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 #define DEBUG_TYPE "gvn"
19 #include "llvm/Transforms/Scalar.h"
20 #include "llvm/ADT/DenseMap.h"
21 #include "llvm/ADT/DepthFirstIterator.h"
22 #include "llvm/ADT/Hashing.h"
23 #include "llvm/ADT/SmallPtrSet.h"
24 #include "llvm/ADT/Statistic.h"
25 #include "llvm/Analysis/AliasAnalysis.h"
26 #include "llvm/Analysis/CFG.h"
27 #include "llvm/Analysis/ConstantFolding.h"
28 #include "llvm/Analysis/Dominators.h"
29 #include "llvm/Analysis/InstructionSimplify.h"
30 #include "llvm/Analysis/Loads.h"
31 #include "llvm/Analysis/MemoryBuiltins.h"
32 #include "llvm/Analysis/MemoryDependenceAnalysis.h"
33 #include "llvm/Analysis/PHITransAddr.h"
34 #include "llvm/Analysis/ValueTracking.h"
35 #include "llvm/Assembly/Writer.h"
36 #include "llvm/IR/DataLayout.h"
37 #include "llvm/IR/GlobalVariable.h"
38 #include "llvm/IR/IRBuilder.h"
39 #include "llvm/IR/IntrinsicInst.h"
40 #include "llvm/IR/LLVMContext.h"
41 #include "llvm/IR/Metadata.h"
42 #include "llvm/Support/Allocator.h"
43 #include "llvm/Support/CommandLine.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/PatternMatch.h"
46 #include "llvm/Target/TargetLibraryInfo.h"
47 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
48 #include "llvm/Transforms/Utils/SSAUpdater.h"
51 using namespace PatternMatch;
53 STATISTIC(NumGVNInstr, "Number of instructions deleted");
54 STATISTIC(NumGVNLoad, "Number of loads deleted");
55 STATISTIC(NumGVNPRE, "Number of instructions PRE'd");
56 STATISTIC(NumGVNBlocks, "Number of blocks merged");
57 STATISTIC(NumGVNSimpl, "Number of instructions simplified");
58 STATISTIC(NumGVNEqProp, "Number of equalities propagated");
59 STATISTIC(NumPRELoad, "Number of loads PRE'd");
61 static cl::opt<bool> EnablePRE("enable-pre",
62 cl::init(true), cl::Hidden);
63 static cl::opt<bool> EnableLoadPRE("enable-load-pre", cl::init(true));
65 // Maximum allowed recursion depth.
66 static cl::opt<uint32_t>
67 MaxRecurseDepth("max-recurse-depth", cl::Hidden, cl::init(1000), cl::ZeroOrMore,
68 cl::desc("Max recurse depth (default = 1000)"));
70 //===----------------------------------------------------------------------===//
72 //===----------------------------------------------------------------------===//
74 /// This class holds the mapping between values and value numbers. It is used
75 /// as an efficient mechanism to determine the expression-wise equivalence of
81 SmallVector<uint32_t, 4> varargs;
83 Expression(uint32_t o = ~2U) : opcode(o) { }
85 bool operator==(const Expression &other) const {
86 if (opcode != other.opcode)
88 if (opcode == ~0U || opcode == ~1U)
90 if (type != other.type)
92 if (varargs != other.varargs)
97 friend hash_code hash_value(const Expression &Value) {
98 return hash_combine(Value.opcode, Value.type,
99 hash_combine_range(Value.varargs.begin(),
100 Value.varargs.end()));
105 DenseMap<Value*, uint32_t> valueNumbering;
106 DenseMap<Expression, uint32_t> expressionNumbering;
108 MemoryDependenceAnalysis *MD;
111 uint32_t nextValueNumber;
113 Expression create_expression(Instruction* I);
114 Expression create_cmp_expression(unsigned Opcode,
115 CmpInst::Predicate Predicate,
116 Value *LHS, Value *RHS);
117 Expression create_extractvalue_expression(ExtractValueInst* EI);
118 uint32_t lookup_or_add_call(CallInst* C);
120 ValueTable() : nextValueNumber(1) { }
121 uint32_t lookup_or_add(Value *V);
122 uint32_t lookup(Value *V) const;
123 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred,
124 Value *LHS, Value *RHS);
125 void add(Value *V, uint32_t num);
127 void erase(Value *v);
128 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
129 AliasAnalysis *getAliasAnalysis() const { return AA; }
130 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
131 void setDomTree(DominatorTree* D) { DT = D; }
132 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
133 void verifyRemoved(const Value *) const;
138 template <> struct DenseMapInfo<Expression> {
139 static inline Expression getEmptyKey() {
143 static inline Expression getTombstoneKey() {
147 static unsigned getHashValue(const Expression e) {
148 using llvm::hash_value;
149 return static_cast<unsigned>(hash_value(e));
151 static bool isEqual(const Expression &LHS, const Expression &RHS) {
158 //===----------------------------------------------------------------------===//
159 // ValueTable Internal Functions
160 //===----------------------------------------------------------------------===//
162 Expression ValueTable::create_expression(Instruction *I) {
164 e.type = I->getType();
165 e.opcode = I->getOpcode();
166 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
168 e.varargs.push_back(lookup_or_add(*OI));
169 if (I->isCommutative()) {
170 // Ensure that commutative instructions that only differ by a permutation
171 // of their operands get the same value number by sorting the operand value
172 // numbers. Since all commutative instructions have two operands it is more
173 // efficient to sort by hand rather than using, say, std::sort.
174 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
175 if (e.varargs[0] > e.varargs[1])
176 std::swap(e.varargs[0], e.varargs[1]);
179 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
180 // Sort the operand value numbers so x<y and y>x get the same value number.
181 CmpInst::Predicate Predicate = C->getPredicate();
182 if (e.varargs[0] > e.varargs[1]) {
183 std::swap(e.varargs[0], e.varargs[1]);
184 Predicate = CmpInst::getSwappedPredicate(Predicate);
186 e.opcode = (C->getOpcode() << 8) | Predicate;
187 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
188 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
190 e.varargs.push_back(*II);
196 Expression ValueTable::create_cmp_expression(unsigned Opcode,
197 CmpInst::Predicate Predicate,
198 Value *LHS, Value *RHS) {
199 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
200 "Not a comparison!");
202 e.type = CmpInst::makeCmpResultType(LHS->getType());
203 e.varargs.push_back(lookup_or_add(LHS));
204 e.varargs.push_back(lookup_or_add(RHS));
206 // Sort the operand value numbers so x<y and y>x get the same value number.
207 if (e.varargs[0] > e.varargs[1]) {
208 std::swap(e.varargs[0], e.varargs[1]);
209 Predicate = CmpInst::getSwappedPredicate(Predicate);
211 e.opcode = (Opcode << 8) | Predicate;
215 Expression ValueTable::create_extractvalue_expression(ExtractValueInst *EI) {
216 assert(EI != 0 && "Not an ExtractValueInst?");
218 e.type = EI->getType();
221 IntrinsicInst *I = dyn_cast<IntrinsicInst>(EI->getAggregateOperand());
222 if (I != 0 && EI->getNumIndices() == 1 && *EI->idx_begin() == 0 ) {
223 // EI might be an extract from one of our recognised intrinsics. If it
224 // is we'll synthesize a semantically equivalent expression instead on
225 // an extract value expression.
226 switch (I->getIntrinsicID()) {
227 case Intrinsic::sadd_with_overflow:
228 case Intrinsic::uadd_with_overflow:
229 e.opcode = Instruction::Add;
231 case Intrinsic::ssub_with_overflow:
232 case Intrinsic::usub_with_overflow:
233 e.opcode = Instruction::Sub;
235 case Intrinsic::smul_with_overflow:
236 case Intrinsic::umul_with_overflow:
237 e.opcode = Instruction::Mul;
244 // Intrinsic recognized. Grab its args to finish building the expression.
245 assert(I->getNumArgOperands() == 2 &&
246 "Expect two args for recognised intrinsics.");
247 e.varargs.push_back(lookup_or_add(I->getArgOperand(0)));
248 e.varargs.push_back(lookup_or_add(I->getArgOperand(1)));
253 // Not a recognised intrinsic. Fall back to producing an extract value
255 e.opcode = EI->getOpcode();
256 for (Instruction::op_iterator OI = EI->op_begin(), OE = EI->op_end();
258 e.varargs.push_back(lookup_or_add(*OI));
260 for (ExtractValueInst::idx_iterator II = EI->idx_begin(), IE = EI->idx_end();
262 e.varargs.push_back(*II);
267 //===----------------------------------------------------------------------===//
268 // ValueTable External Functions
269 //===----------------------------------------------------------------------===//
271 /// add - Insert a value into the table with a specified value number.
