Home | History | Annotate | Download | only in Scalar
      1 //===-- LoopUnswitch.cpp - Hoist loop-invariant conditionals in loop ------===//
      2 //
      3 //                     The LLVM Compiler Infrastructure
      4 //
      5 // This file is distributed under the University of Illinois Open Source
      6 // License. See LICENSE.TXT for details.
      7 //
      8 //===----------------------------------------------------------------------===//
      9 //
     10 // This pass transforms loops that contain branches on loop-invariant conditions
     11 // to have multiple loops.  For example, it turns the left into the right code:
     12 //
     13 //  for (...)                  if (lic)
     14 //    A                          for (...)
     15 //    if (lic)                     A; B; C
     16 //      B                      else
     17 //    C                          for (...)
     18 //                                 A; C
     19 //
     20 // This can increase the size of the code exponentially (doubling it every time
     21 // a loop is unswitched) so we only unswitch if the resultant code will be
     22 // smaller than a threshold.
     23 //
     24 // This pass expects LICM to be run before it to hoist invariant conditions out
     25 // of the loop, to make the unswitching opportunity obvious.
     26 //
     27 //===----------------------------------------------------------------------===//
     28 
     29 #define DEBUG_TYPE "loop-unswitch"
     30 #include "llvm/Transforms/Scalar.h"
     31 #include "llvm/ADT/STLExtras.h"
     32 #include "llvm/ADT/SmallPtrSet.h"
     33 #include "llvm/ADT/Statistic.h"
     34 #include "llvm/Analysis/CodeMetrics.h"
     35 #include "llvm/Analysis/Dominators.h"
     36 #include "llvm/Analysis/InstructionSimplify.h"
     37 #include "llvm/Analysis/LoopInfo.h"
     38 #include "llvm/Analysis/LoopPass.h"
     39 #include "llvm/Analysis/ScalarEvolution.h"
     40 #include "llvm/Analysis/TargetTransformInfo.h"
     41 #include "llvm/IR/Constants.h"
     42 #include "llvm/IR/DerivedTypes.h"
     43 #include "llvm/IR/Function.h"
     44 #include "llvm/IR/Instructions.h"
     45 #include "llvm/Support/CommandLine.h"
     46 #include "llvm/Support/Debug.h"
     47 #include "llvm/Support/raw_ostream.h"
     48 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
     49 #include "llvm/Transforms/Utils/Cloning.h"
     50 #include "llvm/Transforms/Utils/Local.h"
     51 #include <algorithm>
     52 #include <map>
     53 #include <set>
     54 using namespace llvm;
     55 
     56 STATISTIC(NumBranches, "Number of branches unswitched");
     57 STATISTIC(NumSwitches, "Number of switches unswitched");
     58 STATISTIC(NumSelects , "Number of selects unswitched");
     59 STATISTIC(NumTrivial , "Number of unswitches that are trivial");
     60 STATISTIC(NumSimplify, "Number of simplifications of unswitched code");
     61 STATISTIC(TotalInsts,  "Total number of instructions analyzed");
     62 
     63 // The specific value of 100 here was chosen based only on intuition and a
     64 // few specific examples.
     65 static cl::opt<unsigned>
     66 Threshold("loop-unswitch-threshold", cl::desc("Max loop size to unswitch"),
     67           cl::init(100), cl::Hidden);
     68 
     69 namespace {
     70 
     71   class LUAnalysisCache {
     72 
     73     typedef DenseMap<const SwitchInst*, SmallPtrSet<const Value *, 8> >
     74       UnswitchedValsMap;
     75 
     76     typedef UnswitchedValsMap::iterator UnswitchedValsIt;
     77 
     78     struct LoopProperties {
     79       unsigned CanBeUnswitchedCount;
     80       unsigned SizeEstimation;
     81       UnswitchedValsMap UnswitchedVals;
     82     };
     83 
     84     // Here we use std::map instead of DenseMap, since we need to keep valid
     85     // LoopProperties pointer for current loop for better performance.
     86     typedef std::map<const Loop*, LoopProperties> LoopPropsMap;
     87     typedef LoopPropsMap::iterator LoopPropsMapIt;
     88 
     89     LoopPropsMap LoopsProperties;
     90     UnswitchedValsMap *CurLoopInstructions;
     91     LoopProperties *CurrentLoopProperties;
     92 
     93     // Max size of code we can produce on remained iterations.
     94     unsigned MaxSize;
     95 
     96     public:
     97 
     98       LUAnalysisCache() :
     99         CurLoopInstructions(0), CurrentLoopProperties(0),
    100         MaxSize(Threshold)
    101       {}
    102 
    103       // Analyze loop. Check its size, calculate is it possible to unswitch
    104       // it. Returns true if we can unswitch this loop.
    105       bool countLoop(const Loop *L, const TargetTransformInfo &TTI);
    106 
    107       // Clean all data related to given loop.
    108       void forgetLoop(const Loop *L);
    109 
    110       // Mark case value as unswitched.
    111       // Since SI instruction can be partly unswitched, in order to avoid
    112       // extra unswitching in cloned loops keep track all unswitched values.
    113       void setUnswitched(const SwitchInst *SI, const Value *V);
    114 
    115       // Check was this case value unswitched before or not.
    116       bool isUnswitched(const SwitchInst *SI, const Value *V);
    117 
    118       // Clone all loop-unswitch related loop properties.
    119       // Redistribute unswitching quotas.
    120       // Note, that new loop data is stored inside the VMap.
    121       void cloneData(const Loop *NewLoop, const Loop *OldLoop,
    122                      const ValueToValueMapTy &VMap);
    123   };
    124 
    125   class LoopUnswitch : public LoopPass {
    126     LoopInfo *LI;  // Loop information
    127     LPPassManager *LPM;
    128 
    129     // LoopProcessWorklist - Used to check if second loop needs processing
    130     // after RewriteLoopBodyWithConditionConstant rewrites first loop.
    131     std::vector<Loop*> LoopProcessWorklist;
    132 
    133     LUAnalysisCache BranchesInfo;
    134 
    135     bool OptimizeForSize;
    136     bool redoLoop;
    137 
    138     Loop *currentLoop;
    139     DominatorTree *DT;
    140     BasicBlock *loopHeader;
    141     BasicBlock *loopPreheader;
    142 
    143     // LoopBlocks contains all of the basic blocks of the loop, including the
    144     // preheader of the loop, the body of the loop, and the exit blocks of the
    145     // loop, in that order.
    146     std::vector<BasicBlock*> LoopBlocks;
    147     // NewBlocks contained cloned copy of basic blocks from LoopBlocks.
    148     std::vector<BasicBlock*> NewBlocks;
    149 
    150   public:
    151     static char ID; // Pass ID, replacement for typeid
    152     explicit LoopUnswitch(bool Os = false) :
    153       LoopPass(ID), OptimizeForSize(Os), redoLoop(false),
    154       currentLoop(0), DT(0), loopHeader(0),
    155       loopPreheader(0) {
    156         initializeLoopUnswitchPass(*PassRegistry::getPassRegistry());
    157       }
    158 
    159     bool runOnLoop(Loop *L, LPPassManager &LPM);
    160     bool processCurrentLoop();
    161 
    162     /// This transformation requires natural loop information & requires that
    163     /// loop preheaders be inserted into the CFG.
