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 // Block-local map of equivalent values to their leader, does not
612 // propagate to any successors. Entries added mid-block are applied
613 // to the remaining instructions in the block.
614 SmallMapVector<llvm::Value *, llvm::Constant *, 4> ReplaceWithConstMap;
615 SmallVector<Instruction*, 8> InstrsToErase;
617 typedef SmallVector<NonLocalDepResult, 64> LoadDepVect;
618 typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect;
619 typedef SmallVector<BasicBlock*, 64> UnavailBlkVect;
622 static char ID; // Pass identification, replacement for typeid
623 explicit GVN(bool noloads = false)
624 : FunctionPass(ID), NoLoads(noloads), MD(nullptr) {
625 initializeGVNPass(*PassRegistry::getPassRegistry());
628 bool runOnFunction(Function &F) override;
630 /// This removes the specified instruction from
631 /// our various maps and marks it for deletion.
632 void markInstructionForDeletion(Instruction *I) {
634 InstrsToErase.push_back(I);
637 DominatorTree &getDominatorTree() const { return *DT; }
638 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
639 MemoryDependenceAnalysis &getMemDep() const { return *MD; }
641 /// Push a new Value to the LeaderTable onto the list for its value number.
642 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) {
643 LeaderTableEntry &Curr = LeaderTable[N];
650 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
653 Node->Next = Curr.Next;
657 /// Scan the list of values corresponding to a given
658 /// value number, and remove the given instruction if encountered.
659 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) {
660 LeaderTableEntry* Prev = nullptr;
661 LeaderTableEntry* Curr = &LeaderTable[N];
663 while (Curr && (Curr->Val != I || Curr->BB != BB)) {
672 Prev->Next = Curr->Next;
678 LeaderTableEntry* Next = Curr->Next;
679 Curr->Val = Next->Val;
681 Curr->Next = Next->Next;
686 // List of critical edges to be split between iterations.
687 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
689 // This transformation requires dominator postdominator info
690 void getAnalysisUsage(AnalysisUsage &AU) const override {
691 AU.addRequired<AssumptionCacheTracker>();
692 AU.addRequired<DominatorTreeWrapperPass>();
693 AU.addRequired<TargetLibraryInfoWrapperPass>();
695 AU.addRequired<MemoryDependenceAnalysis>();
696 AU.addRequired<AliasAnalysis>();
698 AU.addPreserved<DominatorTreeWrapperPass>();
699 AU.addPreserved<AliasAnalysis>();
703 // Helper functions of redundant load elimination
704 bool processLoad(LoadInst *L);
705 bool processNonLocalLoad(LoadInst *L);
706 bool processAssumeIntrinsic(IntrinsicInst *II);
707 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
708 AvailValInBlkVect &ValuesPerBlock,
709 UnavailBlkVect &UnavailableBlocks);
710 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
711 UnavailBlkVect &UnavailableBlocks);
713 // Other helper routines
714 bool processInstruction(Instruction *I);
715 bool processBlock(BasicBlock *BB);
716 void dump(DenseMap<uint32_t, Value*> &d);
717 bool iterateOnFunction(Function &F);
718 bool performPRE(Function &F);
719 bool performScalarPRE(Instruction *I);
720 bool performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
722 Value *findLeader(const BasicBlock *BB, uint32_t num);
723 void cleanupGlobalSets();
724 void verifyRemoved(const Instruction *I) const;
725 bool splitCriticalEdges();
726 BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ);
727 bool replaceOperandsWithConsts(Instruction *I) const;
728 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
729 bool DominatesByEdge);
730 bool processFoldableCondBr(BranchInst *BI);
731 void addDeadBlock(BasicBlock *BB);
732 void assignValNumForDeadCode();
738 // The public interface to this file...
739 FunctionPass *llvm::createGVNPass(bool NoLoads) {
740 return new GVN(NoLoads);
743 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
744 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
745 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
746 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
747 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
748 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
749 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
751 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
752 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
754 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
755 E = d.end(); I != E; ++I) {
756 errs() << I->first << "\n";
763 /// Return true if we can prove that the value
764 /// we're analyzing is fully available in the specified block. As we go, keep
765 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
766 /// map is actually a tri-state map with the following values:
767 /// 0) we know the block *is not* fully available.
768 /// 1) we know the block *is* fully available.
769 /// 2) we do not know whether the block is fully available or not, but we are
770 /// currently speculating that it will be.
771 /// 3) we are speculating for this block and have used that to speculate for
773 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
774 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
775 uint32_t RecurseDepth) {
776 if (RecurseDepth > MaxRecurseDepth)
779 // Optimistically assume that the block is fully available and check to see
780 // if we already know about this block in one lookup.
781 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
782 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
784 // If the entry already existed for this block, return the precomputed value.
786 // If this is a speculative "available" value, mark it as being used for
787 // speculation of other blocks.
788 if (IV.first->second == 2)
789 IV.first->second = 3;
790 return IV.first->second != 0;
793 // Otherwise, see if it is fully available in all predecessors.
794 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
796 // If this block has no predecessors, it isn't live-in here.
798 goto SpeculationFailure;
800 for (; PI != PE; ++PI)
801 // If the value isn't fully available in one of our predecessors, then it
802 // isn't fully available in this block either. Undo our previous
803 // optimistic assumption and bail out.
804 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
805 goto SpeculationFailure;
809 // If we get here, we found out that this is not, after
810 // all, a fully-available block. We have a problem if we speculated on this and
811 // used the speculation to mark other blocks as available.
813 char &BBVal = FullyAvailableBlocks[BB];
815 // If we didn't speculate on this, just return with it set to false.
821 // If we did speculate on this value, we could have blocks set to 1 that are
822 // incorrect. Walk the (transitive) successors of this block and mark them as
824 SmallVector<BasicBlock*, 32> BBWorklist;
825 BBWorklist.push_back(BB);
828 BasicBlock *Entry = BBWorklist.pop_back_val();
829 // Note that this sets blocks to 0 (unavailable) if they happen to not
830 // already be in FullyAvailableBlocks. This is safe.
831 char &EntryVal = FullyAvailableBlocks[Entry];
832 if (EntryVal == 0) continue; // Already unavailable.
834 // Mark as unavailable.
837 BBWorklist.append(succ_begin(Entry), succ_end(Entry));
838 } while (!BBWorklist.empty());
844 /// Return true if CoerceAvailableValueToLoadType will succeed.
845 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
847 const DataLayout &DL) {
848 // If the loaded or stored value is an first class array or struct, don't try
849 // to transform them. We need to be able to bitcast to integer.
850 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
851 StoredVal->getType()->isStructTy() ||
852 StoredVal->getType()->isArrayTy())
855 // The store has to be at least as big as the load.
856 if (DL.getTypeSizeInBits(StoredVal->getType()) <
857 DL.getTypeSizeInBits(LoadTy))
863 /// If we saw a store of a value to memory, and
864 /// then a load from a must-aliased pointer of a different type, try to coerce
865 /// the stored value. LoadedTy is the type of the load we want to replace.
866 /// IRB is IRBuilder used to insert new instructions.
868 /// If we can't do it, return null.
869 static Value *CoerceAvailableValueToLoadType(Value *StoredVal, Type *LoadedTy,
871 const DataLayout &DL) {
872 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL))
875 // If this is already the right type, just return it.
876 Type *StoredValTy = StoredVal->getType();
878 uint64_t StoreSize = DL.getTypeSizeInBits(StoredValTy);
879 uint64_t LoadSize = DL.getTypeSizeInBits(LoadedTy);
881 // If the store and reload are the same size, we can always reuse it.
882 if (StoreSize == LoadSize) {
883 // Pointer to Pointer -> use bitcast.
884 if (StoredValTy->getScalarType()->isPointerTy() &&
885 LoadedTy->getScalarType()->isPointerTy())
886 return IRB.CreateBitCast(StoredVal, LoadedTy);
888 // Convert source pointers to integers, which can be bitcast.
889 if (StoredValTy->getScalarType()->isPointerTy()) {
890 StoredValTy = DL.getIntPtrType(StoredValTy);
891 StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
894 Type *TypeToCastTo = LoadedTy;
895 if (TypeToCastTo->getScalarType()->isPointerTy())
896 TypeToCastTo = DL.getIntPtrType(TypeToCastTo);
898 if (StoredValTy != TypeToCastTo)
899 StoredVal = IRB.CreateBitCast(StoredVal, TypeToCastTo);
901 // Cast to pointer if the load needs a pointer type.
902 if (LoadedTy->getScalarType()->isPointerTy())
903 StoredVal = IRB.CreateIntToPtr(StoredVal, LoadedTy);
908 // If the loaded value is smaller than the available value, then we can
909 // extract out a piece from it. If the available value is too small, then we
910 // can't do anything.