272 void ValueTable::add(Value *V, uint32_t num) {
273 valueNumbering.insert(std::make_pair(V, num));
276 uint32_t ValueTable::lookup_or_add_call(CallInst *C) {
277 if (AA->doesNotAccessMemory(C)) {
278 Expression exp = create_expression(C);
279 uint32_t &e = expressionNumbering[exp];
280 if (!e) e = nextValueNumber++;
281 valueNumbering[C] = e;
283 } else if (AA->onlyReadsMemory(C)) {
284 Expression exp = create_expression(C);
285 uint32_t &e = expressionNumbering[exp];
287 e = nextValueNumber++;
288 valueNumbering[C] = e;
292 e = nextValueNumber++;
293 valueNumbering[C] = e;
297 MemDepResult local_dep = MD->getDependency(C);
299 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
300 valueNumbering[C] = nextValueNumber;
301 return nextValueNumber++;
304 if (local_dep.isDef()) {
305 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
307 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
308 valueNumbering[C] = nextValueNumber;
309 return nextValueNumber++;
312 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
313 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
314 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
316 valueNumbering[C] = nextValueNumber;
317 return nextValueNumber++;
321 uint32_t v = lookup_or_add(local_cdep);
322 valueNumbering[C] = v;
327 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
328 MD->getNonLocalCallDependency(CallSite(C));
329 // FIXME: Move the checking logic to MemDep!
332 // Check to see if we have a single dominating call instruction that is
334 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
335 const NonLocalDepEntry *I = &deps[i];
336 if (I->getResult().isNonLocal())
339 // We don't handle non-definitions. If we already have a call, reject
340 // instruction dependencies.
341 if (!I->getResult().isDef() || cdep != 0) {
346 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
347 // FIXME: All duplicated with non-local case.
348 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
349 cdep = NonLocalDepCall;
358 valueNumbering[C] = nextValueNumber;
359 return nextValueNumber++;
362 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
363 valueNumbering[C] = nextValueNumber;
364 return nextValueNumber++;
366 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
367 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
368 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
370 valueNumbering[C] = nextValueNumber;
371 return nextValueNumber++;
375 uint32_t v = lookup_or_add(cdep);
376 valueNumbering[C] = v;
380 valueNumbering[C] = nextValueNumber;
381 return nextValueNumber++;
385 /// lookup_or_add - Returns the value number for the specified value, assigning
386 /// it a new number if it did not have one before.
387 uint32_t ValueTable::lookup_or_add(Value *V) {
388 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
389 if (VI != valueNumbering.end())
392 if (!isa<Instruction>(V)) {
393 valueNumbering[V] = nextValueNumber;
394 return nextValueNumber++;
397 Instruction* I = cast<Instruction>(V);
399 switch (I->getOpcode()) {
400 case Instruction::Call:
401 return lookup_or_add_call(cast<CallInst>(I));
402 case Instruction::Add:
403 case Instruction::FAdd:
404 case Instruction::Sub:
405 case Instruction::FSub:
406 case Instruction::Mul:
407 case Instruction::FMul:
408 case Instruction::UDiv:
409 case Instruction::SDiv:
410 case Instruction::FDiv:
411 case Instruction::URem:
412 case Instruction::SRem:
413 case Instruction::FRem:
414 case Instruction::Shl:
415 case Instruction::LShr:
416 case Instruction::AShr:
417 case Instruction::And:
418 case Instruction::Or:
419 case Instruction::Xor:
420 case Instruction::ICmp:
421 case Instruction::FCmp:
422 case Instruction::Trunc:
423 case Instruction::ZExt:
424 case Instruction::SExt:
425 case Instruction::FPToUI:
426 case Instruction::FPToSI:
427 case Instruction::UIToFP:
428 case Instruction::SIToFP:
429 case Instruction::FPTrunc:
430 case Instruction::FPExt:
431 case Instruction::PtrToInt:
432 case Instruction::IntToPtr:
433 case Instruction::BitCast:
434 case Instruction::Select:
435 case Instruction::ExtractElement:
436 case Instruction::InsertElement:
437 case Instruction::ShuffleVector:
438 case Instruction::InsertValue:
439 case Instruction::GetElementPtr:
440 exp = create_expression(I);
442 case Instruction::ExtractValue:
443 exp = create_extractvalue_expression(cast<ExtractValueInst>(I));
446 valueNumbering[V] = nextValueNumber;
447 return nextValueNumber++;
450 uint32_t& e = expressionNumbering[exp];
451 if (!e) e = nextValueNumber++;
452 valueNumbering[V] = e;
456 /// lookup - Returns the value number of the specified value. Fails if
457 /// the value has not yet been numbered.
458 uint32_t ValueTable::lookup(Value *V) const {
459 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
460 assert(VI != valueNumbering.end() && "Value not numbered?");
464 /// lookup_or_add_cmp - Returns the value number of the given comparison,
465 /// assigning it a new number if it did not have one before. Useful when
466 /// we deduced the result of a comparison, but don't immediately have an
467 /// instruction realizing that comparison to hand.
468 uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode,
469 CmpInst::Predicate Predicate,
470 Value *LHS, Value *RHS) {
471 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS);
472 uint32_t& e = expressionNumbering[exp];
473 if (!e) e = nextValueNumber++;
477 /// clear - Remove all entries from the ValueTable.
478 void ValueTable::clear() {
479 valueNumbering.clear();
480 expressionNumbering.clear();
484 /// erase - Remove a value from the value numbering.
485 void ValueTable::erase(Value *V) {
486 valueNumbering.erase(V);
489 /// verifyRemoved - Verify that the value is removed from all internal data
491 void ValueTable::verifyRemoved(const Value *V) const {
492 for (DenseMap<Value*, uint32_t>::const_iterator
493 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
494 assert(I->first != V && "Inst still occurs in value numbering map!");
498 //===----------------------------------------------------------------------===//
500 //===----------------------------------------------------------------------===//
504 struct AvailableValueInBlock {
505 /// BB - The basic block in question.
508 SimpleVal, // A simple offsetted value that is accessed.
509 LoadVal, // A value produced by a load.
510 MemIntrin // A memory intrinsic which is loaded from.
513 /// V - The value that is live out of the block.
514 PointerIntPair<Value *, 2, ValType> Val;
516 /// Offset - The byte offset in Val that is interesting for the load query.
519 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
520 unsigned Offset = 0) {
521 AvailableValueInBlock Res;
523 Res.Val.setPointer(V);
524 Res.Val.setInt(SimpleVal);
529 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
530 unsigned Offset = 0) {
531 AvailableValueInBlock Res;
533 Res.Val.setPointer(MI);
534 Res.Val.setInt(MemIntrin);
539 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
540 unsigned Offset = 0) {
541 AvailableValueInBlock Res;
543 Res.Val.setPointer(LI);
544 Res.Val.setInt(LoadVal);
549 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
550 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
551 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
553 Value *getSimpleValue() const {
554 assert(isSimpleValue() && "Wrong accessor");
555 return Val.getPointer();
558 LoadInst *getCoercedLoadValue() const {
559 assert(isCoercedLoadValue() && "Wrong accessor");
560 return cast<LoadInst>(Val.getPointer());
563 MemIntrinsic *getMemIntrinValue() const {
564 assert(isMemIntrinValue() && "Wrong accessor");
565 return cast<MemIntrinsic>(Val.getPointer());
568 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
569 /// defined here to the specified type. This handles various coercion cases.
570 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const;
573 class GVN : public FunctionPass {
575 MemoryDependenceAnalysis *MD;
577 const DataLayout *TD;
578 const TargetLibraryInfo *TLI;
582 /// LeaderTable - A mapping from value numbers to lists of Value*'s that
583 /// have that value number. Use findLeader to query it.
584 struct LeaderTableEntry {
586 const BasicBlock *BB;
587 LeaderTableEntry *Next;
589 DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
590 BumpPtrAllocator TableAllocator;
592 SmallVector<Instruction*, 8> InstrsToErase;
594 typedef SmallVector<NonLocalDepResult, 64> LoadDepVect;
595 typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect;
596 typedef SmallVector<BasicBlock*, 64> UnavailBlkVect;
599 static char ID; // Pass identification, replacement for typeid
600 explicit GVN(bool noloads = false)
601 : FunctionPass(ID), NoLoads(noloads), MD(0) {
602 initializeGVNPass(*PassRegistry::getPassRegistry());
605 bool runOnFunction(Function &F);
607 /// markInstructionForDeletion - This removes the specified instruction from
608 /// our various maps and marks it for deletion.
609 void markInstructionForDeletion(Instruction *I) {
611 InstrsToErase.push_back(I);
614 const DataLayout *getDataLayout() const { return TD; }
615 DominatorTree &getDominatorTree() const { return *DT; }
616 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
617 MemoryDependenceAnalysis &getMemDep() const { return *MD; }
619 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for
620 /// its value number.
621 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) {
622 LeaderTableEntry &Curr = LeaderTable[N];
629 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
632 Node->Next = Curr.Next;
636 /// removeFromLeaderTable - Scan the list of values corresponding to a given
637 /// value number, and remove the given instruction if encountered.
638 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) {
639 LeaderTableEntry* Prev = 0;
640 LeaderTableEntry* Curr = &LeaderTable[N];
642 while (Curr->Val != I || Curr->BB != BB) {
648 Prev->Next = Curr->Next;
654 LeaderTableEntry* Next = Curr->Next;
655 Curr->Val = Next->Val;
657 Curr->Next = Next->Next;
662 // List of critical edges to be split between iterations.