    164     ///
    165     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
    166       AU.addRequiredID(LoopSimplifyID);
    167       AU.addPreservedID(LoopSimplifyID);
    168       AU.addRequired<LoopInfo>();
    169       AU.addPreserved<LoopInfo>();
    170       AU.addRequiredID(LCSSAID);
    171       AU.addPreservedID(LCSSAID);
    172       AU.addPreserved<DominatorTree>();
    173       AU.addPreserved<ScalarEvolution>();
    174       AU.addRequired<TargetTransformInfo>();
    175     }
    176 
    177   private:
    178 
    179     virtual void releaseMemory() {
    180       BranchesInfo.forgetLoop(currentLoop);
    181     }
    182 
    183     /// RemoveLoopFromWorklist - If the specified loop is on the loop worklist,
    184     /// remove it.
    185     void RemoveLoopFromWorklist(Loop *L) {
    186       std::vector<Loop*>::iterator I = std::find(LoopProcessWorklist.begin(),
    187                                                  LoopProcessWorklist.end(), L);
    188       if (I != LoopProcessWorklist.end())
    189         LoopProcessWorklist.erase(I);
    190     }
    191 
    192     void initLoopData() {
    193       loopHeader = currentLoop->getHeader();
    194       loopPreheader = currentLoop->getLoopPreheader();
    195     }
    196 
    197     /// Split all of the edges from inside the loop to their exit blocks.
    198     /// Update the appropriate Phi nodes as we do so.
    199     void SplitExitEdges(Loop *L, const SmallVectorImpl<BasicBlock *> &ExitBlocks);
    200 
    201     bool UnswitchIfProfitable(Value *LoopCond, Constant *Val);
    202     void UnswitchTrivialCondition(Loop *L, Value *Cond, Constant *Val,
    203                                   BasicBlock *ExitBlock);
    204     void UnswitchNontrivialCondition(Value *LIC, Constant *OnVal, Loop *L);
    205 
    206     void RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
    207                                               Constant *Val, bool isEqual);
    208 
    209     void EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
    210                                         BasicBlock *TrueDest,
    211                                         BasicBlock *FalseDest,
    212                                         Instruction *InsertPt);
    213 
    214     void SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L);
    215     void RemoveBlockIfDead(BasicBlock *BB,
    216                            std::vector<Instruction*> &Worklist, Loop *l);
    217     void RemoveLoopFromHierarchy(Loop *L);
    218     bool IsTrivialUnswitchCondition(Value *Cond, Constant **Val = 0,
    219                                     BasicBlock **LoopExit = 0);
    220 
    221   };
    222 }
    223 
    224 // Analyze loop. Check its size, calculate is it possible to unswitch
    225 // it. Returns true if we can unswitch this loop.
    226 bool LUAnalysisCache::countLoop(const Loop *L, const TargetTransformInfo &TTI) {
    227 
    228   LoopPropsMapIt PropsIt;
    229   bool Inserted;
    230   llvm::tie(PropsIt, Inserted) =
    231       LoopsProperties.insert(std::make_pair(L, LoopProperties()));
    232 
    233   LoopProperties &Props = PropsIt->second;
    234 
    235   if (Inserted) {
    236     // New loop.
    237 
    238     // Limit the number of instructions to avoid causing significant code
    239     // expansion, and the number of basic blocks, to avoid loops with
    240     // large numbers of branches which cause loop unswitching to go crazy.
    241     // This is a very ad-hoc heuristic.
    242 
    243     // FIXME: This is overly conservative because it does not take into
    244     // consideration code simplification opportunities and code that can
    245     // be shared by the resultant unswitched loops.
    246     CodeMetrics Metrics;
    247     for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
    248          I != E; ++I)
    249       Metrics.analyzeBasicBlock(*I, TTI);
    250 
    251     Props.SizeEstimation = std::min(Metrics.NumInsts, Metrics.NumBlocks * 5);
    252     Props.CanBeUnswitchedCount = MaxSize / (Props.SizeEstimation);
    253     MaxSize -= Props.SizeEstimation * Props.CanBeUnswitchedCount;
    254 
    255     if (Metrics.notDuplicatable) {
    256       DEBUG(dbgs() << "NOT unswitching loop %"
    257                    << L->getHeader()->getName() << ", contents cannot be "
    258                    << "duplicated!\n");
    259       return false;
    260     }
    261   }
    262 
    263   if (!Props.CanBeUnswitchedCount) {
    264     DEBUG(dbgs() << "NOT unswitching loop %"
    265                  << L->getHeader()->getName() << ", cost too high: "
    266                  << L->getBlocks().size() << "\n");
    267     return false;
    268   }
    269 
    270   // Be careful. This links are good only before new loop addition.
    271   CurrentLoopProperties = &Props;
    272   CurLoopInstructions = &Props.UnswitchedVals;
    273 
    274   return true;
    275 }
    276 
    277 // Clean all data related to given loop.
    278 void LUAnalysisCache::forgetLoop(const Loop *L) {
    279 
    280   LoopPropsMapIt LIt = LoopsProperties.find(L);
    281 
    282   if (LIt != LoopsProperties.end()) {
    283     LoopProperties &Props = LIt->second;
    284     MaxSize += Props.CanBeUnswitchedCount * Props.SizeEstimation;
    285     LoopsProperties.erase(LIt);
    286   }
    287 
    288   CurrentLoopProperties = 0;
    289   CurLoopInstructions = 0;
    290 }
    291 
    292 // Mark case value as unswitched.
    293 // Since SI instruction can be partly unswitched, in order to avoid
    294 // extra unswitching in cloned loops keep track all unswitched values.
    295 void LUAnalysisCache::setUnswitched(const SwitchInst *SI, const Value *V) {
    296   (*CurLoopInstructions)[SI].insert(V);
    297 }
    298 
    299 // Check was this case value unswitched before or not.
    300 bool LUAnalysisCache::isUnswitched(const SwitchInst *SI, const Value *V) {
    301   return (*CurLoopInstructions)[SI].count(V);
    302 }
    303 
    304 // Clone all loop-unswitch related loop properties.
    305 // Redistribute unswitching quotas.
    306 // Note, that new loop data is stored inside the VMap.
    307 void LUAnalysisCache::cloneData(const Loop *NewLoop, const Loop *OldLoop,
    308                                 const ValueToValueMapTy &VMap) {
    309 
    310   LoopProperties &NewLoopProps = LoopsProperties[NewLoop];
    311   LoopProperties &OldLoopProps = *CurrentLoopProperties;
    312   UnswitchedValsMap &Insts = OldLoopProps.UnswitchedVals;
    313 
    314   // Reallocate "can-be-unswitched quota"
    315 
    316   --OldLoopProps.CanBeUnswitchedCount;
    317   unsigned Quota = OldLoopProps.CanBeUnswitchedCount;
    318   NewLoopProps.CanBeUnswitchedCount = Quota / 2;
    319   OldLoopProps.CanBeUnswitchedCount = Quota - Quota / 2;
    320 
    321   NewLoopProps.SizeEstimation = OldLoopProps.SizeEstimation;
    322 
    323   // Clone unswitched values info:
    324   // for new loop switches we clone info about values that was
    325   // already unswitched and has redundant successors.
    326   for (UnswitchedValsIt I = Insts.begin(); I != Insts.end(); ++I) {
    327     const SwitchInst *OldInst = I->first;
    328     Value *NewI = VMap.lookup(OldInst);
    329     const SwitchInst *NewInst = cast_or_null<SwitchInst>(NewI);
    330     assert(NewInst && "All instructions that are in SrcBB must be in VMap.");
    331 
    332     NewLoopProps.UnswitchedVals[NewInst] = OldLoopProps.UnswitchedVals[OldInst];
    333   }
    334 }
    335 
    336 char LoopUnswitch::ID = 0;
    337 INITIALIZE_PASS_BEGIN(LoopUnswitch, "loop-unswitch", "Unswitch loops",
    338                       false, false)
    339 INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
    340 INITIALIZE_PASS_DEPENDENCY(LoopSimplify)
    341 INITIALIZE_PASS_DEPENDENCY(LoopInfo)
    342 INITIALIZE_PASS_DEPENDENCY(LCSSA)
    343 INITIALIZE_PASS_END(LoopUnswitch, "loop-unswitch", "Unswitch loops",
    344                       false, false)
    345 
    346 Pass *llvm::createLoopUnswitchPass(bool Os) {
    347   return new LoopUnswitch(Os);
    348 }
    349 
    350 /// FindLIVLoopCondition - Cond is a condition that occurs in L.  If it is
    351 /// invariant in the loop, or has an invariant piece, return the invariant.
