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/SetVector.h"
24 #include "llvm/ADT/SmallPtrSet.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/Analysis/AliasAnalysis.h"
27 #include "llvm/Analysis/CFG.h"
28 #include "llvm/Analysis/ConstantFolding.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/IR/DataLayout.h"
36 #include "llvm/IR/Dominators.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/IR/PatternMatch.h"
43 #include "llvm/Support/Allocator.h"
44 #include "llvm/Support/CommandLine.h"
45 #include "llvm/Support/Debug.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_intrinsic_expression(CallInst *C, uint32_t opcode,
116 Expression create_cmp_expression(unsigned Opcode,
117 CmpInst::Predicate Predicate,
118 Value *LHS, Value *RHS);
119 uint32_t lookup_or_add_call(CallInst* C);
121 ValueTable() : nextValueNumber(1) { }
122 uint32_t lookup_or_add(Value *V);
123 uint32_t lookup(Value *V) const;
124 uint32_t lookup_or_add_cmp(unsigned Opcode, CmpInst::Predicate Pred,
125 Value *LHS, Value *RHS);
126 void add(Value *V, uint32_t num);
128 void erase(Value *v);
129 void setAliasAnalysis(AliasAnalysis* A) { AA = A; }
130 AliasAnalysis *getAliasAnalysis() const { return AA; }
131 void setMemDep(MemoryDependenceAnalysis* M) { MD = M; }
132 void setDomTree(DominatorTree* D) { DT = D; }
133 uint32_t getNextUnusedValueNumber() { return nextValueNumber; }
134 void verifyRemoved(const Value *) const;
139 template <> struct DenseMapInfo<Expression> {
140 static inline Expression getEmptyKey() {
144 static inline Expression getTombstoneKey() {
148 static unsigned getHashValue(const Expression e) {
149 using llvm::hash_value;
150 return static_cast<unsigned>(hash_value(e));
152 static bool isEqual(const Expression &LHS, const Expression &RHS) {
159 //===----------------------------------------------------------------------===//
160 // ValueTable Internal Functions
161 //===----------------------------------------------------------------------===//
163 Expression ValueTable::create_expression(Instruction *I) {
165 e.type = I->getType();
166 e.opcode = I->getOpcode();
167 for (Instruction::op_iterator OI = I->op_begin(), OE = I->op_end();
169 e.varargs.push_back(lookup_or_add(*OI));
170 if (I->isCommutative()) {
171 // Ensure that commutative instructions that only differ by a permutation
172 // of their operands get the same value number by sorting the operand value
173 // numbers. Since all commutative instructions have two operands it is more
174 // efficient to sort by hand rather than using, say, std::sort.
175 assert(I->getNumOperands() == 2 && "Unsupported commutative instruction!");
176 if (e.varargs[0] > e.varargs[1])
177 std::swap(e.varargs[0], e.varargs[1]);
180 if (CmpInst *C = dyn_cast<CmpInst>(I)) {
181 // Sort the operand value numbers so x<y and y>x get the same value number.
182 CmpInst::Predicate Predicate = C->getPredicate();
183 if (e.varargs[0] > e.varargs[1]) {
184 std::swap(e.varargs[0], e.varargs[1]);
185 Predicate = CmpInst::getSwappedPredicate(Predicate);
187 e.opcode = (C->getOpcode() << 8) | Predicate;
188 } else if (InsertValueInst *E = dyn_cast<InsertValueInst>(I)) {
189 for (InsertValueInst::idx_iterator II = E->idx_begin(), IE = E->idx_end();
191 e.varargs.push_back(*II);
197 Expression ValueTable::create_intrinsic_expression(CallInst *C, uint32_t opcode,
198 bool IsCommutative) {
201 StructType *ST = cast<StructType>(C->getType());
203 e.type = *ST->element_begin();
205 for (unsigned i = 0, ei = C->getNumArgOperands(); i < ei; ++i)
206 e.varargs.push_back(lookup_or_add(C->getArgOperand(i)));
208 // Ensure that commutative instructions that only differ by a permutation
209 // of their operands get the same value number by sorting the operand value
210 // numbers. Since all commutative instructions have two operands it is more
211 // efficient to sort by hand rather than using, say, std::sort.
212 assert(C->getNumArgOperands() == 2 && "Unsupported commutative instruction!");
213 if (e.varargs[0] > e.varargs[1])
214 std::swap(e.varargs[0], e.varargs[1]);
220 Expression ValueTable::create_cmp_expression(unsigned Opcode,
221 CmpInst::Predicate Predicate,
222 Value *LHS, Value *RHS) {
223 assert((Opcode == Instruction::ICmp || Opcode == Instruction::FCmp) &&
224 "Not a comparison!");
226 e.type = CmpInst::makeCmpResultType(LHS->getType());
227 e.varargs.push_back(lookup_or_add(LHS));
228 e.varargs.push_back(lookup_or_add(RHS));
230 // Sort the operand value numbers so x<y and y>x get the same value number.
231 if (e.varargs[0] > e.varargs[1]) {
232 std::swap(e.varargs[0], e.varargs[1]);
233 Predicate = CmpInst::getSwappedPredicate(Predicate);
235 e.opcode = (Opcode << 8) | Predicate;
239 //===----------------------------------------------------------------------===//
240 // ValueTable External Functions
241 //===----------------------------------------------------------------------===//
243 /// add - Insert a value into the table with a specified value number.
244 void ValueTable::add(Value *V, uint32_t num) {
245 valueNumbering.insert(std::make_pair(V, num));
248 uint32_t ValueTable::lookup_or_add_call(CallInst *C) {
249 if (AA->doesNotAccessMemory(C)) {
250 Expression exp = create_expression(C);
251 uint32_t &e = expressionNumbering[exp];
252 if (!e) e = nextValueNumber++;
253 valueNumbering[C] = e;
255 } else if (AA->onlyReadsMemory(C)) {
256 Expression exp = create_expression(C);
257 uint32_t &e = expressionNumbering[exp];
259 e = nextValueNumber++;
260 valueNumbering[C] = e;
264 e = nextValueNumber++;
265 valueNumbering[C] = e;
269 MemDepResult local_dep = MD->getDependency(C);
271 if (!local_dep.isDef() && !local_dep.isNonLocal()) {
272 valueNumbering[C] = nextValueNumber;
273 return nextValueNumber++;
276 if (local_dep.isDef()) {
277 CallInst* local_cdep = cast<CallInst>(local_dep.getInst());
279 if (local_cdep->getNumArgOperands() != C->getNumArgOperands()) {
280 valueNumbering[C] = nextValueNumber;
281 return nextValueNumber++;
284 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
285 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
286 uint32_t cd_vn = lookup_or_add(local_cdep->getArgOperand(i));
288 valueNumbering[C] = nextValueNumber;
289 return nextValueNumber++;
293 uint32_t v = lookup_or_add(local_cdep);
294 valueNumbering[C] = v;
299 const MemoryDependenceAnalysis::NonLocalDepInfo &deps =
300 MD->getNonLocalCallDependency(CallSite(C));
301 // FIXME: Move the checking logic to MemDep!
304 // Check to see if we have a single dominating call instruction that is
306 for (unsigned i = 0, e = deps.size(); i != e; ++i) {
307 const NonLocalDepEntry *I = &deps[i];
308 if (I->getResult().isNonLocal())
311 // We don't handle non-definitions. If we already have a call, reject
312 // instruction dependencies.
313 if (!I->getResult().isDef() || cdep != 0) {
318 CallInst *NonLocalDepCall = dyn_cast<CallInst>(I->getResult().getInst());
319 // FIXME: All duplicated with non-local case.
320 if (NonLocalDepCall && DT->properlyDominates(I->getBB(), C->getParent())){
321 cdep = NonLocalDepCall;
330 valueNumbering[C] = nextValueNumber;
331 return nextValueNumber++;
334 if (cdep->getNumArgOperands() != C->getNumArgOperands()) {
335 valueNumbering[C] = nextValueNumber;
336 return nextValueNumber++;
338 for (unsigned i = 0, e = C->getNumArgOperands(); i < e; ++i) {
339 uint32_t c_vn = lookup_or_add(C->getArgOperand(i));
340 uint32_t cd_vn = lookup_or_add(cdep->getArgOperand(i));
342 valueNumbering[C] = nextValueNumber;
343 return nextValueNumber++;
347 uint32_t v = lookup_or_add(cdep);
348 valueNumbering[C] = v;
352 valueNumbering[C] = nextValueNumber;
353 return nextValueNumber++;
357 /// lookup_or_add - Returns the value number for the specified value, assigning
358 /// it a new number if it did not have one before.
