• 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 #define DEBUG_TYPE "gvn"
19 #include "llvm/Transforms/Scalar.h"
20 #include "llvm/GlobalVariable.h"
21 #include "llvm/IntrinsicInst.h"
22 #include "llvm/LLVMContext.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/ConstantFolding.h"
25 #include "llvm/Analysis/Dominators.h"
26 #include "llvm/Analysis/InstructionSimplify.h"
27 #include "llvm/Analysis/Loads.h"
28 #include "llvm/Analysis/MemoryBuiltins.h"
29 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
30 #include "llvm/Analysis/PHITransAddr.h"
31 #include "llvm/Analysis/ValueTracking.h"
32 #include "llvm/Assembly/Writer.h"
33 #include "llvm/Target/TargetData.h"
34 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
35 #include "llvm/Transforms/Utils/SSAUpdater.h"
36 #include "llvm/ADT/DenseMap.h"
37 #include "llvm/ADT/DepthFirstIterator.h"
38 #include "llvm/ADT/SmallPtrSet.h"
39 #include "llvm/ADT/Statistic.h"
40 #include "llvm/Support/Allocator.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/IRBuilder.h"
44 #include "llvm/Support/PatternMatch.h"
45 using namespace llvm;
46 using namespace PatternMatch;
47 
48 STATISTIC(NumGVNInstr,  "Number of instructions deleted");
49 STATISTIC(NumGVNLoad,   "Number of loads deleted");
50 STATISTIC(NumGVNPRE,    "Number of instructions PRE'd");
51 STATISTIC(NumGVNBlocks, "Number of blocks merged");
52 STATISTIC(NumGVNSimpl,  "Number of instructions simplified");
53 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
54 STATISTIC(NumPRELoad,   "Number of loads PRE'd");
55 
56 static cl::opt<bool> EnablePRE("enable-pre",
57                                cl::init(true), cl::Hidden);
58 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
59 
60 //===----------------------------------------------------------------------===//
61 //                         ValueTable Class
62 //===----------------------------------------------------------------------===//
63 
64 /// This class holds the mapping between values and value numbers.  It is used
65 /// as an efficient mechanism to determine the expression-wise equivalence of
66 /// two values.
67 namespace {
68   struct Expression {
69     uint32_t opcode;
70     Type *type;
71     SmallVector<uint32_t, 4> varargs;
72 
Expression__anonadea466b0111::Expression73     Expression(uint32_t o = ~2U) : opcode(o) { }
74 
operator ==__anonadea466b0111::Expression75     bool operator==(const Expression &other) const {
76       if (opcode != other.opcode)
77         return false;
78       if (opcode == ~0U || opcode == ~1U)
79         return true;
80       if (type != other.type)
81         return false;
82       if (varargs != other.varargs)
83         return false;
84       return true;
85     }
86   };
87 
88   class ValueTable {
89     DenseMap<Value*, uint32_t> valueNumbering;
90     DenseMap<Expression, uint32_t> expressionNumbering;
91     AliasAnalysis *AA;
92     MemoryDependenceAnalysis *MD;
93     DominatorTree *DT;
94 
95     uint32_t nextValueNumber;
96 
97     Expression create_expression(Instruction* I);
98     Expression create_extractvalue_expression(ExtractValueInst* EI);
99     uint32_t lookup_or_add_call(CallInst* C);
100   public:
ValueTable()101     ValueTable() : nextValueNumber(1) { }
102     uint32_t lookup_or_add(Value *V);
103     uint32_t lookup(Value *V) const;
104     void add(Value *V, uint32_t num);
105     void clear();
106     void erase(Value *v);
setAliasAnalysis(AliasAnalysis * A)107     void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
getAliasAnalysis() const108     AliasAnalysis *getAliasAnalysis() const { return AA; }
setMemDep(MemoryDependenceAnalysis * M)109     void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
setDomTree(DominatorTree * D)110     void setDomTree(DominatorTree* D) { DT = D; }
getNextUnusedValueNumber()111     uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
112     void verifyRemoved(const Value *) const;
113   };
114 }
115 
116 namespace llvm {
117 template <> struct DenseMapInfo<Expression> {
getEmptyKeyllvm::DenseMapInfo118   static inline Expression getEmptyKey() {
119     return ~0U;
120   }
121 
getTombstoneKeyllvm::DenseMapInfo122   static inline Expression getTombstoneKey() {
123     return ~1U;
124   }
125 
getHashValuellvm::DenseMapInfo126   static unsigned getHashValue(const Expression e) {
127     unsigned hash = e.opcode;
128 
129     hash = ((unsigned)((uintptr_t)e.type >> 4) ^
130             (unsigned)((uintptr_t)e.type >> 9));
131 
132     for (SmallVector<uint32_t, 4>::const_iterator I = e.varargs.begin(),
133          E = e.varargs.end(); I != E; ++I)
134       hash = *I + hash * 37;
135 
136     return hash;
137   }
isEqualllvm::DenseMapInfo138   static bool isEqual(const Expression &LHS, const Expression &RHS) {
139     return LHS == RHS;
140   }
141 };
142 
143 }
144 
145 //===----------------------------------------------------------------------===//
146 //                     ValueTable Internal Functions
147 //===----------------------------------------------------------------------===//
148 
create_expression(Instruction * I)149 Expression ValueTable::create_expression(Instruction *I) {
150   Expression e;
151   e.type = I->getType();
152   e.opcode = I->getOpcode();
153   for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
154        OI != OE; ++OI)
155     e.varargs.push_back(lookup_or_add(*OI));
156 
157   if (CmpInst *C = dyn_cast<CmpInst>(I)) {
158     e.opcode = (C->getOpcode() << 8) | C->getPredicate();
159   } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
160     for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
161          II != IE; ++II)
162       e.varargs.push_back(*II);
163   }
164 
165   return e;
166 }
167 
create_extractvalue_expression(ExtractValueInst * EI)168 Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) {
169   assert(EI != 0 && "Not an ExtractValueInst?");
170   Expression e;
171   e.type = EI->getType();
172   e.opcode = 0;
173 
174   IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
175   if (I != 0 && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
176     // EI might be an extract from one of our recognised intrinsics. If it
177     // is we'll synthesize a semantically equivalent expression instead on
178     // an extract value expression.
179     switch (I->getIntrinsicID()) {
180       case Intrinsic::sadd_with_overflow:
181       case Intrinsic::uadd_with_overflow:
182         e.opcode = Instruction::Add;
183         break;
184       case Intrinsic::ssub_with_overflow:
185       case Intrinsic::usub_with_overflow:
186         e.opcode = Instruction::Sub;
187         break;
188       case Intrinsic::smul_with_overflow:
189       case Intrinsic::umul_with_overflow:
190         e.opcode = Instruction::Mul;
191         break;
192       default:
193         break;
194     }
195 
196     if (e.opcode != 0) {
197       // Intrinsic recognized. Grab its args to finish building the expression.
198       assert(I->getNumArgOperands() == 2 &&
199              "Expect two args for recognised intrinsics.");
200       e.varargs.push_back(lookup_or_add(I->getArgOperand(0)));
201       e.varargs.push_back(lookup_or_add(I->getArgOperand(1)));
202       return e;
203     }
204   }
205 
206   // Not a recognised intrinsic. Fall back to producing an extract value
207   // expression.
208   e.opcode = EI->getOpcode();
209   for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
210        OI != OE; ++OI)
211     e.varargs.push_back(lookup_or_add(*OI));
212 
213   for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
214          II != IE; ++II)
215     e.varargs.push_back(*II);
216 
217   return e;
218 }
219 
220 //===----------------------------------------------------------------------===//
221 //                     ValueTable External Functions
222 //===----------------------------------------------------------------------===//
223 
224 /// add - Insert a value into the table with a specified value number.
add(Value * V,uint32_t num)225 void ValueTable::add(Value *V, uint32_t num) {
226   valueNumbering.insert(std::make_pair(V, num));
227 }
228 
lookup_or_add_call(CallInst * C)229 uint32_t ValueTable::lookup_or_add_call(CallInst* C) {
230   if (AA->doesNotAccessMemory(C)) {
231     Expression exp = create_expression(C);
232     uint32_t& e = expressionNumbering[exp];
233     if (!e) e = nextValueNumber++;
234     valueNumbering[C] = e;
235     return e;
236   } else if (AA->onlyReadsMemory(C)) {
237     Expression exp = create_expression(C);
238     uint32_t& e = expressionNumbering[exp];
239     if (!e) {
240       e = nextValueNumber++;
241       valueNumbering[C] = e;
242       return e;
243     }
244     if (!MD) {
245       e = nextValueNumber++;
246       valueNumbering[C] = e;
247       return e;
248     }
249 
250     MemDepResult local_dep = MD->getDependency(C);
251 
252     if (!local_dep.isDef() && !local_dep.isNonLocal()) {
253       valueNumbering[C] =  nextValueNumber;
254       return nextValueNumber++;
255     }
256 
257     if (local_dep.isDef()) {
258       CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
259 
260       if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
261         valueNumbering[C] = nextValueNumber;
262         return nextValueNumber++;
263       }
264 
265       for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
266         uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
267         uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
268         if (c_vn != cd_vn) {
269           valueNumbering[C] = nextValueNumber;
270           return nextValueNumber++;
271         }
272       }
273 
274       uint32_t v = lookup_or_add(local_cdep);
275       valueNumbering[C] = v;
276       return v;
277     }
278 
279     // Non-local case.
280     const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
281       MD->getNonLocalCallDependency(CallSite(C));
282     // FIXME: Move the checking logic to MemDep!
283     CallInst* cdep = 0;
284 
285     // Check to see if we have a single dominating call instruction that is
286     // identical to C.
287     for (unsigned i = 0, e = deps.size(); i != e; ++i) {
288       const NonLocalDepEntry *I = &deps[i];
289       if (I->getResult().isNonLocal())
290         continue;
291 
292       // We don't handle non-definitions.  If we already have a call, reject
293       // instruction dependencies.
294       if (!I->getResult().isDef() || cdep != 0) {
295         cdep = 0;
296         break;
297       }
298 
299       CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
300       // FIXME: All duplicated with non-local case.
