LLVM  mainline
GVN.cpp
Go to the documentation of this file.
00001 //===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
00002 //
00003 //                     The LLVM Compiler Infrastructure
00004 //
00005 // This file is distributed under the University of Illinois Open Source
00006 // License. See LICENSE.TXT for details.
00007 //
00008 //===----------------------------------------------------------------------===//
00009 //
00010 // This pass performs global value numbering to eliminate fully redundant
00011 // instructions.  It also performs simple dead load elimination.
00012 //
00013 // Note that this pass does the value numbering itself; it does not use the
00014 // ValueNumbering analysis passes.
00015 //
00016 //===----------------------------------------------------------------------===//
00017 
00018 #include "llvm/Transforms/Scalar.h"
00019 #include "llvm/ADT/DenseMap.h"
00020 #include "llvm/ADT/DepthFirstIterator.h"
00021 #include "llvm/ADT/Hashing.h"
00022 #include "llvm/ADT/MapVector.h"
00023 #include "llvm/ADT/PostOrderIterator.h"
00024 #include "llvm/ADT/SetVector.h"
00025 #include "llvm/ADT/SmallPtrSet.h"
00026 #include "llvm/ADT/Statistic.h"
00027 #include "llvm/Analysis/AliasAnalysis.h"
00028 #include "llvm/Analysis/AssumptionCache.h"
00029 #include "llvm/Analysis/CFG.h"
00030 #include "llvm/Analysis/ConstantFolding.h"
00031 #include "llvm/Analysis/InstructionSimplify.h"
00032 #include "llvm/Analysis/Loads.h"
00033 #include "llvm/Analysis/MemoryBuiltins.h"
00034 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
00035 #include "llvm/Analysis/PHITransAddr.h"
00036 #include "llvm/Analysis/TargetLibraryInfo.h"
00037 #include "llvm/Analysis/ValueTracking.h"
00038 #include "llvm/IR/DataLayout.h"
00039 #include "llvm/IR/Dominators.h"
00040 #include "llvm/IR/GlobalVariable.h"
00041 #include "llvm/IR/IRBuilder.h"
00042 #include "llvm/IR/IntrinsicInst.h"
00043 #include "llvm/IR/LLVMContext.h"
00044 #include "llvm/IR/Metadata.h"
00045 #include "llvm/IR/PatternMatch.h"
00046 #include "llvm/Support/Allocator.h"
00047 #include "llvm/Support/CommandLine.h"
00048 #include "llvm/Support/Debug.h"
00049 #include "llvm/Support/raw_ostream.h"
00050 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
00051 #include "llvm/Transforms/Utils/Local.h"
00052 #include "llvm/Transforms/Utils/SSAUpdater.h"
00053 #include <vector>
00054 using namespace llvm;
00055 using namespace PatternMatch;
00056 
00057 #define DEBUG_TYPE "gvn"
00058 
00059 STATISTIC(NumGVNInstr,  "Number of instructions deleted");
00060 STATISTIC(NumGVNLoad,   "Number of loads deleted");
00061 STATISTIC(NumGVNPRE,    "Number of instructions PRE'd");
00062 STATISTIC(NumGVNBlocks, "Number of blocks merged");
00063 STATISTIC(NumGVNSimpl,  "Number of instructions simplified");
00064 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
00065 STATISTIC(NumPRELoad,   "Number of loads PRE'd");
00066 
00067 static cl::opt<bool> EnablePRE("enable-pre",
00068                                cl::init(true), cl::Hidden);
00069 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
00070 
00071 // Maximum allowed recursion depth.
00072 static cl::opt<uint32_t>
00073 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
00074                 cl::desc("Max recurse depth (default = 1000)"));
00075 
00076 //===----------------------------------------------------------------------===//
00077 //                         ValueTable Class
00078 //===----------------------------------------------------------------------===//
00079 
00080 /// This class holds the mapping between values and value numbers.  It is used
00081 /// as an efficient mechanism to determine the expression-wise equivalence of
00082 /// two values.
00083 namespace {
00084   struct Expression {
00085     uint32_t opcode;
00086     Type *type;
00087     SmallVector<uint32_t, 4> varargs;
00088 
00089     Expression(uint32_t o = ~2U) : opcode(o) { }
00090 
00091     bool operator==(const Expression &other) const {
00092       if (opcode != other.opcode)
00093         return false;
00094       if (opcode == ~0U || opcode == ~1U)
00095         return true;
00096       if (type != other.type)
00097         return false;
00098       if (varargs != other.varargs)
00099         return false;
00100       return true;
00101     }
00102 
00103     friend hash_code hash_value(const Expression &Value) {
00104       return hash_combine(Value.opcode, Value.type,
00105                           hash_combine_range(Value.varargs.begin(),
00106                                              Value.varargs.end()));
00107     }
00108   };
00109 
00110   class ValueTable {
00111     DenseMap<Value*, uint32_t> valueNumbering;
00112     DenseMap<Expression, uint32_t> expressionNumbering;
00113     AliasAnalysis *AA;
00114     MemoryDependenceAnalysis *MD;
00115     DominatorTree *DT;
00116 
00117     uint32_t nextValueNumber;
00118 
00119     Expression create_expression(Instruction* I);
00120     Expression create_cmp_expression(unsigned Opcode,
00121                                      CmpInst::Predicate Predicate,
00122                                      Value *LHS, Value *RHS);
00123     Expression create_extractvalue_expression(ExtractValueInst* EI);
00124     uint32_t lookup_or_add_call(CallInst* C);
00125   public:
00126     ValueTable() : nextValueNumber(1) { }
00127     uint32_t lookup_or_add(Value *V);
00128     uint32_t lookup(Value *V) const;
00129     uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred,
00130                                Value *LHS, Value *RHS);
00131     void add(Value *V, uint32_t num);
00132     void clear();
00133     void erase(Value *v);
00134     void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
00135     AliasAnalysis *getAliasAnalysis() const { return AA; }
00136     void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
00137     void setDomTree(DominatorTree* D) { DT = D; }
00138     uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
00139     void verifyRemoved(const Value *) const;
00140   };
00141 }
00142 
00143 namespace llvm {
00144 template <> struct DenseMapInfo<Expression> {
00145   static inline Expression getEmptyKey() {
00146     return ~0U;
00147   }
00148 
00149   static inline Expression getTombstoneKey() {
00150     return ~1U;
00151   }
00152 
00153   static unsigned getHashValue(const Expression e) {
00154     using llvm::hash_value;
00155     return static_cast<unsigned>(hash_value(e));
00156   }
00157   static bool isEqual(const Expression &LHS, const Expression &RHS) {
00158     return LHS == RHS;
00159   }
00160 };
00161 
00162 }
00163 
00164 //===----------------------------------------------------------------------===//
00165 //                     ValueTable Internal Functions
00166 //===----------------------------------------------------------------------===//
00167 
00168 Expression ValueTable::create_expression(Instruction *I) {
00169   Expression e;
00170   e.type = I->getType();
00171   e.opcode = I->getOpcode();
00172   for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
00173        OI != OE; ++OI)
00174     e.varargs.push_back(lookup_or_add(*OI));
00175   if (I->isCommutative()) {
00176     // Ensure that commutative instructions that only differ by a permutation
00177     // of their operands get the same value number by sorting the operand value
00178     // numbers.  Since all commutative instructions have two operands it is more
00179     // efficient to sort by hand rather than using, say, std::sort.
00180     assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
00181     if (e.varargs[0] > e.varargs[1])
00182       std::swap(e.varargs[0], e.varargs[1]);
00183   }
00184 
00185   if (CmpInst *C = dyn_cast<CmpInst>(I)) {
00186     // Sort the operand value numbers so x<y and y>x get the same value number.
00187     CmpInst::Predicate Predicate = C->getPredicate();
00188     if (e.varargs[0] > e.varargs[1]) {
00189       std::swap(e.varargs[0], e.varargs[1]);
00190       Predicate = CmpInst::getSwappedPredicate(Predicate);
00191     }
00192     e.opcode = (C->getOpcode() << 8) | Predicate;
00193   } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
00194     for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
00195          II != IE; ++II)
00196       e.varargs.push_back(*II);
00197   }
00198 
00199   return e;
00200 }
00201 
00202 Expression ValueTable::create_cmp_expression(unsigned Opcode,
00203                                              CmpInst::Predicate Predicate,
00204                                              Value *LHS, Value *RHS) {
00205   assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
00206          "Not a comparison!");
00207   Expression e;
00208   e.type = CmpInst::makeCmpResultType(LHS->getType());
00209   e.varargs.push_back(lookup_or_add(LHS));
00210   e.varargs.push_back(lookup_or_add(RHS));
00211 
00212   // Sort the operand value numbers so x<y and y>x get the same value number.
00213   if (e.varargs[0] > e.varargs[1]) {
00214     std::swap(e.varargs[0], e.varargs[1]);
00215     Predicate = CmpInst::getSwappedPredicate(Predicate);
00216   }
00217   e.opcode = (Opcode << 8) | Predicate;
00218   return e;
00219 }
00220 
00221 Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) {
00222   assert(EI && "Not an ExtractValueInst?");
00223   Expression e;
00224   e.type = EI->getType();
00225   e.opcode = 0;
00226 
00227   IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
00228   if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
00229     // EI might be an extract from one of our recognised intrinsics. If it
00230     // is we'll synthesize a semantically equivalent expression instead on
00231     // an extract value expression.
00232     switch (I->getIntrinsicID()) {
00233       case Intrinsic::sadd_with_overflow:
00234       case Intrinsic::uadd_with_overflow:
00235         e.opcode = Instruction::Add;
00236         break;
00237       case Intrinsic::ssub_with_overflow:
00238       case Intrinsic::usub_with_overflow:
00239         e.opcode = Instruction::Sub;
00240         break;
00241       case Intrinsic::smul_with_overflow:
00242       case Intrinsic::umul_with_overflow:
00243         e.opcode = Instruction::Mul;
00244         break;
00245       default:
00246         break;
00247     }
00248 
00249     if (e.opcode != 0) {
00250       // Intrinsic recognized. Grab its args to finish building the expression.
00251       assert(I->getNumArgOperands() == 2 &&
00252              "Expect two args for recognised intrinsics.");
00253       e.varargs.push_back(lookup_or_add(I->getArgOperand(0)));
00254       e.varargs.push_back(lookup_or_add(I->getArgOperand(1)));
00255       return e;
00256     }
00257   }
00258 
00259   // Not a recognised intrinsic. Fall back to producing an extract value
00260   // expression.
00261   e.opcode = EI->getOpcode();
00262   for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
00263        OI != OE; ++OI)
00264     e.varargs.push_back(lookup_or_add(*OI));
00265 
00266   for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
00267          II != IE; ++II)
00268     e.varargs.push_back(*II);
00269 
00270   return e;
00271 }
00272 
00273 //===----------------------------------------------------------------------===//
00274 //                     ValueTable External Functions
00275 //===----------------------------------------------------------------------===//
00276 
00277 /// add - Insert a value into the table with a specified value number.
00278 void ValueTable::add(Value *V, uint32_t num) {
00279   valueNumbering.insert(std::make_pair(V, num));
00280 }
00281 
00282 uint32_t ValueTable::lookup_or_add_call(CallInst *C) {
00283   if (AA->doesNotAccessMemory(C)) {
00284     Expression exp = create_expression(C);
00285     uint32_t &e = expressionNumbering[exp];
00286     if (!e) e = nextValueNumber++;
00287     valueNumbering[C] = e;
00288     return e;
00289   } else if (AA->onlyReadsMemory(C)) {
00290     Expression exp = create_expression(C);
00291     uint32_t &e = expressionNumbering[exp];
00292     if (!e) {
00293       e = nextValueNumber++;
00294       valueNumbering[C] = e;
00295       return e;
00296     }
00297     if (!MD) {
00298       e = nextValueNumber++;
00299       valueNumbering[C] = e;
00300       return e;
00301     }
00302 
00303     MemDepResult local_dep = MD->getDependency(C);
00304 
00305     if (!local_dep.isDef() && !local_dep.isNonLocal()) {
00306       valueNumbering[C] =  nextValueNumber;
00307       return nextValueNumber++;
00308     }
00309 
00310     if (local_dep.isDef()) {
00311       CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
00312 
00313       if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
00314         valueNumbering[C] = nextValueNumber;
00315         return nextValueNumber++;
00316       }
00317 
00318       for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
00319         uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
00320         uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
00321         if (c_vn != cd_vn) {
00322           valueNumbering[C] = nextValueNumber;
00323           return nextValueNumber++;
00324         }
00325       }
00326 
00327       uint32_t v = lookup_or_add(local_cdep);
00328       valueNumbering[C] = v;
00329       return v;
00330     }
00331 
00332     // Non-local case.
00333     const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
00334       MD->getNonLocalCallDependency(CallSite(C));
00335     // FIXME: Move the checking logic to MemDep!
00336     CallInst* cdep = nullptr;
00337 
00338     // Check to see if we have a single dominating call instruction that is
00339     // identical to C.