911 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
913 // Convert source pointers to integers, which can be manipulated.
914 if (StoredValTy->getScalarType()->isPointerTy()) {
915 StoredValTy = DL.getIntPtrType(StoredValTy);
916 StoredVal = IRB.CreatePtrToInt(StoredVal, StoredValTy);
919 // Convert vectors and fp to integer, which can be manipulated.
920 if (!StoredValTy->isIntegerTy()) {
921 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
922 StoredVal = IRB.CreateBitCast(StoredVal, StoredValTy);
925 // If this is a big-endian system, we need to shift the value down to the low
926 // bits so that a truncate will work.
927 if (DL.isBigEndian()) {
928 StoredVal = IRB.CreateLShr(StoredVal, StoreSize - LoadSize, "tmp");
931 // Truncate the integer to the right size now.
932 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
933 StoredVal = IRB.CreateTrunc(StoredVal, NewIntTy, "trunc");
935 if (LoadedTy == NewIntTy)
938 // If the result is a pointer, inttoptr.
939 if (LoadedTy->getScalarType()->isPointerTy())
940 return IRB.CreateIntToPtr(StoredVal, LoadedTy, "inttoptr");
942 // Otherwise, bitcast.
943 return IRB.CreateBitCast(StoredVal, LoadedTy, "bitcast");
946 /// This function is called when we have a
947 /// memdep query of a load that ends up being a clobbering memory write (store,
948 /// memset, memcpy, memmove). This means that the write *may* provide bits used
949 /// by the load but we can't be sure because the pointers don't mustalias.
951 /// Check this case to see if there is anything more we can do before we give
952 /// up. This returns -1 if we have to give up, or a byte number in the stored
953 /// value of the piece that feeds the load.
954 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
956 uint64_t WriteSizeInBits,
957 const DataLayout &DL) {
958 // If the loaded or stored value is a first class array or struct, don't try
959 // to transform them. We need to be able to bitcast to integer.
960 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
963 int64_t StoreOffset = 0, LoadOffset = 0;
965 GetPointerBaseWithConstantOffset(WritePtr, StoreOffset, DL);
966 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, DL);
967 if (StoreBase != LoadBase)
970 // If the load and store are to the exact same address, they should have been
971 // a must alias. AA must have gotten confused.
972 // FIXME: Study to see if/when this happens. One case is forwarding a memset
973 // to a load from the base of the memset.
975 if (LoadOffset == StoreOffset) {
976 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
977 << "Base = " << *StoreBase << "\n"
978 << "Store Ptr = " << *WritePtr << "\n"
979 << "Store Offs = " << StoreOffset << "\n"
980 << "Load Ptr = " << *LoadPtr << "\n";
985 // If the load and store don't overlap at all, the store doesn't provide
986 // anything to the load. In this case, they really don't alias at all, AA
987 // must have gotten confused.
988 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy);
990 if ((WriteSizeInBits & 7) | (LoadSize & 7))
992 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
996 bool isAAFailure = false;
997 if (StoreOffset < LoadOffset)
998 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
1000 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
1004 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
1005 << "Base = " << *StoreBase << "\n"
1006 << "Store Ptr = " << *WritePtr << "\n"
1007 << "Store Offs = " << StoreOffset << "\n"
1008 << "Load Ptr = " << *LoadPtr << "\n";
1014 // If the Load isn't completely contained within the stored bits, we don't
1015 // have all the bits to feed it. We could do something crazy in the future
1016 // (issue a smaller load then merge the bits in) but this seems unlikely to be
1018 if (StoreOffset > LoadOffset ||
1019 StoreOffset+StoreSize < LoadOffset+LoadSize)
1022 // Okay, we can do this transformation. Return the number of bytes into the
1023 // store that the load is.
1024 return LoadOffset-StoreOffset;
1027 /// This function is called when we have a
1028 /// memdep query of a load that ends up being a clobbering store.
1029 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
1031 // Cannot handle reading from store of first-class aggregate yet.
1032 if (DepSI->getValueOperand()->getType()->isStructTy() ||
1033 DepSI->getValueOperand()->getType()->isArrayTy())
1036 const DataLayout &DL = DepSI->getModule()->getDataLayout();
1037 Value *StorePtr = DepSI->getPointerOperand();
1038 uint64_t StoreSize =DL.getTypeSizeInBits(DepSI->getValueOperand()->getType());
1039 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1040 StorePtr, StoreSize, DL);
1043 /// This function is called when we have a
1044 /// memdep query of a load that ends up being clobbered by another load. See if
1045 /// the other load can feed into the second load.
1046 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
1047 LoadInst *DepLI, const DataLayout &DL){
1048 // Cannot handle reading from store of first-class aggregate yet.
1049 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
1052 Value *DepPtr = DepLI->getPointerOperand();
1053 uint64_t DepSize = DL.getTypeSizeInBits(DepLI->getType());
1054 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL);
1055 if (R != -1) return R;
1057 // If we have a load/load clobber an DepLI can be widened to cover this load,
1058 // then we should widen it!
1059 int64_t LoadOffs = 0;
1060 const Value *LoadBase =
1061 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, DL);
1062 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1064 unsigned Size = MemoryDependenceAnalysis::getLoadLoadClobberFullWidthSize(
1065 LoadBase, LoadOffs, LoadSize, DepLI);
1066 if (Size == 0) return -1;
1068 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, DL);
1073 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
1075 const DataLayout &DL) {
1076 // If the mem operation is a non-constant size, we can't handle it.
1077 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1078 if (!SizeCst) return -1;
1079 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1081 // If this is memset, we just need to see if the offset is valid in the size
1083 if (MI->getIntrinsicID() == Intrinsic::memset)
1084 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
1087 // If we have a memcpy/memmove, the only case we can handle is if this is a
1088 // copy from constant memory. In that case, we can read directly from the
1090 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1092 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1093 if (!Src) return -1;
1095 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, DL));
1096 if (!GV || !GV->isConstant()) return -1;
1098 // See if the access is within the bounds of the transfer.
1099 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1100 MI->getDest(), MemSizeInBits, DL);
1104 unsigned AS = Src->getType()->getPointerAddressSpace();
1105 // Otherwise, see if we can constant fold a load from the constant with the
1106 // offset applied as appropriate.
1107 Src = ConstantExpr::getBitCast(Src,
1108 Type::getInt8PtrTy(Src->getContext(), AS));
1109 Constant *OffsetCst =
1110 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1111 Src = ConstantExpr::getGetElementPtr(Type::getInt8Ty(Src->getContext()), Src,
1113 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1114 if (ConstantFoldLoadFromConstPtr(Src, DL))
1120 /// This function is called when we have a
1121 /// memdep query of a load that ends up being a clobbering store. This means
1122 /// that the store provides bits used by the load but we the pointers don't
1123 /// mustalias. Check this case to see if there is anything more we can do
1124 /// before we give up.
1125 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1127 Instruction *InsertPt, const DataLayout &DL){
1128 LLVMContext &Ctx = SrcVal->getType()->getContext();
1130 uint64_t StoreSize = (DL.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1131 uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy) + 7) / 8;
1133 IRBuilder<> Builder(InsertPt);
1135 // Compute which bits of the stored value are being used by the load. Convert
1136 // to an integer type to start with.
1137 if (SrcVal->getType()->getScalarType()->isPointerTy())
1138 SrcVal = Builder.CreatePtrToInt(SrcVal,
1139 DL.getIntPtrType(SrcVal->getType()));
1140 if (!SrcVal->getType()->isIntegerTy())
1141 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1143 // Shift the bits to the least significant depending on endianness.
1145 if (DL.isLittleEndian())
1146 ShiftAmt = Offset*8;
1148 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1151 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1153 if (LoadSize != StoreSize)
1154 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1156 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, Builder, DL);
1159 /// This function is called when we have a
1160 /// memdep query of a load that ends up being a clobbering load. This means
1161 /// that the load *may* provide bits used by the load but we can't be sure
1162 /// because the pointers don't mustalias. Check this case to see if there is
1163 /// anything more we can do before we give up.
1164 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1165 Type *LoadTy, Instruction *InsertPt,
1167 const DataLayout &DL = SrcVal->getModule()->getDataLayout();
1168 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1169 // widen SrcVal out to a larger load.
1170 unsigned SrcValSize = DL.getTypeStoreSize(SrcVal->getType());
1171 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1172 if (Offset+LoadSize > SrcValSize) {
1173 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1174 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1175 // If we have a load/load clobber an DepLI can be widened to cover this
1176 // load, then we should widen it to the next power of 2 size big enough!
