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/TargetLibraryInfo.h"
37 #include "llvm/Analysis/ValueTracking.h"
38 #include "llvm/IR/DataLayout.h"
39 #include "llvm/IR/Dominators.h"
40 #include "llvm/IR/GlobalVariable.h"
41 #include "llvm/IR/IRBuilder.h"
42 #include "llvm/IR/IntrinsicInst.h"
43 #include "llvm/IR/LLVMContext.h"
44 #include "llvm/IR/Metadata.h"
45 #include "llvm/IR/PatternMatch.h"
46 #include "llvm/Support/Allocator.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/raw_ostream.h"
50 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
51 #include "llvm/Transforms/Utils/Local.h"
52 #include "llvm/Transforms/Utils/SSAUpdater.h"
55 using namespace PatternMatch;
57 #define DEBUG_TYPE "gvn"
59 STATISTIC(NumGVNInstr, "Number of instructions deleted");
60 STATISTIC(NumGVNLoad, "Number of loads deleted");
61 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
62 STATISTIC(NumGVNBlocks, "Number of blocks merged");
63 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
64 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
65 STATISTIC(NumPRELoad, "Number of loads PRE'd");
67 static cl::opt<bool> EnablePRE("enable-pre",
68 cl::init(true), cl::Hidden);
69 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
71 // Maximum allowed recursion depth.
72 static cl::opt<uint32_t>
73 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
74 cl::desc("Max recurse depth (default = 1000)"));
76 //===----------------------------------------------------------------------===//
78 //===----------------------------------------------------------------------===//
80 /// This class holds the mapping between values and value numbers. It is used
81 /// as an efficient mechanism to determine the expression-wise equivalence of
87 SmallVector<uint32_t, 4> varargs;
89 Expression(uint32_t o = ~2U) : opcode(o) { }
91 bool operator==(const Expression &other) const {
92 if (opcode != other.opcode)
94 if (opcode == ~0U || opcode == ~1U)
96 if (type != other.type)
98 if (varargs != other.varargs)
103 friend hash_code hash_value(const Expression &Value) {
104 return hash_combine(Value.opcode, Value.type,
105 hash_combine_range(Value.varargs.begin(),
106 Value.varargs.end()));
111 DenseMap<Value*, uint32_t> valueNumbering;
112 DenseMap<Expression, uint32_t> expressionNumbering;
114 MemoryDependenceAnalysis *MD;
117 uint32_t nextValueNumber;
119 Expression create_expression(Instruction* I);
120 Expression create_cmp_expression(unsigned Opcode,
121 CmpInst::Predicate Predicate,
122 Value *LHS, Value *RHS);
123 Expression create_extractvalue_expression(ExtractValueInst* EI);
124 uint32_t lookup_or_add_call(CallInst* C);
126 ValueTable() : nextValueNumber(1) { }
127 uint32_t lookup_or_add(Value *V);
128 uint32_t lookup(Value *V) const;
129 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred,
130 Value *LHS, Value *RHS);
131 void add(Value *V, uint32_t num);
133 void erase(Value *v);
134 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
135 AliasAnalysis *getAliasAnalysis() const { return AA; }
136 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
137 void setDomTree(DominatorTree* D) { DT = D; }
138 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
139 void verifyRemoved(const Value *) const;
144 template <> struct DenseMapInfo<Expression> {
145 static inline Expression getEmptyKey() {
149 static inline Expression getTombstoneKey() {
153 static unsigned getHashValue(const Expression e) {
154 using llvm::hash_value;
155 return static_cast<unsigned>(hash_value(e));
157 static bool isEqual(const Expression &LHS, const Expression &RHS) {
164 //===----------------------------------------------------------------------===//
165 // ValueTable Internal Functions
166 //===----------------------------------------------------------------------===//
168 Expression ValueTable::create_expression(Instruction *I) {
170 e.type = I->getType();
171 e.opcode = I->getOpcode();
172 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
174 e.varargs.push_back(lookup_or_add(*OI));
175 if (I->isCommutative()) {
176 // Ensure that commutative instructions that only differ by a permutation
177 // of their operands get the same value number by sorting the operand value
178 // numbers. Since all commutative instructions have two operands it is more
179 // efficient to sort by hand rather than using, say, std::sort.
180 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
181 if (e.varargs[0] > e.varargs[1])
182 std::swap(e.varargs[0], e.varargs[1]);
185 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
186 // Sort the operand value numbers so x<y and y>x get the same value number.
187 CmpInst::Predicate Predicate = C->getPredicate();
188 if (e.varargs[0] > e.varargs[1]) {
189 std::swap(e.varargs[0], e.varargs[1]);
190 Predicate = CmpInst::getSwappedPredicate(Predicate);
192 e.opcode = (C->getOpcode() << 8) | Predicate;
193 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
194 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
196 e.varargs.push_back(*II);
202 Expression ValueTable::create_cmp_expression(unsigned Opcode,
203 CmpInst::Predicate Predicate,
204 Value *LHS, Value *RHS) {
205 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
206 "Not a comparison!");
208 e.type = CmpInst::makeCmpResultType(LHS->getType());
209 e.varargs.push_back(lookup_or_add(LHS));
210 e.varargs.push_back(lookup_or_add(RHS));
212 // Sort the operand value numbers so x<y and y>x get the same value number.
213 if (e.varargs[0] > e.varargs[1]) {
214 std::swap(e.varargs[0], e.varargs[1]);
215 Predicate = CmpInst::getSwappedPredicate(Predicate);
217 e.opcode = (Opcode << 8) | Predicate;
221 Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) {
222 assert(EI && "Not an ExtractValueInst?");
224 e.type = EI->getType();
227 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
228 if (I != nullptr && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
229 // EI might be an extract from one of our recognised intrinsics. If it
230 // is we'll synthesize a semantically equivalent expression instead on
231 // an extract value expression.
232 switch (I->getIntrinsicID()) {
233 case Intrinsic::sadd_with_overflow:
234 case Intrinsic::uadd_with_overflow:
235 e.opcode = Instruction::Add;
237 case Intrinsic::ssub_with_overflow:
238 case Intrinsic::usub_with_overflow:
239 e.opcode = Instruction::Sub;
241 case Intrinsic::smul_with_overflow:
242 case Intrinsic::umul_with_overflow:
243 e.opcode = Instruction::Mul;
250 // Intrinsic recognized. Grab its args to finish building the expression.
251 assert(I->getNumArgOperands() == 2 &&
252 "Expect two args for recognised intrinsics.");
253 e.varargs.push_back(lookup_or_add(I->getArgOperand(0)));
254 e.varargs.push_back(lookup_or_add(I->getArgOperand(1)));
259 // Not a recognised intrinsic. Fall back to producing an extract value
261 e.opcode = EI->getOpcode();
262 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
264 e.varargs.push_back(lookup_or_add(*OI));
266 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
268 e.varargs.push_back(*II);
273 //===----------------------------------------------------------------------===//
274 // ValueTable External Functions
275 //===----------------------------------------------------------------------===//
277 /// add - Insert a value into the table with a specified value number.
278 void ValueTable::add(Value *V, uint32_t num) {
279 valueNumbering.insert(std::make_pair(V, num));
282 uint32_t ValueTable::lookup_or_add_call(CallInst *C) {
283 if (AA->doesNotAccessMemory(C)) {
284 Expression exp = create_expression(C);
285 uint32_t &e = expressionNumbering[exp];
286 if (!e) e = nextValueNumber++;
287 valueNumbering[C] = e;
289 } else if (AA->onlyReadsMemory(C)) {
290 Expression exp = create_expression(C);
291 uint32_t &e = expressionNumbering[exp];
293 e = nextValueNumber++;
294 valueNumbering[C] = e;
298 e = nextValueNumber++;
299 valueNumbering[C] = e;
303 MemDepResult local_dep = MD->getDependency(C);
305 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
306 valueNumbering[C] = nextValueNumber;
307 return nextValueNumber++;
310 if (local_dep.isDef()) {
311 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
313 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
314 valueNumbering[C] = nextValueNumber;
315 return nextValueNumber++;
318 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
319 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
320 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
322 valueNumbering[C] = nextValueNumber;
323 return nextValueNumber++;
327 uint32_t v = lookup_or_add(local_cdep);
328 valueNumbering[C] = v;
333 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
334 MD->getNonLocalCallDependency(CallSite(C));
335 // FIXME: Move the checking logic to MemDep!
336 CallInst* cdep = nullptr;
338 // Check to see if we have a single dominating call instruction that is
340 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
341 const NonLocalDepEntry *I = &deps[i];
342 if (I->getResult().isNonLocal())
345 // We don't handle non-definitions. If we already have a call, reject
346 // instruction dependencies.
347 if (!I->getResult().isDef() || cdep != nullptr) {
352 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
353 // FIXME: All duplicated with non-local case.
