X-Git-Url: http://plrg.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FScalar%2FLoopUnswitch.cpp;h=95d7f8a3beda23b180d350f3835f2e2c187e639e;hb=2a949077a97c3bcc57be2507ad8c9b5035f14a3d;hp=458949c8444d2a20a9d753af8df8e1eddbd8fe8d;hpb=a9390a4d5f5d568059a80970d22194b165d097a7;p=oota-llvm.git diff --git a/lib/Transforms/Scalar/LoopUnswitch.cpp b/lib/Transforms/Scalar/LoopUnswitch.cpp index 458949c8444..95d7f8a3bed 100644 --- a/lib/Transforms/Scalar/LoopUnswitch.cpp +++ b/lib/Transforms/Scalar/LoopUnswitch.cpp @@ -26,53 +26,159 @@ // //===----------------------------------------------------------------------===// -#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/InlineCost.h" +#include "llvm/ADT/STLExtras.h" +#include "llvm/ADT/SmallPtrSet.h" +#include "llvm/ADT/Statistic.h" +#include "llvm/Analysis/GlobalsModRef.h" +#include "llvm/Analysis/AssumptionCache.h" +#include "llvm/Analysis/CodeMetrics.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" -#include "llvm/ADT/Statistic.h" -#include "llvm/ADT/SmallPtrSet.h" -#include "llvm/ADT/STLExtras.h" +#include "llvm/Analysis/TargetTransformInfo.h" +#include "llvm/Analysis/BlockFrequencyInfoImpl.h" +#include "llvm/Analysis/BlockFrequencyInfo.h" +#include "llvm/Analysis/BranchProbabilityInfo.h" +#include "llvm/Support/BranchProbability.h" +#include "llvm/IR/Constants.h" +#include "llvm/IR/DerivedTypes.h" +#include "llvm/IR/Dominators.h" +#include "llvm/IR/Function.h" +#include "llvm/IR/Instructions.h" +#include "llvm/IR/Module.h" +#include "llvm/IR/MDBuilder.h" #include "llvm/Support/CommandLine.h" #include "llvm/Support/Debug.h" #include "llvm/Support/raw_ostream.h" +#include "llvm/Transforms/Utils/BasicBlockUtils.h" +#include "llvm/Transforms/Utils/Cloning.h" +#include "llvm/Transforms/Utils/Local.h" #include +#include #include using namespace llvm; +#define DEBUG_TYPE "loop-unswitch" + 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"); +STATISTIC(TotalInsts, "Total number of instructions analyzed"); -// The specific value of 50 here was chosen based only on intuition and a +// The specific value of 100 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); + +static cl::opt +LoopUnswitchWithBlockFrequency("loop-unswitch-with-block-frequency", + cl::init(false), cl::Hidden, + cl::desc("Enable the use of the block frequency analysis to access PGO " + "heuristics to minimize code growth in cold regions.")); + +static cl::opt +ColdnessThreshold("loop-unswitch-coldness-threshold", cl::init(1), cl::Hidden, + cl::desc("Coldness threshold in percentage. The loop header frequency " + "(relative to the entry frequency) is compared with this " + "threshold to determine if non-trivial unswitching should be " + "enabled.")); + namespace { + + class LUAnalysisCache { + + typedef DenseMap > + UnswitchedValsMap; + + typedef UnswitchedValsMap::iterator UnswitchedValsIt; + + struct LoopProperties { + unsigned CanBeUnswitchedCount; + unsigned WasUnswitchedCount; + unsigned SizeEstimation; + UnswitchedValsMap UnswitchedVals; + }; + + // Here we use std::map instead of DenseMap, since we need to keep valid + // LoopProperties pointer for current loop for better performance. + typedef std::map LoopPropsMap; + typedef LoopPropsMap::iterator LoopPropsMapIt; + + LoopPropsMap LoopsProperties; + UnswitchedValsMap *CurLoopInstructions; + LoopProperties *CurrentLoopProperties; + + // A loop unswitching with an estimated cost above this threshold + // is not performed. MaxSize is turned into unswitching quota for + // the current loop, and reduced correspondingly, though note that + // the quota is returned by releaseMemory() when the loop has been + // processed, so that MaxSize will return to its previous + // value. So in most cases MaxSize will equal the Threshold flag + // when a new loop is processed. An exception to that is that + // MaxSize will have a smaller value while processing nested loops + // that were introduced due to loop unswitching of an outer loop. + // + // FIXME: The way that MaxSize works is subtle and depends on the + // pass manager processing loops and calling releaseMemory() in a + // specific order. It would be good to find a more straightforward + // way of doing what MaxSize does. + unsigned MaxSize; + + public: + LUAnalysisCache() + : CurLoopInstructions(nullptr), CurrentLoopProperties(nullptr), + 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, const TargetTransformInfo &TTI, + AssumptionCache *AC); + + // 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); + + // Returns true if another unswitching could be done within the cost + // threshold. + bool CostAllowsUnswitching(); + + // 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; + AssumptionCache *AC; - // LoopProcessWorklist - Used to check if second loop needs processing - // after RewriteLoopBodyWithConditionConstant rewrites first loop. + // Used to check if second loop needs processing after + // RewriteLoopBodyWithConditionConstant rewrites first loop. std::vector LoopProcessWorklist; - SmallPtrSet UnswitchedVals; - + + LUAnalysisCache BranchesInfo; + + bool EnabledPGO; + + // BFI and ColdEntryFreq are only used when PGO and + // LoopUnswitchWithBlockFrequency are enabled. + BlockFrequencyInfo BFI; + BlockFrequency ColdEntryFreq; + bool OptimizeForSize; bool redoLoop; @@ -80,9 +186,9 @@ namespace { DominatorTree *DT; BasicBlock *loopHeader; BasicBlock *loopPreheader; - + // 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 + // preheader of the loop, the body of the loop, and the exit blocks of the // loop, in that order. std::vector LoopBlocks; // NewBlocks contained cloned copy of basic blocks from LoopBlocks. @@ -90,43 +196,38 @@ namespace { public: static char ID; // Pass ID, replacement for typeid - explicit LoopUnswitch(bool Os = false) : - LoopPass(ID), OptimizeForSize(Os), redoLoop(false), - currentLoop(NULL), DT(NULL), loopHeader(NULL), - loopPreheader(NULL) { + explicit LoopUnswitch(bool Os = false) : + LoopPass(ID), OptimizeForSize(Os), redoLoop(false), + currentLoop(nullptr), DT(nullptr), loopHeader(nullptr), + loopPreheader(nullptr) { initializeLoopUnswitchPass(*PassRegistry::getPassRegistry()); } - bool runOnLoop(Loop *L, LPPassManager &LPM); + bool runOnLoop(Loop *L, LPPassManager &LPM) override; bool processCurrentLoop(); /// This transformation