1177 unsigned NewLoadSize = Offset+LoadSize;
1178 if (!isPowerOf2_32(NewLoadSize))
1179 NewLoadSize = NextPowerOf2(NewLoadSize);
1181 Value *PtrVal = SrcVal->getPointerOperand();
1183 // Insert the new load after the old load. This ensures that subsequent
1184 // memdep queries will find the new load. We can't easily remove the old
1185 // load completely because it is already in the value numbering table.
1186 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1188 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1189 DestPTy = PointerType::get(DestPTy,
1190 PtrVal->getType()->getPointerAddressSpace());
1191 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1192 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1193 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1194 NewLoad->takeName(SrcVal);
1195 NewLoad->setAlignment(SrcVal->getAlignment());
1197 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1198 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1200 // Replace uses of the original load with the wider load. On a big endian
1201 // system, we need to shift down to get the relevant bits.
1202 Value *RV = NewLoad;
1203 if (DL.isBigEndian())
1204 RV = Builder.CreateLShr(RV,
1205 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
1206 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1207 SrcVal->replaceAllUsesWith(RV);
1209 // We would like to use gvn.markInstructionForDeletion here, but we can't
1210 // because the load is already memoized into the leader map table that GVN
1211 // tracks. It is potentially possible to remove the load from the table,
1212 // but then there all of the operations based on it would need to be
1213 // rehashed. Just leave the dead load around.
1214 gvn.getMemDep().removeInstruction(SrcVal);
1218 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, DL);
1222 /// This function is called when we have a
1223 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1224 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1225 Type *LoadTy, Instruction *InsertPt,
1226 const DataLayout &DL){
1227 LLVMContext &Ctx = LoadTy->getContext();
1228 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy)/8;
1230 IRBuilder<> Builder(InsertPt);
1232 // We know that this method is only called when the mem transfer fully
1233 // provides the bits for the load.
1234 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1235 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1236 // independently of what the offset is.
1237 Value *Val = MSI->getValue();
1239 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1241 Value *OneElt = Val;
1243 // Splat the value out to the right number of bits.
1244 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1245 // If we can double the number of bytes set, do it.
1246 if (NumBytesSet*2 <= LoadSize) {
1247 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1248 Val = Builder.CreateOr(Val, ShVal);
1253 // Otherwise insert one byte at a time.
1254 Value *ShVal = Builder.CreateShl(Val, 1*8);
1255 Val = Builder.CreateOr(OneElt, ShVal);
1259 return CoerceAvailableValueToLoadType(Val, LoadTy, Builder, DL);
1262 // Otherwise, this is a memcpy/memmove from a constant global.
1263 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1264 Constant *Src = cast<Constant>(MTI->getSource());
1265 unsigned AS = Src->getType()->getPointerAddressSpace();
1267 // Otherwise, see if we can constant fold a load from the constant with the
1268 // offset applied as appropriate.
1269 Src = ConstantExpr::getBitCast(Src,
1270 Type::getInt8PtrTy(Src->getContext(), AS));
1271 Constant *OffsetCst =
1272 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1273 Src = ConstantExpr::getGetElementPtr(Type::getInt8Ty(Src->getContext()), Src,
1275 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1276 return ConstantFoldLoadFromConstPtr(Src, DL);
1280 /// Given a set of loads specified by ValuesPerBlock,
1281 /// construct SSA form, allowing us to eliminate LI. This returns the value
1282 /// that should be used at LI's definition site.
1283 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1284 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1286 // Check for the fully redundant, dominating load case. In this case, we can
1287 // just use the dominating value directly.
1288 if (ValuesPerBlock.size() == 1 &&
1289 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1291 assert(!ValuesPerBlock[0].isUndefValue() && "Dead BB dominate this block");
1292 return ValuesPerBlock[0].MaterializeAdjustedValue(LI, gvn);
1295 // Otherwise, we have to construct SSA form.
1296 SmallVector<PHINode*, 8> NewPHIs;
1297 SSAUpdater SSAUpdate(&NewPHIs);
1298 SSAUpdate.Initialize(LI->getType(), LI->getName());
1300 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1301 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1302 BasicBlock *BB = AV.BB;
1304 if (SSAUpdate.HasValueForBlock(BB))
1307 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LI, gvn));
1310 // Perform PHI construction.
1311 return SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1314 Value *AvailableValueInBlock::MaterializeAdjustedValue(LoadInst *LI,
1317 Type *LoadTy = LI->getType();
1318 const DataLayout &DL = LI->getModule()->getDataLayout();
1319 if (isSimpleValue()) {
1320 Res = getSimpleValue();
1321 if (Res->getType() != LoadTy) {
1322 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(), DL);
1324 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1325 << *getSimpleValue() << '\n'
1326 << *Res << '\n' << "\n\n\n");
1328 } else if (isCoercedLoadValue()) {
1329 LoadInst *Load = getCoercedLoadValue();
1330 if (Load->getType() == LoadTy && Offset == 0) {
1333 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1336 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1337 << *getCoercedLoadValue() << '\n'
1338 << *Res << '\n' << "\n\n\n");
1340 } else if (isMemIntrinValue()) {
1341 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset, LoadTy,
1342 BB->getTerminator(), DL);
1343 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1344 << " " << *getMemIntrinValue() << '\n'
1345 << *Res << '\n' << "\n\n\n");
1347 assert(isUndefValue() && "Should be UndefVal");
1348 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
1349 return UndefValue::get(LoadTy);
1354 static bool isLifetimeStart(const Instruction *Inst) {
1355 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1356 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1360 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1361 AvailValInBlkVect &ValuesPerBlock,
1362 UnavailBlkVect &UnavailableBlocks) {
1364 // Filter out useless results (non-locals, etc). Keep track of the blocks
1365 // where we have a value available in repl, also keep track of whether we see
1366 // dependencies that produce an unknown value for the load (such as a call
1367 // that could potentially clobber the load).
1368 unsigned NumDeps = Deps.size();
1369 const DataLayout &DL = LI->getModule()->getDataLayout();
1370 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1371 BasicBlock *DepBB = Deps[i].getBB();
1372 MemDepResult DepInfo = Deps[i].getResult();
1374 if (DeadBlocks.count(DepBB)) {
1375 // Dead dependent mem-op disguise as a load evaluating the same value
1376 // as the load in question.
1377 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1381 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1382 UnavailableBlocks.push_back(DepBB);
1386 if (DepInfo.isClobber()) {
1387 // The address being loaded in this non-local block may not be the same as
1388 // the pointer operand of the load if PHI translation occurs. Make sure
1389 // to consider the right address.
1390 Value *Address = Deps[i].getAddress();
1392 // If the dependence is to a store that writes to a superset of the bits
1393 // read by the load, we can extract the bits we need for the load from the
1395 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1398 AnalyzeLoadFromClobberingStore(LI->getType(), Address, DepSI);
1400 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1401 DepSI->getValueOperand(),
1408 // Check to see if we have something like this:
1411 // if we have this, replace the later with an extraction from the former.
1412 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1413 // If this is a clobber and L is the first instruction in its block, then
1414 // we have the first instruction in the entry block.
1415 if (DepLI != LI && Address) {
1417 AnalyzeLoadFromClobberingLoad(LI->getType(), Address, DepLI, DL);
1420 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1427 // If the clobbering value is a memset/memcpy/memmove, see if we can
1428 // forward a value on from it.
1429 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1431 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1434 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1441 UnavailableBlocks.push_back(DepBB);
1445 // DepInfo.isDef() here
1447 Instruction *DepInst = DepInfo.getInst();
1449 // Loading the allocation -> undef.
1450 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
1451 // Loading immediately after lifetime begin -> undef.
1452 isLifetimeStart(DepInst)) {
1453 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1454 UndefValue::get(LI->getType())));
1458 // Loading from calloc (which zero initializes memory) -> zero
1459 if (isCallocLikeFn(DepInst, TLI)) {
1460 ValuesPerBlock.push_back(AvailableValueInBlock::get(
1461 DepBB, Constant::getNullValue(LI->getType())));
1465 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1466 // Reject loads and stores that are to the same address but are of
1467 // different types if we have to.
1468 if (S->getValueOperand()->getType() != LI->getType()) {
1469 // If the stored value is larger or equal to the loaded value, we can
1471 if (!CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1472 LI->getType(), DL)) {
1473 UnavailableBlocks.push_back(DepBB);
1478 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1479 S->getValueOperand()));
1483 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1484 // If the types mismatch and we can't handle it, reject reuse of the load.
1485 if (LD->getType() != LI->getType()) {
1486 // If the stored value is larger or equal to the loaded value, we can
1488 if (!CanCoerceMustAliasedValueToLoad(LD, LI->getType(), DL)) {
1489 UnavailableBlocks.push_back(DepBB);
1493 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1497 UnavailableBlocks.push_back(DepBB);
1501 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1502 UnavailBlkVect &UnavailableBlocks) {
1503 // Okay, we have *some* definitions of the value. This means that the value
1504 // is available in some of our (transitive) predecessors. Lets think about
1505 // doing PRE of this load. This will involve inserting a new load into the
1506 // predecessor when it's not available. We could do this in general, but
1507 // prefer to not increase code size. As such, we only do this when we know
1508 // that we only have to insert *one* load (which means we're basically moving
1509 // the load, not inserting a new one).