    352 /// Otherwise, return null.
    353 static Value *FindLIVLoopCondition(Value *Cond, Loop *L, bool &Changed) {
    354 
    355   // We started analyze new instruction, increment scanned instructions counter.
    356   ++TotalInsts;
    357 
    358   // We can never unswitch on vector conditions.
    359   if (Cond->getType()->isVectorTy())
    360     return 0;
    361 
    362   // Constants should be folded, not unswitched on!
    363   if (isa<Constant>(Cond)) return 0;
    364 
    365   // TODO: Handle: br (VARIANT|INVARIANT).
    366 
    367   // Hoist simple values out.
    368   if (L->makeLoopInvariant(Cond, Changed))
    369     return Cond;
    370 
    371   if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Cond))
    372     if (BO->getOpcode() == Instruction::And ||
    373         BO->getOpcode() == Instruction::Or) {
    374       // If either the left or right side is invariant, we can unswitch on this,
    375       // which will cause the branch to go away in one loop and the condition to
    376       // simplify in the other one.
    377       if (Value *LHS = FindLIVLoopCondition(BO->getOperand(0), L, Changed))
    378         return LHS;
    379       if (Value *RHS = FindLIVLoopCondition(BO->getOperand(1), L, Changed))
    380         return RHS;
    381     }
    382 
    383   return 0;
    384 }
    385 
    386 bool LoopUnswitch::runOnLoop(Loop *L, LPPassManager &LPM_Ref) {
    387   LI = &getAnalysis<LoopInfo>();
    388   LPM = &LPM_Ref;
    389   DT = getAnalysisIfAvailable<DominatorTree>();
    390   currentLoop = L;
    391   Function *F = currentLoop->getHeader()->getParent();
    392   bool Changed = false;
    393   do {
    394     assert(currentLoop->isLCSSAForm(*DT));
    395     redoLoop = false;
    396     Changed |= processCurrentLoop();
    397   } while(redoLoop);
    398 
    399   if (Changed) {
    400     // FIXME: Reconstruct dom info, because it is not preserved properly.
    401     if (DT)
    402       DT->runOnFunction(*F);
    403   }
    404   return Changed;
    405 }
    406 
    407 /// processCurrentLoop - Do actual work and unswitch loop if possible
    408 /// and profitable.
    409 bool LoopUnswitch::processCurrentLoop() {
    410   bool Changed = false;
    411 
    412   initLoopData();
    413 
    414   // If LoopSimplify was unable to form a preheader, don't do any unswitching.
    415   if (!loopPreheader)
    416     return false;
    417 
    418   // Loops with indirectbr cannot be cloned.
    419   if (!currentLoop->isSafeToClone())
    420     return false;
    421 
    422   // Without dedicated exits, splitting the exit edge may fail.
    423   if (!currentLoop->hasDedicatedExits())
    424     return false;
    425 
    426   LLVMContext &Context = loopHeader->getContext();
    427 
    428   // Probably we reach the quota of branches for this loop. If so
    429   // stop unswitching.
    430   if (!BranchesInfo.countLoop(currentLoop, getAnalysis<TargetTransformInfo>()))
    431     return false;
    432 
    433   // Loop over all of the basic blocks in the loop.  If we find an interior
    434   // block that is branching on a loop-invariant condition, we can unswitch this
    435   // loop.
    436   for (Loop::block_iterator I = currentLoop->block_begin(),
    437          E = currentLoop->block_end(); I != E; ++I) {
    438     TerminatorInst *TI = (*I)->getTerminator();
    439     if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
    440       // If this isn't branching on an invariant condition, we can't unswitch
    441       // it.
    442       if (BI->isConditional()) {
    443         // See if this, or some part of it, is loop invariant.  If so, we can
    444         // unswitch on it if we desire.
    445         Value *LoopCond = FindLIVLoopCondition(BI->getCondition(),
    446                                                currentLoop, Changed);
    447         if (LoopCond && UnswitchIfProfitable(LoopCond,
    448                                              ConstantInt::getTrue(Context))) {
    449           ++NumBranches;
    450           return true;
    451         }
    452       }
    453     } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
    454       Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
    455                                              currentLoop, Changed);
    456       unsigned NumCases = SI->getNumCases();
    457       if (LoopCond && NumCases) {
    458         // Find a value to unswitch on:
    459         // FIXME: this should chose the most expensive case!
    460         // FIXME: scan for a case with a non-critical edge?
    461         Constant *UnswitchVal = 0;
    462 
    463         // Do not process same value again and again.
    464         // At this point we have some cases already unswitched and
    465         // some not yet unswitched. Let's find the first not yet unswitched one.
    466         for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
    467              i != e; ++i) {
    468           Constant *UnswitchValCandidate = i.getCaseValue();
    469           if (!BranchesInfo.isUnswitched(SI, UnswitchValCandidate)) {
    470             UnswitchVal = UnswitchValCandidate;
    471             break;
    472           }
    473         }
    474 
    475         if (!UnswitchVal)
    476           continue;
    477 
    478         if (UnswitchIfProfitable(LoopCond, UnswitchVal)) {
    479           ++NumSwitches;
    480           return true;
    481         }
    482       }
    483     }
    484 
    485     // Scan the instructions to check for unswitchable values.
    486     for (BasicBlock::iterator BBI = (*I)->begin(), E = (*I)->end();
    487          BBI != E; ++BBI)
    488       if (SelectInst *SI = dyn_cast<SelectInst>(BBI)) {
    489         Value *LoopCond = FindLIVLoopCondition(SI->getCondition(),
    490                                                currentLoop, Changed);
    491         if (LoopCond && UnswitchIfProfitable(LoopCond,
    492                                              ConstantInt::getTrue(Context))) {
    493           ++NumSelects;
    494           return true;
    495         }
    496       }
    497   }
    498   return Changed;
    499 }
    500 
    501 /// isTrivialLoopExitBlock - Check to see if all paths from BB exit the
    502 /// loop with no side effects (including infinite loops).
    503 ///
    504 /// If true, we return true and set ExitBB to the block we
    505 /// exit through.
    506 ///
    507 static bool isTrivialLoopExitBlockHelper(Loop *L, BasicBlock *BB,
    508                                          BasicBlock *&ExitBB,
    509                                          std::set<BasicBlock*> &Visited) {
    510   if (!Visited.insert(BB).second) {
    511     // Already visited. Without more analysis, this could indicate an infinite
    512     // loop.
    513     return false;
    514   }
    515   if (!L->contains(BB)) {
    516     // Otherwise, this is a loop exit, this is fine so long as this is the
    517     // first exit.
    518     if (ExitBB != 0) return false;
    519     ExitBB = BB;
    520     return true;
    521   }
    522 
    523   // Otherwise, this is an unvisited intra-loop node.  Check all successors.
    524   for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI) {
    525     // Check to see if the successor is a trivial loop exit.