663 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
665 // This transformation requires dominator postdominator info
666 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
667 AU.addRequired<DominatorTree>();
668 AU.addRequired<TargetLibraryInfo>();
670 AU.addRequired<MemoryDependenceAnalysis>();
671 AU.addRequired<AliasAnalysis>();
673 AU.addPreserved<DominatorTree>();
674 AU.addPreserved<AliasAnalysis>();
678 // Helper fuctions of redundant load elimination
679 bool processLoad(LoadInst *L);
680 bool processNonLocalLoad(LoadInst *L);
681 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
682 AvailValInBlkVect &ValuesPerBlock,
683 UnavailBlkVect &UnavailableBlocks);
684 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
685 UnavailBlkVect &UnavailableBlocks);
687 // Other helper routines
688 bool processInstruction(Instruction *I);
689 bool processBlock(BasicBlock *BB);
690 void dump(DenseMap<uint32_t, Value*> &d);
691 bool iterateOnFunction(Function &F);
692 bool performPRE(Function &F);
693 Value *findLeader(const BasicBlock *BB, uint32_t num);
694 void cleanupGlobalSets();
695 void verifyRemoved(const Instruction *I) const;
696 bool splitCriticalEdges();
697 BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ);
698 unsigned replaceAllDominatedUsesWith(Value *From, Value *To,
699 const BasicBlockEdge &Root);
700 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root);
706 // createGVNPass - The public interface to this file...
707 FunctionPass *llvm::createGVNPass(bool NoLoads) {
708 return new GVN(NoLoads);
711 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
712 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
713 INITIALIZE_PASS_DEPENDENCY(DominatorTree)
714 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
715 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
716 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
718 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
719 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
721 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
722 E = d.end(); I != E; ++I) {
723 errs() << I->first << "\n";
730 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
731 /// we're analyzing is fully available in the specified block. As we go, keep
732 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
733 /// map is actually a tri-state map with the following values:
734 /// 0) we know the block *is not* fully available.
735 /// 1) we know the block *is* fully available.
736 /// 2) we do not know whether the block is fully available or not, but we are
737 /// currently speculating that it will be.
738 /// 3) we are speculating for this block and have used that to speculate for
740 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
741 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
742 uint32_t RecurseDepth) {
743 if (RecurseDepth > MaxRecurseDepth)
746 // Optimistically assume that the block is fully available and check to see
747 // if we already know about this block in one lookup.
748 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
749 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
751 // If the entry already existed for this block, return the precomputed value.
753 // If this is a speculative "available" value, mark it as being used for
754 // speculation of other blocks.
755 if (IV.first->second == 2)
756 IV.first->second = 3;
757 return IV.first->second != 0;
760 // Otherwise, see if it is fully available in all predecessors.
761 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
763 // If this block has no predecessors, it isn't live-in here.
765 goto SpeculationFailure;
767 for (; PI != PE; ++PI)
768 // If the value isn't fully available in one of our predecessors, then it
769 // isn't fully available in this block either. Undo our previous
770 // optimistic assumption and bail out.
771 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
772 goto SpeculationFailure;
776 // SpeculationFailure - If we get here, we found out that this is not, after
777 // all, a fully-available block. We have a problem if we speculated on this and
778 // used the speculation to mark other blocks as available.
780 char &BBVal = FullyAvailableBlocks[BB];
782 // If we didn't speculate on this, just return with it set to false.
788 // If we did speculate on this value, we could have blocks set to 1 that are
789 // incorrect. Walk the (transitive) successors of this block and mark them as
791 SmallVector<BasicBlock*, 32> BBWorklist;
792 BBWorklist.push_back(BB);
795 BasicBlock *Entry = BBWorklist.pop_back_val();
796 // Note that this sets blocks to 0 (unavailable) if they happen to not
797 // already be in FullyAvailableBlocks. This is safe.
798 char &EntryVal = FullyAvailableBlocks[Entry];
799 if (EntryVal == 0) continue; // Already unavailable.
801 // Mark as unavailable.
804 for (succ_iterator I = succ_begin(Entry), E = succ_end(Entry); I != E; ++I)
805 BBWorklist.push_back(*I);
806 } while (!BBWorklist.empty());
812 /// CanCoerceMustAliasedValueToLoad - Return true if
813 /// CoerceAvailableValueToLoadType will succeed.
814 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
816 const DataLayout &TD) {
817 // If the loaded or stored value is an first class array or struct, don't try
818 // to transform them. We need to be able to bitcast to integer.
819 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
820 StoredVal->getType()->isStructTy() ||
821 StoredVal->getType()->isArrayTy())
824 // The store has to be at least as big as the load.
825 if (TD.getTypeSizeInBits(StoredVal->getType()) <
826 TD.getTypeSizeInBits(LoadTy))
832 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
833 /// then a load from a must-aliased pointer of a different type, try to coerce
834 /// the stored value. LoadedTy is the type of the load we want to replace and
835 /// InsertPt is the place to insert new instructions.
837 /// If we can't do it, return null.
838 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
840 Instruction *InsertPt,
841 const DataLayout &TD) {
842 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, TD))
845 // If this is already the right type, just return it.
846 Type *StoredValTy = StoredVal->getType();
848 uint64_t StoreSize = TD.getTypeSizeInBits(StoredValTy);
849 uint64_t LoadSize = TD.getTypeSizeInBits(LoadedTy);
851 // If the store and reload are the same size, we can always reuse it.
852 if (StoreSize == LoadSize) {
853 // Pointer to Pointer -> use bitcast.
854 if (StoredValTy->getScalarType()->isPointerTy() &&
855 LoadedTy->getScalarType()->isPointerTy())
856 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
858 // Convert source pointers to integers, which can be bitcast.
859 if (StoredValTy->getScalarType()->isPointerTy()) {
860 StoredValTy = TD.getIntPtrType(StoredValTy);
861 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
864 Type *TypeToCastTo = LoadedTy;
865 if (TypeToCastTo->getScalarType()->isPointerTy())
866 TypeToCastTo = TD.getIntPtrType(TypeToCastTo);
868 if (StoredValTy != TypeToCastTo)
869 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
871 // Cast to pointer if the load needs a pointer type.
872 if (LoadedTy->getScalarType()->isPointerTy())
873 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
878 // If the loaded value is smaller than the available value, then we can
879 // extract out a piece from it. If the available value is too small, then we
880 // can't do anything.
881 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
883 // Convert source pointers to integers, which can be manipulated.
884 if (StoredValTy->getScalarType()->isPointerTy()) {
885 StoredValTy = TD.getIntPtrType(StoredValTy);
886 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
889 // Convert vectors and fp to integer, which can be manipulated.
890 if (!StoredValTy->isIntegerTy()) {
891 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
892 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
895 // If this is a big-endian system, we need to shift the value down to the low
896 // bits so that a truncate will work.
897 if (TD.isBigEndian()) {
898 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
899 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
902 // Truncate the integer to the right size now.
903 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
904 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
906 if (LoadedTy == NewIntTy)
909 // If the result is a pointer, inttoptr.
910 if (LoadedTy->getScalarType()->isPointerTy())
911 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
913 // Otherwise, bitcast.
914 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
917 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
918 /// memdep query of a load that ends up being a clobbering memory write (store,
919 /// memset, memcpy, memmove). This means that the write *may* provide bits used
920 /// by the load but we can't be sure because the pointers don't mustalias.
922 /// Check this case to see if there is anything more we can do before we give
923 /// up. This returns -1 if we have to give up, or a byte number in the stored
924 /// value of the piece that feeds the load.
925 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
927 uint64_t WriteSizeInBits,
928 const DataLayout &TD) {
929 // If the loaded or stored value is a first class array or struct, don't try
930 // to transform them. We need to be able to bitcast to integer.
931 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
934 int64_t StoreOffset = 0, LoadOffset = 0;
935 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr,StoreOffset,&TD);
936 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, &TD);
937 if (StoreBase != LoadBase)
940 // If the load and store are to the exact same address, they should have been
941 // a must alias. AA must have gotten confused.
942 // FIXME: Study to see if/when this happens. One case is forwarding a memset
943 // to a load from the base of the memset.
945 if (LoadOffset == StoreOffset) {
946 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
947 << "Base = " << *StoreBase << "\n"
948 << "Store Ptr = " << *WritePtr << "\n"
949 << "Store Offs = " << StoreOffset << "\n"
950 << "Load Ptr = " << *LoadPtr << "\n";
955 // If the load and store don't overlap at all, the store doesn't provide
956 // anything to the load. In this case, they really don't alias at all, AA
957 // must have gotten confused.
958 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy);
960 if ((WriteSizeInBits & 7) | (LoadSize & 7))
962 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
966 bool isAAFailure = false;
967 if (StoreOffset < LoadOffset)
968 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
970 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
974 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
975 << "Base = " << *StoreBase << "\n"
976 << "Store Ptr = " << *WritePtr << "\n"
977 << "Store Offs = " << StoreOffset << "\n"
978 << "Load Ptr = " << *LoadPtr << "\n";
984 // If the Load isn't completely contained within the stored bits, we don't
985 // have all the bits to feed it. We could do something crazy in the future
986 // (issue a smaller load then merge the bits in) but this seems unlikely to be
988 if (StoreOffset > LoadOffset ||
989 StoreOffset+StoreSize < LoadOffset+LoadSize)
992 // Okay, we can do this transformation. Return the number of bytes into the
993 // store that the load is.