354 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
355 cdep = NonLocalDepCall;
364 valueNumbering[C] = nextValueNumber;
365 return nextValueNumber++;
368 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
369 valueNumbering[C] = nextValueNumber;
370 return nextValueNumber++;
372 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
373 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
374 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
376 valueNumbering[C] = nextValueNumber;
377 return nextValueNumber++;
381 uint32_t v = lookup_or_add(cdep);
382 valueNumbering[C] = v;
386 valueNumbering[C] = nextValueNumber;
387 return nextValueNumber++;
391 /// lookup_or_add - Returns the value number for the specified value, assigning
392 /// it a new number if it did not have one before.
393 uint32_t ValueTable::lookup_or_add(Value *V) {
394 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
395 if (VI != valueNumbering.end())
398 if (!isa<Instruction>(V)) {
399 valueNumbering[V] = nextValueNumber;
400 return nextValueNumber++;
403 Instruction* I = cast<Instruction>(V);
405 switch (I->getOpcode()) {
406 case Instruction::Call:
407 return lookup_or_add_call(cast<CallInst>(I));
408 case Instruction::Add:
409 case Instruction::FAdd:
410 case Instruction::Sub:
411 case Instruction::FSub:
412 case Instruction::Mul:
413 case Instruction::FMul:
414 case Instruction::UDiv:
415 case Instruction::SDiv:
416 case Instruction::FDiv:
417 case Instruction::URem:
418 case Instruction::SRem:
419 case Instruction::FRem:
420 case Instruction::Shl:
421 case Instruction::LShr:
422 case Instruction::AShr:
423 case Instruction::And:
424 case Instruction::Or:
425 case Instruction::Xor:
426 case Instruction::ICmp:
427 case Instruction::FCmp:
428 case Instruction::Trunc:
429 case Instruction::ZExt:
430 case Instruction::SExt:
431 case Instruction::FPToUI:
432 case Instruction::FPToSI:
433 case Instruction::UIToFP:
434 case Instruction::SIToFP:
435 case Instruction::FPTrunc:
436 case Instruction::FPExt:
437 case Instruction::PtrToInt:
438 case Instruction::IntToPtr:
439 case Instruction::BitCast:
440 case Instruction::Select:
441 case Instruction::ExtractElement:
442 case Instruction::InsertElement:
443 case Instruction::ShuffleVector:
444 case Instruction::InsertValue:
445 case Instruction::GetElementPtr:
446 exp = create_expression(I);
448 case Instruction::ExtractValue:
449 exp = create_extractvalue_expression(cast<ExtractValueInst>(I));
452 valueNumbering[V] = nextValueNumber;
453 return nextValueNumber++;
456 uint32_t& e = expressionNumbering[exp];
457 if (!e) e = nextValueNumber++;
458 valueNumbering[V] = e;
462 /// Returns the value number of the specified value. Fails if
463 /// the value has not yet been numbered.
464 uint32_t ValueTable::lookup(Value *V) const {
465 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
466 assert(VI != valueNumbering.end() && "Value not numbered?");
470 /// Returns the value number of the given comparison,
471 /// assigning it a new number if it did not have one before. Useful when
472 /// we deduced the result of a comparison, but don't immediately have an
473 /// instruction realizing that comparison to hand.
474 uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode,
475 CmpInst::Predicate Predicate,
476 Value *LHS, Value *RHS) {
477 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS);
478 uint32_t& e = expressionNumbering[exp];
479 if (!e) e = nextValueNumber++;
483 /// Remove all entries from the ValueTable.
484 void ValueTable::clear() {
485 valueNumbering.clear();
486 expressionNumbering.clear();
490 /// Remove a value from the value numbering.
491 void ValueTable::erase(Value *V) {
492 valueNumbering.erase(V);
495 /// verifyRemoved - Verify that the value is removed from all internal data
497 void ValueTable::verifyRemoved(const Value *V) const {
498 for (DenseMap<Value*, uint32_t>::const_iterator
499 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
500 assert(I->first != V && "Inst still occurs in value numbering map!");
504 //===----------------------------------------------------------------------===//
506 //===----------------------------------------------------------------------===//
510 struct AvailableValueInBlock {
511 /// BB - The basic block in question.
514 SimpleVal, // A simple offsetted value that is accessed.
515 LoadVal, // A value produced by a load.
516 MemIntrin, // A memory intrinsic which is loaded from.
517 UndefVal // A UndefValue representing a value from dead block (which
518 // is not yet physically removed from the CFG).
521 /// V - The value that is live out of the block.
522 PointerIntPair<Value *, 2, ValType> Val;
524 /// Offset - The byte offset in Val that is interesting for the load query.
527 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
528 unsigned Offset = 0) {
529 AvailableValueInBlock Res;
531 Res.Val.setPointer(V);
532 Res.Val.setInt(SimpleVal);
537 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
538 unsigned Offset = 0) {
539 AvailableValueInBlock Res;
541 Res.Val.setPointer(MI);
542 Res.Val.setInt(MemIntrin);
547 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
548 unsigned Offset = 0) {
549 AvailableValueInBlock Res;
551 Res.Val.setPointer(LI);
552 Res.Val.setInt(LoadVal);
557 static AvailableValueInBlock getUndef(BasicBlock *BB) {
558 AvailableValueInBlock Res;
560 Res.Val.setPointer(nullptr);
561 Res.Val.setInt(UndefVal);
566 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
567 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
568 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
569 bool isUndefValue() const { return Val.getInt() == UndefVal; }
571 Value *getSimpleValue() const {
572 assert(isSimpleValue() && "Wrong accessor");
573 return Val.getPointer();
576 LoadInst *getCoercedLoadValue() const {
577 assert(isCoercedLoadValue() && "Wrong accessor");
578 return cast<LoadInst>(Val.getPointer());
581 MemIntrinsic *getMemIntrinValue() const {
582 assert(isMemIntrinValue() && "Wrong accessor");
583 return cast<MemIntrinsic>(Val.getPointer());
586 /// Emit code into this block to adjust the value defined here to the
587 /// specified type. This handles various coercion cases.
588 Value *MaterializeAdjustedValue(LoadInst *LI, GVN &gvn) const;
591 class GVN : public FunctionPass {
593 MemoryDependenceAnalysis *MD;
595 const TargetLibraryInfo *TLI;
597 SetVector<BasicBlock *> DeadBlocks;
601 /// A mapping from value numbers to lists of Value*'s that
602 /// have that value number. Use findLeader to query it.
603 struct LeaderTableEntry {
605 const BasicBlock *BB;
606 LeaderTableEntry *Next;
608 DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
609 BumpPtrAllocator TableAllocator;
611 SmallVector<Instruction*, 8> InstrsToErase;
613 typedef SmallVector<NonLocalDepResult, 64> LoadDepVect;
614 typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect;
615 typedef SmallVector<BasicBlock*, 64> UnavailBlkVect;
618 static char ID; // Pass identification, replacement for typeid
619 explicit GVN(bool noloads = false)
620 : FunctionPass(ID), NoLoads(noloads), MD(nullptr) {
621 initializeGVNPass(*PassRegistry::getPassRegistry());
624 bool runOnFunction(Function &F) override;
626 /// This removes the specified instruction from
627 /// our various maps and marks it for deletion.
628 void markInstructionForDeletion(Instruction *I) {
630 InstrsToErase.push_back(I);
633 DominatorTree &getDominatorTree() const { return *DT; }
634 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
635 MemoryDependenceAnalysis &getMemDep() const { return *MD; }
637 /// Push a new Value to the LeaderTable onto the list for its value number.
638 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) {
639 LeaderTableEntry &Curr = LeaderTable[N];
646 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
649 Node->Next = Curr.Next;
653 /// Scan the list of values corresponding to a given
654 /// value number, and remove the given instruction if encountered.
655 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) {
656 LeaderTableEntry* Prev = nullptr;
657 LeaderTableEntry* Curr = &LeaderTable[N];
659 while (Curr && (Curr->Val != I || Curr->BB != BB)) {
668 Prev->Next = Curr->Next;
674 LeaderTableEntry* Next = Curr->Next;
675 Curr->Val = Next->Val;
677 Curr->Next = Next->Next;
682 // List of critical edges to be split between iterations.
683 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
685 // This transformation requires dominator postdominator info
686 void getAnalysisUsage(AnalysisUsage &AU) const override {
687 AU.addRequired<AssumptionCacheTracker>();
688 AU.addRequired<DominatorTreeWrapperPass>();
689 AU.addRequired<TargetLibraryInfoWrapperPass>();
691 AU.addRequired<MemoryDependenceAnalysis>();
692 AU.addRequired<AliasAnalysis>();
694 AU.addPreserved<DominatorTreeWrapperPass>();
695 AU.addPreserved<AliasAnalysis>();
699 // Helper fuctions of redundant load elimination
700 bool processLoad(LoadInst *L);
701 bool processNonLocalLoad(LoadInst *L);
702 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
703 AvailValInBlkVect &ValuesPerBlock,
704 UnavailBlkVect &UnavailableBlocks);
705 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
706 UnavailBlkVect &UnavailableBlocks);
708 // Other helper routines
709 bool processInstruction(Instruction *I);
710 bool processBlock(BasicBlock *BB);
711 void dump(DenseMap<uint32_t, Value*> &d);
712 bool iterateOnFunction(Function &F);
713 bool performPRE(Function &F);
714 bool performScalarPRE(Instruction *I);
715 bool performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
717 Value *findLeader(const BasicBlock *BB, uint32_t num);
718 void cleanupGlobalSets();
719 void verifyRemoved(const Instruction *I) const;
720 bool splitCriticalEdges();
721 BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ);
722 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root);
723 bool processFoldableCondBr(BranchInst *BI);
724 void addDeadBlock(BasicBlock *BB);
725 void assignValNumForDeadCode();
731 // The public interface to this file...
