1 //===- GVN.cpp - Eliminate redundant values and loads ---------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass performs global value numbering to eliminate fully redundant
11 // instructions. It also performs simple dead load elimination.
13 // Note that this pass does the value numbering itself; it does not use the
14 // ValueNumbering analysis passes.
16 //===----------------------------------------------------------------------===//
18 #include "llvm/Transforms/Scalar.h"
19 #include "llvm/ADT/DenseMap.h"
20 #include "llvm/ADT/DepthFirstIterator.h"
21 #include "llvm/ADT/Hashing.h"
22 #include "llvm/ADT/MapVector.h"
23 #include "llvm/ADT/PostOrderIterator.h"
24 #include "llvm/ADT/SetVector.h"
25 #include "llvm/ADT/SmallPtrSet.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/AssumptionCache.h"
29 #include "llvm/Analysis/CFG.h"
30 #include "llvm/Analysis/ConstantFolding.h"
31 #include "llvm/Analysis/InstructionSimplify.h"
32 #include "llvm/Analysis/Loads.h"
33 #include "llvm/Analysis/MemoryBuiltins.h"
34 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
35 #include "llvm/Analysis/PHITransAddr.h"
36 #include "llvm/Analysis/ValueTracking.h"
37 #include "llvm/IR/DataLayout.h"
38 #include "llvm/IR/Dominators.h"
39 #include "llvm/IR/GlobalVariable.h"
40 #include "llvm/IR/IRBuilder.h"
41 #include "llvm/IR/IntrinsicInst.h"
42 #include "llvm/IR/LLVMContext.h"
43 #include "llvm/IR/Metadata.h"
44 #include "llvm/IR/PatternMatch.h"
45 #include "llvm/Support/Allocator.h"
46 #include "llvm/Support/CommandLine.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Analysis/TargetLibraryInfo.h"
49 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
50 #include "llvm/Transforms/Utils/Local.h"
51 #include "llvm/Transforms/Utils/SSAUpdater.h"
54 using namespace PatternMatch;
56 #define DEBUG_TYPE "gvn"
58 STATISTIC(NumGVNInstr, "Number of instructions deleted");
59 STATISTIC(NumGVNLoad, "Number of loads deleted");
60 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
61 STATISTIC(NumGVNBlocks, "Number of blocks merged");
62 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
63 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
64 STATISTIC(NumPRELoad, "Number of loads PRE'd");
66 static cl::opt<bool> EnablePRE("enable-pre",
67 cl::init(true), cl::Hidden);
68 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
70 // Maximum allowed recursion depth.
71 static cl::opt<uint32_t>
72 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
73 cl::desc("Max recurse depth (default = 1000)"));
75 //===----------------------------------------------------------------------===//
77 //===----------------------------------------------------------------------===//
79 /// This class holds the mapping between values and value numbers. It is used
80 /// as an efficient mechanism to determine the expression-wise equivalence of
86 SmallVector<uint32_t, 4> varargs;
88 Expression(uint32_t o = ~2U) : opcode(o) { }
90 bool operator==(const Expression &other) const {
91 if (opcode != other.opcode)
93 if (opcode == ~0U || opcode == ~1U)
95 if (type != other.type)
97 if (varargs != other.varargs)
102 friend hash_code hash_value(const Expression &Value) {
103 return hash_combine(Value.opcode, Value.type,
104 hash_combine_range(Value.varargs.begin(),
105 Value.varargs.end()));
110 DenseMap<Value*, uint32_t> valueNumbering;
111 DenseMap<Expression, uint32_t> expressionNumbering;
113 MemoryDependenceAnalysis *MD;
116 uint32_t nextValueNumber;
118 Expression create_expression(Instruction* I);
119 Expression create_cmp_expression(unsigned Opcode,
120 CmpInst::Predicate Predicate,
121 Value *LHS, Value *RHS);
122 Expression create_extractvalue_expression(ExtractValueInst* EI);
123 uint32_t lookup_or_add_call(CallInst* C);
125 ValueTable() : nextValueNumber(1) { }
126 uint32_t lookup_or_add(Value *V);
127 uint32_t lookup(Value *V) const;
128 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred,
129 Value *LHS, Value *RHS);
130 void add(Value *V, uint32_t num);
132 void erase(Value *v);
133 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
134 AliasAnalysis *getAliasAnalysis() const { return AA; }
135 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
136 void setDomTree(DominatorTree* D) { DT = D; }
137 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
138 void verifyRemoved(const Value *) const;
143 template <> struct DenseMapInfo<Expression> {
144 static inline Expression getEmptyKey() {
148 static inline Expression getTombstoneKey() {
152 static unsigned getHashValue(const Expression e) {
153 using llvm::hash_value;
154 return static_cast<unsigned>(hash_value(e));
156 static bool isEqual(const Expression &LHS, const Expression &RHS) {
163 //===----------------------------------------------------------------------===//
164 // ValueTable Internal Functions
165 //===----------------------------------------------------------------------===//
167 Expression ValueTable::create_expression(Instruction *I) {
169 e.type = I->getType();
170 e.opcode = I->getOpcode();
171 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
173 e.varargs.push_back(lookup_or_add(*OI));
174 if (I->isCommutative()) {
175 // Ensure that commutative instructions that only differ by a permutation
176 // of their operands get the same value number by sorting the operand value
177 // numbers. Since all commutative instructions have two operands it is more
178 // efficient to sort by hand rather than using, say, std::sort.
179 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
180 if (e.varargs[0] > e.varargs[1])
181 std::swap(e.varargs[0], e.varargs[1]);
184 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
185 // Sort the operand value numbers so x<y and y>x get the same value number.
186 CmpInst::Predicate Predicate = C->getPredicate();
187 if (e.varargs[0] > e.varargs[1]) {
188 std::swap(e.varargs[0], e.varargs[1]);
189 Predicate = CmpInst::getSwappedPredicate(Predicate);
191 e.opcode = (C->getOpcode() << 8) | Predicate;
192 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
193 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
195 e.varargs.push_back(*II);
201 Expression ValueTable::create_cmp_expression(unsigned Opcode,
202 CmpInst::Predicate Predicate,
203 Value *LHS, Value *RHS) {
204 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
205 "Not a comparison!");
207 e.type = CmpInst::makeCmpResultType(LHS->getType());
208 e.varargs.push_back(lookup_or_add(LHS));
209 e.varargs.push_back(lookup_or_add(RHS));
211 // Sort the operand value numbers so x<y and y>x get the same value number.
212 if (e.varargs[0] > e.varargs[1]) {
213 std::swap(e.varargs[0], e.varargs[1]);
214 Predicate = CmpInst::getSwappedPredicate(Predicate);
216 e.opcode = (Opcode << 8) | Predicate;
220 Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) {
221 assert(EI && "Not an ExtractValueInst?");
223 e.type = EI->getType();
226 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
227 if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
228 // EI might be an extract from one of our recognised intrinsics. If it
229 // is we'll synthesize a semantically equivalent expression instead on
230 // an extract value expression.
231 switch (I->getIntrinsicID()) {
232 case Intrinsic::sadd_with_overflow:
233 case Intrinsic::uadd_with_overflow:
234 e.opcode = Instruction::Add;
236 case Intrinsic::ssub_with_overflow:
237 case Intrinsic::usub_with_overflow:
238 e.opcode = Instruction::Sub;
240 case Intrinsic::smul_with_overflow:
241 case Intrinsic::umul_with_overflow:
242 e.opcode = Instruction::Mul;
249 // Intrinsic recognized. Grab its args to finish building the expression.
250 assert(I->getNumArgOperands() == 2 &&
251 "Expect two args for recognised intrinsics.");
252 e.varargs.push_back(lookup_or_add(I->getArgOperand(0)));
253 e.varargs.push_back(lookup_or_add(I->getArgOperand(1)));
258 // Not a recognised intrinsic. Fall back to producing an extract value
260 e.opcode = EI->getOpcode();
261 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
263 e.varargs.push_back(lookup_or_add(*OI));
265 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
267 e.varargs.push_back(*II);
272 //===----------------------------------------------------------------------===//
273 // ValueTable External Functions
274 //===----------------------------------------------------------------------===//
276 /// add - Insert a value into the table with a specified value number.
277 void ValueTable::add(Value *V, uint32_t num) {
278 valueNumbering.insert(std::make_pair(V, num));
281 uint32_t ValueTable::lookup_or_add_call(CallInst *C) {
282 if (AA->doesNotAccessMemory(C)) {
283 Expression exp = create_expression(C);
284 uint32_t &e = expressionNumbering[exp];
285 if (!e) e = nextValueNumber++;
286 valueNumbering[C] = e;
288 } else if (AA->onlyReadsMemory(C)) {
289 Expression exp = create_expression(C);
290 uint32_t &e = expressionNumbering[exp];
292 e = nextValueNumber++;
293 valueNumbering[C] = e;
297 e = nextValueNumber++;
298 valueNumbering[C] = e;
302 MemDepResult local_dep = MD->getDependency(C);
304 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
305 valueNumbering[C] = nextValueNumber;
306 return nextValueNumber++;
309 if (local_dep.isDef()) {
310 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
312 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
313 valueNumbering[C] = nextValueNumber;
314 return nextValueNumber++;
317 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
318 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
319 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
321 valueNumbering[C] = nextValueNumber;
322 return nextValueNumber++;
326 uint32_t v = lookup_or_add(local_cdep);
327 valueNumbering[C] = v;
332 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
333 MD->getNonLocalCallDependency(CallSite(C));
334 // FIXME: Move the checking logic to MemDep!
335 CallInst* cdep = nullptr;
337 // Check to see if we have a single dominating call instruction that is
339 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
340 const NonLocalDepEntry *I = &deps[i];
341 if (I->getResult().isNonLocal())
344 // We don't handle non-definitions. If we already have a call, reject
345 // instruction dependencies.
346 if (!I->getResult().isDef() || cdep != nullptr) {
351 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
352 // FIXME: All duplicated with non-local case.
