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->Val != I || Curr->BB != BB) {
665 Prev->Next = Curr->Next;
671 LeaderTableEntry* Next = Curr->Next;
672 Curr->Val = Next->Val;
674 Curr->Next = Next->Next;
679 // List of critical edges to be split between iterations.
680 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
682 // This transformation requires dominator postdominator info
683 void getAnalysisUsage(AnalysisUsage &AU) const override {
684 AU.addRequired<AssumptionCacheTracker>();
685 AU.addRequired<DominatorTreeWrapperPass>();
686 AU.addRequired<TargetLibraryInfoWrapperPass>();
688 AU.addRequired<MemoryDependenceAnalysis>();
689 AU.addRequired<AliasAnalysis>();
691 AU.addPreserved<DominatorTreeWrapperPass>();
692 AU.addPreserved<AliasAnalysis>();
696 // Helper fuctions of redundant load elimination
697 bool processLoad(LoadInst *L);
698 bool processNonLocalLoad(LoadInst *L);
699 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
700 AvailValInBlkVect &ValuesPerBlock,
701 UnavailBlkVect &UnavailableBlocks);
702 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
703 UnavailBlkVect &UnavailableBlocks);
705 // Other helper routines
706 bool processInstruction(Instruction *I);
707 bool processBlock(BasicBlock *BB);
708 void dump(DenseMap<uint32_t, Value*> &d);
709 bool iterateOnFunction(Function &F);
710 bool performPRE(Function &F);
711 bool performScalarPRE(Instruction *I);
712 bool performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
714 Value *findLeader(const BasicBlock *BB, uint32_t num);
715 void cleanupGlobalSets();
716 void verifyRemoved(const Instruction *I) const;
717 bool splitCriticalEdges();
718 BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ);
719 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root);
720 bool processFoldableCondBr(BranchInst *BI);
721 void addDeadBlock(BasicBlock *BB);
722 void assignValNumForDeadCode();
728 // The public interface to this file...
729 FunctionPass *llvm::createGVNPass(bool NoLoads) {
730 return new GVN(NoLoads);
733 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
734 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
735 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
736 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
737 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
738 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
739 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
741 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
742 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
744 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
745 E = d.end(); I != E; ++I) {
746 errs() << I->first << "\n";
753 /// Return true if we can prove that the value
754 /// we're analyzing is fully available in the specified block. As we go, keep
755 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
756 /// map is actually a tri-state map with the following values:
757 /// 0) we know the block *is not* fully available.
758 /// 1) we know the block *is* fully available.
759 /// 2) we do not know whether the block is fully available or not, but we are
760 /// currently speculating that it will be.
761 /// 3) we are speculating for this block and have used that to speculate for
763 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
764 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
765 uint32_t RecurseDepth) {
766 if (RecurseDepth > MaxRecurseDepth)
769 // Optimistically assume that the block is fully available and check to see
770 // if we already know about this block in one lookup.
771 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
772 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
774 // If the entry already existed for this block, return the precomputed value.
776 // If this is a speculative "available" value, mark it as being used for
777 // speculation of other blocks.
778 if (IV.first->second == 2)
779 IV.first->second = 3;
780 return IV.first->second != 0;
783 // Otherwise, see if it is fully available in all predecessors.
784 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
786 // If this block has no predecessors, it isn't live-in here.
788 goto SpeculationFailure;
790 for (; PI != PE; ++PI)
791 // If the value isn't fully available in one of our predecessors, then it
792 // isn't fully available in this block either. Undo our previous
793 // optimistic assumption and bail out.
794 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
795 goto SpeculationFailure;
799 // If we get here, we found out that this is not, after
800 // all, a fully-available block. We have a problem if we speculated on this and
801 // used the speculation to mark other blocks as available.
803 char &BBVal = FullyAvailableBlocks[BB];
805 // If we didn't speculate on this, just return with it set to false.
811 // If we did speculate on this value, we could have blocks set to 1 that are
812 // incorrect. Walk the (transitive) successors of this block and mark them as
814 SmallVector<BasicBlock*, 32> BBWorklist;
815 BBWorklist.push_back(BB);
818 BasicBlock *Entry = BBWorklist.pop_back_val();
819 // Note that this sets blocks to 0 (unavailable) if they happen to not
820 // already be in FullyAvailableBlocks. This is safe.
821 char &EntryVal = FullyAvailableBlocks[Entry];
822 if (EntryVal == 0) continue; // Already unavailable.
824 // Mark as unavailable.
827 BBWorklist.append(succ_begin(Entry), succ_end(Entry));
828 } while (!BBWorklist.empty());
834 /// Return true if CoerceAvailableValueToLoadType will succeed.
835 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
837 const DataLayout &DL) {
838 // If the loaded or stored value is an first class array or struct, don't try
839 // to transform them. We need to be able to bitcast to integer.
840 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
841 StoredVal->getType()->isStructTy() ||
842 StoredVal->getType()->isArrayTy())
845 // The store has to be at least as big as the load.
846 if (DL.getTypeSizeInBits(StoredVal->getType()) <
847 DL.getTypeSizeInBits(LoadTy))
853 /// If we saw a store of a value to memory, and
854 /// then a load from a must-aliased pointer of a different type, try to coerce
855 /// the stored value. LoadedTy is the type of the load we want to replace.
856 /// IRB is IRBuilder used to insert new instructions.
858 /// If we can't do it, return null.
859 static Value *CoerceAvailableValueToLoadType(Value *StoredVal, Type *LoadedTy,
861 const DataLayout &DL) {
862 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL))
865 // If this is already the right type, just return it.
866 Type *StoredValTy = StoredVal->getType();
868 uint64_t StoreSize = DL.getTypeSizeInBits(StoredValTy);
869 uint64_t LoadSize = DL.getTypeSizeInBits(LoadedTy);
871 // If the store and reload are the same size, we can always reuse it.
872 if (StoreSize == LoadSize) {
873 // Pointer to Pointer -> use bitcast.
874 if (StoredValTy->getScalarType()->isPointerTy() &&
875 LoadedTy->getScalarType()->isPointerTy())
876 return IRB.CreateBitCast(StoredVal, LoadedTy);
878 // Convert source pointers to integers, which can be bitcast.
879 if (StoredValTy->getScalarType()->isPointerTy()) {
880 StoredValTy = DL.getIntPtrType(StoredValTy);
881 StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
884 Type *TypeToCastTo = LoadedTy;
885 if (TypeToCastTo->getScalarType()->isPointerTy())
886 TypeToCastTo = DL.getIntPtrType(TypeToCastTo);
888 if (StoredValTy != TypeToCastTo)
889 StoredVal = IRB.CreateBitCast(StoredVal, TypeToCastTo);
891 // Cast to pointer if the load needs a pointer type.
892 if (LoadedTy->getScalarType()->isPointerTy())
893 StoredVal = IRB.CreateIntToPtr(StoredVal, LoadedTy);
898 // If the loaded value is smaller than the available value, then we can
899 // extract out a piece from it. If the available value is too small, then we
900 // can't do anything.
901 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
903 // Convert source pointers to integers, which can be manipulated.
904 if (StoredValTy->getScalarType()->isPointerTy()) {
905 StoredValTy = DL.getIntPtrType(StoredValTy);
906 StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
909 // Convert vectors and fp to integer, which can be manipulated.
910 if (!StoredValTy->isIntegerTy()) {
911 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
912 StoredVal = IRB.CreateBitCast(StoredVal, StoredValTy);
915 // If this is a big-endian system, we need to shift the value down to the low
916 // bits so that a truncate will work.
917 if (DL.isBigEndian()) {
918 StoredVal = IRB.CreateLShr(StoredVal, StoreSize - LoadSize, "tmp");
921 // Truncate the integer to the right size now.
922 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
923 StoredVal = IRB.CreateTrunc(StoredVal, NewIntTy, "trunc");
925 if (LoadedTy == NewIntTy)
928 // If the result is a pointer, inttoptr.