732 FunctionPass *llvm::createGVNPass(bool NoLoads) {
733 return new GVN(NoLoads);
736 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
737 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
738 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
739 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
740 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
741 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
742 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
744 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
745 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
747 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
748 E = d.end(); I != E; ++I) {
749 errs() << I->first << "\n";
756 /// Return true if we can prove that the value
757 /// we're analyzing is fully available in the specified block. As we go, keep
758 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
759 /// map is actually a tri-state map with the following values:
760 /// 0) we know the block *is not* fully available.
761 /// 1) we know the block *is* fully available.
762 /// 2) we do not know whether the block is fully available or not, but we are
763 /// currently speculating that it will be.
764 /// 3) we are speculating for this block and have used that to speculate for
766 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
767 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
768 uint32_t RecurseDepth) {
769 if (RecurseDepth > MaxRecurseDepth)
772 // Optimistically assume that the block is fully available and check to see
773 // if we already know about this block in one lookup.
774 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
775 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
777 // If the entry already existed for this block, return the precomputed value.
779 // If this is a speculative "available" value, mark it as being used for
780 // speculation of other blocks.
781 if (IV.first->second == 2)
782 IV.first->second = 3;
783 return IV.first->second != 0;
786 // Otherwise, see if it is fully available in all predecessors.
787 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
789 // If this block has no predecessors, it isn't live-in here.
791 goto SpeculationFailure;
793 for (; PI != PE; ++PI)
794 // If the value isn't fully available in one of our predecessors, then it
795 // isn't fully available in this block either. Undo our previous
796 // optimistic assumption and bail out.
797 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
798 goto SpeculationFailure;
802 // If we get here, we found out that this is not, after
803 // all, a fully-available block. We have a problem if we speculated on this and
804 // used the speculation to mark other blocks as available.
806 char &BBVal = FullyAvailableBlocks[BB];
808 // If we didn't speculate on this, just return with it set to false.
814 // If we did speculate on this value, we could have blocks set to 1 that are
815 // incorrect. Walk the (transitive) successors of this block and mark them as
817 SmallVector<BasicBlock*, 32> BBWorklist;
818 BBWorklist.push_back(BB);
821 BasicBlock *Entry = BBWorklist.pop_back_val();
822 // Note that this sets blocks to 0 (unavailable) if they happen to not
823 // already be in FullyAvailableBlocks. This is safe.
824 char &EntryVal = FullyAvailableBlocks[Entry];
825 if (EntryVal == 0) continue; // Already unavailable.
827 // Mark as unavailable.
830 BBWorklist.append(succ_begin(Entry), succ_end(Entry));
831 } while (!BBWorklist.empty());
837 /// Return true if CoerceAvailableValueToLoadType will succeed.
838 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
840 const DataLayout &DL) {
841 // If the loaded or stored value is an first class array or struct, don't try
842 // to transform them. We need to be able to bitcast to integer.
843 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
844 StoredVal->getType()->isStructTy() ||
845 StoredVal->getType()->isArrayTy())
848 // The store has to be at least as big as the load.
849 if (DL.getTypeSizeInBits(StoredVal->getType()) <
850 DL.getTypeSizeInBits(LoadTy))
856 /// If we saw a store of a value to memory, and
857 /// then a load from a must-aliased pointer of a different type, try to coerce
858 /// the stored value. LoadedTy is the type of the load we want to replace.
859 /// IRB is IRBuilder used to insert new instructions.
861 /// If we can't do it, return null.
862 static Value *CoerceAvailableValueToLoadType(Value *StoredVal, Type *LoadedTy,
864 const DataLayout &DL) {
865 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL))
868 // If this is already the right type, just return it.
869 Type *StoredValTy = StoredVal->getType();
871 uint64_t StoreSize = DL.getTypeSizeInBits(StoredValTy);
872 uint64_t LoadSize = DL.getTypeSizeInBits(LoadedTy);
874 // If the store and reload are the same size, we can always reuse it.
875 if (StoreSize == LoadSize) {
876 // Pointer to Pointer -> use bitcast.
877 if (StoredValTy->getScalarType()->isPointerTy() &&
878 LoadedTy->getScalarType()->isPointerTy())
879 return IRB.CreateBitCast(StoredVal, LoadedTy);
881 // Convert source pointers to integers, which can be bitcast.
882 if (StoredValTy->getScalarType()->isPointerTy()) {
883 StoredValTy = DL.getIntPtrType(StoredValTy);
884 StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
887 Type *TypeToCastTo = LoadedTy;
888 if (TypeToCastTo->getScalarType()->isPointerTy())
889 TypeToCastTo = DL.getIntPtrType(TypeToCastTo);
891 if (StoredValTy != TypeToCastTo)
892 StoredVal = IRB.CreateBitCast(StoredVal, TypeToCastTo);
894 // Cast to pointer if the load needs a pointer type.
895 if (LoadedTy->getScalarType()->isPointerTy())
896 StoredVal = IRB.CreateIntToPtr(StoredVal, LoadedTy);
901 // If the loaded value is smaller than the available value, then we can
902 // extract out a piece from it. If the available value is too small, then we
903 // can't do anything.
904 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
906 // Convert source pointers to integers, which can be manipulated.
907 if (StoredValTy->getScalarType()->isPointerTy()) {
908 StoredValTy = DL.getIntPtrType(StoredValTy);
909 StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
912 // Convert vectors and fp to integer, which can be manipulated.
913 if (!StoredValTy->isIntegerTy()) {
914 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
915 StoredVal = IRB.CreateBitCast(StoredVal, StoredValTy);
918 // If this is a big-endian system, we need to shift the value down to the low
919 // bits so that a truncate will work.
920 if (DL.isBigEndian()) {
921 StoredVal = IRB.CreateLShr(StoredVal, StoreSize - LoadSize, "tmp");
924 // Truncate the integer to the right size now.
925 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
926 StoredVal = IRB.CreateTrunc(StoredVal, NewIntTy, "trunc");
928 if (LoadedTy == NewIntTy)
931 // If the result is a pointer, inttoptr.
932 if (LoadedTy->getScalarType()->isPointerTy())
933 return IRB.CreateIntToPtr(StoredVal, LoadedTy, "inttoptr");
935 // Otherwise, bitcast.
936 return IRB.CreateBitCast(StoredVal, LoadedTy, "bitcast");
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(Type::getInt8Ty(Src->getContext()), Src,
1106 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1107 if (ConstantFoldLoadFromConstPtr(Src, DL))
1113 /// This function is called when we have a
1114 /// memdep query of a load that ends up being a clobbering store. This means
1115 /// that the store provides bits used by the load but we the pointers don't
1116 /// mustalias. Check this case to see if there is anything more we can do
1117 /// before we give up.
1118 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1120 Instruction *InsertPt, const DataLayout &DL){
1121 LLVMContext &Ctx = SrcVal->getType()->getContext();
1123 uint64_t StoreSize = (DL.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1124 uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy) + 7) / 8;
1126 IRBuilder<> Builder(InsertPt);
1128 // Compute which bits of the stored value are being used by the load. Convert
1129 // to an integer type to start with.
1130 if (SrcVal->getType()->getScalarType()->isPointerTy())
1131 SrcVal = Builder.CreatePtrToInt(SrcVal,
1132 DL.getIntPtrType(SrcVal->getType()));
1133 if (!SrcVal->getType()->isIntegerTy())
1134 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1136 // Shift the bits to the least significant depending on endianness.
1138 if (DL.isLittleEndian())
1139 ShiftAmt = Offset*8;
1141 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1144 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1146 if (LoadSize != StoreSize)
1147 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1149 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, Builder, DL);
1152 /// This function is called when we have a
1153 /// memdep query of a load that ends up being a clobbering load. This means
1154 /// that the load *may* provide bits used by the load but we can't be sure
1155 /// because the pointers don't mustalias. Check this case to see if there is
1156 /// anything more we can do before we give up.
1157 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1158 Type *LoadTy, Instruction *InsertPt,
1160 const DataLayout &DL = SrcVal->getModule()->getDataLayout();
1161 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1162 // widen SrcVal out to a larger load.
1163 unsigned SrcValSize = DL.getTypeStoreSize(SrcVal->getType());
1164 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1165 if (Offset+LoadSize > SrcValSize) {
1166 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1167 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1168 // If we have a load/load clobber an DepLI can be widened to cover this
1169 // load, then we should widen it to the next power of 2 size big enough!