00340     for (unsigned i = 0, e = deps.size(); i != e; ++i) {
00341       const NonLocalDepEntry *I = &deps[i];
00342       if (I->getResult().isNonLocal())
00343         continue;
00344 
00345       // We don't handle non-definitions.  If we already have a call, reject
00346       // instruction dependencies.
00347       if (!I->getResult().isDef() || cdep != nullptr) {
00348         cdep = nullptr;
00349         break;
00350       }
00351 
00352       CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
00353       // FIXME: All duplicated with non-local case.
00354       if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
00355         cdep = NonLocalDepCall;
00356         continue;
00357       }
00358 
00359       cdep = nullptr;
00360       break;
00361     }
00362 
00363     if (!cdep) {
00364       valueNumbering[C] = nextValueNumber;
00365       return nextValueNumber++;
00366     }
00367 
00368     if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
00369       valueNumbering[C] = nextValueNumber;
00370       return nextValueNumber++;
00371     }
00372     for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
00373       uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
00374       uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
00375       if (c_vn != cd_vn) {
00376         valueNumbering[C] = nextValueNumber;
00377         return nextValueNumber++;
00378       }
00379     }
00380 
00381     uint32_t v = lookup_or_add(cdep);
00382     valueNumbering[C] = v;
00383     return v;
00384 
00385   } else {
00386     valueNumbering[C] = nextValueNumber;
00387     return nextValueNumber++;
00388   }
00389 }
00390 
00391 /// lookup_or_add - Returns the value number for the specified value, assigning
00392 /// it a new number if it did not have one before.
00393 uint32_t ValueTable::lookup_or_add(Value *V) {
00394   DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
00395   if (VI != valueNumbering.end())
00396     return VI->second;
00397 
00398   if (!isa<Instruction>(V)) {
00399     valueNumbering[V] = nextValueNumber;
00400     return nextValueNumber++;
00401   }
00402 
00403   Instruction* I = cast<Instruction>(V);
00404   Expression exp;
00405   switch (I->getOpcode()) {
00406     case Instruction::Call:
00407       return lookup_or_add_call(cast<CallInst>(I));
00408     case Instruction::Add:
00409     case Instruction::FAdd:
00410     case Instruction::Sub:
00411     case Instruction::FSub:
00412     case Instruction::Mul:
00413     case Instruction::FMul:
00414     case Instruction::UDiv:
00415     case Instruction::SDiv:
00416     case Instruction::FDiv:
00417     case Instruction::URem:
00418     case Instruction::SRem:
00419     case Instruction::FRem:
00420     case Instruction::Shl:
00421     case Instruction::LShr:
00422     case Instruction::AShr:
00423     case Instruction::And:
00424     case Instruction::Or:
00425     case Instruction::Xor:
00426     case Instruction::ICmp:
00427     case Instruction::FCmp:
00428     case Instruction::Trunc:
00429     case Instruction::ZExt:
00430     case Instruction::SExt:
00431     case Instruction::FPToUI:
00432     case Instruction::FPToSI:
00433     case Instruction::UIToFP:
00434     case Instruction::SIToFP:
00435     case Instruction::FPTrunc:
00436     case Instruction::FPExt:
00437     case Instruction::PtrToInt:
00438     case Instruction::IntToPtr:
00439     case Instruction::BitCast:
00440     case Instruction::Select:
00441     case Instruction::ExtractElement:
00442     case Instruction::InsertElement:
00443     case Instruction::ShuffleVector:
00444     case Instruction::InsertValue:
00445     case Instruction::GetElementPtr:
00446       exp = create_expression(I);
00447       break;
00448     case Instruction::ExtractValue:
00449       exp = create_extractvalue_expression(cast<ExtractValueInst>(I));
00450       break;
00451     default:
00452       valueNumbering[V] = nextValueNumber;
00453       return nextValueNumber++;
00454   }
00455 
00456   uint32_t& e = expressionNumbering[exp];
00457   if (!e) e = nextValueNumber++;
00458   valueNumbering[V] = e;
00459   return e;
00460 }
00461 
00462 /// Returns the value number of the specified value. Fails if
00463 /// the value has not yet been numbered.
00464 uint32_t ValueTable::lookup(Value *V) const {
00465   DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
00466   assert(VI != valueNumbering.end() && "Value not numbered?");
00467   return VI->second;
00468 }
00469 
00470 /// Returns the value number of the given comparison,
00471 /// assigning it a new number if it did not have one before.  Useful when
00472 /// we deduced the result of a comparison, but don't immediately have an
00473 /// instruction realizing that comparison to hand.
00474 uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode,
00475                                        CmpInst::Predicate Predicate,
00476                                        Value *LHS, Value *RHS) {
00477   Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS);
00478   uint32_t& e = expressionNumbering[exp];
00479   if (!e) e = nextValueNumber++;
00480   return e;
00481 }
00482 
00483 /// Remove all entries from the ValueTable.
00484 void ValueTable::clear() {
00485   valueNumbering.clear();
00486   expressionNumbering.clear();
00487   nextValueNumber = 1;
00488 }
00489 
00490 /// Remove a value from the value numbering.
00491 void ValueTable::erase(Value *V) {
00492   valueNumbering.erase(V);
00493 }
00494 
00495 /// verifyRemoved - Verify that the value is removed from all internal data
00496 /// structures.
00497 void ValueTable::verifyRemoved(const Value *V) const {
00498   for (DenseMap<Value*, uint32_t>::const_iterator
00499          I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
00500     assert(I->first != V && "Inst still occurs in value numbering map!");
00501   }
00502 }
00503 
00504 //===----------------------------------------------------------------------===//
00505 //                                GVN Pass
00506 //===----------------------------------------------------------------------===//
00507 
00508 namespace {
00509   class GVN;
00510   struct AvailableValueInBlock {
00511     /// BB - The basic block in question.
00512     BasicBlock *BB;
00513     enum ValType {
00514       SimpleVal,  // A simple offsetted value that is accessed.
00515       LoadVal,    // A value produced by a load.
00516       MemIntrin,  // A memory intrinsic which is loaded from.
00517       UndefVal    // A UndefValue representing a value from dead block (which
00518                   // is not yet physically removed from the CFG). 
00519     };
00520   
00521     /// V - The value that is live out of the block.
00522     PointerIntPair<Value *, 2, ValType> Val;
00523   
00524     /// Offset - The byte offset in Val that is interesting for the load query.
00525     unsigned Offset;
00526   
00527     static AvailableValueInBlock get(BasicBlock *BB, Value *V,
00528                                      unsigned Offset = 0) {
00529       AvailableValueInBlock Res;
00530       Res.BB = BB;
00531       Res.Val.setPointer(V);
00532       Res.Val.setInt(SimpleVal);
00533       Res.Offset = Offset;
00534       return Res;
00535     }
00536   
00537     static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
00538                                        unsigned Offset = 0) {
00539       AvailableValueInBlock Res;
00540       Res.BB = BB;
00541       Res.Val.setPointer(MI);
00542       Res.Val.setInt(MemIntrin);
00543       Res.Offset = Offset;
00544       return Res;
00545     }
00546   
00547     static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
00548                                          unsigned Offset = 0) {
00549       AvailableValueInBlock Res;
00550       Res.BB = BB;
00551       Res.Val.setPointer(LI);
00552       Res.Val.setInt(LoadVal);
00553       Res.Offset = Offset;
00554       return Res;
00555     }
00556 
00557     static AvailableValueInBlock getUndef(BasicBlock *BB) {
00558       AvailableValueInBlock Res;
00559       Res.BB = BB;
00560       Res.Val.setPointer(nullptr);
00561       Res.Val.setInt(UndefVal);
00562       Res.Offset = 0;
00563       return Res;
00564     }
00565 
00566     bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
00567     bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
00568     bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
00569     bool isUndefValue() const { return Val.getInt() == UndefVal; }
00570   
00571     Value *getSimpleValue() const {
00572       assert(isSimpleValue() && "Wrong accessor");
00573       return Val.getPointer();
00574     }
00575   
00576     LoadInst *getCoercedLoadValue() const {
00577       assert(isCoercedLoadValue() && "Wrong accessor");
00578       return cast<LoadInst>(Val.getPointer());
00579     }
00580   
00581     MemIntrinsic *getMemIntrinValue() const {
00582       assert(isMemIntrinValue() && "Wrong accessor");
00583       return cast<MemIntrinsic>(Val.getPointer());
00584     }
00585   
00586     /// Emit code into this block to adjust the value defined here to the
00587     /// specified type. This handles various coercion cases.
00588     Value *MaterializeAdjustedValue(LoadInst *LI, GVN &gvn) const;
00589   };
00590 
00591   class GVN : public FunctionPass {
00592     bool NoLoads;
00593     MemoryDependenceAnalysis *MD;
00594     DominatorTree *DT;
00595     const TargetLibraryInfo *TLI;
00596     AssumptionCache *AC;
00597     SetVector<BasicBlock *> DeadBlocks;
00598 
00599     ValueTable VN;
00600 
00601     /// A mapping from value numbers to lists of Value*'s that
00602     /// have that value number.  Use findLeader to query it.
00603     struct LeaderTableEntry {
00604       Value *Val;
00605       const BasicBlock *BB;
00606       LeaderTableEntry *Next;
00607     };
00608     DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
00609     BumpPtrAllocator TableAllocator;
00610 
00611     SmallVector<Instruction*, 8> InstrsToErase;
00612 
00613     typedef SmallVector<NonLocalDepResult, 64> LoadDepVect;
00614     typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect;
00615     typedef SmallVector<BasicBlock*, 64> UnavailBlkVect;
00616 
00617   public:
00618     static char ID; // Pass identification, replacement for typeid
00619     explicit GVN(bool noloads = false)
00620         : FunctionPass(ID), NoLoads(noloads), MD(nullptr) {
00621       initializeGVNPass(*PassRegistry::getPassRegistry());
00622     }
00623 
00624     bool runOnFunction(Function &F) override;
00625 
00626     /// This removes the specified instruction from
00627     /// our various maps and marks it for deletion.
00628     void markInstructionForDeletion(Instruction *I) {
00629       VN.erase(I);
00630       InstrsToErase.push_back(I);
00631     }
00632 
00633     DominatorTree &getDominatorTree() const { return *DT; }
00634     AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
00635     MemoryDependenceAnalysis &getMemDep() const { return *MD; }
00636   private:
00637     /// Push a new Value to the LeaderTable onto the list for its value number.
00638     void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) {
00639       LeaderTableEntry &Curr = LeaderTable[N];
00640       if (!Curr.Val) {
00641         Curr.Val = V;
00642         Curr.BB = BB;
00643         return;
00644       }
00645 
00646       LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
00647       Node->Val = V;
00648       Node->BB = BB;
00649       Node->Next = Curr.Next;
00650       Curr.Next = Node;
00651     }
00652 
00653     /// Scan the list of values corresponding to a given
00654     /// value number, and remove the given instruction if encountered.
00655     void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) {
00656       LeaderTableEntry* Prev = nullptr;
00657       LeaderTableEntry* Curr = &LeaderTable[N];
00658 
00659       while (Curr->Val != I || Curr->BB != BB) {
00660         Prev = Curr;
00661         Curr = Curr->Next;
00662       }
00663 
00664       if (Prev) {
00665         Prev->Next = Curr->Next;
00666       } else {
00667         if (!Curr->Next) {
00668           Curr->Val = nullptr;
00669           Curr->BB = nullptr;
00670         } else {
00671           LeaderTableEntry* Next = Curr->Next;
00672           Curr->Val = Next->Val;
00673           Curr->BB = Next->BB;
00674           Curr->Next = Next->Next;
00675         }
00676       }
00677     }
00678 
00679     // List of critical edges to be split between iterations.
00680     SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
00681 
00682     // This transformation requires dominator postdominator info
00683     void getAnalysisUsage(AnalysisUsage &AU) const override {
00684       AU.addRequired<AssumptionCacheTracker>();
00685       AU.addRequired<DominatorTreeWrapperPass>();
00686       AU.addRequired<TargetLibraryInfoWrapperPass>();
00687       if (!NoLoads)
00688         AU.addRequired<MemoryDependenceAnalysis>();
00689       AU.addRequired<AliasAnalysis>();
00690 
00691       AU.addPreserved<DominatorTreeWrapperPass>();
00692       AU.addPreserved<AliasAnalysis>();
00693     }
00694 
00695 
00696     // Helper fuctions of redundant load elimination 
00697     bool processLoad(LoadInst *L);
00698     bool processNonLocalLoad(LoadInst *L);
00699     void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps, 
00700                                  AvailValInBlkVect &ValuesPerBlock,
00701                                  UnavailBlkVect &UnavailableBlocks);
00702     bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock, 
00703                         UnavailBlkVect &UnavailableBlocks);
00704 
00705     // Other helper routines
00706     bool processInstruction(Instruction *I);
00707     bool processBlock(BasicBlock *BB);
00708     void dump(DenseMap<uint32_t, Value*> &d);
00709     bool iterateOnFunction(Function &F);
00710     bool performPRE(Function &F);
00711     bool performScalarPRE(Instruction *I);
00712     bool performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
00713                                    unsigned int ValNo);
00714     Value *findLeader(const BasicBlock *BB, uint32_t num);
00715     void cleanupGlobalSets();
00716     void verifyRemoved(const Instruction *I) const;
00717     bool splitCriticalEdges();
00718     BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ);
00719     bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root);
00720     bool processFoldableCondBr(BranchInst *BI);
00721     void addDeadBlock(BasicBlock *BB);
00722     void assignValNumForDeadCode();
00723   };
00724 
00725   char GVN::ID = 0;
00726 }
00727 
00728 // The public interface to this file...