929 if (LoadedTy->getScalarType()->isPointerTy())
930 return IRB.CreateIntToPtr(StoredVal, LoadedTy, "inttoptr");
932 // Otherwise, bitcast.
933 return IRB.CreateBitCast(StoredVal, LoadedTy, "bitcast");
936 /// This function is called when we have a
937 /// memdep query of a load that ends up being a clobbering memory write (store,
938 /// memset, memcpy, memmove). This means that the write *may* provide bits used
939 /// by the load but we can't be sure because the pointers don't mustalias.
941 /// Check this case to see if there is anything more we can do before we give
942 /// up. This returns -1 if we have to give up, or a byte number in the stored
943 /// value of the piece that feeds the load.
944 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
946 uint64_t WriteSizeInBits,
947 const DataLayout &DL) {
948 // If the loaded or stored value is a first class array or struct, don't try
949 // to transform them. We need to be able to bitcast to integer.
950 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
953 int64_t StoreOffset = 0, LoadOffset = 0;
955 GetPointerBaseWithConstantOffset(WritePtr, StoreOffset, DL);
956 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, DL);
957 if (StoreBase != LoadBase)
960 // If the load and store are to the exact same address, they should have been
961 // a must alias. AA must have gotten confused.
962 // FIXME: Study to see if/when this happens. One case is forwarding a memset
963 // to a load from the base of the memset.
965 if (LoadOffset == StoreOffset) {
966 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
967 << "Base = " << *StoreBase << "\n"
968 << "Store Ptr = " << *WritePtr << "\n"
969 << "Store Offs = " << StoreOffset << "\n"
970 << "Load Ptr = " << *LoadPtr << "\n";
975 // If the load and store don't overlap at all, the store doesn't provide
976 // anything to the load. In this case, they really don't alias at all, AA
977 // must have gotten confused.
978 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy);
980 if ((WriteSizeInBits & 7) | (LoadSize & 7))
982 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
986 bool isAAFailure = false;
987 if (StoreOffset < LoadOffset)
988 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
990 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
994 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
995 << "Base = " << *StoreBase << "\n"
996 << "Store Ptr = " << *WritePtr << "\n"
997 << "Store Offs = " << StoreOffset << "\n"
998 << "Load Ptr = " << *LoadPtr << "\n";
1004 // If the Load isn't completely contained within the stored bits, we don't
1005 // have all the bits to feed it. We could do something crazy in the future
1006 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1008 if (StoreOffset > LoadOffset ||
1009 StoreOffset+StoreSize < LoadOffset+LoadSize)
1012 // Okay, we can do this transformation. Return the number of bytes into the
1013 // store that the load is.
1014 return LoadOffset-StoreOffset;
1017 /// This function is called when we have a
1018 /// memdep query of a load that ends up being a clobbering store.
1019 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
1021 // Cannot handle reading from store of first-class aggregate yet.
1022 if (DepSI->getValueOperand()->getType()->isStructTy() ||
1023 DepSI->getValueOperand()->getType()->isArrayTy())
1026 const DataLayout &DL = DepSI->getModule()->getDataLayout();
1027 Value *StorePtr = DepSI->getPointerOperand();
1028 uint64_t StoreSize =DL.getTypeSizeInBits(DepSI->getValueOperand()->getType());
1029 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1030 StorePtr, StoreSize, DL);
1033 /// This function is called when we have a
1034 /// memdep query of a load that ends up being clobbered by another load. See if
1035 /// the other load can feed into the second load.
1036 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
1037 LoadInst *DepLI, const DataLayout &DL){
1038 // Cannot handle reading from store of first-class aggregate yet.
1039 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
1042 Value *DepPtr = DepLI->getPointerOperand();
1043 uint64_t DepSize = DL.getTypeSizeInBits(DepLI->getType());
1044 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL);
1045 if (R != -1) return R;
1047 // If we have a load/load clobber an DepLI can be widened to cover this load,
1048 // then we should widen it!
1049 int64_t LoadOffs = 0;
1050 const Value *LoadBase =
1051 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, DL);
1052 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1054 unsigned Size = MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize(
1055 LoadBase, LoadOffs, LoadSize, DepLI);
1056 if (Size == 0) return -1;
1058 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, DL);
1063 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
1065 const DataLayout &DL) {
1066 // If the mem operation is a non-constant size, we can't handle it.
1067 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1068 if (!SizeCst) return -1;
1069 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1071 // If this is memset, we just need to see if the offset is valid in the size
1073 if (MI->getIntrinsicID() == Intrinsic::memset)
1074 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
1077 // If we have a memcpy/memmove, the only case we can handle is if this is a
1078 // copy from constant memory. In that case, we can read directly from the
1080 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1082 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1083 if (!Src) return -1;
1085 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, DL));
1086 if (!GV || !GV->isConstant()) return -1;
1088 // See if the access is within the bounds of the transfer.
1089 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1090 MI->getDest(), MemSizeInBits, DL);
1094 unsigned AS = Src->getType()->getPointerAddressSpace();
1095 // Otherwise, see if we can constant fold a load from the constant with the
1096 // offset applied as appropriate.
1097 Src = ConstantExpr::getBitCast(Src,
1098 Type::getInt8PtrTy(Src->getContext(), AS));
1099 Constant *OffsetCst =
1100 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1101 Src = ConstantExpr::getGetElementPtr(Type::getInt8Ty(Src->getContext()), Src,
1103 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1104 if (ConstantFoldLoadFromConstPtr(Src, DL))
1110 /// This function is called when we have a
1111 /// memdep query of a load that ends up being a clobbering store. This means
1112 /// that the store provides bits used by the load but we the pointers don't
1113 /// mustalias. Check this case to see if there is anything more we can do
1114 /// before we give up.
1115 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1117 Instruction *InsertPt, const DataLayout &DL){
1118 LLVMContext &Ctx = SrcVal->getType()->getContext();
1120 uint64_t StoreSize = (DL.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1121 uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy) + 7) / 8;
1123 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1125 // Compute which bits of the stored value are being used by the load. Convert
1126 // to an integer type to start with.
1127 if (SrcVal->getType()->getScalarType()->isPointerTy())
1128 SrcVal = Builder.CreatePtrToInt(SrcVal,
1129 DL.getIntPtrType(SrcVal->getType()));
1130 if (!SrcVal->getType()->isIntegerTy())
1131 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1133 // Shift the bits to the least significant depending on endianness.
1135 if (DL.isLittleEndian())
1136 ShiftAmt = Offset*8;
1138 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1141 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1143 if (LoadSize != StoreSize)
1144 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1146 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, Builder, DL);
1149 /// This function is called when we have a
1150 /// memdep query of a load that ends up being a clobbering load. This means
1151 /// that the load *may* provide bits used by the load but we can't be sure
1152 /// because the pointers don't mustalias. Check this case to see if there is
1153 /// anything more we can do before we give up.
1154 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1155 Type *LoadTy, Instruction *InsertPt,
1157 const DataLayout &DL = SrcVal->getModule()->getDataLayout();
1158 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1159 // widen SrcVal out to a larger load.
1160 unsigned SrcValSize = DL.getTypeStoreSize(SrcVal->getType());
1161 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1162 if (Offset+LoadSize > SrcValSize) {
1163 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1164 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1165 // If we have a load/load clobber an DepLI can be widened to cover this
1166 // load, then we should widen it to the next power of 2 size big enough!
1167 unsigned NewLoadSize = Offset+LoadSize;
1168 if (!isPowerOf2_32(NewLoadSize))
1169 NewLoadSize = NextPowerOf2(NewLoadSize);
1171 Value *PtrVal = SrcVal->getPointerOperand();
1173 // Insert the new load after the old load. This ensures that subsequent
1174 // memdep queries will find the new load. We can't easily remove the old
1175 // load completely because it is already in the value numbering table.