1511 SmallPtrSet<BasicBlock *, 4> Blockers;
1512 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1513 Blockers.insert(UnavailableBlocks[i]);
1515 // Let's find the first basic block with more than one predecessor. Walk
1516 // backwards through predecessors if needed.
1517 BasicBlock *LoadBB = LI->getParent();
1518 BasicBlock *TmpBB = LoadBB;
1520 while (TmpBB->getSinglePredecessor()) {
1521 TmpBB = TmpBB->getSinglePredecessor();
1522 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1524 if (Blockers.count(TmpBB))
1527 // If any of these blocks has more than one successor (i.e. if the edge we
1528 // just traversed was critical), then there are other paths through this
1529 // block along which the load may not be anticipated. Hoisting the load
1530 // above this block would be adding the load to execution paths along
1531 // which it was not previously executed.
1532 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1539 // Check to see how many predecessors have the loaded value fully
1541 MapVector<BasicBlock *, Value *> PredLoads;
1542 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1543 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1544 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1545 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1546 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1548 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1549 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1551 BasicBlock *Pred = *PI;
1552 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1556 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1557 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1558 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1559 << Pred->getName() << "': " << *LI << '\n');
1563 if (LoadBB->isEHPad()) {
1565 << "COULD NOT PRE LOAD BECAUSE OF AN EH PAD CRITICAL EDGE '"
1566 << Pred->getName() << "': " << *LI << '\n');
1570 CriticalEdgePred.push_back(Pred);
1572 // Only add the predecessors that will not be split for now.
1573 PredLoads[Pred] = nullptr;
1577 // Decide whether PRE is profitable for this load.
1578 unsigned NumUnavailablePreds = PredLoads.size() + CriticalEdgePred.size();
1579 assert(NumUnavailablePreds != 0 &&
1580 "Fully available value should already be eliminated!");
1582 // If this load is unavailable in multiple predecessors, reject it.
1583 // FIXME: If we could restructure the CFG, we could make a common pred with
1584 // all the preds that don't have an available LI and insert a new load into
1586 if (NumUnavailablePreds != 1)
1589 // Split critical edges, and update the unavailable predecessors accordingly.
1590 for (BasicBlock *OrigPred : CriticalEdgePred) {
1591 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1592 assert(!PredLoads.count(OrigPred) && "Split edges shouldn't be in map!");
1593 PredLoads[NewPred] = nullptr;
1594 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1595 << LoadBB->getName() << '\n');
1598 // Check if the load can safely be moved to all the unavailable predecessors.
1599 bool CanDoPRE = true;
1600 const DataLayout &DL = LI->getModule()->getDataLayout();
1601 SmallVector<Instruction*, 8> NewInsts;
1602 for (auto &PredLoad : PredLoads) {
1603 BasicBlock *UnavailablePred = PredLoad.first;
1605 // Do PHI translation to get its value in the predecessor if necessary. The
1606 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1608 // If all preds have a single successor, then we know it is safe to insert
1609 // the load on the pred (?!?), so we can insert code to materialize the
1610 // pointer if it is not available.
1611 PHITransAddr Address(LI->getPointerOperand(), DL, AC);
1612 Value *LoadPtr = nullptr;
1613 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1616 // If we couldn't find or insert a computation of this phi translated value,
1619 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1620 << *LI->getPointerOperand() << "\n");
1625 PredLoad.second = LoadPtr;
1629 while (!NewInsts.empty()) {
1630 Instruction *I = NewInsts.pop_back_val();
1631 if (MD) MD->removeInstruction(I);
1632 I->eraseFromParent();
1634 // HINT: Don't revert the edge-splitting as following transformation may
1635 // also need to split these critical edges.
1636 return !CriticalEdgePred.empty();
1639 // Okay, we can eliminate this load by inserting a reload in the predecessor
1640 // and using PHI construction to get the value in the other predecessors, do
1642 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1643 DEBUG(if (!NewInsts.empty())
1644 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1645 << *NewInsts.back() << '\n');
1647 // Assign value numbers to the new instructions.
1648 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1649 // FIXME: We really _ought_ to insert these value numbers into their
1650 // parent's availability map. However, in doing so, we risk getting into
1651 // ordering issues. If a block hasn't been processed yet, we would be
1652 // marking a value as AVAIL-IN, which isn't what we intend.
1653 VN.lookup_or_add(NewInsts[i]);
1656 for (const auto &PredLoad : PredLoads) {
1657 BasicBlock *UnavailablePred = PredLoad.first;
1658 Value *LoadPtr = PredLoad.second;
1660 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1662 UnavailablePred->getTerminator());
1664 // Transfer the old load's AA tags to the new load.
1666 LI->getAAMetadata(Tags);
1668 NewLoad->setAAMetadata(Tags);
1670 // Transfer DebugLoc.
1671 NewLoad->setDebugLoc(LI->getDebugLoc());
1673 // Add the newly created load.
1674 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1676 MD->invalidateCachedPointerInfo(LoadPtr);
1677 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1680 // Perform PHI construction.
1681 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1682 LI->replaceAllUsesWith(V);
1683 if (isa<PHINode>(V))
1685 if (Instruction *I = dyn_cast<Instruction>(V))
1686 I->setDebugLoc(LI->getDebugLoc());
1687 if (V->getType()->getScalarType()->isPointerTy())
1688 MD->invalidateCachedPointerInfo(V);
1689 markInstructionForDeletion(LI);
1694 /// Attempt to eliminate a load whose dependencies are
1695 /// non-local by performing PHI construction.
1696 bool GVN::processNonLocalLoad(LoadInst *LI) {
1697 // Step 1: Find the non-local dependencies of the load.
1699 MD->getNonLocalPointerDependency(LI, Deps);
1701 // If we had to process more than one hundred blocks to find the
1702 // dependencies, this load isn't worth worrying about. Optimizing
1703 // it will be too expensive.
1704 unsigned NumDeps = Deps.size();
1708 // If we had a phi translation failure, we'll have a single entry which is a
1709 // clobber in the current block. Reject this early.
1711 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1713 dbgs() << "GVN: non-local load ";
1714 LI->printAsOperand(dbgs());
1715 dbgs() << " has unknown dependencies\n";
1720 // If this load follows a GEP, see if we can PRE the indices before analyzing.
1721 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(LI->getOperand(0))) {
1722 for (GetElementPtrInst::op_iterator OI = GEP->idx_begin(),
1723 OE = GEP->idx_end();
1725 if (Instruction *I = dyn_cast<Instruction>(OI->get()))
1726 performScalarPRE(I);
1729 // Step 2: Analyze the availability of the load
1730 AvailValInBlkVect ValuesPerBlock;
1731 UnavailBlkVect UnavailableBlocks;
1732 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1734 // If we have no predecessors that produce a known value for this load, exit
1736 if (ValuesPerBlock.empty())
1739 // Step 3: Eliminate fully redundancy.
1741 // If all of the instructions we depend on produce a known value for this
1742 // load, then it is fully redundant and we can use PHI insertion to compute
1743 // its value. Insert PHIs and remove the fully redundant value now.
1744 if (UnavailableBlocks.empty()) {
1745 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1747 // Perform PHI construction.
1748 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1749 LI->replaceAllUsesWith(V);
1751 if (isa<PHINode>(V))
1753 if (Instruction *I = dyn_cast<Instruction>(V))
1754 if (LI->getDebugLoc())
1755 I->setDebugLoc(LI->getDebugLoc());
1756 if (V->getType()->getScalarType()->isPointerTy())
1757 MD->invalidateCachedPointerInfo(V);
1758 markInstructionForDeletion(LI);
1763 // Step 4: Eliminate partial redundancy.
1764 if (!EnablePRE || !EnableLoadPRE)
1767 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1770 bool GVN::processAssumeIntrinsic(IntrinsicInst *IntrinsicI) {
1771 assert(IntrinsicI->getIntrinsicID() == Intrinsic::assume &&
1772 "This function can only be called with llvm.assume intrinsic");
1773 Value *V = IntrinsicI->getArgOperand(0);
1775 if (ConstantInt *Cond = dyn_cast<ConstantInt>(V)) {
1776 if (Cond->isZero()) {
1777 Type *Int8Ty = Type::getInt8Ty(V->getContext());
1778 // Insert a new store to null instruction before the load to indicate that
1779 // this code is not reachable. FIXME: We could insert unreachable
1780 // instruction directly because we can modify the CFG.