requires natural loop information & requires that /// loop preheaders be inserted into the CFG. /// - virtual void getAnalysisUsage(AnalysisUsage &AU) const { + void getAnalysisUsage(AnalysisUsage &AU) const override { + AU.addRequired(); AU.addRequiredID(LoopSimplifyID); AU.addPreservedID(LoopSimplifyID); - AU.addRequired(); - AU.addPreserved(); + AU.addRequired(); + AU.addPreserved(); AU.addRequiredID(LCSSAID); AU.addPreservedID(LCSSAID); - AU.addPreserved(); - AU.addPreserved(); + AU.addRequired(); + AU.addPreserved(); + AU.addPreserved(); + AU.addRequired(); + AU.addPreserved(); } private: - virtual void releaseMemory() { - UnswitchedVals.clear(); - } - - /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist, - /// remove it. - void RemoveLoopFromWorklist(Loop *L) { - std::vector::iterator I = std::find(LoopProcessWorklist.begin(), - LoopProcessWorklist.end(), L); - if (I != LoopProcessWorklist.end()) - LoopProcessWorklist.erase(I); + void releaseMemory() override { + BranchesInfo.forgetLoop(currentLoop); } void initLoopData() { @@ -136,53 +237,176 @@ namespace { /// 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); + void SplitExitEdges(Loop *L, + const SmallVectorImpl &ExitBlocks); - bool UnswitchIfProfitable(Value *LoopCond, Constant *Val); + bool TryTrivialLoopUnswitch(bool &Changed); + + bool UnswitchIfProfitable(Value *LoopCond, Constant *Val, + TerminatorInst *TI = nullptr); void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val, - BasicBlock *ExitBlock); - void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L); + BasicBlock *ExitBlock, TerminatorInst *TI); + void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L, + TerminatorInst *TI); void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, Constant *Val, bool isEqual); void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, - BasicBlock *TrueDest, + BasicBlock *TrueDest, BasicBlock *FalseDest, - Instruction *InsertPt); + Instruction *InsertPt, + TerminatorInst *TI); void SimplifyCode(std::vector &Worklist, Loop *L); - void RemoveBlockIfDead(BasicBlock *BB, - std::vector &Worklist, Loop *l); - void RemoveLoopFromHierarchy(Loop *L); - bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0, - BasicBlock **LoopExit = 0); - }; } + +// 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, const TargetTransformInfo &TTI, + AssumptionCache *AC) { + + LoopPropsMapIt PropsIt; + bool Inserted; + std::tie(PropsIt, Inserted) = + LoopsProperties.insert(std::make_pair(L, LoopProperties())); + + LoopProperties &Props = PropsIt->second; + + if (Inserted) { + // 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. + + SmallPtrSet EphValues; + CodeMetrics::collectEphemeralValues(L, AC, EphValues); + + // 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, TTI, EphValues); + + Props.SizeEstimation = Metrics.NumInsts; + Props.CanBeUnswitchedCount = MaxSize / (Props.SizeEstimation); + Props.WasUnswitchedCount = 0; + MaxSize -= Props.SizeEstimation * Props.CanBeUnswitchedCount; + + if (Metrics.notDuplicatable) { + DEBUG(dbgs() << "NOT unswitching loop %" + << L->getHeader()->getName() << ", contents cannot be " + << "duplicated!\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.WasUnswitchedCount) * + Props.SizeEstimation; + LoopsProperties.erase(LIt); + } + + CurrentLoopProperties = nullptr; + CurLoopInstructions = nullptr; +} + +// 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); +} + +bool LUAnalysisCache::CostAllowsUnswitching() { + return CurrentLoopProperties->CanBeUnswitchedCount > 0; +} + +// 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; + ++OldLoopProps.WasUnswitchedCount; + NewLoopProps.WasUnswitchedCount = 0; + 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; INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops", false, false) +INITIALIZE_PASS_DEPENDENCY(TargetTransformInfoWrapperPass) +INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker) INITIALIZE_PASS_DEPENDENCY(LoopSimplify) -INITIALIZE_PASS_DEPENDENCY(LoopInfo) +INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass) INITIALIZE_PASS_DEPENDENCY(LCSSA) INITIALIZE_PASS_END(LoopUnswitch, "loop-unswitch", "Unswitch loops", false, false) -Pass *llvm::createLoopUnswitchPass(bool Os) { - return new LoopUnswitch(Os); +Pass *llvm::createLoopUnswitchPass(bool Os) { + return new LoopUnswitch(Os); } -/// FindLIVLoopCondition - Cond is a condition that occurs in L. If it is -/// invariant in the loop, or has an invariant piece, return the invariant. -/// Otherwise, return null. +/// Cond is a condition that occurs in L. If it is 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; + return nullptr; // Constants should be folded, not unswitched on! - if (isa(Cond)) return 0; + if (isa(Cond)) return nullptr; // TODO: Handle: br (VARIANT|INVARIANT). @@ -201,16 +425,35 @@ static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed)) return RHS; } - - return 0; + + return nullptr; } bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { - LI = &getAnalysis(); + if (skipOptnoneFunction(L)) + return false; + + AC = &getAnalysis().getAssumptionCache( + *L->getHeader()->getParent()); + LI = &getAnalysis().getLoopInfo(); LPM = &LPM_Ref; - DT = getAnalysisIfAvailable(); + DT = &getAnalysis().getDomTree(); currentLoop = L; Function *F = currentLoop->getHeader()->getParent(); + + EnabledPGO = F->getEntryCount().hasValue(); + + if (LoopUnswitchWithBlockFrequency && EnabledPGO) { + BranchProbabilityInfo BPI(*F, *LI); + BFI.calculate(*L->getHeader()->getParent(), BPI, *LI); + + // Use BranchProbability to compute a minimum frequency based on + // function entry baseline frequency. Loops with headers below this + // frequency are considered as cold. + const BranchProbability ColdProb(ColdnessThreshold, 100); + ColdEntryFreq = BlockFrequency(BFI.getEntryFreq()) * ColdProb; + } + bool Changed = false; do { assert(currentLoop->isLCSSAForm(*DT)); @@ -218,24 +461,78 @@ bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { Changed |= processCurrentLoop(); } while(redoLoop); - if (Changed) { - // FIXME: Reconstruct dom info, because it is not preserved properly. - if (DT) - DT->runOnFunction(*F); - } + // FIXME: Reconstruct dom info, because it is not preserved properly. + if (Changed) + DT->recalculate(*F); return Changed; } -/// processCurrentLoop - Do actual work and unswitch loop if possible -/// and profitable. +/// Do actual work and unswitch loop if possible 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; + + // Loops with indirectbr cannot be cloned. + if (!currentLoop->isSafeToClone()) + return false; + + // Without dedicated exits, splitting the exit edge may fail. + if (!currentLoop->hasDedicatedExits()) + return false; + + LLVMContext &Context = loopHeader->getContext(); + + // Analyze loop cost, and stop unswitching if loop content can not be duplicated. + if (!BranchesInfo.countLoop( + currentLoop, getAnalysis().getTTI( + *currentLoop->getHeader()->getParent()), + AC)) + return false; + + // Try trivial unswitch first before loop over other basic blocks in the loop. + if (TryTrivialLoopUnswitch(Changed)) { + return true; + } + + // Do not unswitch loops containing convergent operations, as we might be + // making them control dependent on the unswitch value when they were not + // before. + // FIXME: This could be refined to only bail if the convergent operation is + // not already control-dependent on the unswitch value. + for (const auto BB : currentLoop->blocks()) { + for (auto &I : *BB) { + auto CS = CallSite(&I); + if (!CS) continue; + if (CS.hasFnAttr(Attribute::Convergent)) + return false; + } + } + + // Do not do non-trivial unswitch while optimizing for size. + // FIXME: Use Function::optForSize(). + if (OptimizeForSize || + loopHeader->getParent()->hasFnAttribute(Attribute::OptimizeForSize)) + return false; + + if (LoopUnswitchWithBlockFrequency && EnabledPGO) { + // Compute the weighted frequency of the hottest block in the + // loop (loopHeader in this case since inner loops should be + // processed before outer loop). If it is less than ColdFrequency, + // we should not unswitch. + BlockFrequency LoopEntryFreq = BFI.getBlockFreq(loopHeader); + if (LoopEntryFreq < ColdEntryFreq) + 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 // loop. - for (Loop::block_iterator I = currentLoop->block_begin(), + for (Loop::block_iterator I = currentLoop->block_begin(), E = currentLoop->block_end(); I != E; ++I) { TerminatorInst *TI = (*I)->getTerminator(); if (BranchInst *BI = dyn_cast(TI)) { @@ -244,24 +541,37 @@ bool LoopUnswitch::processCurrentLoop() { if (BI->isConditional()) { // 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(), + Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), currentLoop, Changed); - if (LoopCond && UnswitchIfProfitable(LoopCond, - ConstantInt::getTrue(Context))) { + if (LoopCond && + UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(Context), TI)) { ++NumBranches; return true; } - } + } } else if (SwitchInst *SI = dyn_cast(TI)) { - Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), + Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), currentLoop, Changed); - if (LoopCond && SI->getNumCases() > 1) { + unsigned NumCases = SI->getNumCases(); + if (LoopCond && NumCases) { // Find a value to unswitch on: // FIXME: this should chose the most expensive case! // FIXME: scan for a case with a non-critical edge? - Constant *UnswitchVal = SI->getCaseValue(1); + Constant *UnswitchVal = nullptr; + // 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 (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); + i != e; ++i) { + Constant *UnswitchValCandidate = i.getCaseValue(); + if (!BranchesInfo.isUnswitched(SI, UnswitchValCandidate)) { + UnswitchVal = UnswitchValCandidate; + break; + } + } + + if (!UnswitchVal) continue; if (UnswitchIfProfitable(LoopCond, UnswitchVal)) { @@ -270,14 +580,14 @@ bool LoopUnswitch::processCurrentLoop() { } } } - + // Scan the instructions to check for unswitchable values. - for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end(); + for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end(); BBI != E; ++BBI) if (SelectInst *SI = dyn_cast(BBI)) { - Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), + Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), currentLoop, Changed); - if (LoopCond && UnswitchIfProfitable(LoopCond, + if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(Context))) { ++NumSelects; return true; @@ -287,8 +597,8 @@ bool LoopUnswitch::processCurrentLoop() { return Changed; } -/// isTrivialLoopExitBlock - Check to see if all paths from BB exit the -/// loop with no side effects (including infinite loops). +/// Check to see if all paths from BB exit the loop with no side effects +/// (including infinite loops). /// /// If true, we return true and set ExitBB to the block we /// exit through. @@ -297,16 +607,18 @@ static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB, BasicBlock *&ExitBB, std::set &Visited) { if (!Visited.insert(BB).second) { - // Already visited. Without more analysis, this could indicate an infinte loop. + // Already visited. Without more analysis, this could indicate an infinite + // loop. return false; - } else if (!L->contains(BB)) { + } + if (!L->contains(BB)) { // Otherwise, this is a loop exit, this is fine so long as this is the // first exit. - if (ExitBB != 0) return false; + if (ExitBB) return false; ExitBB = BB; return true; } - + // Otherwise, this is an unvisited intra-loop node. Check all successors. for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) { // Check to see if the successor is a trivial loop exit. @@ -319,232 +631,167 @@ static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB, for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) if (I->mayHaveSideEffects()) return false; - + return true; } -/// isTrivialLoopExitBlock - Return true if the specified block unconditionally -/// leads to an exit from the specified loop, and has no side-effects in the -/// process. If so, return the block that is exited to, otherwise return null. +/// Return true if the specified block unconditionally leads to an exit from +/// the specified loop, and has no side-effects in the 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 make infinite loops. - BasicBlock *ExitBB = 0; + BasicBlock *ExitBB = nullptr; if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited)) return ExitBB; - return 0; -} - -/// IsTrivialUnswitchCondition - Check to see if this unswitch condition is -/// trivial: that is, that the condition controls whether or not the loop does -/// anything at all. If this is a trivial condition, unswitching produces no -/// code duplications (equivalently, it produces a simpler loop and a new empty -/// loop, which gets deleted). -/// -/// If this is a trivial condition, return true, otherwise return false. When -/// returning true, this sets Cond and Val to the condition that controls the -/// trivial condition: when Cond dynamically equals Val, the loop is known to -/// exit. Finally, this sets LoopExit to the BB that the loop exits to when -/// Cond == Val. -/// -bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val, - BasicBlock **LoopExit) { - BasicBlock *Header = currentLoop->getHeader(); - TerminatorInst *HeaderTerm = Header->getTerminator(); - LLVMContext &Context = Header->getContext(); - - BasicBlock *LoopExitBB = 0; - if (BranchInst *BI = dyn_cast(HeaderTerm)) { - // If the header block doesn't end with a conditional branch on Cond, we - // can't handle it. - if (!BI->isConditional() || BI->getCondition() != Cond) - return false; - - // 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, - BI->getSuccessor(0)))) { - if (Val) *Val = ConstantInt::getTrue(Context); - } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, - BI->getSuccessor(1)))) { - if (Val) *Val = ConstantInt::getFalse(Context); - } - } else if (SwitchInst *SI = dyn_cast(HeaderTerm)) { - // If this isn't a switch on Cond, we can't handle it. - if (SI->getCondition() != Cond) return false; - - // 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, - SI->getSuccessor(i)))) { - // Okay, we found a trivial case, remember the value that is trivial. - if (Val) *Val = SI->getCaseValue(i); - break; - } - } - - // If we didn't find a single unique LoopExit block, or if the loop exit block - // contains phi nodes, this isn't trivial. - if (!LoopExitBB || isa(LoopExitBB->begin())) - return false; // Can't handle this. - - if (LoopExit) *LoopExit = LoopExitBB; - - // We already know that nothing uses any scalar values defined inside of this - // loop. As such, we just have to check to see if this loop will execute any - // side-effecting instructions (e.g. stores, calls, volatile loads) in the - // part of the loop that the code *would* execute. We already checked the - // tail, check the header now. - for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I) - if (I->mayHaveSideEffects()) - return false; - return true; + return nullptr; } -/// UnswitchIfProfitable - We have found that we can unswitch currentLoop when -/// LoopCond == Val to simplify the loop. If we decide that this is profitable, +/// We have found that we can unswitch currentLoop when LoopCond == Val to +/// simplify the loop. If we decide that this is profitable, /// 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; - BasicBlock *ExitBlock = 0; - if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) { - // If the condition is trivial, always unswitch. There is no code growth - // for this case. - UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock); - return true; - } - +bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val, + TerminatorInst *TI) { // Check to see if it would be profitable to unswitch current loop. - - // Do not do non-trivial unswitch while optimizing for size. - 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.containsIndirectBr || Metrics.isRecursive) { + if (!BranchesInfo.CostAllowsUnswitching()) { DEBUG(dbgs() << "NOT unswitching loop %" - << currentLoop->getHeader()->getName() << ", cost too high: " - << currentLoop->getBlocks().size() << "\n"); + << currentLoop->getHeader()->getName() + << " at non-trivial condition '" << *Val + << "' == " << *LoopCond << "\n" + << ". Cost too high.\n"); return false; } - UnswitchNontrivialCondition(LoopCond, Val, currentLoop); + UnswitchNontrivialCondition(LoopCond, Val, currentLoop, TI); return true; } -/// CloneLoop - Recursively clone the specified loop and all of its children, +/// Recursively clone the specified loop and all of its children, /// mapping the blocks with the specified map. static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM, LoopInfo *LI, LPPassManager *LPM) { - Loop *New = new Loop(); - LPM->insertLoop(New, PL); + Loop &New = LPM->addLoop(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->getBase()); + New.addBasicBlockToLoop(cast(VM[*I]), *LI); // 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, LPM); + CloneLoop(*I, &New, VM, LI, LPM); - return New; + return &New; } -/// 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 copyMetadata(Instruction *DstInst, const Instruction *SrcInst, + bool Swapped) { + if (!SrcInst || !SrcInst->hasMetadata()) + return; + + SmallVector, 4> MDs; + SrcInst->getAllMetadata(MDs); + for (auto &MD : MDs) { + switch (MD.first) { + default: + break; + case LLVMContext::MD_prof: + if (Swapped && MD.second->getNumOperands() == 3 && + isa(MD.second->getOperand(0))) { + MDString *MDName = cast(MD.second->getOperand(0)); + if (MDName->getString() == "branch_weights") { + auto *ValT = cast_or_null( + MD.second->getOperand(1))->getValue(); + auto *ValF = cast_or_null( + MD.second->getOperand(2))->getValue(); + assert(ValT && ValF && "Invalid Operands of branch_weights"); + auto NewMD = + MDBuilder(DstInst->getParent()->getContext()) + .createBranchWeights(cast(ValF)->getZExtValue(), + cast(ValT)->getZExtValue()); + MD.second = NewMD; + } + } + // fallthrough. + case LLVMContext::MD_make_implicit: + case LLVMContext::MD_dbg: + DstInst->setMetadata(MD.first, MD.second); + } + } +} + +/// Emit a conditional branch on two values if LIC == Val, branch to TrueDst, +/// otherwise branch to FalseDest. Insert the code immediately before InsertPt. void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, BasicBlock *TrueDest, BasicBlock *FalseDest, - Instruction *InsertPt) { + Instruction *InsertPt, + TerminatorInst *TI) { // 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; + bool Swapped = false; if (!isa(Val) || Val->getType() != Type::getInt1Ty(LIC->getContext())) BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val); - else if (Val != ConstantInt::getTrue(Val->getContext())) + else if (Val != ConstantInt::getTrue(Val->getContext())) { // We want to enter the new loop when the condition is true. std::swap(TrueDest, FalseDest); + Swapped = true; + } // Insert