00729 FunctionPass *llvm::createGVNPass(bool NoLoads) {
00730   return new GVN(NoLoads);
00731 }
00732 
00733 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
00734 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
00735 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
00736 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
00737 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
00738 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
00739 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
00740 
00741 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
00742 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
00743   errs() << "{\n";
00744   for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
00745        E = d.end(); I != E; ++I) {
00746       errs() << I->first << "\n";
00747       I->second->dump();
00748   }
00749   errs() << "}\n";
00750 }
00751 #endif
00752 
00753 /// Return true if we can prove that the value
00754 /// we're analyzing is fully available in the specified block.  As we go, keep
00755 /// track of which blocks we know are fully alive in FullyAvailableBlocks.  This
00756 /// map is actually a tri-state map with the following values:
00757 ///   0) we know the block *is not* fully available.
00758 ///   1) we know the block *is* fully available.
00759 ///   2) we do not know whether the block is fully available or not, but we are
00760 ///      currently speculating that it will be.
00761 ///   3) we are speculating for this block and have used that to speculate for
00762 ///      other blocks.
00763 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
00764                             DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
00765                             uint32_t RecurseDepth) {
00766   if (RecurseDepth > MaxRecurseDepth)
00767     return false;
00768 
00769   // Optimistically assume that the block is fully available and check to see
00770   // if we already know about this block in one lookup.
00771   std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
00772     FullyAvailableBlocks.insert(std::make_pair(BB, 2));
00773 
00774   // If the entry already existed for this block, return the precomputed value.
00775   if (!IV.second) {
00776     // If this is a speculative "available" value, mark it as being used for
00777     // speculation of other blocks.
00778     if (IV.first->second == 2)
00779       IV.first->second = 3;
00780     return IV.first->second != 0;
00781   }
00782 
00783   // Otherwise, see if it is fully available in all predecessors.
00784   pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
00785 
00786   // If this block has no predecessors, it isn't live-in here.
00787   if (PI == PE)
00788     goto SpeculationFailure;
00789 
00790   for (; PI != PE; ++PI)
00791     // If the value isn't fully available in one of our predecessors, then it
00792     // isn't fully available in this block either.  Undo our previous
00793     // optimistic assumption and bail out.
00794     if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
00795       goto SpeculationFailure;
00796 
00797   return true;
00798 
00799 // If we get here, we found out that this is not, after
00800 // all, a fully-available block.  We have a problem if we speculated on this and
00801 // used the speculation to mark other blocks as available.
00802 SpeculationFailure:
00803   char &BBVal = FullyAvailableBlocks[BB];
00804 
00805   // If we didn't speculate on this, just return with it set to false.
00806   if (BBVal == 2) {
00807     BBVal = 0;
00808     return false;
00809   }
00810 
00811   // If we did speculate on this value, we could have blocks set to 1 that are
00812   // incorrect.  Walk the (transitive) successors of this block and mark them as
00813   // 0 if set to one.
00814   SmallVector<BasicBlock*, 32> BBWorklist;
00815   BBWorklist.push_back(BB);
00816 
00817   do {
00818     BasicBlock *Entry = BBWorklist.pop_back_val();
00819     // Note that this sets blocks to 0 (unavailable) if they happen to not
00820     // already be in FullyAvailableBlocks.  This is safe.
00821     char &EntryVal = FullyAvailableBlocks[Entry];
00822     if (EntryVal == 0) continue;  // Already unavailable.
00823 
00824     // Mark as unavailable.
00825     EntryVal = 0;
00826 
00827     BBWorklist.append(succ_begin(Entry), succ_end(Entry));
00828   } while (!BBWorklist.empty());
00829 
00830   return false;
00831 }
00832 
00833 
00834 /// Return true if CoerceAvailableValueToLoadType will succeed.
00835 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
00836                                             Type *LoadTy,
00837                                             const DataLayout &DL) {
00838   // If the loaded or stored value is an first class array or struct, don't try
00839   // to transform them.  We need to be able to bitcast to integer.
00840   if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
00841       StoredVal->getType()->isStructTy() ||
00842       StoredVal->getType()->isArrayTy())
00843     return false;
00844 
00845   // The store has to be at least as big as the load.
00846   if (DL.getTypeSizeInBits(StoredVal->getType()) <
00847         DL.getTypeSizeInBits(LoadTy))
00848     return false;
00849 
00850   return true;
00851 }
00852 
00853 /// If we saw a store of a value to memory, and
00854 /// then a load from a must-aliased pointer of a different type, try to coerce
00855 /// the stored value.  LoadedTy is the type of the load we want to replace.
00856 /// IRB is IRBuilder used to insert new instructions.
00857 ///
00858 /// If we can't do it, return null.
00859 static Value *CoerceAvailableValueToLoadType(Value *StoredVal, Type *LoadedTy,
00860                                              IRBuilder<> &IRB,
00861                                              const DataLayout &DL) {
00862   if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL))
00863     return nullptr;
00864 
00865   // If this is already the right type, just return it.
00866   Type *StoredValTy = StoredVal->getType();
00867 
00868   uint64_t StoreSize = DL.getTypeSizeInBits(StoredValTy);
00869   uint64_t LoadSize = DL.getTypeSizeInBits(LoadedTy);
00870 
00871   // If the store and reload are the same size, we can always reuse it.
00872   if (StoreSize == LoadSize) {
00873     // Pointer to Pointer -> use bitcast.
00874     if (StoredValTy->getScalarType()->isPointerTy() &&
00875         LoadedTy->getScalarType()->isPointerTy())
00876       return IRB.CreateBitCast(StoredVal, LoadedTy);
00877 
00878     // Convert source pointers to integers, which can be bitcast.
00879     if (StoredValTy->getScalarType()->isPointerTy()) {
00880       StoredValTy = DL.getIntPtrType(StoredValTy);
00881       StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
00882     }
00883 
00884     Type *TypeToCastTo = LoadedTy;
00885     if (TypeToCastTo->getScalarType()->isPointerTy())
00886       TypeToCastTo = DL.getIntPtrType(TypeToCastTo);
00887 
00888     if (StoredValTy != TypeToCastTo)
00889       StoredVal = IRB.CreateBitCast(StoredVal, TypeToCastTo);
00890 
00891     // Cast to pointer if the load needs a pointer type.
00892     if (LoadedTy->getScalarType()->isPointerTy())
00893       StoredVal = IRB.CreateIntToPtr(StoredVal, LoadedTy);
00894 
00895     return StoredVal;
00896   }
00897 
00898   // If the loaded value is smaller than the available value, then we can
00899   // extract out a piece from it.  If the available value is too small, then we
00900   // can't do anything.
00901   assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
00902 
00903   // Convert source pointers to integers, which can be manipulated.
00904   if (StoredValTy->getScalarType()->isPointerTy()) {
00905     StoredValTy = DL.getIntPtrType(StoredValTy);
00906     StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
00907   }
00908 
00909   // Convert vectors and fp to integer, which can be manipulated.
00910   if (!StoredValTy->isIntegerTy()) {
00911     StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
00912     StoredVal = IRB.CreateBitCast(StoredVal, StoredValTy);
00913   }
00914 
00915   // If this is a big-endian system, we need to shift the value down to the low
00916   // bits so that a truncate will work.
00917   if (DL.isBigEndian()) {
00918     StoredVal = IRB.CreateLShr(StoredVal, StoreSize - LoadSize, "tmp");
00919   }
00920 
00921   // Truncate the integer to the right size now.
00922   Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
00923   StoredVal  = IRB.CreateTrunc(StoredVal, NewIntTy, "trunc");
00924 
00925   if (LoadedTy == NewIntTy)
00926     return StoredVal;
00927 
00928   // If the result is a pointer, inttoptr.
00929   if (LoadedTy->getScalarType()->isPointerTy())
00930     return IRB.CreateIntToPtr(StoredVal, LoadedTy, "inttoptr");
00931 
00932   // Otherwise, bitcast.
00933   return IRB.CreateBitCast(StoredVal, LoadedTy, "bitcast");
00934 }
00935 
00936 /// This function is called when we have a
00937 /// memdep query of a load that ends up being a clobbering memory write (store,
00938 /// memset, memcpy, memmove).  This means that the write *may* provide bits used
00939 /// by the load but we can't be sure because the pointers don't mustalias.
00940 ///
00941 /// Check this case to see if there is anything more we can do before we give
00942 /// up.  This returns -1 if we have to give up, or a byte number in the stored
00943 /// value of the piece that feeds the load.
00944 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
00945                                           Value *WritePtr,
00946                                           uint64_t WriteSizeInBits,
00947                                           const DataLayout &DL) {
00948   // If the loaded or stored value is a first class array or struct, don't try
00949   // to transform them.  We need to be able to bitcast to integer.
00950   if (LoadTy->isStructTy() || LoadTy->isArrayTy())
00951     return -1;
00952 
00953   int64_t StoreOffset = 0, LoadOffset = 0;
00954   Value *StoreBase =
00955       GetPointerBaseWithConstantOffset(WritePtr, StoreOffset, DL);
00956   Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, DL);
00957   if (StoreBase != LoadBase)
00958     return -1;
00959 
00960   // If the load and store are to the exact same address, they should have been
00961   // a must alias.  AA must have gotten confused.
00962   // FIXME: Study to see if/when this happens.  One case is forwarding a memset
00963   // to a load from the base of the memset.
00964 #if 0
00965   if (LoadOffset == StoreOffset) {
00966     dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
00967     << "Base       = " << *StoreBase << "\n"
00968     << "Store Ptr  = " << *WritePtr << "\n"
00969     << "Store Offs = " << StoreOffset << "\n"
00970     << "Load Ptr   = " << *LoadPtr << "\n";
00971     abort();
00972   }
00973 #endif
00974 
00975   // If the load and store don't overlap at all, the store doesn't provide
00976   // anything to the load.  In this case, they really don't alias at all, AA
00977   // must have gotten confused.
00978   uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy);
00979 
00980   if ((WriteSizeInBits & 7) | (LoadSize & 7))
00981     return -1;
00982   uint64_t StoreSize = WriteSizeInBits >> 3;  // Convert to bytes.
00983   LoadSize >>= 3;
00984 
00985 
00986   bool isAAFailure = false;
00987   if (StoreOffset < LoadOffset)
00988     isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
00989   else
00990     isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
00991 
00992   if (isAAFailure) {
00993 #if 0
00994     dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
00995     << "Base       = " << *StoreBase << "\n"
00996     << "Store Ptr  = " << *WritePtr << "\n"
00997     << "Store Offs = " << StoreOffset << "\n"
00998     << "Load Ptr   = " << *LoadPtr << "\n";
00999     abort();
01000 #endif
01001     return -1;
01002   }
01003 
01004   // If the Load isn't completely contained within the stored bits, we don't
01005   // have all the bits to feed it.  We could do something crazy in the future
01006   // (issue a smaller load then merge the bits in) but this seems unlikely to be
01007   // valuable.
01008   if (StoreOffset > LoadOffset ||
01009       StoreOffset+StoreSize < LoadOffset+LoadSize)
01010     return -1;
01011 
01012   // Okay, we can do this transformation.  Return the number of bytes into the
01013   // store that the load is.
01014   return LoadOffset-StoreOffset;
01015 }
01016 
01017 /// This function is called when we have a
01018 /// memdep query of a load that ends up being a clobbering store.
01019 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
01020                                           StoreInst *DepSI) {
01021   // Cannot handle reading from store of first-class aggregate yet.
01022   if (DepSI->getValueOperand()->getType()->isStructTy() ||
01023       DepSI->getValueOperand()->getType()->isArrayTy())
01024     return -1;
01025 
01026   const DataLayout &DL = DepSI->getModule()->getDataLayout();
01027   Value *StorePtr = DepSI->getPointerOperand();
01028   uint64_t StoreSize =DL.getTypeSizeInBits(DepSI->getValueOperand()->getType());
01029   return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
01030                                         StorePtr, StoreSize, DL);
01031 }
01032 
01033 /// This function is called when we have a
01034 /// memdep query of a load that ends up being clobbered by another load.  See if
01035 /// the other load can feed into the second load.
01036 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
01037                                          LoadInst *DepLI, const DataLayout &DL){
01038   // Cannot handle reading from store of first-class aggregate yet.
01039   if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
01040     return -1;
01041 
01042   Value *DepPtr = DepLI->getPointerOperand();
01043   uint64_t DepSize = DL.getTypeSizeInBits(DepLI->getType());
01044   int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL);
01045   if (R != -1) return R;
01046 
01047   // If we have a load/load clobber an DepLI can be widened to cover this load,
01048   // then we should widen it!