1176 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1178 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1179 DestPTy = PointerType::get(DestPTy,
1180 PtrVal->getType()->getPointerAddressSpace());
1181 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1182 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1183 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1184 NewLoad->takeName(SrcVal);
1185 NewLoad->setAlignment(SrcVal->getAlignment());
1187 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1188 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1190 // Replace uses of the original load with the wider load. On a big endian
1191 // system, we need to shift down to get the relevant bits.
1192 Value *RV = NewLoad;
1193 if (DL.isBigEndian())
1194 RV = Builder.CreateLShr(RV,
1195 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
1196 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1197 SrcVal->replaceAllUsesWith(RV);
1199 // We would like to use gvn.markInstructionForDeletion here, but we can't
1200 // because the load is already memoized into the leader map table that GVN
1201 // tracks. It is potentially possible to remove the load from the table,
1202 // but then there all of the operations based on it would need to be
1203 // rehashed. Just leave the dead load around.
1204 gvn.getMemDep().removeInstruction(SrcVal);
1208 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, DL);
1212 /// This function is called when we have a
1213 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1214 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1215 Type *LoadTy, Instruction *InsertPt,
1216 const DataLayout &DL){
1217 LLVMContext &Ctx = LoadTy->getContext();
1218 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy)/8;
1220 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1222 // We know that this method is only called when the mem transfer fully
1223 // provides the bits for the load.
1224 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1225 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1226 // independently of what the offset is.
1227 Value *Val = MSI->getValue();
1229 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1231 Value *OneElt = Val;
1233 // Splat the value out to the right number of bits.
1234 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1235 // If we can double the number of bytes set, do it.
1236 if (NumBytesSet*2 <= LoadSize) {
1237 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1238 Val = Builder.CreateOr(Val, ShVal);
1243 // Otherwise insert one byte at a time.
1244 Value *ShVal = Builder.CreateShl(Val, 1*8);
1245 Val = Builder.CreateOr(OneElt, ShVal);
1249 return CoerceAvailableValueToLoadType(Val, LoadTy, Builder, DL);
1252 // Otherwise, this is a memcpy/memmove from a constant global.
1253 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1254 Constant *Src = cast<Constant>(MTI->getSource());
1255 unsigned AS = Src->getType()->getPointerAddressSpace();
1257 // Otherwise, see if we can constant fold a load from the constant with the
1258 // offset applied as appropriate.
1259 Src = ConstantExpr::getBitCast(Src,
1260 Type::getInt8PtrTy(Src->getContext(), AS));
1261 Constant *OffsetCst =
1262 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1263 Src = ConstantExpr::getGetElementPtr(Type::getInt8Ty(Src->getContext()), Src,
1265 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1266 return ConstantFoldLoadFromConstPtr(Src, DL);
1270 /// Given a set of loads specified by ValuesPerBlock,
1271 /// construct SSA form, allowing us to eliminate LI. This returns the value
1272 /// that should be used at LI's definition site.
1273 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1274 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1276 // Check for the fully redundant, dominating load case. In this case, we can
1277 // just use the dominating value directly.
1278 if (ValuesPerBlock.size() == 1 &&
1279 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1281 assert(!ValuesPerBlock[0].isUndefValue() && "Dead BB dominate this block");
1282 return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
1285 // Otherwise, we have to construct SSA form.
1286 SmallVector<PHINode*, 8> NewPHIs;
1287 SSAUpdater SSAUpdate(&NewPHIs);
1288 SSAUpdate.Initialize(LI->getType(), LI->getName());
1290 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1291 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1292 BasicBlock *BB = AV.BB;
1294 if (SSAUpdate.HasValueForBlock(BB))
1297 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
1300 // Perform PHI construction.
1301 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1303 // If new PHI nodes were created, notify alias analysis.
1304 if (V->getType()->getScalarType()->isPointerTy()) {
1305 AliasAnalysis *AA = gvn.getAliasAnalysis();
1307 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1308 AA->copyValue(LI, NewPHIs[i]);
1310 // Now that we've copied information to the new PHIs, scan through
1311 // them again and inform alias analysis that we've added potentially
1312 // escaping uses to any values that are operands to these PHIs.
1313 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
1314 PHINode *P = NewPHIs[i];
1315 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) {
1316 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
1317 AA->addEscapingUse(P->getOperandUse(jj));
1325 Value *AvailableValueInBlock::MaterializeAdjustedValue(LoadInst *LI,
1328 Type *LoadTy = LI->getType();
1329 const DataLayout &DL = LI->getModule()->getDataLayout();
1330 if (isSimpleValue()) {
1331 Res = getSimpleValue();
1332 if (Res->getType() != LoadTy) {
1333 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), DL);
1335 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1336 << *getSimpleValue() << '\n'
1337 << *Res << '\n' << "\n\n\n");
1339 } else if (isCoercedLoadValue()) {
1340 LoadInst *Load = getCoercedLoadValue();
1341 if (Load->getType() == LoadTy && Offset == 0) {
1344 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1347 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1348 << *getCoercedLoadValue() << '\n'
1349 << *Res << '\n' << "\n\n\n");
1351 } else if (isMemIntrinValue()) {
1352 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
1353 BB->getTerminator(), DL);
1354 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1355 << " " << *getMemIntrinValue() << '\n'
1356 << *Res << '\n' << "\n\n\n");
1358 assert(isUndefValue() && "Should be UndefVal");
1359 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
1360 return UndefValue::get(LoadTy);
1365 static bool isLifetimeStart(const Instruction *Inst) {
1366 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1367 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1371 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1372 AvailValInBlkVect &ValuesPerBlock,
1373 UnavailBlkVect &UnavailableBlocks) {
1375 // Filter out useless results (non-locals, etc). Keep track of the blocks
1376 // where we have a value available in repl, also keep track of whether we see
1377 // dependencies that produce an unknown value for the load (such as a call
1378 // that could potentially clobber the load).
1379 unsigned NumDeps = Deps.size();
1380 const DataLayout &DL = LI->getModule()->getDataLayout();
1381 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1382 BasicBlock *DepBB = Deps[i].getBB();
1383 MemDepResult DepInfo = Deps[i].getResult();
1385 if (DeadBlocks.count(DepBB)) {
1386 // Dead dependent mem-op disguise as a load evaluating the same value
1387 // as the load in question.
1388 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1392 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1393 UnavailableBlocks.push_back(DepBB);
1397 if (DepInfo.isClobber()) {
1398 // The address being loaded in this non-local block may not be the same as
1399 // the pointer operand of the load if PHI translation occurs. Make sure
1400 // to consider the right address.
1401 Value *Address = Deps[i].getAddress();
1403 // If the dependence is to a store that writes to a superset of the bits
1404 // read by the load, we can extract the bits we need for the load from the
1406 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1409 AnalyzeLoadFromClobberingStore(LI->getType(), Address, DepSI);
1411 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1412 DepSI->getValueOperand(),
1419 // Check to see if we have something like this:
1422 // if we have this, replace the later with an extraction from the former.
1423 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1424 // If this is a clobber and L is the first instruction in its block, then
1425 // we have the first instruction in the entry block.
1426 if (DepLI != LI && Address) {
1428 AnalyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
1431 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1438 // If the clobbering value is a memset/memcpy/memmove, see if we can
1439 // forward a value on from it.
1440 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1442 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1445 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1452 UnavailableBlocks.push_back(DepBB);
1456 // DepInfo.isDef() here
1458 Instruction *DepInst = DepInfo.getInst();
1460 // Loading the allocation -> undef.
1461 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
1462 // Loading immediately after lifetime begin -> undef.
1463 isLifetimeStart(DepInst)) {
1464 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1465 UndefValue::get(LI->getType())));
1469 // Loading from calloc (which zero initializes memory) -> zero
1470 if (isCallocLikeFn(DepInst, TLI)) {
1471 ValuesPerBlock.push_back(AvailableValueInBlock::get(
1472 DepBB, Constant::getNullValue(LI->getType())));
1476 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1477 // Reject loads and stores that are to the same address but are of
1478 // different types if we have to.