the new branch. BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt); + copyMetadata(BI, TI, Swapped); // If either edge is critical, split it. This helps preserve LoopSimplify // form for enclosing loops. - SplitCriticalEdge(BI, 0, this); - SplitCriticalEdge(BI, 1, this); + auto Options = CriticalEdgeSplittingOptions(DT, LI).setPreserveLCSSA(); + SplitCriticalEdge(BI, 0, Options); + SplitCriticalEdge(BI, 1, Options); } -/// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable -/// condition in it (a cond branch from its header block to its latch block, -/// where the path through the loop that doesn't execute its body has no -/// side-effects), unswitch it. This doesn't involve any code duplication, just -/// moving the conditional branch outside of the loop and updating loop info. -void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, - Constant *Val, - BasicBlock *ExitBlock) { +/// Given a loop that has a trivial unswitchable condition in it (a cond branch +/// from its header block to its latch block, where the path through the loop +/// that doesn't execute its body has no side-effects), unswitch it. This +/// doesn't involve any code duplication, just moving the conditional branch +/// outside of the loop and updating loop info. +void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val, + BasicBlock *ExitBlock, + TerminatorInst *TI) { DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %" - << loopHeader->getName() << " [" << L->getBlocks().size() - << " blocks] in Function " << L->getHeader()->getParent()->getName() - << " on cond: " << *Val << " == " << *Cond << "\n"); - + << loopHeader->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 loopPreheader to have a // conditional branch on Cond. - BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this); + BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, DT, LI); // 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 // short-circuit to. - + // Split this block now, so that the loop maintains its exit block, and so // that the jump from the preheader can execute the contents of the exit block // 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(), this); - - // Okay, now we have a position to branch from and a position to branch to, + BasicBlock *NewExit = SplitBlock(ExitBlock, &ExitBlock->front(), DT, LI); + + // 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, - loopPreheader->getTerminator()); + EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, + loopPreheader->getTerminator(), TI); LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L); loopPreheader->getTerminator()->eraseFromParent(); // We need to reprocess this loop, it could be unswitched again. 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 // at least eliminate the old branch. @@ -552,10 +799,157 @@ void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, ++NumTrivial; } -/// 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){ +/// Check if the first non-constant condition starting from the loop header is +/// a trivial unswitch condition: that is, a condition controls whether or not +/// the loop does anything at all. If it is a trivial condition, unswitching +/// produces no code duplications (equivalently, it produces a simpler loop and +/// a new empty loop, which gets deleted). Therefore always unswitch trivial +/// condition. +bool LoopUnswitch::TryTrivialLoopUnswitch(bool &Changed) { + BasicBlock *CurrentBB = currentLoop->getHeader(); + TerminatorInst *CurrentTerm = CurrentBB->getTerminator(); + LLVMContext &Context = CurrentBB->getContext(); + + // If loop header has only one reachable successor (currently via an + // unconditional branch or constant foldable conditional branch, but + // should also consider adding constant foldable switch instruction in + // future), we should keep looking for trivial condition candidates in + // the successor as well. An alternative is to constant fold conditions + // and merge successors into loop header (then we only need to check header's + // terminator). The reason for not doing this in LoopUnswitch pass is that + // it could potentially break LoopPassManager's invariants. Folding dead + // branches could either eliminate the current loop or make other loops + // unreachable. LCSSA form might also not be preserved after deleting + // branches. The following code keeps traversing loop header's successors + // until it finds the trivial condition candidate (condition that is not a + // constant). Since unswitching generates branches with constant conditions, + // this scenario could be very common in practice. + SmallSet Visited; + + while (true) { + // If we exit loop or reach a previous visited block, then + // we can not reach any trivial condition candidates (unfoldable + // branch instructions or switch instructions) and no unswitch + // can happen. Exit and return false. + if (!currentLoop->contains(CurrentBB) || !Visited.insert(CurrentBB).second) + return false; + + // Check if this loop will execute any side-effecting instructions (e.g. + // stores, calls, volatile loads) in the part of the loop that the code + // *would* execute. Check the header first. + for (Instruction &I : *CurrentBB) + if (I.mayHaveSideEffects()) + return false; + + // FIXME: add check for constant foldable switch instructions. + if (BranchInst *BI = dyn_cast(CurrentTerm)) { + if (BI->isUnconditional()) { + CurrentBB = BI->getSuccessor(0); + } else if (BI->getCondition() == ConstantInt::getTrue(Context)) { + CurrentBB = BI->getSuccessor(0); + } else if (BI->getCondition() == ConstantInt::getFalse(Context)) { + CurrentBB = BI->getSuccessor(1); + } else { + // Found a trivial condition candidate: non-foldable conditional branch. + break; + } + } else { + break; + } + + CurrentTerm = CurrentBB->getTerminator(); + } + + // CondVal is the condition that controls the trivial condition. + // LoopExitBB is the BasicBlock that loop exits when meets trivial condition. + Constant *CondVal = nullptr; + BasicBlock *LoopExitBB = nullptr; + + if (BranchInst *BI = dyn_cast(CurrentTerm)) { + // If this isn't branching on an invariant condition, we can't unswitch it. + if (!BI->isConditional()) + return false; + + Value *LoopCond = FindLIVLoopCondition(BI->getCondition(), + currentLoop, Changed); + + // Unswitch