    526     if (!isTrivialLoopExitBlockHelper(L, *SI, ExitBB, Visited))
    527       return false;
    528   }
    529 
    530   // Okay, everything after this looks good, check to make sure that this block
    531   // doesn't include any side effects.
    532   for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
    533     if (I->mayHaveSideEffects())
    534       return false;
    535 
    536   return true;
    537 }
    538 
    539 /// isTrivialLoopExitBlock - Return true if the specified block unconditionally
    540 /// leads to an exit from the specified loop, and has no side-effects in the
    541 /// process.  If so, return the block that is exited to, otherwise return null.
    542 static BasicBlock *isTrivialLoopExitBlock(Loop *L, BasicBlock *BB) {
    543   std::set<BasicBlock*> Visited;
    544   Visited.insert(L->getHeader());  // Branches to header make infinite loops.
    545   BasicBlock *ExitBB = 0;
    546   if (isTrivialLoopExitBlockHelper(L, BB, ExitBB, Visited))
    547     return ExitBB;
    548   return 0;
    549 }
    550 
    551 /// IsTrivialUnswitchCondition - Check to see if this unswitch condition is
    552 /// trivial: that is, that the condition controls whether or not the loop does
    553 /// anything at all.  If this is a trivial condition, unswitching produces no
    554 /// code duplications (equivalently, it produces a simpler loop and a new empty
    555 /// loop, which gets deleted).
    556 ///
    557 /// If this is a trivial condition, return true, otherwise return false.  When
    558 /// returning true, this sets Cond and Val to the condition that controls the
    559 /// trivial condition: when Cond dynamically equals Val, the loop is known to
    560 /// exit.  Finally, this sets LoopExit to the BB that the loop exits to when
    561 /// Cond == Val.
    562 ///
    563 bool LoopUnswitch::IsTrivialUnswitchCondition(Value *Cond, Constant **Val,
    564                                        BasicBlock **LoopExit) {
    565   BasicBlock *Header = currentLoop->getHeader();
    566   TerminatorInst *HeaderTerm = Header->getTerminator();
    567   LLVMContext &Context = Header->getContext();
    568 
    569   BasicBlock *LoopExitBB = 0;
    570   if (BranchInst *BI = dyn_cast<BranchInst>(HeaderTerm)) {
    571     // If the header block doesn't end with a conditional branch on Cond, we
    572     // can't handle it.
    573     if (!BI->isConditional() || BI->getCondition() != Cond)
    574       return false;
    575 
    576     // Check to see if a successor of the branch is guaranteed to
    577     // exit through a unique exit block without having any
    578     // side-effects.  If so, determine the value of Cond that causes it to do
    579     // this.
    580     if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
    581                                              BI->getSuccessor(0)))) {
    582       if (Val) *Val = ConstantInt::getTrue(Context);
    583     } else if ((LoopExitBB = isTrivialLoopExitBlock(currentLoop,
    584                                                     BI->getSuccessor(1)))) {
    585       if (Val) *Val = ConstantInt::getFalse(Context);
    586     }
    587   } else if (SwitchInst *SI = dyn_cast<SwitchInst>(HeaderTerm)) {
    588     // If this isn't a switch on Cond, we can't handle it.
    589     if (SI->getCondition() != Cond) return false;
    590 
    591     // Check to see if a successor of the switch is guaranteed to go to the
    592     // latch block or exit through a one exit block without having any
    593     // side-effects.  If so, determine the value of Cond that causes it to do
    594     // this.
    595     // Note that we can't trivially unswitch on the default case or
    596     // on already unswitched cases.
    597     for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
    598          i != e; ++i) {
    599       BasicBlock *LoopExitCandidate;
    600       if ((LoopExitCandidate = isTrivialLoopExitBlock(currentLoop,
    601                                                i.getCaseSuccessor()))) {
    602         // Okay, we found a trivial case, remember the value that is trivial.
    603         ConstantInt *CaseVal = i.getCaseValue();
    604 
    605         // Check that it was not unswitched before, since already unswitched
    606         // trivial vals are looks trivial too.
    607         if (BranchesInfo.isUnswitched(SI, CaseVal))
    608           continue;
    609         LoopExitBB = LoopExitCandidate;
    610         if (Val) *Val = CaseVal;
    611         break;
    612       }
    613     }
    614   }
    615 
    616   // If we didn't find a single unique LoopExit block, or if the loop exit block
    617   // contains phi nodes, this isn't trivial.
    618   if (!LoopExitBB || isa<PHINode>(LoopExitBB->begin()))
    619     return false;   // Can't handle this.
    620 
    621   if (LoopExit) *LoopExit = LoopExitBB;
    622 
    623   // We already know that nothing uses any scalar values defined inside of this
    624   // loop.  As such, we just have to check to see if this loop will execute any
    625   // side-effecting instructions (e.g. stores, calls, volatile loads) in the
    626   // part of the loop that the code *would* execute.  We already checked the
    627   // tail, check the header now.
    628   for (BasicBlock::iterator I = Header->begin(), E = Header->end(); I != E; ++I)
    629     if (I->mayHaveSideEffects())
    630       return false;
    631   return true;
    632 }
    633 
    634 /// UnswitchIfProfitable - We have found that we can unswitch currentLoop when
    635 /// LoopCond == Val to simplify the loop.  If we decide that this is profitable,
    636 /// unswitch the loop, reprocess the pieces, then return true.
    637 bool LoopUnswitch::UnswitchIfProfitable(Value *LoopCond, Constant *Val) {
    638   Function *F = loopHeader->getParent();
    639   Constant *CondVal = 0;
    640   BasicBlock *ExitBlock = 0;
    641 
    642   if (IsTrivialUnswitchCondition(LoopCond, &CondVal, &ExitBlock)) {
    643     // If the condition is trivial, always unswitch. There is no code growth
    644     // for this case.
    645     UnswitchTrivialCondition(currentLoop, LoopCond, CondVal, ExitBlock);
    646     return true;
    647   }
    648 
    649   // Check to see if it would be profitable to unswitch current loop.
    650 
    651   // Do not do non-trivial unswitch while optimizing for size.
    652   if (OptimizeForSize ||
    653       F->getAttributes().hasAttribute(AttributeSet::FunctionIndex,
    654                                       Attribute::OptimizeForSize))
    655     return false;
    656 
    657   UnswitchNontrivialCondition(LoopCond, Val, currentLoop);
    658   return true;
    659 }
    660 
    661 /// CloneLoop - Recursively clone the specified loop and all of its children,
    662 /// mapping the blocks with the specified map.
    663 static Loop *CloneLoop(Loop *L, Loop *PL, ValueToValueMapTy &VM,
    664                        LoopInfo *LI, LPPassManager *LPM) {
    665   Loop *New = new Loop();
    666   LPM->insertLoop(New, PL);
    667 
    668   // Add all of the blocks in L to the new loop.
    669   for (Loop::block_iterator I = L->block_begin(), E = L->block_end();
    670        I != E; ++I)
    671     if (LI->getLoopFor(*I) == L)
    672       New->addBasicBlockToLoop(cast<BasicBlock>(VM[*I]), LI->getBase());
    673 
    674   // Add all of the subloops to the new loop.
    675   for (Loop::iterator I = L->begin(), E = L->end(); I != E; ++I)
    676     CloneLoop(*I, New, VM, LI, LPM);
    677 
    678   return New;
    679 }
    680 
    681 /// EmitPreheaderBranchOnCondition - Emit a conditional branch on two values
    682 /// if LIC == Val, branch to TrueDst, otherwise branch to FalseDest.  Insert the
    683 /// code immediately before InsertPt.