1781 new StoreInst(UndefValue::get(Int8Ty),
1782 Constant::getNullValue(Int8Ty->getPointerTo()),
1785 markInstructionForDeletion(IntrinsicI);
1789 Constant *True = ConstantInt::getTrue(V->getContext());
1790 bool Changed = false;
1792 for (BasicBlock *Successor : successors(IntrinsicI->getParent())) {
1793 BasicBlockEdge Edge(IntrinsicI->getParent(), Successor);
1795 // This property is only true in dominated successors, propagateEquality
1796 // will check dominance for us.
1797 Changed |= propagateEquality(V, True, Edge, false);
1800 // We can replace assume value with true, which covers cases like this:
1801 // call void @llvm.assume(i1 %cmp)
1802 // br i1 %cmp, label %bb1, label %bb2 ; will change %cmp to true
1803 ReplaceWithConstMap[V] = True;
1805 // If one of *cmp *eq operand is const, adding it to map will cover this:
1806 // %cmp = fcmp oeq float 3.000000e+00, %0 ; const on lhs could happen
1807 // call void @llvm.assume(i1 %cmp)
1808 // ret float %0 ; will change it to ret float 3.000000e+00
1809 if (auto *CmpI = dyn_cast<CmpInst>(V)) {
1810 if (CmpI->getPredicate() == CmpInst::Predicate::ICMP_EQ ||
1811 CmpI->getPredicate() == CmpInst::Predicate::FCMP_OEQ ||
1812 (CmpI->getPredicate() == CmpInst::Predicate::FCMP_UEQ &&
1813 CmpI->getFastMathFlags().noNaNs())) {
1814 Value *CmpLHS = CmpI->getOperand(0);
1815 Value *CmpRHS = CmpI->getOperand(1);
1816 if (isa<Constant>(CmpLHS))
1817 std::swap(CmpLHS, CmpRHS);
1818 auto *RHSConst = dyn_cast<Constant>(CmpRHS);
1820 // If only one operand is constant.
1821 if (RHSConst != nullptr && !isa<Constant>(CmpLHS))
1822 ReplaceWithConstMap[CmpLHS] = RHSConst;
1828 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1829 // Patch the replacement so that it is not more restrictive than the value
1831 BinaryOperator *Op = dyn_cast<BinaryOperator>(I);
1832 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl);
1834 ReplOp->andIRFlags(Op);
1836 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) {
1837 // FIXME: If both the original and replacement value are part of the
1838 // same control-flow region (meaning that the execution of one
1839 // guarantees the execution of the other), then we can combine the
1840 // noalias scopes here and do better than the general conservative
1841 // answer used in combineMetadata().
1843 // In general, GVN unifies expressions over different control-flow
1844 // regions, and so we need a conservative combination of the noalias
1846 static const unsigned KnownIDs[] = {
1847 LLVMContext::MD_tbaa,
1848 LLVMContext::MD_alias_scope,
1849 LLVMContext::MD_noalias,
1850 LLVMContext::MD_range,
1851 LLVMContext::MD_fpmath,
1852 LLVMContext::MD_invariant_load,
1854 combineMetadata(ReplInst, I, KnownIDs);
1858 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1859 patchReplacementInstruction(I, Repl);
1860 I->replaceAllUsesWith(Repl);
1863 /// Attempt to eliminate a load, first by eliminating it
1864 /// locally, and then attempting non-local elimination if that fails.
1865 bool GVN::processLoad(LoadInst *L) {
1872 if (L->use_empty()) {
1873 markInstructionForDeletion(L);
1877 // ... to a pointer that has been loaded from before...
1878 MemDepResult Dep = MD->getDependency(L);
1879 const DataLayout &DL = L->getModule()->getDataLayout();
1881 // If we have a clobber and target data is around, see if this is a clobber
1882 // that we can fix up through code synthesis.
1883 if (Dep.isClobber()) {
1884 // Check to see if we have something like this:
1885 // store i32 123, i32* %P
1886 // %A = bitcast i32* %P to i8*
1887 // %B = gep i8* %A, i32 1
1890 // We could do that by recognizing if the clobber instructions are obviously
1891 // a common base + constant offset, and if the previous store (or memset)
1892 // completely covers this load. This sort of thing can happen in bitfield
1894 Value *AvailVal = nullptr;
1895 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1896 int Offset = AnalyzeLoadFromClobberingStore(
1897 L->getType(), L->getPointerOperand(), DepSI);
1899 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1900 L->getType(), L, DL);
1903 // Check to see if we have something like this:
1906 // if we have this, replace the later with an extraction from the former.
1907 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1908 // If this is a clobber and L is the first instruction in its block, then
1909 // we have the first instruction in the entry block.
1913 int Offset = AnalyzeLoadFromClobberingLoad(
1914 L->getType(), L->getPointerOperand(), DepLI, DL);
1916 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1919 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1920 // a value on from it.
1921 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1922 int Offset = AnalyzeLoadFromClobberingMemInst(
1923 L->getType(), L->getPointerOperand(), DepMI, DL);
1925 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, DL);
1929 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1930 << *AvailVal << '\n' << *L << "\n\n\n");
1932 // Replace the load!
1933 L->replaceAllUsesWith(AvailVal);
1934 if (AvailVal->getType()->getScalarType()->isPointerTy())
1935 MD->invalidateCachedPointerInfo(AvailVal);
1936 markInstructionForDeletion(L);
1942 // If the value isn't available, don't do anything!
1943 if (Dep.isClobber()) {
1945 // fast print dep, using operator<< on instruction is too slow.
1946 dbgs() << "GVN: load ";
1947 L->printAsOperand(dbgs());
1948 Instruction *I = Dep.getInst();
1949 dbgs() << " is clobbered by " << *I << '\n';
1954 // If it is defined in another block, try harder.
1955 if (Dep.isNonLocal())
1956 return processNonLocalLoad(L);
1960 // fast print dep, using operator<< on instruction is too slow.
1961 dbgs() << "GVN: load ";
1962 L->printAsOperand(dbgs());
1963 dbgs() << " has unknown dependence\n";
1968 Instruction *DepInst = Dep.getInst();
1969 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1970 Value *StoredVal = DepSI->getValueOperand();
1972 // The store and load are to a must-aliased pointer, but they may not
1973 // actually have the same type. See if we know how to reuse the stored
1974 // value (depending on its type).
1975 if (StoredVal->getType() != L->getType()) {
1976 IRBuilder<> Builder(L);
1978 CoerceAvailableValueToLoadType(StoredVal, L->getType(), Builder, DL);
1982 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1983 << '\n' << *L << "\n\n\n");
1987 L->replaceAllUsesWith(StoredVal);
1988 if (StoredVal->getType()->getScalarType()->isPointerTy())
1989 MD->invalidateCachedPointerInfo(StoredVal);
1990 markInstructionForDeletion(L);
1995 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1996 Value *AvailableVal = DepLI;
1998 // The loads are of a must-aliased pointer, but they may not actually have
1999 // the same type. See if we know how to reuse the previously loaded value
2000 // (depending on its type).
2001 if (DepLI->getType() != L->getType()) {
2002 IRBuilder<> Builder(L);
2004 CoerceAvailableValueToLoadType(DepLI, L->getType(), Builder, DL);
2008 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
2009 << "\n" << *L << "\n\n\n");
2013 patchAndReplaceAllUsesWith(L, AvailableVal);
2014 if (DepLI->getType()->getScalarType()->isPointerTy())
2015 MD->invalidateCachedPointerInfo(DepLI);
2016 markInstructionForDeletion(L);
2021 // If this load really doesn't depend on anything, then we must be loading an
2022 // undef value. This can happen when loading for a fresh allocation with no
2023 // intervening stores, for example.
2024 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
2025 L->replaceAllUsesWith(UndefValue::get(L->getType()));
2026 markInstructionForDeletion(L);
2031 // If this load occurs either right after a lifetime begin,
2032 // then the loaded value is undefined.
2033 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
2034 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
2035 L->replaceAllUsesWith(UndefValue::get(L->getType()));
2036 markInstructionForDeletion(L);
2042 // If this load follows a calloc (which zero initializes memory),
2043 // then the loaded value is zero
2044 if (isCallocLikeFn(DepInst, TLI)) {
2045 L->replaceAllUsesWith(Constant::getNullValue(L->getType()));
2046 markInstructionForDeletion(L);
2054 // In order to find a leader for a given value number at a
2055 // specific basic block, we first obtain the list of all Values for that number,
2056 // and then scan the list to find one whose block dominates the block in
2057 // question. This is fast because dominator tree queries consist of only
2058 // a few comparisons of DFS numbers.