353 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
354 cdep = NonLocalDepCall;
363 valueNumbering[C] = nextValueNumber;
364 return nextValueNumber++;
367 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
368 valueNumbering[C] = nextValueNumber;
369 return nextValueNumber++;
371 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
372 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
373 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
375 valueNumbering[C] = nextValueNumber;
376 return nextValueNumber++;
380 uint32_t v = lookup_or_add(cdep);
381 valueNumbering[C] = v;
385 valueNumbering[C] = nextValueNumber;
386 return nextValueNumber++;
390 /// lookup_or_add - Returns the value number for the specified value, assigning
391 /// it a new number if it did not have one before.
392 uint32_t ValueTable::lookup_or_add(Value *V) {
393 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
394 if (VI != valueNumbering.end())
397 if (!isa<Instruction>(V)) {
398 valueNumbering[V] = nextValueNumber;
399 return nextValueNumber++;
402 Instruction* I = cast<Instruction>(V);
404 switch (I->getOpcode()) {
405 case Instruction::Call:
406 return lookup_or_add_call(cast<CallInst>(I));
407 case Instruction::Add:
408 case Instruction::FAdd:
409 case Instruction::Sub:
410 case Instruction::FSub:
411 case Instruction::Mul:
412 case Instruction::FMul:
413 case Instruction::UDiv:
414 case Instruction::SDiv:
415 case Instruction::FDiv:
416 case Instruction::URem:
417 case Instruction::SRem:
418 case Instruction::FRem:
419 case Instruction::Shl:
420 case Instruction::LShr:
421 case Instruction::AShr:
422 case Instruction::And:
423 case Instruction::Or:
424 case Instruction::Xor:
425 case Instruction::ICmp:
426 case Instruction::FCmp:
427 case Instruction::Trunc:
428 case Instruction::ZExt:
429 case Instruction::SExt:
430 case Instruction::FPToUI:
431 case Instruction::FPToSI:
432 case Instruction::UIToFP:
433 case Instruction::SIToFP:
434 case Instruction::FPTrunc:
435 case Instruction::FPExt:
436 case Instruction::PtrToInt:
437 case Instruction::IntToPtr:
438 case Instruction::BitCast:
439 case Instruction::Select:
440 case Instruction::ExtractElement:
441 case Instruction::InsertElement:
442 case Instruction::ShuffleVector:
443 case Instruction::InsertValue:
444 case Instruction::GetElementPtr:
445 exp = create_expression(I);
447 case Instruction::ExtractValue:
448 exp = create_extractvalue_expression(cast<ExtractValueInst>(I));
451 valueNumbering[V] = nextValueNumber;
452 return nextValueNumber++;
455 uint32_t& e = expressionNumbering[exp];
456 if (!e) e = nextValueNumber++;
457 valueNumbering[V] = e;
461 /// Returns the value number of the specified value. Fails if
462 /// the value has not yet been numbered.
463 uint32_t ValueTable::lookup(Value *V) const {
464 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
465 assert(VI != valueNumbering.end() && "Value not numbered?");
469 /// Returns the value number of the given comparison,
470 /// assigning it a new number if it did not have one before. Useful when
471 /// we deduced the result of a comparison, but don't immediately have an
472 /// instruction realizing that comparison to hand.
473 uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode,
474 CmpInst::Predicate Predicate,
475 Value *LHS, Value *RHS) {
476 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS);
477 uint32_t& e = expressionNumbering[exp];
478 if (!e) e = nextValueNumber++;
482 /// Remove all entries from the ValueTable.
483 void ValueTable::clear() {
484 valueNumbering.clear();
485 expressionNumbering.clear();
489 /// Remove a value from the value numbering.
490 void ValueTable::erase(Value *V) {
491 valueNumbering.erase(V);
494 /// verifyRemoved - Verify that the value is removed from all internal data
496 void ValueTable::verifyRemoved(const Value *V) const {
497 for (DenseMap<Value*, uint32_t>::const_iterator
498 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
499 assert(I->first != V && "Inst still occurs in value numbering map!");
503 //===----------------------------------------------------------------------===//
505 //===----------------------------------------------------------------------===//
509 struct AvailableValueInBlock {
510 /// BB - The basic block in question.
513 SimpleVal, // A simple offsetted value that is accessed.
514 LoadVal, // A value produced by a load.
515 MemIntrin, // A memory intrinsic which is loaded from.
516 UndefVal // A UndefValue representing a value from dead block (which
517 // is not yet physically removed from the CFG).
520 /// V - The value that is live out of the block.
521 PointerIntPair<Value *, 2, ValType> Val;
523 /// Offset - The byte offset in Val that is interesting for the load query.
526 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
527 unsigned Offset = 0) {
528 AvailableValueInBlock Res;
530 Res.Val.setPointer(V);
531 Res.Val.setInt(SimpleVal);
536 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
537 unsigned Offset = 0) {
538 AvailableValueInBlock Res;
540 Res.Val.setPointer(MI);
541 Res.Val.setInt(MemIntrin);
546 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
547 unsigned Offset = 0) {
548 AvailableValueInBlock Res;
550 Res.Val.setPointer(LI);
551 Res.Val.setInt(LoadVal);
556 static AvailableValueInBlock getUndef(BasicBlock *BB) {
557 AvailableValueInBlock Res;
559 Res.Val.setPointer(nullptr);
560 Res.Val.setInt(UndefVal);
565 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
566 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
567 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
568 bool isUndefValue() const { return Val.getInt() == UndefVal; }
570 Value *getSimpleValue() const {
571 assert(isSimpleValue() && "Wrong accessor");
572 return Val.getPointer();
575 LoadInst *getCoercedLoadValue() const {
576 assert(isCoercedLoadValue() && "Wrong accessor");
577 return cast<LoadInst>(Val.getPointer());
580 MemIntrinsic *getMemIntrinValue() const {
581 assert(isMemIntrinValue() && "Wrong accessor");
582 return cast<MemIntrinsic>(Val.getPointer());
585 /// Emit code into this block to adjust the value defined here to the
586 /// specified type. This handles various coercion cases.
587 Value *MaterializeAdjustedValue(LoadInst *LI, GVN &gvn) const;
590 class GVN : public FunctionPass {
592 MemoryDependenceAnalysis *MD;
594 const TargetLibraryInfo *TLI;
596 SetVector<BasicBlock *> DeadBlocks;
600 /// A mapping from value numbers to lists of Value*'s that
601 /// have that value number. Use findLeader to query it.
602 struct LeaderTableEntry {
604 const BasicBlock *BB;
605 LeaderTableEntry *Next;
607 DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
608 BumpPtrAllocator TableAllocator;
610 SmallVector<Instruction*, 8> InstrsToErase;
612 typedef SmallVector<NonLocalDepResult, 64> LoadDepVect;
613 typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect;
614 typedef SmallVector<BasicBlock*, 64> UnavailBlkVect;
617 static char ID; // Pass identification, replacement for typeid
618 explicit GVN(bool noloads = false)
619 : FunctionPass(ID), NoLoads(noloads), MD(nullptr) {
620 initializeGVNPass(*PassRegistry::getPassRegistry());
623 bool runOnFunction(Function &F) override;
625 /// This removes the specified instruction from
626 /// our various maps and marks it for deletion.
627 void markInstructionForDeletion(Instruction *I) {
629 InstrsToErase.push_back(I);
632 DominatorTree &getDominatorTree() const { return *DT; }
633 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
634 MemoryDependenceAnalysis &getMemDep() const { return *MD; }
636 /// Push a new Value to the LeaderTable onto the list for its value number.
637 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) {
638 LeaderTableEntry &Curr = LeaderTable[N];
645 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
648 Node->Next = Curr.Next;
652 /// Scan the list of values corresponding to a given
653 /// value number, and remove the given instruction if encountered.
654 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) {
655 LeaderTableEntry* Prev = nullptr;
656 LeaderTableEntry* Curr = &LeaderTable[N];
658 while (Curr->Val != I || Curr->BB != BB) {
664 Prev->Next = Curr->Next;
670 LeaderTableEntry* Next = Curr->Next;
671 Curr->Val = Next->Val;
673 Curr->Next = Next->Next;
678 // List of critical edges to be split between iterations.
679 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
681 // This transformation requires dominator postdominator info
682 void getAnalysisUsage(AnalysisUsage &AU) const override {
683 AU.addRequired<AssumptionCacheTracker>();
684 AU.addRequired<DominatorTreeWrapperPass>();
685 AU.addRequired<TargetLibraryInfoWrapperPass>();
687 AU.addRequired<MemoryDependenceAnalysis>();
688 AU.addRequired<AliasAnalysis>();
690 AU.addPreserved<DominatorTreeWrapperPass>();
691 AU.addPreserved<AliasAnalysis>();
695 // Helper fuctions of redundant load elimination
696 bool processLoad(LoadInst *L);
697 bool processNonLocalLoad(LoadInst *L);
698 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
699 AvailValInBlkVect &ValuesPerBlock,
700 UnavailBlkVect &UnavailableBlocks);
701 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
702 UnavailBlkVect &UnavailableBlocks);
704 // Other helper routines
705 bool processInstruction(Instruction *I);
706 bool processBlock(BasicBlock *BB);
707 void dump(DenseMap<uint32_t, Value*> &d);
708 bool iterateOnFunction(Function &F);
709 bool performPRE(Function &F);
710 bool performScalarPRE(Instruction *I);
711 bool performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
713 Value *findLeader(const BasicBlock *BB, uint32_t num);
714 void cleanupGlobalSets();
715 void verifyRemoved(const Instruction *I) const;
716 bool splitCriticalEdges();
717 BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ);
718 unsigned replaceAllDominatedUsesWith(Value *From, Value *To,
719 const BasicBlockEdge &Root);
720 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root);
721 bool processFoldableCondBr(BranchInst *BI);
722 void addDeadBlock(BasicBlock *BB);
723 void assignValNumForDeadCode();
729 // The public interface to this file...