    684 void LoopUnswitch::EmitPreheaderBranchOnCondition(Value *LIC, Constant *Val,
    685                                                   BasicBlock *TrueDest,
    686                                                   BasicBlock *FalseDest,
    687                                                   Instruction *InsertPt) {
    688   // Insert a conditional branch on LIC to the two preheaders.  The original
    689   // code is the true version and the new code is the false version.
    690   Value *BranchVal = LIC;
    691   if (!isa<ConstantInt>(Val) ||
    692       Val->getType() != Type::getInt1Ty(LIC->getContext()))
    693     BranchVal = new ICmpInst(InsertPt, ICmpInst::ICMP_EQ, LIC, Val);
    694   else if (Val != ConstantInt::getTrue(Val->getContext()))
    695     // We want to enter the new loop when the condition is true.
    696     std::swap(TrueDest, FalseDest);
    697 
    698   // Insert the new branch.
    699   BranchInst *BI = BranchInst::Create(TrueDest, FalseDest, BranchVal, InsertPt);
    700 
    701   // If either edge is critical, split it. This helps preserve LoopSimplify
    702   // form for enclosing loops.
    703   SplitCriticalEdge(BI, 0, this, false, false, true);
    704   SplitCriticalEdge(BI, 1, this, false, false, true);
    705 }
    706 
    707 /// UnswitchTrivialCondition - Given a loop that has a trivial unswitchable
    708 /// condition in it (a cond branch from its header block to its latch block,
    709 /// where the path through the loop that doesn't execute its body has no
    710 /// side-effects), unswitch it.  This doesn't involve any code duplication, just
    711 /// moving the conditional branch outside of the loop and updating loop info.
    712 void LoopUnswitch::UnswitchTrivialCondition(Loop *L, Value *Cond,
    713                                             Constant *Val,
    714                                             BasicBlock *ExitBlock) {
    715   DEBUG(dbgs() << "loop-unswitch: Trivial-Unswitch loop %"
    716         << loopHeader->getName() << " [" << L->getBlocks().size()
    717         << " blocks] in Function " << L->getHeader()->getParent()->getName()
    718         << " on cond: " << *Val << " == " << *Cond << "\n");
    719 
    720   // First step, split the preheader, so that we know that there is a safe place
    721   // to insert the conditional branch.  We will change loopPreheader to have a
    722   // conditional branch on Cond.
    723   BasicBlock *NewPH = SplitEdge(loopPreheader, loopHeader, this);
    724 
    725   // Now that we have a place to insert the conditional branch, create a place
    726   // to branch to: this is the exit block out of the loop that we should
    727   // short-circuit to.
    728 
    729   // Split this block now, so that the loop maintains its exit block, and so
    730   // that the jump from the preheader can execute the contents of the exit block
    731   // without actually branching to it (the exit block should be dominated by the
    732   // loop header, not the preheader).
    733   assert(!L->contains(ExitBlock) && "Exit block is in the loop?");
    734   BasicBlock *NewExit = SplitBlock(ExitBlock, ExitBlock->begin(), this);
    735 
    736   // Okay, now we have a position to branch from and a position to branch to,
    737   // insert the new conditional branch.
    738   EmitPreheaderBranchOnCondition(Cond, Val, NewExit, NewPH,
    739                                  loopPreheader->getTerminator());
    740   LPM->deleteSimpleAnalysisValue(loopPreheader->getTerminator(), L);
    741   loopPreheader->getTerminator()->eraseFromParent();
    742 
    743   // We need to reprocess this loop, it could be unswitched again.
    744   redoLoop = true;
    745 
    746   // Now that we know that the loop is never entered when this condition is a
    747   // particular value, rewrite the loop with this info.  We know that this will
    748   // at least eliminate the old branch.
    749   RewriteLoopBodyWithConditionConstant(L, Cond, Val, false);
    750   ++NumTrivial;
    751 }
    752 
    753 /// SplitExitEdges - Split all of the edges from inside the loop to their exit
    754 /// blocks.  Update the appropriate Phi nodes as we do so.
    755 void LoopUnswitch::SplitExitEdges(Loop *L,
    756                                const SmallVectorImpl<BasicBlock *> &ExitBlocks){
    757 
    758   for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
    759     BasicBlock *ExitBlock = ExitBlocks[i];
    760     SmallVector<BasicBlock *, 4> Preds(pred_begin(ExitBlock),
    761                                        pred_end(ExitBlock));
    762 
    763     // Although SplitBlockPredecessors doesn't preserve loop-simplify in
    764     // general, if we call it on all predecessors of all exits then it does.
    765     if (!ExitBlock->isLandingPad()) {
    766       SplitBlockPredecessors(ExitBlock, Preds, ".us-lcssa", this);
    767     } else {
    768       SmallVector<BasicBlock*, 2> NewBBs;
    769       SplitLandingPadPredecessors(ExitBlock, Preds, ".us-lcssa", ".us-lcssa",
    770                                   this, NewBBs);
    771     }
    772   }
    773 }
    774 
    775 /// UnswitchNontrivialCondition - We determined that the loop is profitable
    776 /// to unswitch when LIC equal Val.  Split it into loop versions and test the
    777 /// condition outside of either loop.  Return the loops created as Out1/Out2.
    778 void LoopUnswitch::UnswitchNontrivialCondition(Value *LIC, Constant *Val,
    779                                                Loop *L) {
    780   Function *F = loopHeader->getParent();
    781   DEBUG(dbgs() << "loop-unswitch: Unswitching loop %"
    782         << loopHeader->getName() << " [" << L->getBlocks().size()
    783         << " blocks] in Function " << F->getName()
    784         << " when '" << *Val << "' == " << *LIC << "\n");
    785 
    786   if (ScalarEvolution *SE = getAnalysisIfAvailable<ScalarEvolution>())
    787     SE->forgetLoop(L);
    788 
    789   LoopBlocks.clear();
    790   NewBlocks.clear();
    791 
    792   // First step, split the preheader and exit blocks, and add these blocks to
    793   // the LoopBlocks list.
    794   BasicBlock *NewPreheader = SplitEdge(loopPreheader, loopHeader, this);
    795   LoopBlocks.push_back(NewPreheader);
    796 
    797   // We want the loop to come after the preheader, but before the exit blocks.
    798   LoopBlocks.insert(LoopBlocks.end(), L->block_begin(), L->block_end());
    799 
    800   SmallVector<BasicBlock*, 8> ExitBlocks;
    801   L->getUniqueExitBlocks(ExitBlocks);
    802 
    803   // Split all of the edges from inside the loop to their exit blocks.  Update
    804   // the appropriate Phi nodes as we do so.
    805   SplitExitEdges(L, ExitBlocks);
    806 
    807   // The exit blocks may have been changed due to edge splitting, recompute.
    808   ExitBlocks.clear();
    809   L->getUniqueExitBlocks(ExitBlocks);
    810 
    811   // Add exit blocks to the loop blocks.
    812   LoopBlocks.insert(LoopBlocks.end(), ExitBlocks.begin(), ExitBlocks.end());
    813 
    814   // Next step, clone all of the basic blocks that make up the loop (including
    815   // the loop preheader and exit blocks), keeping track of the mapping between
    816   // the instructions and blocks.
    817   NewBlocks.reserve(LoopBlocks.size());
    818   ValueToValueMapTy VMap;
    819   for (unsigned i = 0, e = LoopBlocks.size(); i != e; ++i) {
    820     BasicBlock *NewBB = CloneBasicBlock(LoopBlocks[i], VMap, ".us", F);
    821 
    822     NewBlocks.push_back(NewBB);
    823     VMap[LoopBlocks[i]] = NewBB;  // Keep the BB mapping.
    824     LPM->cloneBasicBlockSimpleAnalysis(LoopBlocks[i], NewBB, L);
    825   }
    826 
    827   // Splice the newly inserted blocks into the function right before the
    828   // original preheader.