only if the trivial condition itself is an LIV (not + // partial LIV which could occur in and/or) + if (!LoopCond || LoopCond != BI->getCondition()) + return false; + + // 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, + BI->getSuccessor(0)))) { + CondVal = ConstantInt::getTrue(Context); + } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, + BI->getSuccessor(1)))) { + CondVal = ConstantInt::getFalse(Context); + } + + // If we didn't find a single unique LoopExit block, or if the loop exit + // block contains phi nodes, this isn't trivial. + if (!LoopExitBB || isa(LoopExitBB->begin())) + return false; // Can't handle this. + + UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, LoopExitBB, + CurrentTerm); + ++NumBranches; + return true; + } else if (SwitchInst *SI = dyn_cast(CurrentTerm)) { + // If this isn't switching on an invariant condition, we can't unswitch it. + Value *LoopCond = FindLIVLoopCondition(SI->getCondition(), + currentLoop, Changed); + + // Unswitch only if the trivial condition itself is an LIV (not + // partial LIV which could occur in and/or) + if (!LoopCond || LoopCond != SI->getCondition()) + return false; + + // 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 or + // on already unswitched cases. + for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end(); + i != e; ++i) { + BasicBlock *LoopExitCandidate; + if ((LoopExitCandidate = isTrivialLoopExitBlock(currentLoop, + i.getCaseSuccessor()))) { + // Okay, we found a trivial case, remember the value that is trivial. + ConstantInt *CaseVal = i.getCaseValue(); + + // Check that it was not unswitched before, since already unswitched + // trivial vals are looks trivial too. + if (BranchesInfo.isUnswitched(SI, CaseVal)) + continue; + LoopExitBB = LoopExitCandidate; + CondVal = CaseVal; + break; + } + } + + // If we didn't find a single unique LoopExit block, or if the loop exit + // block contains phi nodes, this isn't trivial. + if (!LoopExitBB || isa(LoopExitBB->begin())) + return false; // Can't handle this. + + UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, LoopExitBB, + nullptr); + ++NumSwitches; + return true; + } + return false; +} + +/// 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 SmallVectorImpl &ExitBlocks){ for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { BasicBlock *ExitBlock = ExitBlocks[i]; @@ -564,37 +958,31 @@ void LoopUnswitch::SplitExitEdges(Loop *L, // 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.data(), Preds.size(), - ".us-lcssa", this); - } else { - SmallVector NewBBs; - SplitLandingPadPredecessors(ExitBlock, Preds, ".us-lcssa", ".us-lcssa", - this, NewBBs); - } + SplitBlockPredecessors(ExitBlock, Preds, ".us-lcssa", DT, LI, + /*PreserveLCSSA*/ true); } } -/// 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) { +/// 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, TerminatorInst *TI) { Function *F = loopHeader->getParent(); DEBUG(dbgs() << "loop-unswitch: Unswitching loop %" << loopHeader->getName() << " [" << L->getBlocks().size() << " blocks] in Function " << F->getName() << " when '" << *Val << "' == " << *LIC << "\n"); - if (ScalarEvolution *SE = getAnalysisIfAvailable()) - SE->forgetLoop(L); + if (auto *SEWP = getAnalysisIfAvailable()) + SEWP->getSE().forgetLoop(L); LoopBlocks.clear(); NewBlocks.clear(); // First step, split the preheader and exit blocks, and add these blocks to // the LoopBlocks list. - BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this); + BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, DT, LI); LoopBlocks.push_back(NewPreheader); // We want the loop to come after the preheader, but before the exit blocks. @@ -621,6 +1009,7 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, ValueToValueMapTy VMap; for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) { BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F); + NewBlocks.push_back(NewBB); VMap[LoopBlocks[i]] = NewBB; // Keep the BB mapping. LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L); @@ -628,33 +1017,42 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, // Splice the newly inserted blocks into the function right before the // original preheader. - F->getBasicBlockList().splice(NewPreheader, F->getBasicBlockList(), - NewBlocks[0], F->end()); + F->getBasicBlockList().splice(NewPreheader->getIterator(), + F->getBasicBlockList(), + NewBlocks[0]->getIterator(), F->end()); + + // FIXME: We could register any cloned assumptions instead of clearing the + // whole function's cache. + AC->clear(); // Now we create the new Loop object for the versioned loop. 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()); + ParentLoop->addBasicBlockToLoop(NewBlocks[0], *LI); } for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++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()); - + ExitBBLoop->addBasicBlockToLoop(NewExit, *LI); + assert(NewExit->getTerminator()->getNumSuccessors() == 1 && "Exit block should have been split to have one successor!"); BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0); // If the successor of the exit block had PHI nodes, add an entry for // NewExit. - PHINode *PN; - for (BasicBlock::iterator I = ExitSucc->begin(); isa(I); ++I) { - PN = cast(I); + for (BasicBlock::iterator I = ExitSucc->begin(); + PHINode *PN = dyn_cast(I); ++I) { Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]); ValueToValueMapTy::iterator It = VMap.find(V); if (It != VMap.end()) V = It->second; @@ -662,8 +1060,8 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, } if (LandingPadInst *LPad = NewExit->getLandingPadInst()) { - PN = PHINode::Create(LPad->getType(), 0, "", - ExitSucc->getFirstInsertionPt()); + PHINode *PN = PHINode::Create(LPad->getType(), 0, "", + &*ExitSucc->getFirstInsertionPt()); for (pred_iterator I = pred_begin(ExitSucc), E = pred_end(ExitSucc); I != E; ++I) { @@ -679,15 +1077,17 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, 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, VMap,RF_NoModuleLevelChanges|RF_IgnoreMissingEntries); - + RemapInstruction(&*I, VMap, + RF_NoModuleLevelChanges | RF_IgnoreMissingEntries); + // Rewrite the original preheader to select between versions of the loop. BranchInst *OldBR = cast(loopPreheader->getTerminator()); assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] && "Preheader