730 FunctionPass *llvm::createGVNPass(bool NoLoads) {
731 return new GVN(NoLoads);
734 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
735 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
736 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
737 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
738 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
739 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
740 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
742 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
743 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
745 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
746 E = d.end(); I != E; ++I) {
747 errs() << I->first << "\n";
754 /// Return true if we can prove that the value
755 /// we're analyzing is fully available in the specified block. As we go, keep
756 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
757 /// map is actually a tri-state map with the following values:
758 /// 0) we know the block *is not* fully available.
759 /// 1) we know the block *is* fully available.
760 /// 2) we do not know whether the block is fully available or not, but we are
761 /// currently speculating that it will be.
762 /// 3) we are speculating for this block and have used that to speculate for
764 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
765 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
766 uint32_t RecurseDepth) {
767 if (RecurseDepth > MaxRecurseDepth)
770 // Optimistically assume that the block is fully available and check to see
771 // if we already know about this block in one lookup.
772 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
773 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
775 // If the entry already existed for this block, return the precomputed value.
777 // If this is a speculative "available" value, mark it as being used for
778 // speculation of other blocks.
779 if (IV.first->second == 2)
780 IV.first->second = 3;
781 return IV.first->second != 0;
784 // Otherwise, see if it is fully available in all predecessors.
785 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
787 // If this block has no predecessors, it isn't live-in here.
789 goto SpeculationFailure;
791 for (; PI != PE; ++PI)
792 // If the value isn't fully available in one of our predecessors, then it
793 // isn't fully available in this block either. Undo our previous
794 // optimistic assumption and bail out.
795 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
796 goto SpeculationFailure;
800 // If we get here, we found out that this is not, after
801 // all, a fully-available block. We have a problem if we speculated on this and
802 // used the speculation to mark other blocks as available.
804 char &BBVal = FullyAvailableBlocks[BB];
806 // If we didn't speculate on this, just return with it set to false.
812 // If we did speculate on this value, we could have blocks set to 1 that are
813 // incorrect. Walk the (transitive) successors of this block and mark them as
815 SmallVector<BasicBlock*, 32> BBWorklist;
816 BBWorklist.push_back(BB);
819 BasicBlock *Entry = BBWorklist.pop_back_val();
820 // Note that this sets blocks to 0 (unavailable) if they happen to not
821 // already be in FullyAvailableBlocks. This is safe.
822 char &EntryVal = FullyAvailableBlocks[Entry];
823 if (EntryVal == 0) continue; // Already unavailable.
825 // Mark as unavailable.
828 BBWorklist.append(succ_begin(Entry), succ_end(Entry));
829 } while (!BBWorklist.empty());
835 /// Return true if CoerceAvailableValueToLoadType will succeed.
836 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
838 const DataLayout &DL) {
839 // If the loaded or stored value is an first class array or struct, don't try
840 // to transform them. We need to be able to bitcast to integer.
841 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
842 StoredVal->getType()->isStructTy() ||
843 StoredVal->getType()->isArrayTy())
846 // The store has to be at least as big as the load.
847 if (DL.getTypeSizeInBits(StoredVal->getType()) <
848 DL.getTypeSizeInBits(LoadTy))
854 /// If we saw a store of a value to memory, and
855 /// then a load from a must-aliased pointer of a different type, try to coerce
856 /// the stored value. LoadedTy is the type of the load we want to replace and
857 /// InsertPt is the place to insert new instructions.
859 /// If we can't do it, return null.
860 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
862 Instruction *InsertPt,
863 const DataLayout &DL) {
864 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL))
867 // If this is already the right type, just return it.
868 Type *StoredValTy = StoredVal->getType();
870 uint64_t StoreSize = DL.getTypeSizeInBits(StoredValTy);
871 uint64_t LoadSize = DL.getTypeSizeInBits(LoadedTy);
873 // If the store and reload are the same size, we can always reuse it.
874 if (StoreSize == LoadSize) {
875 // Pointer to Pointer -> use bitcast.
876 if (StoredValTy->getScalarType()->isPointerTy() &&
877 LoadedTy->getScalarType()->isPointerTy())
878 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
880 // Convert source pointers to integers, which can be bitcast.
881 if (StoredValTy->getScalarType()->isPointerTy()) {
882 StoredValTy = DL.getIntPtrType(StoredValTy);
883 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
886 Type *TypeToCastTo = LoadedTy;
887 if (TypeToCastTo->getScalarType()->isPointerTy())
888 TypeToCastTo = DL.getIntPtrType(TypeToCastTo);
890 if (StoredValTy != TypeToCastTo)
891 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
893 // Cast to pointer if the load needs a pointer type.
894 if (LoadedTy->getScalarType()->isPointerTy())
895 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
900 // If the loaded value is smaller than the available value, then we can
901 // extract out a piece from it. If the available value is too small, then we
902 // can't do anything.
903 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
905 // Convert source pointers to integers, which can be manipulated.
906 if (StoredValTy->getScalarType()->isPointerTy()) {
907 StoredValTy = DL.getIntPtrType(StoredValTy);
908 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
911 // Convert vectors and fp to integer, which can be manipulated.
912 if (!StoredValTy->isIntegerTy()) {
913 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
914 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
917 // If this is a big-endian system, we need to shift the value down to the low
918 // bits so that a truncate will work.
919 if (DL.isBigEndian()) {
920 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
921 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
924 // Truncate the integer to the right size now.
925 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
926 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
928 if (LoadedTy == NewIntTy)
931 // If the result is a pointer, inttoptr.
932 if (LoadedTy->getScalarType()->isPointerTy())
933 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
935 // Otherwise, bitcast.
936 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
939 /// This function is called when we have a
940 /// memdep query of a load that ends up being a clobbering memory write (store,
941 /// memset, memcpy, memmove). This means that the write *may* provide bits used
942 /// by the load but we can't be sure because the pointers don't mustalias.
944 /// Check this case to see if there is anything more we can do before we give
945 /// up. This returns -1 if we have to give up, or a byte number in the stored
946 /// value of the piece that feeds the load.
947 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
949 uint64_t WriteSizeInBits,
950 const DataLayout &DL) {
951 // If the loaded or stored value is a first class array or struct, don't try
952 // to transform them. We need to be able to bitcast to integer.
953 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
956 int64_t StoreOffset = 0, LoadOffset = 0;
958 GetPointerBaseWithConstantOffset(WritePtr, StoreOffset, DL);
959 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, DL);
960 if (StoreBase != LoadBase)
963 // If the load and store are to the exact same address, they should have been
964 // a must alias. AA must have gotten confused.
965 // FIXME: Study to see if/when this happens. One case is forwarding a memset
966 // to a load from the base of the memset.
968 if (LoadOffset == StoreOffset) {
969 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
970 << "Base = " << *StoreBase << "\n"
971 << "Store Ptr = " << *WritePtr << "\n"
972 << "Store Offs = " << StoreOffset << "\n"
973 << "Load Ptr = " << *LoadPtr << "\n";
978 // If the load and store don't overlap at all, the store doesn't provide
979 // anything to the load. In this case, they really don't alias at all, AA
980 // must have gotten confused.
981 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy);
983 if ((WriteSizeInBits & 7) | (LoadSize & 7))
985 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
989 bool isAAFailure = false;
990 if (StoreOffset < LoadOffset)
991 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
993 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
997 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
998 << "Base = " << *StoreBase << "\n"
999 << "Store Ptr = " << *WritePtr << "\n"
1000 << "Store Offs = " << StoreOffset << "\n"
1001 << "Load Ptr = " << *LoadPtr << "\n";
1007 // If the Load isn't completely contained within the stored bits, we don't
1008 // have all the bits to feed it. We could do something crazy in the future
1009 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1011 if (StoreOffset > LoadOffset ||
1012 StoreOffset+StoreSize < LoadOffset+LoadSize)
1015 // Okay, we can do this transformation. Return the number of bytes into the
1016 // store that the load is.
1017 return LoadOffset-StoreOffset;
1020 /// This function is called when we have a
1021 /// memdep query of a load that ends up being a clobbering store.
1022 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
1024 // Cannot handle reading from store of first-class aggregate yet.
1025 if (DepSI->getValueOperand()->getType()->isStructTy() ||
1026 DepSI->getValueOperand()->getType()->isArrayTy())
1029 const DataLayout &DL = DepSI->getModule()->getDataLayout();
1030 Value *StorePtr = DepSI->getPointerOperand();
1031 uint64_t StoreSize =DL.getTypeSizeInBits(DepSI->getValueOperand()->getType());
1032 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1033 StorePtr, StoreSize, DL);
1036 /// This function is called when we have a
1037 /// memdep query of a load that ends up being clobbered by another load. See if
1038 /// the other load can feed into the second load.
1039 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
1040 LoadInst *DepLI, const DataLayout &DL){
1041 // Cannot handle reading from store of first-class aggregate yet.
1042 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
1045 Value *DepPtr = DepLI->getPointerOperand();
1046 uint64_t DepSize = DL.getTypeSizeInBits(DepLI->getType());
1047 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL);
1048 if (R != -1) return R;
1050 // If we have a load/load clobber an DepLI can be widened to cover this load,
1051 // then we should widen it!
1052 int64_t LoadOffs = 0;
1053 const Value *LoadBase =
1054 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, DL);
1055 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1057 unsigned Size = MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize(
1058 LoadBase, LoadOffs, LoadSize, DepLI);
1059 if (Size == 0) return -1;
1061 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, DL);
1066 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
1068 const DataLayout &DL) {
1069 // If the mem operation is a non-constant size, we can't handle it.
1070 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1071 if (!SizeCst) return -1;
1072 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1074 // If this is memset, we just need to see if the offset is valid in the size
1076 if (MI->getIntrinsicID() == Intrinsic::memset)
1077 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
1080 // If we have a memcpy/memmove, the only case we can handle is if this is a
1081 // copy from constant memory. In that case, we can read directly from the
1083 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1085 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1086 if (!Src) return -1;
1088 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, DL));
1089 if (!GV || !GV->isConstant()) return -1;
1091 // See if the access is within the bounds of the transfer.