359 uint32_t ValueTable::lookup_or_add(Value *V) {
360 DenseMap<Value*, uint32_t>::iterator VI = valueNumbering.find(V);
361 if (VI != valueNumbering.end())
364 if (!isa<Instruction>(V)) {
365 valueNumbering[V] = nextValueNumber;
366 return nextValueNumber++;
369 Instruction* I = cast<Instruction>(V);
371 switch (I->getOpcode()) {
372 case Instruction::Call: {
373 CallInst *C = cast<CallInst>(I);
374 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(C)) {
375 switch (II->getIntrinsicID()) {
376 case Intrinsic::sadd_with_overflow:
377 case Intrinsic::uadd_with_overflow:
378 exp = create_intrinsic_expression(C, Instruction::Add, true);
380 case Intrinsic::ssub_with_overflow:
381 case Intrinsic::usub_with_overflow:
382 exp = create_intrinsic_expression(C, Instruction::Sub, false);
384 case Intrinsic::smul_with_overflow:
385 case Intrinsic::umul_with_overflow:
386 exp = create_intrinsic_expression(C, Instruction::Mul, true);
389 return lookup_or_add_call(C);
392 return lookup_or_add_call(C);
395 case Instruction::Add:
396 case Instruction::FAdd:
397 case Instruction::Sub:
398 case Instruction::FSub:
399 case Instruction::Mul:
400 case Instruction::FMul:
401 case Instruction::UDiv:
402 case Instruction::SDiv:
403 case Instruction::FDiv:
404 case Instruction::URem:
405 case Instruction::SRem:
406 case Instruction::FRem:
407 case Instruction::Shl:
408 case Instruction::LShr:
409 case Instruction::AShr:
410 case Instruction::And:
411 case Instruction::Or:
412 case Instruction::Xor:
413 case Instruction::ICmp:
414 case Instruction::FCmp:
415 case Instruction::Trunc:
416 case Instruction::ZExt:
417 case Instruction::SExt:
418 case Instruction::FPToUI:
419 case Instruction::FPToSI:
420 case Instruction::UIToFP:
421 case Instruction::SIToFP:
422 case Instruction::FPTrunc:
423 case Instruction::FPExt:
424 case Instruction::PtrToInt:
425 case Instruction::IntToPtr:
426 case Instruction::BitCast:
427 case Instruction::Select:
428 case Instruction::ExtractElement:
429 case Instruction::InsertElement:
430 case Instruction::ShuffleVector:
431 case Instruction::InsertValue:
432 case Instruction::GetElementPtr:
433 case Instruction::ExtractValue:
434 exp = create_expression(I);
437 valueNumbering[V] = nextValueNumber;
438 return nextValueNumber++;
441 uint32_t& e = expressionNumbering[exp];
442 if (!e) e = nextValueNumber++;
443 valueNumbering[V] = e;
447 /// lookup - Returns the value number of the specified value. Fails if
448 /// the value has not yet been numbered.
449 uint32_t ValueTable::lookup(Value *V) const {
450 DenseMap<Value*, uint32_t>::const_iterator VI = valueNumbering.find(V);
451 assert(VI != valueNumbering.end() && "Value not numbered?");
455 /// lookup_or_add_cmp - Returns the value number of the given comparison,
456 /// assigning it a new number if it did not have one before. Useful when
457 /// we deduced the result of a comparison, but don't immediately have an
458 /// instruction realizing that comparison to hand.
459 uint32_t ValueTable::lookup_or_add_cmp(unsigned Opcode,
460 CmpInst::Predicate Predicate,
461 Value *LHS, Value *RHS) {
462 Expression exp = create_cmp_expression(Opcode, Predicate, LHS, RHS);
463 uint32_t& e = expressionNumbering[exp];
464 if (!e) e = nextValueNumber++;
468 /// clear - Remove all entries from the ValueTable.
469 void ValueTable::clear() {
470 valueNumbering.clear();
471 expressionNumbering.clear();
475 /// erase - Remove a value from the value numbering.
476 void ValueTable::erase(Value *V) {
477 valueNumbering.erase(V);
480 /// verifyRemoved - Verify that the value is removed from all internal data
482 void ValueTable::verifyRemoved(const Value *V) const {
483 for (DenseMap<Value*, uint32_t>::const_iterator
484 I = valueNumbering.begin(), E = valueNumbering.end(); I != E; ++I) {
485 assert(I->first != V && "Inst still occurs in value numbering map!");
489 //===----------------------------------------------------------------------===//
491 //===----------------------------------------------------------------------===//
495 struct AvailableValueInBlock {
496 /// BB - The basic block in question.
499 SimpleVal, // A simple offsetted value that is accessed.
500 LoadVal, // A value produced by a load.
501 MemIntrin, // A memory intrinsic which is loaded from.
502 UndefVal // A UndefValue representing a value from dead block (which
503 // is not yet physically removed from the CFG).
506 /// V - The value that is live out of the block.
507 PointerIntPair<Value *, 2, ValType> Val;
509 /// Offset - The byte offset in Val that is interesting for the load query.
512 static AvailableValueInBlock get(BasicBlock *BB, Value *V,
513 unsigned Offset = 0) {
514 AvailableValueInBlock Res;
516 Res.Val.setPointer(V);
517 Res.Val.setInt(SimpleVal);
522 static AvailableValueInBlock getMI(BasicBlock *BB, MemIntrinsic *MI,
523 unsigned Offset = 0) {
524 AvailableValueInBlock Res;
526 Res.Val.setPointer(MI);
527 Res.Val.setInt(MemIntrin);
532 static AvailableValueInBlock getLoad(BasicBlock *BB, LoadInst *LI,
533 unsigned Offset = 0) {
534 AvailableValueInBlock Res;
536 Res.Val.setPointer(LI);
537 Res.Val.setInt(LoadVal);
542 static AvailableValueInBlock getUndef(BasicBlock *BB) {
543 AvailableValueInBlock Res;
545 Res.Val.setPointer(0);
546 Res.Val.setInt(UndefVal);
551 bool isSimpleValue() const { return Val.getInt() == SimpleVal; }
552 bool isCoercedLoadValue() const { return Val.getInt() == LoadVal; }
553 bool isMemIntrinValue() const { return Val.getInt() == MemIntrin; }
554 bool isUndefValue() const { return Val.getInt() == UndefVal; }
556 Value *getSimpleValue() const {
557 assert(isSimpleValue() && "Wrong accessor");
558 return Val.getPointer();
561 LoadInst *getCoercedLoadValue() const {
562 assert(isCoercedLoadValue() && "Wrong accessor");
563 return cast<LoadInst>(Val.getPointer());
566 MemIntrinsic *getMemIntrinValue() const {
567 assert(isMemIntrinValue() && "Wrong accessor");
568 return cast<MemIntrinsic>(Val.getPointer());
571 /// MaterializeAdjustedValue - Emit code into this block to adjust the value
572 /// defined here to the specified type. This handles various coercion cases.
573 Value *MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const;
576 class GVN : public FunctionPass {
578 MemoryDependenceAnalysis *MD;
580 const DataLayout *DL;
581 const TargetLibraryInfo *TLI;
582 SetVector<BasicBlock *> DeadBlocks;
586 /// LeaderTable - A mapping from value numbers to lists of Value*'s that
587 /// have that value number. Use findLeader to query it.
588 struct LeaderTableEntry {
590 const BasicBlock *BB;
591 LeaderTableEntry *Next;
593 DenseMap<uint32_t, LeaderTableEntry> LeaderTable;
594 BumpPtrAllocator TableAllocator;
596 SmallVector<Instruction*, 8> InstrsToErase;
598 typedef SmallVector<NonLocalDepResult, 64> LoadDepVect;
599 typedef SmallVector<AvailableValueInBlock, 64> AvailValInBlkVect;
600 typedef SmallVector<BasicBlock*, 64> UnavailBlkVect;
603 static char ID; // Pass identification, replacement for typeid
604 explicit GVN(bool noloads = false)
605 : FunctionPass(ID), NoLoads(noloads), MD(0) {
606 initializeGVNPass(*PassRegistry::getPassRegistry());
609 bool runOnFunction(Function &F) override;
611 /// markInstructionForDeletion - This removes the specified instruction from
612 /// our various maps and marks it for deletion.
613 void markInstructionForDeletion(Instruction *I) {
615 InstrsToErase.push_back(I);
618 const DataLayout *getDataLayout() const { return DL; }
619 DominatorTree &getDominatorTree() const { return *DT; }
620 AliasAnalysis *getAliasAnalysis() const { return VN.getAliasAnalysis(); }
621 MemoryDependenceAnalysis &getMemDep() const { return *MD; }
623 /// addToLeaderTable - Push a new Value to the LeaderTable onto the list for
624 /// its value number.
625 void addToLeaderTable(uint32_t N, Value *V, const BasicBlock *BB) {
626 LeaderTableEntry &Curr = LeaderTable[N];
633 LeaderTableEntry *Node = TableAllocator.Allocate<LeaderTableEntry>();
636 Node->Next = Curr.Next;
640 /// removeFromLeaderTable - Scan the list of values corresponding to a given
641 /// value number, and remove the given instruction if encountered.
642 void removeFromLeaderTable(uint32_t N, Instruction *I, BasicBlock *BB) {
643 LeaderTableEntry* Prev = 0;
644 LeaderTableEntry* Curr = &LeaderTable[N];
646 while (Curr->Val != I || Curr->BB != BB) {
652 Prev->Next = Curr->Next;
658 LeaderTableEntry* Next = Curr->Next;
659 Curr->Val = Next->Val;
661 Curr->Next = Next->Next;
666 // List of critical edges to be split between iterations.