    829   F->getBasicBlockList().splice(NewPreheader, F->getBasicBlockList(),
    830                                 NewBlocks[0], F->end());
    831 
    832   // Now we create the new Loop object for the versioned loop.
    833   Loop *NewLoop = CloneLoop(L, L->getParentLoop(), VMap, LI, LPM);
    834 
    835   // Recalculate unswitching quota, inherit simplified switches info for NewBB,
    836   // Probably clone more loop-unswitch related loop properties.
    837   BranchesInfo.cloneData(NewLoop, L, VMap);
    838 
    839   Loop *ParentLoop = L->getParentLoop();
    840   if (ParentLoop) {
    841     // Make sure to add the cloned preheader and exit blocks to the parent loop
    842     // as well.
    843     ParentLoop->addBasicBlockToLoop(NewBlocks[0], LI->getBase());
    844   }
    845 
    846   for (unsigned i = 0, e = ExitBlocks.size(); i != e; ++i) {
    847     BasicBlock *NewExit = cast<BasicBlock>(VMap[ExitBlocks[i]]);
    848     // The new exit block should be in the same loop as the old one.
    849     if (Loop *ExitBBLoop = LI->getLoopFor(ExitBlocks[i]))
    850       ExitBBLoop->addBasicBlockToLoop(NewExit, LI->getBase());
    851 
    852     assert(NewExit->getTerminator()->getNumSuccessors() == 1 &&
    853            "Exit block should have been split to have one successor!");
    854     BasicBlock *ExitSucc = NewExit->getTerminator()->getSuccessor(0);
    855 
    856     // If the successor of the exit block had PHI nodes, add an entry for
    857     // NewExit.
    858     for (BasicBlock::iterator I = ExitSucc->begin();
    859          PHINode *PN = dyn_cast<PHINode>(I); ++I) {
    860       Value *V = PN->getIncomingValueForBlock(ExitBlocks[i]);
    861       ValueToValueMapTy::iterator It = VMap.find(V);
    862       if (It != VMap.end()) V = It->second;
    863       PN->addIncoming(V, NewExit);
    864     }
    865 
    866     if (LandingPadInst *LPad = NewExit->getLandingPadInst()) {
    867       PHINode *PN = PHINode::Create(LPad->getType(), 0, "",
    868                                     ExitSucc->getFirstInsertionPt());
    869 
    870       for (pred_iterator I = pred_begin(ExitSucc), E = pred_end(ExitSucc);
    871            I != E; ++I) {
    872         BasicBlock *BB = *I;
    873         LandingPadInst *LPI = BB->getLandingPadInst();
    874         LPI->replaceAllUsesWith(PN);
    875         PN->addIncoming(LPI, BB);
    876       }
    877     }
    878   }
    879 
    880   // Rewrite the code to refer to itself.
    881   for (unsigned i = 0, e = NewBlocks.size(); i != e; ++i)
    882     for (BasicBlock::iterator I = NewBlocks[i]->begin(),
    883            E = NewBlocks[i]->end(); I != E; ++I)
    884       RemapInstruction(I, VMap,RF_NoModuleLevelChanges|RF_IgnoreMissingEntries);
    885 
    886   // Rewrite the original preheader to select between versions of the loop.
    887   BranchInst *OldBR = cast<BranchInst>(loopPreheader->getTerminator());
    888   assert(OldBR->isUnconditional() && OldBR->getSuccessor(0) == LoopBlocks[0] &&
    889          "Preheader splitting did not work correctly!");
    890 
    891   // Emit the new branch that selects between the two versions of this loop.
    892   EmitPreheaderBranchOnCondition(LIC, Val, NewBlocks[0], LoopBlocks[0], OldBR);
    893   LPM->deleteSimpleAnalysisValue(OldBR, L);
    894   OldBR->eraseFromParent();
    895 
    896   LoopProcessWorklist.push_back(NewLoop);
    897   redoLoop = true;
    898 
    899   // Keep a WeakVH holding onto LIC.  If the first call to RewriteLoopBody
    900   // deletes the instruction (for example by simplifying a PHI that feeds into
    901   // the condition that we're unswitching on), we don't rewrite the second
    902   // iteration.
    903   WeakVH LICHandle(LIC);
    904 
    905   // Now we rewrite the original code to know that the condition is true and the
    906   // new code to know that the condition is false.
    907   RewriteLoopBodyWithConditionConstant(L, LIC, Val, false);
    908 
    909   // It's possible that simplifying one loop could cause the other to be
    910   // changed to another value or a constant.  If its a constant, don't simplify
    911   // it.
    912   if (!LoopProcessWorklist.empty() && LoopProcessWorklist.back() == NewLoop &&
    913       LICHandle && !isa<Constant>(LICHandle))
    914     RewriteLoopBodyWithConditionConstant(NewLoop, LICHandle, Val, true);
    915 }
    916 
    917 /// RemoveFromWorklist - Remove all instances of I from the worklist vector
    918 /// specified.
    919 static void RemoveFromWorklist(Instruction *I,
    920                                std::vector<Instruction*> &Worklist) {
    921 
    922   Worklist.erase(std::remove(Worklist.begin(), Worklist.end(), I),
    923                  Worklist.end());
    924 }
    925 
    926 /// ReplaceUsesOfWith - When we find that I really equals V, remove I from the
    927 /// program, replacing all uses with V and update the worklist.
    928 static void ReplaceUsesOfWith(Instruction *I, Value *V,
    929                               std::vector<Instruction*> &Worklist,
    930                               Loop *L, LPPassManager *LPM) {
    931   DEBUG(dbgs() << "Replace with '" << *V << "': " << *I);
    932 
    933   // Add uses to the worklist, which may be dead now.
    934   for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
    935     if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
    936       Worklist.push_back(Use);
    937 
    938   // Add users to the worklist which may be simplified now.
    939   for (Value::use_iterator UI = I->use_begin(), E = I->use_end();
    940        UI != E; ++UI)
    941     Worklist.push_back(cast<Instruction>(*UI));
    942   LPM->deleteSimpleAnalysisValue(I, L);
    943   RemoveFromWorklist(I, Worklist);
    944   I->replaceAllUsesWith(V);
    945   I->eraseFromParent();
    946   ++NumSimplify;
    947 }
    948 
    949 /// RemoveBlockIfDead - If the specified block is dead, remove it, update loop
    950 /// information, and remove any dead successors it has.
    951 ///
    952 void LoopUnswitch::RemoveBlockIfDead(BasicBlock *BB,
    953                                      std::vector<Instruction*> &Worklist,
    954                                      Loop *L) {
    955   if (pred_begin(BB) != pred_end(BB)) {
    956     // This block isn't dead, since an edge to BB was just removed, see if there
    957     // are any easy simplifications we can do now.
    958     if (BasicBlock *Pred = BB->getSinglePredecessor()) {
    959       // If it has one pred, fold phi nodes in BB.
    960       while (PHINode *PN = dyn_cast<PHINode>(BB->begin()))
    961         ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
    962 
    963       // If this is the header of a loop and the only pred is the latch, we now
    964       // have an unreachable loop.
    965       if (Loop *L = LI->getLoopFor(BB))
    966         if (loopHeader == BB && L->contains(Pred)) {
    967           // Remove the branch from the latch to the header block, this makes
    968           // the header dead, which will make the latch dead (because the header
    969           // dominates the latch).
