• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 //===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This pass performs global value numbering to eliminate fully redundant
11 // instructions.  It also performs simple dead load elimination.
12 //
13 // Note that this pass does the value numbering itself; it does not use the
14 // ValueNumbering analysis passes.
15 //
16 //===----------------------------------------------------------------------===//
17 
18 #include "llvm/Transforms/Scalar/GVN.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/DepthFirstIterator.h"
21 #include "llvm/ADT/Hashing.h"
22 #include "llvm/ADT/MapVector.h"
23 #include "llvm/ADT/PostOrderIterator.h"
24 #include "llvm/ADT/SetVector.h"
25 #include "llvm/ADT/SmallPtrSet.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/AssumptionCache.h"
29 #include "llvm/Analysis/CFG.h"
30 #include "llvm/Analysis/ConstantFolding.h"
31 #include "llvm/Analysis/GlobalsModRef.h"
32 #include "llvm/Analysis/InstructionSimplify.h"
33 #include "llvm/Analysis/Loads.h"
34 #include "llvm/Analysis/MemoryBuiltins.h"
35 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
36 #include "llvm/Analysis/PHITransAddr.h"
37 #include "llvm/Analysis/TargetLibraryInfo.h"
38 #include "llvm/Analysis/ValueTracking.h"
39 #include "llvm/IR/DataLayout.h"
40 #include "llvm/IR/Dominators.h"
41 #include "llvm/IR/GlobalVariable.h"
42 #include "llvm/IR/IRBuilder.h"
43 #include "llvm/IR/IntrinsicInst.h"
44 #include "llvm/IR/LLVMContext.h"
45 #include "llvm/IR/Metadata.h"
46 #include "llvm/IR/PatternMatch.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
51 #include "llvm/Transforms/Utils/Local.h"
52 #include "llvm/Transforms/Utils/SSAUpdater.h"
53 #include <vector>
54 using namespace llvm;
55 using namespace llvm::gvn;
56 using namespace PatternMatch;
57 
58 #define DEBUG_TYPE "gvn"
59 
60 STATISTIC(NumGVNInstr,  "Number of instructions deleted");
61 STATISTIC(NumGVNLoad,   "Number of loads deleted");
62 STATISTIC(NumGVNPRE,    "Number of instructions PRE'd");
63 STATISTIC(NumGVNBlocks, "Number of blocks merged");
64 STATISTIC(NumGVNSimpl,  "Number of instructions simplified");
65 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
66 STATISTIC(NumPRELoad,   "Number of loads PRE'd");
67 
68 static cl::opt<bool> EnablePRE("enable-pre",
69                                cl::init(true), cl::Hidden);
70 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
71 
72 // Maximum allowed recursion depth.
73 static cl::opt<uint32_t>
74 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
75                 cl::desc("Max recurse depth (default = 1000)"));
76 
77 struct llvm::GVN::Expression {
78   uint32_t opcode;
79   Type *type;
80   SmallVector<uint32_t, 4> varargs;
81 
Expressionllvm::GVN::Expression82   Expression(uint32_t o = ~2U) : opcode(o) {}
83 
operator ==llvm::GVN::Expression84   bool operator==(const Expression &other) const {
85     if (opcode != other.opcode)
86       return false;
87     if (opcode == ~0U || opcode == ~1U)
88       return true;
89     if (type != other.type)
90       return false;
91     if (varargs != other.varargs)
92       return false;
93     return true;
94   }
95 
hash_value(const Expression & Value)96   friend hash_code hash_value(const Expression &Value) {
97     return hash_combine(
98         Value.opcode, Value.type,
99         hash_combine_range(Value.varargs.begin(), Value.varargs.end()));
100   }
101 };
102 
103 namespace llvm {
104 template <> struct DenseMapInfo<GVN::Expression> {
getEmptyKeyllvm::DenseMapInfo105   static inline GVN::Expression getEmptyKey() { return ~0U; }
106 
getTombstoneKeyllvm::DenseMapInfo107   static inline GVN::Expression getTombstoneKey() { return ~1U; }
108 
getHashValuellvm::DenseMapInfo109   static unsigned getHashValue(const GVN::Expression &e) {
110     using llvm::hash_value;
111     return static_cast<unsigned>(hash_value(e));
112   }
isEqualllvm::DenseMapInfo113   static bool isEqual(const GVN::Expression &LHS, const GVN::Expression &RHS) {
114     return LHS == RHS;
115   }
116 };
117 } // End llvm namespace.
118 
119 /// Represents a particular available value that we know how to materialize.
120 /// Materialization of an AvailableValue never fails.  An AvailableValue is
121 /// implicitly associated with a rematerialization point which is the
122 /// location of the instruction from which it was formed.
123 struct llvm::gvn::AvailableValue {
124   enum ValType {
125     SimpleVal, // A simple offsetted value that is accessed.
126     LoadVal,   // A value produced by a load.
127     MemIntrin, // A memory intrinsic which is loaded from.
128     UndefVal   // A UndefValue representing a value from dead block (which
129                // is not yet physically removed from the CFG).
130   };
131 
132   /// V - The value that is live out of the block.
133   PointerIntPair<Value *, 2, ValType> Val;
134 
135   /// Offset - The byte offset in Val that is interesting for the load query.
136   unsigned Offset;
137 
getllvm::gvn::AvailableValue138   static AvailableValue get(Value *V, unsigned Offset = 0) {
139     AvailableValue Res;
140     Res.Val.setPointer(V);
141     Res.Val.setInt(SimpleVal);
142     Res.Offset = Offset;
143     return Res;
144   }
145 
getMIllvm::gvn::AvailableValue146   static AvailableValue getMI(MemIntrinsic *MI, unsigned Offset = 0) {
147     AvailableValue Res;
148     Res.Val.setPointer(MI);
149     Res.Val.setInt(MemIntrin);
150     Res.Offset = Offset;
151     return Res;
152   }
153 
getLoadllvm::gvn::AvailableValue154   static AvailableValue getLoad(LoadInst *LI, unsigned Offset = 0) {
155     AvailableValue Res;
156     Res.Val.setPointer(LI);
157     Res.Val.setInt(LoadVal);
158     Res.Offset = Offset;
159     return Res;
160   }
161 
getUndefllvm::gvn::AvailableValue162   static AvailableValue getUndef() {
163     AvailableValue Res;
164     Res.Val.setPointer(nullptr);
165     Res.Val.setInt(UndefVal);
166     Res.Offset = 0;
167     return Res;
168   }
169 
isSimpleValuellvm::gvn::AvailableValue170   bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
isCoercedLoadValuellvm::gvn::AvailableValue171   bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
isMemIntrinValuellvm::gvn::AvailableValue172   bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
isUndefValuellvm::gvn::AvailableValue173   bool isUndefValue() const { return Val.getInt() == UndefVal; }
174 
getSimpleValuellvm::gvn::AvailableValue175   Value *getSimpleValue() const {
176     assert(isSimpleValue() && "Wrong accessor");
177     return Val.getPointer();
178   }
179 
getCoercedLoadValuellvm::gvn::AvailableValue180   LoadInst *getCoercedLoadValue() const {
181     assert(isCoercedLoadValue() && "Wrong accessor");
182     return cast<LoadInst>(Val.getPointer());
183   }
184 
getMemIntrinValuellvm::gvn::AvailableValue185   MemIntrinsic *getMemIntrinValue() const {
186     assert(isMemIntrinValue() && "Wrong accessor");
187     return cast<MemIntrinsic>(Val.getPointer());
188   }
189 
190   /// Emit code at the specified insertion point to adjust the value defined
191   /// here to the specified type. This handles various coercion cases.
192   Value *MaterializeAdjustedValue(LoadInst *LI, Instruction *InsertPt,
193                                   GVN &gvn) const;
194 };
195 
196 /// Represents an AvailableValue which can be rematerialized at the end of
197 /// the associated BasicBlock.
198 struct llvm::gvn::AvailableValueInBlock {
199   /// BB - The basic block in question.
200   BasicBlock *BB;
201 
202   /// AV - The actual available value
203   AvailableValue AV;
204 
getllvm::gvn::AvailableValueInBlock205   static AvailableValueInBlock get(BasicBlock *BB, AvailableValue &&AV) {
206     AvailableValueInBlock Res;
207     Res.BB = BB;
208     Res.AV = std::move(AV);
209     return Res;
210   }
211 
getllvm::gvn::AvailableValueInBlock212   static AvailableValueInBlock get(BasicBlock *BB, Value *V,
213                                    unsigned Offset = 0) {
214     return get(BB, AvailableValue::get(V, Offset));
215   }
getUndefllvm::gvn::AvailableValueInBlock216   static AvailableValueInBlock getUndef(BasicBlock *BB) {
217     return get(BB, AvailableValue::getUndef());
218   }
219 
220   /// Emit code at the end of this block to adjust the value defined here to
221   /// the specified type. This handles various coercion cases.
MaterializeAdjustedValuellvm::gvn::AvailableValueInBlock222   Value *MaterializeAdjustedValue(LoadInst *LI, GVN &gvn) const {
223     return AV.MaterializeAdjustedValue(LI, BB->getTerminator(), gvn);
224   }
225 };
226 
227 //===----------------------------------------------------------------------===//
228 //                     ValueTable Internal Functions
229 //===----------------------------------------------------------------------===//
230 
createExpr(Instruction * I)231 GVN::Expression GVN::ValueTable::createExpr(Instruction *I) {
232   Expression e;
233   e.type = I->getType();
234   e.opcode = I->getOpcode();
235   for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
236        OI != OE; ++OI)
237     e.varargs.push_back(lookupOrAdd(*OI));
238   if (I->isCommutative()) {
239     // Ensure that commutative instructions that only differ by a permutation
240     // of their operands get the same value number by sorting the operand value
241     // numbers.  Since all commutative instructions have two operands it is more
242     // efficient to sort by hand rather than using, say, std::sort.
243     assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
244     if (e.varargs[0] > e.varargs[1])
245       std::swap(e.varargs[0], e.varargs[1]);
246   }
247 
248   if (CmpInst *C = dyn_cast<CmpInst>(I)) {
249     // Sort the operand value numbers so x<y and y>x get the same value number.
250     CmpInst::Predicate Predicate = C->getPredicate();
251     if (e.varargs[0] > e.varargs[1]) {
252       std::swap(e.varargs[0], e.varargs[1]);
253       Predicate = CmpInst::getSwappedPredicate(Predicate);
254     }
255     e.opcode = (C->getOpcode() << 8) | Predicate;
256   } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
257     for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
258          II != IE; ++II)
259       e.varargs.push_back(*II);
260   }
261 
262   return e;
263 }
264 
createCmpExpr(unsigned Opcode,CmpInst::Predicate Predicate,Value * LHS,Value * RHS)265 GVN::Expression GVN::ValueTable::createCmpExpr(unsigned Opcode,
266                                                CmpInst::Predicate Predicate,
267                                                Value *LHS, Value *RHS) {
268   assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
269          "Not a comparison!");
270   Expression e;
271   e.type = CmpInst::makeCmpResultType(LHS->getType());
272   e.varargs.push_back(lookupOrAdd(LHS));
273   e.varargs.push_back(lookupOrAdd(RHS));
274 
275   // Sort the operand value numbers so x<y and y>x get the same value number.
276   if (e.varargs[0] > e.varargs[1]) {
277     std::swap(e.varargs[0], e.varargs[1]);
278     Predicate = CmpInst::getSwappedPredicate(Predicate);
279   }
280   e.opcode = (Opcode << 8) | Predicate;
281   return e;
282 }
283 
createExtractvalueExpr(ExtractValueInst * EI)284 GVN::Expression GVN::ValueTable::createExtractvalueExpr(ExtractValueInst *EI) {
285   assert(EI && "Not an ExtractValueInst?");
286   Expression e;
287   e.type = EI->getType();
288   e.opcode = 0;
289 
290   IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
291   if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
292     // EI might be an extract from one of our recognised intrinsics. If it
293     // is we'll synthesize a semantically equivalent expression instead on
294     // an extract value expression.
295     switch (I->getIntrinsicID()) {
296       case Intrinsic::sadd_with_overflow:
297       case Intrinsic::uadd_with_overflow:
298         e.opcode = Instruction::Add;
299         break;
300       case Intrinsic::ssub_with_overflow:
301       case Intrinsic::usub_with_overflow:
302         e.opcode = Instruction::Sub;
303         break;
304       case Intrinsic::smul_with_overflow:
305       case Intrinsic::umul_with_overflow:
306         e.opcode = Instruction::Mul;
307         break;
308       default:
309         break;
310     }
311 
312     if (e.opcode != 0) {
313       // Intrinsic recognized. Grab its args to finish building the expression.
314       assert(I->getNumArgOperands() == 2 &&
315              "Expect two args for recognised intrinsics.");
316       e.varargs.push_back(lookupOrAdd(I->getArgOperand(0)));
317       e.varargs.push_back(lookupOrAdd(I->getArgOperand(1)));
318       return e;
319     }
320   }
321 
322   // Not a recognised intrinsic. Fall back to producing an extract value
323   // expression.