1092 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1093 MI->getDest(), MemSizeInBits, DL);
1097 unsigned AS = Src->getType()->getPointerAddressSpace();
1098 // Otherwise, see if we can constant fold a load from the constant with the
1099 // offset applied as appropriate.
1100 Src = ConstantExpr::getBitCast(Src,
1101 Type::getInt8PtrTy(Src->getContext(), AS));
1102 Constant *OffsetCst =
1103 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1104 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1105 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1106 if (ConstantFoldLoadFromConstPtr(Src, DL))
1112 /// This function is called when we have a
1113 /// memdep query of a load that ends up being a clobbering store. This means
1114 /// that the store provides bits used by the load but we the pointers don't
1115 /// mustalias. Check this case to see if there is anything more we can do
1116 /// before we give up.
1117 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1119 Instruction *InsertPt, const DataLayout &DL){
1120 LLVMContext &Ctx = SrcVal->getType()->getContext();
1122 uint64_t StoreSize = (DL.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1123 uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy) + 7) / 8;
1125 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1127 // Compute which bits of the stored value are being used by the load. Convert
1128 // to an integer type to start with.
1129 if (SrcVal->getType()->getScalarType()->isPointerTy())
1130 SrcVal = Builder.CreatePtrToInt(SrcVal,
1131 DL.getIntPtrType(SrcVal->getType()));
1132 if (!SrcVal->getType()->isIntegerTy())
1133 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1135 // Shift the bits to the least significant depending on endianness.
1137 if (DL.isLittleEndian())
1138 ShiftAmt = Offset*8;
1140 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1143 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1145 if (LoadSize != StoreSize)
1146 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1148 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, DL);
1151 /// This function is called when we have a
1152 /// memdep query of a load that ends up being a clobbering load. This means
1153 /// that the load *may* provide bits used by the load but we can't be sure
1154 /// because the pointers don't mustalias. Check this case to see if there is
1155 /// anything more we can do before we give up.
1156 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1157 Type *LoadTy, Instruction *InsertPt,
1159 const DataLayout &DL = SrcVal->getModule()->getDataLayout();
1160 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1161 // widen SrcVal out to a larger load.
1162 unsigned SrcValSize = DL.getTypeStoreSize(SrcVal->getType());
1163 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1164 if (Offset+LoadSize > SrcValSize) {
1165 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1166 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1167 // If we have a load/load clobber an DepLI can be widened to cover this
1168 // load, then we should widen it to the next power of 2 size big enough!
1169 unsigned NewLoadSize = Offset+LoadSize;
1170 if (!isPowerOf2_32(NewLoadSize))
1171 NewLoadSize = NextPowerOf2(NewLoadSize);
1173 Value *PtrVal = SrcVal->getPointerOperand();
1175 // Insert the new load after the old load. This ensures that subsequent
1176 // memdep queries will find the new load. We can't easily remove the old
1177 // load completely because it is already in the value numbering table.
1178 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1180 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1181 DestPTy = PointerType::get(DestPTy,
1182 PtrVal->getType()->getPointerAddressSpace());
1183 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1184 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1185 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1186 NewLoad->takeName(SrcVal);
1187 NewLoad->setAlignment(SrcVal->getAlignment());
1189 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1190 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1192 // Replace uses of the original load with the wider load. On a big endian
1193 // system, we need to shift down to get the relevant bits.
1194 Value *RV = NewLoad;
1195 if (DL.isBigEndian())
1196 RV = Builder.CreateLShr(RV,
1197 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
1198 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1199 SrcVal->replaceAllUsesWith(RV);
1201 // We would like to use gvn.markInstructionForDeletion here, but we can't
1202 // because the load is already memoized into the leader map table that GVN
1203 // tracks. It is potentially possible to remove the load from the table,
1204 // but then there all of the operations based on it would need to be
1205 // rehashed. Just leave the dead load around.
1206 gvn.getMemDep().removeInstruction(SrcVal);
1210 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, DL);
1214 /// This function is called when we have a
1215 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1216 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1217 Type *LoadTy, Instruction *InsertPt,
1218 const DataLayout &DL){
1219 LLVMContext &Ctx = LoadTy->getContext();
1220 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy)/8;
1222 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1224 // We know that this method is only called when the mem transfer fully
1225 // provides the bits for the load.
1226 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1227 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1228 // independently of what the offset is.
1229 Value *Val = MSI->getValue();
1231 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1233 Value *OneElt = Val;
1235 // Splat the value out to the right number of bits.
1236 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1237 // If we can double the number of bytes set, do it.
1238 if (NumBytesSet*2 <= LoadSize) {
1239 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1240 Val = Builder.CreateOr(Val, ShVal);
1245 // Otherwise insert one byte at a time.
1246 Value *ShVal = Builder.CreateShl(Val, 1*8);
1247 Val = Builder.CreateOr(OneElt, ShVal);
1251 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, DL);
1254 // Otherwise, this is a memcpy/memmove from a constant global.
1255 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1256 Constant *Src = cast<Constant>(MTI->getSource());
1257 unsigned AS = Src->getType()->getPointerAddressSpace();
1259 // Otherwise, see if we can constant fold a load from the constant with the
1260 // offset applied as appropriate.
1261 Src = ConstantExpr::getBitCast(Src,
1262 Type::getInt8PtrTy(Src->getContext(), AS));
1263 Constant *OffsetCst =
1264 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1265 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1266 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1267 return ConstantFoldLoadFromConstPtr(Src, DL);
1271 /// Given a set of loads specified by ValuesPerBlock,
1272 /// construct SSA form, allowing us to eliminate LI. This returns the value
1273 /// that should be used at LI's definition site.
1274 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1275 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1277 // Check for the fully redundant, dominating load case. In this case, we can
1278 // just use the dominating value directly.
1279 if (ValuesPerBlock.size() == 1 &&
1280 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1282 assert(!ValuesPerBlock[0].isUndefValue() && "Dead BB dominate this block");
1283 return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
1286 // Otherwise, we have to construct SSA form.
1287 SmallVector<PHINode*, 8> NewPHIs;
1288 SSAUpdater SSAUpdate(&NewPHIs);
1289 SSAUpdate.Initialize(LI->getType(), LI->getName());
1291 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1292 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1293 BasicBlock *BB = AV.BB;
1295 if (SSAUpdate.HasValueForBlock(BB))
1298 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
1301 // Perform PHI construction.
1302 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1304 // If new PHI nodes were created, notify alias analysis.
1305 if (V->getType()->getScalarType()->isPointerTy()) {
1306 AliasAnalysis *AA = gvn.getAliasAnalysis();
1308 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1309 AA->copyValue(LI, NewPHIs[i]);
1311 // Now that we've copied information to the new PHIs, scan through
1312 // them again and inform alias analysis that we've added potentially
1313 // escaping uses to any values that are operands to these PHIs.
1314 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
1315 PHINode *P = NewPHIs[i];
1316 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) {
1317 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
1318 AA->addEscapingUse(P->getOperandUse(jj));
1326 Value *AvailableValueInBlock::MaterializeAdjustedValue(LoadInst *LI,
1329 Type *LoadTy = LI->getType();
1330 const DataLayout &DL = LI->getModule()->getDataLayout();
1331 if (isSimpleValue()) {
1332 Res = getSimpleValue();
1333 if (Res->getType() != LoadTy) {
1334 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), DL);
1336 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1337 << *getSimpleValue() << '\n'
1338 << *Res << '\n' << "\n\n\n");
1340 } else if (isCoercedLoadValue()) {
1341 LoadInst *Load = getCoercedLoadValue();
1342 if (Load->getType() == LoadTy && Offset == 0) {
1345 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1348 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1349 << *getCoercedLoadValue() << '\n'
1350 << *Res << '\n' << "\n\n\n");
1352 } else if (isMemIntrinValue()) {
1353 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
1354 BB->getTerminator(), DL);
1355 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1356 << " " << *getMemIntrinValue() << '\n'
1357 << *Res << '\n' << "\n\n\n");
1359 assert(isUndefValue() && "Should be UndefVal");
1360 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
1361 return UndefValue::get(LoadTy);
1366 static bool isLifetimeStart(const Instruction *Inst) {
1367 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1368 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1372 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1373 AvailValInBlkVect &ValuesPerBlock,
1374 UnavailBlkVect &UnavailableBlocks) {
1376 // Filter out useless results (non-locals, etc). Keep track of the blocks
1377 // where we have a value available in repl, also keep track of whether we see
1378 // dependencies that produce an unknown value for the load (such as a call
1379 // that could potentially clobber the load).
1380 unsigned NumDeps = Deps.size();
1381 const DataLayout &DL = LI->getModule()->getDataLayout();
1382 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1383 BasicBlock *DepBB = Deps[i].getBB();
1384 MemDepResult DepInfo = Deps[i].getResult();
1386 if (DeadBlocks.count(DepBB)) {
1387 // Dead dependent mem-op disguise as a load evaluating the same value
1388 // as the load in question.
1389 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1393 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1394 UnavailableBlocks.push_back(DepBB);
1398 if (DepInfo.isClobber()) {
1399 // The address being loaded in this non-local block may not be the same as
1400 // the pointer operand of the load if PHI translation occurs. Make sure
1401 // to consider the right address.
1402 Value *Address = Deps[i].getAddress();
1404 // If the dependence is to a store that writes to a superset of the bits
1405 // read by the load, we can extract the bits we need for the load from the
1407 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1410 AnalyzeLoadFromClobberingStore(LI->getType(), Address, DepSI);
1412 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1413 DepSI->getValueOperand(),
1420 // Check to see if we have something like this:
1423 // if we have this, replace the later with an extraction from the former.
1424 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1425 // If this is a clobber and L is the first instruction in its block, then
1426 // we have the first instruction in the entry block.