667 SmallVector<std::pair<TerminatorInst*, unsigned>, 4> toSplit;
669 // This transformation requires dominator postdominator info
670 void getAnalysisUsage(AnalysisUsage &AU) const override {
671 AU.addRequired<DominatorTreeWrapperPass>();
672 AU.addRequired<TargetLibraryInfo>();
674 AU.addRequired<MemoryDependenceAnalysis>();
675 AU.addRequired<AliasAnalysis>();
677 AU.addPreserved<DominatorTreeWrapperPass>();
678 AU.addPreserved<AliasAnalysis>();
682 // Helper fuctions of redundant load elimination
683 bool processLoad(LoadInst *L);
684 bool processNonLocalLoad(LoadInst *L);
685 void AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
686 AvailValInBlkVect &ValuesPerBlock,
687 UnavailBlkVect &UnavailableBlocks);
688 bool PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
689 UnavailBlkVect &UnavailableBlocks);
691 // Other helper routines
692 bool processInstruction(Instruction *I);
693 bool processBlock(BasicBlock *BB);
694 void dump(DenseMap<uint32_t, Value*> &d);
695 bool iterateOnFunction(Function &F);
696 bool performPRE(Function &F);
697 Value *findLeader(const BasicBlock *BB, uint32_t num);
698 void cleanupGlobalSets();
699 void verifyRemoved(const Instruction *I) const;
700 bool splitCriticalEdges();
701 BasicBlock *splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ);
702 unsigned replaceAllDominatedUsesWith(Value *From, Value *To,
703 const BasicBlockEdge &Root);
704 bool propagateEquality(Value *LHS, Value *RHS, const BasicBlockEdge &Root);
705 bool processFoldableCondBr(BranchInst *BI);
706 void addDeadBlock(BasicBlock *BB);
707 void assignValNumForDeadCode();
713 // createGVNPass - The public interface to this file...
714 FunctionPass *llvm::createGVNPass(bool NoLoads) {
715 return new GVN(NoLoads);
718 INITIALIZE_PASS_BEGIN(GVN, "gvn", "Global Value Numbering", false, false)
719 INITIALIZE_PASS_DEPENDENCY(MemoryDependenceAnalysis)
720 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
721 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfo)
722 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
723 INITIALIZE_PASS_END(GVN, "gvn", "Global Value Numbering", false, false)
725 #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP)
726 void GVN::dump(DenseMap<uint32_t, Value*>& d) {
728 for (DenseMap<uint32_t, Value*>::iterator I = d.begin(),
729 E = d.end(); I != E; ++I) {
730 errs() << I->first << "\n";
737 /// IsValueFullyAvailableInBlock - Return true if we can prove that the value
738 /// we're analyzing is fully available in the specified block. As we go, keep
739 /// track of which blocks we know are fully alive in FullyAvailableBlocks. This
740 /// map is actually a tri-state map with the following values:
741 /// 0) we know the block *is not* fully available.
742 /// 1) we know the block *is* fully available.
743 /// 2) we do not know whether the block is fully available or not, but we are
744 /// currently speculating that it will be.
745 /// 3) we are speculating for this block and have used that to speculate for
747 static bool IsValueFullyAvailableInBlock(BasicBlock *BB,
748 DenseMap<BasicBlock*, char> &FullyAvailableBlocks,
749 uint32_t RecurseDepth) {
750 if (RecurseDepth > MaxRecurseDepth)
753 // Optimistically assume that the block is fully available and check to see
754 // if we already know about this block in one lookup.
755 std::pair<DenseMap<BasicBlock*, char>::iterator, char> IV =
756 FullyAvailableBlocks.insert(std::make_pair(BB, 2));
758 // If the entry already existed for this block, return the precomputed value.
760 // If this is a speculative "available" value, mark it as being used for
761 // speculation of other blocks.
762 if (IV.first->second == 2)
763 IV.first->second = 3;
764 return IV.first->second != 0;
767 // Otherwise, see if it is fully available in all predecessors.
768 pred_iterator PI = pred_begin(BB), PE = pred_end(BB);
770 // If this block has no predecessors, it isn't live-in here.
772 goto SpeculationFailure;
774 for (; PI != PE; ++PI)
775 // If the value isn't fully available in one of our predecessors, then it
776 // isn't fully available in this block either. Undo our previous
777 // optimistic assumption and bail out.
778 if (!IsValueFullyAvailableInBlock(*PI, FullyAvailableBlocks,RecurseDepth+1))
779 goto SpeculationFailure;
783 // SpeculationFailure - If we get here, we found out that this is not, after
784 // all, a fully-available block. We have a problem if we speculated on this and
785 // used the speculation to mark other blocks as available.
787 char &BBVal = FullyAvailableBlocks[BB];
789 // If we didn't speculate on this, just return with it set to false.
795 // If we did speculate on this value, we could have blocks set to 1 that are
796 // incorrect. Walk the (transitive) successors of this block and mark them as
798 SmallVector<BasicBlock*, 32> BBWorklist;
799 BBWorklist.push_back(BB);
802 BasicBlock *Entry = BBWorklist.pop_back_val();
803 // Note that this sets blocks to 0 (unavailable) if they happen to not
804 // already be in FullyAvailableBlocks. This is safe.
805 char &EntryVal = FullyAvailableBlocks[Entry];
806 if (EntryVal == 0) continue; // Already unavailable.
808 // Mark as unavailable.
811 BBWorklist.append(succ_begin(Entry), succ_end(Entry));
812 } while (!BBWorklist.empty());
818 /// CanCoerceMustAliasedValueToLoad - Return true if
819 /// CoerceAvailableValueToLoadType will succeed.
820 static bool CanCoerceMustAliasedValueToLoad(Value *StoredVal,
822 const DataLayout &DL) {
823 // If the loaded or stored value is an first class array or struct, don't try
824 // to transform them. We need to be able to bitcast to integer.
825 if (LoadTy->isStructTy() || LoadTy->isArrayTy() ||
826 StoredVal->getType()->isStructTy() ||
827 StoredVal->getType()->isArrayTy())
830 // The store has to be at least as big as the load.
831 if (DL.getTypeSizeInBits(StoredVal->getType()) <
832 DL.getTypeSizeInBits(LoadTy))
838 /// CoerceAvailableValueToLoadType - If we saw a store of a value to memory, and
839 /// then a load from a must-aliased pointer of a different type, try to coerce
840 /// the stored value. LoadedTy is the type of the load we want to replace and
841 /// InsertPt is the place to insert new instructions.
843 /// If we can't do it, return null.
844 static Value *CoerceAvailableValueToLoadType(Value *StoredVal,
846 Instruction *InsertPt,
847 const DataLayout &DL) {
848 if (!CanCoerceMustAliasedValueToLoad(StoredVal, LoadedTy, DL))
851 // If this is already the right type, just return it.
852 Type *StoredValTy = StoredVal->getType();
854 uint64_t StoreSize = DL.getTypeSizeInBits(StoredValTy);
855 uint64_t LoadSize = DL.getTypeSizeInBits(LoadedTy);
857 // If the store and reload are the same size, we can always reuse it.
858 if (StoreSize == LoadSize) {
859 // Pointer to Pointer -> use bitcast.
860 if (StoredValTy->getScalarType()->isPointerTy() &&
861 LoadedTy->getScalarType()->isPointerTy())
862 return new BitCastInst(StoredVal, LoadedTy, "", InsertPt);
864 // Convert source pointers to integers, which can be bitcast.
865 if (StoredValTy->getScalarType()->isPointerTy()) {
866 StoredValTy = DL.getIntPtrType(StoredValTy);
867 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
870 Type *TypeToCastTo = LoadedTy;
871 if (TypeToCastTo->getScalarType()->isPointerTy())
872 TypeToCastTo = DL.getIntPtrType(TypeToCastTo);
874 if (StoredValTy != TypeToCastTo)
875 StoredVal = new BitCastInst(StoredVal, TypeToCastTo, "", InsertPt);
877 // Cast to pointer if the load needs a pointer type.
878 if (LoadedTy->getScalarType()->isPointerTy())
879 StoredVal = new IntToPtrInst(StoredVal, LoadedTy, "", InsertPt);
884 // If the loaded value is smaller than the available value, then we can
885 // extract out a piece from it. If the available value is too small, then we
886 // can't do anything.
887 assert(StoreSize >= LoadSize && "CanCoerceMustAliasedValueToLoad fail");
889 // Convert source pointers to integers, which can be manipulated.
890 if (StoredValTy->getScalarType()->isPointerTy()) {
891 StoredValTy = DL.getIntPtrType(StoredValTy);
892 StoredVal = new PtrToIntInst(StoredVal, StoredValTy, "", InsertPt);
895 // Convert vectors and fp to integer, which can be manipulated.
896 if (!StoredValTy->isIntegerTy()) {
897 StoredValTy = IntegerType::get(StoredValTy->getContext(), StoreSize);
898 StoredVal = new BitCastInst(StoredVal, StoredValTy, "", InsertPt);
901 // If this is a big-endian system, we need to shift the value down to the low
902 // bits so that a truncate will work.
903 if (DL.isBigEndian()) {
904 Constant *Val = ConstantInt::get(StoredVal->getType(), StoreSize-LoadSize);
905 StoredVal = BinaryOperator::CreateLShr(StoredVal, Val, "tmp", InsertPt);
908 // Truncate the integer to the right size now.