    970           LPM->deleteSimpleAnalysisValue(Pred->getTerminator(), L);
    971           Pred->getTerminator()->eraseFromParent();
    972           new UnreachableInst(BB->getContext(), Pred);
    973 
    974           // The loop is now broken, remove it from LI.
    975           RemoveLoopFromHierarchy(L);
    976 
    977           // Reprocess the header, which now IS dead.
    978           RemoveBlockIfDead(BB, Worklist, L);
    979           return;
    980         }
    981 
    982       // If pred ends in a uncond branch, add uncond branch to worklist so that
    983       // the two blocks will get merged.
    984       if (BranchInst *BI = dyn_cast<BranchInst>(Pred->getTerminator()))
    985         if (BI->isUnconditional())
    986           Worklist.push_back(BI);
    987     }
    988     return;
    989   }
    990 
    991   DEBUG(dbgs() << "Nuking dead block: " << *BB);
    992 
    993   // Remove the instructions in the basic block from the worklist.
    994   for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) {
    995     RemoveFromWorklist(I, Worklist);
    996 
    997     // Anything that uses the instructions in this basic block should have their
    998     // uses replaced with undefs.
    999     // If I is not void type then replaceAllUsesWith undef.
   1000     // This allows ValueHandlers and custom metadata to adjust itself.
   1001     if (!I->getType()->isVoidTy())
   1002       I->replaceAllUsesWith(UndefValue::get(I->getType()));
   1003   }
   1004 
   1005   // If this is the edge to the header block for a loop, remove the loop and
   1006   // promote all subloops.
   1007   if (Loop *BBLoop = LI->getLoopFor(BB)) {
   1008     if (BBLoop->getLoopLatch() == BB) {
   1009       RemoveLoopFromHierarchy(BBLoop);
   1010       if (currentLoop == BBLoop) {
   1011         currentLoop = 0;
   1012         redoLoop = false;
   1013       }
   1014     }
   1015   }
   1016 
   1017   // Remove the block from the loop info, which removes it from any loops it
   1018   // was in.
   1019   LI->removeBlock(BB);
   1020 
   1021   // Remove phi node entries in successors for this block.
   1022   TerminatorInst *TI = BB->getTerminator();
   1023   SmallVector<BasicBlock*, 4> Succs;
   1024   for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i) {
   1025     Succs.push_back(TI->getSuccessor(i));
   1026     TI->getSuccessor(i)->removePredecessor(BB);
   1027   }
   1028 
   1029   // Unique the successors, remove anything with multiple uses.
   1030   array_pod_sort(Succs.begin(), Succs.end());
   1031   Succs.erase(std::unique(Succs.begin(), Succs.end()), Succs.end());
   1032 
   1033   // Remove the basic block, including all of the instructions contained in it.
   1034   LPM->deleteSimpleAnalysisValue(BB, L);
   1035   BB->eraseFromParent();
   1036   // Remove successor blocks here that are not dead, so that we know we only
   1037   // have dead blocks in this list.  Nondead blocks have a way of becoming dead,
   1038   // then getting removed before we revisit them, which is badness.
   1039   //
   1040   for (unsigned i = 0; i != Succs.size(); ++i)
   1041     if (pred_begin(Succs[i]) != pred_end(Succs[i])) {
   1042       // One exception is loop headers.  If this block was the preheader for a
   1043       // loop, then we DO want to visit the loop so the loop gets deleted.
   1044       // We know that if the successor is a loop header, that this loop had to
   1045       // be the preheader: the case where this was the latch block was handled
   1046       // above and headers can only have two predecessors.
   1047       if (!LI->isLoopHeader(Succs[i])) {
   1048         Succs.erase(Succs.begin()+i);
   1049         --i;
   1050       }
   1051     }
   1052 
   1053   for (unsigned i = 0, e = Succs.size(); i != e; ++i)
   1054     RemoveBlockIfDead(Succs[i], Worklist, L);
   1055 }
   1056 
   1057 /// RemoveLoopFromHierarchy - We have discovered that the specified loop has
   1058 /// become unwrapped, either because the backedge was deleted, or because the
   1059 /// edge into the header was removed.  If the edge into the header from the
   1060 /// latch block was removed, the loop is unwrapped but subloops are still alive,
   1061 /// so they just reparent loops.  If the loops are actually dead, they will be
   1062 /// removed later.
   1063 void LoopUnswitch::RemoveLoopFromHierarchy(Loop *L) {
   1064   LPM->deleteLoopFromQueue(L);
   1065   RemoveLoopFromWorklist(L);
   1066 }
   1067 
   1068 // RewriteLoopBodyWithConditionConstant - We know either that the value LIC has
   1069 // the value specified by Val in the specified loop, or we know it does NOT have
   1070 // that value.  Rewrite any uses of LIC or of properties correlated to it.
   1071 void LoopUnswitch::RewriteLoopBodyWithConditionConstant(Loop *L, Value *LIC,
   1072                                                         Constant *Val,
   1073                                                         bool IsEqual) {
   1074   assert(!isa<Constant>(LIC) && "Why are we unswitching on a constant?");
   1075 
   1076   // FIXME: Support correlated properties, like:
   1077   //  for (...)
   1078   //    if (li1 < li2)
   1079   //      ...
   1080   //    if (li1 > li2)
   1081   //      ...
   1082 
   1083   // FOLD boolean conditions (X|LIC), (X&LIC).  Fold conditional branches,
   1084   // selects, switches.
   1085   std::vector<Instruction*> Worklist;
   1086   LLVMContext &Context = Val->getContext();
   1087 
   1088   // If we know that LIC == Val, or that LIC == NotVal, just replace uses of LIC
   1089   // in the loop with the appropriate one directly.
   1090   if (IsEqual || (isa<ConstantInt>(Val) &&
   1091       Val->getType()->isIntegerTy(1))) {
   1092     Value *Replacement;
   1093     if (IsEqual)
   1094       Replacement = Val;
   1095     else
   1096       Replacement = ConstantInt::get(Type::getInt1Ty(Val->getContext()),
   1097                                      !cast<ConstantInt>(Val)->getZExtValue());
   1098 
   1099     for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end();
   1100          UI != E; ++UI) {
   1101       Instruction *U = dyn_cast<Instruction>(*UI);
   1102       if (!U || !L->contains(U))
   1103         continue;
   1104       Worklist.push_back(U);
   1105     }
   1106 
   1107     for (std::vector<Instruction*>::iterator UI = Worklist.begin(),
   1108          UE = Worklist.end(); UI != UE; ++UI)
   1109       (*UI)->replaceUsesOfWith(LIC, Replacement);
   1110 
   1111     SimplifyCode(Worklist, L);
   1112     return;
   1113   }
   1114 
   1115   // Otherwise, we don't know the precise value of LIC, but we do know that it
   1116   // is certainly NOT "Val".  As such, simplify any uses in the loop that we
   1117   // can.  This case occurs when we unswitch switch statements.
   1118   for (Value::use_iterator UI = LIC->use_begin(), E = LIC->use_end();
   1119        UI != E; ++UI) {
   1120     Instruction *U = dyn_cast<Instruction>(*UI);
   1121     if (!U || !L->contains(U))
   1122       continue;
   1123 
   1124     Worklist.push_back(U);
   1125 
   1126     // TODO: We could do other simplifications, for example, turning
   1127     // 'icmp eq LIC, Val' -> false.
   1128 
   1129     // If we know that LIC is not Val, use this info to simplify code.
   1130     SwitchInst *SI = dyn_cast<SwitchInst>(U);
   1131     if (SI == 0 || !isa<ConstantInt>(Val)) continue;
   1132 
   1133     SwitchInst::CaseIt DeadCase = SI->findCaseValue(cast<ConstantInt>(Val));
   1134     // Default case is live for multiple values.