324   e.opcode = EI->getOpcode();
325   for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
326        OI != OE; ++OI)
327     e.varargs.push_back(lookupOrAdd(*OI));
328 
329   for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
330          II != IE; ++II)
331     e.varargs.push_back(*II);
332 
333   return e;
334 }
335 
336 //===----------------------------------------------------------------------===//
337 //                     ValueTable External Functions
338 //===----------------------------------------------------------------------===//
339 
ValueTable()340 GVN::ValueTable::ValueTable() : nextValueNumber(1) {}
ValueTable(const ValueTable & Arg)341 GVN::ValueTable::ValueTable(const ValueTable &Arg)
342     : valueNumbering(Arg.valueNumbering),
343       expressionNumbering(Arg.expressionNumbering), AA(Arg.AA), MD(Arg.MD),
344       DT(Arg.DT), nextValueNumber(Arg.nextValueNumber) {}
ValueTable(ValueTable && Arg)345 GVN::ValueTable::ValueTable(ValueTable &&Arg)
346     : valueNumbering(std::move(Arg.valueNumbering)),
347       expressionNumbering(std::move(Arg.expressionNumbering)),
348       AA(std::move(Arg.AA)), MD(std::move(Arg.MD)), DT(std::move(Arg.DT)),
349       nextValueNumber(std::move(Arg.nextValueNumber)) {}
~ValueTable()350 GVN::ValueTable::~ValueTable() {}
351 
352 /// add - Insert a value into the table with a specified value number.
add(Value * V,uint32_t num)353 void GVN::ValueTable::add(Value *V, uint32_t num) {
354   valueNumbering.insert(std::make_pair(V, num));
355 }
356 
lookupOrAddCall(CallInst * C)357 uint32_t GVN::ValueTable::lookupOrAddCall(CallInst *C) {
358   if (AA->doesNotAccessMemory(C)) {
359     Expression exp = createExpr(C);
360     uint32_t &e = expressionNumbering[exp];
361     if (!e) e = nextValueNumber++;
362     valueNumbering[C] = e;
363     return e;
364   } else if (AA->onlyReadsMemory(C)) {
365     Expression exp = createExpr(C);
366     uint32_t &e = expressionNumbering[exp];
367     if (!e) {
368       e = nextValueNumber++;
369       valueNumbering[C] = e;
370       return e;
371     }
372     if (!MD) {
373       e = nextValueNumber++;
374       valueNumbering[C] = e;
375       return e;
376     }
377 
378     MemDepResult local_dep = MD->getDependency(C);
379 
380     if (!local_dep.isDef() && !local_dep.isNonLocal()) {
381       valueNumbering[C] =  nextValueNumber;
382       return nextValueNumber++;
383     }
384 
385     if (local_dep.isDef()) {
386       CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
387 
388       if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
389         valueNumbering[C] = nextValueNumber;
390         return nextValueNumber++;
391       }
392 
393       for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
394         uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
395         uint32_t cd_vn = lookupOrAdd(local_cdep->getArgOperand(i));
396         if (c_vn != cd_vn) {
397           valueNumbering[C] = nextValueNumber;
398           return nextValueNumber++;
399         }
400       }
401 
402       uint32_t v = lookupOrAdd(local_cdep);
403       valueNumbering[C] = v;
404       return v;
405     }
406 
407     // Non-local case.
408     const MemoryDependenceResults::NonLocalDepInfo &deps =
409       MD->getNonLocalCallDependency(CallSite(C));
410     // FIXME: Move the checking logic to MemDep!
411     CallInst* cdep = nullptr;
412 
413     // Check to see if we have a single dominating call instruction that is
414     // identical to C.
415     for (unsigned i = 0, e = deps.size(); i != e; ++i) {
416       const NonLocalDepEntry *I = &deps[i];
417       if (I->getResult().isNonLocal())
418         continue;
419 
420       // We don't handle non-definitions.  If we already have a call, reject
421       // instruction dependencies.
422       if (!I->getResult().isDef() || cdep != nullptr) {
423         cdep = nullptr;
424         break;
425       }
426 
427       CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
428       // FIXME: All duplicated with non-local case.
429       if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
430         cdep = NonLocalDepCall;
431         continue;
432       }
433 
434       cdep = nullptr;
435       break;
436     }
437 
438     if (!cdep) {
439       valueNumbering[C] = nextValueNumber;
440       return nextValueNumber++;
441     }
442 
443     if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
444       valueNumbering[C] = nextValueNumber;
445       return nextValueNumber++;
446     }
447     for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
448       uint32_t c_vn = lookupOrAdd(C->getArgOperand(i));
449       uint32_t cd_vn = lookupOrAdd(cdep->getArgOperand(i));
450       if (c_vn != cd_vn) {
451         valueNumbering[C] = nextValueNumber;
452         return nextValueNumber++;
453       }
454     }
455 
456     uint32_t v = lookupOrAdd(cdep);
457     valueNumbering[C] = v;
458     return v;
459 
460   } else {
461     valueNumbering[C] = nextValueNumber;
462     return nextValueNumber++;
463   }
464 }
465 
466 /// Returns true if a value number exists for the specified value.
exists(Value * V) const467 bool GVN::ValueTable::exists(Value *V) const { return valueNumbering.count(V) != 0; }
468 
469 /// lookup_or_add - Returns the value number for the specified value, assigning
470 /// it a new number if it did not have one before.
lookupOrAdd(Value * V)471 uint32_t GVN::ValueTable::lookupOrAdd(Value *V) {
472   DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
473   if (VI != valueNumbering.end())
474     return VI->second;
475 
476   if (!isa<Instruction>(V)) {
477     valueNumbering[V] = nextValueNumber;
478     return nextValueNumber++;
479   }
480 
481   Instruction* I = cast<Instruction>(V);
482   Expression exp;
483   switch (I->getOpcode()) {
484     case Instruction::Call:
485       return lookupOrAddCall(cast<CallInst>(I));
486     case Instruction::Add:
487     case Instruction::FAdd:
488     case Instruction::Sub:
489     case Instruction::FSub:
490     case Instruction::Mul:
491     case Instruction::FMul:
492     case Instruction::UDiv:
493     case Instruction::SDiv:
494     case Instruction::FDiv:
495     case Instruction::URem:
496     case Instruction::SRem:
497     case Instruction::FRem:
498     case Instruction::Shl:
499     case Instruction::LShr:
500     case Instruction::AShr:
501     case Instruction::And:
502     case Instruction::Or:
503     case Instruction::Xor:
504     case Instruction::ICmp:
505     case Instruction::FCmp:
506     case Instruction::Trunc:
507     case Instruction::ZExt:
508     case Instruction::SExt:
509     case Instruction::FPToUI:
510     case Instruction::FPToSI:
511     case Instruction::UIToFP:
512     case Instruction::SIToFP:
513     case Instruction::FPTrunc:
514     case Instruction::FPExt:
515     case Instruction::PtrToInt:
516     case Instruction::IntToPtr:
517     case Instruction::BitCast:
518     case Instruction::Select:
519     case Instruction::ExtractElement:
520     case Instruction::InsertElement:
521     case Instruction::ShuffleVector:
522     case Instruction::InsertValue:
523     case Instruction::GetElementPtr:
524       exp = createExpr(I);
525       break;
526     case Instruction::ExtractValue:
527       exp = createExtractvalueExpr(cast<ExtractValueInst>(I));
528       break;
529     default:
530       valueNumbering[V] = nextValueNumber;
531       return nextValueNumber++;
532   }
533 
534   uint32_t& e = expressionNumbering[exp];
535   if (!e) e = nextValueNumber++;
536   valueNumbering[V] = e;
537   return e;
538 }
539 
540 /// Returns the value number of the specified value. Fails if
541 /// the value has not yet been numbered.
lookup(Value * V) const542 uint32_t GVN::ValueTable::lookup(Value *V) const {
543   DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
544   assert(VI != valueNumbering.end() && "Value not numbered?");
545   return VI->second;
546 }
547 
548 /// Returns the value number of the given comparison,
549 /// assigning it a new number if it did not have one before.  Useful when
550 /// we deduced the result of a comparison, but don't immediately have an
551 /// instruction realizing that comparison to hand.
lookupOrAddCmp(unsigned Opcode,CmpInst::Predicate Predicate,Value * LHS,Value * RHS)552 uint32_t GVN::ValueTable::lookupOrAddCmp(unsigned Opcode,
553                                          CmpInst::Predicate Predicate,
554                                          Value *LHS, Value *RHS) {
555   Expression exp = createCmpExpr(Opcode, Predicate, LHS, RHS);
556   uint32_t& e = expressionNumbering[exp];
557   if (!e) e = nextValueNumber++;
558   return e;
559 }
560 
561 /// Remove all entries from the ValueTable.
clear()562 void GVN::ValueTable::clear() {
563   valueNumbering.clear();
564   expressionNumbering.clear();
565   nextValueNumber = 1;
566 }
567 
568 /// Remove a value from the value numbering.
erase(Value * V)569 void GVN::ValueTable::erase(Value *V) {
570   valueNumbering.erase(V);
571 }
572 
573 /// verifyRemoved - Verify that the value is removed from all internal data
574 /// structures.
verifyRemoved(const Value * V) const575 void GVN::ValueTable::verifyRemoved(const Value *V) const {
576   for (DenseMap<Value*, uint32_t>::const_iterator
577          I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
578     assert(I->first != V && "Inst still occurs in value numbering map!");
579   }
580 }
581 
582 //===----------------------------------------------------------------------===//
583 //                                GVN Pass
584 //===----------------------------------------------------------------------===//
585 
run(Function & F,AnalysisManager<Function> & AM)586 PreservedAnalyses GVN::run(Function &F, AnalysisManager<Function> &AM) {
587   // FIXME: The order of evaluation of these 'getResult' calls is very
588   // significant! Re-ordering these variables will cause GVN when run alone to
589   // be less effective! We should fix memdep and basic-aa to not exhibit this
590   // behavior, but until then don't change the order here.
591   auto &AC = AM.getResult<AssumptionAnalysis>(F);
592   auto &DT = AM.getResult<DominatorTreeAnalysis>(F);
593   auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
594   auto &AA = AM.getResult<AAManager>(F);
595   auto &MemDep = AM.getResult<MemoryDependenceAnalysis>(F);
596   bool Changed = runImpl(F, AC, DT, TLI, AA, &MemDep);
597   if (!Changed)
598     return PreservedAnalyses::all();
599   PreservedAnalyses PA;
600   PA.preserve<DominatorTreeAnalysis>();
601   PA.preserve<GlobalsAA>();
602   return PA;
603 }
604 
605 LLVM_DUMP_METHOD
dump(DenseMap<uint32_t,Value * > & d)606 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
607   errs() << "{\n";
608   for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
609        E = d.end(); I != E; ++I) {
610       errs() << I->first << "\n";
611       I->second->dump();
612   }
613   errs() << "}\n";
614 }
615 
616 /// Return true if we can prove that the value
617 /// we're analyzing is fully available in the specified block.  As we go, keep
618 /// track of which blocks we know are fully alive in FullyAvailableBlocks.  This
619 /// map is actually a tri-state map with the following values:
620 ///   0) we know the block *is not* fully available.
621 ///   1) we know the block *is* fully available.
622 ///   2) we do not know whether the block is fully available or not, but we are
623 ///      currently speculating that it will be.
624 ///   3) we are speculating for this block and have used that to speculate for
625 ///      other blocks.
IsValueFullyAvailableInBlock(BasicBlock * BB,DenseMap<BasicBlock *,char> & FullyAvailableBlocks,uint32_t RecurseDepth)626 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
627                             DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
628                             uint32_t RecurseDepth) {
629   if (RecurseDepth > MaxRecurseDepth)
630     return false;
631 
632   // Optimistically assume that the block is fully available and check to see
633   // if we already know about this block in one lookup.
634   std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
635     FullyAvailableBlocks.insert(std::make_pair(BB, 2));
636 
637   // If the entry already existed for this block, return the precomputed value.
638   if (!IV.second) {
639     // If this is a speculative "available" value, mark it as being used for
640     // speculation of other blocks.
641     if (IV.first->second == 2)
642       IV.first->second = 3;
643     return IV.first->second != 0;
644   }
645 
646   // Otherwise, see if it is fully available in all predecessors.
647   pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
648 
649   // If this block has no predecessors, it isn't live-in here.
650   if (PI == PE)
651     goto SpeculationFailure;
652 
653   for (; PI != PE; ++PI)
654     // If the value isn't fully available in one of our predecessors, then it
655     // isn't fully available in this block either.  Undo our previous
656     // optimistic assumption and bail out.
657     if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
658       goto SpeculationFailure;
659 
660   return true;
661 
662 // If we get here, we found out that this is not, after
663 // all, a fully-available block.  We have a problem if we speculated on this and
664 // used the speculation to mark other blocks as available.
665 SpeculationFailure:
666   char &BBVal = FullyAvailableBlocks[BB];
667 
668   // If we didn't speculate on this, just return with it set to false.
669   if (BBVal == 2) {
670     BBVal = 0;
671     return false;
672   }
673 
674   // If we did speculate on this value, we could have blocks set to 1 that are
675   // incorrect.  Walk the (transitive) successors of this block and mark them as
676   // 0 if set to one.
677   SmallVector<BasicBlock*, 32> BBWorklist;
678   BBWorklist.push_back(BB);
679 
680   do {
681     BasicBlock *Entry = BBWorklist.pop_back_val();
682     // Note that this sets blocks to 0 (unavailable) if they happen to not
683     // already be in FullyAvailableBlocks.  This is safe.
684     char &EntryVal = FullyAvailableBlocks[Entry];
685     if (EntryVal == 0) continue;  // Already unavailable.
686 
687     // Mark as unavailable.