994 return LoadOffset-StoreOffset;
997 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
998 /// memdep query of a load that ends up being a clobbering store.
999 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
1001 const DataLayout &TD) {
1002 // Cannot handle reading from store of first-class aggregate yet.
1003 if (DepSI->getValueOperand()->getType()->isStructTy() ||
1004 DepSI->getValueOperand()->getType()->isArrayTy())
1007 Value *StorePtr = DepSI->getPointerOperand();
1008 uint64_t StoreSize =TD.getTypeSizeInBits(DepSI->getValueOperand()->getType());
1009 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1010 StorePtr, StoreSize, TD);
1013 /// AnalyzeLoadFromClobberingLoad - This function is called when we have a
1014 /// memdep query of a load that ends up being clobbered by another load. See if
1015 /// the other load can feed into the second load.
1016 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
1017 LoadInst *DepLI, const DataLayout &TD){
1018 // Cannot handle reading from store of first-class aggregate yet.
1019 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
1022 Value *DepPtr = DepLI->getPointerOperand();
1023 uint64_t DepSize = TD.getTypeSizeInBits(DepLI->getType());
1024 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, TD);
1025 if (R != -1) return R;
1027 // If we have a load/load clobber an DepLI can be widened to cover this load,
1028 // then we should widen it!
1029 int64_t LoadOffs = 0;
1030 const Value *LoadBase =
1031 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, &TD);
1032 unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
1034 unsigned Size = MemoryDependenceAnalysis::
1035 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, TD);
1036 if (Size == 0) return -1;
1038 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, TD);
1043 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
1045 const DataLayout &TD) {
1046 // If the mem operation is a non-constant size, we can't handle it.
1047 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1048 if (SizeCst == 0) return -1;
1049 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1051 // If this is memset, we just need to see if the offset is valid in the size
1053 if (MI->getIntrinsicID() == Intrinsic::memset)
1054 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
1057 // If we have a memcpy/memmove, the only case we can handle is if this is a
1058 // copy from constant memory. In that case, we can read directly from the
1060 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1062 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1063 if (Src == 0) return -1;
1065 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &TD));
1066 if (GV == 0 || !GV->isConstant()) return -1;
1068 // See if the access is within the bounds of the transfer.
1069 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1070 MI->getDest(), MemSizeInBits, TD);
1074 unsigned AS = Src->getType()->getPointerAddressSpace();
1075 // Otherwise, see if we can constant fold a load from the constant with the
1076 // offset applied as appropriate.
1077 Src = ConstantExpr::getBitCast(Src,
1078 Type::getInt8PtrTy(Src->getContext(), AS));
1079 Constant *OffsetCst =
1080 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1081 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1082 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1083 if (ConstantFoldLoadFromConstPtr(Src, &TD))
1089 /// GetStoreValueForLoad - This function is called when we have a
1090 /// memdep query of a load that ends up being a clobbering store. This means
1091 /// that the store provides bits used by the load but we the pointers don't
1092 /// mustalias. Check this case to see if there is anything more we can do
1093 /// before we give up.
1094 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1096 Instruction *InsertPt, const DataLayout &TD){
1097 LLVMContext &Ctx = SrcVal->getType()->getContext();
1099 uint64_t StoreSize = (TD.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1100 uint64_t LoadSize = (TD.getTypeSizeInBits(LoadTy) + 7) / 8;
1102 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1104 // Compute which bits of the stored value are being used by the load. Convert
1105 // to an integer type to start with.
1106 if (SrcVal->getType()->getScalarType()->isPointerTy())
1107 SrcVal = Builder.CreatePtrToInt(SrcVal,
1108 TD.getIntPtrType(SrcVal->getType()));
1109 if (!SrcVal->getType()->isIntegerTy())
1110 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1112 // Shift the bits to the least significant depending on endianness.
1114 if (TD.isLittleEndian())
1115 ShiftAmt = Offset*8;
1117 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1120 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1122 if (LoadSize != StoreSize)
1123 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1125 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, TD);
1128 /// GetLoadValueForLoad - This function is called when we have a
1129 /// memdep query of a load that ends up being a clobbering load. This means
1130 /// that the load *may* provide bits used by the load but we can't be sure
1131 /// because the pointers don't mustalias. Check this case to see if there is
1132 /// anything more we can do before we give up.
1133 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1134 Type *LoadTy, Instruction *InsertPt,
1136 const DataLayout &TD = *gvn.getDataLayout();
1137 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1138 // widen SrcVal out to a larger load.
1139 unsigned SrcValSize = TD.getTypeStoreSize(SrcVal->getType());
1140 unsigned LoadSize = TD.getTypeStoreSize(LoadTy);
1141 if (Offset+LoadSize > SrcValSize) {
1142 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1143 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1144 // If we have a load/load clobber an DepLI can be widened to cover this
1145 // load, then we should widen it to the next power of 2 size big enough!
1146 unsigned NewLoadSize = Offset+LoadSize;
1147 if (!isPowerOf2_32(NewLoadSize))
1148 NewLoadSize = NextPowerOf2(NewLoadSize);
1150 Value *PtrVal = SrcVal->getPointerOperand();
1152 // Insert the new load after the old load. This ensures that subsequent
1153 // memdep queries will find the new load. We can't easily remove the old
1154 // load completely because it is already in the value numbering table.
1155 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1157 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1158 DestPTy = PointerType::get(DestPTy,
1159 PtrVal->getType()->getPointerAddressSpace());
1160 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1161 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1162 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1163 NewLoad->takeName(SrcVal);
1164 NewLoad->setAlignment(SrcVal->getAlignment());
1166 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1167 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1169 // Replace uses of the original load with the wider load. On a big endian
1170 // system, we need to shift down to get the relevant bits.
1171 Value *RV = NewLoad;
1172 if (TD.isBigEndian())
1173 RV = Builder.CreateLShr(RV,
1174 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
1175 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1176 SrcVal->replaceAllUsesWith(RV);
1178 // We would like to use gvn.markInstructionForDeletion here, but we can't
1179 // because the load is already memoized into the leader map table that GVN
1180 // tracks. It is potentially possible to remove the load from the table,
1181 // but then there all of the operations based on it would need to be
1182 // rehashed. Just leave the dead load around.
1183 gvn.getMemDep().removeInstruction(SrcVal);
1187 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, TD);
1191 /// GetMemInstValueForLoad - This function is called when we have a
1192 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1193 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1194 Type *LoadTy, Instruction *InsertPt,
1195 const DataLayout &TD){
1196 LLVMContext &Ctx = LoadTy->getContext();
1197 uint64_t LoadSize = TD.getTypeSizeInBits(LoadTy)/8;
1199 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1201 // We know that this method is only called when the mem transfer fully
1202 // provides the bits for the load.
1203 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1204 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1205 // independently of what the offset is.
1206 Value *Val = MSI->getValue();
1208 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1210 Value *OneElt = Val;
1212 // Splat the value out to the right number of bits.
1213 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1214 // If we can double the number of bytes set, do it.
1215 if (NumBytesSet*2 <= LoadSize) {
1216 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1217 Val = Builder.CreateOr(Val, ShVal);
1222 // Otherwise insert one byte at a time.
1223 Value *ShVal = Builder.CreateShl(Val, 1*8);
1224 Val = Builder.CreateOr(OneElt, ShVal);
1228 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, TD);
1231 // Otherwise, this is a memcpy/memmove from a constant global.
1232 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1233 Constant *Src = cast<Constant>(MTI->getSource());
1234 unsigned AS = Src->getType()->getPointerAddressSpace();
1236 // Otherwise, see if we can constant fold a load from the constant with the
1237 // offset applied as appropriate.
1238 Src = ConstantExpr::getBitCast(Src,
1239 Type::getInt8PtrTy(Src->getContext(), AS));
1240 Constant *OffsetCst =
1241 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1242 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1243 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1244 return ConstantFoldLoadFromConstPtr(Src, &TD);
1248 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1249 /// construct SSA form, allowing us to eliminate LI. This returns the value
1250 /// that should be used at LI's definition site.
1251 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1252 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1254 // Check for the fully redundant, dominating load case. In this case, we can
1255 // just use the dominating value directly.
1256 if (ValuesPerBlock.size() == 1 &&
1257 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1259 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn);
1261 // Otherwise, we have to construct SSA form.
1262 SmallVector<PHINode*, 8> NewPHIs;
1263 SSAUpdater SSAUpdate(&NewPHIs);
1264 SSAUpdate.Initialize(LI->getType(), LI->getName());
1266 Type *LoadTy = LI->getType();
1268 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1269 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1270 BasicBlock *BB = AV.BB;
1272 if (SSAUpdate.HasValueForBlock(BB))
1275 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn));
1278 // Perform PHI construction.
1279 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1281 // If new PHI nodes were created, notify alias analysis.