301       if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
302         cdep = NonLocalDepCall;
303         continue;
304       }
305 
306       cdep = 0;
307       break;
308     }
309 
310     if (!cdep) {
311       valueNumbering[C] = nextValueNumber;
312       return nextValueNumber++;
313     }
314 
315     if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
316       valueNumbering[C] = nextValueNumber;
317       return nextValueNumber++;
318     }
319     for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
320       uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
321       uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
322       if (c_vn != cd_vn) {
323         valueNumbering[C] = nextValueNumber;
324         return nextValueNumber++;
325       }
326     }
327 
328     uint32_t v = lookup_or_add(cdep);
329     valueNumbering[C] = v;
330     return v;
331 
332   } else {
333     valueNumbering[C] = nextValueNumber;
334     return nextValueNumber++;
335   }
336 }
337 
338 /// lookup_or_add - Returns the value number for the specified value, assigning
339 /// it a new number if it did not have one before.
lookup_or_add(Value * V)340 uint32_t ValueTable::lookup_or_add(Value *V) {
341   DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
342   if (VI != valueNumbering.end())
343     return VI->second;
344 
345   if (!isa<Instruction>(V)) {
346     valueNumbering[V] = nextValueNumber;
347     return nextValueNumber++;
348   }
349 
350   Instruction* I = cast<Instruction>(V);
351   Expression exp;
352   switch (I->getOpcode()) {
353     case Instruction::Call:
354       return lookup_or_add_call(cast<CallInst>(I));
355     case Instruction::Add:
356     case Instruction::FAdd:
357     case Instruction::Sub:
358     case Instruction::FSub:
359     case Instruction::Mul:
360     case Instruction::FMul:
361     case Instruction::UDiv:
362     case Instruction::SDiv:
363     case Instruction::FDiv:
364     case Instruction::URem:
365     case Instruction::SRem:
366     case Instruction::FRem:
367     case Instruction::Shl:
368     case Instruction::LShr:
369     case Instruction::AShr:
370     case Instruction::And:
371     case Instruction::Or :
372     case Instruction::Xor:
373     case Instruction::ICmp:
374     case Instruction::FCmp:
375     case Instruction::Trunc:
376     case Instruction::ZExt:
377     case Instruction::SExt:
378     case Instruction::FPToUI:
379     case Instruction::FPToSI:
380     case Instruction::UIToFP:
381     case Instruction::SIToFP:
382     case Instruction::FPTrunc:
383     case Instruction::FPExt:
384     case Instruction::PtrToInt:
385     case Instruction::IntToPtr:
386     case Instruction::BitCast:
387     case Instruction::Select:
388     case Instruction::ExtractElement:
389     case Instruction::InsertElement:
390     case Instruction::ShuffleVector:
391     case Instruction::InsertValue:
392     case Instruction::GetElementPtr:
393       exp = create_expression(I);
394       break;
395     case Instruction::ExtractValue:
396       exp = create_extractvalue_expression(cast<ExtractValueInst>(I));
397       break;
398     default:
399       valueNumbering[V] = nextValueNumber;
400       return nextValueNumber++;
401   }
402 
403   uint32_t& e = expressionNumbering[exp];
404   if (!e) e = nextValueNumber++;
405   valueNumbering[V] = e;
406   return e;
407 }
408 
409 /// lookup - Returns the value number of the specified value. Fails if
410 /// the value has not yet been numbered.
lookup(Value * V) const411 uint32_t ValueTable::lookup(Value *V) const {
412   DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
413   assert(VI != valueNumbering.end() && "Value not numbered?");
414   return VI->second;
415 }
416 
417 /// clear - Remove all entries from the ValueTable.
clear()418 void ValueTable::clear() {
419   valueNumbering.clear();
420   expressionNumbering.clear();
421   nextValueNumber = 1;
422 }
423 
424 /// erase - Remove a value from the value numbering.
erase(Value * V)425 void ValueTable::erase(Value *V) {
426   valueNumbering.erase(V);
427 }
428 
429 /// verifyRemoved - Verify that the value is removed from all internal data
430 /// structures.
verifyRemoved(const Value * V) const431 void ValueTable::verifyRemoved(const Value *V) const {
432   for (DenseMap<Value*, uint32_t>::const_iterator
433          I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
434     assert(I->first != V && "Inst still occurs in value numbering map!");
435   }
436 }
437 
438 //===----------------------------------------------------------------------===//
439 //                                GVN Pass
440 //===----------------------------------------------------------------------===//
441 
442 namespace {
443 
444   class GVN : public FunctionPass {
445     bool NoLoads;
446     MemoryDependenceAnalysis *MD;
447     DominatorTree *DT;
448     const TargetData *TD;
449 
450     ValueTable VN;
451 
452     /// LeaderTable - A mapping from value numbers to lists of Value*'s that
453     /// have that value number.  Use findLeader to query it.
454     struct LeaderTableEntry {
455       Value *Val;
456       BasicBlock *BB;
457       LeaderTableEntry *Next;
458     };
459     DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
460     BumpPtrAllocator TableAllocator;
461 
462     SmallVector<Instruction*, 8> InstrsToErase;
463   public:
464     static char ID; // Pass identification, replacement for typeid
GVN(bool noloads=false)465     explicit GVN(bool noloads = false)
466         : FunctionPass(ID), NoLoads(noloads), MD(0) {
467       initializeGVNPass(*PassRegistry::getPassRegistry());
468     }
469 
470     bool runOnFunction(Function &F);
471 
472     /// markInstructionForDeletion - This removes the specified instruction from
473     /// our various maps and marks it for deletion.
markInstructionForDeletion(Instruction * I)474     void markInstructionForDeletion(Instruction *I) {
475       VN.erase(I);
476       InstrsToErase.push_back(I);
477     }
478 
getTargetData() const479     const TargetData *getTargetData() const { return TD; }
getDominatorTree() const480     DominatorTree &getDominatorTree() const { return *DT; }
getAliasAnalysis() const481     AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
getMemDep() const482     MemoryDependenceAnalysis &getMemDep() const { return *MD; }
483   private:
484     /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for
485     /// its value number.
addToLeaderTable(uint32_t N,Value * V,BasicBlock * BB)486     void addToLeaderTable(uint32_t N, Value *V, BasicBlock *BB) {
487       LeaderTableEntry &Curr = LeaderTable[N];
488       if (!Curr.Val) {
489         Curr.Val = V;
490         Curr.BB = BB;
491         return;
492       }
493 
494       LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
495       Node->Val = V;
496       Node->BB = BB;
497       Node->Next = Curr.Next;
498       Curr.Next = Node;
499     }
500 
501     /// removeFromLeaderTable - Scan the list of values corresponding to a given
502     /// value number, and remove the given value if encountered.
removeFromLeaderTable(uint32_t N,Value * V,BasicBlock * BB)503     void removeFromLeaderTable(uint32_t N, Value *V, BasicBlock *BB) {
504       LeaderTableEntry* Prev = 0;
505       LeaderTableEntry* Curr = &LeaderTable[N];
506 
507       while (Curr->Val != V || Curr->BB != BB) {
508         Prev = Curr;
509         Curr = Curr->Next;
510       }
511 
512       if (Prev) {
513         Prev->Next = Curr->Next;
514       } else {
515         if (!Curr->Next) {
516           Curr->Val = 0;
517           Curr->BB = 0;
518         } else {
519           LeaderTableEntry* Next = Curr->Next;
520           Curr->Val = Next->Val;
521           Curr->BB = Next->BB;
522           Curr->Next = Next->Next;
523         }
524       }
525     }
526 
527     // List of critical edges to be split between iterations.
528     SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
529 
530     // This transformation requires dominator postdominator info
getAnalysisUsage(AnalysisUsage & AU) const531     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
532       AU.addRequired<DominatorTree>();
533       if (!NoLoads)
534         AU.addRequired<MemoryDependenceAnalysis>();
535       AU.addRequired<AliasAnalysis>();
536 
537       AU.addPreserved<DominatorTree>();
538       AU.addPreserved<AliasAnalysis>();
539     }
540 
541 
542     // Helper fuctions
543     // FIXME: eliminate or document these better
544     bool processLoad(LoadInst *L);
545     bool processInstruction(Instruction *I);
546     bool processNonLocalLoad(LoadInst *L);
547     bool processBlock(BasicBlock *BB);
548     void dump(DenseMap<uint32_t, Value*> &d);
549     bool iterateOnFunction(Function &F);
550     bool performPRE(Function &F);
551     Value *findLeader(BasicBlock *BB, uint32_t num);
552     void cleanupGlobalSets();
553     void verifyRemoved(const Instruction *I) const;
554     bool splitCriticalEdges();
555     unsigned replaceAllDominatedUsesWith(Value *From, Value *To,
556                                          BasicBlock *Root);
557     bool propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root);
558   };
559 
560   char GVN::ID = 0;
561 }
562 
563 // createGVNPass - The public interface to this file...
createGVNPass(bool NoLoads)564 FunctionPass *llvm::createGVNPass(bool NoLoads) {
565   return new GVN(NoLoads);
566 }
567 
568 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)569 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
570 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
571 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
572 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
573 
574 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
575   errs() << "{\n";
576   for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
577        E = d.end(); I != E; ++I) {
578       errs() << I->first << "\n";
579       I->second->dump();
580   }
581   errs() << "}\n";
582 }
583 
584 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
585 /// we're analyzing is fully available in the specified block.  As we go, keep
586 /// track of which blocks we know are fully alive in FullyAvailableBlocks.  This
587 /// map is actually a tri-state map with the following values:
588 ///   0) we know the block *is not* fully available.
589 ///   1) we know the block *is* fully available.
590 ///   2) we do not know whether the block is fully available or not, but we are
591 ///      currently speculating that it will be.
592 ///   3) we are speculating for this block and have used that to speculate for
593 ///      other blocks.
IsValueFullyAvailableInBlock(BasicBlock * BB,DenseMap<BasicBlock *,char> & FullyAvailableBlocks)594 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
595                             DenseMap<BasicBlock*, char> &FullyAvailableBlocks) {
596   // Optimistically assume that the block is fully available and check to see
597   // if we already know about this block in one lookup.
598   std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
599     FullyAvailableBlocks.insert(std::make_pair(BB, 2));
600 
601   // If the entry already existed for this block, return the precomputed value.
602   if (!IV.second) {
603     // If this is a speculative "available" value, mark it as being used for
604     // speculation of other blocks.
605     if (IV.first->second == 2)
606       IV.first->second = 3;
607     return IV.first->second != 0;
608   }
609 
610   // Otherwise, see if it is fully available in all predecessors.
611   pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
612 
613   // If this block has no predecessors, it isn't live-in here.
614   if (PI == PE)
615     goto SpeculationFailure;
616 
617   for (; PI != PE; ++PI)
618     // If the value isn't fully available in one of our predecessors, then it
619     // isn't fully available in this block either.  Undo our previous
620     // optimistic assumption and bail out.
621     if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks))
622       goto SpeculationFailure;
623 
624   return true;
625 
626 // SpeculationFailure - If we get here, we found out that this is not, after
627 // all, a fully-available block.  We have a problem if we speculated on this and
628 // used the speculation to mark other blocks as available.
629 SpeculationFailure:
630   char &BBVal = FullyAvailableBlocks[BB];
631 
632   // If we didn't speculate on this, just return with it set to false.
633   if (BBVal == 2) {
634     BBVal = 0;
635     return false;
636   }
637 
638   // If we did speculate on this value, we could have blocks set to 1 that are
639   // incorrect.  Walk the (transitive) successors of this block and mark them as
640   // 0 if set to one.
641   SmallVector<BasicBlock*, 32> BBWorklist;
642   BBWorklist.push_back(BB);
643 
644   do {
645     BasicBlock *Entry = BBWorklist.pop_back_val();
646     // Note that this sets blocks to 0 (unavailable) if they happen to not
647     // already be in FullyAvailableBlocks.  This is safe.