2059 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
2060 LeaderTableEntry Vals = LeaderTable[num];
2061 if (!Vals.Val) return nullptr;
2063 Value *Val = nullptr;
2064 if (DT->dominates(Vals.BB, BB)) {
2066 if (isa<Constant>(Val)) return Val;
2069 LeaderTableEntry* Next = Vals.Next;
2071 if (DT->dominates(Next->BB, BB)) {
2072 if (isa<Constant>(Next->Val)) return Next->Val;
2073 if (!Val) Val = Next->Val;
2082 /// There is an edge from 'Src' to 'Dst'. Return
2083 /// true if every path from the entry block to 'Dst' passes via this edge. In
2084 /// particular 'Dst' must not be reachable via another edge from 'Src'.
2085 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
2086 DominatorTree *DT) {
2087 // While in theory it is interesting to consider the case in which Dst has
2088 // more than one predecessor, because Dst might be part of a loop which is
2089 // only reachable from Src, in practice it is pointless since at the time
2090 // GVN runs all such loops have preheaders, which means that Dst will have
2091 // been changed to have only one predecessor, namely Src.
2092 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
2093 const BasicBlock *Src = E.getStart();
2094 assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
2096 return Pred != nullptr;
2099 // Tries to replace instruction with const, using information from
2100 // ReplaceWithConstMap.
2101 bool GVN::replaceOperandsWithConsts(Instruction *Instr) const {
2102 bool Changed = false;
2103 for (unsigned OpNum = 0; OpNum < Instr->getNumOperands(); ++OpNum) {
2104 Value *Operand = Instr->getOperand(OpNum);
2105 auto it = ReplaceWithConstMap.find(Operand);
2106 if (it != ReplaceWithConstMap.end()) {
2107 assert(!isa<Constant>(Operand) &&
2108 "Replacing constants with constants is invalid");
2109 Instr->setOperand(OpNum, it->second);
2116 /// The given values are known to be equal in every block
2117 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
2118 /// 'RHS' everywhere in the scope. Returns whether a change was made.
2119 /// If DominatesByEdge is false, then it means that it is dominated by Root.End.
2120 bool GVN::propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root,
2121 bool DominatesByEdge) {
2122 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
2123 Worklist.push_back(std::make_pair(LHS, RHS));
2124 bool Changed = false;
2125 // For speed, compute a conservative fast approximation to
2126 // DT->dominates(Root, Root.getEnd());
2127 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
2129 while (!Worklist.empty()) {
2130 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
2131 LHS = Item.first; RHS = Item.second;
2135 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
2137 // Don't try to propagate equalities between constants.
2138 if (isa<Constant>(LHS) && isa<Constant>(RHS))
2141 // Prefer a constant on the right-hand side, or an Argument if no constants.
2142 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
2143 std::swap(LHS, RHS);
2144 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
2146 // If there is no obvious reason to prefer the left-hand side over the
2147 // right-hand side, ensure the longest lived term is on the right-hand side,
2148 // so the shortest lived term will be replaced by the longest lived.
2149 // This tends to expose more simplifications.
2150 uint32_t LVN = VN.lookup_or_add(LHS);
2151 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
2152 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
2153 // Move the 'oldest' value to the right-hand side, using the value number
2154 // as a proxy for age.
2155 uint32_t RVN = VN.lookup_or_add(RHS);
2157 std::swap(LHS, RHS);
2162 // If value numbering later sees that an instruction in the scope is equal
2163 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
2164 // the invariant that instructions only occur in the leader table for their
2165 // own value number (this is used by removeFromLeaderTable), do not do this
2166 // if RHS is an instruction (if an instruction in the scope is morphed into
2167 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
2168 // using the leader table is about compiling faster, not optimizing better).
2169 // The leader table only tracks basic blocks, not edges. Only add to if we
2170 // have the simple case where the edge dominates the end.
2171 if (RootDominatesEnd && !isa<Instruction>(RHS))
2172 addToLeaderTable(LVN, RHS, Root.getEnd());
2174 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
2175 // LHS always has at least one use that is not dominated by Root, this will
2176 // never do anything if LHS has only one use.
2177 if (!LHS->hasOneUse()) {
2178 unsigned NumReplacements =
2180 ? replaceDominatedUsesWith(LHS, RHS, *DT, Root)
2181 : replaceDominatedUsesWith(LHS, RHS, *DT, Root.getEnd());
2183 Changed |= NumReplacements > 0;
2184 NumGVNEqProp += NumReplacements;
2187 // Now try to deduce additional equalities from this one. For example, if
2188 // the known equality was "(A != B)" == "false" then it follows that A and B
2189 // are equal in the scope. Only boolean equalities with an explicit true or
2190 // false RHS are currently supported.
2191 if (!RHS->getType()->isIntegerTy(1))
2192 // Not a boolean equality - bail out.
2194 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
2196 // RHS neither 'true' nor 'false' - bail out.
2198 // Whether RHS equals 'true'. Otherwise it equals 'false'.
2199 bool isKnownTrue = CI->isAllOnesValue();
2200 bool isKnownFalse = !isKnownTrue;
2202 // If "A && B" is known true then both A and B are known true. If "A || B"
2203 // is known false then both A and B are known false.
2205 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
2206 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
2207 Worklist.push_back(std::make_pair(A, RHS));
2208 Worklist.push_back(std::make_pair(B, RHS));
2212 // If we are propagating an equality like "(A == B)" == "true" then also
2213 // propagate the equality A == B. When propagating a comparison such as
2214 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
2215 if (CmpInst *Cmp = dyn_cast<CmpInst>(LHS)) {
2216 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
2218 // If "A == B" is known true, or "A != B" is known false, then replace
2219 // A with B everywhere in the scope.
2220 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2221 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2222 Worklist.push_back(std::make_pair(Op0, Op1));
2224 // Handle the floating point versions of equality comparisons too.
2225 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::FCMP_OEQ) ||
2226 (isKnownFalse && Cmp->getPredicate() == CmpInst::FCMP_UNE)) {
2228 // Floating point -0.0 and 0.0 compare equal, so we can only
2229 // propagate values if we know that we have a constant and that
2230 // its value is non-zero.
2232 // FIXME: We should do this optimization if 'no signed zeros' is
2233 // applicable via an instruction-level fast-math-flag or some other
2234 // indicator that relaxed FP semantics are being used.
2236 if (isa<ConstantFP>(Op1) && !cast<ConstantFP>(Op1)->isZero())
2237 Worklist.push_back(std::make_pair(Op0, Op1));
2240 // If "A >= B" is known true, replace "A < B" with false everywhere.
2241 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2242 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2243 // Since we don't have the instruction "A < B" immediately to hand, work
2244 // out the value number that it would have and use that to find an
2245 // appropriate instruction (if any).
2246 uint32_t NextNum = VN.getNextUnusedValueNumber();
2247 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2248 // If the number we were assigned was brand new then there is no point in
2249 // looking for an instruction realizing it: there cannot be one!
2250 if (Num < NextNum) {
2251 Value *NotCmp = findLeader(Root.getEnd(), Num);
2252 if (NotCmp && isa<Instruction>(NotCmp)) {
2253 unsigned NumReplacements =
2255 ? replaceDominatedUsesWith(NotCmp, NotVal, *DT, Root)
2256 : replaceDominatedUsesWith(NotCmp, NotVal, *DT,
2258 Changed |= NumReplacements > 0;
2259 NumGVNEqProp += NumReplacements;
2262 // Ensure that any instruction in scope that gets the "A < B" value number
2263 // is replaced with false.
2264 // The leader table only tracks basic blocks, not edges. Only add to if we
2265 // have the simple case where the edge dominates the end.
2266 if (RootDominatesEnd)
2267 addToLeaderTable(Num, NotVal, Root.getEnd());
2276 /// When calculating availability, handle an instruction
2277 /// by inserting it into the appropriate sets
2278 bool GVN::processInstruction(Instruction *I) {
2279 // Ignore dbg info intrinsics.
2280 if (isa<DbgInfoIntrinsic>(I))
2283 // If the instruction can be easily simplified then do so now in preference
2284 // to value numbering it. Value numbering often exposes redundancies, for
2285 // example if it determines that %y is equal to %x then the instruction
2286 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2287 const DataLayout &DL = I->getModule()->getDataLayout();
2288 if (Value *V = SimplifyInstruction(I, DL, TLI, DT, AC)) {
2289 I->replaceAllUsesWith(V);
2290 if (MD && V->getType()->getScalarType()->isPointerTy())
2291 MD->invalidateCachedPointerInfo(V);
2292 markInstructionForDeletion(I);
2297 if (IntrinsicInst *IntrinsicI = dyn_cast<IntrinsicInst>(I))
2298 if (IntrinsicI->getIntrinsicID() == Intrinsic::assume)
2299 return processAssumeIntrinsic(IntrinsicI);
2301 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2302 if (processLoad(LI))
2305 unsigned Num = VN.lookup_or_add(LI);
2306 addToLeaderTable(Num, LI, LI->getParent());
2310 // For conditional branches, we can perform simple conditional propagation on
2311 // the condition value itself.