1170 unsigned NewLoadSize = Offset+LoadSize;
1171 if (!isPowerOf2_32(NewLoadSize))
1172 NewLoadSize = NextPowerOf2(NewLoadSize);
1174 Value *PtrVal = SrcVal->getPointerOperand();
1176 // Insert the new load after the old load. This ensures that subsequent
1177 // memdep queries will find the new load. We can't easily remove the old
1178 // load completely because it is already in the value numbering table.
1179 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1181 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1182 DestPTy = PointerType::get(DestPTy,
1183 PtrVal->getType()->getPointerAddressSpace());
1184 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1185 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1186 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1187 NewLoad->takeName(SrcVal);
1188 NewLoad->setAlignment(SrcVal->getAlignment());
1190 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1191 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1193 // Replace uses of the original load with the wider load. On a big endian
1194 // system, we need to shift down to get the relevant bits.
1195 Value *RV = NewLoad;
1196 if (DL.isBigEndian())
1197 RV = Builder.CreateLShr(RV,
1198 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
1199 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1200 SrcVal->replaceAllUsesWith(RV);
1202 // We would like to use gvn.markInstructionForDeletion here, but we can't
1203 // because the load is already memoized into the leader map table that GVN
1204 // tracks. It is potentially possible to remove the load from the table,
1205 // but then there all of the operations based on it would need to be
1206 // rehashed. Just leave the dead load around.
1207 gvn.getMemDep().removeInstruction(SrcVal);
1211 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, DL);
1215 /// This function is called when we have a
1216 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1217 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1218 Type *LoadTy, Instruction *InsertPt,
1219 const DataLayout &DL){
1220 LLVMContext &Ctx = LoadTy->getContext();
1221 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy)/8;
1223 IRBuilder<> Builder(InsertPt);
1225 // We know that this method is only called when the mem transfer fully
1226 // provides the bits for the load.
1227 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1228 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1229 // independently of what the offset is.
1230 Value *Val = MSI->getValue();
1232 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1234 Value *OneElt = Val;
1236 // Splat the value out to the right number of bits.
1237 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1238 // If we can double the number of bytes set, do it.
1239 if (NumBytesSet*2 <= LoadSize) {
1240 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1241 Val = Builder.CreateOr(Val, ShVal);
1246 // Otherwise insert one byte at a time.
1247 Value *ShVal = Builder.CreateShl(Val, 1*8);
1248 Val = Builder.CreateOr(OneElt, ShVal);
1252 return CoerceAvailableValueToLoadType(Val, LoadTy, Builder, DL);
1255 // Otherwise, this is a memcpy/memmove from a constant global.
1256 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1257 Constant *Src = cast<Constant>(MTI->getSource());
1258 unsigned AS = Src->getType()->getPointerAddressSpace();
1260 // Otherwise, see if we can constant fold a load from the constant with the
1261 // offset applied as appropriate.
1262 Src = ConstantExpr::getBitCast(Src,
1263 Type::getInt8PtrTy(Src->getContext(), AS));
1264 Constant *OffsetCst =
1265 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1266 Src = ConstantExpr::getGetElementPtr(Type::getInt8Ty(Src->getContext()), Src,
1268 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1269 return ConstantFoldLoadFromConstPtr(Src, DL);
1273 /// Given a set of loads specified by ValuesPerBlock,
1274 /// construct SSA form, allowing us to eliminate LI. This returns the value
1275 /// that should be used at LI's definition site.
1276 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1277 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1279 // Check for the fully redundant, dominating load case. In this case, we can
1280 // just use the dominating value directly.
1281 if (ValuesPerBlock.size() == 1 &&
1282 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1284 assert(!ValuesPerBlock[0].isUndefValue() && "Dead BB dominate this block");
1285 return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
1288 // Otherwise, we have to construct SSA form.
1289 SmallVector<PHINode*, 8> NewPHIs;
1290 SSAUpdater SSAUpdate(&NewPHIs);
1291 SSAUpdate.Initialize(LI->getType(), LI->getName());
1293 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1294 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1295 BasicBlock *BB = AV.BB;
1297 if (SSAUpdate.HasValueForBlock(BB))
1300 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
1303 // Perform PHI construction.
1304 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1306 // If new PHI nodes were created, notify alias analysis.
1307 if (V->getType()->getScalarType()->isPointerTy()) {
1308 AliasAnalysis *AA = gvn.getAliasAnalysis();
1310 // Scan the new PHIs and inform alias analysis that we've added potentially
1311 // escaping uses to any values that are operands to these PHIs.
1312 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
1313 PHINode *P = NewPHIs[i];
1314 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) {
1315 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
1316 AA->addEscapingUse(P->getOperandUse(jj));
1324 Value *AvailableValueInBlock::MaterializeAdjustedValue(LoadInst *LI,
1327 Type *LoadTy = LI->getType();
1328 const DataLayout &DL = LI->getModule()->getDataLayout();
1329 if (isSimpleValue()) {
1330 Res = getSimpleValue();
1331 if (Res->getType() != LoadTy) {
1332 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), DL);
1334 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1335 << *getSimpleValue() << '\n'
1336 << *Res << '\n' << "\n\n\n");
1338 } else if (isCoercedLoadValue()) {
1339 LoadInst *Load = getCoercedLoadValue();
1340 if (Load->getType() == LoadTy && Offset == 0) {
1343 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1346 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1347 << *getCoercedLoadValue() << '\n'
1348 << *Res << '\n' << "\n\n\n");
1350 } else if (isMemIntrinValue()) {
1351 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
1352 BB->getTerminator(), DL);
1353 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1354 << " " << *getMemIntrinValue() << '\n'
1355 << *Res << '\n' << "\n\n\n");
1357 assert(isUndefValue() && "Should be UndefVal");
1358 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
1359 return UndefValue::get(LoadTy);
1364 static bool isLifetimeStart(const Instruction *Inst) {
1365 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1366 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1370 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1371 AvailValInBlkVect &ValuesPerBlock,
1372 UnavailBlkVect &UnavailableBlocks) {
1374 // Filter out useless results (non-locals, etc). Keep track of the blocks
1375 // where we have a value available in repl, also keep track of whether we see
1376 // dependencies that produce an unknown value for the load (such as a call
1377 // that could potentially clobber the load).
1378 unsigned NumDeps = Deps.size();
1379 const DataLayout &DL = LI->getModule()->getDataLayout();
1380 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1381 BasicBlock *DepBB = Deps[i].getBB();
1382 MemDepResult DepInfo = Deps[i].getResult();
1384 if (DeadBlocks.count(DepBB)) {
1385 // Dead dependent mem-op disguise as a load evaluating the same value
1386 // as the load in question.
1387 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1391 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1392 UnavailableBlocks.push_back(DepBB);
1396 if (DepInfo.isClobber()) {
1397 // The address being loaded in this non-local block may not be the same as
1398 // the pointer operand of the load if PHI translation occurs. Make sure
1399 // to consider the right address.
1400 Value *Address = Deps[i].getAddress();
1402 // If the dependence is to a store that writes to a superset of the bits
1403 // read by the load, we can extract the bits we need for the load from the
1405 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1408 AnalyzeLoadFromClobberingStore(LI->getType(), Address, DepSI);
1410 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1411 DepSI->getValueOperand(),
1418 // Check to see if we have something like this:
1421 // if we have this, replace the later with an extraction from the former.
1422 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1423 // If this is a clobber and L is the first instruction in its block, then
1424 // we have the first instruction in the entry block.
1425 if (DepLI != LI && Address) {
1427 AnalyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
1430 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1437 // If the clobbering value is a memset/memcpy/memmove, see if we can
1438 // forward a value on from it.
1439 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1441 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1444 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1451 UnavailableBlocks.push_back(DepBB);
1455 // DepInfo.isDef() here
1457 Instruction *DepInst = DepInfo.getInst();
1459 // Loading the allocation -> undef.
1460 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
1461 // Loading immediately after lifetime begin -> undef.
1462 isLifetimeStart(DepInst)) {
1463 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1464 UndefValue::get(LI->getType())));
1468 // Loading from calloc (which zero initializes memory) -> zero
1469 if (isCallocLikeFn(DepInst, TLI)) {
1470 ValuesPerBlock.push_back(AvailableValueInBlock::get(
1471 DepBB, Constant::getNullValue(LI->getType())));
1475 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1476 // Reject loads and stores that are to the same address but are of
1477 // different types if we have to.
1478 if (S->getValueOperand()->getType() != LI->getType()) {
1479 // If the stored value is larger or equal to the loaded value, we can
1481 if (!CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1482 LI->getType(), DL)) {
1483 UnavailableBlocks.push_back(DepBB);
1488 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1489 S->getValueOperand()));
1493 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1494 // If the types mismatch and we can't handle it, reject reuse of the load.
1495 if (LD->getType() != LI->getType()) {
1496 // If the stored value is larger or equal to the loaded value, we can
1498 if (!CanCoerceMustAliasedValueToLoad(LD, LI->getType(), DL)) {
1499 UnavailableBlocks.push_back(DepBB);
1503 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1507 UnavailableBlocks.push_back(DepBB);
1511 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1512 UnavailBlkVect &UnavailableBlocks) {
1513 // Okay, we have *some* definitions of the value. This means that the value
1514 // is available in some of our (transitive) predecessors. Lets think about
1515 // doing PRE of this load. This will involve inserting a new load into the
1516 // predecessor when it's not available. We could do this in general, but
1517 // prefer to not increase code size. As such, we only do this when we know
1518 // that we only have to insert *one* load (which means we're basically moving
1519 // the load, not inserting a new one).