648     char &EntryVal = FullyAvailableBlocks[Entry];
649     if (EntryVal == 0) continue;  // Already unavailable.
650 
651     // Mark as unavailable.
652     EntryVal = 0;
653 
654     for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
655       BBWorklist.push_back(*I);
656   } while (!BBWorklist.empty());
657 
658   return false;
659 }
660 
661 
662 /// CanCoerceMustAliasedValueToLoad - Return true if
663 /// CoerceAvailableValueToLoadType will succeed.
CanCoerceMustAliasedValueToLoad(Value * StoredVal,Type * LoadTy,const TargetData & TD)664 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
665                                             Type *LoadTy,
666                                             const TargetData &TD) {
667   // If the loaded or stored value is an first class array or struct, don't try
668   // to transform them.  We need to be able to bitcast to integer.
669   if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
670       StoredVal->getType()->isStructTy() ||
671       StoredVal->getType()->isArrayTy())
672     return false;
673 
674   // The store has to be at least as big as the load.
675   if (TD.getTypeSizeInBits(StoredVal->getType()) <
676         TD.getTypeSizeInBits(LoadTy))
677     return false;
678 
679   return true;
680 }
681 
682 
683 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
684 /// then a load from a must-aliased pointer of a different type, try to coerce
685 /// the stored value.  LoadedTy is the type of the load we want to replace and
686 /// InsertPt is the place to insert new instructions.
687 ///
688 /// If we can't do it, return null.
CoerceAvailableValueToLoadType(Value * StoredVal,Type * LoadedTy,Instruction * InsertPt,const TargetData & TD)689 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
690                                              Type *LoadedTy,
691                                              Instruction *InsertPt,
692                                              const TargetData &TD) {
693   if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
694     return 0;
695 
696   // If this is already the right type, just return it.
697   Type *StoredValTy = StoredVal->getType();
698 
699   uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy);
700   uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
701 
702   // If the store and reload are the same size, we can always reuse it.
703   if (StoreSize == LoadSize) {
704     // Pointer to Pointer -> use bitcast.
705     if (StoredValTy->isPointerTy() && LoadedTy->isPointerTy())
706       return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
707 
708     // Convert source pointers to integers, which can be bitcast.
709     if (StoredValTy->isPointerTy()) {
710       StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
711       StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
712     }
713 
714     Type *TypeToCastTo = LoadedTy;
715     if (TypeToCastTo->isPointerTy())
716       TypeToCastTo = TD.getIntPtrType(StoredValTy->getContext());
717 
718     if (StoredValTy != TypeToCastTo)
719       StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
720 
721     // Cast to pointer if the load needs a pointer type.
722     if (LoadedTy->isPointerTy())
723       StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
724 
725     return StoredVal;
726   }
727 
728   // If the loaded value is smaller than the available value, then we can
729   // extract out a piece from it.  If the available value is too small, then we
730   // can't do anything.
731   assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
732 
733   // Convert source pointers to integers, which can be manipulated.
734   if (StoredValTy->isPointerTy()) {
735     StoredValTy = TD.getIntPtrType(StoredValTy->getContext());
736     StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
737   }
738 
739   // Convert vectors and fp to integer, which can be manipulated.
740   if (!StoredValTy->isIntegerTy()) {
741     StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
742     StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
743   }
744 
745   // If this is a big-endian system, we need to shift the value down to the low
746   // bits so that a truncate will work.
747   if (TD.isBigEndian()) {
748     Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
749     StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
750   }
751 
752   // Truncate the integer to the right size now.
753   Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
754   StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
755 
756   if (LoadedTy == NewIntTy)
757     return StoredVal;
758 
759   // If the result is a pointer, inttoptr.
760   if (LoadedTy->isPointerTy())
761     return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
762 
763   // Otherwise, bitcast.
764   return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
765 }
766 
767 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
768 /// memdep query of a load that ends up being a clobbering memory write (store,
769 /// memset, memcpy, memmove).  This means that the write *may* provide bits used
770 /// by the load but we can't be sure because the pointers don't mustalias.
771 ///
772 /// Check this case to see if there is anything more we can do before we give
773 /// up.  This returns -1 if we have to give up, or a byte number in the stored
774 /// value of the piece that feeds the load.
AnalyzeLoadFromClobberingWrite(Type * LoadTy,Value * LoadPtr,Value * WritePtr,uint64_t WriteSizeInBits,const TargetData & TD)775 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
776                                           Value *WritePtr,
777                                           uint64_t WriteSizeInBits,
778                                           const TargetData &TD) {
779   // If the loaded or stored value is an first class array or struct, don't try
780   // to transform them.  We need to be able to bitcast to integer.
781   if (LoadTy->isStructTy() || LoadTy->isArrayTy())
782     return -1;
783 
784   int64_t StoreOffset = 0, LoadOffset = 0;
785   Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr, StoreOffset,TD);
786   Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, TD);
787   if (StoreBase != LoadBase)
788     return -1;
789 
790   // If the load and store are to the exact same address, they should have been
791   // a must alias.  AA must have gotten confused.
792   // FIXME: Study to see if/when this happens.  One case is forwarding a memset
793   // to a load from the base of the memset.
794 #if 0
795   if (LoadOffset == StoreOffset) {
796     dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
797     << "Base       = " << *StoreBase << "\n"
798     << "Store Ptr  = " << *WritePtr << "\n"
799     << "Store Offs = " << StoreOffset << "\n"
800     << "Load Ptr   = " << *LoadPtr << "\n";
801     abort();
802   }
803 #endif
804 
805   // If the load and store don't overlap at all, the store doesn't provide
806   // anything to the load.  In this case, they really don't alias at all, AA
807   // must have gotten confused.
808   uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
809 
810   if ((WriteSizeInBits & 7) | (LoadSize & 7))
811     return -1;
812   uint64_t StoreSize = WriteSizeInBits >> 3;  // Convert to bytes.
813   LoadSize >>= 3;
814 
815 
816   bool isAAFailure = false;
817   if (StoreOffset < LoadOffset)
818     isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
819   else
820     isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
821 
822   if (isAAFailure) {
823 #if 0
824     dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
825     << "Base       = " << *StoreBase << "\n"
826     << "Store Ptr  = " << *WritePtr << "\n"
827     << "Store Offs = " << StoreOffset << "\n"
828     << "Load Ptr   = " << *LoadPtr << "\n";
829     abort();
830 #endif
831     return -1;
832   }
833 
834   // If the Load isn't completely contained within the stored bits, we don't
835   // have all the bits to feed it.  We could do something crazy in the future
836   // (issue a smaller load then merge the bits in) but this seems unlikely to be
837   // valuable.
838   if (StoreOffset > LoadOffset ||
839       StoreOffset+StoreSize < LoadOffset+LoadSize)
840     return -1;
841 
842   // Okay, we can do this transformation.  Return the number of bytes into the
843   // store that the load is.
844   return LoadOffset-StoreOffset;
845 }
846 
847 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
848 /// memdep query of a load that ends up being a clobbering store.
AnalyzeLoadFromClobberingStore(Type * LoadTy,Value * LoadPtr,StoreInst * DepSI,const TargetData & TD)849 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
850                                           StoreInst *DepSI,
851                                           const TargetData &TD) {
852   // Cannot handle reading from store of first-class aggregate yet.
853   if (DepSI->getValueOperand()->getType()->isStructTy() ||
854       DepSI->getValueOperand()->getType()->isArrayTy())
855     return -1;
856 
857   Value *StorePtr = DepSI->getPointerOperand();
858   uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType());
859   return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
860                                         StorePtr, StoreSize, TD);
861 }
862 
863 /// AnalyzeLoadFromClobberingLoad - This function is called when we have a
864 /// memdep query of a load that ends up being clobbered by another load.  See if
865 /// the other load can feed into the second load.
AnalyzeLoadFromClobberingLoad(Type * LoadTy,Value * LoadPtr,LoadInst * DepLI,const TargetData & TD)866 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
867                                          LoadInst *DepLI, const TargetData &TD){
868   // Cannot handle reading from store of first-class aggregate yet.
869   if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
870     return -1;
871 
872   Value *DepPtr = DepLI->getPointerOperand();
873   uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType());
874   int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD);
875   if (R != -1) return R;
876 
877   // If we have a load/load clobber an DepLI can be widened to cover this load,
878   // then we should widen it!
879   int64_t LoadOffs = 0;
880   const Value *LoadBase =
881     GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, TD);
882   unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
883 
884   unsigned Size = MemoryDependenceAnalysis::
885     getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD);
886   if (Size == 0) return -1;
887 
888   return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD);
889 }
890 
891 
892 
AnalyzeLoadFromClobberingMemInst(Type * LoadTy,Value * LoadPtr,MemIntrinsic * MI,const TargetData & TD)893 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
894                                             MemIntrinsic *MI,
895                                             const TargetData &TD) {
896   // If the mem operation is a non-constant size, we can't handle it.
897   ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
898   if (SizeCst == 0) return -1;
899   uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
900 
901   // If this is memset, we just need to see if the offset is valid in the size
902   // of the memset..
903   if (MI->getIntrinsicID() == Intrinsic::memset)
904     return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
905                                           MemSizeInBits, TD);
906 
907   // If we have a memcpy/memmove, the only case we can handle is if this is a
908   // copy from constant memory.  In that case, we can read directly from the
909   // constant memory.
910   MemTransferInst *MTI = cast<MemTransferInst>(MI);
911 
912   Constant *Src = dyn_cast<Constant>(MTI->getSource());
913   if (Src == 0) return -1;
914 
915   GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD));
916   if (GV == 0 || !GV->isConstant()) return -1;
917 
918   // See if the access is within the bounds of the transfer.
919   int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
920                                               MI->getDest(), MemSizeInBits, TD);
921   if (Offset == -1)
922     return Offset;
923 
924   // Otherwise, see if we can constant fold a load from the constant with the
925   // offset applied as appropriate.
926   Src = ConstantExpr::getBitCast(Src,
927                                  llvm::Type::getInt8PtrTy(Src->getContext()));
928   Constant *OffsetCst =
929     ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
930   Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
931   Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
932   if (ConstantFoldLoadFromConstPtr(Src, &TD))
933     return Offset;
934   return -1;
935 }
936 
937 
938 /// GetStoreValueForLoad - This function is called when we have a
939 /// memdep query of a load that ends up being a clobbering store.  This means
940 /// that the store provides bits used by the load but we the pointers don't
941 /// mustalias.  Check this case to see if there is anything more we can do
942 /// before we give up.
GetStoreValueForLoad(Value * SrcVal,unsigned Offset,Type * LoadTy,Instruction * InsertPt,const TargetData & TD)943 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
944                                    Type *LoadTy,
945                                    Instruction *InsertPt, const TargetData &TD){
946   LLVMContext &Ctx = SrcVal->getType()->getContext();
947 
948   uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
949   uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8;
950 
951   IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
952 
953   // Compute which bits of the stored value are being used by the load.  Convert
954   // to an integer type to start with.