splitting did not work correctly!"); // Emit the new branch that selects between the two versions of this loop. - EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR); + EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR, + TI); LPM->deleteSimpleAnalysisValue(OldBR, L); OldBR->eraseFromParent(); @@ -699,7 +1099,7 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, // the condition that we're unswitching on), we don't rewrite the second // iteration. WeakVH LICHandle(LIC); - + // Now we rewrite the original code to know that the condition is true and the // new code to know that the condition is false. RewriteLoopBodyWithConditionConstant(L, LIC, Val, false); @@ -712,22 +1112,17 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true); } -/// RemoveFromWorklist - Remove all instances of I from the worklist vector -/// specified. -static void RemoveFromWorklist(Instruction *I, +/// Remove all instances of I from the worklist vector specified. +static void RemoveFromWorklist(Instruction *I, std::vector &Worklist) { - std::vector::iterator WI = std::find(Worklist.begin(), - Worklist.end(), I); - while (WI != Worklist.end()) { - unsigned Offset = WI-Worklist.begin(); - Worklist.erase(WI); - WI = std::find(Worklist.begin()+Offset, Worklist.end(), I); - } + + Worklist.erase(std::remove(Worklist.begin(), Worklist.end(), I), + Worklist.end()); } -/// ReplaceUsesOfWith - When we find that I really equals V, remove I from the +/// 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, +static void ReplaceUsesOfWith(Instruction *I, Value *V, std::vector &Worklist, Loop *L, LPPassManager *LPM) { DEBUG(dbgs() << "Replace with '" << *V << "': " << *I); @@ -738,9 +1133,8 @@ static void ReplaceUsesOfWith(Instruction *I, Value *V, Worklist.push_back(Use); // Add users to the worklist which may be simplified now. - for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); - UI != E; ++UI) - Worklist.push_back(cast(*UI)); + for (User *U : I->users()) + Worklist.push_back(cast(U)); LPM->deleteSimpleAnalysisValue(I, L); RemoveFromWorklist(I, Worklist); I->replaceAllUsesWith(V); @@ -748,149 +1142,26 @@ static void ReplaceUsesOfWith(Instruction *I, Value *V, ++NumSimplify; } -/// RemoveBlockIfDead - If the specified block is dead, remove it, update loop -/// information, and remove any dead successors it has. -/// -void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB, - 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. - if (BasicBlock *Pred = BB->getSinglePredecessor()) { - // If it has one pred, fold phi nodes in BB. - while (isa(BB->begin())) - ReplaceUsesOfWith(BB->begin(), - cast(BB->begin())->getIncomingValue(0), - Worklist, L, LPM); - - // If this is the header of a loop and the only pred is the latch, we now - // have an unreachable loop. - if (Loop *L = LI->getLoopFor(BB)) - if (loopHeader == BB && L->contains(Pred)) { - // 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(BB->getContext(), Pred); - - // The loop is now broken, remove it from LI. - RemoveLoopFromHierarchy(L); - - // Reprocess the header, which now IS dead. - RemoveBlockIfDead(BB, Worklist, L); - return; - } - - // If pred ends in a uncond branch, add uncond branch to worklist so that - // the two blocks will get merged. - if (BranchInst *BI = dyn_cast(Pred->getTerminator())) - if (BI->isUnconditional()) - Worklist.push_back(BI); - } - return; - } - - DEBUG(dbgs() << "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) { - RemoveFromWorklist(I, Worklist); - - // Anything that uses the instructions in this basic block should have their - // uses replaced with undefs. - // If I is not void type then replaceAllUsesWith undef. - // This allows ValueHandlers and custom metadata to adjust itself. - if (!I->getType()->isVoidTy()) - I->replaceAllUsesWith(UndefValue::get(I->getType())); - } - - // 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) { - RemoveLoopFromHierarchy(BBLoop); - if (currentLoop == BBLoop) { - currentLoop = 0; - redoLoop = false; - } - } - } - - // Remove the block from the loop info, which removes it from any loops it - // was in. - LI->removeBlock(BB); - - - // Remove phi node entries in successors for this block. - TerminatorInst *TI = BB->getTerminator(); - SmallVector Succs; - for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) { - Succs.push_back(TI->getSuccessor(i)); - TI->getSuccessor(i)->removePredecessor(BB); - } - - // Unique the successors, remove anything with multiple uses. - array_pod_sort(Succs.begin(), Succs.end()); - 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. - // - for (unsigned i = 0; i != Succs.size(); ++i) - if (pred_begin(Succs[i]) != pred_end(Succs[i])) { - // One exception is loop headers. If this block was the preheader for a - // loop, then we DO want to visit the loop so the loop gets deleted. - // We know that if the successor is a loop header, that this loop had to - // be the preheader: the case where this was the latch block was handled - // above and headers can only have two predecessors. - if (!LI->isLoopHeader(Succs[i])) { - Succs.erase(Succs.begin()+i); - --i; - } - } - - for (unsigned i = 0, e = Succs.size(); i != e; ++i) - RemoveBlockIfDead(Succs[i], Worklist, L); -} - -/// RemoveLoopFromHierarchy - We have discovered that the specified loop has -/// become unwrapped, either because the backedge was deleted, or because the -/// edge into the header was removed. If the edge into the header from the -/// latch block was removed, the loop is unwrapped but subloops are still alive, -/// so they just reparent loops. If the loops are actually dead, they will be -/// removed later. -void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) { - LPM->deleteLoopFromQueue(L); - RemoveLoopFromWorklist(L); -} - -// RewriteLoopBodyWithConditionConstant - We know either that the value LIC has -// the value specified by Val in the specified loop, or we know it does NOT have -// that value. Rewrite any uses of LIC or of properties correlated to it. +/// We know either that the value LIC has the value specified by Val in the +/// specified loop, or we know it does NOT have that value. +/// Rewrite any uses of LIC or of properties correlated to it. void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, Constant *Val, bool IsEqual) { assert(!isa(LIC) && "Why are we unswitching on a constant?"); - + // FIXME: Support correlated properties, like: // for (...) // if (li1 < li2) // ... // if (li1 > li2) // ... - + // FOLD boolean conditions (X|LIC), (X&LIC). Fold conditional branches, // selects, switches. std::vector Worklist; LLVMContext &Context = Val->getContext(); - // 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) && @@ -899,49 +1170,55 @@ void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, if (IsEqual) Replacement = Val; else - Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()), + Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()), !cast(Val)->getZExtValue()); - - 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)) + + for (User *U : LIC->users()) { + Instruction *UI = dyn_cast(U); + if (!UI || !L->contains(UI)) continue; - U->replaceUsesOfWith(LIC, Replacement); - Worklist.push_back(U); + Worklist.push_back(UI); } + + for (std::vector::iterator UI = Worklist.begin(), + UE = Worklist.end(); UI != UE; ++UI) + (*UI)->replaceUsesOfWith(LIC, Replacement); + SimplifyCode(Worklist, L); return; } - + // 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 (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end(); - UI != E; ++UI) { - Instruction *U = dyn_cast(*UI); - if (!U || !L->contains(U)) + for (User *U : LIC->users()) { + Instruction *UI = dyn_cast(U); + if (!UI || !L->contains(UI)) continue; - Worklist.push_back(U); + Worklist.push_back(UI); - // TODO: We could do other simplifications, for example, turning + // TODO: We could do other simplifications, for example, turning // 'icmp eq LIC, Val' -> false. // If we know that LIC is not Val, use this info to simplify code. - SwitchInst *SI = dyn_cast(U); - if (SI == 0 || !isa(Val)) continue; - - unsigned DeadCase = SI->findCaseValue(cast(Val)); - if (DeadCase == 0) continue; // Default case is live for multiple values. - - // Found a dead case value. Don't remove PHI nodes in the + SwitchInst *SI = dyn_cast(UI); + if (!SI || !isa(Val)) continue; + + SwitchInst::CaseIt DeadCase = SI->findCaseValue(cast(Val)); + // Default case is live for multiple values. + if (DeadCase == SI->case_default()) continue; + + // 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 *SISucc = DeadCase.getCaseSuccessor(); 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 @@ -953,11 +1230,11 @@ void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, // 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. - SplitEdge(Switch, SISucc, this); + SplitEdge(Switch, SISucc, DT, LI); // Compute the successors instead of relying on the return value // of SplitEdge, since it may have split the switch successor // after PHI nodes. - BasicBlock *NewSISucc = SI->getSuccessor(DeadCase); + BasicBlock *NewSISucc = DeadCase.getCaseSuccessor(); BasicBlock *OldSISucc = *succ_begin(NewSISucc); // Create an "unreachable" destination. BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable", @@ -978,23 +1255,22 @@ void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, // domtree here -- instead we force it to do a full recomputation // after the pass is complete -- but we do need to inform it of // new blocks. - if (DT) - DT->addNewBlock(Abort, NewSISucc); + DT->addNewBlock(Abort, NewSISucc); } - + SimplifyCode(Worklist, L); } -/// SimplifyCode - Okay, now that we have simplified some instructions in the -/// loop, walk over it and constant prop, dce, and fold control flow where -/// possible. Note that this is effectively a very simple loop-structure-aware -/// optimizer. During processing of this loop, L could very well be deleted, so -/// it must not be used. +/// Now that we have simplified some instructions in the loop, walk over it and +/// constant prop, dce, and fold control flow where possible. Note that this is +/// effectively a very simple loop-structure-aware optimizer. During processing +/// of this loop, L could very well be deleted, so it must not be used. /// /// FIXME: When the loop optimizer is more mature, separate this out to a new /// pass. /// void LoopUnswitch::SimplifyCode(std::vector &Worklist, Loop *L) { + const DataLayout &DL = L->getHeader()->getModule()->getDataLayout(); while (!Worklist.empty()) { Instruction *I = Worklist.back(); Worklist.pop_back(); @@ -1002,7 +1278,7 @@ void LoopUnswitch::SimplifyCode(std::vector &Worklist, Loop *L) { // Simple DCE. if (isInstructionTriviallyDead(I)) { DEBUG(dbgs() << "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))) @@ -1016,8 +1292,8 @@ void LoopUnswitch::SimplifyCode(std::vector &Worklist, Loop *L) { // 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, DT)) + // 'false'. TODO: update the domtree properly so we can pass it here. + if (Value *V = SimplifyInstruction(I, DL)) if (LI->replacementPreservesLCSSAForm(I, V)) { ReplaceUsesOfWith(I, V, Worklist, L, LPM); continue; @@ -1034,24 +1310,24 @@ void LoopUnswitch::SimplifyCode(std::vector &Worklist, Loop *L) { if (!SinglePred) continue; // Nothing to do. assert(SinglePred == Pred && "CFG broken"); - DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- " + DEBUG(dbgs() << "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, 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()); + Pred->getInstList().splice(BI->getIterator(), Succ->getInstList(), + Succ->begin(), Succ->end()); LPM->deleteSimpleAnalysisValue(BI, L); BI->eraseFromParent(); RemoveFromWorklist(BI, Worklist); - + // Remove Succ from the loop tree. LI->removeBlock(Succ); LPM->deleteSimpleAnalysisValue(Succ, L); @@ -1059,24 +1335,7 @@ void LoopUnswitch::SimplifyCode(std::vector &Worklist, Loop *L) { ++NumSimplify; continue; } - - if (ConstantInt *CB = dyn_cast(BI->getCondition())){ - // Conditional branch. Turn it into an unconditional branch, then - // remove dead blocks. - continue; // FIXME: Enable. - - DEBUG(dbgs() << "Folded branch: " << *BI); - BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue()); - BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue()); - DeadSucc->removePredecessor(BI->getParent(), true); - Worklist.push_back(BranchInst::Create(LiveSucc, BI)); - LPM->deleteSimpleAnalysisValue(BI, L); - BI->eraseFromParent(); - RemoveFromWorklist(BI, Worklist); - ++NumSimplify; - RemoveBlockIfDead(DeadSucc, Worklist, L); - } continue; } }