1427 if (DepLI != LI && Address) {
1429 AnalyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
1432 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1439 // If the clobbering value is a memset/memcpy/memmove, see if we can
1440 // forward a value on from it.
1441 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1443 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1446 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1453 UnavailableBlocks.push_back(DepBB);
1457 // DepInfo.isDef() here
1459 Instruction *DepInst = DepInfo.getInst();
1461 // Loading the allocation -> undef.
1462 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
1463 // Loading immediately after lifetime begin -> undef.
1464 isLifetimeStart(DepInst)) {
1465 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1466 UndefValue::get(LI->getType())));
1470 // Loading from calloc (which zero initializes memory) -> zero
1471 if (isCallocLikeFn(DepInst, TLI)) {
1472 ValuesPerBlock.push_back(AvailableValueInBlock::get(
1473 DepBB, Constant::getNullValue(LI->getType())));
1477 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1478 // Reject loads and stores that are to the same address but are of
1479 // different types if we have to.
1480 if (S->getValueOperand()->getType() != LI->getType()) {
1481 // If the stored value is larger or equal to the loaded value, we can
1483 if (!CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1484 LI->getType(), DL)) {
1485 UnavailableBlocks.push_back(DepBB);
1490 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1491 S->getValueOperand()));
1495 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1496 // If the types mismatch and we can't handle it, reject reuse of the load.
1497 if (LD->getType() != LI->getType()) {
1498 // If the stored value is larger or equal to the loaded value, we can
1500 if (!CanCoerceMustAliasedValueToLoad(LD, LI->getType(), DL)) {
1501 UnavailableBlocks.push_back(DepBB);
1505 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1509 UnavailableBlocks.push_back(DepBB);
1513 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1514 UnavailBlkVect &UnavailableBlocks) {
1515 // Okay, we have *some* definitions of the value. This means that the value
1516 // is available in some of our (transitive) predecessors. Lets think about
1517 // doing PRE of this load. This will involve inserting a new load into the
1518 // predecessor when it's not available. We could do this in general, but
1519 // prefer to not increase code size. As such, we only do this when we know
1520 // that we only have to insert *one* load (which means we're basically moving
1521 // the load, not inserting a new one).
1523 SmallPtrSet<BasicBlock *, 4> Blockers;
1524 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1525 Blockers.insert(UnavailableBlocks[i]);
1527 // Let's find the first basic block with more than one predecessor. Walk
1528 // backwards through predecessors if needed.
1529 BasicBlock *LoadBB = LI->getParent();
1530 BasicBlock *TmpBB = LoadBB;
1532 while (TmpBB->getSinglePredecessor()) {
1533 TmpBB = TmpBB->getSinglePredecessor();
1534 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1536 if (Blockers.count(TmpBB))
1539 // If any of these blocks has more than one successor (i.e. if the edge we
1540 // just traversed was critical), then there are other paths through this
1541 // block along which the load may not be anticipated. Hoisting the load
1542 // above this block would be adding the load to execution paths along
1543 // which it was not previously executed.
1544 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1551 // Check to see how many predecessors have the loaded value fully
1553 MapVector<BasicBlock *, Value *> PredLoads;
1554 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1555 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1556 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1557 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1558 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1560 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1561 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1563 BasicBlock *Pred = *PI;
1564 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1568 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1569 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1570 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1571 << Pred->getName() << "': " << *LI << '\n');
1575 if (LoadBB->isLandingPad()) {
1577 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '"
1578 << Pred->getName() << "': " << *LI << '\n');
1582 CriticalEdgePred.push_back(Pred);
1584 // Only add the predecessors that will not be split for now.
1585 PredLoads[Pred] = nullptr;
1589 // Decide whether PRE is profitable for this load.
1590 unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1591 assert(NumUnavailablePreds != 0 &&
1592 "Fully available value should already be eliminated!");
1594 // If this load is unavailable in multiple predecessors, reject it.
1595 // FIXME: If we could restructure the CFG, we could make a common pred with
1596 // all the preds that don't have an available LI and insert a new load into
1598 if (NumUnavailablePreds != 1)
1601 // Split critical edges, and update the unavailable predecessors accordingly.
1602 for (BasicBlock *OrigPred : CriticalEdgePred) {
1603 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1604 assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1605 PredLoads[NewPred] = nullptr;
1606 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1607 << LoadBB->getName() << '\n');
1610 // Check if the load can safely be moved to all the unavailable predecessors.
1611 bool CanDoPRE = true;
1612 const DataLayout &DL = LI->getModule()->getDataLayout();
1613 SmallVector<Instruction*, 8> NewInsts;
1614 for (auto &PredLoad : PredLoads) {
1615 BasicBlock *UnavailablePred = PredLoad.first;
1617 // Do PHI translation to get its value in the predecessor if necessary. The
1618 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1620 // If all preds have a single successor, then we know it is safe to insert
1621 // the load on the pred (?!?), so we can insert code to materialize the
1622 // pointer if it is not available.
1623 PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1624 Value *LoadPtr = nullptr;
1625 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1628 // If we couldn't find or insert a computation of this phi translated value,
1631 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1632 << *LI->getPointerOperand() << "\n");
1637 PredLoad.second = LoadPtr;
1641 while (!NewInsts.empty()) {
1642 Instruction *I = NewInsts.pop_back_val();
1643 if (MD) MD->removeInstruction(I);
1644 I->eraseFromParent();
1646 // HINT: Don't revert the edge-splitting as following transformation may
1647 // also need to split these critical edges.
1648 return !CriticalEdgePred.empty();
1651 // Okay, we can eliminate this load by inserting a reload in the predecessor
1652 // and using PHI construction to get the value in the other predecessors, do
1654 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1655 DEBUG(if (!NewInsts.empty())
1656 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1657 << *NewInsts.back() << '\n');
1659 // Assign value numbers to the new instructions.
1660 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1661 // FIXME: We really _ought_ to insert these value numbers into their
1662 // parent's availability map. However, in doing so, we risk getting into
1663 // ordering issues. If a block hasn't been processed yet, we would be
1664 // marking a value as AVAIL-IN, which isn't what we intend.
1665 VN.lookup_or_add(NewInsts[i]);
1668 for (const auto &PredLoad : PredLoads) {
1669 BasicBlock *UnavailablePred = PredLoad.first;
1670 Value *LoadPtr = PredLoad.second;
1672 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1674 UnavailablePred->getTerminator());
1676 // Transfer the old load's AA tags to the new load.
1678 LI->getAAMetadata(Tags);
1680 NewLoad->setAAMetadata(Tags);
1682 // Transfer DebugLoc.
1683 NewLoad->setDebugLoc(LI->getDebugLoc());
1685 // Add the newly created load.
1686 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1688 MD->invalidateCachedPointerInfo(LoadPtr);
1689 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1692 // Perform PHI construction.
1693 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1694 LI->replaceAllUsesWith(V);
1695 if (isa<PHINode>(V))
1697 if (V->getType()->getScalarType()->isPointerTy())
1698 MD->invalidateCachedPointerInfo(V);
1699 markInstructionForDeletion(LI);
1704 /// Attempt to eliminate a load whose dependencies are
1705 /// non-local by performing PHI construction.
1706 bool GVN::processNonLocalLoad(LoadInst *LI) {
1707 // Step 1: Find the non-local dependencies of the load.
1709 MD->getNonLocalPointerDependency(LI, Deps);
1711 // If we had to process more than one hundred blocks to find the
1712 // dependencies, this load isn't worth worrying about. Optimizing
1713 // it will be too expensive.
1714 unsigned NumDeps = Deps.size();
1718 // If we had a phi translation failure, we'll have a single entry which is a
1719 // clobber in the current block. Reject this early.
1721 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1723 dbgs() << "GVN: non-local load ";
1724 LI->printAsOperand(dbgs());
1725 dbgs() << " has unknown dependencies\n";
1730 // If this load follows a GEP, see if we can PRE the indices before analyzing.
1731 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1732 for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1733 OE = GEP->idx_end();
1735 if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1736 performScalarPRE(I);
1739 // Step 2: Analyze the availability of the load
1740 AvailValInBlkVect ValuesPerBlock;
1741 UnavailBlkVect UnavailableBlocks;
1742 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1744 // If we have no predecessors that produce a known value for this load, exit
1746 if (ValuesPerBlock.empty())
1749 // Step 3: Eliminate fully redundancy.
1751 // If all of the instructions we depend on produce a known value for this
1752 // load, then it is fully redundant and we can use PHI insertion to compute
1753 // its value. Insert PHIs and remove the fully redundant value now.
1754 if (UnavailableBlocks.empty()) {
1755 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1757 // Perform PHI construction.
1758 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1759 LI->replaceAllUsesWith(V);
1761 if (isa<PHINode>(V))
1763 if (V->getType()->getScalarType()->isPointerTy())
1764 MD->invalidateCachedPointerInfo(V);
1765 markInstructionForDeletion(LI);
1770 // Step 4: Eliminate partial redundancy.
1771 if (!EnablePRE || !EnableLoadPRE)
1774 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1778 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1779 // Patch the replacement so that it is not more restrictive than the value
1781 BinaryOperator *Op = dyn_cast<BinaryOperator>(I);
1782 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl);
1783 if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) &&
1784 isa<OverflowingBinaryOperator>(ReplOp)) {
1785 if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap())
1786 ReplOp->setHasNoSignedWrap(false);
1787 if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap())
1788 ReplOp->setHasNoUnsignedWrap(false);
1790 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) {
1791 // FIXME: If both the original and replacement value are part of the
1792 // same control-flow region (meaning that the execution of one
1793 // guarentees the executation of the other), then we can combine the
1794 // noalias scopes here and do better than the general conservative
1795 // answer used in combineMetadata().