1282 if (V->getType()->getScalarType()->isPointerTy()) {
1283 AliasAnalysis *AA = gvn.getAliasAnalysis();
1285 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1286 AA->copyValue(LI, NewPHIs[i]);
1288 // Now that we've copied information to the new PHIs, scan through
1289 // them again and inform alias analysis that we've added potentially
1290 // escaping uses to any values that are operands to these PHIs.
1291 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
1292 PHINode *P = NewPHIs[i];
1293 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) {
1294 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
1295 AA->addEscapingUse(P->getOperandUse(jj));
1303 Value *AvailableValueInBlock::MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const {
1305 if (isSimpleValue()) {
1306 Res = getSimpleValue();
1307 if (Res->getType() != LoadTy) {
1308 const DataLayout *TD = gvn.getDataLayout();
1309 assert(TD && "Need target data to handle type mismatch case");
1310 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1313 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1314 << *getSimpleValue() << '\n'
1315 << *Res << '\n' << "\n\n\n");
1317 } else if (isCoercedLoadValue()) {
1318 LoadInst *Load = getCoercedLoadValue();
1319 if (Load->getType() == LoadTy && Offset == 0) {
1322 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1325 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1326 << *getCoercedLoadValue() << '\n'
1327 << *Res << '\n' << "\n\n\n");
1330 const DataLayout *TD = gvn.getDataLayout();
1331 assert(TD && "Need target data to handle type mismatch case");
1332 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1333 LoadTy, BB->getTerminator(), *TD);
1334 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1335 << " " << *getMemIntrinValue() << '\n'
1336 << *Res << '\n' << "\n\n\n");
1341 static bool isLifetimeStart(const Instruction *Inst) {
1342 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1343 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1347 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1348 AvailValInBlkVect &ValuesPerBlock,
1349 UnavailBlkVect &UnavailableBlocks) {
1351 // Filter out useless results (non-locals, etc). Keep track of the blocks
1352 // where we have a value available in repl, also keep track of whether we see
1353 // dependencies that produce an unknown value for the load (such as a call
1354 // that could potentially clobber the load).
1355 unsigned NumDeps = Deps.size();
1356 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1357 BasicBlock *DepBB = Deps[i].getBB();
1358 MemDepResult DepInfo = Deps[i].getResult();
1360 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1361 UnavailableBlocks.push_back(DepBB);
1365 if (DepInfo.isClobber()) {
1366 // The address being loaded in this non-local block may not be the same as
1367 // the pointer operand of the load if PHI translation occurs. Make sure
1368 // to consider the right address.
1369 Value *Address = Deps[i].getAddress();
1371 // If the dependence is to a store that writes to a superset of the bits
1372 // read by the load, we can extract the bits we need for the load from the
1374 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1375 if (TD && Address) {
1376 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1379 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1380 DepSI->getValueOperand(),
1387 // Check to see if we have something like this:
1390 // if we have this, replace the later with an extraction from the former.
1391 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1392 // If this is a clobber and L is the first instruction in its block, then
1393 // we have the first instruction in the entry block.
1394 if (DepLI != LI && Address && TD) {
1395 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(),
1396 LI->getPointerOperand(),
1400 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1407 // If the clobbering value is a memset/memcpy/memmove, see if we can
1408 // forward a value on from it.
1409 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1410 if (TD && Address) {
1411 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1414 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1421 UnavailableBlocks.push_back(DepBB);
1425 // DepInfo.isDef() here
1427 Instruction *DepInst = DepInfo.getInst();
1429 // Loading the allocation -> undef.
1430 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
1431 // Loading immediately after lifetime begin -> undef.
1432 isLifetimeStart(DepInst)) {
1433 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1434 UndefValue::get(LI->getType())));
1438 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1439 // Reject loads and stores that are to the same address but are of
1440 // different types if we have to.
1441 if (S->getValueOperand()->getType() != LI->getType()) {
1442 // If the stored value is larger or equal to the loaded value, we can
1444 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1445 LI->getType(), *TD)) {
1446 UnavailableBlocks.push_back(DepBB);
1451 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1452 S->getValueOperand()));
1456 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1457 // If the types mismatch and we can't handle it, reject reuse of the load.
1458 if (LD->getType() != LI->getType()) {
1459 // If the stored value is larger or equal to the loaded value, we can
1461 if (TD == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*TD)){
1462 UnavailableBlocks.push_back(DepBB);
1466 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1470 UnavailableBlocks.push_back(DepBB);
1474 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1475 UnavailBlkVect &UnavailableBlocks) {
1476 // Okay, we have *some* definitions of the value. This means that the value
1477 // is available in some of our (transitive) predecessors. Lets think about
1478 // doing PRE of this load. This will involve inserting a new load into the
1479 // predecessor when it's not available. We could do this in general, but
1480 // prefer to not increase code size. As such, we only do this when we know
1481 // that we only have to insert *one* load (which means we're basically moving
1482 // the load, not inserting a new one).
1484 SmallPtrSet<BasicBlock *, 4> Blockers;
1485 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1486 Blockers.insert(UnavailableBlocks[i]);
1488 // Let's find the first basic block with more than one predecessor. Walk
1489 // backwards through predecessors if needed.
1490 BasicBlock *LoadBB = LI->getParent();
1491 BasicBlock *TmpBB = LoadBB;
1493 while (TmpBB->getSinglePredecessor()) {
1494 TmpBB = TmpBB->getSinglePredecessor();
1495 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1497 if (Blockers.count(TmpBB))
1500 // If any of these blocks has more than one successor (i.e. if the edge we
1501 // just traversed was critical), then there are other paths through this
1502 // block along which the load may not be anticipated. Hoisting the load
1503 // above this block would be adding the load to execution paths along
1504 // which it was not previously executed.
1505 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1512 // Check to see how many predecessors have the loaded value fully
1514 DenseMap<BasicBlock*, Value*> PredLoads;
1515 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1516 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1517 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1518 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1519 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1521 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1522 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1524 BasicBlock *Pred = *PI;
1525 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1528 PredLoads[Pred] = 0;
1530 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1531 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1532 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1533 << Pred->getName() << "': " << *LI << '\n');
1537 if (LoadBB->isLandingPad()) {
1539 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '"
1540 << Pred->getName() << "': " << *LI << '\n');
1544 CriticalEdgePred.push_back(Pred);
1548 // Decide whether PRE is profitable for this load.
1549 unsigned NumUnavailablePreds = PredLoads.size();
1550 assert(NumUnavailablePreds != 0 &&
1551 "Fully available value should already be eliminated!");
1553 // If this load is unavailable in multiple predecessors, reject it.
1554 // FIXME: If we could restructure the CFG, we could make a common pred with
1555 // all the preds that don't have an available LI and insert a new load into
1557 if (NumUnavailablePreds != 1)
1560 // Split critical edges, and update the unavailable predecessors accordingly.
1561 for (SmallVectorImpl<BasicBlock *>::iterator I = CriticalEdgePred.begin(),
1562 E = CriticalEdgePred.end(); I != E; I++) {
1563 BasicBlock *OrigPred = *I;
1564 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1565 PredLoads.erase(OrigPred);
1566 PredLoads[NewPred] = 0;
1567 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1568 << LoadBB->getName() << '\n');
1571 // Check if the load can safely be moved to all the unavailable predecessors.
1572 bool CanDoPRE = true;
1573 SmallVector<Instruction*, 8> NewInsts;
1574 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1575 E = PredLoads.end(); I != E; ++I) {
1576 BasicBlock *UnavailablePred = I->first;
1578 // Do PHI translation to get its value in the predecessor if necessary. The
1579 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1581 // If all preds have a single successor, then we know it is safe to insert
1582 // the load on the pred (?!?), so we can insert code to materialize the
1583 // pointer if it is not available.
1584 PHITransAddr Address(LI->getPointerOperand(), TD);
1586 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1589 // If we couldn't find or insert a computation of this phi translated value,
1592 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1593 << *LI->getPointerOperand() << "\n");
1598 I->second = LoadPtr;
1602 while (!NewInsts.empty()) {
1603 Instruction *I = NewInsts.pop_back_val();
1604 if (MD) MD->removeInstruction(I);
1605 I->eraseFromParent();
1607 // HINT:Don't revert the edge-splitting as following transformation may
1608 // also need to split these critial edges.
1609 return !CriticalEdgePred.empty();
1612 // Okay, we can eliminate this load by inserting a reload in the predecessor
1613 // and using PHI construction to get the value in the other predecessors, do
1615 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1616 DEBUG(if (!NewInsts.empty())
1617 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1618 << *NewInsts.back() << '\n');
1620 // Assign value numbers to the new instructions.
1621 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1622 // FIXME: We really _ought_ to insert these value numbers into their
1623 // parent's availability map. However, in doing so, we risk getting into
1624 // ordering issues. If a block hasn't been processed yet, we would be
1625 // marking a value as AVAIL-IN, which isn't what we intend.
1626 VN.lookup_or_add(NewInsts[i]);
1629 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1630 E = PredLoads.end(); I != E; ++I) {
1631 BasicBlock *UnavailablePred = I->first;
1632 Value *LoadPtr = I->second;
1634 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1636 UnavailablePred->getTerminator());
1638 // Transfer the old load's TBAA tag to the new load.