955   if (SrcVal->getType()->isPointerTy())
956     SrcVal = Builder.CreatePtrToInt(SrcVal, TD.getIntPtrType(Ctx));
957   if (!SrcVal->getType()->isIntegerTy())
958     SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
959 
960   // Shift the bits to the least significant depending on endianness.
961   unsigned ShiftAmt;
962   if (TD.isLittleEndian())
963     ShiftAmt = Offset*8;
964   else
965     ShiftAmt = (StoreSize-LoadSize-Offset)*8;
966 
967   if (ShiftAmt)
968     SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
969 
970   if (LoadSize != StoreSize)
971     SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
972 
973   return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
974 }
975 
976 /// GetStoreValueForLoad - This function is called when we have a
977 /// memdep query of a load that ends up being a clobbering load.  This means
978 /// that the load *may* provide bits used by the load but we can't be sure
979 /// because the pointers don't mustalias.  Check this case to see if there is
980 /// anything more we can do before we give up.
GetLoadValueForLoad(LoadInst * SrcVal,unsigned Offset,Type * LoadTy,Instruction * InsertPt,GVN & gvn)981 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
982                                   Type *LoadTy, Instruction *InsertPt,
983                                   GVN &gvn) {
984   const TargetData &TD = *gvn.getTargetData();
985   // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
986   // widen SrcVal out to a larger load.
987   unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType());
988   unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
989   if (Offset+LoadSize > SrcValSize) {
990     assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
991     assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
992     // If we have a load/load clobber an DepLI can be widened to cover this
993     // load, then we should widen it to the next power of 2 size big enough!
994     unsigned NewLoadSize = Offset+LoadSize;
995     if (!isPowerOf2_32(NewLoadSize))
996       NewLoadSize = NextPowerOf2(NewLoadSize);
997 
998     Value *PtrVal = SrcVal->getPointerOperand();
999 
1000     // Insert the new load after the old load.  This ensures that subsequent
1001     // memdep queries will find the new load.  We can't easily remove the old
1002     // load completely because it is already in the value numbering table.
1003     IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1004     Type *DestPTy =
1005       IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1006     DestPTy = PointerType::get(DestPTy,
1007                        cast<PointerType>(PtrVal->getType())->getAddressSpace());
1008     Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1009     PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1010     LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1011     NewLoad->takeName(SrcVal);
1012     NewLoad->setAlignment(SrcVal->getAlignment());
1013 
1014     DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1015     DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1016 
1017     // Replace uses of the original load with the wider load.  On a big endian
1018     // system, we need to shift down to get the relevant bits.
1019     Value *RV = NewLoad;
1020     if (TD.isBigEndian())
1021       RV = Builder.CreateLShr(RV,
1022                     NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
1023     RV = Builder.CreateTrunc(RV, SrcVal->getType());
1024     SrcVal->replaceAllUsesWith(RV);
1025 
1026     // We would like to use gvn.markInstructionForDeletion here, but we can't
1027     // because the load is already memoized into the leader map table that GVN
1028     // tracks.  It is potentially possible to remove the load from the table,
1029     // but then there all of the operations based on it would need to be
1030     // rehashed.  Just leave the dead load around.
1031     gvn.getMemDep().removeInstruction(SrcVal);
1032     SrcVal = NewLoad;
1033   }
1034 
1035   return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD);
1036 }
1037 
1038 
1039 /// GetMemInstValueForLoad - This function is called when we have a
1040 /// memdep query of a load that ends up being a clobbering mem intrinsic.
GetMemInstValueForLoad(MemIntrinsic * SrcInst,unsigned Offset,Type * LoadTy,Instruction * InsertPt,const TargetData & TD)1041 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1042                                      Type *LoadTy, Instruction *InsertPt,
1043                                      const TargetData &TD){
1044   LLVMContext &Ctx = LoadTy->getContext();
1045   uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1046 
1047   IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1048 
1049   // We know that this method is only called when the mem transfer fully
1050   // provides the bits for the load.
1051   if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1052     // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1053     // independently of what the offset is.
1054     Value *Val = MSI->getValue();
1055     if (LoadSize != 1)
1056       Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1057 
1058     Value *OneElt = Val;
1059 
1060     // Splat the value out to the right number of bits.
1061     for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1062       // If we can double the number of bytes set, do it.
1063       if (NumBytesSet*2 <= LoadSize) {
1064         Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1065         Val = Builder.CreateOr(Val, ShVal);
1066         NumBytesSet <<= 1;
1067         continue;
1068       }
1069 
1070       // Otherwise insert one byte at a time.
1071       Value *ShVal = Builder.CreateShl(Val, 1*8);
1072       Val = Builder.CreateOr(OneElt, ShVal);
1073       ++NumBytesSet;
1074     }
1075 
1076     return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
1077   }
1078 
1079   // Otherwise, this is a memcpy/memmove from a constant global.
1080   MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1081   Constant *Src = cast<Constant>(MTI->getSource());
1082 
1083   // Otherwise, see if we can constant fold a load from the constant with the
1084   // offset applied as appropriate.
1085   Src = ConstantExpr::getBitCast(Src,
1086                                  llvm::Type::getInt8PtrTy(Src->getContext()));
1087   Constant *OffsetCst =
1088   ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1089   Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1090   Src = ConstantExpr::getBitCast(Src, PointerType::getUnqual(LoadTy));
1091   return ConstantFoldLoadFromConstPtr(Src, &TD);
1092 }
1093 
1094 namespace {
1095 
1096 struct AvailableValueInBlock {
1097   /// BB - The basic block in question.
1098   BasicBlock *BB;
1099   enum ValType {
1100     SimpleVal,  // A simple offsetted value that is accessed.
1101     LoadVal,    // A value produced by a load.
1102     MemIntrin   // A memory intrinsic which is loaded from.
1103   };
1104 
1105   /// V - The value that is live out of the block.
1106   PointerIntPair<Value *, 2, ValType> Val;
1107 
1108   /// Offset - The byte offset in Val that is interesting for the load query.
1109   unsigned Offset;
1110 
get__anonadea466b0311::AvailableValueInBlock1111   static AvailableValueInBlock get(BasicBlock *BB, Value *V,
1112                                    unsigned Offset = 0) {
1113     AvailableValueInBlock Res;
1114     Res.BB = BB;
1115     Res.Val.setPointer(V);
1116     Res.Val.setInt(SimpleVal);
1117     Res.Offset = Offset;
1118     return Res;
1119   }
1120 
getMI__anonadea466b0311::AvailableValueInBlock1121   static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
1122                                      unsigned Offset = 0) {
1123     AvailableValueInBlock Res;
1124     Res.BB = BB;
1125     Res.Val.setPointer(MI);
1126     Res.Val.setInt(MemIntrin);
1127     Res.Offset = Offset;
1128     return Res;
1129   }
1130 
getLoad__anonadea466b0311::AvailableValueInBlock1131   static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
1132                                        unsigned Offset = 0) {
1133     AvailableValueInBlock Res;
1134     Res.BB = BB;
1135     Res.Val.setPointer(LI);
1136     Res.Val.setInt(LoadVal);
1137     Res.Offset = Offset;
1138     return Res;
1139   }
1140 
isSimpleValue__anonadea466b0311::AvailableValueInBlock1141   bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
isCoercedLoadValue__anonadea466b0311::AvailableValueInBlock1142   bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
isMemIntrinValue__anonadea466b0311::AvailableValueInBlock1143   bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
1144 
getSimpleValue__anonadea466b0311::AvailableValueInBlock1145   Value *getSimpleValue() const {
1146     assert(isSimpleValue() && "Wrong accessor");
1147     return Val.getPointer();
1148   }
1149 
getCoercedLoadValue__anonadea466b0311::AvailableValueInBlock1150   LoadInst *getCoercedLoadValue() const {
1151     assert(isCoercedLoadValue() && "Wrong accessor");
1152     return cast<LoadInst>(Val.getPointer());
1153   }
1154 
getMemIntrinValue__anonadea466b0311::AvailableValueInBlock1155   MemIntrinsic *getMemIntrinValue() const {
1156     assert(isMemIntrinValue() && "Wrong accessor");
1157     return cast<MemIntrinsic>(Val.getPointer());
1158   }
1159 
1160   /// MaterializeAdjustedValue - Emit code into this block to adjust the value
1161   /// defined here to the specified type.  This handles various coercion cases.
MaterializeAdjustedValue__anonadea466b0311::AvailableValueInBlock1162   Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const {
1163     Value *Res;
1164     if (isSimpleValue()) {
1165       Res = getSimpleValue();
1166       if (Res->getType() != LoadTy) {
1167         const TargetData *TD = gvn.getTargetData();
1168         assert(TD && "Need target data to handle type mismatch case");
1169         Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1170                                    *TD);
1171 
1172         DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << "  "
1173                      << *getSimpleValue() << '\n'
1174                      << *Res << '\n' << "\n\n\n");
1175       }
1176     } else if (isCoercedLoadValue()) {
1177       LoadInst *Load = getCoercedLoadValue();
1178       if (Load->getType() == LoadTy && Offset == 0) {
1179         Res = Load;
1180       } else {
1181         Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1182                                   gvn);
1183 
1184         DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << "  "
1185                      << *getCoercedLoadValue() << '\n'
1186                      << *Res << '\n' << "\n\n\n");
1187       }
1188     } else {
1189       const TargetData *TD = gvn.getTargetData();
1190       assert(TD && "Need target data to handle type mismatch case");
1191       Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1192                                    LoadTy, BB->getTerminator(), *TD);
1193       DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1194                    << "  " << *getMemIntrinValue() << '\n'
1195                    << *Res << '\n' << "\n\n\n");
1196     }
1197     return Res;
1198   }
1199 };
1200 
1201 } // end anonymous namespace
1202 
1203 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1204 /// construct SSA form, allowing us to eliminate LI.  This returns the value
1205 /// that should be used at LI's definition site.
ConstructSSAForLoadSet(LoadInst * LI,SmallVectorImpl<AvailableValueInBlock> & ValuesPerBlock,GVN & gvn)1206 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1207                          SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1208                                      GVN &gvn) {
1209   // Check for the fully redundant, dominating load case.  In this case, we can
1210   // just use the dominating value directly.
1211   if (ValuesPerBlock.size() == 1 &&
1212       gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1213                                                LI->getParent()))
1214     return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn);
1215 
1216   // Otherwise, we have to construct SSA form.
1217   SmallVector<PHINode*, 8> NewPHIs;
1218   SSAUpdater SSAUpdate(&NewPHIs);
1219   SSAUpdate.Initialize(LI->getType(), LI->getName());
1220 
1221   Type *LoadTy = LI->getType();
1222 
1223   for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1224     const AvailableValueInBlock &AV = ValuesPerBlock[i];
1225     BasicBlock *BB = AV.BB;
1226 
1227     if (SSAUpdate.HasValueForBlock(BB))
1228       continue;
1229 
1230     SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn));
1231   }
1232 
1233   // Perform PHI construction.