688     EntryVal = 0;
689 
690     BBWorklist.append(succ_begin(Entry), succ_end(Entry));
691   } while (!BBWorklist.empty());
692 
693   return false;
694 }
695 
696 
697 /// Return true if CoerceAvailableValueToLoadType will succeed.
CanCoerceMustAliasedValueToLoad(Value * StoredVal,Type * LoadTy,const DataLayout & DL)698 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
699                                             Type *LoadTy,
700                                             const DataLayout &DL) {
701   // If the loaded or stored value is an first class array or struct, don't try
702   // to transform them.  We need to be able to bitcast to integer.
703   if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
704       StoredVal->getType()->isStructTy() ||
705       StoredVal->getType()->isArrayTy())
706     return false;
707 
708   // The store has to be at least as big as the load.
709   if (DL.getTypeSizeInBits(StoredVal->getType()) <
710         DL.getTypeSizeInBits(LoadTy))
711     return false;
712 
713   return true;
714 }
715 
716 /// If we saw a store of a value to memory, and
717 /// then a load from a must-aliased pointer of a different type, try to coerce
718 /// the stored value.  LoadedTy is the type of the load we want to replace.
719 /// IRB is IRBuilder used to insert new instructions.
720 ///
721 /// If we can't do it, return null.
CoerceAvailableValueToLoadType(Value * StoredVal,Type * LoadedTy,IRBuilder<> & IRB,const DataLayout & DL)722 static Value *CoerceAvailableValueToLoadType(Value *StoredVal, Type *LoadedTy,
723                                              IRBuilder<> &IRB,
724                                              const DataLayout &DL) {
725   assert(CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL) &&
726          "precondition violation - materialization can't fail");
727 
728   // If this is already the right type, just return it.
729   Type *StoredValTy = StoredVal->getType();
730 
731   uint64_t StoredValSize = DL.getTypeSizeInBits(StoredValTy);
732   uint64_t LoadedValSize = DL.getTypeSizeInBits(LoadedTy);
733 
734   // If the store and reload are the same size, we can always reuse it.
735   if (StoredValSize == LoadedValSize) {
736     // Pointer to Pointer -> use bitcast.
737     if (StoredValTy->getScalarType()->isPointerTy() &&
738         LoadedTy->getScalarType()->isPointerTy())
739       return IRB.CreateBitCast(StoredVal, LoadedTy);
740 
741     // Convert source pointers to integers, which can be bitcast.
742     if (StoredValTy->getScalarType()->isPointerTy()) {
743       StoredValTy = DL.getIntPtrType(StoredValTy);
744       StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
745     }
746 
747     Type *TypeToCastTo = LoadedTy;
748     if (TypeToCastTo->getScalarType()->isPointerTy())
749       TypeToCastTo = DL.getIntPtrType(TypeToCastTo);
750 
751     if (StoredValTy != TypeToCastTo)
752       StoredVal = IRB.CreateBitCast(StoredVal, TypeToCastTo);
753 
754     // Cast to pointer if the load needs a pointer type.
755     if (LoadedTy->getScalarType()->isPointerTy())
756       StoredVal = IRB.CreateIntToPtr(StoredVal, LoadedTy);
757 
758     return StoredVal;
759   }
760 
761   // If the loaded value is smaller than the available value, then we can
762   // extract out a piece from it.  If the available value is too small, then we
763   // can't do anything.
764   assert(StoredValSize >= LoadedValSize &&
765          "CanCoerceMustAliasedValueToLoad fail");
766 
767   // Convert source pointers to integers, which can be manipulated.
768   if (StoredValTy->getScalarType()->isPointerTy()) {
769     StoredValTy = DL.getIntPtrType(StoredValTy);
770     StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
771   }
772 
773   // Convert vectors and fp to integer, which can be manipulated.
774   if (!StoredValTy->isIntegerTy()) {
775     StoredValTy = IntegerType::get(StoredValTy->getContext(), StoredValSize);
776     StoredVal = IRB.CreateBitCast(StoredVal, StoredValTy);
777   }
778 
779   // If this is a big-endian system, we need to shift the value down to the low
780   // bits so that a truncate will work.
781   if (DL.isBigEndian()) {
782     uint64_t ShiftAmt = DL.getTypeStoreSizeInBits(StoredValTy) -
783                         DL.getTypeStoreSizeInBits(LoadedTy);
784     StoredVal = IRB.CreateLShr(StoredVal, ShiftAmt, "tmp");
785   }
786 
787   // Truncate the integer to the right size now.
788   Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadedValSize);
789   StoredVal  = IRB.CreateTrunc(StoredVal, NewIntTy, "trunc");
790 
791   if (LoadedTy == NewIntTy)
792     return StoredVal;
793 
794   // If the result is a pointer, inttoptr.
795   if (LoadedTy->getScalarType()->isPointerTy())
796     return IRB.CreateIntToPtr(StoredVal, LoadedTy, "inttoptr");
797 
798   // Otherwise, bitcast.
799   return IRB.CreateBitCast(StoredVal, LoadedTy, "bitcast");
800 }
801 
802 /// This function is called when we have a
803 /// memdep query of a load that ends up being a clobbering memory write (store,
804 /// memset, memcpy, memmove).  This means that the write *may* provide bits used
805 /// by the load but we can't be sure because the pointers don't mustalias.
806 ///
807 /// Check this case to see if there is anything more we can do before we give
808 /// up.  This returns -1 if we have to give up, or a byte number in the stored
809 /// value of the piece that feeds the load.
AnalyzeLoadFromClobberingWrite(Type * LoadTy,Value * LoadPtr,Value * WritePtr,uint64_t WriteSizeInBits,const DataLayout & DL)810 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
811                                           Value *WritePtr,
812                                           uint64_t WriteSizeInBits,
813                                           const DataLayout &DL) {
814   // If the loaded or stored value is a first class array or struct, don't try
815   // to transform them.  We need to be able to bitcast to integer.
816   if (LoadTy->isStructTy() || LoadTy->isArrayTy())
817     return -1;
818 
819   int64_t StoreOffset = 0, LoadOffset = 0;
820   Value *StoreBase =
821       GetPointerBaseWithConstantOffset(WritePtr, StoreOffset, DL);
822   Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, DL);
823   if (StoreBase != LoadBase)
824     return -1;
825 
826   // If the load and store are to the exact same address, they should have been
827   // a must alias.  AA must have gotten confused.
828   // FIXME: Study to see if/when this happens.  One case is forwarding a memset
829   // to a load from the base of the memset.
830 #if 0
831   if (LoadOffset == StoreOffset) {
832     dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
833     << "Base       = " << *StoreBase << "\n"
834     << "Store Ptr  = " << *WritePtr << "\n"
835     << "Store Offs = " << StoreOffset << "\n"
836     << "Load Ptr   = " << *LoadPtr << "\n";
837     abort();
838   }
839 #endif
840 
841   // If the load and store don't overlap at all, the store doesn't provide
842   // anything to the load.  In this case, they really don't alias at all, AA
843   // must have gotten confused.
844   uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy);
845 
846   if ((WriteSizeInBits & 7) | (LoadSize & 7))
847     return -1;
848   uint64_t StoreSize = WriteSizeInBits >> 3;  // Convert to bytes.
849   LoadSize >>= 3;
850 
851 
852   bool isAAFailure = false;
853   if (StoreOffset < LoadOffset)
854     isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
855   else
856     isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
857 
858   if (isAAFailure) {
859 #if 0
860     dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
861     << "Base       = " << *StoreBase << "\n"
862     << "Store Ptr  = " << *WritePtr << "\n"
863     << "Store Offs = " << StoreOffset << "\n"
864     << "Load Ptr   = " << *LoadPtr << "\n";
865     abort();
866 #endif
867     return -1;
868   }
869 
870   // If the Load isn't completely contained within the stored bits, we don't
871   // have all the bits to feed it.  We could do something crazy in the future
872   // (issue a smaller load then merge the bits in) but this seems unlikely to be
873   // valuable.
874   if (StoreOffset > LoadOffset ||
875       StoreOffset+StoreSize < LoadOffset+LoadSize)
876     return -1;
877 
878   // Okay, we can do this transformation.  Return the number of bytes into the
879   // store that the load is.
880   return LoadOffset-StoreOffset;
881 }
882 
883 /// This function is called when we have a
884 /// memdep query of a load that ends up being a clobbering store.
AnalyzeLoadFromClobberingStore(Type * LoadTy,Value * LoadPtr,StoreInst * DepSI)885 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
886                                           StoreInst *DepSI) {
887   // Cannot handle reading from store of first-class aggregate yet.
888   if (DepSI->getValueOperand()->getType()->isStructTy() ||
889       DepSI->getValueOperand()->getType()->isArrayTy())
890     return -1;
891 
892   const DataLayout &DL = DepSI->getModule()->getDataLayout();
893   Value *StorePtr = DepSI->getPointerOperand();
894   uint64_t StoreSize =DL.getTypeSizeInBits(DepSI->getValueOperand()->getType());
895   return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
896                                         StorePtr, StoreSize, DL);
897 }
898 
899 /// This function is called when we have a
900 /// memdep query of a load that ends up being clobbered by another load.  See if
901 /// the other load can feed into the second load.
AnalyzeLoadFromClobberingLoad(Type * LoadTy,Value * LoadPtr,LoadInst * DepLI,const DataLayout & DL)902 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
903                                          LoadInst *DepLI, const DataLayout &DL){
904   // Cannot handle reading from store of first-class aggregate yet.
905   if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
906     return -1;
907 
908   Value *DepPtr = DepLI->getPointerOperand();
909   uint64_t DepSize = DL.getTypeSizeInBits(DepLI->getType());
910   int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL);
911   if (R != -1) return R;
912 
913   // If we have a load/load clobber an DepLI can be widened to cover this load,
914   // then we should widen it!
915   int64_t LoadOffs = 0;
916   const Value *LoadBase =
917       GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, DL);
918   unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
919 
920   unsigned Size = MemoryDependenceResults::getLoadLoadClobberFullWidthSize(
921       LoadBase, LoadOffs, LoadSize, DepLI);
922   if (Size == 0) return -1;
923 
924   // Check non-obvious conditions enforced by MDA which we rely on for being
925   // able to materialize this potentially available value
926   assert(DepLI->isSimple() && "Cannot widen volatile/atomic load!");
927   assert(DepLI->getType()->isIntegerTy() && "Can't widen non-integer load");
928 
929   return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, DL);
930 }
931 
932 
933 
AnalyzeLoadFromClobberingMemInst(Type * LoadTy,Value * LoadPtr,MemIntrinsic * MI,const DataLayout & DL)934 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
935                                             MemIntrinsic *MI,
936                                             const DataLayout &DL) {
937   // If the mem operation is a non-constant size, we can't handle it.
938   ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
939   if (!SizeCst) return -1;
940   uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
941 
942   // If this is memset, we just need to see if the offset is valid in the size
943   // of the memset..
944   if (MI->getIntrinsicID() == Intrinsic::memset)
945     return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
946                                           MemSizeInBits, DL);
947 
948   // If we have a memcpy/memmove, the only case we can handle is if this is a
949   // copy from constant memory.  In that case, we can read directly from the
950   // constant memory.
951   MemTransferInst *MTI = cast<MemTransferInst>(MI);
952 
953   Constant *Src = dyn_cast<Constant>(MTI->getSource());
954   if (!Src) return -1;
955 
956   GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, DL));
957   if (!GV || !GV->isConstant()) return -1;
958 
959   // See if the access is within the bounds of the transfer.
960   int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
961                                               MI->getDest(), MemSizeInBits, DL);
962   if (Offset == -1)
963     return Offset;
964 
965   unsigned AS = Src->getType()->getPointerAddressSpace();
966   // Otherwise, see if we can constant fold a load from the constant with the
967   // offset applied as appropriate.
968   Src = ConstantExpr::getBitCast(Src,
969                                  Type::getInt8PtrTy(Src->getContext(), AS));
970   Constant *OffsetCst =
971     ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
972   Src = ConstantExpr::getGetElementPtr(Type::getInt8Ty(Src->getContext()), Src,
973                                        OffsetCst);
974   Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
975   if (ConstantFoldLoadFromConstPtr(Src, LoadTy, DL))
976     return Offset;
977   return -1;
978 }
979 
980 
981 /// This function is called when we have a
982 /// memdep query of a load that ends up being a clobbering store.  This means
983 /// that the store provides bits used by the load but we the pointers don't
984 /// mustalias.  Check this case to see if there is anything more we can do
985 /// before we give up.
GetStoreValueForLoad(Value * SrcVal,unsigned Offset,Type * LoadTy,Instruction * InsertPt,const DataLayout & DL)986 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
987                                    Type *LoadTy,
988                                    Instruction *InsertPt, const DataLayout &DL){
989   LLVMContext &Ctx = SrcVal->getType()->getContext();
990 
991   uint64_t StoreSize = (DL.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
992   uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy) + 7) / 8;
993 
994   IRBuilder<> Builder(InsertPt);
995 
996   // Compute which bits of the stored value are being used by the load.  Convert
997   // to an integer type to start with.
998   if (SrcVal->getType()->getScalarType()->isPointerTy())
999     SrcVal = Builder.CreatePtrToInt(SrcVal,
1000         DL.getIntPtrType(SrcVal->getType()));
1001   if (!SrcVal->getType()->isIntegerTy())
1002     SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1003 
1004   // Shift the bits to the least significant depending on endianness.