1639 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa))
1640 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1642 // Transfer DebugLoc.
1643 NewLoad->setDebugLoc(LI->getDebugLoc());
1645 // Add the newly created load.
1646 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1648 MD->invalidateCachedPointerInfo(LoadPtr);
1649 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1652 // Perform PHI construction.
1653 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1654 LI->replaceAllUsesWith(V);
1655 if (isa<PHINode>(V))
1657 if (V->getType()->getScalarType()->isPointerTy())
1658 MD->invalidateCachedPointerInfo(V);
1659 markInstructionForDeletion(LI);
1664 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1665 /// non-local by performing PHI construction.
1666 bool GVN::processNonLocalLoad(LoadInst *LI) {
1667 // Step 1: Find the non-local dependencies of the load.
1669 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
1670 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
1672 // If we had to process more than one hundred blocks to find the
1673 // dependencies, this load isn't worth worrying about. Optimizing
1674 // it will be too expensive.
1675 unsigned NumDeps = Deps.size();
1679 // If we had a phi translation failure, we'll have a single entry which is a
1680 // clobber in the current block. Reject this early.
1682 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1684 dbgs() << "GVN: non-local load ";
1685 WriteAsOperand(dbgs(), LI);
1686 dbgs() << " has unknown dependencies\n";
1691 // Step 2: Analyze the availability of the load
1692 AvailValInBlkVect ValuesPerBlock;
1693 UnavailBlkVect UnavailableBlocks;
1694 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1696 // If we have no predecessors that produce a known value for this load, exit
1698 if (ValuesPerBlock.empty())
1701 // Step 3: Eliminate fully redundancy.
1703 // If all of the instructions we depend on produce a known value for this
1704 // load, then it is fully redundant and we can use PHI insertion to compute
1705 // its value. Insert PHIs and remove the fully redundant value now.
1706 if (UnavailableBlocks.empty()) {
1707 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1709 // Perform PHI construction.
1710 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1711 LI->replaceAllUsesWith(V);
1713 if (isa<PHINode>(V))
1715 if (V->getType()->getScalarType()->isPointerTy())
1716 MD->invalidateCachedPointerInfo(V);
1717 markInstructionForDeletion(LI);
1722 // Step 4: Eliminate partial redundancy.
1723 if (!EnablePRE || !EnableLoadPRE)
1726 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1730 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1731 // Patch the replacement so that it is not more restrictive than the value
1733 BinaryOperator *Op = dyn_cast<BinaryOperator>(I);
1734 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl);
1735 if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) &&
1736 isa<OverflowingBinaryOperator>(ReplOp)) {
1737 if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap())
1738 ReplOp->setHasNoSignedWrap(false);
1739 if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap())
1740 ReplOp->setHasNoUnsignedWrap(false);
1742 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) {
1743 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
1744 ReplInst->getAllMetadataOtherThanDebugLoc(Metadata);
1745 for (int i = 0, n = Metadata.size(); i < n; ++i) {
1746 unsigned Kind = Metadata[i].first;
1747 MDNode *IMD = I->getMetadata(Kind);
1748 MDNode *ReplMD = Metadata[i].second;
1751 ReplInst->setMetadata(Kind, NULL); // Remove unknown metadata
1753 case LLVMContext::MD_dbg:
1754 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg");
1755 case LLVMContext::MD_tbaa:
1756 ReplInst->setMetadata(Kind, MDNode::getMostGenericTBAA(IMD, ReplMD));
1758 case LLVMContext::MD_range:
1759 ReplInst->setMetadata(Kind, MDNode::getMostGenericRange(IMD, ReplMD));
1761 case LLVMContext::MD_prof:
1762 llvm_unreachable("MD_prof in a non terminator instruction");
1764 case LLVMContext::MD_fpmath:
1765 ReplInst->setMetadata(Kind, MDNode::getMostGenericFPMath(IMD, ReplMD));
1772 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1773 patchReplacementInstruction(I, Repl);
1774 I->replaceAllUsesWith(Repl);
1777 /// processLoad - Attempt to eliminate a load, first by eliminating it
1778 /// locally, and then attempting non-local elimination if that fails.
1779 bool GVN::processLoad(LoadInst *L) {
1786 if (L->use_empty()) {
1787 markInstructionForDeletion(L);
1791 // ... to a pointer that has been loaded from before...
1792 MemDepResult Dep = MD->getDependency(L);
1794 // If we have a clobber and target data is around, see if this is a clobber
1795 // that we can fix up through code synthesis.
1796 if (Dep.isClobber() && TD) {
1797 // Check to see if we have something like this:
1798 // store i32 123, i32* %P
1799 // %A = bitcast i32* %P to i8*
1800 // %B = gep i8* %A, i32 1
1803 // We could do that by recognizing if the clobber instructions are obviously
1804 // a common base + constant offset, and if the previous store (or memset)
1805 // completely covers this load. This sort of thing can happen in bitfield
1807 Value *AvailVal = 0;
1808 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1809 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1810 L->getPointerOperand(),
1813 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1814 L->getType(), L, *TD);
1817 // Check to see if we have something like this:
1820 // if we have this, replace the later with an extraction from the former.
1821 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1822 // If this is a clobber and L is the first instruction in its block, then
1823 // we have the first instruction in the entry block.
1827 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(),
1828 L->getPointerOperand(),
1831 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1834 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1835 // a value on from it.
1836 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1837 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1838 L->getPointerOperand(),
1841 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *TD);
1845 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1846 << *AvailVal << '\n' << *L << "\n\n\n");
1848 // Replace the load!
1849 L->replaceAllUsesWith(AvailVal);
1850 if (AvailVal->getType()->getScalarType()->isPointerTy())
1851 MD->invalidateCachedPointerInfo(AvailVal);
1852 markInstructionForDeletion(L);
1858 // If the value isn't available, don't do anything!
1859 if (Dep.isClobber()) {
1861 // fast print dep, using operator<< on instruction is too slow.
1862 dbgs() << "GVN: load ";
1863 WriteAsOperand(dbgs(), L);
1864 Instruction *I = Dep.getInst();
1865 dbgs() << " is clobbered by " << *I << '\n';
1870 // If it is defined in another block, try harder.
1871 if (Dep.isNonLocal())
1872 return processNonLocalLoad(L);
1876 // fast print dep, using operator<< on instruction is too slow.
1877 dbgs() << "GVN: load ";
1878 WriteAsOperand(dbgs(), L);
1879 dbgs() << " has unknown dependence\n";
1884 Instruction *DepInst = Dep.getInst();
1885 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1886 Value *StoredVal = DepSI->getValueOperand();
1888 // The store and load are to a must-aliased pointer, but they may not
1889 // actually have the same type. See if we know how to reuse the stored
1890 // value (depending on its type).
1891 if (StoredVal->getType() != L->getType()) {
1893 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1898 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1899 << '\n' << *L << "\n\n\n");
1906 L->replaceAllUsesWith(StoredVal);
1907 if (StoredVal->getType()->getScalarType()->isPointerTy())
1908 MD->invalidateCachedPointerInfo(StoredVal);
1909 markInstructionForDeletion(L);
1914 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1915 Value *AvailableVal = DepLI;
1917 // The loads are of a must-aliased pointer, but they may not actually have
1918 // the same type. See if we know how to reuse the previously loaded value
1919 // (depending on its type).
1920 if (DepLI->getType() != L->getType()) {
1922 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(),
1924 if (AvailableVal == 0)
1927 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1928 << "\n" << *L << "\n\n\n");
1935 patchAndReplaceAllUsesWith(L, AvailableVal);
1936 if (DepLI->getType()->getScalarType()->isPointerTy())
1937 MD->invalidateCachedPointerInfo(DepLI);
1938 markInstructionForDeletion(L);
1943 // If this load really doesn't depend on anything, then we must be loading an
1944 // undef value. This can happen when loading for a fresh allocation with no
1945 // intervening stores, for example.
1946 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
1947 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1948 markInstructionForDeletion(L);
1953 // If this load occurs either right after a lifetime begin,
1954 // then the loaded value is undefined.
1955 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
1956 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1957 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1958 markInstructionForDeletion(L);
1967 // findLeader - In order to find a leader for a given value number at a
1968 // specific basic block, we first obtain the list of all Values for that number,
1969 // and then scan the list to find one whose block dominates the block in
1970 // question. This is fast because dominator tree queries consist of only
1971 // a few comparisons of DFS numbers.
1972 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1973 LeaderTableEntry Vals = LeaderTable[num];
1974 if (!Vals.Val) return 0;
1977 if (DT->dominates(Vals.BB, BB)) {
1979 if (isa<Constant>(Val)) return Val;
1982 LeaderTableEntry* Next = Vals.Next;
1984 if (DT->dominates(Next->BB, BB)) {
1985 if (isa<Constant>(Next->Val)) return Next->Val;
1986 if (!Val) Val = Next->Val;
1995 /// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the
1996 /// use is dominated by the given basic block. Returns the number of uses that
1998 unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To,
1999 const BasicBlockEdge &Root) {
2001 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
2003 Use &U = (UI++).getUse();
2005 if (DT->dominates(Root, U)) {
2013 /// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return
2014 /// true if every path from the entry block to 'Dst' passes via this edge. In
2015 /// particular 'Dst' must not be reachable via another edge from 'Src'.