1479 if (S->getValueOperand()->getType() != LI->getType()) {
1480 // If the stored value is larger or equal to the loaded value, we can
1482 if (!CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1483 LI->getType(), DL)) {
1484 UnavailableBlocks.push_back(DepBB);
1489 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1490 S->getValueOperand()));
1494 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1495 // If the types mismatch and we can't handle it, reject reuse of the load.
1496 if (LD->getType() != LI->getType()) {
1497 // If the stored value is larger or equal to the loaded value, we can
1499 if (!CanCoerceMustAliasedValueToLoad(LD, LI->getType(), DL)) {
1500 UnavailableBlocks.push_back(DepBB);
1504 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1508 UnavailableBlocks.push_back(DepBB);
1512 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1513 UnavailBlkVect &UnavailableBlocks) {
1514 // Okay, we have *some* definitions of the value. This means that the value
1515 // is available in some of our (transitive) predecessors. Lets think about
1516 // doing PRE of this load. This will involve inserting a new load into the
1517 // predecessor when it's not available. We could do this in general, but
1518 // prefer to not increase code size. As such, we only do this when we know
1519 // that we only have to insert *one* load (which means we're basically moving
1520 // the load, not inserting a new one).
1522 SmallPtrSet<BasicBlock *, 4> Blockers;
1523 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1524 Blockers.insert(UnavailableBlocks[i]);
1526 // Let's find the first basic block with more than one predecessor. Walk
1527 // backwards through predecessors if needed.
1528 BasicBlock *LoadBB = LI->getParent();
1529 BasicBlock *TmpBB = LoadBB;
1531 while (TmpBB->getSinglePredecessor()) {
1532 TmpBB = TmpBB->getSinglePredecessor();
1533 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1535 if (Blockers.count(TmpBB))
1538 // If any of these blocks has more than one successor (i.e. if the edge we
1539 // just traversed was critical), then there are other paths through this
1540 // block along which the load may not be anticipated. Hoisting the load
1541 // above this block would be adding the load to execution paths along
1542 // which it was not previously executed.
1543 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1550 // Check to see how many predecessors have the loaded value fully
1552 MapVector<BasicBlock *, Value *> PredLoads;
1553 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1554 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1555 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1556 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1557 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1559 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1560 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1562 BasicBlock *Pred = *PI;
1563 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1567 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1568 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1569 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1570 << Pred->getName() << "': " << *LI << '\n');
1574 if (LoadBB->isLandingPad()) {
1576 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '"
1577 << Pred->getName() << "': " << *LI << '\n');
1581 CriticalEdgePred.push_back(Pred);
1583 // Only add the predecessors that will not be split for now.
1584 PredLoads[Pred] = nullptr;
1588 // Decide whether PRE is profitable for this load.
1589 unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1590 assert(NumUnavailablePreds != 0 &&
1591 "Fully available value should already be eliminated!");
1593 // If this load is unavailable in multiple predecessors, reject it.
1594 // FIXME: If we could restructure the CFG, we could make a common pred with
1595 // all the preds that don't have an available LI and insert a new load into
1597 if (NumUnavailablePreds != 1)
1600 // Split critical edges, and update the unavailable predecessors accordingly.
1601 for (BasicBlock *OrigPred : CriticalEdgePred) {
1602 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1603 assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1604 PredLoads[NewPred] = nullptr;
1605 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1606 << LoadBB->getName() << '\n');
1609 // Check if the load can safely be moved to all the unavailable predecessors.
1610 bool CanDoPRE = true;
1611 const DataLayout &DL = LI->getModule()->getDataLayout();
1612 SmallVector<Instruction*, 8> NewInsts;
1613 for (auto &PredLoad : PredLoads) {
1614 BasicBlock *UnavailablePred = PredLoad.first;
1616 // Do PHI translation to get its value in the predecessor if necessary. The
1617 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1619 // If all preds have a single successor, then we know it is safe to insert
1620 // the load on the pred (?!?), so we can insert code to materialize the
1621 // pointer if it is not available.
1622 PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1623 Value *LoadPtr = nullptr;
1624 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1627 // If we couldn't find or insert a computation of this phi translated value,
1630 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1631 << *LI->getPointerOperand() << "\n");
1636 PredLoad.second = LoadPtr;
1640 while (!NewInsts.empty()) {
1641 Instruction *I = NewInsts.pop_back_val();
1642 if (MD) MD->removeInstruction(I);
1643 I->eraseFromParent();
1645 // HINT: Don't revert the edge-splitting as following transformation may
1646 // also need to split these critical edges.
1647 return !CriticalEdgePred.empty();
1650 // Okay, we can eliminate this load by inserting a reload in the predecessor
1651 // and using PHI construction to get the value in the other predecessors, do
1653 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1654 DEBUG(if (!NewInsts.empty())
1655 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1656 << *NewInsts.back() << '\n');
1658 // Assign value numbers to the new instructions.
1659 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1660 // FIXME: We really _ought_ to insert these value numbers into their
1661 // parent's availability map. However, in doing so, we risk getting into
1662 // ordering issues. If a block hasn't been processed yet, we would be
1663 // marking a value as AVAIL-IN, which isn't what we intend.
1664 VN.lookup_or_add(NewInsts[i]);
1667 for (const auto &PredLoad : PredLoads) {
1668 BasicBlock *UnavailablePred = PredLoad.first;
1669 Value *LoadPtr = PredLoad.second;
1671 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1673 UnavailablePred->getTerminator());
1675 // Transfer the old load's AA tags to the new load.
1677 LI->getAAMetadata(Tags);
1679 NewLoad->setAAMetadata(Tags);
1681 // Transfer DebugLoc.
1682 NewLoad->setDebugLoc(LI->getDebugLoc());
1684 // Add the newly created load.
1685 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1687 MD->invalidateCachedPointerInfo(LoadPtr);
1688 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1691 // Perform PHI construction.
1692 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1693 LI->replaceAllUsesWith(V);
1694 if (isa<PHINode>(V))
1696 if (Instruction *I = dyn_cast<Instruction>(V))
1697 I->setDebugLoc(LI->getDebugLoc());
1698 if (V->getType()->getScalarType()->isPointerTy())
1699 MD->invalidateCachedPointerInfo(V);
1700 markInstructionForDeletion(LI);
1705 /// Attempt to eliminate a load whose dependencies are
1706 /// non-local by performing PHI construction.
1707 bool GVN::processNonLocalLoad(LoadInst *LI) {
1708 // Step 1: Find the non-local dependencies of the load.
1710 MD->getNonLocalPointerDependency(LI, Deps);
1712 // If we had to process more than one hundred blocks to find the
1713 // dependencies, this load isn't worth worrying about. Optimizing
1714 // it will be too expensive.
1715 unsigned NumDeps = Deps.size();
1719 // If we had a phi translation failure, we'll have a single entry which is a
1720 // clobber in the current block. Reject this early.
1722 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1724 dbgs() << "GVN: non-local load ";
1725 LI->printAsOperand(dbgs());
1726 dbgs() << " has unknown dependencies\n";
1731 // If this load follows a GEP, see if we can PRE the indices before analyzing.
1732 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1733 for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1734 OE = GEP->idx_end();
1736 if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1737 performScalarPRE(I);
1740 // Step 2: Analyze the availability of the load
1741 AvailValInBlkVect ValuesPerBlock;
1742 UnavailBlkVect UnavailableBlocks;
1743 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1745 // If we have no predecessors that produce a known value for this load, exit
1747 if (ValuesPerBlock.empty())
1750 // Step 3: Eliminate fully redundancy.
1752 // If all of the instructions we depend on produce a known value for this
1753 // load, then it is fully redundant and we can use PHI insertion to compute
1754 // its value. Insert PHIs and remove the fully redundant value now.