909 Type *NewIntTy = IntegerType::get(StoredValTy->getContext(), LoadSize);
910 StoredVal = new TruncInst(StoredVal, NewIntTy, "trunc", InsertPt);
912 if (LoadedTy == NewIntTy)
915 // If the result is a pointer, inttoptr.
916 if (LoadedTy->getScalarType()->isPointerTy())
917 return new IntToPtrInst(StoredVal, LoadedTy, "inttoptr", InsertPt);
919 // Otherwise, bitcast.
920 return new BitCastInst(StoredVal, LoadedTy, "bitcast", InsertPt);
923 /// AnalyzeLoadFromClobberingWrite - This function is called when we have a
924 /// memdep query of a load that ends up being a clobbering memory write (store,
925 /// memset, memcpy, memmove). This means that the write *may* provide bits used
926 /// by the load but we can't be sure because the pointers don't mustalias.
928 /// Check this case to see if there is anything more we can do before we give
929 /// up. This returns -1 if we have to give up, or a byte number in the stored
930 /// value of the piece that feeds the load.
931 static int AnalyzeLoadFromClobberingWrite(Type *LoadTy, Value *LoadPtr,
933 uint64_t WriteSizeInBits,
934 const DataLayout &DL) {
935 // If the loaded or stored value is a first class array or struct, don't try
936 // to transform them. We need to be able to bitcast to integer.
937 if (LoadTy->isStructTy() || LoadTy->isArrayTy())
940 int64_t StoreOffset = 0, LoadOffset = 0;
941 Value *StoreBase = GetPointerBaseWithConstantOffset(WritePtr,StoreOffset,&DL);
942 Value *LoadBase = GetPointerBaseWithConstantOffset(LoadPtr, LoadOffset, &DL);
943 if (StoreBase != LoadBase)
946 // If the load and store are to the exact same address, they should have been
947 // a must alias. AA must have gotten confused.
948 // FIXME: Study to see if/when this happens. One case is forwarding a memset
949 // to a load from the base of the memset.
951 if (LoadOffset == StoreOffset) {
952 dbgs() << "STORE/LOAD DEP WITH COMMON POINTER MISSED:\n"
953 << "Base = " << *StoreBase << "\n"
954 << "Store Ptr = " << *WritePtr << "\n"
955 << "Store Offs = " << StoreOffset << "\n"
956 << "Load Ptr = " << *LoadPtr << "\n";
961 // If the load and store don't overlap at all, the store doesn't provide
962 // anything to the load. In this case, they really don't alias at all, AA
963 // must have gotten confused.
964 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy);
966 if ((WriteSizeInBits & 7) | (LoadSize & 7))
968 uint64_t StoreSize = WriteSizeInBits >> 3; // Convert to bytes.
972 bool isAAFailure = false;
973 if (StoreOffset < LoadOffset)
974 isAAFailure = StoreOffset+int64_t(StoreSize) <= LoadOffset;
976 isAAFailure = LoadOffset+int64_t(LoadSize) <= StoreOffset;
980 dbgs() << "STORE LOAD DEP WITH COMMON BASE:\n"
981 << "Base = " << *StoreBase << "\n"
982 << "Store Ptr = " << *WritePtr << "\n"
983 << "Store Offs = " << StoreOffset << "\n"
984 << "Load Ptr = " << *LoadPtr << "\n";
990 // If the Load isn't completely contained within the stored bits, we don't
991 // have all the bits to feed it. We could do something crazy in the future
992 // (issue a smaller load then merge the bits in) but this seems unlikely to be
994 if (StoreOffset > LoadOffset ||
995 StoreOffset+StoreSize < LoadOffset+LoadSize)
998 // Okay, we can do this transformation. Return the number of bytes into the
999 // store that the load is.
1000 return LoadOffset-StoreOffset;
1003 /// AnalyzeLoadFromClobberingStore - This function is called when we have a
1004 /// memdep query of a load that ends up being a clobbering store.
1005 static int AnalyzeLoadFromClobberingStore(Type *LoadTy, Value *LoadPtr,
1007 const DataLayout &DL) {
1008 // Cannot handle reading from store of first-class aggregate yet.
1009 if (DepSI->getValueOperand()->getType()->isStructTy() ||
1010 DepSI->getValueOperand()->getType()->isArrayTy())
1013 Value *StorePtr = DepSI->getPointerOperand();
1014 uint64_t StoreSize =DL.getTypeSizeInBits(DepSI->getValueOperand()->getType());
1015 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1016 StorePtr, StoreSize, DL);
1019 /// AnalyzeLoadFromClobberingLoad - This function is called when we have a
1020 /// memdep query of a load that ends up being clobbered by another load. See if
1021 /// the other load can feed into the second load.
1022 static int AnalyzeLoadFromClobberingLoad(Type *LoadTy, Value *LoadPtr,
1023 LoadInst *DepLI, const DataLayout &DL){
1024 // Cannot handle reading from store of first-class aggregate yet.
1025 if (DepLI->getType()->isStructTy() || DepLI->getType()->isArrayTy())
1028 Value *DepPtr = DepLI->getPointerOperand();
1029 uint64_t DepSize = DL.getTypeSizeInBits(DepLI->getType());
1030 int R = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, DepSize, DL);
1031 if (R != -1) return R;
1033 // If we have a load/load clobber an DepLI can be widened to cover this load,
1034 // then we should widen it!
1035 int64_t LoadOffs = 0;
1036 const Value *LoadBase =
1037 GetPointerBaseWithConstantOffset(LoadPtr, LoadOffs, &DL);
1038 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1040 unsigned Size = MemoryDependenceAnalysis::
1041 getLoadLoadClobberFullWidthSize(LoadBase, LoadOffs, LoadSize, DepLI, DL);
1042 if (Size == 0) return -1;
1044 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, DepPtr, Size*8, DL);
1049 static int AnalyzeLoadFromClobberingMemInst(Type *LoadTy, Value *LoadPtr,
1051 const DataLayout &DL) {
1052 // If the mem operation is a non-constant size, we can't handle it.
1053 ConstantInt *SizeCst = dyn_cast<ConstantInt>(MI->getLength());
1054 if (SizeCst == 0) return -1;
1055 uint64_t MemSizeInBits = SizeCst->getZExtValue()*8;
1057 // If this is memset, we just need to see if the offset is valid in the size
1059 if (MI->getIntrinsicID() == Intrinsic::memset)
1060 return AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr, MI->getDest(),
1063 // If we have a memcpy/memmove, the only case we can handle is if this is a
1064 // copy from constant memory. In that case, we can read directly from the
1066 MemTransferInst *MTI = cast<MemTransferInst>(MI);
1068 Constant *Src = dyn_cast<Constant>(MTI->getSource());
1069 if (Src == 0) return -1;
1071 GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Src, &DL));
1072 if (GV == 0 || !GV->isConstant()) return -1;
1074 // See if the access is within the bounds of the transfer.
1075 int Offset = AnalyzeLoadFromClobberingWrite(LoadTy, LoadPtr,
1076 MI->getDest(), MemSizeInBits, DL);
1080 unsigned AS = Src->getType()->getPointerAddressSpace();
1081 // Otherwise, see if we can constant fold a load from the constant with the
1082 // offset applied as appropriate.
1083 Src = ConstantExpr::getBitCast(Src,
1084 Type::getInt8PtrTy(Src->getContext(), AS));
1085 Constant *OffsetCst =
1086 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1087 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1088 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1089 if (ConstantFoldLoadFromConstPtr(Src, &DL))
1095 /// GetStoreValueForLoad - This function is called when we have a
1096 /// memdep query of a load that ends up being a clobbering store. This means
1097 /// that the store provides bits used by the load but we the pointers don't
1098 /// mustalias. Check this case to see if there is anything more we can do
1099 /// before we give up.
1100 static Value *GetStoreValueForLoad(Value *SrcVal, unsigned Offset,
1102 Instruction *InsertPt, const DataLayout &DL){
1103 LLVMContext &Ctx = SrcVal->getType()->getContext();
1105 uint64_t StoreSize = (DL.getTypeSizeInBits(SrcVal->getType()) + 7) / 8;
1106 uint64_t LoadSize = (DL.getTypeSizeInBits(LoadTy) + 7) / 8;
1108 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1110 // Compute which bits of the stored value are being used by the load. Convert
1111 // to an integer type to start with.
1112 if (SrcVal->getType()->getScalarType()->isPointerTy())
1113 SrcVal = Builder.CreatePtrToInt(SrcVal,
1114 DL.getIntPtrType(SrcVal->getType()));
1115 if (!SrcVal->getType()->isIntegerTy())
1116 SrcVal = Builder.CreateBitCast(SrcVal, IntegerType::get(Ctx, StoreSize*8));
1118 // Shift the bits to the least significant depending on endianness.
