X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTransforms%2FScalar%2FLoopUnswitch.cpp;h=5954f4af51fbe563c64eb42d65948f573fb3843e;hb=0d5c0629bbd49977ed53a093fd96ed3fd2c234f5;hp=de124d10aaae08acec35c6afd02ee4344a5959f7;hpb=a78130c3207e59bfb2ec0b77ca3f48677685d9ae;p=oota-llvm.git diff --git a/lib/Transforms/Scalar/LoopUnswitch.cpp b/lib/Transforms/Scalar/LoopUnswitch.cpp index de124d10aaa..5954f4af51f 100644 --- a/lib/Transforms/Scalar/LoopUnswitch.cpp +++ b/lib/Transforms/Scalar/LoopUnswitch.cpp @@ -28,25 +28,28 @@ #define DEBUG_TYPE "loop-unswitch" #include "llvm/Transforms/Scalar.h" -#include "llvm/Constants.h" -#include "llvm/DerivedTypes.h" -#include "llvm/Function.h" -#include "llvm/Instructions.h" -#include "llvm/Analysis/ConstantFolding.h" -#include "llvm/Analysis/InlineCost.h" +#include "llvm/ADT/STLExtras.h" +#include "llvm/ADT/SmallPtrSet.h" +#include "llvm/ADT/Statistic.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/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/ScalarEvolution.h" +#include "llvm/Analysis/TargetTransformInfo.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/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; @@ -55,14 +58,70 @@ 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); + namespace { + + class LUAnalysisCache { + + typedef DenseMap > + UnswitchedValsMap; + + typedef UnswitchedValsMap::iterator UnswitchedValsIt; + + struct LoopProperties { + unsigned CanBeUnswitchedCount; + 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; + + // Max size of code we can produce on remained iterations. + unsigned MaxSize; + + public: + + LUAnalysisCache() : + CurLoopInstructions(0), CurrentLoopProperties(0), + 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); + + // 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; @@ -70,19 +129,19 @@ namespace { // LoopProcessWorklist - Used to check if second loop needs processing // after RewriteLoopBodyWithConditionConstant rewrites first loop. std::vector LoopProcessWorklist; - SmallPtrSet UnswitchedVals; - + + LUAnalysisCache BranchesInfo; + bool OptimizeForSize; bool redoLoop; Loop *currentLoop; - DominanceFrontier *DF; 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. @@ -90,32 +149,35 @@ 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) {} + explicit LoopUnswitch(bool Os = false) : + LoopPass(ID), OptimizeForSize(Os), redoLoop(false), + currentLoop(0), DT(0), loopHeader(0), + loopPreheader(0) { + 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... + /// loop preheaders be inserted into the CFG. /// - virtual void getAnalysisUsage(AnalysisUsage &AU) const { + void getAnalysisUsage(AnalysisUsage &AU) const override { AU.addRequiredID(LoopSimplifyID); AU.addPreservedID(LoopSimplifyID); AU.addRequired(); AU.addPreserved(); AU.addRequiredID(LCSSAID); AU.addPreservedID(LCSSAID); - AU.addPreserved(); - AU.addPreserved(); + AU.addPreserved(); + AU.addPreserved(); + AU.addRequired(); } private: - virtual void releaseMemory() { - UnswitchedVals.clear(); + void releaseMemory() override { + BranchesInfo.forgetLoop(currentLoop); } /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist, @@ -134,7 +196,7 @@ 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); void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val, @@ -145,30 +207,152 @@ namespace { Constant *Val, bool isEqual); void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, - BasicBlock *TrueDest, + BasicBlock *TrueDest, BasicBlock *FalseDest, Instruction *InsertPt); 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); }; } -char LoopUnswitch::ID = 0; -static RegisterPass X("loop-unswitch", "Unswitch loops"); -Pass *llvm::createLoopUnswitchPass(bool Os) { - return new LoopUnswitch(Os); +// 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) { + + 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. + + // 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); + + Props.SizeEstimation = std::min(Metrics.NumInsts, Metrics.NumBlocks * 5); + Props.CanBeUnswitchedCount = MaxSize / (Props.SizeEstimation); + MaxSize -= Props.SizeEstimation * Props.CanBeUnswitchedCount; + + if (Metrics.notDuplicatable) { + DEBUG(dbgs() << "NOT unswitching loop %" + << L->getHeader()->getName() << ", contents cannot be " + << "duplicated!\n"); + return false; + } + } + + 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 = 0; + CurLoopInstructions = 0; +} + +// 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; +INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops", + false, false) +INITIALIZE_AG_DEPENDENCY(TargetTransformInfo) +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); } /// 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. 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; @@ -193,15 +377,19 @@ static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) { if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed)) return RHS; } - + return 0; } bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { + if (skipOptnoneFunction(L)) + return false; + LI = &getAnalysis(); LPM = &LPM_Ref; - DF = getAnalysisIfAvailable(); - DT = getAnalysisIfAvailable(); + DominatorTreeWrapperPass *DTWP = + getAnalysisIfAvailable(); + DT = DTWP ? &DTWP->getDomTree() : 0; currentLoop = L; Function *F = currentLoop->getHeader()->getParent(); bool Changed = false; @@ -214,23 +402,41 @@ bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) { if (Changed) { // FIXME: Reconstruct dom info, because it is not preserved properly. if (DT) - DT->runOnFunction(*F); - if (DF) - DF->runOnFunction(*F); + DT->recalculate(*F); } return Changed; } -/// processCurrentLoop - Do actual work and unswitch loop if possible +/// processCurrentLoop - 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(); + + // Probably we reach the quota of branches for this loop. If so + // stop unswitching. + if (!BranchesInfo.countLoop(currentLoop, getAnalysis())) + 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)) { @@ -239,23 +445,37 @@ 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, + if (LoopCond && UnswitchIfProfitable(LoopCond, ConstantInt::getTrue(Context))) { ++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! - Constant *UnswitchVal = SI->getCaseValue(1); + // FIXME: scan for a case with a non-critical edge? + Constant *UnswitchVal = 0; + // 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)) { @@ -264,14 +484,14 @@ 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; @@ -281,27 +501,28 @@ 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; - } else if (!L->contains(BB)) { + // Already visited. Without