1234   Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1235 
1236   // If new PHI nodes were created, notify alias analysis.
1237   if (V->getType()->isPointerTy()) {
1238     AliasAnalysis *AA = gvn.getAliasAnalysis();
1239 
1240     for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1241       AA->copyValue(LI, NewPHIs[i]);
1242 
1243     // Now that we've copied information to the new PHIs, scan through
1244     // them again and inform alias analysis that we've added potentially
1245     // escaping uses to any values that are operands to these PHIs.
1246     for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
1247       PHINode *P = NewPHIs[i];
1248       for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) {
1249         unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
1250         AA->addEscapingUse(P->getOperandUse(jj));
1251       }
1252     }
1253   }
1254 
1255   return V;
1256 }
1257 
isLifetimeStart(const Instruction * Inst)1258 static bool isLifetimeStart(const Instruction *Inst) {
1259   if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1260     return II->getIntrinsicID() == Intrinsic::lifetime_start;
1261   return false;
1262 }
1263 
1264 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1265 /// non-local by performing PHI construction.
processNonLocalLoad(LoadInst * LI)1266 bool GVN::processNonLocalLoad(LoadInst *LI) {
1267   // Find the non-local dependencies of the load.
1268   SmallVector<NonLocalDepResult, 64> Deps;
1269   AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
1270   MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
1271   //DEBUG(dbgs() << "INVESTIGATING NONLOCAL LOAD: "
1272   //             << Deps.size() << *LI << '\n');
1273 
1274   // If we had to process more than one hundred blocks to find the
1275   // dependencies, this load isn't worth worrying about.  Optimizing
1276   // it will be too expensive.
1277   if (Deps.size() > 100)
1278     return false;
1279 
1280   // If we had a phi translation failure, we'll have a single entry which is a
1281   // clobber in the current block.  Reject this early.
1282   if (Deps.size() == 1
1283       && !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber())
1284   {
1285     DEBUG(
1286       dbgs() << "GVN: non-local load ";
1287       WriteAsOperand(dbgs(), LI);
1288       dbgs() << " has unknown dependencies\n";
1289     );
1290     return false;
1291   }
1292 
1293   // Filter out useless results (non-locals, etc).  Keep track of the blocks
1294   // where we have a value available in repl, also keep track of whether we see
1295   // dependencies that produce an unknown value for the load (such as a call
1296   // that could potentially clobber the load).
1297   SmallVector<AvailableValueInBlock, 16> ValuesPerBlock;
1298   SmallVector<BasicBlock*, 16> UnavailableBlocks;
1299 
1300   for (unsigned i = 0, e = Deps.size(); i != e; ++i) {
1301     BasicBlock *DepBB = Deps[i].getBB();
1302     MemDepResult DepInfo = Deps[i].getResult();
1303 
1304     if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1305       UnavailableBlocks.push_back(DepBB);
1306       continue;
1307     }
1308 
1309     if (DepInfo.isClobber()) {
1310       // The address being loaded in this non-local block may not be the same as
1311       // the pointer operand of the load if PHI translation occurs.  Make sure
1312       // to consider the right address.
1313       Value *Address = Deps[i].getAddress();
1314 
1315       // If the dependence is to a store that writes to a superset of the bits
1316       // read by the load, we can extract the bits we need for the load from the
1317       // stored value.
1318       if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1319         if (TD && Address) {
1320           int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1321                                                       DepSI, *TD);
1322           if (Offset != -1) {
1323             ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1324                                                        DepSI->getValueOperand(),
1325                                                                 Offset));
1326             continue;
1327           }
1328         }
1329       }
1330 
1331       // Check to see if we have something like this:
1332       //    load i32* P
1333       //    load i8* (P+1)
1334       // if we have this, replace the later with an extraction from the former.
1335       if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1336         // If this is a clobber and L is the first instruction in its block, then
1337         // we have the first instruction in the entry block.
1338         if (DepLI != LI && Address && TD) {
1339           int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(),
1340                                                      LI->getPointerOperand(),
1341                                                      DepLI, *TD);
1342 
1343           if (Offset != -1) {
1344             ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1345                                                                     Offset));
1346             continue;
1347           }
1348         }
1349       }
1350 
1351       // If the clobbering value is a memset/memcpy/memmove, see if we can
1352       // forward a value on from it.
1353       if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1354         if (TD && Address) {
1355           int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1356                                                         DepMI, *TD);
1357           if (Offset != -1) {
1358             ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1359                                                                   Offset));
1360             continue;
1361           }
1362         }
1363       }
1364 
1365       UnavailableBlocks.push_back(DepBB);
1366       continue;
1367     }
1368 
1369     // DepInfo.isDef() here
1370 
1371     Instruction *DepInst = DepInfo.getInst();
1372 
1373     // Loading the allocation -> undef.
1374     if (isa<AllocaInst>(DepInst) || isMalloc(DepInst) ||
1375         // Loading immediately after lifetime begin -> undef.
1376         isLifetimeStart(DepInst)) {
1377       ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1378                                              UndefValue::get(LI->getType())));
1379       continue;
1380     }
1381 
1382     if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1383       // Reject loads and stores that are to the same address but are of
1384       // different types if we have to.
1385       if (S->getValueOperand()->getType() != LI->getType()) {
1386         // If the stored value is larger or equal to the loaded value, we can
1387         // reuse it.
1388         if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1389                                                         LI->getType(), *TD)) {
1390           UnavailableBlocks.push_back(DepBB);
1391           continue;
1392         }
1393       }
1394 
1395       ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1396                                                          S->getValueOperand()));
1397       continue;
1398     }
1399 
1400     if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1401       // If the types mismatch and we can't handle it, reject reuse of the load.
1402       if (LD->getType() != LI->getType()) {
1403         // If the stored value is larger or equal to the loaded value, we can
1404         // reuse it.
1405         if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1406           UnavailableBlocks.push_back(DepBB);
1407           continue;
1408         }
1409       }
1410       ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1411       continue;
1412     }
1413 
1414     UnavailableBlocks.push_back(DepBB);
1415     continue;
1416   }
1417 
1418   // If we have no predecessors that produce a known value for this load, exit
1419   // early.
1420   if (ValuesPerBlock.empty()) return false;
1421 
1422   // If all of the instructions we depend on produce a known value for this
1423   // load, then it is fully redundant and we can use PHI insertion to compute
1424   // its value.  Insert PHIs and remove the fully redundant value now.
1425   if (UnavailableBlocks.empty()) {
1426     DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1427 
1428     // Perform PHI construction.
1429     Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1430     LI->replaceAllUsesWith(V);
1431 
1432     if (isa<PHINode>(V))
1433       V->takeName(LI);
1434     if (V->getType()->isPointerTy())
1435       MD->invalidateCachedPointerInfo(V);
1436     markInstructionForDeletion(LI);
1437     ++NumGVNLoad;
1438     return true;
1439   }
1440 
1441   if (!EnablePRE || !EnableLoadPRE)
1442     return false;
1443 
1444   // Okay, we have *some* definitions of the value.  This means that the value
1445   // is available in some of our (transitive) predecessors.  Lets think about
1446   // doing PRE of this load.  This will involve inserting a new load into the
1447   // predecessor when it's not available.  We could do this in general, but
1448   // prefer to not increase code size.  As such, we only do this when we know
1449   // that we only have to insert *one* load (which means we're basically moving
1450   // the load, not inserting a new one).
1451 
1452   SmallPtrSet<BasicBlock *, 4> Blockers;
1453   for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1454     Blockers.insert(UnavailableBlocks[i]);
1455 
1456   // Let's find the first basic block with more than one predecessor.  Walk
1457   // backwards through predecessors if needed.
1458   BasicBlock *LoadBB = LI->getParent();
1459   BasicBlock *TmpBB = LoadBB;
1460 
1461   bool isSinglePred = false;
1462   bool allSingleSucc = true;
1463   while (TmpBB->getSinglePredecessor()) {
1464     isSinglePred = true;
1465     TmpBB = TmpBB->getSinglePredecessor();
1466     if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1467       return false;
1468     if (Blockers.count(TmpBB))
1469       return false;
1470 
1471     // If any of these blocks has more than one successor (i.e. if the edge we
1472     // just traversed was critical), then there are other paths through this
1473     // block along which the load may not be anticipated.  Hoisting the load
1474     // above this block would be adding the load to execution paths along
1475     // which it was not previously executed.
1476     if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1477       return false;
1478   }
1479 
1480   assert(TmpBB);
1481   LoadBB = TmpBB;
1482 
1483   // FIXME: It is extremely unclear what this loop is doing, other than
1484   // artificially restricting loadpre.
1485   if (isSinglePred) {
1486     bool isHot = false;
1487     for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1488       const AvailableValueInBlock &AV = ValuesPerBlock[i];
1489       if (AV.isSimpleValue())
1490         // "Hot" Instruction is in some loop (because it dominates its dep.
1491         // instruction).
1492         if (Instruction *I = dyn_cast<Instruction>(AV.getSimpleValue()))
1493           if (DT->dominates(LI, I)) {
1494             isHot = true;
1495             break;
1496           }
1497     }
1498 
1499     // We are interested only in "hot" instructions. We don't want to do any
1500     // mis-optimizations here.
1501     if (!isHot)
1502       return false;
1503   }
1504 
1505   // Check to see how many predecessors have the loaded value fully
1506   // available.
1507   DenseMap<BasicBlock*, Value*> PredLoads;
1508   DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1509   for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1510     FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1511   for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1512     FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1513 
1514   SmallVector<std::pair<TerminatorInst*, unsigned>, 4> NeedToSplit;
1515   for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1516        PI != E; ++PI) {
1517     BasicBlock *Pred = *PI;
1518     if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks)) {
1519       continue;
1520     }
1521     PredLoads[Pred] = 0;
1522 
1523     if (Pred->getTerminator()->getNumSuccessors() != 1) {
1524       if (isa<IndirectBrInst>(Pred->getTerminator())) {
1525         DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1526               << Pred->getName() << "': " << *LI << '\n');
1527         return false;
1528       }
1529 
1530       if (LoadBB->isLandingPad()) {
1531         DEBUG(dbgs()
1532               << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '"
1533               << Pred->getName() << "': " << *LI << '\n');
1534         return false;
1535       }
1536 
1537       unsigned SuccNum = GetSuccessorNumber(Pred, LoadBB);
1538       NeedToSplit.push_back(std::make_pair(Pred->getTerminator(), SuccNum));
1539     }
1540   }
1541 
1542   if (!NeedToSplit.empty()) {
1543     toSplit.append(NeedToSplit.begin(), NeedToSplit.end());
1544     return false;
1545   }
1546 
1547   // Decide whether PRE is profitable for this load.
1548   unsigned NumUnavailablePreds = PredLoads.size();
1549   assert(NumUnavailablePreds != 0 &&
1550          "Fully available value should be eliminated above!");
1551 
1552   // If this load is unavailable in multiple predecessors, reject it.