1005   unsigned ShiftAmt;
1006   if (DL.isLittleEndian())
1007     ShiftAmt = Offset*8;
1008   else
1009     ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1010 
1011   if (ShiftAmt)
1012     SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1013 
1014   if (LoadSize != StoreSize)
1015     SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1016 
1017   return CoerceAvailableValueToLoadType(SrcVal, LoadTy, Builder, DL);
1018 }
1019 
1020 /// This function is called when we have a
1021 /// memdep query of a load that ends up being a clobbering load.  This means
1022 /// that the load *may* provide bits used by the load but we can't be sure
1023 /// because the pointers don't mustalias.  Check this case to see if there is
1024 /// anything more we can do before we give up.
GetLoadValueForLoad(LoadInst * SrcVal,unsigned Offset,Type * LoadTy,Instruction * InsertPt,GVN & gvn)1025 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1026                                   Type *LoadTy, Instruction *InsertPt,
1027                                   GVN &gvn) {
1028   const DataLayout &DL = SrcVal->getModule()->getDataLayout();
1029   // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1030   // widen SrcVal out to a larger load.
1031   unsigned SrcValStoreSize = DL.getTypeStoreSize(SrcVal->getType());
1032   unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1033   if (Offset+LoadSize > SrcValStoreSize) {
1034     assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1035     assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1036     // If we have a load/load clobber an DepLI can be widened to cover this
1037     // load, then we should widen it to the next power of 2 size big enough!
1038     unsigned NewLoadSize = Offset+LoadSize;
1039     if (!isPowerOf2_32(NewLoadSize))
1040       NewLoadSize = NextPowerOf2(NewLoadSize);
1041 
1042     Value *PtrVal = SrcVal->getPointerOperand();
1043 
1044     // Insert the new load after the old load.  This ensures that subsequent
1045     // memdep queries will find the new load.  We can't easily remove the old
1046     // load completely because it is already in the value numbering table.
1047     IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1048     Type *DestPTy =
1049       IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1050     DestPTy = PointerType::get(DestPTy,
1051                                PtrVal->getType()->getPointerAddressSpace());
1052     Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1053     PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1054     LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1055     NewLoad->takeName(SrcVal);
1056     NewLoad->setAlignment(SrcVal->getAlignment());
1057 
1058     DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1059     DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1060 
1061     // Replace uses of the original load with the wider load.  On a big endian
1062     // system, we need to shift down to get the relevant bits.
1063     Value *RV = NewLoad;
1064     if (DL.isBigEndian())
1065       RV = Builder.CreateLShr(RV, (NewLoadSize - SrcValStoreSize) * 8);
1066     RV = Builder.CreateTrunc(RV, SrcVal->getType());
1067     SrcVal->replaceAllUsesWith(RV);
1068 
1069     // We would like to use gvn.markInstructionForDeletion here, but we can't
1070     // because the load is already memoized into the leader map table that GVN
1071     // tracks.  It is potentially possible to remove the load from the table,
1072     // but then there all of the operations based on it would need to be
1073     // rehashed.  Just leave the dead load around.
1074     gvn.getMemDep().removeInstruction(SrcVal);
1075     SrcVal = NewLoad;
1076   }
1077 
1078   return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, DL);
1079 }
1080 
1081 
1082 /// This function is called when we have a
1083 /// memdep query of a load that ends up being a clobbering mem intrinsic.
GetMemInstValueForLoad(MemIntrinsic * SrcInst,unsigned Offset,Type * LoadTy,Instruction * InsertPt,const DataLayout & DL)1084 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1085                                      Type *LoadTy, Instruction *InsertPt,
1086                                      const DataLayout &DL){
1087   LLVMContext &Ctx = LoadTy->getContext();
1088   uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy)/8;
1089 
1090   IRBuilder<> Builder(InsertPt);
1091 
1092   // We know that this method is only called when the mem transfer fully
1093   // provides the bits for the load.
1094   if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1095     // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1096     // independently of what the offset is.
1097     Value *Val = MSI->getValue();
1098     if (LoadSize != 1)
1099       Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1100 
1101     Value *OneElt = Val;
1102 
1103     // Splat the value out to the right number of bits.
1104     for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1105       // If we can double the number of bytes set, do it.
1106       if (NumBytesSet*2 <= LoadSize) {
1107         Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1108         Val = Builder.CreateOr(Val, ShVal);
1109         NumBytesSet <<= 1;
1110         continue;
1111       }
1112 
1113       // Otherwise insert one byte at a time.
1114       Value *ShVal = Builder.CreateShl(Val, 1*8);
1115       Val = Builder.CreateOr(OneElt, ShVal);
1116       ++NumBytesSet;
1117     }
1118 
1119     return CoerceAvailableValueToLoadType(Val, LoadTy, Builder, DL);
1120   }
1121 
1122   // Otherwise, this is a memcpy/memmove from a constant global.
1123   MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1124   Constant *Src = cast<Constant>(MTI->getSource());
1125   unsigned AS = Src->getType()->getPointerAddressSpace();
1126 
1127   // Otherwise, see if we can constant fold a load from the constant with the
1128   // offset applied as appropriate.
1129   Src = ConstantExpr::getBitCast(Src,
1130                                  Type::getInt8PtrTy(Src->getContext(), AS));
1131   Constant *OffsetCst =
1132     ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1133   Src = ConstantExpr::getGetElementPtr(Type::getInt8Ty(Src->getContext()), Src,
1134                                        OffsetCst);
1135   Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1136   return ConstantFoldLoadFromConstPtr(Src, LoadTy, DL);
1137 }
1138 
1139 
1140 /// Given a set of loads specified by ValuesPerBlock,
1141 /// construct SSA form, allowing us to eliminate LI.  This returns the value
1142 /// that should be used at LI's definition site.
ConstructSSAForLoadSet(LoadInst * LI,SmallVectorImpl<AvailableValueInBlock> & ValuesPerBlock,GVN & gvn)1143 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1144                          SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1145                                      GVN &gvn) {
1146   // Check for the fully redundant, dominating load case.  In this case, we can
1147   // just use the dominating value directly.
1148   if (ValuesPerBlock.size() == 1 &&
1149       gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1150                                                LI->getParent())) {
1151     assert(!ValuesPerBlock[0].AV.isUndefValue() &&
1152            "Dead BB dominate this block");
1153     return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
1154   }
1155 
1156   // Otherwise, we have to construct SSA form.
1157   SmallVector<PHINode*, 8> NewPHIs;
1158   SSAUpdater SSAUpdate(&NewPHIs);
1159   SSAUpdate.Initialize(LI->getType(), LI->getName());
1160 
1161   for (const AvailableValueInBlock &AV : ValuesPerBlock) {
1162     BasicBlock *BB = AV.BB;
1163 
1164     if (SSAUpdate.HasValueForBlock(BB))
1165       continue;
1166 
1167     SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
1168   }
1169 
1170   // Perform PHI construction.
1171   return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1172 }
1173 
MaterializeAdjustedValue(LoadInst * LI,Instruction * InsertPt,GVN & gvn) const1174 Value *AvailableValue::MaterializeAdjustedValue(LoadInst *LI,
1175                                                 Instruction *InsertPt,
1176                                                 GVN &gvn) const {
1177   Value *Res;
1178   Type *LoadTy = LI->getType();
1179   const DataLayout &DL = LI->getModule()->getDataLayout();
1180   if (isSimpleValue()) {
1181     Res = getSimpleValue();
1182     if (Res->getType() != LoadTy) {
1183       Res = GetStoreValueForLoad(Res, Offset, LoadTy, InsertPt, DL);
1184 
1185       DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << "  "
1186                    << *getSimpleValue() << '\n'
1187                    << *Res << '\n' << "\n\n\n");
1188     }
1189   } else if (isCoercedLoadValue()) {
1190     LoadInst *Load = getCoercedLoadValue();
1191     if (Load->getType() == LoadTy && Offset == 0) {
1192       Res = Load;
1193     } else {
1194       Res = GetLoadValueForLoad(Load, Offset, LoadTy, InsertPt, gvn);
1195 
1196       DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << "  "
1197                    << *getCoercedLoadValue() << '\n'
1198                    << *Res << '\n' << "\n\n\n");
1199     }
1200   } else if (isMemIntrinValue()) {
1201     Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
1202                                  InsertPt, DL);
1203     DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1204                  << "  " << *getMemIntrinValue() << '\n'
1205                  << *Res << '\n' << "\n\n\n");
1206   } else {
1207     assert(isUndefValue() && "Should be UndefVal");
1208     DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
1209     return UndefValue::get(LoadTy);
1210   }
1211   assert(Res && "failed to materialize?");
1212   return Res;
1213 }
1214 
isLifetimeStart(const Instruction * Inst)1215 static bool isLifetimeStart(const Instruction *Inst) {
1216   if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1217     return II->getIntrinsicID() == Intrinsic::lifetime_start;
1218   return false;
1219 }
1220 
AnalyzeLoadAvailability(LoadInst * LI,MemDepResult DepInfo,Value * Address,AvailableValue & Res)1221 bool GVN::AnalyzeLoadAvailability(LoadInst *LI, MemDepResult DepInfo,
1222                                   Value *Address, AvailableValue &Res) {
1223 
1224   assert((DepInfo.isDef() || DepInfo.isClobber()) &&
1225          "expected a local dependence");
1226   assert(LI->isUnordered() && "rules below are incorrect for ordered access");
1227 
1228   const DataLayout &DL = LI->getModule()->getDataLayout();
1229 
1230   if (DepInfo.isClobber()) {
1231     // If the dependence is to a store that writes to a superset of the bits
1232     // read by the load, we can extract the bits we need for the load from the
1233     // stored value.
1234     if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1235       // Can't forward from non-atomic to atomic without violating memory model.
1236       if (Address && LI->isAtomic() <= DepSI->isAtomic()) {
1237         int Offset =
1238           AnalyzeLoadFromClobberingStore(LI->getType(), Address, DepSI);
1239         if (Offset != -1) {
1240           Res = AvailableValue::get(DepSI->getValueOperand(), Offset);
1241           return true;
1242         }
1243       }
1244     }
1245 
1246     // Check to see if we have something like this:
1247     //    load i32* P
1248     //    load i8* (P+1)
1249     // if we have this, replace the later with an extraction from the former.
1250     if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1251       // If this is a clobber and L is the first instruction in its block, then
1252       // we have the first instruction in the entry block.
1253       // Can't forward from non-atomic to atomic without violating memory model.
1254       if (DepLI != LI && Address && LI->isAtomic() <= DepLI->isAtomic()) {
1255         int Offset =
1256           AnalyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
1257 
1258         if (Offset != -1) {
1259           Res = AvailableValue::getLoad(DepLI, Offset);
1260           return true;
1261         }
1262       }
1263     }
1264 
1265     // If the clobbering value is a memset/memcpy/memmove, see if we can
1266     // forward a value on from it.
1267     if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1268       if (Address && !LI->isAtomic()) {
1269         int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1270                                                       DepMI, DL);
1271         if (Offset != -1) {
1272           Res = AvailableValue::getMI(DepMI, Offset);
1273           return true;
1274         }
1275       }
1276     }
1277     // Nothing known about this clobber, have to be conservative
1278     DEBUG(
1279       // fast print dep, using operator<< on instruction is too slow.
1280       dbgs() << "GVN: load ";
1281       LI->printAsOperand(dbgs());
1282       Instruction *I = DepInfo.getInst();
1283       dbgs() << " is clobbered by " << *I << '\n';
1284     );
1285     return false;
1286   }
1287   assert(DepInfo.isDef() && "follows from above");
1288 
1289   Instruction *DepInst = DepInfo.getInst();
1290 
1291   // Loading the allocation -> undef.
1292   if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
1293       // Loading immediately after lifetime begin -> undef.
1294       isLifetimeStart(DepInst)) {
1295     Res = AvailableValue::get(UndefValue::get(LI->getType()));
1296     return true;
1297   }
1298 
1299   // Loading from calloc (which zero initializes memory) -> zero
1300   if (isCallocLikeFn(DepInst, TLI)) {
1301     Res = AvailableValue::get(Constant::getNullValue(LI->getType()));
1302     return true;
1303   }
1304 
1305   if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1306     // Reject loads and stores that are to the same address but are of
1307     // different types if we have to. If the stored value is larger or equal to
1308     // the loaded value, we can reuse it.
1309     if (S->getValueOperand()->getType() != LI->getType() &&
1310         !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1311                                          LI->getType(), DL))
1312       return false;
1313 
1314     // Can't forward from non-atomic to atomic without violating memory model.
1315     if (S->isAtomic() < LI->isAtomic())
1316       return false;
1317 
1318     Res = AvailableValue::get(S->getValueOperand());
1319     return true;
1320   }
1321 
1322   if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1323     // If the types mismatch and we can't handle it, reject reuse of the load.
1324     // If the stored value is larger or equal to the loaded value, we can reuse
1325     // it.
1326     if (LD->getType() != LI->getType() &&
1327         !CanCoerceMustAliasedValueToLoad(LD, LI->getType(), DL))
1328       return false;
1329 
1330     // Can't forward from non-atomic to atomic without violating memory model.
1331     if (LD->isAtomic() < LI->isAtomic())
1332       return false;
1333 
1334     Res = AvailableValue::getLoad(LD);
1335     return true;
1336   }
1337 
1338   // Unknown def - must be conservative
1339   DEBUG(
1340     // fast print dep, using operator<< on instruction is too slow.