1797 // In general, GVN unifies expressions over different control-flow
1798 // regions, and so we need a conservative combination of the noalias
1800 unsigned KnownIDs[] = {
1801 LLVMContext::MD_tbaa,
1802 LLVMContext::MD_alias_scope,
1803 LLVMContext::MD_noalias,
1804 LLVMContext::MD_range,
1805 LLVMContext::MD_fpmath,
1806 LLVMContext::MD_invariant_load,
1808 combineMetadata(ReplInst, I, KnownIDs);
1812 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1813 patchReplacementInstruction(I, Repl);
1814 I->replaceAllUsesWith(Repl);
1817 /// Attempt to eliminate a load, first by eliminating it
1818 /// locally, and then attempting non-local elimination if that fails.
1819 bool GVN::processLoad(LoadInst *L) {
1826 if (L->use_empty()) {
1827 markInstructionForDeletion(L);
1831 // ... to a pointer that has been loaded from before...
1832 MemDepResult Dep = MD->getDependency(L);
1833 const DataLayout &DL = L->getModule()->getDataLayout();
1835 // If we have a clobber and target data is around, see if this is a clobber
1836 // that we can fix up through code synthesis.
1837 if (Dep.isClobber()) {
1838 // Check to see if we have something like this:
1839 // store i32 123, i32* %P
1840 // %A = bitcast i32* %P to i8*
1841 // %B = gep i8* %A, i32 1
1844 // We could do that by recognizing if the clobber instructions are obviously
1845 // a common base + constant offset, and if the previous store (or memset)
1846 // completely covers this load. This sort of thing can happen in bitfield
1848 Value *AvailVal = nullptr;
1849 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1850 int Offset = AnalyzeLoadFromClobberingStore(
1851 L->getType(), L->getPointerOperand(), DepSI);
1853 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1854 L->getType(), L, DL);
1857 // Check to see if we have something like this:
1860 // if we have this, replace the later with an extraction from the former.
1861 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1862 // If this is a clobber and L is the first instruction in its block, then
1863 // we have the first instruction in the entry block.
1867 int Offset = AnalyzeLoadFromClobberingLoad(
1868 L->getType(), L->getPointerOperand(), DepLI, DL);
1870 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1873 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1874 // a value on from it.
1875 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1876 int Offset = AnalyzeLoadFromClobberingMemInst(
1877 L->getType(), L->getPointerOperand(), DepMI, DL);
1879 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, DL);
1883 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1884 << *AvailVal << '\n' << *L << "\n\n\n");
1886 // Replace the load!
1887 L->replaceAllUsesWith(AvailVal);
1888 if (AvailVal->getType()->getScalarType()->isPointerTy())
1889 MD->invalidateCachedPointerInfo(AvailVal);
1890 markInstructionForDeletion(L);
1896 // If the value isn't available, don't do anything!
1897 if (Dep.isClobber()) {
1899 // fast print dep, using operator<< on instruction is too slow.
1900 dbgs() << "GVN: load ";
1901 L->printAsOperand(dbgs());
1902 Instruction *I = Dep.getInst();
1903 dbgs() << " is clobbered by " << *I << '\n';
1908 // If it is defined in another block, try harder.
1909 if (Dep.isNonLocal())
1910 return processNonLocalLoad(L);
1914 // fast print dep, using operator<< on instruction is too slow.
1915 dbgs() << "GVN: load ";
1916 L->printAsOperand(dbgs());
1917 dbgs() << " has unknown dependence\n";
1922 Instruction *DepInst = Dep.getInst();
1923 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1924 Value *StoredVal = DepSI->getValueOperand();
1926 // The store and load are to a must-aliased pointer, but they may not
1927 // actually have the same type. See if we know how to reuse the stored
1928 // value (depending on its type).
1929 if (StoredVal->getType() != L->getType()) {
1931 CoerceAvailableValueToLoadType(StoredVal, L->getType(), L, DL);
1935 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1936 << '\n' << *L << "\n\n\n");
1940 L->replaceAllUsesWith(StoredVal);
1941 if (StoredVal->getType()->getScalarType()->isPointerTy())
1942 MD->invalidateCachedPointerInfo(StoredVal);
1943 markInstructionForDeletion(L);
1948 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1949 Value *AvailableVal = DepLI;
1951 // The loads are of a must-aliased pointer, but they may not actually have
1952 // the same type. See if we know how to reuse the previously loaded value
1953 // (depending on its type).
1954 if (DepLI->getType() != L->getType()) {
1955 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(), L, DL);
1959 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1960 << "\n" << *L << "\n\n\n");
1964 patchAndReplaceAllUsesWith(L, AvailableVal);
1965 if (DepLI->getType()->getScalarType()->isPointerTy())
1966 MD->invalidateCachedPointerInfo(DepLI);
1967 markInstructionForDeletion(L);
1972 // If this load really doesn't depend on anything, then we must be loading an
1973 // undef value. This can happen when loading for a fresh allocation with no
1974 // intervening stores, for example.
1975 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
1976 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1977 markInstructionForDeletion(L);
1982 // If this load occurs either right after a lifetime begin,
1983 // then the loaded value is undefined.
1984 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
1985 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1986 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1987 markInstructionForDeletion(L);
1993 // If this load follows a calloc (which zero initializes memory),
1994 // then the loaded value is zero
1995 if (isCallocLikeFn(DepInst, TLI)) {
1996 L->replaceAllUsesWith(Constant::getNullValue(L->getType()));
1997 markInstructionForDeletion(L);
2005 // In order to find a leader for a given value number at a
2006 // specific basic block, we first obtain the list of all Values for that number,
2007 // and then scan the list to find one whose block dominates the block in
2008 // question. This is fast because dominator tree queries consist of only
2009 // a few comparisons of DFS numbers.
2010 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
2011 LeaderTableEntry Vals = LeaderTable[num];
2012 if (!Vals.Val) return nullptr;
2014 Value *Val = nullptr;
2015 if (DT->dominates(Vals.BB, BB)) {
2017 if (isa<Constant>(Val)) return Val;
2020 LeaderTableEntry* Next = Vals.Next;
2022 if (DT->dominates(Next->BB, BB)) {
2023 if (isa<Constant>(Next->Val)) return Next->Val;
2024 if (!Val) Val = Next->Val;
2033 /// Replace all uses of 'From' with 'To' if the use is dominated by the given
2034 /// basic block. Returns the number of uses that were replaced.
2035 unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To,
2036 const BasicBlockEdge &Root) {
2038 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
2042 if (DT->dominates(Root, U)) {
2050 /// There is an edge from 'Src' to 'Dst'. Return
2051 /// true if every path from the entry block to 'Dst' passes via this edge. In
2052 /// particular 'Dst' must not be reachable via another edge from 'Src'.
2053 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
2054 DominatorTree *DT) {
2055 // While in theory it is interesting to consider the case in which Dst has
2056 // more than one predecessor, because Dst might be part of a loop which is
2057 // only reachable from Src, in practice it is pointless since at the time
2058 // GVN runs all such loops have preheaders, which means that Dst will have
2059 // been changed to have only one predecessor, namely Src.
2060 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
2061 const BasicBlock *Src = E.getStart();
2062 assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
2064 return Pred != nullptr;
2067 /// The given values are known to be equal in every block
2068 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
2069 /// 'RHS' everywhere in the scope. Returns whether a change was made.
2070 bool GVN::propagateEquality(Value *LHS, Value *RHS,
2071 const BasicBlockEdge &Root) {
2072 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
2073 Worklist.push_back(std::make_pair(LHS, RHS));
2074 bool Changed = false;
2075 // For speed, compute a conservative fast approximation to
2076 // DT->dominates(Root, Root.getEnd());
2077 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
2079 while (!Worklist.empty()) {
2080 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
2081 LHS = Item.first; RHS = Item.second;
2083 if (LHS == RHS) continue;
2084 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
2086 // Don't try to propagate equalities between constants.
2087 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue;
2089 // Prefer a constant on the right-hand side, or an Argument if no constants.
2090 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
2091 std::swap(LHS, RHS);
2092 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
2094 // If there is no obvious reason to prefer the left-hand side over the
2095 // right-hand side, ensure the longest lived term is on the right-hand side,
2096 // so the shortest lived term will be replaced by the longest lived.
2097 // This tends to expose more simplifications.
2098 uint32_t LVN = VN.lookup_or_add(LHS);
2099 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
2100 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
2101 // Move the 'oldest' value to the right-hand side, using the value number
2102 // as a proxy for age.
2103 uint32_t RVN = VN.lookup_or_add(RHS);
2105 std::swap(LHS, RHS);
2110 // If value numbering later sees that an instruction in the scope is equal
2111 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
2112 // the invariant that instructions only occur in the leader table for their
2113 // own value number (this is used by removeFromLeaderTable), do not do this
2114 // if RHS is an instruction (if an instruction in the scope is morphed into
2115 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
2116 // using the leader table is about compiling faster, not optimizing better).
2117 // The leader table only tracks basic blocks, not edges. Only add to if we
2118 // have the simple case where the edge dominates the end.
2119 if (RootDominatesEnd && !isa<Instruction>(RHS))
2120 addToLeaderTable(LVN, RHS, Root.getEnd());
2122 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
2123 // LHS always has at least one use that is not dominated by Root, this will
2124 // never do anything if LHS has only one use.
2125 if (!LHS->hasOneUse()) {
2126 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root);
2127 Changed |= NumReplacements > 0;
2128 NumGVNEqProp += NumReplacements;
2131 // Now try to deduce additional equalities from this one. For example, if
2132 // the known equality was "(A != B)" == "false" then it follows that A and B
2133 // are equal in the scope. Only boolean equalities with an explicit true or
2134 // false RHS are currently supported.
2135 if (!RHS->getType()->isIntegerTy(1))
2136 // Not a boolean equality - bail out.
2138 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
2140 // RHS neither 'true' nor 'false' - bail out.
2142 // Whether RHS equals 'true'. Otherwise it equals 'false'.
2143 bool isKnownTrue = CI->isAllOnesValue();
2144 bool isKnownFalse = !isKnownTrue;
2146 // If "A && B" is known true then both A and B are known true. If "A || B"
2147 // is known false then both A and B are known false.