1553   // FIXME: If we could restructure the CFG, we could make a common pred with
1554   // all the preds that don't have an available LI and insert a new load into
1555   // that one block.
1556   if (NumUnavailablePreds != 1)
1557       return false;
1558 
1559   // Check if the load can safely be moved to all the unavailable predecessors.
1560   bool CanDoPRE = true;
1561   SmallVector<Instruction*, 8> NewInsts;
1562   for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1563          E = PredLoads.end(); I != E; ++I) {
1564     BasicBlock *UnavailablePred = I->first;
1565 
1566     // Do PHI translation to get its value in the predecessor if necessary.  The
1567     // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1568 
1569     // If all preds have a single successor, then we know it is safe to insert
1570     // the load on the pred (?!?), so we can insert code to materialize the
1571     // pointer if it is not available.
1572     PHITransAddr Address(LI->getPointerOperand(), TD);
1573     Value *LoadPtr = 0;
1574     if (allSingleSucc) {
1575       LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1576                                                   *DT, NewInsts);
1577     } else {
1578       Address.PHITranslateValue(LoadBB, UnavailablePred, DT);
1579       LoadPtr = Address.getAddr();
1580     }
1581 
1582     // If we couldn't find or insert a computation of this phi translated value,
1583     // we fail PRE.
1584     if (LoadPtr == 0) {
1585       DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1586             << *LI->getPointerOperand() << "\n");
1587       CanDoPRE = false;
1588       break;
1589     }
1590 
1591     // Make sure it is valid to move this load here.  We have to watch out for:
1592     //  @1 = getelementptr (i8* p, ...
1593     //  test p and branch if == 0
1594     //  load @1
1595     // It is valid to have the getelementptr before the test, even if p can
1596     // be 0, as getelementptr only does address arithmetic.
1597     // If we are not pushing the value through any multiple-successor blocks
1598     // we do not have this case.  Otherwise, check that the load is safe to
1599     // put anywhere; this can be improved, but should be conservatively safe.
1600     if (!allSingleSucc &&
1601         // FIXME: REEVALUTE THIS.
1602         !isSafeToLoadUnconditionally(LoadPtr,
1603                                      UnavailablePred->getTerminator(),
1604                                      LI->getAlignment(), TD)) {
1605       CanDoPRE = false;
1606       break;
1607     }
1608 
1609     I->second = LoadPtr;
1610   }
1611 
1612   if (!CanDoPRE) {
1613     while (!NewInsts.empty()) {
1614       Instruction *I = NewInsts.pop_back_val();
1615       if (MD) MD->removeInstruction(I);
1616       I->eraseFromParent();
1617     }
1618     return false;
1619   }
1620 
1621   // Okay, we can eliminate this load by inserting a reload in the predecessor
1622   // and using PHI construction to get the value in the other predecessors, do
1623   // it.
1624   DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1625   DEBUG(if (!NewInsts.empty())
1626           dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1627                  << *NewInsts.back() << '\n');
1628 
1629   // Assign value numbers to the new instructions.
1630   for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1631     // FIXME: We really _ought_ to insert these value numbers into their
1632     // parent's availability map.  However, in doing so, we risk getting into
1633     // ordering issues.  If a block hasn't been processed yet, we would be
1634     // marking a value as AVAIL-IN, which isn't what we intend.
1635     VN.lookup_or_add(NewInsts[i]);
1636   }
1637 
1638   for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1639          E = PredLoads.end(); I != E; ++I) {
1640     BasicBlock *UnavailablePred = I->first;
1641     Value *LoadPtr = I->second;
1642 
1643     Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1644                                         LI->getAlignment(),
1645                                         UnavailablePred->getTerminator());
1646 
1647     // Transfer the old load's TBAA tag to the new load.
1648     if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa))
1649       NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1650 
1651     // Transfer DebugLoc.
1652     NewLoad->setDebugLoc(LI->getDebugLoc());
1653 
1654     // Add the newly created load.
1655     ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1656                                                         NewLoad));
1657     MD->invalidateCachedPointerInfo(LoadPtr);
1658     DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1659   }
1660 
1661   // Perform PHI construction.
1662   Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1663   LI->replaceAllUsesWith(V);
1664   if (isa<PHINode>(V))
1665     V->takeName(LI);
1666   if (V->getType()->isPointerTy())
1667     MD->invalidateCachedPointerInfo(V);
1668   markInstructionForDeletion(LI);
1669   ++NumPRELoad;
1670   return true;
1671 }
1672 
1673 /// processLoad - Attempt to eliminate a load, first by eliminating it
1674 /// locally, and then attempting non-local elimination if that fails.
processLoad(LoadInst * L)1675 bool GVN::processLoad(LoadInst *L) {
1676   if (!MD)
1677     return false;
1678 
1679   if (!L->isSimple())
1680     return false;
1681 
1682   if (L->use_empty()) {
1683     markInstructionForDeletion(L);
1684     return true;
1685   }
1686 
1687   // ... to a pointer that has been loaded from before...
1688   MemDepResult Dep = MD->getDependency(L);
1689 
1690   // If we have a clobber and target data is around, see if this is a clobber
1691   // that we can fix up through code synthesis.
1692   if (Dep.isClobber() && TD) {
1693     // Check to see if we have something like this:
1694     //   store i32 123, i32* %P
1695     //   %A = bitcast i32* %P to i8*
1696     //   %B = gep i8* %A, i32 1
1697     //   %C = load i8* %B
1698     //
1699     // We could do that by recognizing if the clobber instructions are obviously
1700     // a common base + constant offset, and if the previous store (or memset)
1701     // completely covers this load.  This sort of thing can happen in bitfield
1702     // access code.
1703     Value *AvailVal = 0;
1704     if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1705       int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1706                                                   L->getPointerOperand(),
1707                                                   DepSI, *TD);
1708       if (Offset != -1)
1709         AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1710                                         L->getType(), L, *TD);
1711     }
1712 
1713     // Check to see if we have something like this:
1714     //    load i32* P
1715     //    load i8* (P+1)
1716     // if we have this, replace the later with an extraction from the former.
1717     if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1718       // If this is a clobber and L is the first instruction in its block, then
1719       // we have the first instruction in the entry block.
1720       if (DepLI == L)
1721         return false;
1722 
1723       int Offset = AnalyzeLoadFromClobberingLoad(L->getType(),
1724                                                  L->getPointerOperand(),
1725                                                  DepLI, *TD);
1726       if (Offset != -1)
1727         AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1728     }
1729 
1730     // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1731     // a value on from it.
1732     if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1733       int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1734                                                     L->getPointerOperand(),
1735                                                     DepMI, *TD);
1736       if (Offset != -1)
1737         AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD);
1738     }
1739 
1740     if (AvailVal) {
1741       DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1742             << *AvailVal << '\n' << *L << "\n\n\n");
1743 
1744       // Replace the load!
1745       L->replaceAllUsesWith(AvailVal);
1746       if (AvailVal->getType()->isPointerTy())
1747         MD->invalidateCachedPointerInfo(AvailVal);
1748       markInstructionForDeletion(L);
1749       ++NumGVNLoad;
1750       return true;
1751     }
1752   }
1753 
1754   // If the value isn't available, don't do anything!
1755   if (Dep.isClobber()) {
1756     DEBUG(
1757       // fast print dep, using operator<< on instruction is too slow.
1758       dbgs() << "GVN: load ";
1759       WriteAsOperand(dbgs(), L);
1760       Instruction *I = Dep.getInst();
1761       dbgs() << " is clobbered by " << *I << '\n';
1762     );
1763     return false;
1764   }
1765 
1766   // If it is defined in another block, try harder.
1767   if (Dep.isNonLocal())
1768     return processNonLocalLoad(L);
1769 
1770   if (!Dep.isDef()) {
1771     DEBUG(
1772       // fast print dep, using operator<< on instruction is too slow.
1773       dbgs() << "GVN: load ";
1774       WriteAsOperand(dbgs(), L);
1775       dbgs() << " has unknown dependence\n";
1776     );
1777     return false;
1778   }
1779 
1780   Instruction *DepInst = Dep.getInst();
1781   if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1782     Value *StoredVal = DepSI->getValueOperand();
1783 
1784     // The store and load are to a must-aliased pointer, but they may not
1785     // actually have the same type.  See if we know how to reuse the stored
1786     // value (depending on its type).
1787     if (StoredVal->getType() != L->getType()) {
1788       if (TD) {
1789         StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1790                                                    L, *TD);
1791         if (StoredVal == 0)
1792           return false;
1793 
1794         DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1795                      << '\n' << *L << "\n\n\n");
1796       }
1797       else
1798         return false;
1799     }
1800 
1801     // Remove it!
1802     L->replaceAllUsesWith(StoredVal);
1803     if (StoredVal->getType()->isPointerTy())
1804       MD->invalidateCachedPointerInfo(StoredVal);
1805     markInstructionForDeletion(L);
1806     ++NumGVNLoad;
1807     return true;
1808   }
1809 
1810   if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1811     Value *AvailableVal = DepLI;
1812 
1813     // The loads are of a must-aliased pointer, but they may not actually have
1814     // the same type.  See if we know how to reuse the previously loaded value
1815     // (depending on its type).
1816     if (DepLI->getType() != L->getType()) {
1817       if (TD) {
1818         AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(),
1819                                                       L, *TD);
1820         if (AvailableVal == 0)
1821           return false;
1822 
1823         DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1824                      << "\n" << *L << "\n\n\n");
1825       }
1826       else
1827         return false;
1828     }
1829 
1830     // Remove it!
1831     L->replaceAllUsesWith(AvailableVal);
1832     if (DepLI->getType()->isPointerTy())
1833       MD->invalidateCachedPointerInfo(DepLI);
1834     markInstructionForDeletion(L);
1835     ++NumGVNLoad;
1836     return true;
1837   }
1838 
1839   // If this load really doesn't depend on anything, then we must be loading an
1840   // undef value.  This can happen when loading for a fresh allocation with no
1841   // intervening stores, for example.
1842   if (isa<AllocaInst>(DepInst) || isMalloc(DepInst)) {
1843     L->replaceAllUsesWith(UndefValue::get(L->getType()));
1844     markInstructionForDeletion(L);
1845     ++NumGVNLoad;
1846     return true;
1847   }
1848 
1849   // If this load occurs either right after a lifetime begin,
1850   // then the loaded value is undefined.
