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