2312 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2313 if (!BI->isConditional())
2316 if (isa<Constant>(BI->getCondition()))
2317 return processFoldableCondBr(BI);
2319 Value *BranchCond = BI->getCondition();
2320 BasicBlock *TrueSucc = BI->getSuccessor(0);
2321 BasicBlock *FalseSucc = BI->getSuccessor(1);
2322 // Avoid multiple edges early.
2323 if (TrueSucc == FalseSucc)
2326 BasicBlock *Parent = BI->getParent();
2327 bool Changed = false;
2329 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
2330 BasicBlockEdge TrueE(Parent, TrueSucc);
2331 Changed |= propagateEquality(BranchCond, TrueVal, TrueE, true);
2333 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
2334 BasicBlockEdge FalseE(Parent, FalseSucc);
2335 Changed |= propagateEquality(BranchCond, FalseVal, FalseE, true);
2340 // For switches, propagate the case values into the case destinations.
2341 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2342 Value *SwitchCond = SI->getCondition();
2343 BasicBlock *Parent = SI->getParent();
2344 bool Changed = false;
2346 // Remember how many outgoing edges there are to every successor.
2347 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2348 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
2349 ++SwitchEdges[SI->getSuccessor(i)];
2351 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2353 BasicBlock *Dst = i.getCaseSuccessor();
2354 // If there is only a single edge, propagate the case value into it.
2355 if (SwitchEdges.lookup(Dst) == 1) {
2356 BasicBlockEdge E(Parent, Dst);
2357 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E, true);
2363 // Instructions with void type don't return a value, so there's
2364 // no point in trying to find redundancies in them.
2365 if (I->getType()->isVoidTy())
2368 uint32_t NextNum = VN.getNextUnusedValueNumber();
2369 unsigned Num = VN.lookup_or_add(I);
2371 // Allocations are always uniquely numbered, so we can save time and memory
2372 // by fast failing them.
2373 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2374 addToLeaderTable(Num, I, I->getParent());
2378 // If the number we were assigned was a brand new VN, then we don't
2379 // need to do a lookup to see if the number already exists
2380 // somewhere in the domtree: it can't!
2381 if (Num >= NextNum) {
2382 addToLeaderTable(Num, I, I->getParent());
2386 // Perform fast-path value-number based elimination of values inherited from
2388 Value *repl = findLeader(I->getParent(), Num);
2390 // Failure, just remember this instance for future use.
2391 addToLeaderTable(Num, I, I->getParent());
2396 patchAndReplaceAllUsesWith(I, repl);
2397 if (MD && repl->getType()->getScalarType()->isPointerTy())
2398 MD->invalidateCachedPointerInfo(repl);
2399 markInstructionForDeletion(I);
2403 /// runOnFunction - This is the main transformation entry point for a function.
2404 bool GVN::runOnFunction(Function& F) {
2405 if (skipOptnoneFunction(F))
2409 MD = &getAnalysis<MemoryDependenceAnalysis>();
2410 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2411 AC = &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(F);
2412 TLI = &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
2413 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2417 bool Changed = false;
2418 bool ShouldContinue = true;
2420 // Merge unconditional branches, allowing PRE to catch more
2421 // optimization opportunities.
2422 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2423 BasicBlock *BB = FI++;
2426 MergeBlockIntoPredecessor(BB, DT, /* LoopInfo */ nullptr, MD);
2427 if (removedBlock) ++NumGVNBlocks;
2429 Changed |= removedBlock;
2432 unsigned Iteration = 0;
2433 while (ShouldContinue) {
2434 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2435 ShouldContinue = iterateOnFunction(F);
2436 Changed |= ShouldContinue;
2441 // Fabricate val-num for dead-code in order to suppress assertion in
2443 assignValNumForDeadCode();
2444 bool PREChanged = true;
2445 while (PREChanged) {
2446 PREChanged = performPRE(F);
2447 Changed |= PREChanged;
2451 // FIXME: Should perform GVN again after PRE does something. PRE can move
2452 // computations into blocks where they become fully redundant. Note that
2453 // we can't do this until PRE's critical edge splitting updates memdep.
2454 // Actually, when this happens, we should just fully integrate PRE into GVN.
2456 cleanupGlobalSets();
2457 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2465 bool GVN::processBlock(BasicBlock *BB) {
2466 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2467 // (and incrementing BI before processing an instruction).
2468 assert(InstrsToErase.empty() &&
2469 "We expect InstrsToErase to be empty across iterations");
2470 if (DeadBlocks.count(BB))
2473 // Clearing map before every BB because it can be used only for single BB.
2474 ReplaceWithConstMap.clear();
2475 bool ChangedFunction = false;
2477 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2479 if (!ReplaceWithConstMap.empty())
2480 ChangedFunction |= replaceOperandsWithConsts(BI);
2481 ChangedFunction |= processInstruction(BI);
2483 if (InstrsToErase.empty()) {
2488 // If we need some instructions deleted, do it now.
2489 NumGVNInstr += InstrsToErase.size();
2491 // Avoid iterator invalidation.
2492 bool AtStart = BI == BB->begin();
2496 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2497 E = InstrsToErase.end(); I != E; ++I) {
2498 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2499 if (MD) MD->removeInstruction(*I);
2500 DEBUG(verifyRemoved(*I));
2501 (*I)->eraseFromParent();
2503 InstrsToErase.clear();
2511 return ChangedFunction;
2514 // Instantiate an expression in a predecessor that lacked it.
2515 bool GVN::performScalarPREInsertion(Instruction *Instr, BasicBlock *Pred,
2516 unsigned int ValNo) {
2517 // Because we are going top-down through the block, all value numbers
2518 // will be available in the predecessor by the time we need them. Any
2519 // that weren't originally present will have been instantiated earlier
2521 bool success = true;
2522 for (unsigned i = 0, e = Instr->getNumOperands(); i != e; ++i) {
2523 Value *Op = Instr->getOperand(i);
2524 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2527 if (Value *V = findLeader(Pred, VN.lookup(Op))) {
2528 Instr->setOperand(i, V);
2535 // Fail out if we encounter an operand that is not available in
2536 // the PRE predecessor. This is typically because of loads which
2537 // are not value numbered precisely.
2541 Instr->insertBefore(Pred->getTerminator());
2542 Instr->setName(Instr->getName() + ".pre");
2543 Instr->setDebugLoc(Instr->getDebugLoc());
2544 VN.add(Instr, ValNo);
2546 // Update the availability map to include the new instruction.
2547 addToLeaderTable(ValNo, Instr, Pred);
2551 bool GVN::performScalarPRE(Instruction *CurInst) {
2552 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap;
2554 if (isa<AllocaInst>(CurInst) || isa<TerminatorInst>(CurInst) ||
2555 isa<PHINode>(CurInst) || CurInst->getType()->isVoidTy() ||
2556 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2557 isa<DbgInfoIntrinsic>(CurInst))
2560 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2561 // sinking the compare again, and it would force the code generator to
2562 // move the i1 from processor flags or predicate registers into a general
2563 // purpose register.
2564 if (isa<CmpInst>(CurInst))
2567 // We don't currently value number ANY inline asm calls.
2568 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2569 if (CallI->isInlineAsm())
2572 uint32_t ValNo = VN.lookup(CurInst);
2574 // Look for the predecessors for PRE opportunities. We're
2575 // only trying to solve the basic diamond case, where
2576 // a value is computed in the successor and one predecessor,
2577 // but not the other. We also explicitly disallow cases
2578 // where the successor is its own predecessor, because they're
2579 // more complicated to get right.
2580 unsigned NumWith = 0;
2581 unsigned NumWithout = 0;
2582 BasicBlock *PREPred = nullptr;
2583 BasicBlock *CurrentBlock = CurInst->getParent();
2586 for (pred_iterator PI = pred_begin(CurrentBlock), PE = pred_end(CurrentBlock);
2588 BasicBlock *P = *PI;
2589 // We're not interested in PRE where the block is its
2590 // own predecessor, or in blocks with predecessors
2591 // that are not reachable.
2592 if (P == CurrentBlock) {
2595 } else if (!DT->isReachableFromEntry(P)) {
2600 Value *predV = findLeader(P, ValNo);
2602 predMap.push_back(std::make_pair(static_cast<Value *>(nullptr), P));
2605 } else if (predV == CurInst) {
2606 /* CurInst dominates this predecessor. */
2610 predMap.push_back(std::make_pair(predV, P));
2615 // Don't do PRE when it might increase code size, i.e. when
2616 // we would need to insert instructions in more than one pred.