01049   int64_t LoadOffs = 0;
01050   const Value *LoadBase =
01051       GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, DL);
01052   unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
01053 
01054   unsigned Size = MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize(
01055       LoadBase, LoadOffs, LoadSize, DepLI);
01056   if (Size == 0) return -1;
01057 
01058   return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, DL);
01059 }
01060 
01061 
01062 
01063 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
01064                                             MemIntrinsic *MI,
01065                                             const DataLayout &DL) {
01066   // If the mem operation is a non-constant size, we can't handle it.
01067   ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
01068   if (!SizeCst) return -1;
01069   uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
01070 
01071   // If this is memset, we just need to see if the offset is valid in the size
01072   // of the memset..
01073   if (MI->getIntrinsicID() == Intrinsic::memset)
01074     return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
01075                                           MemSizeInBits, DL);
01076 
01077   // If we have a memcpy/memmove, the only case we can handle is if this is a
01078   // copy from constant memory.  In that case, we can read directly from the
01079   // constant memory.
01080   MemTransferInst *MTI = cast<MemTransferInst>(MI);
01081 
01082   Constant *Src = dyn_cast<Constant>(MTI->getSource());
01083   if (!Src) return -1;
01084 
01085   GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, DL));
01086   if (!GV || !GV->isConstant()) return -1;
01087 
01088   // See if the access is within the bounds of the transfer.
01089   int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
01090                                               MI->getDest(), MemSizeInBits, DL);
01091   if (Offset == -1)
01092     return Offset;
01093 
01094   unsigned AS = Src->getType()->getPointerAddressSpace();
01095   // Otherwise, see if we can constant fold a load from the constant with the
01096   // offset applied as appropriate.
01097   Src = ConstantExpr::getBitCast(Src,
01098                                  Type::getInt8PtrTy(Src->getContext(), AS));
01099   Constant *OffsetCst =
01100     ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
01101   Src = ConstantExpr::getGetElementPtr(Type::getInt8Ty(Src->getContext()), Src,
01102                                        OffsetCst);
01103   Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
01104   if (ConstantFoldLoadFromConstPtr(Src, DL))
01105     return Offset;
01106   return -1;
01107 }
01108 
01109 
01110 /// This function is called when we have a
01111 /// memdep query of a load that ends up being a clobbering store.  This means
01112 /// that the store provides bits used by the load but we the pointers don't
01113 /// mustalias.  Check this case to see if there is anything more we can do
01114 /// before we give up.
01115 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
01116                                    Type *LoadTy,
01117                                    Instruction *InsertPt, const DataLayout &DL){
01118   LLVMContext &Ctx = SrcVal->getType()->getContext();
01119 
01120   uint64_t StoreSize = (DL.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
01121   uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy) + 7) / 8;
01122 
01123   IRBuilder<> Builder(InsertPt);
01124 
01125   // Compute which bits of the stored value are being used by the load.  Convert
01126   // to an integer type to start with.
01127   if (SrcVal->getType()->getScalarType()->isPointerTy())
01128     SrcVal = Builder.CreatePtrToInt(SrcVal,
01129         DL.getIntPtrType(SrcVal->getType()));
01130   if (!SrcVal->getType()->isIntegerTy())
01131     SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
01132 
01133   // Shift the bits to the least significant depending on endianness.
01134   unsigned ShiftAmt;
01135   if (DL.isLittleEndian())
01136     ShiftAmt = Offset*8;
01137   else
01138     ShiftAmt = (StoreSize-LoadSize-Offset)*8;
01139 
01140   if (ShiftAmt)
01141     SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
01142 
01143   if (LoadSize != StoreSize)
01144     SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
01145 
01146   return CoerceAvailableValueToLoadType(SrcVal, LoadTy, Builder, DL);
01147 }
01148 
01149 /// This function is called when we have a
01150 /// memdep query of a load that ends up being a clobbering load.  This means
01151 /// that the load *may* provide bits used by the load but we can't be sure
01152 /// because the pointers don't mustalias.  Check this case to see if there is
01153 /// anything more we can do before we give up.
01154 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
01155                                   Type *LoadTy, Instruction *InsertPt,
01156                                   GVN &gvn) {
01157   const DataLayout &DL = SrcVal->getModule()->getDataLayout();
01158   // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
01159   // widen SrcVal out to a larger load.
01160   unsigned SrcValSize = DL.getTypeStoreSize(SrcVal->getType());
01161   unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
01162   if (Offset+LoadSize > SrcValSize) {
01163     assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
01164     assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
01165     // If we have a load/load clobber an DepLI can be widened to cover this
01166     // load, then we should widen it to the next power of 2 size big enough!
01167     unsigned NewLoadSize = Offset+LoadSize;
01168     if (!isPowerOf2_32(NewLoadSize))
01169       NewLoadSize = NextPowerOf2(NewLoadSize);
01170 
01171     Value *PtrVal = SrcVal->getPointerOperand();
01172 
01173     // Insert the new load after the old load.  This ensures that subsequent
01174     // memdep queries will find the new load.  We can't easily remove the old
01175     // load completely because it is already in the value numbering table.
01176     IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
01177     Type *DestPTy =
01178       IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
01179     DestPTy = PointerType::get(DestPTy,
01180                                PtrVal->getType()->getPointerAddressSpace());
01181     Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
01182     PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
01183     LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
01184     NewLoad->takeName(SrcVal);
01185     NewLoad->setAlignment(SrcVal->getAlignment());
01186 
01187     DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
01188     DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
01189 
01190     // Replace uses of the original load with the wider load.  On a big endian
01191     // system, we need to shift down to get the relevant bits.
01192     Value *RV = NewLoad;
01193     if (DL.isBigEndian())
01194       RV = Builder.CreateLShr(RV,
01195                     NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
01196     RV = Builder.CreateTrunc(RV, SrcVal->getType());
01197     SrcVal->replaceAllUsesWith(RV);
01198 
01199     // We would like to use gvn.markInstructionForDeletion here, but we can't
01200     // because the load is already memoized into the leader map table that GVN
01201     // tracks.  It is potentially possible to remove the load from the table,
01202     // but then there all of the operations based on it would need to be
01203     // rehashed.  Just leave the dead load around.
01204     gvn.getMemDep().removeInstruction(SrcVal);
01205     SrcVal = NewLoad;
01206   }
01207 
01208   return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, DL);
01209 }
01210 
01211 
01212 /// This function is called when we have a
01213 /// memdep query of a load that ends up being a clobbering mem intrinsic.
01214 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
01215                                      Type *LoadTy, Instruction *InsertPt,
01216                                      const DataLayout &DL){
01217   LLVMContext &Ctx = LoadTy->getContext();
01218   uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy)/8;
01219 
01220   IRBuilder<> Builder(InsertPt);
01221 
01222   // We know that this method is only called when the mem transfer fully
01223   // provides the bits for the load.
01224   if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
01225     // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
01226     // independently of what the offset is.
01227     Value *Val = MSI->getValue();
01228     if (LoadSize != 1)
01229       Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
01230 
01231     Value *OneElt = Val;
01232 
01233     // Splat the value out to the right number of bits.
01234     for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
01235       // If we can double the number of bytes set, do it.
01236       if (NumBytesSet*2 <= LoadSize) {
01237         Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
01238         Val = Builder.CreateOr(Val, ShVal);
01239         NumBytesSet <<= 1;
01240         continue;
01241       }
01242 
01243       // Otherwise insert one byte at a time.
01244       Value *ShVal = Builder.CreateShl(Val, 1*8);
01245       Val = Builder.CreateOr(OneElt, ShVal);
01246       ++NumBytesSet;
01247     }
01248 
01249     return CoerceAvailableValueToLoadType(Val, LoadTy, Builder, DL);
01250   }
01251 
01252   // Otherwise, this is a memcpy/memmove from a constant global.
01253   MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
01254   Constant *Src = cast<Constant>(MTI->getSource());
01255   unsigned AS = Src->getType()->getPointerAddressSpace();
01256 
01257   // Otherwise, see if we can constant fold a load from the constant with the
01258   // offset applied as appropriate.
01259   Src = ConstantExpr::getBitCast(Src,
01260                                  Type::getInt8PtrTy(Src->getContext(), AS));
01261   Constant *OffsetCst =
01262     ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
01263   Src = ConstantExpr::getGetElementPtr(Type::getInt8Ty(Src->getContext()), Src,
01264                                        OffsetCst);
01265   Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
01266   return ConstantFoldLoadFromConstPtr(Src, DL);
01267 }
01268 
01269 
01270 /// Given a set of loads specified by ValuesPerBlock,
01271 /// construct SSA form, allowing us to eliminate LI.  This returns the value
01272 /// that should be used at LI's definition site.
01273 static Value *ConstructSSAForLoadSet(LoadInst *LI,
01274                          SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
01275                                      GVN &gvn) {
01276   // Check for the fully redundant, dominating load case.  In this case, we can
01277   // just use the dominating value directly.
01278   if (ValuesPerBlock.size() == 1 &&
01279       gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
01280                                                LI->getParent())) {
01281     assert(!ValuesPerBlock[0].isUndefValue() && "Dead BB dominate this block");
01282     return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
01283   }
01284 
01285   // Otherwise, we have to construct SSA form.
01286   SmallVector<PHINode*, 8> NewPHIs;
01287   SSAUpdater SSAUpdate(&NewPHIs);
01288   SSAUpdate.Initialize(LI->getType(), LI->getName());
01289 
01290   for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
01291     const AvailableValueInBlock &AV = ValuesPerBlock[i];
01292     BasicBlock *BB = AV.BB;
01293 
01294     if (SSAUpdate.HasValueForBlock(BB))
01295       continue;
01296 
01297     SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
01298   }
01299 
01300   // Perform PHI construction.
01301   Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
01302 
01303   // If new PHI nodes were created, notify alias analysis.
01304   if (V->getType()->getScalarType()->isPointerTy()) {
01305     AliasAnalysis *AA = gvn.getAliasAnalysis();
01306 
01307     for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
01308       AA->copyValue(LI, NewPHIs[i]);
01309 
01310     // Now that we've copied information to the new PHIs, scan through
01311     // them again and inform alias analysis that we've added potentially
01312     // escaping uses to any values that are operands to these PHIs.
01313     for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
01314       PHINode *P = NewPHIs[i];
01315       for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) {
01316         unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
01317         AA->addEscapingUse(P->getOperandUse(jj));
01318       }
01319     }
01320   }
01321 
01322   return V;
01323 }
01324 
01325 Value *AvailableValueInBlock::MaterializeAdjustedValue(LoadInst *LI,
01326                                                        GVN &gvn) const {
01327   Value *Res;
01328   Type *LoadTy = LI->getType();
01329   const DataLayout &DL = LI->getModule()->getDataLayout();
01330   if (isSimpleValue()) {
01331     Res = getSimpleValue();
01332     if (Res->getType() != LoadTy) {
01333       Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), DL);
01334 
01335       DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << "  "
01336                    << *getSimpleValue() << '\n'
01337                    << *Res << '\n' << "\n\n\n");
01338     }
01339   } else if (isCoercedLoadValue()) {
01340     LoadInst *Load = getCoercedLoadValue();
01341     if (Load->getType() == LoadTy && Offset == 0) {
01342       Res = Load;
01343     } else {
01344       Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
01345                                 gvn);
01346   
01347       DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << "  "
01348                    << *getCoercedLoadValue() << '\n'
01349                    << *Res << '\n' << "\n\n\n");
01350     }
01351   } else if (isMemIntrinValue()) {
01352     Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
01353                                  BB->getTerminator(), DL);
01354     DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
01355                  << "  " << *getMemIntrinValue() << '\n'
01356                  << *Res << '\n' << "\n\n\n");
01357   } else {
01358     assert(isUndefValue() && "Should be UndefVal");
01359     DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
01360     return UndefValue::get(LoadTy);
01361   }
01362   return Res;
01363 }
01364 
01365 static bool isLifetimeStart(const Instruction *Inst) {
01366   if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
01367     return II->getIntrinsicID() == Intrinsic::lifetime_start;
01368   return false;
01369 }
01370 
01371 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps, 
01372                                   AvailValInBlkVect &ValuesPerBlock,
01373                                   UnavailBlkVect &UnavailableBlocks) {
01374 
01375   // Filter out useless results (non-locals, etc).  Keep track of the blocks
01376   // where we have a value available in repl, also keep track of whether we see
01377   // dependencies that produce an unknown value for the load (such as a call
01378   // that could potentially clobber the load).
01379   unsigned NumDeps = Deps.size();
01380   const DataLayout &DL = LI->getModule()->getDataLayout();
01381   for (unsigned i = 0, e = NumDeps; i != e; ++i) {
01382     BasicBlock *DepBB = Deps[i].getBB();
01383     MemDepResult DepInfo = Deps[i].getResult();
01384 
01385     if (DeadBlocks.count(DepBB)) {
01386       // Dead dependent mem-op disguise as a load evaluating the same value
01387       // as the load in question.
01388       ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
01389       continue;
01390     }
01391 
01392     if (!DepInfo.isDef() && !DepInfo.isClobber()) {
01393       UnavailableBlocks.push_back(DepBB);
01394       continue;
01395     }
01396 
01397     if (DepInfo.isClobber()) {
01398       // The address being loaded in this non-local block may not be the same as
01399       // the pointer operand of the load if PHI translation occurs.  Make sure
01400       // to consider the right address.