1521 SmallPtrSet<BasicBlock *, 4> Blockers;
1522 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1523 Blockers.insert(UnavailableBlocks[i]);
1525 // Let's find the first basic block with more than one predecessor. Walk
1526 // backwards through predecessors if needed.
1527 BasicBlock *LoadBB = LI->getParent();
1528 BasicBlock *TmpBB = LoadBB;
1530 while (TmpBB->getSinglePredecessor()) {
1531 TmpBB = TmpBB->getSinglePredecessor();
1532 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1534 if (Blockers.count(TmpBB))
1537 // If any of these blocks has more than one successor (i.e. if the edge we
1538 // just traversed was critical), then there are other paths through this
1539 // block along which the load may not be anticipated. Hoisting the load
1540 // above this block would be adding the load to execution paths along
1541 // which it was not previously executed.
1542 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1549 // Check to see how many predecessors have the loaded value fully
1551 MapVector<BasicBlock *, Value *> PredLoads;
1552 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1553 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1554 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1555 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1556 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1558 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1559 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1561 BasicBlock *Pred = *PI;
1562 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1566 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1567 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1568 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1569 << Pred->getName() << "': " << *LI << '\n');
1573 if (LoadBB->isLandingPad()) {
1575 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '"
1576 << Pred->getName() << "': " << *LI << '\n');
1580 CriticalEdgePred.push_back(Pred);
1582 // Only add the predecessors that will not be split for now.
1583 PredLoads[Pred] = nullptr;
1587 // Decide whether PRE is profitable for this load.
1588 unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1589 assert(NumUnavailablePreds != 0 &&
1590 "Fully available value should already be eliminated!");
1592 // If this load is unavailable in multiple predecessors, reject it.
1593 // FIXME: If we could restructure the CFG, we could make a common pred with
1594 // all the preds that don't have an available LI and insert a new load into
1596 if (NumUnavailablePreds != 1)
1599 // Split critical edges, and update the unavailable predecessors accordingly.
1600 for (BasicBlock *OrigPred : CriticalEdgePred) {
1601 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1602 assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1603 PredLoads[NewPred] = nullptr;
1604 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1605 << LoadBB->getName() << '\n');
1608 // Check if the load can safely be moved to all the unavailable predecessors.
1609 bool CanDoPRE = true;
1610 const DataLayout &DL = LI->getModule()->getDataLayout();
1611 SmallVector<Instruction*, 8> NewInsts;
1612 for (auto &PredLoad : PredLoads) {
1613 BasicBlock *UnavailablePred = PredLoad.first;
1615 // Do PHI translation to get its value in the predecessor if necessary. The
1616 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1618 // If all preds have a single successor, then we know it is safe to insert
1619 // the load on the pred (?!?), so we can insert code to materialize the
1620 // pointer if it is not available.
1621 PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1622 Value *LoadPtr = nullptr;
1623 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1626 // If we couldn't find or insert a computation of this phi translated value,
1629 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1630 << *LI->getPointerOperand() << "\n");
1635 PredLoad.second = LoadPtr;
1639 while (!NewInsts.empty()) {
1640 Instruction *I = NewInsts.pop_back_val();
1641 if (MD) MD->removeInstruction(I);
1642 I->eraseFromParent();
1644 // HINT: Don't revert the edge-splitting as following transformation may
1645 // also need to split these critical edges.
1646 return !CriticalEdgePred.empty();
1649 // Okay, we can eliminate this load by inserting a reload in the predecessor
1650 // and using PHI construction to get the value in the other predecessors, do
1652 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1653 DEBUG(if (!NewInsts.empty())
1654 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1655 << *NewInsts.back() << '\n');
1657 // Assign value numbers to the new instructions.
1658 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1659 // FIXME: We really _ought_ to insert these value numbers into their
1660 // parent's availability map. However, in doing so, we risk getting into
1661 // ordering issues. If a block hasn't been processed yet, we would be
1662 // marking a value as AVAIL-IN, which isn't what we intend.
1663 VN.lookup_or_add(NewInsts[i]);
1666 for (const auto &PredLoad : PredLoads) {
1667 BasicBlock *UnavailablePred = PredLoad.first;
1668 Value *LoadPtr = PredLoad.second;
1670 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1672 UnavailablePred->getTerminator());
1674 // Transfer the old load's AA tags to the new load.
1676 LI->getAAMetadata(Tags);
1678 NewLoad->setAAMetadata(Tags);
1680 // Transfer DebugLoc.
1681 NewLoad->setDebugLoc(LI->getDebugLoc());
1683 // Add the newly created load.
1684 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1686 MD->invalidateCachedPointerInfo(LoadPtr);
1687 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1690 // Perform PHI construction.
1691 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1692 LI->replaceAllUsesWith(V);
1693 if (isa<PHINode>(V))
1695 if (Instruction *I = dyn_cast<Instruction>(V))
1696 I->setDebugLoc(LI->getDebugLoc());
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 (Instruction *I = dyn_cast<Instruction>(V))
1764 I->setDebugLoc(LI->getDebugLoc());
1765 if (V->getType()->getScalarType()->isPointerTy())
1766 MD->invalidateCachedPointerInfo(V);
1767 markInstructionForDeletion(LI);
1772 // Step 4: Eliminate partial redundancy.
1773 if (!EnablePRE || !EnableLoadPRE)
1776 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1780 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1781 // Patch the replacement so that it is not more restrictive than the value
1783 BinaryOperator *Op = dyn_cast<BinaryOperator>(I);
1784 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl);
1786 ReplOp->andIRFlags(Op);
1788 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) {
1789 // FIXME: If both the original and replacement value are part of the
1790 // same control-flow region (meaning that the execution of one
1791 // guarentees the executation of the other), then we can combine the
1792 // noalias scopes here and do better than the general conservative
1793 // answer used in combineMetadata().
1795 // In general, GVN unifies expressions over different control-flow
1796 // regions, and so we need a conservative combination of the noalias
1798 unsigned KnownIDs[] = {
1799 LLVMContext::MD_tbaa,
1800 LLVMContext::MD_alias_scope,
1801 LLVMContext::MD_noalias,
1802 LLVMContext::MD_range,
1803 LLVMContext::MD_fpmath,
1804 LLVMContext::MD_invariant_load,
1806 combineMetadata(ReplInst, I, KnownIDs);
1810 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1811 patchReplacementInstruction(I, Repl);
1812 I->replaceAllUsesWith(Repl);
1815 /// Attempt to eliminate a load, first by eliminating it
1816 /// locally, and then attempting non-local elimination if that fails.
1817 bool GVN::processLoad(LoadInst *L) {
1824 if (L->use_empty()) {
1825 markInstructionForDeletion(L);
1829 // ... to a pointer that has been loaded from before...
1830 MemDepResult Dep = MD->getDependency(L);
1831 const DataLayout &DL = L->getModule()->getDataLayout();
1833 // If we have a clobber and target data is around, see if this is a clobber
1834 // that we can fix up through code synthesis.
1835 if (Dep.isClobber()) {
1836 // Check to see if we have something like this:
1837 // store i32 123, i32* %P
1838 // %A = bitcast i32* %P to i8*
1839 // %B = gep i8* %A, i32 1
1842 // We could do that by recognizing if the clobber instructions are obviously
1843 // a common base + constant offset, and if the previous store (or memset)
1844 // completely covers this load. This sort of thing can happen in bitfield
1846 Value *AvailVal = nullptr;
1847 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1848 int Offset = AnalyzeLoadFromClobberingStore(
1849 L->getType(), L->getPointerOperand(), DepSI);
1851 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1852 L->getType(), L, DL);
1855 // Check to see if we have something like this:
1858 // if we have this, replace the later with an extraction from the former.
1859 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1860 // If this is a clobber and L is the first instruction in its block, then
1861 // we have the first instruction in the entry block.
1865 int Offset = AnalyzeLoadFromClobberingLoad(
1866 L->getType(), L->getPointerOperand(), DepLI, DL);
1868 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1871 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1872 // a value on from it.
1873 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1874 int Offset = AnalyzeLoadFromClobberingMemInst(
1875 L->getType(), L->getPointerOperand(), DepMI, DL);
1877 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, DL);
1881 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1882 << *AvailVal << '\n' << *L << "\n\n\n");
1884 // Replace the load!
1885 L->replaceAllUsesWith(AvailVal);
1886 if (AvailVal->getType()->getScalarType()->isPointerTy())
1887 MD->invalidateCachedPointerInfo(AvailVal);
1888 markInstructionForDeletion(L);
1894 // If the value isn't available, don't do anything!
1895 if (Dep.isClobber()) {
1897 // fast print dep, using operator<< on instruction is too slow.