1341     dbgs() << "GVN: load ";
1342     LI->printAsOperand(dbgs());
1343     dbgs() << " has unknown def " << *DepInst << '\n';
1344   );
1345   return false;
1346 }
1347 
AnalyzeLoadAvailability(LoadInst * LI,LoadDepVect & Deps,AvailValInBlkVect & ValuesPerBlock,UnavailBlkVect & UnavailableBlocks)1348 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1349                                   AvailValInBlkVect &ValuesPerBlock,
1350                                   UnavailBlkVect &UnavailableBlocks) {
1351 
1352   // Filter out useless results (non-locals, etc).  Keep track of the blocks
1353   // where we have a value available in repl, also keep track of whether we see
1354   // dependencies that produce an unknown value for the load (such as a call
1355   // that could potentially clobber the load).
1356   unsigned NumDeps = Deps.size();
1357   for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1358     BasicBlock *DepBB = Deps[i].getBB();
1359     MemDepResult DepInfo = Deps[i].getResult();
1360 
1361     if (DeadBlocks.count(DepBB)) {
1362       // Dead dependent mem-op disguise as a load evaluating the same value
1363       // as the load in question.
1364       ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1365       continue;
1366     }
1367 
1368     if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1369       UnavailableBlocks.push_back(DepBB);
1370       continue;
1371     }
1372 
1373     // The address being loaded in this non-local block may not be the same as
1374     // the pointer operand of the load if PHI translation occurs.  Make sure
1375     // to consider the right address.
1376     Value *Address = Deps[i].getAddress();
1377 
1378     AvailableValue AV;
1379     if (AnalyzeLoadAvailability(LI, DepInfo, Address, AV)) {
1380       // subtlety: because we know this was a non-local dependency, we know
1381       // it's safe to materialize anywhere between the instruction within
1382       // DepInfo and the end of it's block.
1383       ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1384                                                           std::move(AV)));
1385     } else {
1386       UnavailableBlocks.push_back(DepBB);
1387     }
1388   }
1389 
1390   assert(NumDeps == ValuesPerBlock.size() + UnavailableBlocks.size() &&
1391          "post condition violation");
1392 }
1393 
PerformLoadPRE(LoadInst * LI,AvailValInBlkVect & ValuesPerBlock,UnavailBlkVect & UnavailableBlocks)1394 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1395                          UnavailBlkVect &UnavailableBlocks) {
1396   // Okay, we have *some* definitions of the value.  This means that the value
1397   // is available in some of our (transitive) predecessors.  Lets think about
1398   // doing PRE of this load.  This will involve inserting a new load into the
1399   // predecessor when it's not available.  We could do this in general, but
1400   // prefer to not increase code size.  As such, we only do this when we know
1401   // that we only have to insert *one* load (which means we're basically moving
1402   // the load, not inserting a new one).
1403 
1404   SmallPtrSet<BasicBlock *, 4> Blockers(UnavailableBlocks.begin(),
1405                                         UnavailableBlocks.end());
1406 
1407   // Let's find the first basic block with more than one predecessor.  Walk
1408   // backwards through predecessors if needed.
1409   BasicBlock *LoadBB = LI->getParent();
1410   BasicBlock *TmpBB = LoadBB;
1411 
1412   while (TmpBB->getSinglePredecessor()) {
1413     TmpBB = TmpBB->getSinglePredecessor();
1414     if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1415       return false;
1416     if (Blockers.count(TmpBB))
1417       return false;
1418 
1419     // If any of these blocks has more than one successor (i.e. if the edge we
1420     // just traversed was critical), then there are other paths through this
1421     // block along which the load may not be anticipated.  Hoisting the load
1422     // above this block would be adding the load to execution paths along
1423     // which it was not previously executed.
1424     if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1425       return false;
1426   }
1427 
1428   assert(TmpBB);
1429   LoadBB = TmpBB;
1430 
1431   // Check to see how many predecessors have the loaded value fully
1432   // available.
1433   MapVector<BasicBlock *, Value *> PredLoads;
1434   DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1435   for (const AvailableValueInBlock &AV : ValuesPerBlock)
1436     FullyAvailableBlocks[AV.BB] = true;
1437   for (BasicBlock *UnavailableBB : UnavailableBlocks)
1438     FullyAvailableBlocks[UnavailableBB] = false;
1439 
1440   SmallVector<BasicBlock *, 4> CriticalEdgePred;
1441   for (BasicBlock *Pred : predecessors(LoadBB)) {
1442     // If any predecessor block is an EH pad that does not allow non-PHI
1443     // instructions before the terminator, we can't PRE the load.
1444     if (Pred->getTerminator()->isEHPad()) {
1445       DEBUG(dbgs()
1446             << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD PREDECESSOR '"
1447             << Pred->getName() << "': " << *LI << '\n');
1448       return false;
1449     }
1450 
1451     if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1452       continue;
1453     }
1454 
1455     if (Pred->getTerminator()->getNumSuccessors() != 1) {
1456       if (isa<IndirectBrInst>(Pred->getTerminator())) {
1457         DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1458               << Pred->getName() << "': " << *LI << '\n');
1459         return false;
1460       }
1461 
1462       if (LoadBB->isEHPad()) {
1463         DEBUG(dbgs()
1464               << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1465               << Pred->getName() << "': " << *LI << '\n');
1466         return false;
1467       }
1468 
1469       CriticalEdgePred.push_back(Pred);
1470     } else {
1471       // Only add the predecessors that will not be split for now.
1472       PredLoads[Pred] = nullptr;
1473     }
1474   }
1475 
1476   // Decide whether PRE is profitable for this load.
1477   unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1478   assert(NumUnavailablePreds != 0 &&
1479          "Fully available value should already be eliminated!");
1480 
1481   // If this load is unavailable in multiple predecessors, reject it.
1482   // FIXME: If we could restructure the CFG, we could make a common pred with
1483   // all the preds that don't have an available LI and insert a new load into
1484   // that one block.
1485   if (NumUnavailablePreds != 1)
1486       return false;
1487 
1488   // Split critical edges, and update the unavailable predecessors accordingly.
1489   for (BasicBlock *OrigPred : CriticalEdgePred) {
1490     BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1491     assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1492     PredLoads[NewPred] = nullptr;
1493     DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1494                  << LoadBB->getName() << '\n');
1495   }
1496 
1497   // Check if the load can safely be moved to all the unavailable predecessors.
1498   bool CanDoPRE = true;
1499   const DataLayout &DL = LI->getModule()->getDataLayout();
1500   SmallVector<Instruction*, 8> NewInsts;
1501   for (auto &PredLoad : PredLoads) {
1502     BasicBlock *UnavailablePred = PredLoad.first;
1503 
1504     // Do PHI translation to get its value in the predecessor if necessary.  The
1505     // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1506 
1507     // If all preds have a single successor, then we know it is safe to insert
1508     // the load on the pred (?!?), so we can insert code to materialize the
1509     // pointer if it is not available.
1510     PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1511     Value *LoadPtr = nullptr;
1512     LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1513                                                 *DT, NewInsts);
1514 
1515     // If we couldn't find or insert a computation of this phi translated value,
1516     // we fail PRE.
1517     if (!LoadPtr) {
1518       DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1519             << *LI->getPointerOperand() << "\n");
1520       CanDoPRE = false;
1521       break;
1522     }
1523 
1524     PredLoad.second = LoadPtr;
1525   }
1526 
1527   if (!CanDoPRE) {
1528     while (!NewInsts.empty()) {
1529       Instruction *I = NewInsts.pop_back_val();
1530       if (MD) MD->removeInstruction(I);
1531       I->eraseFromParent();
1532     }
1533     // HINT: Don't revert the edge-splitting as following transformation may
1534     // also need to split these critical edges.
1535     return !CriticalEdgePred.empty();
1536   }
1537 
1538   // Okay, we can eliminate this load by inserting a reload in the predecessor
1539   // and using PHI construction to get the value in the other predecessors, do
1540   // it.
1541   DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1542   DEBUG(if (!NewInsts.empty())
1543           dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1544                  << *NewInsts.back() << '\n');
1545 
1546   // Assign value numbers to the new instructions.
1547   for (Instruction *I : NewInsts) {
1548     // FIXME: We really _ought_ to insert these value numbers into their
1549     // parent's availability map.  However, in doing so, we risk getting into
1550     // ordering issues.  If a block hasn't been processed yet, we would be
1551     // marking a value as AVAIL-IN, which isn't what we intend.
1552     VN.lookupOrAdd(I);
1553   }
1554 
1555   for (const auto &PredLoad : PredLoads) {
1556     BasicBlock *UnavailablePred = PredLoad.first;
1557     Value *LoadPtr = PredLoad.second;
1558 
1559     auto *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre",
1560                                  LI->isVolatile(), LI->getAlignment(),
1561                                  LI->getOrdering(), LI->getSynchScope(),
1562                                  UnavailablePred->getTerminator());
1563 
1564     // Transfer the old load's AA tags to the new load.
1565     AAMDNodes Tags;
1566     LI->getAAMetadata(Tags);
1567     if (Tags)
1568       NewLoad->setAAMetadata(Tags);
1569 
1570     if (auto *MD = LI->getMetadata(LLVMContext::MD_invariant_load))
1571       NewLoad->setMetadata(LLVMContext::MD_invariant_load, MD);
1572     if (auto *InvGroupMD = LI->getMetadata(LLVMContext::MD_invariant_group))
1573       NewLoad->setMetadata(LLVMContext::MD_invariant_group, InvGroupMD);
1574     if (auto *RangeMD = LI->getMetadata(LLVMContext::MD_range))
1575       NewLoad->setMetadata(LLVMContext::MD_range, RangeMD);
1576 
1577     // Transfer DebugLoc.
1578     NewLoad->setDebugLoc(LI->getDebugLoc());
1579 
1580     // Add the newly created load.
1581     ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1582                                                         NewLoad));
1583     MD->invalidateCachedPointerInfo(LoadPtr);
1584     DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1585   }
1586 
1587   // Perform PHI construction.
1588   Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1589   LI->replaceAllUsesWith(V);
1590   if (isa<PHINode>(V))
1591     V->takeName(LI);
1592   if (Instruction *I = dyn_cast<Instruction>(V))
1593     I->setDebugLoc(LI->getDebugLoc());
1594   if (V->getType()->getScalarType()->isPointerTy())
1595     MD->invalidateCachedPointerInfo(V);
1596   markInstructionForDeletion(LI);
1597   ++NumPRELoad;
1598   return true;
1599 }
1600 
1601 /// Attempt to eliminate a load whose dependencies are
1602 /// non-local by performing PHI construction.
processNonLocalLoad(LoadInst * LI)1603 bool GVN::processNonLocalLoad(LoadInst *LI) {
1604   // non-local speculations are not allowed under asan.
1605   if (LI->getParent()->getParent()->hasFnAttribute(Attribute::SanitizeAddress))
1606     return false;
1607 
1608   // Step 1: Find the non-local dependencies of the load.
1609   LoadDepVect Deps;
1610   MD->getNonLocalPointerDependency(LI, Deps);
1611 
1612   // If we had to process more than one hundred blocks to find the
1613   // dependencies, this load isn't worth worrying about.  Optimizing
1614   // it will be too expensive.
1615   unsigned NumDeps = Deps.size();
1616   if (NumDeps > 100)
1617     return false;
1618 
1619   // If we had a phi translation failure, we'll have a single entry which is a
1620   // clobber in the current block.  Reject this early.
1621   if (NumDeps == 1 &&
1622       !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1623     DEBUG(
1624       dbgs() << "GVN: non-local load ";
1625       LI->printAsOperand(dbgs());
1626       dbgs() << " has unknown dependencies\n";
1627     );
1628     return false;
1629   }
1630 
1631   // If this load follows a GEP, see if we can PRE the indices before analyzing.
1632   if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1633     for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1634                                         OE = GEP->idx_end();
1635          OI != OE; ++OI)
1636       if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1637         performScalarPRE(I);
1638   }
1639 
1640   // Step 2: Analyze the availability of the load
1641   AvailValInBlkVect ValuesPerBlock;
1642   UnavailBlkVect UnavailableBlocks;
1643   AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1644 
1645   // If we have no predecessors that produce a known value for this load, exit
1646   // early.
1647   if (ValuesPerBlock.empty())
1648     return false;
1649 
1650   // Step 3: Eliminate fully redundancy.
1651   //
1652   // If all of the instructions we depend on produce a known value for this
1653   // load, then it is fully redundant and we can use PHI insertion to compute
1654   // its value.  Insert PHIs and remove the fully redundant value now.
1655   if (UnavailableBlocks.empty()) {
1656     DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1657 
1658     // Perform PHI construction.
1659     Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1660     LI->replaceAllUsesWith(V);
1661 
1662     if (isa<PHINode>(V))
1663       V->takeName(LI);
1664     if (Instruction *I = dyn_cast<Instruction>(V))
1665       if (LI->getDebugLoc())
1666         I->setDebugLoc(LI->getDebugLoc());
1667     if (V->getType()->getScalarType()->isPointerTy())
1668       MD->invalidateCachedPointerInfo(V);
1669     markInstructionForDeletion(LI);
1670     ++NumGVNLoad;
1671     return true;
1672   }
1673 
1674   // Step 4: Eliminate partial redundancy.
1675   if (!EnablePRE || !EnableLoadPRE)
1676     return false;
1677 
1678   return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1679 }
1680 
processAssumeIntrinsic(IntrinsicInst * IntrinsicI)1681 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1682   assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1683          "This function can only be called with llvm.assume intrinsic");
1684   Value *V = IntrinsicI->getArgOperand(0);
1685 
1686   if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1687     if (Cond->isZero()) {
1688       Type *Int8Ty = Type::getInt8Ty(V->getContext());
1689       // Insert a new store to null instruction before the load to indicate that
1690       // this code is not reachable.  FIXME: We could insert unreachable
1691       // instruction directly because we can modify the CFG.