01401       Value *Address = Deps[i].getAddress();
01402 
01403       // If the dependence is to a store that writes to a superset of the bits
01404       // read by the load, we can extract the bits we need for the load from the
01405       // stored value.
01406       if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
01407         if (Address) {
01408           int Offset =
01409               AnalyzeLoadFromClobberingStore(LI->getType(), Address, DepSI);
01410           if (Offset != -1) {
01411             ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
01412                                                        DepSI->getValueOperand(),
01413                                                                 Offset));
01414             continue;
01415           }
01416         }
01417       }
01418 
01419       // Check to see if we have something like this:
01420       //    load i32* P
01421       //    load i8* (P+1)
01422       // if we have this, replace the later with an extraction from the former.
01423       if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
01424         // If this is a clobber and L is the first instruction in its block, then
01425         // we have the first instruction in the entry block.
01426         if (DepLI != LI && Address) {
01427           int Offset =
01428               AnalyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
01429 
01430           if (Offset != -1) {
01431             ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
01432                                                                     Offset));
01433             continue;
01434           }
01435         }
01436       }
01437 
01438       // If the clobbering value is a memset/memcpy/memmove, see if we can
01439       // forward a value on from it.
01440       if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
01441         if (Address) {
01442           int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
01443                                                         DepMI, DL);
01444           if (Offset != -1) {
01445             ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
01446                                                                   Offset));
01447             continue;
01448           }
01449         }
01450       }
01451 
01452       UnavailableBlocks.push_back(DepBB);
01453       continue;
01454     }
01455 
01456     // DepInfo.isDef() here
01457 
01458     Instruction *DepInst = DepInfo.getInst();
01459 
01460     // Loading the allocation -> undef.
01461     if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
01462         // Loading immediately after lifetime begin -> undef.
01463         isLifetimeStart(DepInst)) {
01464       ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
01465                                              UndefValue::get(LI->getType())));
01466       continue;
01467     }
01468 
01469     // Loading from calloc (which zero initializes memory) -> zero
01470     if (isCallocLikeFn(DepInst, TLI)) {
01471       ValuesPerBlock.push_back(AvailableValueInBlock::get(
01472           DepBB, Constant::getNullValue(LI->getType())));
01473       continue;
01474     }
01475 
01476     if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
01477       // Reject loads and stores that are to the same address but are of
01478       // different types if we have to.
01479       if (S->getValueOperand()->getType() != LI->getType()) {
01480         // If the stored value is larger or equal to the loaded value, we can
01481         // reuse it.
01482         if (!CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
01483                                              LI->getType(), DL)) {
01484           UnavailableBlocks.push_back(DepBB);
01485           continue;
01486         }
01487       }
01488 
01489       ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
01490                                                          S->getValueOperand()));
01491       continue;
01492     }
01493 
01494     if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
01495       // If the types mismatch and we can't handle it, reject reuse of the load.
01496       if (LD->getType() != LI->getType()) {
01497         // If the stored value is larger or equal to the loaded value, we can
01498         // reuse it.
01499         if (!CanCoerceMustAliasedValueToLoad(LD, LI->getType(), DL)) {
01500           UnavailableBlocks.push_back(DepBB);
01501           continue;
01502         }
01503       }
01504       ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
01505       continue;
01506     }
01507 
01508     UnavailableBlocks.push_back(DepBB);
01509   }
01510 }
01511 
01512 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock, 
01513                          UnavailBlkVect &UnavailableBlocks) {
01514   // Okay, we have *some* definitions of the value.  This means that the value
01515   // is available in some of our (transitive) predecessors.  Lets think about
01516   // doing PRE of this load.  This will involve inserting a new load into the
01517   // predecessor when it's not available.  We could do this in general, but
01518   // prefer to not increase code size.  As such, we only do this when we know
01519   // that we only have to insert *one* load (which means we're basically moving
01520   // the load, not inserting a new one).
01521 
01522   SmallPtrSet<BasicBlock *, 4> Blockers;
01523   for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
01524     Blockers.insert(UnavailableBlocks[i]);
01525 
01526   // Let's find the first basic block with more than one predecessor.  Walk
01527   // backwards through predecessors if needed.
01528   BasicBlock *LoadBB = LI->getParent();
01529   BasicBlock *TmpBB = LoadBB;
01530 
01531   while (TmpBB->getSinglePredecessor()) {
01532     TmpBB = TmpBB->getSinglePredecessor();
01533     if (TmpBB == LoadBB) // Infinite (unreachable) loop.
01534       return false;
01535     if (Blockers.count(TmpBB))
01536       return false;
01537 
01538     // If any of these blocks has more than one successor (i.e. if the edge we
01539     // just traversed was critical), then there are other paths through this
01540     // block along which the load may not be anticipated.  Hoisting the load
01541     // above this block would be adding the load to execution paths along
01542     // which it was not previously executed.
01543     if (TmpBB->getTerminator()->getNumSuccessors() != 1)
01544       return false;
01545   }
01546 
01547   assert(TmpBB);
01548   LoadBB = TmpBB;
01549 
01550   // Check to see how many predecessors have the loaded value fully
01551   // available.
01552   MapVector<BasicBlock *, Value *> PredLoads;
01553   DenseMap<BasicBlock*, char> FullyAvailableBlocks;
01554   for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
01555     FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
01556   for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
01557     FullyAvailableBlocks[UnavailableBlocks[i]] = false;
01558 
01559   SmallVector<BasicBlock *, 4> CriticalEdgePred;
01560   for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
01561        PI != E; ++PI) {
01562     BasicBlock *Pred = *PI;
01563     if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
01564       continue;
01565     }
01566 
01567     if (Pred->getTerminator()->getNumSuccessors() != 1) {
01568       if (isa<IndirectBrInst>(Pred->getTerminator())) {
01569         DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
01570               << Pred->getName() << "': " << *LI << '\n');
01571         return false;
01572       }
01573 
01574       if (LoadBB->isLandingPad()) {
01575         DEBUG(dbgs()
01576               << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '"
01577               << Pred->getName() << "': " << *LI << '\n');
01578         return false;
01579       }
01580 
01581       CriticalEdgePred.push_back(Pred);
01582     } else {
01583       // Only add the predecessors that will not be split for now.
01584       PredLoads[Pred] = nullptr;
01585     }
01586   }
01587 
01588   // Decide whether PRE is profitable for this load.
01589   unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
01590   assert(NumUnavailablePreds != 0 &&
01591          "Fully available value should already be eliminated!");
01592 
01593   // If this load is unavailable in multiple predecessors, reject it.
01594   // FIXME: If we could restructure the CFG, we could make a common pred with
01595   // all the preds that don't have an available LI and insert a new load into
01596   // that one block.
01597   if (NumUnavailablePreds != 1)
01598       return false;
01599 
01600   // Split critical edges, and update the unavailable predecessors accordingly.
01601   for (BasicBlock *OrigPred : CriticalEdgePred) {
01602     BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
01603     assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
01604     PredLoads[NewPred] = nullptr;
01605     DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
01606                  << LoadBB->getName() << '\n');
01607   }
01608 
01609   // Check if the load can safely be moved to all the unavailable predecessors.
01610   bool CanDoPRE = true;
01611   const DataLayout &DL = LI->getModule()->getDataLayout();
01612   SmallVector<Instruction*, 8> NewInsts;
01613   for (auto &PredLoad : PredLoads) {
01614     BasicBlock *UnavailablePred = PredLoad.first;
01615 
01616     // Do PHI translation to get its value in the predecessor if necessary.  The
01617     // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
01618 
01619     // If all preds have a single successor, then we know it is safe to insert
01620     // the load on the pred (?!?), so we can insert code to materialize the
01621     // pointer if it is not available.
01622     PHITransAddr Address(LI->getPointerOperand(), DL, AC);
01623     Value *LoadPtr = nullptr;
01624     LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
01625                                                 *DT, NewInsts);
01626 
01627     // If we couldn't find or insert a computation of this phi translated value,
01628     // we fail PRE.
01629     if (!LoadPtr) {
01630       DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
01631             << *LI->getPointerOperand() << "\n");
01632       CanDoPRE = false;
01633       break;
01634     }
01635 
01636     PredLoad.second = LoadPtr;
01637   }
01638 
01639   if (!CanDoPRE) {
01640     while (!NewInsts.empty()) {
01641       Instruction *I = NewInsts.pop_back_val();
01642       if (MD) MD->removeInstruction(I);
01643       I->eraseFromParent();
01644     }
01645     // HINT: Don't revert the edge-splitting as following transformation may
01646     // also need to split these critical edges.
01647     return !CriticalEdgePred.empty();
01648   }
01649 
01650   // Okay, we can eliminate this load by inserting a reload in the predecessor
01651   // and using PHI construction to get the value in the other predecessors, do
01652   // it.
01653   DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
01654   DEBUG(if (!NewInsts.empty())
01655           dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
01656                  << *NewInsts.back() << '\n');
01657 
01658   // Assign value numbers to the new instructions.
01659   for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
01660     // FIXME: We really _ought_ to insert these value numbers into their
01661     // parent's availability map.  However, in doing so, we risk getting into
01662     // ordering issues.  If a block hasn't been processed yet, we would be
01663     // marking a value as AVAIL-IN, which isn't what we intend.
01664     VN.lookup_or_add(NewInsts[i]);
01665   }
01666 
01667   for (const auto &PredLoad : PredLoads) {
01668     BasicBlock *UnavailablePred = PredLoad.first;
01669     Value *LoadPtr = PredLoad.second;
01670 
01671     Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
01672                                         LI->getAlignment(),
01673                                         UnavailablePred->getTerminator());
01674 
01675     // Transfer the old load's AA tags to the new load.
01676     AAMDNodes Tags;
01677     LI->getAAMetadata(Tags);
01678     if (Tags)
01679       NewLoad->setAAMetadata(Tags);
01680 
01681     // Transfer DebugLoc.
01682     NewLoad->setDebugLoc(LI->getDebugLoc());
01683 
01684     // Add the newly created load.
01685     ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
01686                                                         NewLoad));
01687     MD->invalidateCachedPointerInfo(LoadPtr);
01688     DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
01689   }
01690 
01691   // Perform PHI construction.
01692   Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
01693   LI->replaceAllUsesWith(V);
01694   if (isa<PHINode>(V))
01695     V->takeName(LI);
01696   if (Instruction *I = dyn_cast<Instruction>(V))
01697     I->setDebugLoc(LI->getDebugLoc());
01698   if (V->getType()->getScalarType()->isPointerTy())
01699     MD->invalidateCachedPointerInfo(V);
01700   markInstructionForDeletion(LI);
01701   ++NumPRELoad;
01702   return true;
01703 }
01704 
01705 /// Attempt to eliminate a load whose dependencies are
01706 /// non-local by performing PHI construction.
01707 bool GVN::processNonLocalLoad(LoadInst *LI) {
01708   // Step 1: Find the non-local dependencies of the load.
01709   LoadDepVect Deps;
01710   MD->getNonLocalPointerDependency(LI, Deps);
01711 
01712   // If we had to process more than one hundred blocks to find the
01713   // dependencies, this load isn't worth worrying about.  Optimizing
01714   // it will be too expensive.
01715   unsigned NumDeps = Deps.size();
01716   if (NumDeps > 100)
01717     return false;
01718 
01719   // If we had a phi translation failure, we'll have a single entry which is a
01720   // clobber in the current block.  Reject this early.
01721   if (NumDeps == 1 &&
01722       !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
01723     DEBUG(
01724       dbgs() << "GVN: non-local load ";
01725       LI->printAsOperand(dbgs());
01726       dbgs() << " has unknown dependencies\n";
01727     );
01728     return false;
01729   }
01730 
01731   // If this load follows a GEP, see if we can PRE the indices before analyzing.
01732   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
01733     for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
01734                                         OE = GEP->idx_end();
01735          OI != OE; ++OI)
01736       if (Instruction *I = dyn_cast<Instruction>(OI->get()))
01737         performScalarPRE(I);
01738   }
01739 
01740   // Step 2: Analyze the availability of the load
01741   AvailValInBlkVect ValuesPerBlock;
01742   UnavailBlkVect UnavailableBlocks;
01743   AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
01744 
01745   // If we have no predecessors that produce a known value for this load, exit
01746   // early.
01747   if (ValuesPerBlock.empty())
01748     return false;
01749 
01750   // Step 3: Eliminate fully redundancy.
01751   //
01752   // If all of the instructions we depend on produce a known value for this
01753   // load, then it is fully redundant and we can use PHI insertion to compute
01754   // its value.  Insert PHIs and remove the fully redundant value now.
01755   if (UnavailableBlocks.empty()) {
01756     DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
01757 
01758     // Perform PHI construction.
01759     Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
01760     LI->replaceAllUsesWith(V);
01761 
01762     if (isa<PHINode>(V))
01763       V->takeName(LI);
01764     if (Instruction *I = dyn_cast<Instruction>(V))
01765       I->setDebugLoc(LI->getDebugLoc());
01766     if (V->getType()->getScalarType()->isPointerTy())
01767       MD->invalidateCachedPointerInfo(V);
01768     markInstructionForDeletion(LI);
01769     ++NumGVNLoad;
01770     return true;
01771   }
01772 
01773   // Step 4: Eliminate partial redundancy.