1755 if (UnavailableBlocks.empty()) {
1756 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1758 // Perform PHI construction.
1759 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1760 LI->replaceAllUsesWith(V);
1762 if (isa<PHINode>(V))
1764 if (Instruction *I = dyn_cast<Instruction>(V))
1765 I->setDebugLoc(LI->getDebugLoc());
1766 if (V->getType()->getScalarType()->isPointerTy())
1767 MD->invalidateCachedPointerInfo(V);
1768 markInstructionForDeletion(LI);
1773 // Step 4: Eliminate partial redundancy.
1774 if (!EnablePRE || !EnableLoadPRE)
1777 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1781 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1782 // Patch the replacement so that it is not more restrictive than the value
1784 BinaryOperator *Op = dyn_cast<BinaryOperator>(I);
1785 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl);
1786 if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) &&
1787 isa<OverflowingBinaryOperator>(ReplOp)) {
1788 if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap())
1789 ReplOp->setHasNoSignedWrap(false);
1790 if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap())
1791 ReplOp->setHasNoUnsignedWrap(false);
1793 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) {
1794 // FIXME: If both the original and replacement value are part of the
1795 // same control-flow region (meaning that the execution of one
1796 // guarentees the executation of the other), then we can combine the
1797 // noalias scopes here and do better than the general conservative
1798 // answer used in combineMetadata().
1800 // In general, GVN unifies expressions over different control-flow
1801 // regions, and so we need a conservative combination of the noalias
1803 unsigned KnownIDs[] = {
1804 LLVMContext::MD_tbaa,
1805 LLVMContext::MD_alias_scope,
1806 LLVMContext::MD_noalias,
1807 LLVMContext::MD_range,
1808 LLVMContext::MD_fpmath,
1809 LLVMContext::MD_invariant_load,
1811 combineMetadata(ReplInst, I, KnownIDs);
1815 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1816 patchReplacementInstruction(I, Repl);
1817 I->replaceAllUsesWith(Repl);
1820 /// Attempt to eliminate a load, first by eliminating it
1821 /// locally, and then attempting non-local elimination if that fails.
1822 bool GVN::processLoad(LoadInst *L) {
1829 if (L->use_empty()) {
1830 markInstructionForDeletion(L);
1834 // ... to a pointer that has been loaded from before...
1835 MemDepResult Dep = MD->getDependency(L);
1836 const DataLayout &DL = L->getModule()->getDataLayout();
1838 // If we have a clobber and target data is around, see if this is a clobber
1839 // that we can fix up through code synthesis.
1840 if (Dep.isClobber()) {
1841 // Check to see if we have something like this:
1842 // store i32 123, i32* %P
1843 // %A = bitcast i32* %P to i8*
1844 // %B = gep i8* %A, i32 1
1847 // We could do that by recognizing if the clobber instructions are obviously
1848 // a common base + constant offset, and if the previous store (or memset)
1849 // completely covers this load. This sort of thing can happen in bitfield
1851 Value *AvailVal = nullptr;
1852 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1853 int Offset = AnalyzeLoadFromClobberingStore(
1854 L->getType(), L->getPointerOperand(), DepSI);
1856 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1857 L->getType(), L, DL);
1860 // Check to see if we have something like this:
1863 // if we have this, replace the later with an extraction from the former.
1864 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1865 // If this is a clobber and L is the first instruction in its block, then
1866 // we have the first instruction in the entry block.
1870 int Offset = AnalyzeLoadFromClobberingLoad(
1871 L->getType(), L->getPointerOperand(), DepLI, DL);
1873 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1876 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1877 // a value on from it.
1878 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1879 int Offset = AnalyzeLoadFromClobberingMemInst(
1880 L->getType(), L->getPointerOperand(), DepMI, DL);
1882 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, DL);
1886 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1887 << *AvailVal << '\n' << *L << "\n\n\n");
1889 // Replace the load!
1890 L->replaceAllUsesWith(AvailVal);
1891 if (AvailVal->getType()->getScalarType()->isPointerTy())
1892 MD->invalidateCachedPointerInfo(AvailVal);
1893 markInstructionForDeletion(L);
1899 // If the value isn't available, don't do anything!
1900 if (Dep.isClobber()) {
1902 // fast print dep, using operator<< on instruction is too slow.
1903 dbgs() << "GVN: load ";
1904 L->printAsOperand(dbgs());
1905 Instruction *I = Dep.getInst();
1906 dbgs() << " is clobbered by " << *I << '\n';
1911 // If it is defined in another block, try harder.
1912 if (Dep.isNonLocal())
1913 return processNonLocalLoad(L);
1917 // fast print dep, using operator<< on instruction is too slow.
1918 dbgs() << "GVN: load ";
1919 L->printAsOperand(dbgs());
1920 dbgs() << " has unknown dependence\n";
1925 Instruction *DepInst = Dep.getInst();
1926 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1927 Value *StoredVal = DepSI->getValueOperand();
1929 // The store and load are to a must-aliased pointer, but they may not
1930 // actually have the same type. See if we know how to reuse the stored
1931 // value (depending on its type).
1932 if (StoredVal->getType() != L->getType()) {
1933 IRBuilder<> Builder(L);
1935 CoerceAvailableValueToLoadType(StoredVal, L->getType(), Builder, DL);
1939 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1940 << '\n' << *L << "\n\n\n");
1944 L->replaceAllUsesWith(StoredVal);
1945 if (StoredVal->getType()->getScalarType()->isPointerTy())
1946 MD->invalidateCachedPointerInfo(StoredVal);
1947 markInstructionForDeletion(L);
1952 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1953 Value *AvailableVal = DepLI;
1955 // The loads are of a must-aliased pointer, but they may not actually have
1956 // the same type. See if we know how to reuse the previously loaded value
1957 // (depending on its type).
1958 if (DepLI->getType() != L->getType()) {
1959 IRBuilder<> Builder(L);
1961 CoerceAvailableValueToLoadType(DepLI, L->getType(), Builder, DL);
1965 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1966 << "\n" << *L << "\n\n\n");
1970 patchAndReplaceAllUsesWith(L, AvailableVal);
1971 if (DepLI->getType()->getScalarType()->isPointerTy())
1972 MD->invalidateCachedPointerInfo(DepLI);
1973 markInstructionForDeletion(L);
1978 // If this load really doesn't depend on anything, then we must be loading an
1979 // undef value. This can happen when loading for a fresh allocation with no
1980 // intervening stores, for example.
1981 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
1982 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1983 markInstructionForDeletion(L);
1988 // If this load occurs either right after a lifetime begin,
1989 // then the loaded value is undefined.
1990 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
1991 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1992 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1993 markInstructionForDeletion(L);
1999 // If this load follows a calloc (which zero initializes memory),
2000 // then the loaded value is zero
2001 if (isCallocLikeFn(DepInst, TLI)) {
2002 L->replaceAllUsesWith(Constant::getNullValue(L->getType()));
2003 markInstructionForDeletion(L);
2011 // In order to find a leader for a given value number at a
2012 // specific basic block, we first obtain the list of all Values for that number,
2013 // and then scan the list to find one whose block dominates the block in
2014 // question. This is fast because dominator tree queries consist of only
2015 // a few comparisons of DFS numbers.
2016 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
2017 LeaderTableEntry Vals = LeaderTable[num];
2018 if (!Vals.Val) return nullptr;
2020 Value *Val = nullptr;
2021 if (DT->dominates(Vals.BB, BB)) {
2023 if (isa<Constant>(Val)) return Val;
2026 LeaderTableEntry* Next = Vals.Next;
2028 if (DT->dominates(Next->BB, BB)) {
2029 if (isa<Constant>(Next->Val)) return Next->Val;
2030 if (!Val) Val = Next->Val;
2039 /// There is an edge from 'Src' to 'Dst'. Return
2040 /// true if every path from the entry block to 'Dst' passes via this edge. In
2041 /// particular 'Dst' must not be reachable via another edge from 'Src'.