1692       new StoreInst(UndefValue::get(Int8Ty),
1693                     Constant::getNullValue(Int8Ty->getPointerTo()),
1694                     IntrinsicI);
1695     }
1696     markInstructionForDeletion(IntrinsicI);
1697     return false;
1698   }
1699 
1700   Constant *True = ConstantInt::getTrue(V->getContext());
1701   bool Changed = false;
1702 
1703   for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1704     BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1705 
1706     // This property is only true in dominated successors, propagateEquality
1707     // will check dominance for us.
1708     Changed |= propagateEquality(V, True, Edge, false);
1709   }
1710 
1711   // We can replace assume value with true, which covers cases like this:
1712   // call void @llvm.assume(i1 %cmp)
1713   // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1714   ReplaceWithConstMap[V] = True;
1715 
1716   // If one of *cmp *eq operand is const, adding it to map will cover this:
1717   // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1718   // call void @llvm.assume(i1 %cmp)
1719   // ret float %0 ; will change it to ret float 3.000000e+00
1720   if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1721     if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1722         CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1723         (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1724          CmpI->getFastMathFlags().noNaNs())) {
1725       Value *CmpLHS = CmpI->getOperand(0);
1726       Value *CmpRHS = CmpI->getOperand(1);
1727       if (isa<Constant>(CmpLHS))
1728         std::swap(CmpLHS, CmpRHS);
1729       auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1730 
1731       // If only one operand is constant.
1732       if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1733         ReplaceWithConstMap[CmpLHS] = RHSConst;
1734     }
1735   }
1736   return Changed;
1737 }
1738 
patchReplacementInstruction(Instruction * I,Value * Repl)1739 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1740   auto *ReplInst = dyn_cast<Instruction>(Repl);
1741   if (!ReplInst)
1742     return;
1743 
1744   // Patch the replacement so that it is not more restrictive than the value
1745   // being replaced.
1746   ReplInst->andIRFlags(I);
1747 
1748   // FIXME: If both the original and replacement value are part of the
1749   // same control-flow region (meaning that the execution of one
1750   // guarantees the execution of the other), then we can combine the
1751   // noalias scopes here and do better than the general conservative
1752   // answer used in combineMetadata().
1753 
1754   // In general, GVN unifies expressions over different control-flow
1755   // regions, and so we need a conservative combination of the noalias
1756   // scopes.
1757   static const unsigned KnownIDs[] = {
1758       LLVMContext::MD_tbaa,           LLVMContext::MD_alias_scope,
1759       LLVMContext::MD_noalias,        LLVMContext::MD_range,
1760       LLVMContext::MD_fpmath,         LLVMContext::MD_invariant_load,
1761       LLVMContext::MD_invariant_group};
1762   combineMetadata(ReplInst, I, KnownIDs);
1763 }
1764 
patchAndReplaceAllUsesWith(Instruction * I,Value * Repl)1765 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1766   patchReplacementInstruction(I, Repl);
1767   I->replaceAllUsesWith(Repl);
1768 }
1769 
1770 /// Attempt to eliminate a load, first by eliminating it
1771 /// locally, and then attempting non-local elimination if that fails.
processLoad(LoadInst * L)1772 bool GVN::processLoad(LoadInst *L) {
1773   if (!MD)
1774     return false;
1775 
1776   // This code hasn't been audited for ordered or volatile memory access
1777   if (!L->isUnordered())
1778     return false;
1779 
1780   if (L->use_empty()) {
1781     markInstructionForDeletion(L);
1782     return true;
1783   }
1784 
1785   // ... to a pointer that has been loaded from before...
1786   MemDepResult Dep = MD->getDependency(L);
1787 
1788   // If it is defined in another block, try harder.
1789   if (Dep.isNonLocal())
1790     return processNonLocalLoad(L);
1791 
1792   // Only handle the local case below
1793   if (!Dep.isDef() && !Dep.isClobber()) {
1794     // This might be a NonFuncLocal or an Unknown
1795     DEBUG(
1796       // fast print dep, using operator<< on instruction is too slow.
1797       dbgs() << "GVN: load ";
1798       L->printAsOperand(dbgs());
1799       dbgs() << " has unknown dependence\n";
1800     );
1801     return false;
1802   }
1803 
1804   AvailableValue AV;
1805   if (AnalyzeLoadAvailability(L, Dep, L->getPointerOperand(), AV)) {
1806     Value *AvailableValue = AV.MaterializeAdjustedValue(L, L, *this);
1807 
1808     // Replace the load!
1809     patchAndReplaceAllUsesWith(L, AvailableValue);
1810     markInstructionForDeletion(L);
1811     ++NumGVNLoad;
1812     // Tell MDA to rexamine the reused pointer since we might have more
1813     // information after forwarding it.
1814     if (MD && AvailableValue->getType()->getScalarType()->isPointerTy())
1815       MD->invalidateCachedPointerInfo(AvailableValue);
1816     return true;
1817   }
1818 
1819   return false;
1820 }
1821 
1822 // In order to find a leader for a given value number at a
1823 // specific basic block, we first obtain the list of all Values for that number,
1824 // and then scan the list to find one whose block dominates the block in
1825 // question.  This is fast because dominator tree queries consist of only
1826 // a few comparisons of DFS numbers.
findLeader(const BasicBlock * BB,uint32_t num)1827 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1828   LeaderTableEntry Vals = LeaderTable[num];
1829   if (!Vals.Val) return nullptr;
1830 
1831   Value *Val = nullptr;
1832   if (DT->dominates(Vals.BB, BB)) {
1833     Val = Vals.Val;
1834     if (isa<Constant>(Val)) return Val;
1835   }
1836 
1837   LeaderTableEntry* Next = Vals.Next;
1838   while (Next) {
1839     if (DT->dominates(Next->BB, BB)) {
1840       if (isa<Constant>(Next->Val)) return Next->Val;
1841       if (!Val) Val = Next->Val;
1842     }
1843 
1844     Next = Next->Next;
1845   }
1846 
1847   return Val;
1848 }
1849 
1850 /// There is an edge from 'Src' to 'Dst'.  Return
1851 /// true if every path from the entry block to 'Dst' passes via this edge.  In
1852 /// particular 'Dst' must not be reachable via another edge from 'Src'.
isOnlyReachableViaThisEdge(const BasicBlockEdge & E,DominatorTree * DT)1853 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
1854                                        DominatorTree *DT) {
1855   // While in theory it is interesting to consider the case in which Dst has
1856   // more than one predecessor, because Dst might be part of a loop which is
1857   // only reachable from Src, in practice it is pointless since at the time
1858   // GVN runs all such loops have preheaders, which means that Dst will have
1859   // been changed to have only one predecessor, namely Src.
1860   const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
1861   assert((!Pred || Pred == E.getStart()) &&
1862          "No edge between these basic blocks!");
1863   return Pred != nullptr;
1864 }
1865 
1866 // Tries to replace instruction with const, using information from
1867 // ReplaceWithConstMap.
replaceOperandsWithConsts(Instruction * Instr) const1868 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
1869   bool Changed = false;
1870   for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
1871     Value *Operand = Instr->getOperand(OpNum);
1872     auto it = ReplaceWithConstMap.find(Operand);
1873     if (it != ReplaceWithConstMap.end()) {
1874       assert(!isa<Constant>(Operand) &&
1875              "Replacing constants with constants is invalid");
1876       DEBUG(dbgs() << "GVN replacing: " << *Operand << " with " << *it->second
1877                    << " in instruction " << *Instr << '\n');
1878       Instr->setOperand(OpNum, it->second);
1879       Changed = true;
1880     }
1881   }
1882   return Changed;
1883 }
1884 
1885 /// The given values are known to be equal in every block
1886 /// dominated by 'Root'.  Exploit this, for example by replacing 'LHS' with
1887 /// 'RHS' everywhere in the scope.  Returns whether a change was made.
1888 /// If DominatesByEdge is false, then it means that we will propagate the RHS
1889 /// value starting from the end of Root.Start.
propagateEquality(Value * LHS,Value * RHS,const BasicBlockEdge & Root,bool DominatesByEdge)1890 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
1891                             bool DominatesByEdge) {
1892   SmallVector<std::pair<Value*, Value*>, 4> Worklist;
1893   Worklist.push_back(std::make_pair(LHS, RHS));
1894   bool Changed = false;
1895   // For speed, compute a conservative fast approximation to
1896   // DT->dominates(Root, Root.getEnd());
1897   const bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
1898 
1899   while (!Worklist.empty()) {
1900     std::pair<Value*, Value*> Item = Worklist.pop_back_val();
1901     LHS = Item.first; RHS = Item.second;
1902 
1903     if (LHS == RHS)
1904       continue;
1905     assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
1906 
1907     // Don't try to propagate equalities between constants.
1908     if (isa<Constant>(LHS) && isa<Constant>(RHS))
1909       continue;
1910 
1911     // Prefer a constant on the right-hand side, or an Argument if no constants.
1912     if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
1913       std::swap(LHS, RHS);
1914     assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
1915 
1916     // If there is no obvious reason to prefer the left-hand side over the
1917     // right-hand side, ensure the longest lived term is on the right-hand side,
1918     // so the shortest lived term will be replaced by the longest lived.
1919     // This tends to expose more simplifications.
1920     uint32_t LVN = VN.lookupOrAdd(LHS);
1921     if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
1922         (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
1923       // Move the 'oldest' value to the right-hand side, using the value number
1924       // as a proxy for age.
1925       uint32_t RVN = VN.lookupOrAdd(RHS);
1926       if (LVN < RVN) {
1927         std::swap(LHS, RHS);
1928         LVN = RVN;
1929       }
1930     }
1931 
1932     // If value numbering later sees that an instruction in the scope is equal
1933     // to 'LHS' then ensure it will be turned into 'RHS'.  In order to preserve
1934     // the invariant that instructions only occur in the leader table for their
1935     // own value number (this is used by removeFromLeaderTable), do not do this
1936     // if RHS is an instruction (if an instruction in the scope is morphed into
1937     // LHS then it will be turned into RHS by the next GVN iteration anyway, so
1938     // using the leader table is about compiling faster, not optimizing better).
1939     // The leader table only tracks basic blocks, not edges. Only add to if we
1940     // have the simple case where the edge dominates the end.
1941     if (RootDominatesEnd && !isa<Instruction>(RHS))
1942       addToLeaderTable(LVN, RHS, Root.getEnd());
1943 
1944     // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope.  As
1945     // LHS always has at least one use that is not dominated by Root, this will
1946     // never do anything if LHS has only one use.
1947     if (!LHS->hasOneUse()) {
1948       unsigned NumReplacements =
1949           DominatesByEdge
1950               ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
1951               : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getStart());
1952 
1953       Changed |= NumReplacements > 0;
1954       NumGVNEqProp += NumReplacements;
1955     }
1956 
1957     // Now try to deduce additional equalities from this one. For example, if
1958     // the known equality was "(A != B)" == "false" then it follows that A and B
1959     // are equal in the scope. Only boolean equalities with an explicit true or
1960     // false RHS are currently supported.
1961     if (!RHS->getType()->isIntegerTy(1))
1962       // Not a boolean equality - bail out.
1963       continue;
1964     ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
1965     if (!CI)
1966       // RHS neither 'true' nor 'false' - bail out.
1967       continue;
1968     // Whether RHS equals 'true'.  Otherwise it equals 'false'.
1969     bool isKnownTrue = CI->isAllOnesValue();
1970     bool isKnownFalse = !isKnownTrue;
1971 
1972     // If "A && B" is known true then both A and B are known true.  If "A || B"
1973     // is known false then both A and B are known false.
1974     Value *A, *B;
1975     if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
1976         (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
1977       Worklist.push_back(std::make_pair(A, RHS));
1978       Worklist.push_back(std::make_pair(B, RHS));
1979       continue;
1980     }
1981 
1982     // If we are propagating an equality like "(A == B)" == "true" then also
1983     // propagate the equality A == B.  When propagating a comparison such as
1984     // "(A >= B)" == "true", replace all instances of "A < B" with "false".
1985     if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
1986       Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1987 
1988       // If "A == B" is known true, or "A != B" is known false, then replace
1989       // A with B everywhere in the scope.
1990       if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
1991           (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
1992         Worklist.push_back(std::make_pair(Op0, Op1));
1993 
1994       // Handle the floating point versions of equality comparisons too.
1995       if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
1996           (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
1997 
1998         // Floating point -0.0 and 0.0 compare equal, so we can only
1999         // propagate values if we know that we have a constant and that
2000         // its value is non-zero.
2001 
2002         // FIXME: We should do this optimization if 'no signed zeros' is
2003         // applicable via an instruction-level fast-math-flag or some other
2004         // indicator that relaxed FP semantics are being used.
2005 
2006         if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
2007           Worklist.push_back(std::make_pair(Op0, Op1));
2008       }
2009 
2010       // If "A >= B" is known true, replace "A < B" with false everywhere.
2011       CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2012       Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2013       // Since we don't have the instruction "A < B" immediately to hand, work
2014       // out the value number that it would have and use that to find an
2015       // appropriate instruction (if any).