01774   if (!EnablePRE || !EnableLoadPRE)
01775     return false;
01776 
01777   return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
01778 }
01779 
01780 
01781 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
01782   // Patch the replacement so that it is not more restrictive than the value
01783   // being replaced.
01784   BinaryOperator *Op = dyn_cast<BinaryOperator>(I);
01785   BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl);
01786   if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) &&
01787       isa<OverflowingBinaryOperator>(ReplOp)) {
01788     if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap())
01789       ReplOp->setHasNoSignedWrap(false);
01790     if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap())
01791       ReplOp->setHasNoUnsignedWrap(false);
01792   }
01793   if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) {
01794     // FIXME: If both the original and replacement value are part of the
01795     // same control-flow region (meaning that the execution of one
01796     // guarentees the executation of the other), then we can combine the
01797     // noalias scopes here and do better than the general conservative
01798     // answer used in combineMetadata().
01799 
01800     // In general, GVN unifies expressions over different control-flow
01801     // regions, and so we need a conservative combination of the noalias
01802     // scopes.
01803     unsigned KnownIDs[] = {
01804       LLVMContext::MD_tbaa,
01805       LLVMContext::MD_alias_scope,
01806       LLVMContext::MD_noalias,
01807       LLVMContext::MD_range,
01808       LLVMContext::MD_fpmath,
01809       LLVMContext::MD_invariant_load,
01810     };
01811     combineMetadata(ReplInst, I, KnownIDs);
01812   }
01813 }
01814 
01815 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
01816   patchReplacementInstruction(I, Repl);
01817   I->replaceAllUsesWith(Repl);
01818 }
01819 
01820 /// Attempt to eliminate a load, first by eliminating it
01821 /// locally, and then attempting non-local elimination if that fails.
01822 bool GVN::processLoad(LoadInst *L) {
01823   if (!MD)
01824     return false;
01825 
01826   if (!L->isSimple())
01827     return false;
01828 
01829   if (L->use_empty()) {
01830     markInstructionForDeletion(L);
01831     return true;
01832   }
01833 
01834   // ... to a pointer that has been loaded from before...
01835   MemDepResult Dep = MD->getDependency(L);
01836   const DataLayout &DL = L->getModule()->getDataLayout();
01837 
01838   // If we have a clobber and target data is around, see if this is a clobber
01839   // that we can fix up through code synthesis.
01840   if (Dep.isClobber()) {
01841     // Check to see if we have something like this:
01842     //   store i32 123, i32* %P
01843     //   %A = bitcast i32* %P to i8*
01844     //   %B = gep i8* %A, i32 1
01845     //   %C = load i8* %B
01846     //
01847     // We could do that by recognizing if the clobber instructions are obviously
01848     // a common base + constant offset, and if the previous store (or memset)
01849     // completely covers this load.  This sort of thing can happen in bitfield
01850     // access code.
01851     Value *AvailVal = nullptr;
01852     if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
01853       int Offset = AnalyzeLoadFromClobberingStore(
01854           L->getType(), L->getPointerOperand(), DepSI);
01855       if (Offset != -1)
01856         AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
01857                                         L->getType(), L, DL);
01858     }
01859 
01860     // Check to see if we have something like this:
01861     //    load i32* P
01862     //    load i8* (P+1)
01863     // if we have this, replace the later with an extraction from the former.
01864     if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
01865       // If this is a clobber and L is the first instruction in its block, then
01866       // we have the first instruction in the entry block.
01867       if (DepLI == L)
01868         return false;
01869 
01870       int Offset = AnalyzeLoadFromClobberingLoad(
01871           L->getType(), L->getPointerOperand(), DepLI, DL);
01872       if (Offset != -1)
01873         AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
01874     }
01875 
01876     // If the clobbering value is a memset/memcpy/memmove, see if we can forward
01877     // a value on from it.
01878     if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
01879       int Offset = AnalyzeLoadFromClobberingMemInst(
01880           L->getType(), L->getPointerOperand(), DepMI, DL);
01881       if (Offset != -1)
01882         AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, DL);
01883     }
01884 
01885     if (AvailVal) {
01886       DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
01887             << *AvailVal << '\n' << *L << "\n\n\n");
01888 
01889       // Replace the load!
01890       L->replaceAllUsesWith(AvailVal);
01891       if (AvailVal->getType()->getScalarType()->isPointerTy())
01892         MD->invalidateCachedPointerInfo(AvailVal);
01893       markInstructionForDeletion(L);
01894       ++NumGVNLoad;
01895       return true;
01896     }
01897   }
01898 
01899   // If the value isn't available, don't do anything!
01900   if (Dep.isClobber()) {
01901     DEBUG(
01902       // fast print dep, using operator<< on instruction is too slow.
01903       dbgs() << "GVN: load ";
01904       L->printAsOperand(dbgs());
01905       Instruction *I = Dep.getInst();
01906       dbgs() << " is clobbered by " << *I << '\n';
01907     );
01908     return false;
01909   }
01910 
01911   // If it is defined in another block, try harder.
01912   if (Dep.isNonLocal())
01913     return processNonLocalLoad(L);
01914 
01915   if (!Dep.isDef()) {
01916     DEBUG(
01917       // fast print dep, using operator<< on instruction is too slow.
01918       dbgs() << "GVN: load ";
01919       L->printAsOperand(dbgs());
01920       dbgs() << " has unknown dependence\n";
01921     );
01922     return false;
01923   }
01924 
01925   Instruction *DepInst = Dep.getInst();
01926   if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
01927     Value *StoredVal = DepSI->getValueOperand();
01928 
01929     // The store and load are to a must-aliased pointer, but they may not
01930     // actually have the same type.  See if we know how to reuse the stored
01931     // value (depending on its type).
01932     if (StoredVal->getType() != L->getType()) {
01933       IRBuilder<> Builder(L);
01934       StoredVal =
01935           CoerceAvailableValueToLoadType(StoredVal, L->getType(), Builder, DL);
01936       if (!StoredVal)
01937         return false;
01938 
01939       DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
01940                    << '\n' << *L << "\n\n\n");
01941     }
01942 
01943     // Remove it!
01944     L->replaceAllUsesWith(StoredVal);
01945     if (StoredVal->getType()->getScalarType()->isPointerTy())
01946       MD->invalidateCachedPointerInfo(StoredVal);
01947     markInstructionForDeletion(L);
01948     ++NumGVNLoad;
01949     return true;
01950   }
01951 
01952   if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
01953     Value *AvailableVal = DepLI;
01954 
01955     // The loads are of a must-aliased pointer, but they may not actually have
01956     // the same type.  See if we know how to reuse the previously loaded value
01957     // (depending on its type).
01958     if (DepLI->getType() != L->getType()) {
01959       IRBuilder<> Builder(L);
01960       AvailableVal =
01961           CoerceAvailableValueToLoadType(DepLI, L->getType(), Builder, DL);
01962       if (!AvailableVal)
01963         return false;
01964 
01965       DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
01966                    << "\n" << *L << "\n\n\n");
01967     }
01968 
01969     // Remove it!
01970     patchAndReplaceAllUsesWith(L, AvailableVal);
01971     if (DepLI->getType()->getScalarType()->isPointerTy())
01972       MD->invalidateCachedPointerInfo(DepLI);
01973     markInstructionForDeletion(L);
01974     ++NumGVNLoad;
01975     return true;
01976   }
01977 
01978   // If this load really doesn't depend on anything, then we must be loading an
01979   // undef value.  This can happen when loading for a fresh allocation with no
01980   // intervening stores, for example.
01981   if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
01982     L->replaceAllUsesWith(UndefValue::get(L->getType()));
01983     markInstructionForDeletion(L);
01984     ++NumGVNLoad;
01985     return true;
01986   }
01987 
01988   // If this load occurs either right after a lifetime begin,
01989   // then the loaded value is undefined.
01990   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
01991     if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
01992       L->replaceAllUsesWith(UndefValue::get(L->getType()));
01993       markInstructionForDeletion(L);
01994       ++NumGVNLoad;
01995       return true;
01996     }
01997   }
01998 
01999   // If this load follows a calloc (which zero initializes memory),
02000   // then the loaded value is zero
02001   if (isCallocLikeFn(DepInst, TLI)) {
02002     L->replaceAllUsesWith(Constant::getNullValue(L->getType()));
02003     markInstructionForDeletion(L);
02004     ++NumGVNLoad;
02005     return true;
02006   }
02007 
02008   return false;
02009 }
02010 
02011 // In order to find a leader for a given value number at a
02012 // specific basic block, we first obtain the list of all Values for that number,
02013 // and then scan the list to find one whose block dominates the block in
02014 // question.  This is fast because dominator tree queries consist of only
02015 // a few comparisons of DFS numbers.
02016 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
02017   LeaderTableEntry Vals = LeaderTable[num];
02018   if (!Vals.Val) return nullptr;
02019 
02020   Value *Val = nullptr;
02021   if (DT->dominates(Vals.BB, BB)) {
02022     Val = Vals.Val;
02023     if (isa<Constant>(Val)) return Val;
02024   }
02025 
02026   LeaderTableEntry* Next = Vals.Next;
02027   while (Next) {
02028     if (DT->dominates(Next->BB, BB)) {
02029       if (isa<Constant>(Next->Val)) return Next->Val;
02030       if (!Val) Val = Next->Val;
02031     }
02032 
02033     Next = Next->Next;
02034   }
02035 
02036   return Val;
02037 }
02038 
02039 /// There is an edge from 'Src' to 'Dst'.  Return
02040 /// true if every path from the entry block to 'Dst' passes via this edge.  In
02041 /// particular 'Dst' must not be reachable via another edge from 'Src'.
02042 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
02043                                        DominatorTree *DT) {
02044   // While in theory it is interesting to consider the case in which Dst has
02045   // more than one predecessor, because Dst might be part of a loop which is
02046   // only reachable from Src, in practice it is pointless since at the time
02047   // GVN runs all such loops have preheaders, which means that Dst will have
02048   // been changed to have only one predecessor, namely Src.
02049   const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
02050   const BasicBlock *Src = E.getStart();
02051   assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
02052   (void)Src;
02053   return Pred != nullptr;
02054 }
02055 
02056 /// The given values are known to be equal in every block
02057 /// dominated by 'Root'.  Exploit this, for example by replacing 'LHS' with
02058 /// 'RHS' everywhere in the scope.  Returns whether a change was made.
02059 bool GVN::propagateEquality(Value *LHS, Value *RHS,
02060                             const BasicBlockEdge &Root) {
02061   SmallVector<std::pair<Value*, Value*>, 4> Worklist;
02062   Worklist.push_back(std::make_pair(LHS, RHS));
02063   bool Changed = false;
02064   // For speed, compute a conservative fast approximation to
02065   // DT->dominates(Root, Root.getEnd());
02066   bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
02067 
02068   while (!Worklist.empty()) {
02069     std::pair<Value*, Value*> Item = Worklist.pop_back_val();
02070     LHS = Item.first; RHS = Item.second;
02071 
02072     if (LHS == RHS) continue;
02073     assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
02074 
02075     // Don't try to propagate equalities between constants.
02076     if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue;
02077 
02078     // Prefer a constant on the right-hand side, or an Argument if no constants.
02079     if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
02080       std::swap(LHS, RHS);
02081     assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
02082 
02083     // If there is no obvious reason to prefer the left-hand side over the
02084     // right-hand side, ensure the longest lived term is on the right-hand side,
02085     // so the shortest lived term will be replaced by the longest lived.
02086     // This tends to expose more simplifications.
02087     uint32_t LVN = VN.lookup_or_add(LHS);
02088     if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
02089         (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
02090       // Move the 'oldest' value to the right-hand side, using the value number
02091       // as a proxy for age.
02092       uint32_t RVN = VN.lookup_or_add(RHS);
02093       if (LVN < RVN) {
02094         std::swap(LHS, RHS);
02095         LVN = RVN;
02096       }
02097     }
02098 
02099     // If value numbering later sees that an instruction in the scope is equal
02100     // to 'LHS' then ensure it will be turned into 'RHS'.  In order to preserve
02101     // the invariant that instructions only occur in the leader table for their
02102     // own value number (this is used by removeFromLeaderTable), do not do this
02103     // if RHS is an instruction (if an instruction in the scope is morphed into
02104     // LHS then it will be turned into RHS by the next GVN iteration anyway, so
02105     // using the leader table is about compiling faster, not optimizing better).
02106     // The leader table only tracks basic blocks, not edges. Only add to if we
02107     // have the simple case where the edge dominates the end.
02108     if (RootDominatesEnd && !isa<Instruction>(RHS))
02109       addToLeaderTable(LVN, RHS, Root.getEnd());
02110 
02111     // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope.  As
02112     // LHS always has at least one use that is not dominated by Root, this will
02113     // never do anything if LHS has only one use.