2617 if (NumWithout > 1 || NumWith == 0)
2620 // We may have a case where all predecessors have the instruction,
2621 // and we just need to insert a phi node. Otherwise, perform
2623 Instruction *PREInstr = nullptr;
2625 if (NumWithout != 0) {
2626 // Don't do PRE across indirect branch.
2627 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2630 // We can't do PRE safely on a critical edge, so instead we schedule
2631 // the edge to be split and perform the PRE the next time we iterate
2633 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2634 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2635 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2638 // We need to insert somewhere, so let's give it a shot
2639 PREInstr = CurInst->clone();
2640 if (!performScalarPREInsertion(PREInstr, PREPred, ValNo)) {
2641 // If we failed insertion, make sure we remove the instruction.
2642 DEBUG(verifyRemoved(PREInstr));
2648 // Either we should have filled in the PRE instruction, or we should
2649 // not have needed insertions.
2650 assert (PREInstr != nullptr || NumWithout == 0);
2654 // Create a PHI to make the value available in this block.
2656 PHINode::Create(CurInst->getType(), predMap.size(),
2657 CurInst->getName() + ".pre-phi", CurrentBlock->begin());
2658 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2659 if (Value *V = predMap[i].first)
2660 Phi->addIncoming(V, predMap[i].second);
2662 Phi->addIncoming(PREInstr, PREPred);
2666 addToLeaderTable(ValNo, Phi, CurrentBlock);
2667 Phi->setDebugLoc(CurInst->getDebugLoc());
2668 CurInst->replaceAllUsesWith(Phi);
2669 if (MD && Phi->getType()->getScalarType()->isPointerTy())
2670 MD->invalidateCachedPointerInfo(Phi);
2672 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2674 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2676 MD->removeInstruction(CurInst);
2677 DEBUG(verifyRemoved(CurInst));
2678 CurInst->eraseFromParent();
2684 /// Perform a purely local form of PRE that looks for diamond
2685 /// control flow patterns and attempts to perform simple PRE at the join point.
2686 bool GVN::performPRE(Function &F) {
2687 bool Changed = false;
2688 for (BasicBlock *CurrentBlock : depth_first(&F.getEntryBlock())) {
2689 // Nothing to PRE in the entry block.
2690 if (CurrentBlock == &F.getEntryBlock())
2693 // Don't perform PRE on an EH pad.
2694 if (CurrentBlock->isEHPad())
2697 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2698 BE = CurrentBlock->end();
2700 Instruction *CurInst = BI++;
2701 Changed = performScalarPRE(CurInst);
2705 if (splitCriticalEdges())
2711 /// Split the critical edge connecting the given two blocks, and return
2712 /// the block inserted to the critical edge.
2713 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2715 SplitCriticalEdge(Pred, Succ, CriticalEdgeSplittingOptions(DT));
2717 MD->invalidateCachedPredecessors();
2721 /// Split critical edges found during the previous
2722 /// iteration that may enable further optimization.
2723 bool GVN::splitCriticalEdges() {
2724 if (toSplit.empty())
2727 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2728 SplitCriticalEdge(Edge.first, Edge.second,
2729 CriticalEdgeSplittingOptions(DT));
2730 } while (!toSplit.empty());
2731 if (MD) MD->invalidateCachedPredecessors();
2735 /// Executes one iteration of GVN
2736 bool GVN::iterateOnFunction(Function &F) {
2737 cleanupGlobalSets();
2739 // Top-down walk of the dominator tree
2740 bool Changed = false;
2741 // Save the blocks this function have before transformation begins. GVN may
2742 // split critical edge, and hence may invalidate the RPO/DT iterator.
2744 std::vector<BasicBlock *> BBVect;
2745 BBVect.reserve(256);
2746 // Needed for value numbering with phi construction to work.
2747 ReversePostOrderTraversal<Function *> RPOT(&F);
2748 for (ReversePostOrderTraversal<Function *>::rpo_iterator RI = RPOT.begin(),
2751 BBVect.push_back(*RI);
2753 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
2755 Changed |= processBlock(*I);
2760 void GVN::cleanupGlobalSets() {
2762 LeaderTable.clear();
2763 TableAllocator.Reset();
2766 /// Verify that the specified instruction does not occur in our
2767 /// internal data structures.
2768 void GVN::verifyRemoved(const Instruction *Inst) const {
2769 VN.verifyRemoved(Inst);
2771 // Walk through the value number scope to make sure the instruction isn't
2772 // ferreted away in it.
2773 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2774 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2775 const LeaderTableEntry *Node = &I->second;
2776 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2778 while (Node->Next) {
2780 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2785 /// BB is declared dead, which implied other blocks become dead as well. This
2786 /// function is to add all these blocks to "DeadBlocks". For the dead blocks'
2787 /// live successors, update their phi nodes by replacing the operands
2788 /// corresponding to dead blocks with UndefVal.
2789 void GVN::addDeadBlock(BasicBlock *BB) {
2790 SmallVector<BasicBlock *, 4> NewDead;
2791 SmallSetVector<BasicBlock *, 4> DF;
2793 NewDead.push_back(BB);
2794 while (!NewDead.empty()) {
2795 BasicBlock *D = NewDead.pop_back_val();
2796 if (DeadBlocks.count(D))
2799 // All blocks dominated by D are dead.
2800 SmallVector<BasicBlock *, 8> Dom;
2801 DT->getDescendants(D, Dom);
2802 DeadBlocks.insert(Dom.begin(), Dom.end());
2804 // Figure out the dominance-frontier(D).
2805 for (SmallVectorImpl<BasicBlock *>::iterator I = Dom.begin(),
2806 E = Dom.end(); I != E; I++) {
2808 for (succ_iterator SI = succ_begin(B), SE = succ_end(B); SI != SE; SI++) {
2809 BasicBlock *S = *SI;
2810 if (DeadBlocks.count(S))
2813 bool AllPredDead = true;
2814 for (pred_iterator PI = pred_begin(S), PE = pred_end(S); PI != PE; PI++)
2815 if (!DeadBlocks.count(*PI)) {
2816 AllPredDead = false;
2821 // S could be proved dead later on. That is why we don't update phi
2822 // operands at this moment.
2825 // While S is not dominated by D, it is dead by now. This could take
2826 // place if S already have a dead predecessor before D is declared
2828 NewDead.push_back(S);
2834 // For the dead blocks' live successors, update their phi nodes by replacing
2835 // the operands corresponding to dead blocks with UndefVal.
2836 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2839 if (DeadBlocks.count(B))
2842 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2843 for (SmallVectorImpl<BasicBlock *>::iterator PI = Preds.begin(),
2844 PE = Preds.end(); PI != PE; PI++) {
2845 BasicBlock *P = *PI;
2847 if (!DeadBlocks.count(P))
2850 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2851 if (BasicBlock *S = splitCriticalEdges(P, B))
2852 DeadBlocks.insert(P = S);
2855 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2856 PHINode &Phi = cast<PHINode>(*II);
2857 Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2858 UndefValue::get(Phi.getType()));
2864 // If the given branch is recognized as a foldable branch (i.e. conditional
2865 // branch with constant condition), it will perform following analyses and
2867 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2868 // R be the target of the dead out-coming edge.
2869 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2870 // edge. The result of this step will be {X| X is dominated by R}
2871 // 2) Identify those blocks which haves at least one dead predecessor. The
2872 // result of this step will be dominance-frontier(R).
2873 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2874 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2876 // Return true iff *NEW* dead code are found.
2877 bool GVN::processFoldableCondBr(BranchInst *BI) {
2878 if (!BI || BI->isUnconditional())
2881 // If a branch has two identical successors, we cannot declare either dead.
2882 if (BI->getSuccessor(0) == BI->getSuccessor(1))
2885 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2889 BasicBlock *DeadRoot = Cond->getZExtValue() ?
2890 BI->getSuccessor(1) : BI->getSuccessor(0);
2891 if (DeadBlocks.count(DeadRoot))
2894 if (!DeadRoot->getSinglePredecessor())
2895 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2897 addDeadBlock(DeadRoot);
2901 // performPRE() will trigger assert if it comes across an instruction without
2902 // associated val-num. As it normally has far more live instructions than dead
2903 // instructions, it makes more sense just to "fabricate" a val-number for the
2904 // dead code than checking if instruction involved is dead or not.
2905 void GVN::assignValNumForDeadCode() {
2906 for (SetVector<BasicBlock *>::iterator I = DeadBlocks.begin(),
2907 E = DeadBlocks.end(); I != E; I++) {
2908 BasicBlock *BB = *I;
2909 for (BasicBlock::iterator II = BB->begin(), EE = BB->end();
2911 Instruction *Inst = &*II;
2912 unsigned ValNum = VN.lookup_or_add(Inst);
2913 addToLeaderTable(ValNum, Inst, BB);