1120 if (DL.isLittleEndian())
1121 ShiftAmt = Offset*8;
1123 ShiftAmt = (StoreSize-LoadSize-Offset)*8;
1126 SrcVal = Builder.CreateLShr(SrcVal, ShiftAmt);
1128 if (LoadSize != StoreSize)
1129 SrcVal = Builder.CreateTrunc(SrcVal, IntegerType::get(Ctx, LoadSize*8));
1131 return CoerceAvailableValueToLoadType(SrcVal, LoadTy, InsertPt, DL);
1134 /// GetLoadValueForLoad - This function is called when we have a
1135 /// memdep query of a load that ends up being a clobbering load. This means
1136 /// that the load *may* provide bits used by the load but we can't be sure
1137 /// because the pointers don't mustalias. Check this case to see if there is
1138 /// anything more we can do before we give up.
1139 static Value *GetLoadValueForLoad(LoadInst *SrcVal, unsigned Offset,
1140 Type *LoadTy, Instruction *InsertPt,
1142 const DataLayout &DL = *gvn.getDataLayout();
1143 // If Offset+LoadTy exceeds the size of SrcVal, then we must be wanting to
1144 // widen SrcVal out to a larger load.
1145 unsigned SrcValSize = DL.getTypeStoreSize(SrcVal->getType());
1146 unsigned LoadSize = DL.getTypeStoreSize(LoadTy);
1147 if (Offset+LoadSize > SrcValSize) {
1148 assert(SrcVal->isSimple() && "Cannot widen volatile/atomic load!");
1149 assert(SrcVal->getType()->isIntegerTy() && "Can't widen non-integer load");
1150 // If we have a load/load clobber an DepLI can be widened to cover this
1151 // load, then we should widen it to the next power of 2 size big enough!
1152 unsigned NewLoadSize = Offset+LoadSize;
1153 if (!isPowerOf2_32(NewLoadSize))
1154 NewLoadSize = NextPowerOf2(NewLoadSize);
1156 Value *PtrVal = SrcVal->getPointerOperand();
1158 // Insert the new load after the old load. This ensures that subsequent
1159 // memdep queries will find the new load. We can't easily remove the old
1160 // load completely because it is already in the value numbering table.
1161 IRBuilder<> Builder(SrcVal->getParent(), ++BasicBlock::iterator(SrcVal));
1163 IntegerType::get(LoadTy->getContext(), NewLoadSize*8);
1164 DestPTy = PointerType::get(DestPTy,
1165 PtrVal->getType()->getPointerAddressSpace());
1166 Builder.SetCurrentDebugLocation(SrcVal->getDebugLoc());
1167 PtrVal = Builder.CreateBitCast(PtrVal, DestPTy);
1168 LoadInst *NewLoad = Builder.CreateLoad(PtrVal);
1169 NewLoad->takeName(SrcVal);
1170 NewLoad->setAlignment(SrcVal->getAlignment());
1172 DEBUG(dbgs() << "GVN WIDENED LOAD: " << *SrcVal << "\n");
1173 DEBUG(dbgs() << "TO: " << *NewLoad << "\n");
1175 // Replace uses of the original load with the wider load. On a big endian
1176 // system, we need to shift down to get the relevant bits.
1177 Value *RV = NewLoad;
1178 if (DL.isBigEndian())
1179 RV = Builder.CreateLShr(RV,
1180 NewLoadSize*8-SrcVal->getType()->getPrimitiveSizeInBits());
1181 RV = Builder.CreateTrunc(RV, SrcVal->getType());
1182 SrcVal->replaceAllUsesWith(RV);
1184 // We would like to use gvn.markInstructionForDeletion here, but we can't
1185 // because the load is already memoized into the leader map table that GVN
1186 // tracks. It is potentially possible to remove the load from the table,
1187 // but then there all of the operations based on it would need to be
1188 // rehashed. Just leave the dead load around.
1189 gvn.getMemDep().removeInstruction(SrcVal);
1193 return GetStoreValueForLoad(SrcVal, Offset, LoadTy, InsertPt, DL);
1197 /// GetMemInstValueForLoad - This function is called when we have a
1198 /// memdep query of a load that ends up being a clobbering mem intrinsic.
1199 static Value *GetMemInstValueForLoad(MemIntrinsic *SrcInst, unsigned Offset,
1200 Type *LoadTy, Instruction *InsertPt,
1201 const DataLayout &DL){
1202 LLVMContext &Ctx = LoadTy->getContext();
1203 uint64_t LoadSize = DL.getTypeSizeInBits(LoadTy)/8;
1205 IRBuilder<> Builder(InsertPt->getParent(), InsertPt);
1207 // We know that this method is only called when the mem transfer fully
1208 // provides the bits for the load.
1209 if (MemSetInst *MSI = dyn_cast<MemSetInst>(SrcInst)) {
1210 // memset(P, 'x', 1234) -> splat('x'), even if x is a variable, and
1211 // independently of what the offset is.
1212 Value *Val = MSI->getValue();
1214 Val = Builder.CreateZExt(Val, IntegerType::get(Ctx, LoadSize*8));
1216 Value *OneElt = Val;
1218 // Splat the value out to the right number of bits.
1219 for (unsigned NumBytesSet = 1; NumBytesSet != LoadSize; ) {
1220 // If we can double the number of bytes set, do it.
1221 if (NumBytesSet*2 <= LoadSize) {
1222 Value *ShVal = Builder.CreateShl(Val, NumBytesSet*8);
1223 Val = Builder.CreateOr(Val, ShVal);
1228 // Otherwise insert one byte at a time.
1229 Value *ShVal = Builder.CreateShl(Val, 1*8);
1230 Val = Builder.CreateOr(OneElt, ShVal);
1234 return CoerceAvailableValueToLoadType(Val, LoadTy, InsertPt, DL);
1237 // Otherwise, this is a memcpy/memmove from a constant global.
1238 MemTransferInst *MTI = cast<MemTransferInst>(SrcInst);
1239 Constant *Src = cast<Constant>(MTI->getSource());
1240 unsigned AS = Src->getType()->getPointerAddressSpace();
1242 // Otherwise, see if we can constant fold a load from the constant with the
1243 // offset applied as appropriate.
1244 Src = ConstantExpr::getBitCast(Src,
1245 Type::getInt8PtrTy(Src->getContext(), AS));
1246 Constant *OffsetCst =
1247 ConstantInt::get(Type::getInt64Ty(Src->getContext()), (unsigned)Offset);
1248 Src = ConstantExpr::getGetElementPtr(Src, OffsetCst);
1249 Src = ConstantExpr::getBitCast(Src, PointerType::get(LoadTy, AS));
1250 return ConstantFoldLoadFromConstPtr(Src, &DL);
1254 /// ConstructSSAForLoadSet - Given a set of loads specified by ValuesPerBlock,
1255 /// construct SSA form, allowing us to eliminate LI. This returns the value
1256 /// that should be used at LI's definition site.
1257 static Value *ConstructSSAForLoadSet(LoadInst *LI,
1258 SmallVectorImpl<AvailableValueInBlock> &ValuesPerBlock,
1260 // Check for the fully redundant, dominating load case. In this case, we can
1261 // just use the dominating value directly.
1262 if (ValuesPerBlock.size() == 1 &&
1263 gvn.getDominatorTree().properlyDominates(ValuesPerBlock[0].BB,
1265 assert(!ValuesPerBlock[0].isUndefValue() && "Dead BB dominate this block");
1266 return ValuesPerBlock[0].MaterializeAdjustedValue(LI->getType(), gvn);
1269 // Otherwise, we have to construct SSA form.
1270 SmallVector<PHINode*, 8> NewPHIs;
1271 SSAUpdater SSAUpdate(&NewPHIs);
1272 SSAUpdate.Initialize(LI->getType(), LI->getName());
1274 Type *LoadTy = LI->getType();
1276 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i) {
1277 const AvailableValueInBlock &AV = ValuesPerBlock[i];
1278 BasicBlock *BB = AV.BB;
1280 if (SSAUpdate.HasValueForBlock(BB))
1283 SSAUpdate.AddAvailableValue(BB, AV.MaterializeAdjustedValue(LoadTy, gvn));
1286 // Perform PHI construction.
1287 Value *V = SSAUpdate.GetValueInMiddleOfBlock(LI->getParent());
1289 // If new PHI nodes were created, notify alias analysis.
1290 if (V->getType()->getScalarType()->isPointerTy()) {
1291 AliasAnalysis *AA = gvn.getAliasAnalysis();
1293 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i)
1294 AA->copyValue(LI, NewPHIs[i]);
1296 // Now that we've copied information to the new PHIs, scan through
1297 // them again and inform alias analysis that we've added potentially
1298 // escaping uses to any values that are operands to these PHIs.