2042 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
2043 DominatorTree *DT) {
2044 // While in theory it is interesting to consider the case in which Dst has
2045 // more than one predecessor, because Dst might be part of a loop which is
2046 // only reachable from Src, in practice it is pointless since at the time
2047 // GVN runs all such loops have preheaders, which means that Dst will have
2048 // been changed to have only one predecessor, namely Src.
2049 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
2050 const BasicBlock *Src = E.getStart();
2051 assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
2053 return Pred != nullptr;
2056 /// The given values are known to be equal in every block
2057 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
2058 /// 'RHS' everywhere in the scope. Returns whether a change was made.
2059 bool GVN::propagateEquality(Value *LHS, Value *RHS,
2060 const BasicBlockEdge &Root) {
2061 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
2062 Worklist.push_back(std::make_pair(LHS, RHS));
2063 bool Changed = false;
2064 // For speed, compute a conservative fast approximation to
2065 // DT->dominates(Root, Root.getEnd());
2066 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
2068 while (!Worklist.empty()) {
2069 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
2070 LHS = Item.first; RHS = Item.second;
2072 if (LHS == RHS) continue;
2073 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
2075 // Don't try to propagate equalities between constants.
2076 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue;
2078 // Prefer a constant on the right-hand side, or an Argument if no constants.
2079 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
2080 std::swap(LHS, RHS);
2081 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
2083 // If there is no obvious reason to prefer the left-hand side over the
2084 // right-hand side, ensure the longest lived term is on the right-hand side,
2085 // so the shortest lived term will be replaced by the longest lived.
2086 // This tends to expose more simplifications.
2087 uint32_t LVN = VN.lookup_or_add(LHS);
2088 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
2089 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
2090 // Move the 'oldest' value to the right-hand side, using the value number
2091 // as a proxy for age.
2092 uint32_t RVN = VN.lookup_or_add(RHS);
2094 std::swap(LHS, RHS);
2099 // If value numbering later sees that an instruction in the scope is equal
2100 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
2101 // the invariant that instructions only occur in the leader table for their
2102 // own value number (this is used by removeFromLeaderTable), do not do this
2103 // if RHS is an instruction (if an instruction in the scope is morphed into
2104 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
2105 // using the leader table is about compiling faster, not optimizing better).
2106 // The leader table only tracks basic blocks, not edges. Only add to if we
2107 // have the simple case where the edge dominates the end.
2108 if (RootDominatesEnd && !isa<Instruction>(RHS))
2109 addToLeaderTable(LVN, RHS, Root.getEnd());
2111 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
2112 // LHS always has at least one use that is not dominated by Root, this will
2113 // never do anything if LHS has only one use.
2114 if (!LHS->hasOneUse()) {
2115 unsigned NumReplacements = replaceDominatedUsesWith(LHS, RHS, *DT, Root);
2116 Changed |= NumReplacements > 0;
2117 NumGVNEqProp += NumReplacements;
2120 // Now try to deduce additional equalities from this one. For example, if
2121 // the known equality was "(A != B)" == "false" then it follows that A and B
2122 // are equal in the scope. Only boolean equalities with an explicit true or
2123 // false RHS are currently supported.
2124 if (!RHS->getType()->isIntegerTy(1))
2125 // Not a boolean equality - bail out.
2127 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
2129 // RHS neither 'true' nor 'false' - bail out.
2131 // Whether RHS equals 'true'. Otherwise it equals 'false'.
2132 bool isKnownTrue = CI->isAllOnesValue();
2133 bool isKnownFalse = !isKnownTrue;
2135 // If "A && B" is known true then both A and B are known true. If "A || B"
2136 // is known false then both A and B are known false.
2138 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
2139 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
2140 Worklist.push_back(std::make_pair(A, RHS));
2141 Worklist.push_back(std::make_pair(B, RHS));
2145 // If we are propagating an equality like "(A == B)" == "true" then also
2146 // propagate the equality A == B. When propagating a comparison such as
2147 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
2148 if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
2149 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
2151 // If "A == B" is known true, or "A != B" is known false, then replace
2152 // A with B everywhere in the scope.
2153 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2154 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2155 Worklist.push_back(std::make_pair(Op0, Op1));
2157 // Handle the floating point versions of equality comparisons too.
2158 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
2159 (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
2161 // Floating point -0.0 and 0.0 compare equal, so we can only
2162 // propagate values if we know that we have a constant and that
2163 // its value is non-zero.
2165 // FIXME: We should do this optimization if 'no signed zeros' is
2166 // applicable via an instruction-level fast-math-flag or some other
2167 // indicator that relaxed FP semantics are being used.
2169 if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
2170 Worklist.push_back(std::make_pair(Op0, Op1));
2173 // If "A >= B" is known true, replace "A < B" with false everywhere.
2174 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2175 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2176 // Since we don't have the instruction "A < B" immediately to hand, work
2177 // out the value number that it would have and use that to find an
2178 // appropriate instruction (if any).
2179 uint32_t NextNum = VN.getNextUnusedValueNumber();
2180 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2181 // If the number we were assigned was brand new then there is no point in
2182 // looking for an instruction realizing it: there cannot be one!
2183 if (Num < NextNum) {
2184 Value *NotCmp = findLeader(Root.getEnd(), Num);
2185 if (NotCmp && isa<Instruction>(NotCmp)) {
2186 unsigned NumReplacements =
2187 replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root);
2188 Changed |= NumReplacements > 0;
2189 NumGVNEqProp += NumReplacements;
2192 // Ensure that any instruction in scope that gets the "A < B" value number
2193 // is replaced with false.
2194 // The leader table only tracks basic blocks, not edges. Only add to if we
2195 // have the simple case where the edge dominates the end.
2196 if (RootDominatesEnd)
2197 addToLeaderTable(Num, NotVal, Root.getEnd());
2206 /// When calculating availability, handle an instruction
2207 /// by inserting it into the appropriate sets
2208 bool GVN::processInstruction(Instruction *I) {
2209 // Ignore dbg info intrinsics.
2210 if (isa<DbgInfoIntrinsic>(I))
2213 // If the instruction can be easily simplified then do so now in preference
2214 // to value numbering it. Value numbering often exposes redundancies, for
2215 // example if it determines that %y is equal to %x then the instruction
2216 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2217 const DataLayout &DL = I->getModule()->getDataLayout();
2218 if (Value *V = SimplifyInstruction(I, DL, TLI, DT, AC)) {
2219 I->replaceAllUsesWith(V);
2220 if (MD && V->getType()->getScalarType()->isPointerTy())
2221 MD->invalidateCachedPointerInfo(V);
2222 markInstructionForDeletion(I);
2227 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2228 if (processLoad(LI))
2231 unsigned Num = VN.lookup_or_add(LI);
2232 addToLeaderTable(Num, LI, LI->getParent());
2236 // For conditional branches, we can perform simple conditional propagation on
2237 // the condition value itself.
2238 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2239 if (!BI->isConditional())
2242 if (isa<Constant>(BI->getCondition()))
2243 return processFoldableCondBr(BI);
2245 Value *BranchCond = BI->getCondition();
2246 BasicBlock *TrueSucc = BI->getSuccessor(0);
2247 BasicBlock *FalseSucc = BI->getSuccessor(1);
2248 // Avoid multiple edges early.
2249 if (TrueSucc == FalseSucc)
2252 BasicBlock *Parent = BI->getParent();
2253 bool Changed = false;
2255 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
2256 BasicBlockEdge TrueE(Parent, TrueSucc);
2257 Changed |= propagateEquality(BranchCond, TrueVal, TrueE);
2259 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
2260 BasicBlockEdge FalseE(Parent, FalseSucc);
2261 Changed |= propagateEquality(BranchCond, FalseVal, FalseE);
2266 // For switches, propagate the case values into the case destinations.