1851   if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
1852     if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1853       L->replaceAllUsesWith(UndefValue::get(L->getType()));
1854       markInstructionForDeletion(L);
1855       ++NumGVNLoad;
1856       return true;
1857     }
1858   }
1859 
1860   return false;
1861 }
1862 
1863 // findLeader - In order to find a leader for a given value number at a
1864 // specific basic block, we first obtain the list of all Values for that number,
1865 // and then scan the list to find one whose block dominates the block in
1866 // question.  This is fast because dominator tree queries consist of only
1867 // a few comparisons of DFS numbers.
findLeader(BasicBlock * BB,uint32_t num)1868 Value *GVN::findLeader(BasicBlock *BB, uint32_t num) {
1869   LeaderTableEntry Vals = LeaderTable[num];
1870   if (!Vals.Val) return 0;
1871 
1872   Value *Val = 0;
1873   if (DT->dominates(Vals.BB, BB)) {
1874     Val = Vals.Val;
1875     if (isa<Constant>(Val)) return Val;
1876   }
1877 
1878   LeaderTableEntry* Next = Vals.Next;
1879   while (Next) {
1880     if (DT->dominates(Next->BB, BB)) {
1881       if (isa<Constant>(Next->Val)) return Next->Val;
1882       if (!Val) Val = Next->Val;
1883     }
1884 
1885     Next = Next->Next;
1886   }
1887 
1888   return Val;
1889 }
1890 
1891 /// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the
1892 /// use is dominated by the given basic block.  Returns the number of uses that
1893 /// were replaced.
replaceAllDominatedUsesWith(Value * From,Value * To,BasicBlock * Root)1894 unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To,
1895                                           BasicBlock *Root) {
1896   unsigned Count = 0;
1897   for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
1898        UI != UE; ) {
1899     Instruction *User = cast<Instruction>(*UI);
1900     unsigned OpNum = UI.getOperandNo();
1901     ++UI;
1902 
1903     if (DT->dominates(Root, User->getParent())) {
1904       User->setOperand(OpNum, To);
1905       ++Count;
1906     }
1907   }
1908   return Count;
1909 }
1910 
1911 /// propagateEquality - The given values are known to be equal in every block
1912 /// dominated by 'Root'.  Exploit this, for example by replacing 'LHS' with
1913 /// 'RHS' everywhere in the scope.  Returns whether a change was made.
propagateEquality(Value * LHS,Value * RHS,BasicBlock * Root)1914 bool GVN::propagateEquality(Value *LHS, Value *RHS, BasicBlock *Root) {
1915   if (LHS == RHS) return false;
1916   assert(LHS->getType() == RHS->getType() && "Equal but types differ!");
1917 
1918   // Don't try to propagate equalities between constants.
1919   if (isa<Constant>(LHS) && isa<Constant>(RHS))
1920     return false;
1921 
1922   // Make sure that any constants are on the right-hand side.  In general the
1923   // best results are obtained by placing the longest lived value on the RHS.
1924   if (isa<Constant>(LHS))
1925     std::swap(LHS, RHS);
1926 
1927   // If neither term is constant then bail out.  This is not for correctness,
1928   // it's just that the non-constant case is much less useful: it occurs just
1929   // as often as the constant case but handling it hardly ever results in an
1930   // improvement.
1931   if (!isa<Constant>(RHS))
1932     return false;
1933 
1934   // If value numbering later deduces that an instruction in the scope is equal
1935   // to 'LHS' then ensure it will be turned into 'RHS'.
1936   addToLeaderTable(VN.lookup_or_add(LHS), RHS, Root);
1937 
1938   // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope.
1939   unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root);
1940   bool Changed = NumReplacements > 0;
1941   NumGVNEqProp += NumReplacements;
1942 
1943   // Now try to deduce additional equalities from this one.  For example, if the
1944   // known equality was "(A != B)" == "false" then it follows that A and B are
1945   // equal in the scope.  Only boolean equalities with an explicit true or false
1946   // RHS are currently supported.
1947   if (!RHS->getType()->isIntegerTy(1))
1948     // Not a boolean equality - bail out.
1949     return Changed;
1950   ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
1951   if (!CI)
1952     // RHS neither 'true' nor 'false' - bail out.
1953     return Changed;
1954   // Whether RHS equals 'true'.  Otherwise it equals 'false'.
1955   bool isKnownTrue = CI->isAllOnesValue();
1956   bool isKnownFalse = !isKnownTrue;
1957 
1958   // If "A && B" is known true then both A and B are known true.  If "A || B"
1959   // is known false then both A and B are known false.
1960   Value *A, *B;
1961   if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
1962       (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
1963     Changed |= propagateEquality(A, RHS, Root);
1964     Changed |= propagateEquality(B, RHS, Root);
1965     return Changed;
1966   }
1967 
1968   // If we are propagating an equality like "(A == B)" == "true" then also
1969   // propagate the equality A == B.
1970   if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) {
1971     // Only equality comparisons are supported.
1972     if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
1973         (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE)) {
1974       Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
1975       Changed |= propagateEquality(Op0, Op1, Root);
1976     }
1977     return Changed;
1978   }
1979 
1980   return Changed;
1981 }
1982 
1983 /// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'.  Return
1984 /// true if every path from the entry block to 'Dst' passes via this edge.  In
1985 /// particular 'Dst' must not be reachable via another edge from 'Src'.
isOnlyReachableViaThisEdge(BasicBlock * Src,BasicBlock * Dst,DominatorTree * DT)1986 static bool isOnlyReachableViaThisEdge(BasicBlock *Src, BasicBlock *Dst,
1987                                        DominatorTree *DT) {
1988   // First off, there must not be more than one edge from Src to Dst, there
1989   // should be exactly one.  So keep track of the number of times Src occurs
1990   // as a predecessor of Dst and fail if it's more than once.  Secondly, any
1991   // other predecessors of Dst should be dominated by Dst (see logic below).
1992   bool SawEdgeFromSrc = false;
1993   for (pred_iterator PI = pred_begin(Dst), PE = pred_end(Dst); PI != PE; ++PI) {
1994     BasicBlock *Pred = *PI;
1995     if (Pred == Src) {
1996       // An edge from Src to Dst.
1997       if (SawEdgeFromSrc)
1998         // There are multiple edges from Src to Dst - fail.
1999         return false;
2000       SawEdgeFromSrc = true;
2001       continue;
2002     }
2003     // If the predecessor is not dominated by Dst, then it must be possible to
2004     // reach it either without passing through Src (and thus not via the edge)
2005     // or by passing through Src but taking a different edge out of Src.  Either
2006     // way it is possible to reach Dst without passing via the edge, so fail.
2007     if (!DT->dominates(Dst, *PI))
2008       return false;
2009   }
2010   assert(SawEdgeFromSrc && "No edge between these basic blocks!");
2011 
2012   // Every path from the entry block to Dst must at some point pass to Dst from
2013   // a predecessor that is not dominated by Dst.  This predecessor can only be
2014   // Src, since all others are dominated by Dst.  As there is only one edge from
2015   // Src to Dst, the path passes by this edge.
2016   return true;
2017 }
2018 
2019 /// processInstruction - When calculating availability, handle an instruction
2020 /// by inserting it into the appropriate sets
processInstruction(Instruction * I)2021 bool GVN::processInstruction(Instruction *I) {
2022   // Ignore dbg info intrinsics.
2023   if (isa<DbgInfoIntrinsic>(I))
2024     return false;
2025 
2026   // If the instruction can be easily simplified then do so now in preference
2027   // to value numbering it.  Value numbering often exposes redundancies, for
2028   // example if it determines that %y is equal to %x then the instruction
2029   // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2030   if (Value *V = SimplifyInstruction(I, TD, DT)) {
2031     I->replaceAllUsesWith(V);
2032     if (MD && V->getType()->isPointerTy())
2033       MD->invalidateCachedPointerInfo(V);
2034     markInstructionForDeletion(I);
2035     ++NumGVNSimpl;
2036     return true;
2037   }
2038 
2039   if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2040     if (processLoad(LI))
2041       return true;
2042 
2043     unsigned Num = VN.lookup_or_add(LI);
2044     addToLeaderTable(Num, LI, LI->getParent());
2045     return false;
2046   }
2047 
2048   // For conditional branches, we can perform simple conditional propagation on
2049   // the condition value itself.
2050   if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2051     if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
2052       return false;
2053 
2054     Value *BranchCond = BI->getCondition();
2055 
2056     BasicBlock *TrueSucc = BI->getSuccessor(0);
2057     BasicBlock *FalseSucc = BI->getSuccessor(1);
2058     BasicBlock *Parent = BI->getParent();
2059     bool Changed = false;
2060 
2061     if (isOnlyReachableViaThisEdge(Parent, TrueSucc, DT))
2062       Changed |= propagateEquality(BranchCond,
2063                                    ConstantInt::getTrue(TrueSucc->getContext()),
2064                                    TrueSucc);
2065 
2066     if (isOnlyReachableViaThisEdge(Parent, FalseSucc, DT))
2067       Changed |= propagateEquality(BranchCond,
2068                                    ConstantInt::getFalse(FalseSucc->getContext()),
2069                                    FalseSucc);
2070 
2071     return Changed;
2072   }
2073 
2074   // For switches, propagate the case values into the case destinations.
2075   if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2076     Value *SwitchCond = SI->getCondition();
2077     BasicBlock *Parent = SI->getParent();
2078     bool Changed = false;
2079     for (unsigned i = 1, e = SI->getNumCases(); i != e; ++i) {
2080       BasicBlock *Dst = SI->getSuccessor(i);
2081       if (isOnlyReachableViaThisEdge(Parent, Dst, DT))
2082         Changed |= propagateEquality(SwitchCond, SI->getCaseValue(i), Dst);
2083     }
2084     return Changed;
2085   }
2086 
2087   // Instructions with void type don't return a value, so there's
2088   // no point in trying to find redudancies in them.
2089   if (I->getType()->isVoidTy()) return false;
2090 
2091   uint32_t NextNum = VN.getNextUnusedValueNumber();
2092   unsigned Num = VN.lookup_or_add(I);
2093 
2094   // Allocations are always uniquely numbered, so we can save time and memory
2095   // by fast failing them.
2096   if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2097     addToLeaderTable(Num, I, I->getParent());
2098     return false;
2099   }
2100 
2101   // If the number we were assigned was a brand new VN, then we don't
2102   // need to do a lookup to see if the number already exists
2103   // somewhere in the domtree: it can't!
2104   if (Num == NextNum) {
2105     addToLeaderTable(Num, I, I->getParent());
2106     return false;
2107   }
2108 
2109   // Perform fast-path value-number based elimination of values inherited from
2110   // dominators.
2111   Value *repl = findLeader(I->getParent(), Num);
2112   if (repl == 0) {
2113     // Failure, just remember this instance for future use.
2114     addToLeaderTable(Num, I, I->getParent());
2115     return false;
2116   }
2117 
2118   // Remove it!
2119   I->replaceAllUsesWith(repl);
2120   if (MD && repl->getType()->isPointerTy())
2121     MD->invalidateCachedPointerInfo(repl);
2122   markInstructionForDeletion(I);
2123   return true;
2124 }
2125 
2126 /// runOnFunction - This is the main transformation entry point for a function.
runOnFunction(Function & F)2127 bool GVN::runOnFunction(Function& F) {
2128   if (!NoLoads)
2129     MD = &getAnalysis<MemoryDependenceAnalysis>();
2130   DT = &getAnalysis<DominatorTree>();
2131   TD = getAnalysisIfAvailable<TargetData>();
2132   VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2133   VN.setMemDep(MD);
2134   VN.setDomTree(DT);
2135 
2136   bool Changed = false;
2137   bool ShouldContinue = true;
2138 
2139   // Merge unconditional branches, allowing PRE to catch more
2140   // optimization opportunities.