02114     if (!LHS->hasOneUse()) {
02115       unsigned NumReplacements = replaceDominatedUsesWith(LHS, RHS, *DT, Root);
02116       Changed |= NumReplacements > 0;
02117       NumGVNEqProp += NumReplacements;
02118     }
02119 
02120     // Now try to deduce additional equalities from this one. For example, if
02121     // the known equality was "(A != B)" == "false" then it follows that A and B
02122     // are equal in the scope. Only boolean equalities with an explicit true or
02123     // false RHS are currently supported.
02124     if (!RHS->getType()->isIntegerTy(1))
02125       // Not a boolean equality - bail out.
02126       continue;
02127     ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
02128     if (!CI)
02129       // RHS neither 'true' nor 'false' - bail out.
02130       continue;
02131     // Whether RHS equals 'true'.  Otherwise it equals 'false'.
02132     bool isKnownTrue = CI->isAllOnesValue();
02133     bool isKnownFalse = !isKnownTrue;
02134 
02135     // If "A && B" is known true then both A and B are known true.  If "A || B"
02136     // is known false then both A and B are known false.
02137     Value *A, *B;
02138     if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
02139         (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
02140       Worklist.push_back(std::make_pair(A, RHS));
02141       Worklist.push_back(std::make_pair(B, RHS));
02142       continue;
02143     }
02144 
02145     // If we are propagating an equality like "(A == B)" == "true" then also
02146     // propagate the equality A == B.  When propagating a comparison such as
02147     // "(A >= B)" == "true", replace all instances of "A < B" with "false".
02148     if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
02149       Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
02150 
02151       // If "A == B" is known true, or "A != B" is known false, then replace
02152       // A with B everywhere in the scope.
02153       if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
02154           (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
02155         Worklist.push_back(std::make_pair(Op0, Op1));
02156 
02157       // Handle the floating point versions of equality comparisons too.
02158       if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
02159           (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
02160 
02161         // Floating point -0.0 and 0.0 compare equal, so we can only
02162         // propagate values if we know that we have a constant and that
02163         // its value is non-zero.
02164         
02165         // FIXME: We should do this optimization if 'no signed zeros' is
02166         // applicable via an instruction-level fast-math-flag or some other
02167         // indicator that relaxed FP semantics are being used.
02168 
02169         if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
02170           Worklist.push_back(std::make_pair(Op0, Op1));
02171       }
02172  
02173       // If "A >= B" is known true, replace "A < B" with false everywhere.
02174       CmpInst::Predicate NotPred = Cmp->getInversePredicate();
02175       Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
02176       // Since we don't have the instruction "A < B" immediately to hand, work
02177       // out the value number that it would have and use that to find an
02178       // appropriate instruction (if any).
02179       uint32_t NextNum = VN.getNextUnusedValueNumber();
02180       uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
02181       // If the number we were assigned was brand new then there is no point in
02182       // looking for an instruction realizing it: there cannot be one!
02183       if (Num < NextNum) {
02184         Value *NotCmp = findLeader(Root.getEnd(), Num);
02185         if (NotCmp && isa<Instruction>(NotCmp)) {
02186           unsigned NumReplacements =
02187             replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root);
02188           Changed |= NumReplacements > 0;
02189           NumGVNEqProp += NumReplacements;
02190         }
02191       }
02192       // Ensure that any instruction in scope that gets the "A < B" value number
02193       // is replaced with false.
02194       // The leader table only tracks basic blocks, not edges. Only add to if we
02195       // have the simple case where the edge dominates the end.
02196       if (RootDominatesEnd)
02197         addToLeaderTable(Num, NotVal, Root.getEnd());
02198 
02199       continue;
02200     }
02201   }
02202 
02203   return Changed;
02204 }
02205 
02206 /// When calculating availability, handle an instruction
02207 /// by inserting it into the appropriate sets
02208 bool GVN::processInstruction(Instruction *I) {
02209   // Ignore dbg info intrinsics.
02210   if (isa<DbgInfoIntrinsic>(I))
02211     return false;
02212 
02213   // If the instruction can be easily simplified then do so now in preference
02214   // to value numbering it.  Value numbering often exposes redundancies, for
02215   // example if it determines that %y is equal to %x then the instruction
02216   // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
02217   const DataLayout &DL = I->getModule()->getDataLayout();
02218   if (Value *V = SimplifyInstruction(I, DL, TLI, DT, AC)) {
02219     I->replaceAllUsesWith(V);
02220     if (MD && V->getType()->getScalarType()->isPointerTy())
02221       MD->invalidateCachedPointerInfo(V);
02222     markInstructionForDeletion(I);
02223     ++NumGVNSimpl;
02224     return true;
02225   }
02226 
02227   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
02228     if (processLoad(LI))
02229       return true;
02230 
02231     unsigned Num = VN.lookup_or_add(LI);
02232     addToLeaderTable(Num, LI, LI->getParent());
02233     return false;
02234   }
02235 
02236   // For conditional branches, we can perform simple conditional propagation on
02237   // the condition value itself.
02238   if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
02239     if (!BI->isConditional())
02240       return false;
02241 
02242     if (isa<Constant>(BI->getCondition()))
02243       return processFoldableCondBr(BI);
02244 
02245     Value *BranchCond = BI->getCondition();
02246     BasicBlock *TrueSucc = BI->getSuccessor(0);
02247     BasicBlock *FalseSucc = BI->getSuccessor(1);
02248     // Avoid multiple edges early.
02249     if (TrueSucc == FalseSucc)
02250       return false;
02251 
02252     BasicBlock *Parent = BI->getParent();
02253     bool Changed = false;
02254 
02255     Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
02256     BasicBlockEdge TrueE(Parent, TrueSucc);
02257     Changed |= propagateEquality(BranchCond, TrueVal, TrueE);
02258 
02259     Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
02260     BasicBlockEdge FalseE(Parent, FalseSucc);
02261     Changed |= propagateEquality(BranchCond, FalseVal, FalseE);
02262 
02263     return Changed;
02264   }
02265 
02266   // For switches, propagate the case values into the case destinations.
02267   if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
02268     Value *SwitchCond = SI->getCondition();
02269     BasicBlock *Parent = SI->getParent();
02270     bool Changed = false;
02271 
02272     // Remember how many outgoing edges there are to every successor.
02273     SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
02274     for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
02275       ++SwitchEdges[SI->getSuccessor(i)];
02276 
02277     for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
02278          i != e; ++i) {
02279       BasicBlock *Dst = i.getCaseSuccessor();
02280       // If there is only a single edge, propagate the case value into it.
02281       if (SwitchEdges.lookup(Dst) == 1) {
02282         BasicBlockEdge E(Parent, Dst);
02283         Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E);
02284       }
02285     }
02286     return Changed;
02287   }
02288 
02289   // Instructions with void type don't return a value, so there's
02290   // no point in trying to find redundancies in them.
02291   if (I->getType()->isVoidTy()) return false;
02292 
02293   uint32_t NextNum = VN.getNextUnusedValueNumber();
02294   unsigned Num = VN.lookup_or_add(I);
02295 
02296   // Allocations are always uniquely numbered, so we can save time and memory
02297   // by fast failing them.
02298   if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
02299     addToLeaderTable(Num, I, I->getParent());
02300     return false;
02301   }
02302 
02303   // If the number we were assigned was a brand new VN, then we don't
02304   // need to do a lookup to see if the number already exists
02305   // somewhere in the domtree: it can't!
02306   if (Num >= NextNum) {
02307     addToLeaderTable(Num, I, I->getParent());
02308     return false;
02309   }
02310 
02311   // Perform fast-path value-number based elimination of values inherited from
02312   // dominators.
02313   Value *repl = findLeader(I->getParent(), Num);
02314   if (!repl) {
02315     // Failure, just remember this instance for future use.
02316     addToLeaderTable(Num, I, I->getParent());
02317     return false;
02318   }
02319 
02320   // Remove it!
02321   patchAndReplaceAllUsesWith(I, repl);
02322   if (MD && repl->getType()->getScalarType()->isPointerTy())
02323     MD->invalidateCachedPointerInfo(repl);
02324   markInstructionForDeletion(I);
02325   return true;
02326 }
02327 
02328 /// runOnFunction - This is the main transformation entry point for a function.
02329 bool GVN::runOnFunction(Function& F) {
02330   if (skipOptnoneFunction(F))
02331     return false;
02332 
02333   if (!NoLoads)
02334     MD = &getAnalysis<MemoryDependenceAnalysis>();
02335   DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
02336   AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
02337   TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
02338   VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
02339   VN.setMemDep(MD);
02340   VN.setDomTree(DT);
02341 
02342   bool Changed = false;
02343   bool ShouldContinue = true;
02344 
02345   // Merge unconditional branches, allowing PRE to catch more
02346   // optimization opportunities.
02347   for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
02348     BasicBlock *BB = FI++;
02349 
02350     bool removedBlock = MergeBlockIntoPredecessor(
02351         BB, DT, /* LoopInfo */ nullptr, VN.getAliasAnalysis(), MD);
02352     if (removedBlock) ++NumGVNBlocks;
02353 
02354     Changed |= removedBlock;
02355   }
02356 
02357   unsigned Iteration = 0;
02358   while (ShouldContinue) {
02359     DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
02360     ShouldContinue = iterateOnFunction(F);
02361     Changed |= ShouldContinue;
02362     ++Iteration;
02363   }
02364 
02365   if (EnablePRE) {
02366     // Fabricate val-num for dead-code in order to suppress assertion in
02367     // performPRE().
02368     assignValNumForDeadCode();
02369     bool PREChanged = true;
02370     while (PREChanged) {
02371       PREChanged = performPRE(F);
02372       Changed |= PREChanged;
02373     }
02374   }
02375 
02376   // FIXME: Should perform GVN again after PRE does something.  PRE can move
02377   // computations into blocks where they become fully redundant.  Note that
02378   // we can't do this until PRE's critical edge splitting updates memdep.
02379   // Actually, when this happens, we should just fully integrate PRE into GVN.
02380 
02381   cleanupGlobalSets();
02382   // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
02383   // iteration. 
02384   DeadBlocks.clear();
02385 
02386   return Changed;
02387 }
02388 
02389 
02390 bool GVN::processBlock(BasicBlock *BB) {
02391   // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
02392   // (and incrementing BI before processing an instruction).
02393   assert(InstrsToErase.empty() &&
02394          "We expect InstrsToErase to be empty across iterations");
02395   if (DeadBlocks.count(BB))
02396     return false;
02397 
02398   bool ChangedFunction = false;
02399 
02400   for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
02401        BI != BE;) {
02402     ChangedFunction |= processInstruction(BI);
02403     if (InstrsToErase.empty()) {
02404       ++BI;
02405       continue;
02406     }
02407 
02408     // If we need some instructions deleted, do it now.
02409     NumGVNInstr += InstrsToErase.size();
02410 
02411     // Avoid iterator invalidation.
02412     bool AtStart = BI == BB->begin();
02413     if (!AtStart)
02414       --BI;
02415 
02416     for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
02417          E = InstrsToErase.end(); I != E; ++I) {
02418       DEBUG(dbgs() << "GVN removed: " << **I << '\n');
02419       if (MD) MD->removeInstruction(*I);
02420       DEBUG(verifyRemoved(*I));
02421       (*I)->eraseFromParent();
02422     }
02423     InstrsToErase.clear();
02424 
02425     if (AtStart)
02426       BI = BB->begin();
02427     else
02428       ++BI;
02429   }
02430 
02431   return ChangedFunction;
02432 }
02433 
02434 // Instantiate an expression in a predecessor that lacked it.
02435 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
02436                                     unsigned int ValNo) {
02437   // Because we are going top-down through the block, all value numbers
02438   // will be available in the predecessor by the time we need them.  Any
02439   // that weren't originally present will have been instantiated earlier
02440   // in this loop.
02441   bool success = true;
02442   for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
02443     Value *Op = Instr->getOperand(i);
02444     if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
02445       continue;
02446 
02447     if (Value *V = findLeader(Pred, VN.lookup(Op))) {
02448       Instr->setOperand(i, V);
02449     } else {
02450       success = false;
02451       break;
02452     }
02453   }
02454 
02455   // Fail out if we encounter an operand that is not available in
02456   // the PRE predecessor.  This is typically because of loads which
02457   // are not value numbered precisely.
02458   if (!success)
02459     return false;
02460 
02461   Instr->insertBefore(Pred->getTerminator());
02462   Instr->setName(Instr->getName() + ".pre");
02463   Instr->setDebugLoc(Instr->getDebugLoc());
02464   VN.add(Instr, ValNo);
02465 
02466   // Update the availability map to include the new instruction.
02467   addToLeaderTable(ValNo, Instr, Pred);
02468   return true;
02469 }
02470 
02471 bool GVN::performScalarPRE(Instruction *CurInst) {
02472   SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap;
02473 
02474   if (isa<AllocaInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
02475       isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
02476       CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
02477       isa<DbgInfoIntrinsic>(CurInst))
02478     return false;
02479 
02480   // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
02481   // sinking the compare again, and it would force the code generator to
02482   // move the i1 from processor flags or predicate registers into a general
02483   // purpose register.
02484   if (isa<CmpInst>(CurInst))
02485     return false;
02486 
02487   // We don't currently value number ANY inline asm calls.