1299 for (unsigned i = 0, e = NewPHIs.size(); i != e; ++i) {
1300 PHINode *P = NewPHIs[i];
1301 for (unsigned ii = 0, ee = P->getNumIncomingValues(); ii != ee; ++ii) {
1302 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
1303 AA->addEscapingUse(P->getOperandUse(jj));
1311 Value *AvailableValueInBlock::MaterializeAdjustedValue(Type *LoadTy, GVN &gvn) const {
1313 if (isSimpleValue()) {
1314 Res = getSimpleValue();
1315 if (Res->getType() != LoadTy) {
1316 const DataLayout *DL = gvn.getDataLayout();
1317 assert(DL && "Need target data to handle type mismatch case");
1318 Res = GetStoreValueForLoad(Res, Offset, LoadTy, BB->getTerminator(),
1321 DEBUG(dbgs() << "GVN COERCED NONLOCAL VAL:\nOffset: " << Offset << " "
1322 << *getSimpleValue() << '\n'
1323 << *Res << '\n' << "\n\n\n");
1325 } else if (isCoercedLoadValue()) {
1326 LoadInst *Load = getCoercedLoadValue();
1327 if (Load->getType() == LoadTy && Offset == 0) {
1330 Res = GetLoadValueForLoad(Load, Offset, LoadTy, BB->getTerminator(),
1333 DEBUG(dbgs() << "GVN COERCED NONLOCAL LOAD:\nOffset: " << Offset << " "
1334 << *getCoercedLoadValue() << '\n'
1335 << *Res << '\n' << "\n\n\n");
1337 } else if (isMemIntrinValue()) {
1338 const DataLayout *DL = gvn.getDataLayout();
1339 assert(DL && "Need target data to handle type mismatch case");
1340 Res = GetMemInstValueForLoad(getMemIntrinValue(), Offset,
1341 LoadTy, BB->getTerminator(), *DL);
1342 DEBUG(dbgs() << "GVN COERCED NONLOCAL MEM INTRIN:\nOffset: " << Offset
1343 << " " << *getMemIntrinValue() << '\n'
1344 << *Res << '\n' << "\n\n\n");
1346 assert(isUndefValue() && "Should be UndefVal");
1347 DEBUG(dbgs() << "GVN COERCED NONLOCAL Undef:\n";);
1348 return UndefValue::get(LoadTy);
1353 static bool isLifetimeStart(const Instruction *Inst) {
1354 if (const IntrinsicInst* II = dyn_cast<IntrinsicInst>(Inst))
1355 return II->getIntrinsicID() == Intrinsic::lifetime_start;
1359 void GVN::AnalyzeLoadAvailability(LoadInst *LI, LoadDepVect &Deps,
1360 AvailValInBlkVect &ValuesPerBlock,
1361 UnavailBlkVect &UnavailableBlocks) {
1363 // Filter out useless results (non-locals, etc). Keep track of the blocks
1364 // where we have a value available in repl, also keep track of whether we see
1365 // dependencies that produce an unknown value for the load (such as a call
1366 // that could potentially clobber the load).
1367 unsigned NumDeps = Deps.size();
1368 for (unsigned i = 0, e = NumDeps; i != e; ++i) {
1369 BasicBlock *DepBB = Deps[i].getBB();
1370 MemDepResult DepInfo = Deps[i].getResult();
1372 if (DeadBlocks.count(DepBB)) {
1373 // Dead dependent mem-op disguise as a load evaluating the same value
1374 // as the load in question.
1375 ValuesPerBlock.push_back(AvailableValueInBlock::getUndef(DepBB));
1379 if (!DepInfo.isDef() && !DepInfo.isClobber()) {
1380 UnavailableBlocks.push_back(DepBB);
1384 if (DepInfo.isClobber()) {
1385 // The address being loaded in this non-local block may not be the same as
1386 // the pointer operand of the load if PHI translation occurs. Make sure
1387 // to consider the right address.
1388 Value *Address = Deps[i].getAddress();
1390 // If the dependence is to a store that writes to a superset of the bits
1391 // read by the load, we can extract the bits we need for the load from the
1393 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInfo.getInst())) {
1394 if (DL && Address) {
1395 int Offset = AnalyzeLoadFromClobberingStore(LI->getType(), Address,
1398 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1399 DepSI->getValueOperand(),
1406 // Check to see if we have something like this:
1409 // if we have this, replace the later with an extraction from the former.
1410 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInfo.getInst())) {
1411 // If this is a clobber and L is the first instruction in its block, then
1412 // we have the first instruction in the entry block.
1413 if (DepLI != LI && Address && DL) {
1414 int Offset = AnalyzeLoadFromClobberingLoad(LI->getType(),
1415 LI->getPointerOperand(),
1419 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB,DepLI,
1426 // If the clobbering value is a memset/memcpy/memmove, see if we can
1427 // forward a value on from it.
1428 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(DepInfo.getInst())) {
1429 if (DL && Address) {
1430 int Offset = AnalyzeLoadFromClobberingMemInst(LI->getType(), Address,
1433 ValuesPerBlock.push_back(AvailableValueInBlock::getMI(DepBB, DepMI,
1440 UnavailableBlocks.push_back(DepBB);
1444 // DepInfo.isDef() here
1446 Instruction *DepInst = DepInfo.getInst();
1448 // Loading the allocation -> undef.
1449 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI) ||
1450 // Loading immediately after lifetime begin -> undef.
1451 isLifetimeStart(DepInst)) {
1452 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1453 UndefValue::get(LI->getType())));
1457 if (StoreInst *S = dyn_cast<StoreInst>(DepInst)) {
1458 // Reject loads and stores that are to the same address but are of
1459 // different types if we have to.
1460 if (S->getValueOperand()->getType() != LI->getType()) {
1461 // If the stored value is larger or equal to the loaded value, we can
1463 if (DL == 0 || !CanCoerceMustAliasedValueToLoad(S->getValueOperand(),
1464 LI->getType(), *DL)) {
1465 UnavailableBlocks.push_back(DepBB);
1470 ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
1471 S->getValueOperand()));
1475 if (LoadInst *LD = dyn_cast<LoadInst>(DepInst)) {
1476 // If the types mismatch and we can't handle it, reject reuse of the load.
1477 if (LD->getType() != LI->getType()) {
1478 // If the stored value is larger or equal to the loaded value, we can
1480 if (DL == 0 || !CanCoerceMustAliasedValueToLoad(LD, LI->getType(),*DL)){
1481 UnavailableBlocks.push_back(DepBB);
1485 ValuesPerBlock.push_back(AvailableValueInBlock::getLoad(DepBB, LD));
1489 UnavailableBlocks.push_back(DepBB);
1493 bool GVN::PerformLoadPRE(LoadInst *LI, AvailValInBlkVect &ValuesPerBlock,
1494 UnavailBlkVect &UnavailableBlocks) {
1495 // Okay, we have *some* definitions of the value. This means that the value
1496 // is available in some of our (transitive) predecessors. Lets think about
1497 // doing PRE of this load. This will involve inserting a new load into the
1498 // predecessor when it's not available. We could do this in general, but
1499 // prefer to not increase code size. As such, we only do this when we know
1500 // that we only have to insert *one* load (which means we're basically moving
1501 // the load, not inserting a new one).
1503 SmallPtrSet<BasicBlock *, 4> Blockers;
1504 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1505 Blockers.insert(UnavailableBlocks[i]);
1507 // Let's find the first basic block with more than one predecessor. Walk
1508 // backwards through predecessors if needed.
1509 BasicBlock *LoadBB = LI->getParent();
1510 BasicBlock *TmpBB = LoadBB;
1512 while (TmpBB->getSinglePredecessor()) {
1513 TmpBB = TmpBB->getSinglePredecessor();
1514 if (TmpBB == LoadBB) // Infinite (unreachable) loop.
1516 if (Blockers.count(TmpBB))
1519 // If any of these blocks has more than one successor (i.e. if the edge we
1520 // just traversed was critical), then there are other paths through this
1521 // block along which the load may not be anticipated. Hoisting the load
1522 // above this block would be adding the load to execution paths along
1523 // which it was not previously executed.
1524 if (TmpBB->getTerminator()->getNumSuccessors() != 1)
1531 // Check to see how many predecessors have the loaded value fully
1533 DenseMap<BasicBlock*, Value*> PredLoads;
1534 DenseMap<BasicBlock*, char> FullyAvailableBlocks;
1535 for (unsigned i = 0, e = ValuesPerBlock.size(); i != e; ++i)
1536 FullyAvailableBlocks[ValuesPerBlock[i].BB] = true;
1537 for (unsigned i = 0, e = UnavailableBlocks.size(); i != e; ++i)
1538 FullyAvailableBlocks[UnavailableBlocks[i]] = false;
1540 SmallVector<BasicBlock *, 4> CriticalEdgePred;
1541 for (pred_iterator PI = pred_begin(LoadBB), E = pred_end(LoadBB);
1543 BasicBlock *Pred = *PI;
1544 if (IsValueFullyAvailableInBlock(Pred, FullyAvailableBlocks, 0)) {
1547 PredLoads[Pred] = 0;
1549 if (Pred->getTerminator()->getNumSuccessors() != 1) {
1550 if (isa<IndirectBrInst>(Pred->getTerminator())) {
1551 DEBUG(dbgs() << "COULD NOT PRE LOAD BECAUSE OF INDBR CRITICAL EDGE '"
1552 << Pred->getName() << "': " << *LI << '\n');
1556 if (LoadBB->isLandingPad()) {
1558 << "COULD NOT PRE LOAD BECAUSE OF LANDING PAD CRITICAL EDGE '"
1559 << Pred->getName() << "': " << *LI << '\n');
1563 CriticalEdgePred.push_back(Pred);
1567 // Decide whether PRE is profitable for this load.
1568 unsigned NumUnavailablePreds = PredLoads.size();
1569 assert(NumUnavailablePreds != 0 &&
1570 "Fully available value should already be eliminated!");
1572 // If this load is unavailable in multiple predecessors, reject it.