2016 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
2017 DominatorTree *DT) {
2018 // While in theory it is interesting to consider the case in which Dst has
2019 // more than one predecessor, because Dst might be part of a loop which is
2020 // only reachable from Src, in practice it is pointless since at the time
2021 // GVN runs all such loops have preheaders, which means that Dst will have
2022 // been changed to have only one predecessor, namely Src.
2023 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
2024 const BasicBlock *Src = E.getStart();
2025 assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
2030 /// propagateEquality - The given values are known to be equal in every block
2031 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
2032 /// 'RHS' everywhere in the scope. Returns whether a change was made.
2033 bool GVN::propagateEquality(Value *LHS, Value *RHS,
2034 const BasicBlockEdge &Root) {
2035 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
2036 Worklist.push_back(std::make_pair(LHS, RHS));
2037 bool Changed = false;
2038 // For speed, compute a conservative fast approximation to
2039 // DT->dominates(Root, Root.getEnd());
2040 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
2042 while (!Worklist.empty()) {
2043 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
2044 LHS = Item.first; RHS = Item.second;
2046 if (LHS == RHS) continue;
2047 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
2049 // Don't try to propagate equalities between constants.
2050 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue;
2052 // Prefer a constant on the right-hand side, or an Argument if no constants.
2053 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
2054 std::swap(LHS, RHS);
2055 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
2057 // If there is no obvious reason to prefer the left-hand side over the right-
2058 // hand side, ensure the longest lived term is on the right-hand side, so the
2059 // shortest lived term will be replaced by the longest lived. This tends to
2060 // expose more simplifications.
2061 uint32_t LVN = VN.lookup_or_add(LHS);
2062 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
2063 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
2064 // Move the 'oldest' value to the right-hand side, using the value number as
2066 uint32_t RVN = VN.lookup_or_add(RHS);
2068 std::swap(LHS, RHS);
2073 // If value numbering later sees that an instruction in the scope is equal
2074 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
2075 // the invariant that instructions only occur in the leader table for their
2076 // own value number (this is used by removeFromLeaderTable), do not do this
2077 // if RHS is an instruction (if an instruction in the scope is morphed into
2078 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
2079 // using the leader table is about compiling faster, not optimizing better).
2080 // The leader table only tracks basic blocks, not edges. Only add to if we
2081 // have the simple case where the edge dominates the end.
2082 if (RootDominatesEnd && !isa<Instruction>(RHS))
2083 addToLeaderTable(LVN, RHS, Root.getEnd());
2085 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
2086 // LHS always has at least one use that is not dominated by Root, this will
2087 // never do anything if LHS has only one use.
2088 if (!LHS->hasOneUse()) {
2089 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root);
2090 Changed |= NumReplacements > 0;
2091 NumGVNEqProp += NumReplacements;
2094 // Now try to deduce additional equalities from this one. For example, if the
2095 // known equality was "(A != B)" == "false" then it follows that A and B are
2096 // equal in the scope. Only boolean equalities with an explicit true or false
2097 // RHS are currently supported.
2098 if (!RHS->getType()->isIntegerTy(1))
2099 // Not a boolean equality - bail out.
2101 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
2103 // RHS neither 'true' nor 'false' - bail out.
2105 // Whether RHS equals 'true'. Otherwise it equals 'false'.
2106 bool isKnownTrue = CI->isAllOnesValue();
2107 bool isKnownFalse = !isKnownTrue;
2109 // If "A && B" is known true then both A and B are known true. If "A || B"
2110 // is known false then both A and B are known false.
2112 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
2113 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
2114 Worklist.push_back(std::make_pair(A, RHS));
2115 Worklist.push_back(std::make_pair(B, RHS));
2119 // If we are propagating an equality like "(A == B)" == "true" then also
2120 // propagate the equality A == B. When propagating a comparison such as
2121 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
2122 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) {
2123 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
2125 // If "A == B" is known true, or "A != B" is known false, then replace
2126 // A with B everywhere in the scope.
2127 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2128 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2129 Worklist.push_back(std::make_pair(Op0, Op1));
2131 // If "A >= B" is known true, replace "A < B" with false everywhere.
2132 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2133 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2134 // Since we don't have the instruction "A < B" immediately to hand, work out
2135 // the value number that it would have and use that to find an appropriate
2136 // instruction (if any).
2137 uint32_t NextNum = VN.getNextUnusedValueNumber();
2138 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2139 // If the number we were assigned was brand new then there is no point in
2140 // looking for an instruction realizing it: there cannot be one!
2141 if (Num < NextNum) {
2142 Value *NotCmp = findLeader(Root.getEnd(), Num);
2143 if (NotCmp && isa<Instruction>(NotCmp)) {
2144 unsigned NumReplacements =
2145 replaceAllDominatedUsesWith(NotCmp, NotVal, Root);
2146 Changed |= NumReplacements > 0;
2147 NumGVNEqProp += NumReplacements;
2150 // Ensure that any instruction in scope that gets the "A < B" value number
2151 // is replaced with false.
2152 // The leader table only tracks basic blocks, not edges. Only add to if we
2153 // have the simple case where the edge dominates the end.
2154 if (RootDominatesEnd)
2155 addToLeaderTable(Num, NotVal, Root.getEnd());
2164 /// processInstruction - When calculating availability, handle an instruction
2165 /// by inserting it into the appropriate sets
2166 bool GVN::processInstruction(Instruction *I) {
2167 // Ignore dbg info intrinsics.
2168 if (isa<DbgInfoIntrinsic>(I))
2171 // If the instruction can be easily simplified then do so now in preference
2172 // to value numbering it. Value numbering often exposes redundancies, for
2173 // example if it determines that %y is equal to %x then the instruction
2174 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2175 if (Value *V = SimplifyInstruction(I, TD, TLI, DT)) {
2176 I->replaceAllUsesWith(V);
2177 if (MD && V->getType()->getScalarType()->isPointerTy())
2178 MD->invalidateCachedPointerInfo(V);
2179 markInstructionForDeletion(I);
2184 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2185 if (processLoad(LI))
2188 unsigned Num = VN.lookup_or_add(LI);
2189 addToLeaderTable(Num, LI, LI->getParent());
2193 // For conditional branches, we can perform simple conditional propagation on
2194 // the condition value itself.
2195 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2196 if (!BI->isConditional() || isa<Constant>(BI->getCondition()))
2199 Value *BranchCond = BI->getCondition();
2201 BasicBlock *TrueSucc = BI->getSuccessor(0);
2202 BasicBlock *FalseSucc = BI->getSuccessor(1);
2203 // Avoid multiple edges early.
2204 if (TrueSucc == FalseSucc)
2207 BasicBlock *Parent = BI->getParent();
2208 bool Changed = false;
2210 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
2211 BasicBlockEdge TrueE(Parent, TrueSucc);
2212 Changed |= propagateEquality(BranchCond, TrueVal, TrueE);
2214 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
2215 BasicBlockEdge FalseE(Parent, FalseSucc);
2216 Changed |= propagateEquality(BranchCond, FalseVal, FalseE);
2221 // For switches, propagate the case values into the case destinations.
2222 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2223 Value *SwitchCond = SI->getCondition();
2224 BasicBlock *Parent = SI->getParent();
2225 bool Changed = false;
2227 // Remember how many outgoing edges there are to every successor.
2228 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2229 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
2230 ++SwitchEdges[SI->getSuccessor(i)];
2232 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2234 BasicBlock *Dst = i.getCaseSuccessor();
2235 // If there is only a single edge, propagate the case value into it.
2236 if (SwitchEdges.lookup(Dst) == 1) {
2237 BasicBlockEdge E(Parent, Dst);
2238 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E);
2244 // Instructions with void type don't return a value, so there's
2245 // no point in trying to find redundancies in them.
2246 if (I->getType()->isVoidTy()) return false;
2248 uint32_t NextNum = VN.getNextUnusedValueNumber();
2249 unsigned Num = VN.lookup_or_add(I);
2251 // Allocations are always uniquely numbered, so we can save time and memory
2252 // by fast failing them.
2253 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2254 addToLeaderTable(Num, I, I->getParent());
2258 // If the number we were assigned was a brand new VN, then we don't
2259 // need to do a lookup to see if the number already exists
2260 // somewhere in the domtree: it can't!
2261 if (Num >= NextNum) {
2262 addToLeaderTable(Num, I, I->getParent());
2266 // Perform fast-path value-number based elimination of values inherited from
2268 Value *repl = findLeader(I->getParent(), Num);
2270 // Failure, just remember this instance for future use.
2271 addToLeaderTable(Num, I, I->getParent());
2276 patchAndReplaceAllUsesWith(I, repl);
2277 if (MD && repl->getType()->getScalarType()->isPointerTy())
2278 MD->invalidateCachedPointerInfo(repl);
2279 markInstructionForDeletion(I);
2283 /// runOnFunction - This is the main transformation entry point for a function.