   1135     if (DeadCase == SI->case_default()) continue;
   1136 
   1137     // Found a dead case value.  Don't remove PHI nodes in the
   1138     // successor if they become single-entry, those PHI nodes may
   1139     // be in the Users list.
   1140 
   1141     BasicBlock *Switch = SI->getParent();
   1142     BasicBlock *SISucc = DeadCase.getCaseSuccessor();
   1143     BasicBlock *Latch = L->getLoopLatch();
   1144 
   1145     BranchesInfo.setUnswitched(SI, Val);
   1146 
   1147     if (!SI->findCaseDest(SISucc)) continue;  // Edge is critical.
   1148     // If the DeadCase successor dominates the loop latch, then the
   1149     // transformation isn't safe since it will delete the sole predecessor edge
   1150     // to the latch.
   1151     if (Latch && DT->dominates(SISucc, Latch))
   1152       continue;
   1153 
   1154     // FIXME: This is a hack.  We need to keep the successor around
   1155     // and hooked up so as to preserve the loop structure, because
   1156     // trying to update it is complicated.  So instead we preserve the
   1157     // loop structure and put the block on a dead code path.
   1158     SplitEdge(Switch, SISucc, this);
   1159     // Compute the successors instead of relying on the return value
   1160     // of SplitEdge, since it may have split the switch successor
   1161     // after PHI nodes.
   1162     BasicBlock *NewSISucc = DeadCase.getCaseSuccessor();
   1163     BasicBlock *OldSISucc = *succ_begin(NewSISucc);
   1164     // Create an "unreachable" destination.
   1165     BasicBlock *Abort = BasicBlock::Create(Context, "us-unreachable",
   1166                                            Switch->getParent(),
   1167                                            OldSISucc);
   1168     new UnreachableInst(Context, Abort);
   1169     // Force the new case destination to branch to the "unreachable"
   1170     // block while maintaining a (dead) CFG edge to the old block.
   1171     NewSISucc->getTerminator()->eraseFromParent();
   1172     BranchInst::Create(Abort, OldSISucc,
   1173                        ConstantInt::getTrue(Context), NewSISucc);
   1174     // Release the PHI operands for this edge.
   1175     for (BasicBlock::iterator II = NewSISucc->begin();
   1176          PHINode *PN = dyn_cast<PHINode>(II); ++II)
   1177       PN->setIncomingValue(PN->getBasicBlockIndex(Switch),
   1178                            UndefValue::get(PN->getType()));
   1179     // Tell the domtree about the new block. We don't fully update the
   1180     // domtree here -- instead we force it to do a full recomputation
   1181     // after the pass is complete -- but we do need to inform it of
   1182     // new blocks.
   1183     if (DT)
   1184       DT->addNewBlock(Abort, NewSISucc);
   1185   }
   1186 
   1187   SimplifyCode(Worklist, L);
   1188 }
   1189 
   1190 /// SimplifyCode - Okay, now that we have simplified some instructions in the
   1191 /// loop, walk over it and constant prop, dce, and fold control flow where
   1192 /// possible.  Note that this is effectively a very simple loop-structure-aware
   1193 /// optimizer.  During processing of this loop, L could very well be deleted, so
   1194 /// it must not be used.
   1195 ///
   1196 /// FIXME: When the loop optimizer is more mature, separate this out to a new
   1197 /// pass.
   1198 ///
   1199 void LoopUnswitch::SimplifyCode(std::vector<Instruction*> &Worklist, Loop *L) {
   1200   while (!Worklist.empty()) {
   1201     Instruction *I = Worklist.back();
   1202     Worklist.pop_back();
   1203 
   1204     // Simple DCE.
   1205     if (isInstructionTriviallyDead(I)) {
   1206       DEBUG(dbgs() << "Remove dead instruction '" << *I);
   1207 
   1208       // Add uses to the worklist, which may be dead now.
   1209       for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
   1210         if (Instruction *Use = dyn_cast<Instruction>(I->getOperand(i)))
   1211           Worklist.push_back(Use);
   1212       LPM->deleteSimpleAnalysisValue(I, L);
   1213       RemoveFromWorklist(I, Worklist);
   1214       I->eraseFromParent();
   1215       ++NumSimplify;
   1216       continue;
   1217     }
   1218 
   1219     // See if instruction simplification can hack this up.  This is common for
   1220     // things like "select false, X, Y" after unswitching made the condition be
   1221     // 'false'.  TODO: update the domtree properly so we can pass it here.
   1222     if (Value *V = SimplifyInstruction(I))
   1223       if (LI->replacementPreservesLCSSAForm(I, V)) {
   1224         ReplaceUsesOfWith(I, V, Worklist, L, LPM);
   1225         continue;
   1226       }
   1227 
   1228     // Special case hacks that appear commonly in unswitched code.
   1229     if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
   1230       if (BI->isUnconditional()) {
   1231         // If BI's parent is the only pred of the successor, fold the two blocks
   1232         // together.
   1233         BasicBlock *Pred = BI->getParent();
   1234         BasicBlock *Succ = BI->getSuccessor(0);
   1235         BasicBlock *SinglePred = Succ->getSinglePredecessor();
   1236         if (!SinglePred) continue;  // Nothing to do.
   1237         assert(SinglePred == Pred && "CFG broken");
   1238 
   1239         DEBUG(dbgs() << "Merging blocks: " << Pred->getName() << " <- "
   1240               << Succ->getName() << "\n");
   1241 
   1242         // Resolve any single entry PHI nodes in Succ.
   1243         while (PHINode *PN = dyn_cast<PHINode>(Succ->begin()))
   1244           ReplaceUsesOfWith(PN, PN->getIncomingValue(0), Worklist, L, LPM);
   1245 
   1246         // If Succ has any successors with PHI nodes, update them to have
   1247         // entries coming from Pred instead of Succ.
   1248         Succ->replaceAllUsesWith(Pred);
   1249 
   1250         // Move all of the successor contents from Succ to Pred.
   1251         Pred->getInstList().splice(BI, Succ->getInstList(), Succ->begin(),
   1252                                    Succ->end());
   1253         LPM->deleteSimpleAnalysisValue(BI, L);
   1254         BI->eraseFromParent();
   1255         RemoveFromWorklist(BI, Worklist);
   1256 
   1257         // Remove Succ from the loop tree.
   1258         LI->removeBlock(Succ);
   1259         LPM->deleteSimpleAnalysisValue(Succ, L);
   1260         Succ->eraseFromParent();
   1261         ++NumSimplify;
   1262         continue;
   1263       }
   1264 
   1265       if (ConstantInt *CB = dyn_cast<ConstantInt>(BI->getCondition())){
   1266         // Conditional branch.  Turn it into an unconditional branch, then
   1267         // remove dead blocks.
   1268         continue;  // FIXME: Enable.
   1269 
   1270         DEBUG(dbgs() << "Folded branch: " << *BI);
   1271         BasicBlock *DeadSucc = BI->getSuccessor(CB->getZExtValue());
   1272         BasicBlock *LiveSucc = BI->getSuccessor(!CB->getZExtValue());
   1273         DeadSucc->removePredecessor(BI->getParent(), true);
   1274         Worklist.push_back(BranchInst::Create(LiveSucc, BI));
   1275         LPM->deleteSimpleAnalysisValue(BI, L);
   1276         BI->eraseFromParent();
   1277         RemoveFromWorklist(BI, Worklist);
   1278         ++NumSimplify;
   1279 
   1280         RemoveBlockIfDead(DeadSucc, Worklist, L);
   1281       }
   1282       continue;
   1283     }
   1284   }
   1285 }
   1286