1573 // FIXME: If we could restructure the CFG, we could make a common pred with
1574 // all the preds that don't have an available LI and insert a new load into
1576 if (NumUnavailablePreds != 1)
1579 // Split critical edges, and update the unavailable predecessors accordingly.
1580 for (SmallVectorImpl<BasicBlock *>::iterator I = CriticalEdgePred.begin(),
1581 E = CriticalEdgePred.end(); I != E; I++) {
1582 BasicBlock *OrigPred = *I;
1583 BasicBlock *NewPred = splitCriticalEdges(OrigPred, LoadBB);
1584 PredLoads.erase(OrigPred);
1585 PredLoads[NewPred] = 0;
1586 DEBUG(dbgs() << "Split critical edge " << OrigPred->getName() << "->"
1587 << LoadBB->getName() << '\n');
1590 // Check if the load can safely be moved to all the unavailable predecessors.
1591 bool CanDoPRE = true;
1592 SmallVector<Instruction*, 8> NewInsts;
1593 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1594 E = PredLoads.end(); I != E; ++I) {
1595 BasicBlock *UnavailablePred = I->first;
1597 // Do PHI translation to get its value in the predecessor if necessary. The
1598 // returned pointer (if non-null) is guaranteed to dominate UnavailablePred.
1600 // If all preds have a single successor, then we know it is safe to insert
1601 // the load on the pred (?!?), so we can insert code to materialize the
1602 // pointer if it is not available.
1603 PHITransAddr Address(LI->getPointerOperand(), DL);
1605 LoadPtr = Address.PHITranslateWithInsertion(LoadBB, UnavailablePred,
1608 // If we couldn't find or insert a computation of this phi translated value,
1611 DEBUG(dbgs() << "COULDN'T INSERT PHI TRANSLATED VALUE OF: "
1612 << *LI->getPointerOperand() << "\n");
1617 I->second = LoadPtr;
1621 while (!NewInsts.empty()) {
1622 Instruction *I = NewInsts.pop_back_val();
1623 if (MD) MD->removeInstruction(I);
1624 I->eraseFromParent();
1626 // HINT:Don't revert the edge-splitting as following transformation may
1627 // also need to split these critial edges.
1628 return !CriticalEdgePred.empty();
1631 // Okay, we can eliminate this load by inserting a reload in the predecessor
1632 // and using PHI construction to get the value in the other predecessors, do
1634 DEBUG(dbgs() << "GVN REMOVING PRE LOAD: " << *LI << '\n');
1635 DEBUG(if (!NewInsts.empty())
1636 dbgs() << "INSERTED " << NewInsts.size() << " INSTS: "
1637 << *NewInsts.back() << '\n');
1639 // Assign value numbers to the new instructions.
1640 for (unsigned i = 0, e = NewInsts.size(); i != e; ++i) {
1641 // FIXME: We really _ought_ to insert these value numbers into their
1642 // parent's availability map. However, in doing so, we risk getting into
1643 // ordering issues. If a block hasn't been processed yet, we would be
1644 // marking a value as AVAIL-IN, which isn't what we intend.
1645 VN.lookup_or_add(NewInsts[i]);
1648 for (DenseMap<BasicBlock*, Value*>::iterator I = PredLoads.begin(),
1649 E = PredLoads.end(); I != E; ++I) {
1650 BasicBlock *UnavailablePred = I->first;
1651 Value *LoadPtr = I->second;
1653 Instruction *NewLoad = new LoadInst(LoadPtr, LI->getName()+".pre", false,
1655 UnavailablePred->getTerminator());
1657 // Transfer the old load's TBAA tag to the new load.
1658 if (MDNode *Tag = LI->getMetadata(LLVMContext::MD_tbaa))
1659 NewLoad->setMetadata(LLVMContext::MD_tbaa, Tag);
1661 // Transfer DebugLoc.
1662 NewLoad->setDebugLoc(LI->getDebugLoc());
1664 // Add the newly created load.
1665 ValuesPerBlock.push_back(AvailableValueInBlock::get(UnavailablePred,
1667 MD->invalidateCachedPointerInfo(LoadPtr);
1668 DEBUG(dbgs() << "GVN INSERTED " << *NewLoad << '\n');
1671 // Perform PHI construction.
1672 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1673 LI->replaceAllUsesWith(V);
1674 if (isa<PHINode>(V))
1676 if (V->getType()->getScalarType()->isPointerTy())
1677 MD->invalidateCachedPointerInfo(V);
1678 markInstructionForDeletion(LI);
1683 /// processNonLocalLoad - Attempt to eliminate a load whose dependencies are
1684 /// non-local by performing PHI construction.
1685 bool GVN::processNonLocalLoad(LoadInst *LI) {
1686 // Step 1: Find the non-local dependencies of the load.
1688 AliasAnalysis::Location Loc = VN.getAliasAnalysis()->getLocation(LI);
1689 MD->getNonLocalPointerDependency(Loc, true, LI->getParent(), Deps);
1691 // If we had to process more than one hundred blocks to find the
1692 // dependencies, this load isn't worth worrying about. Optimizing
1693 // it will be too expensive.
1694 unsigned NumDeps = Deps.size();
1698 // If we had a phi translation failure, we'll have a single entry which is a
1699 // clobber in the current block. Reject this early.
1701 !Deps[0].getResult().isDef() && !Deps[0].getResult().isClobber()) {
1703 dbgs() << "GVN: non-local load ";
1704 LI->printAsOperand(dbgs());
1705 dbgs() << " has unknown dependencies\n";
1710 // Step 2: Analyze the availability of the load
1711 AvailValInBlkVect ValuesPerBlock;
1712 UnavailBlkVect UnavailableBlocks;
1713 AnalyzeLoadAvailability(LI, Deps, ValuesPerBlock, UnavailableBlocks);
1715 // If we have no predecessors that produce a known value for this load, exit
1717 if (ValuesPerBlock.empty())
1720 // Step 3: Eliminate fully redundancy.
1722 // If all of the instructions we depend on produce a known value for this
1723 // load, then it is fully redundant and we can use PHI insertion to compute
1724 // its value. Insert PHIs and remove the fully redundant value now.
1725 if (UnavailableBlocks.empty()) {
1726 DEBUG(dbgs() << "GVN REMOVING NONLOCAL LOAD: " << *LI << '\n');
1728 // Perform PHI construction.
1729 Value *V = ConstructSSAForLoadSet(LI, ValuesPerBlock, *this);
1730 LI->replaceAllUsesWith(V);
1732 if (isa<PHINode>(V))
1734 if (V->getType()->getScalarType()->isPointerTy())
1735 MD->invalidateCachedPointerInfo(V);
1736 markInstructionForDeletion(LI);
1741 // Step 4: Eliminate partial redundancy.
1742 if (!EnablePRE || !EnableLoadPRE)
1745 return PerformLoadPRE(LI, ValuesPerBlock, UnavailableBlocks);
1749 static void patchReplacementInstruction(Instruction *I, Value *Repl) {
1750 // Patch the replacement so that it is not more restrictive than the value
1752 BinaryOperator *Op = dyn_cast<BinaryOperator>(I);
1753 BinaryOperator *ReplOp = dyn_cast<BinaryOperator>(Repl);
1754 if (Op && ReplOp && isa<OverflowingBinaryOperator>(Op) &&
1755 isa<OverflowingBinaryOperator>(ReplOp)) {
1756 if (ReplOp->hasNoSignedWrap() && !Op->hasNoSignedWrap())
1757 ReplOp->setHasNoSignedWrap(false);
1758 if (ReplOp->hasNoUnsignedWrap() && !Op->hasNoUnsignedWrap())
1759 ReplOp->setHasNoUnsignedWrap(false);
1761 if (Instruction *ReplInst = dyn_cast<Instruction>(Repl)) {
1762 SmallVector<std::pair<unsigned, MDNode*>, 4> Metadata;
1763 ReplInst->getAllMetadataOtherThanDebugLoc(Metadata);
1764 for (int i = 0, n = Metadata.size(); i < n; ++i) {
1765 unsigned Kind = Metadata[i].first;
1766 MDNode *IMD = I->getMetadata(Kind);
1767 MDNode *ReplMD = Metadata[i].second;
1770 ReplInst->setMetadata(Kind, NULL); // Remove unknown metadata
1772 case LLVMContext::MD_dbg:
1773 llvm_unreachable("getAllMetadataOtherThanDebugLoc returned a MD_dbg");
1774 case LLVMContext::MD_tbaa:
1775 ReplInst->setMetadata(Kind, MDNode::getMostGenericTBAA(IMD, ReplMD));
1777 case LLVMContext::MD_range:
1778 ReplInst->setMetadata(Kind, MDNode::getMostGenericRange(IMD, ReplMD));
1780 case LLVMContext::MD_prof:
1781 llvm_unreachable("MD_prof in a non-terminator instruction");
1783 case LLVMContext::MD_fpmath:
1784 ReplInst->setMetadata(Kind, MDNode::getMostGenericFPMath(IMD, ReplMD));
1791 static void patchAndReplaceAllUsesWith(Instruction *I, Value *Repl) {
1792 patchReplacementInstruction(I, Repl);
1793 I->replaceAllUsesWith(Repl);
1796 /// processLoad - Attempt to eliminate a load, first by eliminating it
1797 /// locally, and then attempting non-local elimination if that fails.