2149 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
2150 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
2151 Worklist.push_back(std::make_pair(A, RHS));
2152 Worklist.push_back(std::make_pair(B, RHS));
2156 // If we are propagating an equality like "(A == B)" == "true" then also
2157 // propagate the equality A == B. When propagating a comparison such as
2158 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
2159 if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
2160 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
2162 // If "A == B" is known true, or "A != B" is known false, then replace
2163 // A with B everywhere in the scope.
2164 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2165 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2166 Worklist.push_back(std::make_pair(Op0, Op1));
2168 // Handle the floating point versions of equality comparisons too.
2169 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
2170 (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
2172 // Floating point -0.0 and 0.0 compare equal, so we can only
2173 // propagate values if we know that we have a constant and that
2174 // its value is non-zero.
2176 // FIXME: We should do this optimization if 'no signed zeros' is
2177 // applicable via an instruction-level fast-math-flag or some other
2178 // indicator that relaxed FP semantics are being used.
2180 if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
2181 Worklist.push_back(std::make_pair(Op0, Op1));
2184 // If "A >= B" is known true, replace "A < B" with false everywhere.
2185 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2186 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2187 // Since we don't have the instruction "A < B" immediately to hand, work
2188 // out the value number that it would have and use that to find an
2189 // appropriate instruction (if any).
2190 uint32_t NextNum = VN.getNextUnusedValueNumber();
2191 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2192 // If the number we were assigned was brand new then there is no point in
2193 // looking for an instruction realizing it: there cannot be one!
2194 if (Num < NextNum) {
2195 Value *NotCmp = findLeader(Root.getEnd(), Num);
2196 if (NotCmp && isa<Instruction>(NotCmp)) {
2197 unsigned NumReplacements =
2198 replaceAllDominatedUsesWith(NotCmp, NotVal, Root);
2199 Changed |= NumReplacements > 0;
2200 NumGVNEqProp += NumReplacements;
2203 // Ensure that any instruction in scope that gets the "A < B" value number
2204 // is replaced with false.
2205 // The leader table only tracks basic blocks, not edges. Only add to if we
2206 // have the simple case where the edge dominates the end.
2207 if (RootDominatesEnd)
2208 addToLeaderTable(Num, NotVal, Root.getEnd());
2217 /// When calculating availability, handle an instruction
2218 /// by inserting it into the appropriate sets
2219 bool GVN::processInstruction(Instruction *I) {
2220 // Ignore dbg info intrinsics.
2221 if (isa<DbgInfoIntrinsic>(I))
2224 // If the instruction can be easily simplified then do so now in preference
2225 // to value numbering it. Value numbering often exposes redundancies, for
2226 // example if it determines that %y is equal to %x then the instruction
2227 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2228 const DataLayout &DL = I->getModule()->getDataLayout();
2229 if (Value *V = SimplifyInstruction(I, DL, TLI, DT, AC)) {
2230 I->replaceAllUsesWith(V);
2231 if (MD && V->getType()->getScalarType()->isPointerTy())
2232 MD->invalidateCachedPointerInfo(V);
2233 markInstructionForDeletion(I);
2238 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2239 if (processLoad(LI))
2242 unsigned Num = VN.lookup_or_add(LI);
2243 addToLeaderTable(Num, LI, LI->getParent());
2247 // For conditional branches, we can perform simple conditional propagation on
2248 // the condition value itself.
2249 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2250 if (!BI->isConditional())
2253 if (isa<Constant>(BI->getCondition()))
2254 return processFoldableCondBr(BI);
2256 Value *BranchCond = BI->getCondition();
2257 BasicBlock *TrueSucc = BI->getSuccessor(0);
2258 BasicBlock *FalseSucc = BI->getSuccessor(1);
2259 // Avoid multiple edges early.
2260 if (TrueSucc == FalseSucc)
2263 BasicBlock *Parent = BI->getParent();
2264 bool Changed = false;
2266 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
2267 BasicBlockEdge TrueE(Parent, TrueSucc);
2268 Changed |= propagateEquality(BranchCond, TrueVal, TrueE);
2270 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
2271 BasicBlockEdge FalseE(Parent, FalseSucc);
2272 Changed |= propagateEquality(BranchCond, FalseVal, FalseE);
2277 // For switches, propagate the case values into the case destinations.
2278 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2279 Value *SwitchCond = SI->getCondition();
2280 BasicBlock *Parent = SI->getParent();
2281 bool Changed = false;
2283 // Remember how many outgoing edges there are to every successor.
2284 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2285 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
2286 ++SwitchEdges[SI->getSuccessor(i)];
2288 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2290 BasicBlock *Dst = i.getCaseSuccessor();
2291 // If there is only a single edge, propagate the case value into it.
2292 if (SwitchEdges.lookup(Dst) == 1) {
2293 BasicBlockEdge E(Parent, Dst);
2294 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E);
2300 // Instructions with void type don't return a value, so there's
2301 // no point in trying to find redundancies in them.
2302 if (I->getType()->isVoidTy()) return false;
2304 uint32_t NextNum = VN.getNextUnusedValueNumber();
2305 unsigned Num = VN.lookup_or_add(I);
2307 // Allocations are always uniquely numbered, so we can save time and memory
2308 // by fast failing them.
2309 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2310 addToLeaderTable(Num, I, I->getParent());
2314 // If the number we were assigned was a brand new VN, then we don't
2315 // need to do a lookup to see if the number already exists
2316 // somewhere in the domtree: it can't!
2317 if (Num >= NextNum) {
2318 addToLeaderTable(Num, I, I->getParent());
2322 // Perform fast-path value-number based elimination of values inherited from
2324 Value *repl = findLeader(I->getParent(), Num);
2326 // Failure, just remember this instance for future use.
2327 addToLeaderTable(Num, I, I->getParent());
2332 patchAndReplaceAllUsesWith(I, repl);
2333 if (MD && repl->getType()->getScalarType()->isPointerTy())
2334 MD->invalidateCachedPointerInfo(repl);
2335 markInstructionForDeletion(I);
2339 /// runOnFunction - This is the main transformation entry point for a function.
2340 bool GVN::runOnFunction(Function& F) {
2341 if (skipOptnoneFunction(F))
2345 MD = &getAnalysis<MemoryDependenceAnalysis>();
2346 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2347 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2348 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
2349 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2353 bool Changed = false;
2354 bool ShouldContinue = true;
2356 // Merge unconditional branches, allowing PRE to catch more
2357 // optimization opportunities.
2358 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2359 BasicBlock *BB = FI++;
2361 bool removedBlock = MergeBlockIntoPredecessor(
2362 BB, DT, /* LoopInfo */ nullptr, VN.getAliasAnalysis(), MD);
2363 if (removedBlock) ++NumGVNBlocks;
2365 Changed |= removedBlock;
2368 unsigned Iteration = 0;
2369 while (ShouldContinue) {
2370 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2371 ShouldContinue = iterateOnFunction(F);
2372 Changed |= ShouldContinue;
2377 // Fabricate val-num for dead-code in order to suppress assertion in
2379 assignValNumForDeadCode();
2380 bool PREChanged = true;
2381 while (PREChanged) {
2382 PREChanged = performPRE(F);
2383 Changed |= PREChanged;
2387 // FIXME: Should perform GVN again after PRE does something. PRE can move
2388 // computations into blocks where they become fully redundant. Note that
2389 // we can't do this until PRE's critical edge splitting updates memdep.
2390 // Actually, when this happens, we should just fully integrate PRE into GVN.
2392 cleanupGlobalSets();
2393 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2401 bool GVN::processBlock(BasicBlock *BB) {
2402 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2403 // (and incrementing BI before processing an instruction).
2404 assert(InstrsToErase.empty() &&
2405 "We expect InstrsToErase to be empty across iterations");
2406 if (DeadBlocks.count(BB))
2409 bool ChangedFunction = false;
2411 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2413 ChangedFunction |= processInstruction(BI);
2414 if (InstrsToErase.empty()) {
2419 // If we need some instructions deleted, do it now.
2420 NumGVNInstr += InstrsToErase.size();
2422 // Avoid iterator invalidation.
2423 bool AtStart = BI == BB->begin();
2427 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2428 E = InstrsToErase.end(); I != E; ++I) {
2429 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2430 if (MD) MD->removeInstruction(*I);
2431 DEBUG(verifyRemoved(*I));
2432 (*I)->eraseFromParent();
2434 InstrsToErase.clear();
2442 return ChangedFunction;
2445 // Instantiate an expression in a predecessor that lacked it.
2446 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2447 unsigned int ValNo) {
2448 // Because we are going top-down through the block, all value numbers
2449 // will be available in the predecessor by the time we need them. Any
2450 // that weren't originally present will have been instantiated earlier
2452 bool success = true;
2453 for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2454 Value *Op = Instr->getOperand(i);
2455 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2458 if (Value *V = findLeader(Pred, VN.lookup(Op))) {
2459 Instr->setOperand(i, V);
2466 // Fail out if we encounter an operand that is not available in
2467 // the PRE predecessor. This is typically because of loads which
2468 // are not value numbered precisely.
2472 Instr->insertBefore(Pred->getTerminator());
2473 Instr->setName(Instr->getName() + ".pre");
2474 Instr->setDebugLoc(Instr->getDebugLoc());
2475 VN.add(Instr, ValNo);
2477 // Update the availability map to include the new instruction.
2478 addToLeaderTable(ValNo, Instr, Pred);
2482 bool GVN::performScalarPRE(Instruction *CurInst) {
2483 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap;
2485 if (isa<AllocaInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
2486 isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2487 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2488 isa<DbgInfoIntrinsic>(CurInst))
2491 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2492 // sinking the compare again, and it would force the code generator to
2493 // move the i1 from processor flags or predicate registers into a general
2494 // purpose register.
2495 if (isa<CmpInst>(CurInst))
2498 // We don't currently value number ANY inline asm calls.