1898 dbgs() << "GVN: load ";
1899 L->printAsOperand(dbgs());
1900 Instruction *I = Dep.getInst();
1901 dbgs() << " is clobbered by " << *I << '\n';
1906 // If it is defined in another block, try harder.
1907 if (Dep.isNonLocal())
1908 return processNonLocalLoad(L);
1912 // fast print dep, using operator<< on instruction is too slow.
1913 dbgs() << "GVN: load ";
1914 L->printAsOperand(dbgs());
1915 dbgs() << " has unknown dependence\n";
1920 Instruction *DepInst = Dep.getInst();
1921 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1922 Value *StoredVal = DepSI->getValueOperand();
1924 // The store and load are to a must-aliased pointer, but they may not
1925 // actually have the same type. See if we know how to reuse the stored
1926 // value (depending on its type).
1927 if (StoredVal->getType() != L->getType()) {
1928 IRBuilder<> Builder(L);
1930 CoerceAvailableValueToLoadType(StoredVal, L->getType(), Builder, DL);
1934 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1935 << '\n' << *L << "\n\n\n");
1939 L->replaceAllUsesWith(StoredVal);
1940 if (StoredVal->getType()->getScalarType()->isPointerTy())
1941 MD->invalidateCachedPointerInfo(StoredVal);
1942 markInstructionForDeletion(L);
1947 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1948 Value *AvailableVal = DepLI;
1950 // The loads are of a must-aliased pointer, but they may not actually have
1951 // the same type. See if we know how to reuse the previously loaded value
1952 // (depending on its type).
1953 if (DepLI->getType() != L->getType()) {
1954 IRBuilder<> Builder(L);
1956 CoerceAvailableValueToLoadType(DepLI, L->getType(), Builder, DL);
1960 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1961 << "\n" << *L << "\n\n\n");
1965 patchAndReplaceAllUsesWith(L, AvailableVal);
1966 if (DepLI->getType()->getScalarType()->isPointerTy())
1967 MD->invalidateCachedPointerInfo(DepLI);
1968 markInstructionForDeletion(L);
1973 // If this load really doesn't depend on anything, then we must be loading an
1974 // undef value. This can happen when loading for a fresh allocation with no
1975 // intervening stores, for example.
1976 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
1977 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1978 markInstructionForDeletion(L);
1983 // If this load occurs either right after a lifetime begin,
1984 // then the loaded value is undefined.
1985 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
1986 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1987 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1988 markInstructionForDeletion(L);
1994 // If this load follows a calloc (which zero initializes memory),
1995 // then the loaded value is zero
1996 if (isCallocLikeFn(DepInst, TLI)) {
1997 L->replaceAllUsesWith(Constant::getNullValue(L->getType()));
1998 markInstructionForDeletion(L);
2006 // In order to find a leader for a given value number at a
2007 // specific basic block, we first obtain the list of all Values for that number,
2008 // and then scan the list to find one whose block dominates the block in
2009 // question. This is fast because dominator tree queries consist of only
2010 // a few comparisons of DFS numbers.
2011 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
2012 LeaderTableEntry Vals = LeaderTable[num];
2013 if (!Vals.Val) return nullptr;
2015 Value *Val = nullptr;
2016 if (DT->dominates(Vals.BB, BB)) {
2018 if (isa<Constant>(Val)) return Val;
2021 LeaderTableEntry* Next = Vals.Next;
2023 if (DT->dominates(Next->BB, BB)) {
2024 if (isa<Constant>(Next->Val)) return Next->Val;
2025 if (!Val) Val = Next->Val;
2034 /// There is an edge from 'Src' to 'Dst'. Return
2035 /// true if every path from the entry block to 'Dst' passes via this edge. In
2036 /// particular 'Dst' must not be reachable via another edge from 'Src'.
2037 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
2038 DominatorTree *DT) {
2039 // While in theory it is interesting to consider the case in which Dst has
2040 // more than one predecessor, because Dst might be part of a loop which is
2041 // only reachable from Src, in practice it is pointless since at the time
2042 // GVN runs all such loops have preheaders, which means that Dst will have
2043 // been changed to have only one predecessor, namely Src.
2044 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
2045 const BasicBlock *Src = E.getStart();
2046 assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
2048 return Pred != nullptr;
2051 /// The given values are known to be equal in every block
2052 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
2053 /// 'RHS' everywhere in the scope. Returns whether a change was made.
2054 bool GVN::propagateEquality(Value *LHS, Value *RHS,
2055 const BasicBlockEdge &Root) {
2056 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
2057 Worklist.push_back(std::make_pair(LHS, RHS));
2058 bool Changed = false;
2059 // For speed, compute a conservative fast approximation to
2060 // DT->dominates(Root, Root.getEnd());
2061 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
2063 while (!Worklist.empty()) {
2064 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
2065 LHS = Item.first; RHS = Item.second;
2067 if (LHS == RHS) continue;
2068 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
2070 // Don't try to propagate equalities between constants.
2071 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue;
2073 // Prefer a constant on the right-hand side, or an Argument if no constants.
2074 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
2075 std::swap(LHS, RHS);
2076 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
2078 // If there is no obvious reason to prefer the left-hand side over the
2079 // right-hand side, ensure the longest lived term is on the right-hand side,
2080 // so the shortest lived term will be replaced by the longest lived.
2081 // This tends to expose more simplifications.
2082 uint32_t LVN = VN.lookup_or_add(LHS);
2083 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
2084 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
2085 // Move the 'oldest' value to the right-hand side, using the value number
2086 // as a proxy for age.
2087 uint32_t RVN = VN.lookup_or_add(RHS);
2089 std::swap(LHS, RHS);
2094 // If value numbering later sees that an instruction in the scope is equal
2095 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
2096 // the invariant that instructions only occur in the leader table for their
2097 // own value number (this is used by removeFromLeaderTable), do not do this
2098 // if RHS is an instruction (if an instruction in the scope is morphed into
2099 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
2100 // using the leader table is about compiling faster, not optimizing better).
2101 // The leader table only tracks basic blocks, not edges. Only add to if we
2102 // have the simple case where the edge dominates the end.
2103 if (RootDominatesEnd && !isa<Instruction>(RHS))
2104 addToLeaderTable(LVN, RHS, Root.getEnd());
2106 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
2107 // LHS always has at least one use that is not dominated by Root, this will
2108 // never do anything if LHS has only one use.
2109 if (!LHS->hasOneUse()) {
2110 unsigned NumReplacements = replaceDominatedUsesWith(LHS, RHS, *DT, Root);
2111 Changed |= NumReplacements > 0;
2112 NumGVNEqProp += NumReplacements;
2115 // Now try to deduce additional equalities from this one. For example, if
2116 // the known equality was "(A != B)" == "false" then it follows that A and B
2117 // are equal in the scope. Only boolean equalities with an explicit true or
2118 // false RHS are currently supported.
2119 if (!RHS->getType()->isIntegerTy(1))
2120 // Not a boolean equality - bail out.
2122 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
2124 // RHS neither 'true' nor 'false' - bail out.
2126 // Whether RHS equals 'true'. Otherwise it equals 'false'.
2127 bool isKnownTrue = CI->isAllOnesValue();
2128 bool isKnownFalse = !isKnownTrue;
2130 // If "A && B" is known true then both A and B are known true. If "A || B"
2131 // is known false then both A and B are known false.
2133 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
2134 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
2135 Worklist.push_back(std::make_pair(A, RHS));
2136 Worklist.push_back(std::make_pair(B, RHS));
2140 // If we are propagating an equality like "(A == B)" == "true" then also
2141 // propagate the equality A == B. When propagating a comparison such as
2142 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
2143 if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
2144 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
2146 // If "A == B" is known true, or "A != B" is known false, then replace
2147 // A with B everywhere in the scope.
2148 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2149 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2150 Worklist.push_back(std::make_pair(Op0, Op1));
2152 // Handle the floating point versions of equality comparisons too.
2153 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
2154 (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
2156 // Floating point -0.0 and 0.0 compare equal, so we can only
2157 // propagate values if we know that we have a constant and that
2158 // its value is non-zero.
2160 // FIXME: We should do this optimization if 'no signed zeros' is
2161 // applicable via an instruction-level fast-math-flag or some other
2162 // indicator that relaxed FP semantics are being used.
2164 if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
2165 Worklist.push_back(std::make_pair(Op0, Op1));
2168 // If "A >= B" is known true, replace "A < B" with false everywhere.
2169 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2170 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2171 // Since we don't have the instruction "A < B" immediately to hand, work
2172 // out the value number that it would have and use that to find an
2173 // appropriate instruction (if any).
2174 uint32_t NextNum = VN.getNextUnusedValueNumber();
2175 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2176 // If the number we were assigned was brand new then there is no point in
2177 // looking for an instruction realizing it: there cannot be one!
2178 if (Num < NextNum) {
2179 Value *NotCmp = findLeader(Root.getEnd(), Num);
2180 if (NotCmp && isa<Instruction>(NotCmp)) {
2181 unsigned NumReplacements =
2182 replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root);
2183 Changed |= NumReplacements > 0;
2184 NumGVNEqProp += NumReplacements;
2187 // Ensure that any instruction in scope that gets the "A < B" value number
2188 // is replaced with false.