2016       uint32_t NextNum = VN.getNextUnusedValueNumber();
2017       uint32_t Num = VN.lookupOrAddCmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2018       // If the number we were assigned was brand new then there is no point in
2019       // looking for an instruction realizing it: there cannot be one!
2020       if (Num < NextNum) {
2021         Value *NotCmp = findLeader(Root.getEnd(), Num);
2022         if (NotCmp && isa<Instruction>(NotCmp)) {
2023           unsigned NumReplacements =
2024               DominatesByEdge
2025                   ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
2026                   : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
2027                                              Root.getStart());
2028           Changed |= NumReplacements > 0;
2029           NumGVNEqProp += NumReplacements;
2030         }
2031       }
2032       // Ensure that any instruction in scope that gets the "A < B" value number
2033       // is replaced with false.
2034       // The leader table only tracks basic blocks, not edges. Only add to if we
2035       // have the simple case where the edge dominates the end.
2036       if (RootDominatesEnd)
2037         addToLeaderTable(Num, NotVal, Root.getEnd());
2038 
2039       continue;
2040     }
2041   }
2042 
2043   return Changed;
2044 }
2045 
2046 /// When calculating availability, handle an instruction
2047 /// by inserting it into the appropriate sets
processInstruction(Instruction * I)2048 bool GVN::processInstruction(Instruction *I) {
2049   // Ignore dbg info intrinsics.
2050   if (isa<DbgInfoIntrinsic>(I))
2051     return false;
2052 
2053   // If the instruction can be easily simplified then do so now in preference
2054   // to value numbering it.  Value numbering often exposes redundancies, for
2055   // example if it determines that %y is equal to %x then the instruction
2056   // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2057   const DataLayout &DL = I->getModule()->getDataLayout();
2058   if (Value *V = SimplifyInstruction(I, DL, TLI, DT, AC)) {
2059     bool Changed = false;
2060     if (!I->use_empty()) {
2061       I->replaceAllUsesWith(V);
2062       Changed = true;
2063     }
2064     if (isInstructionTriviallyDead(I, TLI)) {
2065       markInstructionForDeletion(I);
2066       Changed = true;
2067     }
2068     if (Changed) {
2069       if (MD && V->getType()->getScalarType()->isPointerTy())
2070         MD->invalidateCachedPointerInfo(V);
2071       ++NumGVNSimpl;
2072       return true;
2073     }
2074   }
2075 
2076   if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
2077     if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
2078       return processAssumeIntrinsic(IntrinsicI);
2079 
2080   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2081     if (processLoad(LI))
2082       return true;
2083 
2084     unsigned Num = VN.lookupOrAdd(LI);
2085     addToLeaderTable(Num, LI, LI->getParent());
2086     return false;
2087   }
2088 
2089   // For conditional branches, we can perform simple conditional propagation on
2090   // the condition value itself.
2091   if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2092     if (!BI->isConditional())
2093       return false;
2094 
2095     if (isa<Constant>(BI->getCondition()))
2096       return processFoldableCondBr(BI);
2097 
2098     Value *BranchCond = BI->getCondition();
2099     BasicBlock *TrueSucc = BI->getSuccessor(0);
2100     BasicBlock *FalseSucc = BI->getSuccessor(1);
2101     // Avoid multiple edges early.
2102     if (TrueSucc == FalseSucc)
2103       return false;
2104 
2105     BasicBlock *Parent = BI->getParent();
2106     bool Changed = false;
2107 
2108     Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
2109     BasicBlockEdge TrueE(Parent, TrueSucc);
2110     Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
2111 
2112     Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
2113     BasicBlockEdge FalseE(Parent, FalseSucc);
2114     Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
2115 
2116     return Changed;
2117   }
2118 
2119   // For switches, propagate the case values into the case destinations.
2120   if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2121     Value *SwitchCond = SI->getCondition();
2122     BasicBlock *Parent = SI->getParent();
2123     bool Changed = false;
2124 
2125     // Remember how many outgoing edges there are to every successor.
2126     SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2127     for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
2128       ++SwitchEdges[SI->getSuccessor(i)];
2129 
2130     for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2131          i != e; ++i) {
2132       BasicBlock *Dst = i.getCaseSuccessor();
2133       // If there is only a single edge, propagate the case value into it.
2134       if (SwitchEdges.lookup(Dst) == 1) {
2135         BasicBlockEdge E(Parent, Dst);
2136         Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E, true);
2137       }
2138     }
2139     return Changed;
2140   }
2141 
2142   // Instructions with void type don't return a value, so there's
2143   // no point in trying to find redundancies in them.
2144   if (I->getType()->isVoidTy())
2145     return false;
2146 
2147   uint32_t NextNum = VN.getNextUnusedValueNumber();
2148   unsigned Num = VN.lookupOrAdd(I);
2149 
2150   // Allocations are always uniquely numbered, so we can save time and memory
2151   // by fast failing them.
2152   if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2153     addToLeaderTable(Num, I, I->getParent());
2154     return false;
2155   }
2156 
2157   // If the number we were assigned was a brand new VN, then we don't
2158   // need to do a lookup to see if the number already exists
2159   // somewhere in the domtree: it can't!
2160   if (Num >= NextNum) {
2161     addToLeaderTable(Num, I, I->getParent());
2162     return false;
2163   }
2164 
2165   // Perform fast-path value-number based elimination of values inherited from
2166   // dominators.
2167   Value *Repl = findLeader(I->getParent(), Num);
2168   if (!Repl) {
2169     // Failure, just remember this instance for future use.
2170     addToLeaderTable(Num, I, I->getParent());
2171     return false;
2172   } else if (Repl == I) {
2173     // If I was the result of a shortcut PRE, it might already be in the table
2174     // and the best replacement for itself. Nothing to do.
2175     return false;
2176   }
2177 
2178   // Remove it!
2179   patchAndReplaceAllUsesWith(I, Repl);
2180   if (MD && Repl->getType()->getScalarType()->isPointerTy())
2181     MD->invalidateCachedPointerInfo(Repl);
2182   markInstructionForDeletion(I);
2183   return true;
2184 }
2185 
2186 /// runOnFunction - This is the main transformation entry point for a function.
runImpl(Function & F,AssumptionCache & RunAC,DominatorTree & RunDT,const TargetLibraryInfo & RunTLI,AAResults & RunAA,MemoryDependenceResults * RunMD)2187 bool GVN::runImpl(Function &F, AssumptionCache &RunAC, DominatorTree &RunDT,
2188                   const TargetLibraryInfo &RunTLI, AAResults &RunAA,
2189                   MemoryDependenceResults *RunMD) {
2190   AC = &RunAC;
2191   DT = &RunDT;
2192   VN.setDomTree(DT);
2193   TLI = &RunTLI;
2194   VN.setAliasAnalysis(&RunAA);
2195   MD = RunMD;
2196   VN.setMemDep(MD);
2197 
2198   bool Changed = false;
2199   bool ShouldContinue = true;
2200 
2201   // Merge unconditional branches, allowing PRE to catch more
2202   // optimization opportunities.
2203   for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2204     BasicBlock *BB = &*FI++;
2205 
2206     bool removedBlock =
2207         MergeBlockIntoPredecessor(BB, DT, /* LoopInfo */ nullptr, MD);
2208     if (removedBlock) ++NumGVNBlocks;
2209 
2210     Changed |= removedBlock;
2211   }
2212 
2213   unsigned Iteration = 0;
2214   while (ShouldContinue) {
2215     DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2216     ShouldContinue = iterateOnFunction(F);
2217     Changed |= ShouldContinue;
2218     ++Iteration;
2219   }
2220 
2221   if (EnablePRE) {
2222     // Fabricate val-num for dead-code in order to suppress assertion in
2223     // performPRE().
2224     assignValNumForDeadCode();
2225     bool PREChanged = true;
2226     while (PREChanged) {
2227       PREChanged = performPRE(F);
2228       Changed |= PREChanged;
2229     }
2230   }
2231 
2232   // FIXME: Should perform GVN again after PRE does something.  PRE can move
2233   // computations into blocks where they become fully redundant.  Note that
2234   // we can't do this until PRE's critical edge splitting updates memdep.
2235   // Actually, when this happens, we should just fully integrate PRE into GVN.
2236 
2237   cleanupGlobalSets();
2238   // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2239   // iteration.
2240   DeadBlocks.clear();
2241 
2242   return Changed;
2243 }
2244 
processBlock(BasicBlock * BB)2245 bool GVN::processBlock(BasicBlock *BB) {
2246   // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2247   // (and incrementing BI before processing an instruction).
2248   assert(InstrsToErase.empty() &&
2249          "We expect InstrsToErase to be empty across iterations");
2250   if (DeadBlocks.count(BB))
2251     return false;
2252 
2253   // Clearing map before every BB because it can be used only for single BB.
2254   ReplaceWithConstMap.clear();
2255   bool ChangedFunction = false;
2256 
2257   for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2258        BI != BE;) {
2259     if (!ReplaceWithConstMap.empty())
2260       ChangedFunction |= replaceOperandsWithConsts(&*BI);
2261     ChangedFunction |= processInstruction(&*BI);
2262 
2263     if (InstrsToErase.empty()) {
2264       ++BI;
2265       continue;
2266     }
2267 
2268     // If we need some instructions deleted, do it now.
2269     NumGVNInstr += InstrsToErase.size();
2270 
2271     // Avoid iterator invalidation.
2272     bool AtStart = BI == BB->begin();
2273     if (!AtStart)
2274       --BI;
2275 
2276     for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2277          E = InstrsToErase.end(); I != E; ++I) {
2278       DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2279       if (MD) MD->removeInstruction(*I);
2280       DEBUG(verifyRemoved(*I));
2281       (*I)->eraseFromParent();
2282     }
2283     InstrsToErase.clear();
2284 
2285     if (AtStart)
2286       BI = BB->begin();
2287     else
2288       ++BI;
2289   }
2290 
2291   return ChangedFunction;
2292 }
2293 
2294 // Instantiate an expression in a predecessor that lacked it.
performScalarPREInsertion(Instruction * Instr,BasicBlock * Pred,unsigned int ValNo)2295 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2296                                     unsigned int ValNo) {
2297   // Because we are going top-down through the block, all value numbers
2298   // will be available in the predecessor by the time we need them.  Any
2299   // that weren't originally present will have been instantiated earlier
2300   // in this loop.
2301   bool success = true;
2302   for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2303     Value *Op = Instr->getOperand(i);
2304     if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2305       continue;
2306     // This could be a newly inserted instruction, in which case, we won't
2307     // find a value number, and should give up before we hurt ourselves.
2308     // FIXME: Rewrite the infrastructure to let it easier to value number
2309     // and process newly inserted instructions.
2310     if (!VN.exists(Op)) {
2311       success = false;
2312       break;
2313     }
2314     if (Value *V = findLeader(Pred, VN.lookup(Op))) {
2315       Instr->setOperand(i, V);
2316     } else {
2317       success = false;
2318       break;
2319     }
2320   }
2321 
2322   // Fail out if we encounter an operand that is not available in
2323   // the PRE predecessor.  This is typically because of loads which
2324   // are not value numbered precisely.
2325   if (!success)
2326     return false;
2327 
2328   Instr->insertBefore(Pred->getTerminator());
2329   Instr->setName(Instr->getName() + ".pre");
2330   Instr->setDebugLoc(Instr->getDebugLoc());
2331   VN.add(Instr, ValNo);
2332 
2333   // Update the availability map to include the new instruction.
2334   addToLeaderTable(ValNo, Instr, Pred);
2335   return true;
2336 }
2337 
performScalarPRE(Instruction * CurInst)2338 bool GVN::performScalarPRE(Instruction *CurInst) {
2339   if (isa<AllocaInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
2340       isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2341       CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2342       isa<DbgInfoIntrinsic>(CurInst))
2343     return false;
2344 
2345   // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2346   // sinking the compare again, and it would force the code generator to
2347   // move the i1 from processor flags or predicate registers into a general
2348   // purpose register.
2349   if (isa<CmpInst>(CurInst))
2350     return false;
2351 
2352   // We don't currently value number ANY inline asm calls.
2353   if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2354     if (CallI->isInlineAsm())
2355       return false;
2356 
2357   uint32_t ValNo = VN.lookup(CurInst);
2358 
2359   // Look for the predecessors for PRE opportunities.  We're
2360   // only trying to solve the basic diamond case, where
2361   // a value is computed in the successor and one predecessor,
2362   // but not the other.  We also explicitly disallow cases
2363   // where the successor is its own predecessor, because they're
2364   // more complicated to get right.
2365   unsigned NumWith = 0;
2366   unsigned NumWithout = 0;
2367   BasicBlock *PREPred = nullptr;
2368   BasicBlock *CurrentBlock = CurInst->getParent();
2369 
2370   SmallVector<std::pair<Value *, BasicBlock *>, 8> predMap;
2371   for (BasicBlock *P : predecessors(CurrentBlock)) {
2372     // We're not interested in PRE where the block is its
2373     // own predecessor, or in blocks with predecessors
2374     // that are not reachable.
2375     if (P == CurrentBlock) {
2376       NumWithout = 2;
2377       break;
2378     } else if (!DT->isReachableFromEntry(P)) {
2379       NumWithout = 2;
2380       break;
2381     }
2382 
2383     Value *predV = findLeader(P, ValNo);
2384     if (!predV) {
2385       predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2386       PREPred = P;
2387       ++NumWithout;
2388     } else if (predV == CurInst) {
2389       /* CurInst dominates this predecessor. */
2390       NumWithout = 2;
2391       break;
2392     } else {
2393       predMap.push_back(std::make_pair(predV, P));
2394       ++NumWith;
2395     }
2396   }
2397 
2398   // Don't do PRE when it might increase code size, i.e. when
2399   // we would need to insert instructions in more than one pred.