2267 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2268 Value *SwitchCond = SI->getCondition();
2269 BasicBlock *Parent = SI->getParent();
2270 bool Changed = false;
2272 // Remember how many outgoing edges there are to every successor.
2273 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2274 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
2275 ++SwitchEdges[SI->getSuccessor(i)];
2277 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2279 BasicBlock *Dst = i.getCaseSuccessor();
2280 // If there is only a single edge, propagate the case value into it.
2281 if (SwitchEdges.lookup(Dst) == 1) {
2282 BasicBlockEdge E(Parent, Dst);
2283 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E);
2289 // Instructions with void type don't return a value, so there's
2290 // no point in trying to find redundancies in them.
2291 if (I->getType()->isVoidTy()) return false;
2293 uint32_t NextNum = VN.getNextUnusedValueNumber();
2294 unsigned Num = VN.lookup_or_add(I);
2296 // Allocations are always uniquely numbered, so we can save time and memory
2297 // by fast failing them.
2298 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2299 addToLeaderTable(Num, I, I->getParent());
2303 // If the number we were assigned was a brand new VN, then we don't
2304 // need to do a lookup to see if the number already exists
2305 // somewhere in the domtree: it can't!
2306 if (Num >= NextNum) {
2307 addToLeaderTable(Num, I, I->getParent());
2311 // Perform fast-path value-number based elimination of values inherited from
2313 Value *repl = findLeader(I->getParent(), Num);
2315 // Failure, just remember this instance for future use.
2316 addToLeaderTable(Num, I, I->getParent());
2321 patchAndReplaceAllUsesWith(I, repl);
2322 if (MD && repl->getType()->getScalarType()->isPointerTy())
2323 MD->invalidateCachedPointerInfo(repl);
2324 markInstructionForDeletion(I);
2328 /// runOnFunction - This is the main transformation entry point for a function.
2329 bool GVN::runOnFunction(Function& F) {
2330 if (skipOptnoneFunction(F))
2334 MD = &getAnalysis<MemoryDependenceAnalysis>();
2335 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2336 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2337 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
2338 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2342 bool Changed = false;
2343 bool ShouldContinue = true;
2345 // Merge unconditional branches, allowing PRE to catch more
2346 // optimization opportunities.
2347 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2348 BasicBlock *BB = FI++;
2350 bool removedBlock = MergeBlockIntoPredecessor(
2351 BB, DT, /* LoopInfo */ nullptr, VN.getAliasAnalysis(), MD);
2352 if (removedBlock) ++NumGVNBlocks;
2354 Changed |= removedBlock;
2357 unsigned Iteration = 0;
2358 while (ShouldContinue) {
2359 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2360 ShouldContinue = iterateOnFunction(F);
2361 Changed |= ShouldContinue;
2366 // Fabricate val-num for dead-code in order to suppress assertion in
2368 assignValNumForDeadCode();
2369 bool PREChanged = true;
2370 while (PREChanged) {
2371 PREChanged = performPRE(F);
2372 Changed |= PREChanged;
2376 // FIXME: Should perform GVN again after PRE does something. PRE can move
2377 // computations into blocks where they become fully redundant. Note that
2378 // we can't do this until PRE's critical edge splitting updates memdep.
2379 // Actually, when this happens, we should just fully integrate PRE into GVN.
2381 cleanupGlobalSets();
2382 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2390 bool GVN::processBlock(BasicBlock *BB) {
2391 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2392 // (and incrementing BI before processing an instruction).
2393 assert(InstrsToErase.empty() &&
2394 "We expect InstrsToErase to be empty across iterations");
2395 if (DeadBlocks.count(BB))
2398 bool ChangedFunction = false;
2400 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2402 ChangedFunction |= processInstruction(BI);
2403 if (InstrsToErase.empty()) {
2408 // If we need some instructions deleted, do it now.
2409 NumGVNInstr += InstrsToErase.size();
2411 // Avoid iterator invalidation.
2412 bool AtStart = BI == BB->begin();
2416 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2417 E = InstrsToErase.end(); I != E; ++I) {
2418 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2419 if (MD) MD->removeInstruction(*I);
2420 DEBUG(verifyRemoved(*I));
2421 (*I)->eraseFromParent();
2423 InstrsToErase.clear();
2431 return ChangedFunction;
2434 // Instantiate an expression in a predecessor that lacked it.
2435 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2436 unsigned int ValNo) {
2437 // Because we are going top-down through the block, all value numbers
2438 // will be available in the predecessor by the time we need them. Any
2439 // that weren't originally present will have been instantiated earlier
2441 bool success = true;
2442 for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2443 Value *Op = Instr->getOperand(i);
2444 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2447 if (Value *V = findLeader(Pred, VN.lookup(Op))) {
2448 Instr->setOperand(i, V);
2455 // Fail out if we encounter an operand that is not available in
2456 // the PRE predecessor. This is typically because of loads which
2457 // are not value numbered precisely.
2461 Instr->insertBefore(Pred->getTerminator());
2462 Instr->setName(Instr->getName() + ".pre");
2463 Instr->setDebugLoc(Instr->getDebugLoc());
2464 VN.add(Instr, ValNo);
2466 // Update the availability map to include the new instruction.
2467 addToLeaderTable(ValNo, Instr, Pred);
2471 bool GVN::performScalarPRE(Instruction *CurInst) {
2472 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap;
2474 if (isa<AllocaInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
2475 isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2476 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2477 isa<DbgInfoIntrinsic>(CurInst))
2480 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2481 // sinking the compare again, and it would force the code generator to
2482 // move the i1 from processor flags or predicate registers into a general
2483 // purpose register.
2484 if (isa<CmpInst>(CurInst))
2487 // We don't currently value number ANY inline asm calls.
2488 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2489 if (CallI->isInlineAsm())
2492 uint32_t ValNo = VN.lookup(CurInst);
2494 // Look for the predecessors for PRE opportunities. We're
2495 // only trying to solve the basic diamond case, where
2496 // a value is computed in the successor and one predecessor,
2497 // but not the other. We also explicitly disallow cases
2498 // where the successor is its own predecessor, because they're
2499 // more complicated to get right.
2500 unsigned NumWith = 0;
2501 unsigned NumWithout = 0;
2502 BasicBlock *PREPred = nullptr;
2503 BasicBlock *CurrentBlock = CurInst->getParent();
2506 for (pred_iterator PI = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock);
2508 BasicBlock *P = *PI;
2509 // We're not interested in PRE where the block is its
2510 // own predecessor, or in blocks with predecessors
2511 // that are not reachable.
2512 if (P == CurrentBlock) {
2515 } else if (!DT->isReachableFromEntry(P)) {
2520 Value *predV = findLeader(P, ValNo);
2522 predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2525 } else if (predV == CurInst) {
2526 /* CurInst dominates this predecessor. */
2530 predMap.push_back(std::make_pair(predV, P));
2535 // Don't do PRE when it might increase code size, i.e. when
2536 // we would need to insert instructions in more than one pred.
2537 if (NumWithout > 1 || NumWith == 0)
2540 // We may have a case where all predecessors have the instruction,
2541 // and we just need to insert a phi node. Otherwise, perform
2543 Instruction *PREInstr = nullptr;
2545 if (NumWithout != 0) {
2546 // Don't do PRE across indirect branch.
2547 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2550 // We can't do PRE safely on a critical edge, so instead we schedule
2551 // the edge to be split and perform the PRE the next time we iterate
2553 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2554 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2555 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2558 // We need to insert somewhere, so let's give it a shot
2559 PREInstr = CurInst->clone();
2560 if (!performScalarPREInsertion(PREInstr, PREPred, ValNo)) {
2561 // If we failed insertion, make sure we remove the instruction.
2562 DEBUG(verifyRemoved(PREInstr));
2568 // Either we should have filled in the PRE instruction, or we should
2569 // not have needed insertions.