2189 // The leader table only tracks basic blocks, not edges. Only add to if we
2190 // have the simple case where the edge dominates the end.
2191 if (RootDominatesEnd)
2192 addToLeaderTable(Num, NotVal, Root.getEnd());
2201 /// When calculating availability, handle an instruction
2202 /// by inserting it into the appropriate sets
2203 bool GVN::processInstruction(Instruction *I) {
2204 // Ignore dbg info intrinsics.
2205 if (isa<DbgInfoIntrinsic>(I))
2208 // If the instruction can be easily simplified then do so now in preference
2209 // to value numbering it. Value numbering often exposes redundancies, for
2210 // example if it determines that %y is equal to %x then the instruction
2211 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2212 const DataLayout &DL = I->getModule()->getDataLayout();
2213 if (Value *V = SimplifyInstruction(I, DL, TLI, DT, AC)) {
2214 I->replaceAllUsesWith(V);
2215 if (MD && V->getType()->getScalarType()->isPointerTy())
2216 MD->invalidateCachedPointerInfo(V);
2217 markInstructionForDeletion(I);
2222 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2223 if (processLoad(LI))
2226 unsigned Num = VN.lookup_or_add(LI);
2227 addToLeaderTable(Num, LI, LI->getParent());
2231 // For conditional branches, we can perform simple conditional propagation on
2232 // the condition value itself.
2233 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2234 if (!BI->isConditional())
2237 if (isa<Constant>(BI->getCondition()))
2238 return processFoldableCondBr(BI);
2240 Value *BranchCond = BI->getCondition();
2241 BasicBlock *TrueSucc = BI->getSuccessor(0);
2242 BasicBlock *FalseSucc = BI->getSuccessor(1);
2243 // Avoid multiple edges early.
2244 if (TrueSucc == FalseSucc)
2247 BasicBlock *Parent = BI->getParent();
2248 bool Changed = false;
2250 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
2251 BasicBlockEdge TrueE(Parent, TrueSucc);
2252 Changed |= propagateEquality(BranchCond, TrueVal, TrueE);
2254 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
2255 BasicBlockEdge FalseE(Parent, FalseSucc);
2256 Changed |= propagateEquality(BranchCond, FalseVal, FalseE);
2261 // For switches, propagate the case values into the case destinations.
2262 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2263 Value *SwitchCond = SI->getCondition();
2264 BasicBlock *Parent = SI->getParent();
2265 bool Changed = false;
2267 // Remember how many outgoing edges there are to every successor.
2268 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2269 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
2270 ++SwitchEdges[SI->getSuccessor(i)];
2272 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2274 BasicBlock *Dst = i.getCaseSuccessor();
2275 // If there is only a single edge, propagate the case value into it.
2276 if (SwitchEdges.lookup(Dst) == 1) {
2277 BasicBlockEdge E(Parent, Dst);
2278 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E);
2284 // Instructions with void type don't return a value, so there's
2285 // no point in trying to find redundancies in them.
2286 if (I->getType()->isVoidTy()) return false;
2288 uint32_t NextNum = VN.getNextUnusedValueNumber();
2289 unsigned Num = VN.lookup_or_add(I);
2291 // Allocations are always uniquely numbered, so we can save time and memory
2292 // by fast failing them.
2293 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2294 addToLeaderTable(Num, I, I->getParent());
2298 // If the number we were assigned was a brand new VN, then we don't
2299 // need to do a lookup to see if the number already exists
2300 // somewhere in the domtree: it can't!
2301 if (Num >= NextNum) {
2302 addToLeaderTable(Num, I, I->getParent());
2306 // Perform fast-path value-number based elimination of values inherited from
2308 Value *repl = findLeader(I->getParent(), Num);
2310 // Failure, just remember this instance for future use.
2311 addToLeaderTable(Num, I, I->getParent());
2316 patchAndReplaceAllUsesWith(I, repl);
2317 if (MD && repl->getType()->getScalarType()->isPointerTy())
2318 MD->invalidateCachedPointerInfo(repl);
2319 markInstructionForDeletion(I);
2323 /// runOnFunction - This is the main transformation entry point for a function.
2324 bool GVN::runOnFunction(Function& F) {
2325 if (skipOptnoneFunction(F))
2329 MD = &getAnalysis<MemoryDependenceAnalysis>();
2330 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2331 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2332 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
2333 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2337 bool Changed = false;
2338 bool ShouldContinue = true;
2340 // Merge unconditional branches, allowing PRE to catch more
2341 // optimization opportunities.
2342 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2343 BasicBlock *BB = FI++;
2345 bool removedBlock = MergeBlockIntoPredecessor(
2346 BB, DT, /* LoopInfo */ nullptr, VN.getAliasAnalysis(), MD);
2347 if (removedBlock) ++NumGVNBlocks;
2349 Changed |= removedBlock;
2352 unsigned Iteration = 0;
2353 while (ShouldContinue) {
2354 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2355 ShouldContinue = iterateOnFunction(F);
2356 Changed |= ShouldContinue;
2361 // Fabricate val-num for dead-code in order to suppress assertion in
2363 assignValNumForDeadCode();
2364 bool PREChanged = true;
2365 while (PREChanged) {
2366 PREChanged = performPRE(F);
2367 Changed |= PREChanged;
2371 // FIXME: Should perform GVN again after PRE does something. PRE can move
2372 // computations into blocks where they become fully redundant. Note that
2373 // we can't do this until PRE's critical edge splitting updates memdep.
2374 // Actually, when this happens, we should just fully integrate PRE into GVN.
2376 cleanupGlobalSets();
2377 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2385 bool GVN::processBlock(BasicBlock *BB) {
2386 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2387 // (and incrementing BI before processing an instruction).
2388 assert(InstrsToErase.empty() &&
2389 "We expect InstrsToErase to be empty across iterations");
2390 if (DeadBlocks.count(BB))
2393 bool ChangedFunction = false;
2395 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2397 ChangedFunction |= processInstruction(BI);
2398 if (InstrsToErase.empty()) {
2403 // If we need some instructions deleted, do it now.
2404 NumGVNInstr += InstrsToErase.size();
2406 // Avoid iterator invalidation.
2407 bool AtStart = BI == BB->begin();
2411 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2412 E = InstrsToErase.end(); I != E; ++I) {
2413 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2414 if (MD) MD->removeInstruction(*I);
2415 DEBUG(verifyRemoved(*I));
2416 (*I)->eraseFromParent();
2418 InstrsToErase.clear();
2426 return ChangedFunction;
2429 // Instantiate an expression in a predecessor that lacked it.
2430 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2431 unsigned int ValNo) {
2432 // Because we are going top-down through the block, all value numbers
2433 // will be available in the predecessor by the time we need them. Any
2434 // that weren't originally present will have been instantiated earlier
2436 bool success = true;
2437 for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2438 Value *Op = Instr->getOperand(i);
2439 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2442 if (Value *V = findLeader(Pred, VN.lookup(Op))) {
2443 Instr->setOperand(i, V);
2450 // Fail out if we encounter an operand that is not available in
2451 // the PRE predecessor. This is typically because of loads which
2452 // are not value numbered precisely.
2456 Instr->insertBefore(Pred->getTerminator());
2457 Instr->setName(Instr->getName() + ".pre");
2458 Instr->setDebugLoc(Instr->getDebugLoc());
2459 VN.add(Instr, ValNo);
2461 // Update the availability map to include the new instruction.
2462 addToLeaderTable(ValNo, Instr, Pred);
2466 bool GVN::performScalarPRE(Instruction *CurInst) {
2467 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap;
2469 if (isa<AllocaInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
2470 isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2471 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2472 isa<DbgInfoIntrinsic>(CurInst))
2475 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2476 // sinking the compare again, and it would force the code generator to
2477 // move the i1 from processor flags or predicate registers into a general
2478 // purpose register.
2479 if (isa<CmpInst>(CurInst))
2482 // We don't currently value number ANY inline asm calls.
2483 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2484 if (CallI->isInlineAsm())
2487 uint32_t ValNo = VN.lookup(CurInst);
2489 // Look for the predecessors for PRE opportunities. We're
2490 // only trying to solve the basic diamond case, where
2491 // a value is computed in the successor and one predecessor,
2492 // but not the other. We also explicitly disallow cases
2493 // where the successor is its own predecessor, because they're
2494 // more complicated to get right.
2495 unsigned NumWith = 0;
2496 unsigned NumWithout = 0;
2497 BasicBlock *PREPred = nullptr;
2498 BasicBlock *CurrentBlock = CurInst->getParent();
2501 for (pred_iterator PI = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock);
2503 BasicBlock *P = *PI;
2504 // We're not interested in PRE where the block is its
2505 // own predecessor, or in blocks with predecessors
2506 // that are not reachable.
2507 if (P == CurrentBlock) {
2510 } else if (!DT->isReachableFromEntry(P)) {
2515 Value *predV = findLeader(P, ValNo);
2517 predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2520 } else if (predV == CurInst) {
2521 /* CurInst dominates this predecessor. */
2525 predMap.push_back(std::make_pair(predV, P));
2530 // Don't do PRE when it might increase code size, i.e. when
2531 // we would need to insert instructions in more than one pred.
