Loop *L) { while (PHINode *PN = dyn_cast(Succ->begin())) ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM); + // If Succ has any successors with PHI nodes, update them to have + // entries coming from Pred instead of Succ. + Succ->replaceAllUsesWith(Pred); + // Move all of the successor contents from Succ to Pred. Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(), Succ->end()); @@ -1052,22 +1250,18 @@ void LoopUnswitch::SimplifyCode(std::vector &Worklist, Loop *L) { BI->eraseFromParent(); RemoveFromWorklist(BI, Worklist); - // If Succ has any successors with PHI nodes, update them to have - // entries coming from Pred instead of Succ. - Succ->replaceAllUsesWith(Pred); - // Remove Succ from the loop tree. LI->removeBlock(Succ); LPM->deleteSimpleAnalysisValue(Succ, L); Succ->eraseFromParent(); ++NumSimplify; - break; + continue; } if (ConstantInt *CB = dyn_cast(BI->getCondition())){ // Conditional branch. Turn it into an unconditional branch, then // remove dead blocks. - break; // FIXME: Enable. + continue; // FIXME: Enable. DEBUG(dbgs() << "Folded branch: " << *BI); BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue()); @@ -1081,8 +1275,7 @@ void LoopUnswitch::SimplifyCode(std::vector &Worklist, Loop *L) { RemoveBlockIfDead(DeadSucc, Worklist, L); } - break; - } + continue; } } }