2499 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2500 if (CallI->isInlineAsm())
2503 uint32_t ValNo = VN.lookup(CurInst);
2505 // Look for the predecessors for PRE opportunities. We're
2506 // only trying to solve the basic diamond case, where
2507 // a value is computed in the successor and one predecessor,
2508 // but not the other. We also explicitly disallow cases
2509 // where the successor is its own predecessor, because they're
2510 // more complicated to get right.
2511 unsigned NumWith = 0;
2512 unsigned NumWithout = 0;
2513 BasicBlock *PREPred = nullptr;
2514 BasicBlock *CurrentBlock = CurInst->getParent();
2517 for (pred_iterator PI = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock);
2519 BasicBlock *P = *PI;
2520 // We're not interested in PRE where the block is its
2521 // own predecessor, or in blocks with predecessors
2522 // that are not reachable.
2523 if (P == CurrentBlock) {
2526 } else if (!DT->isReachableFromEntry(P)) {
2531 Value *predV = findLeader(P, ValNo);
2533 predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2536 } else if (predV == CurInst) {
2537 /* CurInst dominates this predecessor. */
2541 predMap.push_back(std::make_pair(predV, P));
2546 // Don't do PRE when it might increase code size, i.e. when
2547 // we would need to insert instructions in more than one pred.
2548 if (NumWithout > 1 || NumWith == 0)
2551 // We may have a case where all predecessors have the instruction,
2552 // and we just need to insert a phi node. Otherwise, perform
2554 Instruction *PREInstr = nullptr;
2556 if (NumWithout != 0) {
2557 // Don't do PRE across indirect branch.
2558 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2561 // We can't do PRE safely on a critical edge, so instead we schedule
2562 // the edge to be split and perform the PRE the next time we iterate
2564 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2565 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2566 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2569 // We need to insert somewhere, so let's give it a shot
2570 PREInstr = CurInst->clone();
2571 if (!performScalarPREInsertion(PREInstr, PREPred, ValNo)) {
2572 // If we failed insertion, make sure we remove the instruction.
2573 DEBUG(verifyRemoved(PREInstr));
2579 // Either we should have filled in the PRE instruction, or we should
2580 // not have needed insertions.
2581 assert (PREInstr != nullptr || NumWithout == 0);
2585 // Create a PHI to make the value available in this block.
2587 PHINode::Create(CurInst->getType(), predMap.size(),
2588 CurInst->getName() + ".pre-phi", CurrentBlock->begin());
2589 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2590 if (Value *V = predMap[i].first)
2591 Phi->addIncoming(V, predMap[i].second);
2593 Phi->addIncoming(PREInstr, PREPred);
2597 addToLeaderTable(ValNo, Phi, CurrentBlock);
2598 Phi->setDebugLoc(CurInst->getDebugLoc());
2599 CurInst->replaceAllUsesWith(Phi);
2600 if (Phi->getType()->getScalarType()->isPointerTy()) {
2601 // Because we have added a PHI-use of the pointer value, it has now
2602 // "escaped" from alias analysis' perspective. We need to inform
2604 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; ++ii) {
2605 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
2606 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
2610 MD->invalidateCachedPointerInfo(Phi);
2613 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2615 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2617 MD->removeInstruction(CurInst);
2618 DEBUG(verifyRemoved(CurInst));
2619 CurInst->eraseFromParent();
2625 /// Perform a purely local form of PRE that looks for diamond
2626 /// control flow patterns and attempts to perform simple PRE at the join point.
2627 bool GVN::performPRE(Function &F) {
2628 bool Changed = false;
2629 for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2630 // Nothing to PRE in the entry block.
2631 if (CurrentBlock == &F.getEntryBlock())
2634 // Don't perform PRE on a landing pad.
2635 if (CurrentBlock->isLandingPad())
2638 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2639 BE = CurrentBlock->end();
2641 Instruction *CurInst = BI++;
2642 Changed = performScalarPRE(CurInst);
2646 if (splitCriticalEdges())
2652 /// Split the critical edge connecting the given two blocks, and return
2653 /// the block inserted to the critical edge.
2654 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2655 BasicBlock *BB = SplitCriticalEdge(
2656 Pred, Succ, CriticalEdgeSplittingOptions(getAliasAnalysis(), DT));
2658 MD->invalidateCachedPredecessors();
2662 /// Split critical edges found during the previous
2663 /// iteration that may enable further optimization.
2664 bool GVN::splitCriticalEdges() {
2665 if (toSplit.empty())
2668 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2669 SplitCriticalEdge(Edge.first, Edge.second,
2670 CriticalEdgeSplittingOptions(getAliasAnalysis(), DT));
2671 } while (!toSplit.empty());
2672 if (MD) MD->invalidateCachedPredecessors();
2676 /// Executes one iteration of GVN
2677 bool GVN::iterateOnFunction(Function &F) {
2678 cleanupGlobalSets();
2680 // Top-down walk of the dominator tree
2681 bool Changed = false;
2682 // Save the blocks this function have before transformation begins. GVN may
2683 // split critical edge, and hence may invalidate the RPO/DT iterator.
2685 std::vector<BasicBlock *> BBVect;
2686 BBVect.reserve(256);
2687 // Needed for value numbering with phi construction to work.
2688 ReversePostOrderTraversal<Function *> RPOT(&F);
2689 for (ReversePostOrderTraversal<Function *>::rpo_iterator RI = RPOT.begin(),
2692 BBVect.push_back(*RI);
2694 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
2696 Changed |= processBlock(*I);
2701 void GVN::cleanupGlobalSets() {
2703 LeaderTable.clear();
2704 TableAllocator.Reset();
2707 /// Verify that the specified instruction does not occur in our
2708 /// internal data structures.
2709 void GVN::verifyRemoved(const Instruction *Inst) const {
2710 VN.verifyRemoved(Inst);
2712 // Walk through the value number scope to make sure the instruction isn't
2713 // ferreted away in it.
2714 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2715 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2716 const LeaderTableEntry *Node = &I->second;
2717 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2719 while (Node->Next) {
2721 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2726 /// BB is declared dead, which implied other blocks become dead as well. This
2727 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2728 /// live successors, update their phi nodes by replacing the operands
2729 /// corresponding to dead blocks with UndefVal.
2730 void GVN::addDeadBlock(BasicBlock *BB) {
2731 SmallVector<BasicBlock *, 4> NewDead;
2732 SmallSetVector<BasicBlock *, 4> DF;
2734 NewDead.push_back(BB);
2735 while (!NewDead.empty()) {
2736 BasicBlock *D = NewDead.pop_back_val();
2737 if (DeadBlocks.count(D))
2740 // All blocks dominated by D are dead.
2741 SmallVector<BasicBlock *, 8> Dom;
2742 DT->getDescendants(D, Dom);
2743 DeadBlocks.insert(Dom.begin(), Dom.end());
2745 // Figure out the dominance-frontier(D).
2746 for (SmallVectorImpl<BasicBlock *>::iterator I = Dom.begin(),
2747 E = Dom.end(); I != E; I++) {
2749 for (succ_iterator SI = succ_begin(B), SE = succ_end(B); SI != SE; SI++) {
2750 BasicBlock *S = *SI;
2751 if (DeadBlocks.count(S))
2754 bool AllPredDead = true;
2755 for (pred_iterator PI = pred_begin(S), PE = pred_end(S); PI != PE; PI++)
2756 if (!DeadBlocks.count(*PI)) {
2757 AllPredDead = false;
2762 // S could be proved dead later on. That is why we don't update phi
2763 // operands at this moment.
2766 // While S is not dominated by D, it is dead by now. This could take
2767 // place if S already have a dead predecessor before D is declared
2769 NewDead.push_back(S);
2775 // For the dead blocks' live successors, update their phi nodes by replacing
2776 // the operands corresponding to dead blocks with UndefVal.
2777 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2780 if (DeadBlocks.count(B))
2783 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2784 for (SmallVectorImpl<BasicBlock *>::iterator PI = Preds.begin(),
2785 PE = Preds.end(); PI != PE; PI++) {
2786 BasicBlock *P = *PI;
2788 if (!DeadBlocks.count(P))
2791 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2792 if (BasicBlock *S = splitCriticalEdges(P, B))
2793 DeadBlocks.insert(P = S);
2796 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2797 PHINode &Phi = cast<PHINode>(*II);
2798 Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2799 UndefValue::get(Phi.getType()));
2805 // If the given branch is recognized as a foldable branch (i.e. conditional
2806 // branch with constant condition), it will perform following analyses and
2808 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2809 // R be the target of the dead out-coming edge.
2810 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2811 // edge. The result of this step will be {X| X is dominated by R}
2812 // 2) Identify those blocks which haves at least one dead prodecessor. The
2813 // result of this step will be dominance-frontier(R).
2814 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2815 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2817 // Return true iff *NEW* dead code are found.
2818 bool GVN::processFoldableCondBr(BranchInst *BI) {
2819 if (!BI || BI->isUnconditional())
2822 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2826 BasicBlock *DeadRoot = Cond->getZExtValue() ?
2827 BI->getSuccessor(1) : BI->getSuccessor(0);
2828 if (DeadBlocks.count(DeadRoot))
2831 if (!DeadRoot->getSinglePredecessor())
2832 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2834 addDeadBlock(DeadRoot);
2838 // performPRE() will trigger assert if it comes across an instruction without
2839 // associated val-num. As it normally has far more live instructions than dead
2840 // instructions, it makes more sense just to "fabricate" a val-number for the
2841 // dead code than checking if instruction involved is dead or not.
2842 void GVN::assignValNumForDeadCode() {
2843 for (SetVector<BasicBlock *>::iterator I = DeadBlocks.begin(),
2844 E = DeadBlocks.end(); I != E; I++) {
2845 BasicBlock *BB = *I;
2846 for (BasicBlock::iterator II = BB->begin(), EE = BB->end();
2848 Instruction *Inst = &*II;
2849 unsigned ValNum = VN.lookup_or_add(Inst);
2850 addToLeaderTable(ValNum, Inst, BB);