2532 if (NumWithout > 1 || NumWith == 0)
2535 // We may have a case where all predecessors have the instruction,
2536 // and we just need to insert a phi node. Otherwise, perform
2538 Instruction *PREInstr = nullptr;
2540 if (NumWithout != 0) {
2541 // Don't do PRE across indirect branch.
2542 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2545 // We can't do PRE safely on a critical edge, so instead we schedule
2546 // the edge to be split and perform the PRE the next time we iterate
2548 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2549 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2550 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2553 // We need to insert somewhere, so let's give it a shot
2554 PREInstr = CurInst->clone();
2555 if (!performScalarPREInsertion(PREInstr, PREPred, ValNo)) {
2556 // If we failed insertion, make sure we remove the instruction.
2557 DEBUG(verifyRemoved(PREInstr));
2563 // Either we should have filled in the PRE instruction, or we should
2564 // not have needed insertions.
2565 assert (PREInstr != nullptr || NumWithout == 0);
2569 // Create a PHI to make the value available in this block.
2571 PHINode::Create(CurInst->getType(), predMap.size(),
2572 CurInst->getName() + ".pre-phi", CurrentBlock->begin());
2573 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2574 if (Value *V = predMap[i].first)
2575 Phi->addIncoming(V, predMap[i].second);
2577 Phi->addIncoming(PREInstr, PREPred);
2581 addToLeaderTable(ValNo, Phi, CurrentBlock);
2582 Phi->setDebugLoc(CurInst->getDebugLoc());
2583 CurInst->replaceAllUsesWith(Phi);
2584 if (Phi->getType()->getScalarType()->isPointerTy()) {
2585 // Because we have added a PHI-use of the pointer value, it has now
2586 // "escaped" from alias analysis' perspective. We need to inform
2588 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; ++ii) {
2589 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
2590 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
2594 MD->invalidateCachedPointerInfo(Phi);
2597 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2599 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2601 MD->removeInstruction(CurInst);
2602 DEBUG(verifyRemoved(CurInst));
2603 CurInst->eraseFromParent();
2609 /// Perform a purely local form of PRE that looks for diamond
2610 /// control flow patterns and attempts to perform simple PRE at the join point.
2611 bool GVN::performPRE(Function &F) {
2612 bool Changed = false;
2613 for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2614 // Nothing to PRE in the entry block.
2615 if (CurrentBlock == &F.getEntryBlock())
2618 // Don't perform PRE on a landing pad.
2619 if (CurrentBlock->isLandingPad())
2622 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2623 BE = CurrentBlock->end();
2625 Instruction *CurInst = BI++;
2626 Changed = performScalarPRE(CurInst);
2630 if (splitCriticalEdges())
2636 /// Split the critical edge connecting the given two blocks, and return
2637 /// the block inserted to the critical edge.
2638 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2639 BasicBlock *BB = SplitCriticalEdge(
2640 Pred, Succ, CriticalEdgeSplittingOptions(getAliasAnalysis(), DT));
2642 MD->invalidateCachedPredecessors();
2646 /// Split critical edges found during the previous
2647 /// iteration that may enable further optimization.
2648 bool GVN::splitCriticalEdges() {
2649 if (toSplit.empty())
2652 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2653 SplitCriticalEdge(Edge.first, Edge.second,
2654 CriticalEdgeSplittingOptions(getAliasAnalysis(), DT));
2655 } while (!toSplit.empty());
2656 if (MD) MD->invalidateCachedPredecessors();
2660 /// Executes one iteration of GVN
2661 bool GVN::iterateOnFunction(Function &F) {
2662 cleanupGlobalSets();
2664 // Top-down walk of the dominator tree
2665 bool Changed = false;
2666 // Save the blocks this function have before transformation begins. GVN may
2667 // split critical edge, and hence may invalidate the RPO/DT iterator.
2669 std::vector<BasicBlock *> BBVect;
2670 BBVect.reserve(256);
2671 // Needed for value numbering with phi construction to work.
2672 ReversePostOrderTraversal<Function *> RPOT(&F);
2673 for (ReversePostOrderTraversal<Function *>::rpo_iterator RI = RPOT.begin(),
2676 BBVect.push_back(*RI);
2678 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
2680 Changed |= processBlock(*I);
2685 void GVN::cleanupGlobalSets() {
2687 LeaderTable.clear();
2688 TableAllocator.Reset();
2691 /// Verify that the specified instruction does not occur in our
2692 /// internal data structures.
2693 void GVN::verifyRemoved(const Instruction *Inst) const {
2694 VN.verifyRemoved(Inst);
2696 // Walk through the value number scope to make sure the instruction isn't
2697 // ferreted away in it.
2698 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2699 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2700 const LeaderTableEntry *Node = &I->second;
2701 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2703 while (Node->Next) {
2705 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2710 /// BB is declared dead, which implied other blocks become dead as well. This
2711 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2712 /// live successors, update their phi nodes by replacing the operands
2713 /// corresponding to dead blocks with UndefVal.
2714 void GVN::addDeadBlock(BasicBlock *BB) {
2715 SmallVector<BasicBlock *, 4> NewDead;
2716 SmallSetVector<BasicBlock *, 4> DF;
2718 NewDead.push_back(BB);
2719 while (!NewDead.empty()) {
2720 BasicBlock *D = NewDead.pop_back_val();
2721 if (DeadBlocks.count(D))
2724 // All blocks dominated by D are dead.
2725 SmallVector<BasicBlock *, 8> Dom;
2726 DT->getDescendants(D, Dom);
2727 DeadBlocks.insert(Dom.begin(), Dom.end());
2729 // Figure out the dominance-frontier(D).
2730 for (SmallVectorImpl<BasicBlock *>::iterator I = Dom.begin(),
2731 E = Dom.end(); I != E; I++) {
2733 for (succ_iterator SI = succ_begin(B), SE = succ_end(B); SI != SE; SI++) {
2734 BasicBlock *S = *SI;
2735 if (DeadBlocks.count(S))
2738 bool AllPredDead = true;
2739 for (pred_iterator PI = pred_begin(S), PE = pred_end(S); PI != PE; PI++)
2740 if (!DeadBlocks.count(*PI)) {
2741 AllPredDead = false;
2746 // S could be proved dead later on. That is why we don't update phi
2747 // operands at this moment.
2750 // While S is not dominated by D, it is dead by now. This could take
2751 // place if S already have a dead predecessor before D is declared
2753 NewDead.push_back(S);
2759 // For the dead blocks' live successors, update their phi nodes by replacing
2760 // the operands corresponding to dead blocks with UndefVal.
2761 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2764 if (DeadBlocks.count(B))
2767 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2768 for (SmallVectorImpl<BasicBlock *>::iterator PI = Preds.begin(),
2769 PE = Preds.end(); PI != PE; PI++) {
2770 BasicBlock *P = *PI;
2772 if (!DeadBlocks.count(P))
2775 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2776 if (BasicBlock *S = splitCriticalEdges(P, B))
2777 DeadBlocks.insert(P = S);
2780 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2781 PHINode &Phi = cast<PHINode>(*II);
2782 Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2783 UndefValue::get(Phi.getType()));
2789 // If the given branch is recognized as a foldable branch (i.e. conditional
2790 // branch with constant condition), it will perform following analyses and
2792 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2793 // R be the target of the dead out-coming edge.
2794 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2795 // edge. The result of this step will be {X| X is dominated by R}
2796 // 2) Identify those blocks which haves at least one dead prodecessor. The
2797 // result of this step will be dominance-frontier(R).
2798 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2799 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2801 // Return true iff *NEW* dead code are found.
2802 bool GVN::processFoldableCondBr(BranchInst *BI) {
2803 if (!BI || BI->isUnconditional())
2806 // If a branch has two identical successors, we cannot declare either dead.
2807 if (BI->getSuccessor(0) == BI->getSuccessor(1))
2810 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2814 BasicBlock *DeadRoot = Cond->getZExtValue() ?
2815 BI->getSuccessor(1) : BI->getSuccessor(0);
2816 if (DeadBlocks.count(DeadRoot))
2819 if (!DeadRoot->getSinglePredecessor())
2820 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2822 addDeadBlock(DeadRoot);
2826 // performPRE() will trigger assert if it comes across an instruction without
2827 // associated val-num. As it normally has far more live instructions than dead
2828 // instructions, it makes more sense just to "fabricate" a val-number for the
2829 // dead code than checking if instruction involved is dead or not.
2830 void GVN::assignValNumForDeadCode() {
2831 for (SetVector<BasicBlock *>::iterator I = DeadBlocks.begin(),
2832 E = DeadBlocks.end(); I != E; I++) {
2833 BasicBlock *BB = *I;
2834 for (BasicBlock::iterator II = BB->begin(), EE = BB->end();
2836 Instruction *Inst = &*II;
2837 unsigned ValNum = VN.lookup_or_add(Inst);
2838 addToLeaderTable(ValNum, Inst, BB);