2400   if (NumWithout > 1 || NumWith == 0)
2401     return false;
2402 
2403   // We may have a case where all predecessors have the instruction,
2404   // and we just need to insert a phi node. Otherwise, perform
2405   // insertion.
2406   Instruction *PREInstr = nullptr;
2407 
2408   if (NumWithout != 0) {
2409     // Don't do PRE across indirect branch.
2410     if (isa<IndirectBrInst>(PREPred->getTerminator()))
2411       return false;
2412 
2413     // We can't do PRE safely on a critical edge, so instead we schedule
2414     // the edge to be split and perform the PRE the next time we iterate
2415     // on the function.
2416     unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2417     if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2418       toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2419       return false;
2420     }
2421     // We need to insert somewhere, so let's give it a shot
2422     PREInstr = CurInst->clone();
2423     if (!performScalarPREInsertion(PREInstr, PREPred, ValNo)) {
2424       // If we failed insertion, make sure we remove the instruction.
2425       DEBUG(verifyRemoved(PREInstr));
2426       delete PREInstr;
2427       return false;
2428     }
2429   }
2430 
2431   // Either we should have filled in the PRE instruction, or we should
2432   // not have needed insertions.
2433   assert (PREInstr != nullptr || NumWithout == 0);
2434 
2435   ++NumGVNPRE;
2436 
2437   // Create a PHI to make the value available in this block.
2438   PHINode *Phi =
2439       PHINode::Create(CurInst->getType(), predMap.size(),
2440                       CurInst->getName() + ".pre-phi", &CurrentBlock->front());
2441   for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2442     if (Value *V = predMap[i].first)
2443       Phi->addIncoming(V, predMap[i].second);
2444     else
2445       Phi->addIncoming(PREInstr, PREPred);
2446   }
2447 
2448   VN.add(Phi, ValNo);
2449   addToLeaderTable(ValNo, Phi, CurrentBlock);
2450   Phi->setDebugLoc(CurInst->getDebugLoc());
2451   CurInst->replaceAllUsesWith(Phi);
2452   if (MD && Phi->getType()->getScalarType()->isPointerTy())
2453     MD->invalidateCachedPointerInfo(Phi);
2454   VN.erase(CurInst);
2455   removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2456 
2457   DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2458   if (MD)
2459     MD->removeInstruction(CurInst);
2460   DEBUG(verifyRemoved(CurInst));
2461   CurInst->eraseFromParent();
2462   ++NumGVNInstr;
2463 
2464   return true;
2465 }
2466 
2467 /// Perform a purely local form of PRE that looks for diamond
2468 /// control flow patterns and attempts to perform simple PRE at the join point.
performPRE(Function & F)2469 bool GVN::performPRE(Function &F) {
2470   bool Changed = false;
2471   for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2472     // Nothing to PRE in the entry block.
2473     if (CurrentBlock == &F.getEntryBlock())
2474       continue;
2475 
2476     // Don't perform PRE on an EH pad.
2477     if (CurrentBlock->isEHPad())
2478       continue;
2479 
2480     for (BasicBlock::iterator BI = CurrentBlock->begin(),
2481                               BE = CurrentBlock->end();
2482          BI != BE;) {
2483       Instruction *CurInst = &*BI++;
2484       Changed |= performScalarPRE(CurInst);
2485     }
2486   }
2487 
2488   if (splitCriticalEdges())
2489     Changed = true;
2490 
2491   return Changed;
2492 }
2493 
2494 /// Split the critical edge connecting the given two blocks, and return
2495 /// the block inserted to the critical edge.
splitCriticalEdges(BasicBlock * Pred,BasicBlock * Succ)2496 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2497   BasicBlock *BB =
2498       SplitCriticalEdge(Pred, Succ, CriticalEdgeSplittingOptions(DT));
2499   if (MD)
2500     MD->invalidateCachedPredecessors();
2501   return BB;
2502 }
2503 
2504 /// Split critical edges found during the previous
2505 /// iteration that may enable further optimization.
splitCriticalEdges()2506 bool GVN::splitCriticalEdges() {
2507   if (toSplit.empty())
2508     return false;
2509   do {
2510     std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2511     SplitCriticalEdge(Edge.first, Edge.second,
2512                       CriticalEdgeSplittingOptions(DT));
2513   } while (!toSplit.empty());
2514   if (MD) MD->invalidateCachedPredecessors();
2515   return true;
2516 }
2517 
2518 /// Executes one iteration of GVN
iterateOnFunction(Function & F)2519 bool GVN::iterateOnFunction(Function &F) {
2520   cleanupGlobalSets();
2521 
2522   // Top-down walk of the dominator tree
2523   bool Changed = false;
2524   // Save the blocks this function have before transformation begins. GVN may
2525   // split critical edge, and hence may invalidate the RPO/DT iterator.
2526   //
2527   std::vector<BasicBlock *> BBVect;
2528   BBVect.reserve(256);
2529   // Needed for value numbering with phi construction to work.
2530   ReversePostOrderTraversal<Function *> RPOT(&F);
2531   for (ReversePostOrderTraversal<Function *>::rpo_iterator RI = RPOT.begin(),
2532                                                            RE = RPOT.end();
2533        RI != RE; ++RI)
2534     BBVect.push_back(*RI);
2535 
2536   for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
2537        I != E; I++)
2538     Changed |= processBlock(*I);
2539 
2540   return Changed;
2541 }
2542 
cleanupGlobalSets()2543 void GVN::cleanupGlobalSets() {
2544   VN.clear();
2545   LeaderTable.clear();
2546   TableAllocator.Reset();
2547 }
2548 
2549 /// Verify that the specified instruction does not occur in our
2550 /// internal data structures.
verifyRemoved(const Instruction * Inst) const2551 void GVN::verifyRemoved(const Instruction *Inst) const {
2552   VN.verifyRemoved(Inst);
2553 
2554   // Walk through the value number scope to make sure the instruction isn't
2555   // ferreted away in it.
2556   for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2557        I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2558     const LeaderTableEntry *Node = &I->second;
2559     assert(Node->Val != Inst && "Inst still in value numbering scope!");
2560 
2561     while (Node->Next) {
2562       Node = Node->Next;
2563       assert(Node->Val != Inst && "Inst still in value numbering scope!");
2564     }
2565   }
2566 }
2567 
2568 /// BB is declared dead, which implied other blocks become dead as well. This
2569 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2570 /// live successors, update their phi nodes by replacing the operands
2571 /// corresponding to dead blocks with UndefVal.
addDeadBlock(BasicBlock * BB)2572 void GVN::addDeadBlock(BasicBlock *BB) {
2573   SmallVector<BasicBlock *, 4> NewDead;
2574   SmallSetVector<BasicBlock *, 4> DF;
2575 
2576   NewDead.push_back(BB);
2577   while (!NewDead.empty()) {
2578     BasicBlock *D = NewDead.pop_back_val();
2579     if (DeadBlocks.count(D))
2580       continue;
2581 
2582     // All blocks dominated by D are dead.
2583     SmallVector<BasicBlock *, 8> Dom;
2584     DT->getDescendants(D, Dom);
2585     DeadBlocks.insert(Dom.begin(), Dom.end());
2586 
2587     // Figure out the dominance-frontier(D).
2588     for (BasicBlock *B : Dom) {
2589       for (BasicBlock *S : successors(B)) {
2590         if (DeadBlocks.count(S))
2591           continue;
2592 
2593         bool AllPredDead = true;
2594         for (BasicBlock *P : predecessors(S))
2595           if (!DeadBlocks.count(P)) {
2596             AllPredDead = false;
2597             break;
2598           }
2599 
2600         if (!AllPredDead) {
2601           // S could be proved dead later on. That is why we don't update phi
2602           // operands at this moment.
2603           DF.insert(S);
2604         } else {
2605           // While S is not dominated by D, it is dead by now. This could take
2606           // place if S already have a dead predecessor before D is declared
2607           // dead.
2608           NewDead.push_back(S);
2609         }
2610       }
2611     }
2612   }
2613 
2614   // For the dead blocks' live successors, update their phi nodes by replacing
2615   // the operands corresponding to dead blocks with UndefVal.
2616   for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2617         I != E; I++) {
2618     BasicBlock *B = *I;
2619     if (DeadBlocks.count(B))
2620       continue;
2621 
2622     SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2623     for (BasicBlock *P : Preds) {
2624       if (!DeadBlocks.count(P))
2625         continue;
2626 
2627       if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2628         if (BasicBlock *S = splitCriticalEdges(P, B))
2629           DeadBlocks.insert(P = S);
2630       }
2631 
2632       for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2633         PHINode &Phi = cast<PHINode>(*II);
2634         Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2635                              UndefValue::get(Phi.getType()));
2636       }
2637     }
2638   }
2639 }
2640 
2641 // If the given branch is recognized as a foldable branch (i.e. conditional
2642 // branch with constant condition), it will perform following analyses and
2643 // transformation.
2644 //  1) If the dead out-coming edge is a critical-edge, split it. Let
2645 //     R be the target of the dead out-coming edge.
2646 //  1) Identify the set of dead blocks implied by the branch's dead outcoming
2647 //     edge. The result of this step will be {X| X is dominated by R}
2648 //  2) Identify those blocks which haves at least one dead predecessor. The
2649 //     result of this step will be dominance-frontier(R).
2650 //  3) Update the PHIs in DF(R) by replacing the operands corresponding to
2651 //     dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2652 //
2653 // Return true iff *NEW* dead code are found.
processFoldableCondBr(BranchInst * BI)2654 bool GVN::processFoldableCondBr(BranchInst *BI) {
2655   if (!BI || BI->isUnconditional())
2656     return false;
2657 
2658   // If a branch has two identical successors, we cannot declare either dead.
2659   if (BI->getSuccessor(0) == BI->getSuccessor(1))
2660     return false;
2661 
2662   ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2663   if (!Cond)
2664     return false;
2665 
2666   BasicBlock *DeadRoot =
2667       Cond->getZExtValue() ? BI->getSuccessor(1) : BI->getSuccessor(0);
2668   if (DeadBlocks.count(DeadRoot))
2669     return false;
2670 
2671   if (!DeadRoot->getSinglePredecessor())
2672     DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2673 
2674   addDeadBlock(DeadRoot);
2675   return true;
2676 }
2677 
2678 // performPRE() will trigger assert if it comes across an instruction without
2679 // associated val-num. As it normally has far more live instructions than dead
2680 // instructions, it makes more sense just to "fabricate" a val-number for the
2681 // dead code than checking if instruction involved is dead or not.
assignValNumForDeadCode()2682 void GVN::assignValNumForDeadCode() {
2683   for (BasicBlock *BB : DeadBlocks) {
2684     for (Instruction &Inst : *BB) {
2685       unsigned ValNum = VN.lookupOrAdd(&Inst);
2686       addToLeaderTable(ValNum, &Inst, BB);
2687     }
2688   }
2689 }
2690 
2691 class llvm::gvn::GVNLegacyPass : public FunctionPass {
2692 public:
2693   static char ID; // Pass identification, replacement for typeid
GVNLegacyPass(bool NoLoads=false)2694   explicit GVNLegacyPass(bool NoLoads = false)
2695       : FunctionPass(ID), NoLoads(NoLoads) {
2696     initializeGVNLegacyPassPass(*PassRegistry::getPassRegistry());
2697   }
2698 
runOnFunction(Function & F)2699   bool runOnFunction(Function &F) override {
2700     if (skipFunction(F))
2701       return false;
2702 
2703     return Impl.runImpl(
2704         F, getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F),
2705         getAnalysis<DominatorTreeWrapperPass>().getDomTree(),
2706         getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(),
2707         getAnalysis<AAResultsWrapperPass>().getAAResults(),
2708         NoLoads ? nullptr
2709                 : &getAnalysis<MemoryDependenceWrapperPass>().getMemDep());
2710   }
2711 
getAnalysisUsage(AnalysisUsage & AU) const2712   void getAnalysisUsage(AnalysisUsage &AU) const override {
2713     AU.addRequired<AssumptionCacheTracker>();
2714     AU.addRequired<DominatorTreeWrapperPass>();
2715     AU.addRequired<TargetLibraryInfoWrapperPass>();
2716     if (!NoLoads)
2717       AU.addRequired<MemoryDependenceWrapperPass>();
2718     AU.addRequired<AAResultsWrapperPass>();
2719 
2720     AU.addPreserved<DominatorTreeWrapperPass>();
2721     AU.addPreserved<GlobalsAAWrapperPass>();
2722   }
2723 
2724 private:
2725   bool NoLoads;
2726   GVN Impl;
2727 };
2728 
2729 char GVNLegacyPass::ID = 0;
2730 
2731 // The public interface to this file...
createGVNPass(bool NoLoads)2732 FunctionPass *llvm::createGVNPass(bool NoLoads) {
2733   return new GVNLegacyPass(NoLoads);
2734 }
2735 
2736 INITIALIZE_PASS_BEGIN(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2737 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
2738 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceWrapperPass)
2739 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
2740 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
2741 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
2742 INITIALIZE_PASS_DEPENDENCY(GlobalsAAWrapperPass)
2743 INITIALIZE_PASS_END(GVNLegacyPass, "gvn", "Global Value Numbering", false, false)
2744