more analysis, this could indicate an infinite + // loop. + return false; + } + 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; 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. @@ -314,16 +535,16 @@ 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 +/// 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 are ok. + Visited.insert(L->getHeader()); // Branches to header make infinite loops. BasicBlock *ExitBB = 0; if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited)) return ExitBB; @@ -347,49 +568,61 @@ bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val, 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 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, + if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, BI->getSuccessor(0)))) { if (Val) *Val = ConstantInt::getTrue(Context); - } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop, + } 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 + // 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)))) { + // 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. - if (Val) *Val = SI->getCaseValue(i); + 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; + if (Val) *Val = 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. - + 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 @@ -405,17 +638,10 @@ bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val, /// 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. @@ -426,51 +652,18 @@ bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) { // 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.NeverInline) { - DEBUG(dbgs() << "NOT unswitching loop %" - << currentLoop->getHeader()->getName() << ", cost too high: " - << currentLoop->getBlocks().size() << "\n"); + if (OptimizeForSize || + F->getAttributes().hasAttribute(AttributeSet::FunctionIndex, + Attribute::OptimizeForSize)) 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 +693,7 @@ 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); @@ -510,23 +703,23 @@ void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val, // If either edge is critical, split it. This helps preserve LoopSimplify // form for enclosing loops. - SplitCriticalEdge(BI, 0, this); - SplitCriticalEdge(BI, 1, this); + SplitCriticalEdge(BI, 0, this, false, false, true); + SplitCriticalEdge(BI, 1, this, false, false, true); } /// 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 +/// 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, +void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, + Constant *Val, BasicBlock *ExitBlock) { DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %" << 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. @@ -535,24 +728,24 @@ void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, // 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, + + // 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, + EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH, loopPreheader->getTerminator()); 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. @@ -562,22 +755,30 @@ void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond, /// 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){ +void LoopUnswitch::SplitExitEdges(Loop *L, + const SmallVectorImpl &ExitBlocks){ for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) { 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); + } } } -/// UnswitchNontrivialCondition - We determined that the loop is profitable -/// to unswitch when LIC equal Val. Split it into loop versions and test the +/// 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, +void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, Loop *L) { Function *F = loopHeader->getParent(); DEBUG(dbgs() << "loop-unswitch: Unswitching loop %" @@ -585,6 +786,9 @@ 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 +818,12 @@ 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); } @@ -628,42 +833,59 @@ 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()); - + 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]); - 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()) { + PHINode *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()); assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] && @@ -682,7 +904,7 @@ 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); @@ -697,20 +919,16 @@ void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val, /// RemoveFromWorklist - Remove all instances of I from the worklist vector /// specified. -static void RemoveFromWorklist(Instruction *I, +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 /// 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); @@ -721,9 +939,8 @@ 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); @@ -731,112 +948,6 @@ 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); - } - - // 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 @@ -855,21 +966,19 @@ 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 Users(LIC->use_begin(), LIC->use_end()); 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) && @@ -878,54 +987,71 @@ 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 (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 (User *U : LIC->users()) { + Instruction *UI = dyn_cast(U); + if (!UI || !L->contains(UI)) + continue; + 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 (unsigned i = 0, e = Users.size(); i != e; ++i) { - Instruction *U = cast(Users[i]); - if (!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); + SwitchInst *SI = dyn_cast(UI); 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::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 = 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 + // 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. - 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", @@ -949,7 +1075,7 @@ void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC, if (DT) DT->addNewBlock(Abort, NewSISucc); } - + SimplifyCode(Worklist, L); } @@ -966,17 +1092,11 @@ 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); - + // 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))) @@ -987,48 +1107,18 @@ 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'. TODO: update the domtree properly so we can pass it here. + if (Value *V = SimplifyInstruction(I)) + 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. @@ -1038,51 +1128,33 @@ 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()); LPM->deleteSimpleAnalysisValue(BI, 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. - - 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); - } - break; - } + continue; } } }