2141   for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2142     BasicBlock *BB = FI++;
2143 
2144     bool removedBlock = MergeBlockIntoPredecessor(BB, this);
2145     if (removedBlock) ++NumGVNBlocks;
2146 
2147     Changed |= removedBlock;
2148   }
2149 
2150   unsigned Iteration = 0;
2151   while (ShouldContinue) {
2152     DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2153     ShouldContinue = iterateOnFunction(F);
2154     if (splitCriticalEdges())
2155       ShouldContinue = true;
2156     Changed |= ShouldContinue;
2157     ++Iteration;
2158   }
2159 
2160   if (EnablePRE) {
2161     bool PREChanged = true;
2162     while (PREChanged) {
2163       PREChanged = performPRE(F);
2164       Changed |= PREChanged;
2165     }
2166   }
2167   // FIXME: Should perform GVN again after PRE does something.  PRE can move
2168   // computations into blocks where they become fully redundant.  Note that
2169   // we can't do this until PRE's critical edge splitting updates memdep.
2170   // Actually, when this happens, we should just fully integrate PRE into GVN.
2171 
2172   cleanupGlobalSets();
2173 
2174   return Changed;
2175 }
2176 
2177 
processBlock(BasicBlock * BB)2178 bool GVN::processBlock(BasicBlock *BB) {
2179   // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2180   // (and incrementing BI before processing an instruction).
2181   assert(InstrsToErase.empty() &&
2182          "We expect InstrsToErase to be empty across iterations");
2183   bool ChangedFunction = false;
2184 
2185   for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2186        BI != BE;) {
2187     ChangedFunction |= processInstruction(BI);
2188     if (InstrsToErase.empty()) {
2189       ++BI;
2190       continue;
2191     }
2192 
2193     // If we need some instructions deleted, do it now.
2194     NumGVNInstr += InstrsToErase.size();
2195 
2196     // Avoid iterator invalidation.
2197     bool AtStart = BI == BB->begin();
2198     if (!AtStart)
2199       --BI;
2200 
2201     for (SmallVector<Instruction*, 4>::iterator I = InstrsToErase.begin(),
2202          E = InstrsToErase.end(); I != E; ++I) {
2203       DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2204       if (MD) MD->removeInstruction(*I);
2205       (*I)->eraseFromParent();
2206       DEBUG(verifyRemoved(*I));
2207     }
2208     InstrsToErase.clear();
2209 
2210     if (AtStart)
2211       BI = BB->begin();
2212     else
2213       ++BI;
2214   }
2215 
2216   return ChangedFunction;
2217 }
2218 
2219 /// performPRE - Perform a purely local form of PRE that looks for diamond
2220 /// control flow patterns and attempts to perform simple PRE at the join point.
performPRE(Function & F)2221 bool GVN::performPRE(Function &F) {
2222   bool Changed = false;
2223   DenseMap<BasicBlock*, Value*> predMap;
2224   for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
2225        DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
2226     BasicBlock *CurrentBlock = *DI;
2227 
2228     // Nothing to PRE in the entry block.
2229     if (CurrentBlock == &F.getEntryBlock()) continue;
2230 
2231     // Don't perform PRE on a landing pad.
2232     if (CurrentBlock->isLandingPad()) continue;
2233 
2234     for (BasicBlock::iterator BI = CurrentBlock->begin(),
2235          BE = CurrentBlock->end(); BI != BE; ) {
2236       Instruction *CurInst = BI++;
2237 
2238       if (isa<AllocaInst>(CurInst) ||
2239           isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2240           CurInst->getType()->isVoidTy() ||
2241           CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2242           isa<DbgInfoIntrinsic>(CurInst))
2243         continue;
2244 
2245       // We don't currently value number ANY inline asm calls.
2246       if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2247         if (CallI->isInlineAsm())
2248           continue;
2249 
2250       uint32_t ValNo = VN.lookup(CurInst);
2251 
2252       // Look for the predecessors for PRE opportunities.  We're
2253       // only trying to solve the basic diamond case, where
2254       // a value is computed in the successor and one predecessor,
2255       // but not the other.  We also explicitly disallow cases
2256       // where the successor is its own predecessor, because they're
2257       // more complicated to get right.
2258       unsigned NumWith = 0;
2259       unsigned NumWithout = 0;
2260       BasicBlock *PREPred = 0;
2261       predMap.clear();
2262 
2263       for (pred_iterator PI = pred_begin(CurrentBlock),
2264            PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2265         BasicBlock *P = *PI;
2266         // We're not interested in PRE where the block is its
2267         // own predecessor, or in blocks with predecessors
2268         // that are not reachable.
2269         if (P == CurrentBlock) {
2270           NumWithout = 2;
2271           break;
2272         } else if (!DT->dominates(&F.getEntryBlock(), P))  {
2273           NumWithout = 2;
2274           break;
2275         }
2276 
2277         Value* predV = findLeader(P, ValNo);
2278         if (predV == 0) {
2279           PREPred = P;
2280           ++NumWithout;
2281         } else if (predV == CurInst) {
2282           NumWithout = 2;
2283         } else {
2284           predMap[P] = predV;
2285           ++NumWith;
2286         }
2287       }
2288 
2289       // Don't do PRE when it might increase code size, i.e. when
2290       // we would need to insert instructions in more than one pred.
2291       if (NumWithout != 1 || NumWith == 0)
2292         continue;
2293 
2294       // Don't do PRE across indirect branch.
2295       if (isa<IndirectBrInst>(PREPred->getTerminator()))
2296         continue;
2297 
2298       // We can't do PRE safely on a critical edge, so instead we schedule
2299       // the edge to be split and perform the PRE the next time we iterate
2300       // on the function.
2301       unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2302       if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2303         toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2304         continue;
2305       }
2306 
2307       // Instantiate the expression in the predecessor that lacked it.
2308       // Because we are going top-down through the block, all value numbers
2309       // will be available in the predecessor by the time we need them.  Any
2310       // that weren't originally present will have been instantiated earlier
2311       // in this loop.
2312       Instruction *PREInstr = CurInst->clone();
2313       bool success = true;
2314       for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2315         Value *Op = PREInstr->getOperand(i);
2316         if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2317           continue;
2318 
2319         if (Value *V = findLeader(PREPred, VN.lookup(Op))) {
2320           PREInstr->setOperand(i, V);
2321         } else {
2322           success = false;
2323           break;
2324         }
2325       }
2326 
2327       // Fail out if we encounter an operand that is not available in
2328       // the PRE predecessor.  This is typically because of loads which
2329       // are not value numbered precisely.
2330       if (!success) {
2331         delete PREInstr;
2332         DEBUG(verifyRemoved(PREInstr));
2333         continue;
2334       }
2335 
2336       PREInstr->insertBefore(PREPred->getTerminator());
2337       PREInstr->setName(CurInst->getName() + ".pre");
2338       PREInstr->setDebugLoc(CurInst->getDebugLoc());
2339       predMap[PREPred] = PREInstr;
2340       VN.add(PREInstr, ValNo);
2341       ++NumGVNPRE;
2342 
2343       // Update the availability map to include the new instruction.
2344       addToLeaderTable(ValNo, PREInstr, PREPred);
2345 
2346       // Create a PHI to make the value available in this block.
2347       pred_iterator PB = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock);
2348       PHINode* Phi = PHINode::Create(CurInst->getType(), std::distance(PB, PE),
2349                                      CurInst->getName() + ".pre-phi",
2350                                      CurrentBlock->begin());
2351       for (pred_iterator PI = PB; PI != PE; ++PI) {
2352         BasicBlock *P = *PI;
2353         Phi->addIncoming(predMap[P], P);
2354       }
2355 
2356       VN.add(Phi, ValNo);
2357       addToLeaderTable(ValNo, Phi, CurrentBlock);
2358       Phi->setDebugLoc(CurInst->getDebugLoc());
2359       CurInst->replaceAllUsesWith(Phi);
2360       if (Phi->getType()->isPointerTy()) {
2361         // Because we have added a PHI-use of the pointer value, it has now
2362         // "escaped" from alias analysis' perspective.  We need to inform
2363         // AA of this.
2364         for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee;
2365              ++ii) {
2366           unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
2367           VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
2368         }
2369 
2370         if (MD)
2371           MD->invalidateCachedPointerInfo(Phi);
2372       }
2373       VN.erase(CurInst);
2374       removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2375 
2376       DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2377       if (MD) MD->removeInstruction(CurInst);
2378       CurInst->eraseFromParent();
2379       DEBUG(verifyRemoved(CurInst));
2380       Changed = true;
2381     }
2382   }
2383 
2384   if (splitCriticalEdges())
2385     Changed = true;
2386 
2387   return Changed;
2388 }
2389 
2390 /// splitCriticalEdges - Split critical edges found during the previous
2391 /// iteration that may enable further optimization.
splitCriticalEdges()2392 bool GVN::splitCriticalEdges() {
2393   if (toSplit.empty())
2394     return false;
2395   do {
2396     std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2397     SplitCriticalEdge(Edge.first, Edge.second, this);
2398   } while (!toSplit.empty());
2399   if (MD) MD->invalidateCachedPredecessors();
2400   return true;
2401 }
2402 
2403 /// iterateOnFunction - Executes one iteration of GVN
iterateOnFunction(Function & F)2404 bool GVN::iterateOnFunction(Function &F) {
2405   cleanupGlobalSets();
2406 
2407   // Top-down walk of the dominator tree
2408   bool Changed = false;
2409 #if 0
2410   // Needed for value numbering with phi construction to work.
2411   ReversePostOrderTraversal<Function*> RPOT(&F);
2412   for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2413        RE = RPOT.end(); RI != RE; ++RI)
2414     Changed |= processBlock(*RI);
2415 #else
2416   for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2417        DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2418     Changed |= processBlock(DI->getBlock());
2419 #endif
2420 
2421   return Changed;
2422 }
2423 
cleanupGlobalSets()2424 void GVN::cleanupGlobalSets() {
2425   VN.clear();
2426   LeaderTable.clear();
2427   TableAllocator.Reset();
2428 }
2429 
2430 /// verifyRemoved - Verify that the specified instruction does not occur in our
2431 /// internal data structures.
verifyRemoved(const Instruction * Inst) const2432 void GVN::verifyRemoved(const Instruction *Inst) const {
2433   VN.verifyRemoved(Inst);
2434 
2435   // Walk through the value number scope to make sure the instruction isn't
2436   // ferreted away in it.
2437   for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2438        I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2439     const LeaderTableEntry *Node = &I->second;
2440     assert(Node->Val != Inst && "Inst still in value numbering scope!");
2441 
2442     while (Node->Next) {
2443       Node = Node->Next;
2444       assert(Node->Val != Inst && "Inst still in value numbering scope!");
2445     }
2446   }
2447 }
2448