02488   if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
02489     if (CallI->isInlineAsm())
02490       return false;
02491 
02492   uint32_t ValNo = VN.lookup(CurInst);
02493 
02494   // Look for the predecessors for PRE opportunities.  We're
02495   // only trying to solve the basic diamond case, where
02496   // a value is computed in the successor and one predecessor,
02497   // but not the other.  We also explicitly disallow cases
02498   // where the successor is its own predecessor, because they're
02499   // more complicated to get right.
02500   unsigned NumWith = 0;
02501   unsigned NumWithout = 0;
02502   BasicBlock *PREPred = nullptr;
02503   BasicBlock *CurrentBlock = CurInst->getParent();
02504   predMap.clear();
02505 
02506   for (pred_iterator PI = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock);
02507        PI != PE; ++PI) {
02508     BasicBlock *P = *PI;
02509     // We're not interested in PRE where the block is its
02510     // own predecessor, or in blocks with predecessors
02511     // that are not reachable.
02512     if (P == CurrentBlock) {
02513       NumWithout = 2;
02514       break;
02515     } else if (!DT->isReachableFromEntry(P)) {
02516       NumWithout = 2;
02517       break;
02518     }
02519 
02520     Value *predV = findLeader(P, ValNo);
02521     if (!predV) {
02522       predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
02523       PREPred = P;
02524       ++NumWithout;
02525     } else if (predV == CurInst) {
02526       /* CurInst dominates this predecessor. */
02527       NumWithout = 2;
02528       break;
02529     } else {
02530       predMap.push_back(std::make_pair(predV, P));
02531       ++NumWith;
02532     }
02533   }
02534 
02535   // Don't do PRE when it might increase code size, i.e. when
02536   // we would need to insert instructions in more than one pred.
02537   if (NumWithout > 1 || NumWith == 0)
02538     return false;
02539 
02540   // We may have a case where all predecessors have the instruction,
02541   // and we just need to insert a phi node. Otherwise, perform
02542   // insertion.
02543   Instruction *PREInstr = nullptr;
02544 
02545   if (NumWithout != 0) {
02546     // Don't do PRE across indirect branch.
02547     if (isa<IndirectBrInst>(PREPred->getTerminator()))
02548       return false;
02549 
02550     // We can't do PRE safely on a critical edge, so instead we schedule
02551     // the edge to be split and perform the PRE the next time we iterate
02552     // on the function.
02553     unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
02554     if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
02555       toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
02556       return false;
02557     }
02558     // We need to insert somewhere, so let's give it a shot
02559     PREInstr = CurInst->clone();
02560     if (!performScalarPREInsertion(PREInstr, PREPred, ValNo)) {
02561       // If we failed insertion, make sure we remove the instruction.
02562       DEBUG(verifyRemoved(PREInstr));
02563       delete PREInstr;
02564       return false;
02565     }
02566   }
02567 
02568   // Either we should have filled in the PRE instruction, or we should
02569   // not have needed insertions.
02570   assert (PREInstr != nullptr || NumWithout == 0);
02571 
02572   ++NumGVNPRE;
02573 
02574   // Create a PHI to make the value available in this block.
02575   PHINode *Phi =
02576       PHINode::Create(CurInst->getType(), predMap.size(),
02577                       CurInst->getName() + ".pre-phi", CurrentBlock->begin());
02578   for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
02579     if (Value *V = predMap[i].first)
02580       Phi->addIncoming(V, predMap[i].second);
02581     else
02582       Phi->addIncoming(PREInstr, PREPred);
02583   }
02584 
02585   VN.add(Phi, ValNo);
02586   addToLeaderTable(ValNo, Phi, CurrentBlock);
02587   Phi->setDebugLoc(CurInst->getDebugLoc());
02588   CurInst->replaceAllUsesWith(Phi);
02589   if (Phi->getType()->getScalarType()->isPointerTy()) {
02590     // Because we have added a PHI-use of the pointer value, it has now
02591     // "escaped" from alias analysis' perspective.  We need to inform
02592     // AA of this.
02593     for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; ++ii) {
02594       unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
02595       VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
02596     }
02597 
02598     if (MD)
02599       MD->invalidateCachedPointerInfo(Phi);
02600   }
02601   VN.erase(CurInst);
02602   removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
02603 
02604   DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
02605   if (MD)
02606     MD->removeInstruction(CurInst);
02607   DEBUG(verifyRemoved(CurInst));
02608   CurInst->eraseFromParent();
02609   ++NumGVNInstr;
02610   
02611   return true;
02612 }
02613 
02614 /// Perform a purely local form of PRE that looks for diamond
02615 /// control flow patterns and attempts to perform simple PRE at the join point.
02616 bool GVN::performPRE(Function &F) {
02617   bool Changed = false;
02618   for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
02619     // Nothing to PRE in the entry block.
02620     if (CurrentBlock == &F.getEntryBlock())
02621       continue;
02622 
02623     // Don't perform PRE on a landing pad.
02624     if (CurrentBlock->isLandingPad())
02625       continue;
02626 
02627     for (BasicBlock::iterator BI = CurrentBlock->begin(),
02628                               BE = CurrentBlock->end();
02629          BI != BE;) {
02630       Instruction *CurInst = BI++;
02631       Changed = performScalarPRE(CurInst);
02632     }
02633   }
02634 
02635   if (splitCriticalEdges())
02636     Changed = true;
02637 
02638   return Changed;
02639 }
02640 
02641 /// Split the critical edge connecting the given two blocks, and return
02642 /// the block inserted to the critical edge.
02643 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
02644   BasicBlock *BB = SplitCriticalEdge(
02645       Pred, Succ, CriticalEdgeSplittingOptions(getAliasAnalysis(), DT));
02646   if (MD)
02647     MD->invalidateCachedPredecessors();
02648   return BB;
02649 }
02650 
02651 /// Split critical edges found during the previous
02652 /// iteration that may enable further optimization.
02653 bool GVN::splitCriticalEdges() {
02654   if (toSplit.empty())
02655     return false;
02656   do {
02657     std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
02658     SplitCriticalEdge(Edge.first, Edge.second,
02659                       CriticalEdgeSplittingOptions(getAliasAnalysis(), DT));
02660   } while (!toSplit.empty());
02661   if (MD) MD->invalidateCachedPredecessors();
02662   return true;
02663 }
02664 
02665 /// Executes one iteration of GVN
02666 bool GVN::iterateOnFunction(Function &F) {
02667   cleanupGlobalSets();
02668 
02669   // Top-down walk of the dominator tree
02670   bool Changed = false;
02671   // Save the blocks this function have before transformation begins. GVN may
02672   // split critical edge, and hence may invalidate the RPO/DT iterator.
02673   //
02674   std::vector<BasicBlock *> BBVect;
02675   BBVect.reserve(256);
02676   // Needed for value numbering with phi construction to work.
02677   ReversePostOrderTraversal<Function *> RPOT(&F);
02678   for (ReversePostOrderTraversal<Function *>::rpo_iterator RI = RPOT.begin(),
02679                                                            RE = RPOT.end();
02680        RI != RE; ++RI)
02681     BBVect.push_back(*RI);
02682 
02683   for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
02684        I != E; I++)
02685     Changed |= processBlock(*I);
02686 
02687   return Changed;
02688 }
02689 
02690 void GVN::cleanupGlobalSets() {
02691   VN.clear();
02692   LeaderTable.clear();
02693   TableAllocator.Reset();
02694 }
02695 
02696 /// Verify that the specified instruction does not occur in our
02697 /// internal data structures.
02698 void GVN::verifyRemoved(const Instruction *Inst) const {
02699   VN.verifyRemoved(Inst);
02700 
02701   // Walk through the value number scope to make sure the instruction isn't
02702   // ferreted away in it.
02703   for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
02704        I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
02705     const LeaderTableEntry *Node = &I->second;
02706     assert(Node->Val != Inst && "Inst still in value numbering scope!");
02707 
02708     while (Node->Next) {
02709       Node = Node->Next;
02710       assert(Node->Val != Inst && "Inst still in value numbering scope!");
02711     }
02712   }
02713 }
02714 
02715 /// BB is declared dead, which implied other blocks become dead as well. This
02716 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
02717 /// live successors, update their phi nodes by replacing the operands
02718 /// corresponding to dead blocks with UndefVal.
02719 void GVN::addDeadBlock(BasicBlock *BB) {
02720   SmallVector<BasicBlock *, 4> NewDead;
02721   SmallSetVector<BasicBlock *, 4> DF;
02722 
02723   NewDead.push_back(BB);
02724   while (!NewDead.empty()) {
02725     BasicBlock *D = NewDead.pop_back_val();
02726     if (DeadBlocks.count(D))
02727       continue;
02728 
02729     // All blocks dominated by D are dead.
02730     SmallVector<BasicBlock *, 8> Dom;
02731     DT->getDescendants(D, Dom);
02732     DeadBlocks.insert(Dom.begin(), Dom.end());
02733     
02734     // Figure out the dominance-frontier(D).
02735     for (SmallVectorImpl<BasicBlock *>::iterator I = Dom.begin(),
02736            E = Dom.end(); I != E; I++) {
02737       BasicBlock *B = *I;
02738       for (succ_iterator SI = succ_begin(B), SE = succ_end(B); SI != SE; SI++) {
02739         BasicBlock *S = *SI;
02740         if (DeadBlocks.count(S))
02741           continue;
02742 
02743         bool AllPredDead = true;
02744         for (pred_iterator PI = pred_begin(S), PE = pred_end(S); PI != PE; PI++)
02745           if (!DeadBlocks.count(*PI)) {
02746             AllPredDead = false;
02747             break;
02748           }
02749 
02750         if (!AllPredDead) {
02751           // S could be proved dead later on. That is why we don't update phi
02752           // operands at this moment.
02753           DF.insert(S);
02754         } else {
02755           // While S is not dominated by D, it is dead by now. This could take
02756           // place if S already have a dead predecessor before D is declared
02757           // dead.
02758           NewDead.push_back(S);
02759         }
02760       }
02761     }
02762   }
02763 
02764   // For the dead blocks' live successors, update their phi nodes by replacing
02765   // the operands corresponding to dead blocks with UndefVal.
02766   for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
02767         I != E; I++) {
02768     BasicBlock *B = *I;
02769     if (DeadBlocks.count(B))
02770       continue;
02771 
02772     SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
02773     for (SmallVectorImpl<BasicBlock *>::iterator PI = Preds.begin(),
02774            PE = Preds.end(); PI != PE; PI++) {
02775       BasicBlock *P = *PI;
02776 
02777       if (!DeadBlocks.count(P))
02778         continue;
02779 
02780       if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
02781         if (BasicBlock *S = splitCriticalEdges(P, B))
02782           DeadBlocks.insert(P = S);
02783       }
02784 
02785       for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
02786         PHINode &Phi = cast<PHINode>(*II);
02787         Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
02788                              UndefValue::get(Phi.getType()));
02789       }
02790     }
02791   }
02792 }
02793 
02794 // If the given branch is recognized as a foldable branch (i.e. conditional
02795 // branch with constant condition), it will perform following analyses and
02796 // transformation.
02797 //  1) If the dead out-coming edge is a critical-edge, split it. Let 
02798 //     R be the target of the dead out-coming edge.
02799 //  1) Identify the set of dead blocks implied by the branch's dead outcoming
02800 //     edge. The result of this step will be {X| X is dominated by R}
02801 //  2) Identify those blocks which haves at least one dead prodecessor. The
02802 //     result of this step will be dominance-frontier(R).
02803 //  3) Update the PHIs in DF(R) by replacing the operands corresponding to 
02804 //     dead blocks with "UndefVal" in an hope these PHIs will optimized away.
02805 //
02806 // Return true iff *NEW* dead code are found.
02807 bool GVN::processFoldableCondBr(BranchInst *BI) {
02808   if (!BI || BI->isUnconditional())
02809     return false;
02810 
02811   ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
02812   if (!Cond)
02813     return false;
02814 
02815   BasicBlock *DeadRoot = Cond->getZExtValue() ? 
02816                          BI->getSuccessor(1) : BI->getSuccessor(0);
02817   if (DeadBlocks.count(DeadRoot))
02818     return false;
02819 
02820   if (!DeadRoot->getSinglePredecessor())
02821     DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
02822 
02823   addDeadBlock(DeadRoot);
02824   return true;
02825 }
02826 
02827 // performPRE() will trigger assert if it comes across an instruction without
02828 // associated val-num. As it normally has far more live instructions than dead
02829 // instructions, it makes more sense just to "fabricate" a val-number for the
02830 // dead code than checking if instruction involved is dead or not.
02831 void GVN::assignValNumForDeadCode() {
02832   for (SetVector<BasicBlock *>::iterator I = DeadBlocks.begin(),
02833         E = DeadBlocks.end(); I != E; I++) {
02834     BasicBlock *BB = *I;
02835     for (BasicBlock::iterator II = BB->begin(), EE = BB->end();
02836           II != EE; II++) {
02837       Instruction *Inst = &*II;
02838       unsigned ValNum = VN.lookup_or_add(Inst);
02839       addToLeaderTable(ValNum, Inst, BB);
02840     }
02841   }
02842 }