2570 assert (PREInstr != nullptr || NumWithout == 0);
2574 // Create a PHI to make the value available in this block.
2576 PHINode::Create(CurInst->getType(), predMap.size(),
2577 CurInst->getName() + ".pre-phi", CurrentBlock->begin());
2578 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2579 if (Value *V = predMap[i].first)
2580 Phi->addIncoming(V, predMap[i].second);
2582 Phi->addIncoming(PREInstr, PREPred);
2586 addToLeaderTable(ValNo, Phi, CurrentBlock);
2587 Phi->setDebugLoc(CurInst->getDebugLoc());
2588 CurInst->replaceAllUsesWith(Phi);
2589 if (Phi->getType()->getScalarType()->isPointerTy()) {
2590 // Because we have added a PHI-use of the pointer value, it has now
2591 // "escaped" from alias analysis' perspective. We need to inform
2593 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee; ++ii) {
2594 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
2595 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
2599 MD->invalidateCachedPointerInfo(Phi);
2602 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2604 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2606 MD->removeInstruction(CurInst);
2607 DEBUG(verifyRemoved(CurInst));
2608 CurInst->eraseFromParent();
2614 /// Perform a purely local form of PRE that looks for diamond
2615 /// control flow patterns and attempts to perform simple PRE at the join point.
2616 bool GVN::performPRE(Function &F) {
2617 bool Changed = false;
2618 for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2619 // Nothing to PRE in the entry block.
2620 if (CurrentBlock == &F.getEntryBlock())
2623 // Don't perform PRE on a landing pad.
2624 if (CurrentBlock->isLandingPad())
2627 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2628 BE = CurrentBlock->end();
2630 Instruction *CurInst = BI++;
2631 Changed = performScalarPRE(CurInst);
2635 if (splitCriticalEdges())
2641 /// Split the critical edge connecting the given two blocks, and return
2642 /// the block inserted to the critical edge.
2643 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2644 BasicBlock *BB = SplitCriticalEdge(
2645 Pred, Succ, CriticalEdgeSplittingOptions(getAliasAnalysis(), DT));
2647 MD->invalidateCachedPredecessors();
2651 /// Split critical edges found during the previous
2652 /// iteration that may enable further optimization.
2653 bool GVN::splitCriticalEdges() {
2654 if (toSplit.empty())
2657 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2658 SplitCriticalEdge(Edge.first, Edge.second,
2659 CriticalEdgeSplittingOptions(getAliasAnalysis(), DT));
2660 } while (!toSplit.empty());
2661 if (MD) MD->invalidateCachedPredecessors();
2665 /// Executes one iteration of GVN
2666 bool GVN::iterateOnFunction(Function &F) {
2667 cleanupGlobalSets();
2669 // Top-down walk of the dominator tree
2670 bool Changed = false;
2671 // Save the blocks this function have before transformation begins. GVN may
2672 // split critical edge, and hence may invalidate the RPO/DT iterator.
2674 std::vector<BasicBlock *> BBVect;
2675 BBVect.reserve(256);
2676 // Needed for value numbering with phi construction to work.
2677 ReversePostOrderTraversal<Function *> RPOT(&F);
2678 for (ReversePostOrderTraversal<Function *>::rpo_iterator RI = RPOT.begin(),
2681 BBVect.push_back(*RI);
2683 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
2685 Changed |= processBlock(*I);
2690 void GVN::cleanupGlobalSets() {
2692 LeaderTable.clear();
2693 TableAllocator.Reset();
2696 /// Verify that the specified instruction does not occur in our
2697 /// internal data structures.
2698 void GVN::verifyRemoved(const Instruction *Inst) const {
2699 VN.verifyRemoved(Inst);
2701 // Walk through the value number scope to make sure the instruction isn't
2702 // ferreted away in it.
2703 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2704 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2705 const LeaderTableEntry *Node = &I->second;
2706 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2708 while (Node->Next) {
2710 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2715 /// BB is declared dead, which implied other blocks become dead as well. This
2716 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2717 /// live successors, update their phi nodes by replacing the operands
2718 /// corresponding to dead blocks with UndefVal.
2719 void GVN::addDeadBlock(BasicBlock *BB) {
2720 SmallVector<BasicBlock *, 4> NewDead;
2721 SmallSetVector<BasicBlock *, 4> DF;
2723 NewDead.push_back(BB);
2724 while (!NewDead.empty()) {
2725 BasicBlock *D = NewDead.pop_back_val();
2726 if (DeadBlocks.count(D))
2729 // All blocks dominated by D are dead.
2730 SmallVector<BasicBlock *, 8> Dom;
2731 DT->getDescendants(D, Dom);
2732 DeadBlocks.insert(Dom.begin(), Dom.end());
2734 // Figure out the dominance-frontier(D).
2735 for (SmallVectorImpl<BasicBlock *>::iterator I = Dom.begin(),
2736 E = Dom.end(); I != E; I++) {
2738 for (succ_iterator SI = succ_begin(B), SE = succ_end(B); SI != SE; SI++) {
2739 BasicBlock *S = *SI;
2740 if (DeadBlocks.count(S))
2743 bool AllPredDead = true;
2744 for (pred_iterator PI = pred_begin(S), PE = pred_end(S); PI != PE; PI++)
2745 if (!DeadBlocks.count(*PI)) {
2746 AllPredDead = false;
2751 // S could be proved dead later on. That is why we don't update phi
2752 // operands at this moment.
2755 // While S is not dominated by D, it is dead by now. This could take
2756 // place if S already have a dead predecessor before D is declared
2758 NewDead.push_back(S);
2764 // For the dead blocks' live successors, update their phi nodes by replacing
2765 // the operands corresponding to dead blocks with UndefVal.
2766 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2769 if (DeadBlocks.count(B))
2772 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2773 for (SmallVectorImpl<BasicBlock *>::iterator PI = Preds.begin(),
2774 PE = Preds.end(); PI != PE; PI++) {
2775 BasicBlock *P = *PI;
2777 if (!DeadBlocks.count(P))
2780 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2781 if (BasicBlock *S = splitCriticalEdges(P, B))
2782 DeadBlocks.insert(P = S);
2785 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2786 PHINode &Phi = cast<PHINode>(*II);
2787 Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2788 UndefValue::get(Phi.getType()));
2794 // If the given branch is recognized as a foldable branch (i.e. conditional
2795 // branch with constant condition), it will perform following analyses and
2797 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2798 // R be the target of the dead out-coming edge.
2799 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2800 // edge. The result of this step will be {X| X is dominated by R}
2801 // 2) Identify those blocks which haves at least one dead prodecessor. The
2802 // result of this step will be dominance-frontier(R).
2803 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2804 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2806 // Return true iff *NEW* dead code are found.
2807 bool GVN::processFoldableCondBr(BranchInst *BI) {
2808 if (!BI || BI->isUnconditional())
2811 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2815 BasicBlock *DeadRoot = Cond->getZExtValue() ?
2816 BI->getSuccessor(1) : BI->getSuccessor(0);
2817 if (DeadBlocks.count(DeadRoot))
2820 if (!DeadRoot->getSinglePredecessor())
2821 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2823 addDeadBlock(DeadRoot);
2827 // performPRE() will trigger assert if it comes across an instruction without
2828 // associated val-num. As it normally has far more live instructions than dead
2829 // instructions, it makes more sense just to "fabricate" a val-number for the
2830 // dead code than checking if instruction involved is dead or not.
2831 void GVN::assignValNumForDeadCode() {
2832 for (SetVector<BasicBlock *>::iterator I = DeadBlocks.begin(),
2833 E = DeadBlocks.end(); I != E; I++) {
2834 BasicBlock *BB = *I;
2835 for (BasicBlock::iterator II = BB->begin(), EE = BB->end();
2837 Instruction *Inst = &*II;
2838 unsigned ValNum = VN.lookup_or_add(Inst);
2839 addToLeaderTable(ValNum, Inst, BB);