2284 bool GVN::runOnFunction(Function& F) {
2286 MD = &getAnalysis<MemoryDependenceAnalysis>();
2287 DT = &getAnalysis<DominatorTree>();
2288 TD = getAnalysisIfAvailable<DataLayout>();
2289 TLI = &getAnalysis<TargetLibraryInfo>();
2290 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2294 bool Changed = false;
2295 bool ShouldContinue = true;
2297 // Merge unconditional branches, allowing PRE to catch more
2298 // optimization opportunities.
2299 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2300 BasicBlock *BB = FI++;
2302 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
2303 if (removedBlock) ++NumGVNBlocks;
2305 Changed |= removedBlock;
2308 unsigned Iteration = 0;
2309 while (ShouldContinue) {
2310 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2311 ShouldContinue = iterateOnFunction(F);
2312 Changed |= ShouldContinue;
2317 bool PREChanged = true;
2318 while (PREChanged) {
2319 PREChanged = performPRE(F);
2320 Changed |= PREChanged;
2324 // FIXME: Should perform GVN again after PRE does something. PRE can move
2325 // computations into blocks where they become fully redundant. Note that
2326 // we can't do this until PRE's critical edge splitting updates memdep.
2327 // Actually, when this happens, we should just fully integrate PRE into GVN.
2329 cleanupGlobalSets();
2335 bool GVN::processBlock(BasicBlock *BB) {
2336 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2337 // (and incrementing BI before processing an instruction).
2338 assert(InstrsToErase.empty() &&
2339 "We expect InstrsToErase to be empty across iterations");
2340 bool ChangedFunction = false;
2342 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2344 ChangedFunction |= processInstruction(BI);
2345 if (InstrsToErase.empty()) {
2350 // If we need some instructions deleted, do it now.
2351 NumGVNInstr += InstrsToErase.size();
2353 // Avoid iterator invalidation.
2354 bool AtStart = BI == BB->begin();
2358 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2359 E = InstrsToErase.end(); I != E; ++I) {
2360 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2361 if (MD) MD->removeInstruction(*I);
2362 DEBUG(verifyRemoved(*I));
2363 (*I)->eraseFromParent();
2365 InstrsToErase.clear();
2373 return ChangedFunction;
2376 /// performPRE - Perform a purely local form of PRE that looks for diamond
2377 /// control flow patterns and attempts to perform simple PRE at the join point.
2378 bool GVN::performPRE(Function &F) {
2379 bool Changed = false;
2380 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap;
2381 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
2382 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
2383 BasicBlock *CurrentBlock = *DI;
2385 // Nothing to PRE in the entry block.
2386 if (CurrentBlock == &F.getEntryBlock()) continue;
2388 // Don't perform PRE on a landing pad.
2389 if (CurrentBlock->isLandingPad()) continue;
2391 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2392 BE = CurrentBlock->end(); BI != BE; ) {
2393 Instruction *CurInst = BI++;
2395 if (isa<AllocaInst>(CurInst) ||
2396 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2397 CurInst->getType()->isVoidTy() ||
2398 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2399 isa<DbgInfoIntrinsic>(CurInst))
2402 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2403 // sinking the compare again, and it would force the code generator to
2404 // move the i1 from processor flags or predicate registers into a general
2405 // purpose register.
2406 if (isa<CmpInst>(CurInst))
2409 // We don't currently value number ANY inline asm calls.
2410 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2411 if (CallI->isInlineAsm())
2414 uint32_t ValNo = VN.lookup(CurInst);
2416 // Look for the predecessors for PRE opportunities. We're
2417 // only trying to solve the basic diamond case, where
2418 // a value is computed in the successor and one predecessor,
2419 // but not the other. We also explicitly disallow cases
2420 // where the successor is its own predecessor, because they're
2421 // more complicated to get right.
2422 unsigned NumWith = 0;
2423 unsigned NumWithout = 0;
2424 BasicBlock *PREPred = 0;
2427 for (pred_iterator PI = pred_begin(CurrentBlock),
2428 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2429 BasicBlock *P = *PI;
2430 // We're not interested in PRE where the block is its
2431 // own predecessor, or in blocks with predecessors
2432 // that are not reachable.
2433 if (P == CurrentBlock) {
2436 } else if (!DT->isReachableFromEntry(P)) {
2441 Value* predV = findLeader(P, ValNo);
2443 predMap.push_back(std::make_pair(static_cast<Value *>(0), P));
2446 } else if (predV == CurInst) {
2447 /* CurInst dominates this predecessor. */
2451 predMap.push_back(std::make_pair(predV, P));
2456 // Don't do PRE when it might increase code size, i.e. when
2457 // we would need to insert instructions in more than one pred.
2458 if (NumWithout != 1 || NumWith == 0)
2461 // Don't do PRE across indirect branch.
2462 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2465 // We can't do PRE safely on a critical edge, so instead we schedule
2466 // the edge to be split and perform the PRE the next time we iterate
2468 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2469 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2470 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2474 // Instantiate the expression in the predecessor that lacked it.
2475 // Because we are going top-down through the block, all value numbers
2476 // will be available in the predecessor by the time we need them. Any
2477 // that weren't originally present will have been instantiated earlier
2479 Instruction *PREInstr = CurInst->clone();
2480 bool success = true;
2481 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2482 Value *Op = PREInstr->getOperand(i);
2483 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2486 if (Value *V = findLeader(PREPred, VN.lookup(Op))) {
2487 PREInstr->setOperand(i, V);
2494 // Fail out if we encounter an operand that is not available in
2495 // the PRE predecessor. This is typically because of loads which
2496 // are not value numbered precisely.
2498 DEBUG(verifyRemoved(PREInstr));
2503 PREInstr->insertBefore(PREPred->getTerminator());
2504 PREInstr->setName(CurInst->getName() + ".pre");
2505 PREInstr->setDebugLoc(CurInst->getDebugLoc());
2506 VN.add(PREInstr, ValNo);
2509 // Update the availability map to include the new instruction.
2510 addToLeaderTable(ValNo, PREInstr, PREPred);
2512 // Create a PHI to make the value available in this block.
2513 PHINode* Phi = PHINode::Create(CurInst->getType(), predMap.size(),
2514 CurInst->getName() + ".pre-phi",
2515 CurrentBlock->begin());
2516 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2517 if (Value *V = predMap[i].first)
2518 Phi->addIncoming(V, predMap[i].second);
2520 Phi->addIncoming(PREInstr, PREPred);
2524 addToLeaderTable(ValNo, Phi, CurrentBlock);
2525 Phi->setDebugLoc(CurInst->getDebugLoc());
2526 CurInst->replaceAllUsesWith(Phi);
2527 if (Phi->getType()->getScalarType()->isPointerTy()) {
2528 // Because we have added a PHI-use of the pointer value, it has now
2529 // "escaped" from alias analysis' perspective. We need to inform
2531 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee;
2533 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
2534 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
2538 MD->invalidateCachedPointerInfo(Phi);
2541 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2543 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2544 if (MD) MD->removeInstruction(CurInst);
2545 DEBUG(verifyRemoved(CurInst));
2546 CurInst->eraseFromParent();
2551 if (splitCriticalEdges())
2557 /// Split the critical edge connecting the given two blocks, and return
2558 /// the block inserted to the critical edge.
2559 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2560 BasicBlock *BB = SplitCriticalEdge(Pred, Succ, this);
2562 MD->invalidateCachedPredecessors();
2566 /// splitCriticalEdges - Split critical edges found during the previous
2567 /// iteration that may enable further optimization.
2568 bool GVN::splitCriticalEdges() {
2569 if (toSplit.empty())
2572 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2573 SplitCriticalEdge(Edge.first, Edge.second, this);
2574 } while (!toSplit.empty());
2575 if (MD) MD->invalidateCachedPredecessors();
2579 /// iterateOnFunction - Executes one iteration of GVN
2580 bool GVN::iterateOnFunction(Function &F) {
2581 cleanupGlobalSets();
2583 // Top-down walk of the dominator tree
2584 bool Changed = false;
2586 // Needed for value numbering with phi construction to work.
2587 ReversePostOrderTraversal<Function*> RPOT(&F);
2588 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2589 RE = RPOT.end(); RI != RE; ++RI)
2590 Changed |= processBlock(*RI);
2592 // Save the blocks this function have before transformation begins. GVN may
2593 // split critical edge, and hence may invalidate the RPO/DT iterator.
2595 std::vector<BasicBlock *> BBVect;
2596 BBVect.reserve(256);
2597 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2598 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2599 BBVect.push_back(DI->getBlock());
2601 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
2603 Changed |= processBlock(*I);
2609 void GVN::cleanupGlobalSets() {
2611 LeaderTable.clear();
2612 TableAllocator.Reset();
2615 /// verifyRemoved - Verify that the specified instruction does not occur in our
2616 /// internal data structures.
2617 void GVN::verifyRemoved(const Instruction *Inst) const {
2618 VN.verifyRemoved(Inst);
2620 // Walk through the value number scope to make sure the instruction isn't
2621 // ferreted away in it.
2622 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2623 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2624 const LeaderTableEntry *Node = &I->second;
2625 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2627 while (Node->Next) {
2629 assert(Node->Val != Inst && "Inst still in value numbering scope!");