1798 bool GVN::processLoad(LoadInst *L) {
1805 if (L->use_empty()) {
1806 markInstructionForDeletion(L);
1810 // ... to a pointer that has been loaded from before...
1811 MemDepResult Dep = MD->getDependency(L);
1813 // If we have a clobber and target data is around, see if this is a clobber
1814 // that we can fix up through code synthesis.
1815 if (Dep.isClobber() && DL) {
1816 // Check to see if we have something like this:
1817 // store i32 123, i32* %P
1818 // %A = bitcast i32* %P to i8*
1819 // %B = gep i8* %A, i32 1
1822 // We could do that by recognizing if the clobber instructions are obviously
1823 // a common base + constant offset, and if the previous store (or memset)
1824 // completely covers this load. This sort of thing can happen in bitfield
1826 Value *AvailVal = 0;
1827 if (StoreInst *DepSI = dyn_cast<StoreInst>(Dep.getInst())) {
1828 int Offset = AnalyzeLoadFromClobberingStore(L->getType(),
1829 L->getPointerOperand(),
1832 AvailVal = GetStoreValueForLoad(DepSI->getValueOperand(), Offset,
1833 L->getType(), L, *DL);
1836 // Check to see if we have something like this:
1839 // if we have this, replace the later with an extraction from the former.
1840 if (LoadInst *DepLI = dyn_cast<LoadInst>(Dep.getInst())) {
1841 // If this is a clobber and L is the first instruction in its block, then
1842 // we have the first instruction in the entry block.
1846 int Offset = AnalyzeLoadFromClobberingLoad(L->getType(),
1847 L->getPointerOperand(),
1850 AvailVal = GetLoadValueForLoad(DepLI, Offset, L->getType(), L, *this);
1853 // If the clobbering value is a memset/memcpy/memmove, see if we can forward
1854 // a value on from it.
1855 if (MemIntrinsic *DepMI = dyn_cast<MemIntrinsic>(Dep.getInst())) {
1856 int Offset = AnalyzeLoadFromClobberingMemInst(L->getType(),
1857 L->getPointerOperand(),
1860 AvailVal = GetMemInstValueForLoad(DepMI, Offset, L->getType(), L, *DL);
1864 DEBUG(dbgs() << "GVN COERCED INST:\n" << *Dep.getInst() << '\n'
1865 << *AvailVal << '\n' << *L << "\n\n\n");
1867 // Replace the load!
1868 L->replaceAllUsesWith(AvailVal);
1869 if (AvailVal->getType()->getScalarType()->isPointerTy())
1870 MD->invalidateCachedPointerInfo(AvailVal);
1871 markInstructionForDeletion(L);
1877 // If the value isn't available, don't do anything!
1878 if (Dep.isClobber()) {
1880 // fast print dep, using operator<< on instruction is too slow.
1881 dbgs() << "GVN: load ";
1882 L->printAsOperand(dbgs());
1883 Instruction *I = Dep.getInst();
1884 dbgs() << " is clobbered by " << *I << '\n';
1889 // If it is defined in another block, try harder.
1890 if (Dep.isNonLocal())
1891 return processNonLocalLoad(L);
1895 // fast print dep, using operator<< on instruction is too slow.
1896 dbgs() << "GVN: load ";
1897 L->printAsOperand(dbgs());
1898 dbgs() << " has unknown dependence\n";
1903 Instruction *DepInst = Dep.getInst();
1904 if (StoreInst *DepSI = dyn_cast<StoreInst>(DepInst)) {
1905 Value *StoredVal = DepSI->getValueOperand();
1907 // The store and load are to a must-aliased pointer, but they may not
1908 // actually have the same type. See if we know how to reuse the stored
1909 // value (depending on its type).
1910 if (StoredVal->getType() != L->getType()) {
1912 StoredVal = CoerceAvailableValueToLoadType(StoredVal, L->getType(),
1917 DEBUG(dbgs() << "GVN COERCED STORE:\n" << *DepSI << '\n' << *StoredVal
1918 << '\n' << *L << "\n\n\n");
1925 L->replaceAllUsesWith(StoredVal);
1926 if (StoredVal->getType()->getScalarType()->isPointerTy())
1927 MD->invalidateCachedPointerInfo(StoredVal);
1928 markInstructionForDeletion(L);
1933 if (LoadInst *DepLI = dyn_cast<LoadInst>(DepInst)) {
1934 Value *AvailableVal = DepLI;
1936 // The loads are of a must-aliased pointer, but they may not actually have
1937 // the same type. See if we know how to reuse the previously loaded value
1938 // (depending on its type).
1939 if (DepLI->getType() != L->getType()) {
1941 AvailableVal = CoerceAvailableValueToLoadType(DepLI, L->getType(),
1943 if (AvailableVal == 0)
1946 DEBUG(dbgs() << "GVN COERCED LOAD:\n" << *DepLI << "\n" << *AvailableVal
1947 << "\n" << *L << "\n\n\n");
1954 patchAndReplaceAllUsesWith(L, AvailableVal);
1955 if (DepLI->getType()->getScalarType()->isPointerTy())
1956 MD->invalidateCachedPointerInfo(DepLI);
1957 markInstructionForDeletion(L);
1962 // If this load really doesn't depend on anything, then we must be loading an
1963 // undef value. This can happen when loading for a fresh allocation with no
1964 // intervening stores, for example.
1965 if (isa<AllocaInst>(DepInst) || isMallocLikeFn(DepInst, TLI)) {
1966 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1967 markInstructionForDeletion(L);
1972 // If this load occurs either right after a lifetime begin,
1973 // then the loaded value is undefined.
1974 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(DepInst)) {
1975 if (II->getIntrinsicID() == Intrinsic::lifetime_start) {
1976 L->replaceAllUsesWith(UndefValue::get(L->getType()));
1977 markInstructionForDeletion(L);
1986 // findLeader - In order to find a leader for a given value number at a
1987 // specific basic block, we first obtain the list of all Values for that number,
1988 // and then scan the list to find one whose block dominates the block in
1989 // question. This is fast because dominator tree queries consist of only
1990 // a few comparisons of DFS numbers.
1991 Value *GVN::findLeader(const BasicBlock *BB, uint32_t num) {
1992 LeaderTableEntry Vals = LeaderTable[num];
1993 if (!Vals.Val) return 0;
1996 if (DT->dominates(Vals.BB, BB)) {
1998 if (isa<Constant>(Val)) return Val;
2001 LeaderTableEntry* Next = Vals.Next;
2003 if (DT->dominates(Next->BB, BB)) {
2004 if (isa<Constant>(Next->Val)) return Next->Val;
2005 if (!Val) Val = Next->Val;
2014 /// replaceAllDominatedUsesWith - Replace all uses of 'From' with 'To' if the
2015 /// use is dominated by the given basic block. Returns the number of uses that
2017 unsigned GVN::replaceAllDominatedUsesWith(Value *From, Value *To,
2018 const BasicBlockEdge &Root) {
2020 for (Value::use_iterator UI = From->use_begin(), UE = From->use_end();
2024 if (DT->dominates(Root, U)) {
2032 /// isOnlyReachableViaThisEdge - There is an edge from 'Src' to 'Dst'. Return
2033 /// true if every path from the entry block to 'Dst' passes via this edge. In
2034 /// particular 'Dst' must not be reachable via another edge from 'Src'.
2035 static bool isOnlyReachableViaThisEdge(const BasicBlockEdge &E,
2036 DominatorTree *DT) {
2037 // While in theory it is interesting to consider the case in which Dst has
2038 // more than one predecessor, because Dst might be part of a loop which is
2039 // only reachable from Src, in practice it is pointless since at the time
2040 // GVN runs all such loops have preheaders, which means that Dst will have
2041 // been changed to have only one predecessor, namely Src.
2042 const BasicBlock *Pred = E.getEnd()->getSinglePredecessor();
2043 const BasicBlock *Src = E.getStart();
2044 assert((!Pred || Pred == Src) && "No edge between these basic blocks!");
2049 /// propagateEquality - The given values are known to be equal in every block
2050 /// dominated by 'Root'. Exploit this, for example by replacing 'LHS' with
2051 /// 'RHS' everywhere in the scope. Returns whether a change was made.
2052 bool GVN::propagateEquality(Value *LHS, Value *RHS,
2053 const BasicBlockEdge &Root) {
2054 SmallVector<std::pair<Value*, Value*>, 4> Worklist;
2055 Worklist.push_back(std::make_pair(LHS, RHS));
2056 bool Changed = false;
2057 // For speed, compute a conservative fast approximation to
2058 // DT->dominates(Root, Root.getEnd());
2059 bool RootDominatesEnd = isOnlyReachableViaThisEdge(Root, DT);
2061 while (!Worklist.empty()) {
2062 std::pair<Value*, Value*> Item = Worklist.pop_back_val();
2063 LHS = Item.first; RHS = Item.second;
2065 if (LHS == RHS) continue;
2066 assert(LHS->getType() == RHS->getType() && "Equality but unequal types!");
2068 // Don't try to propagate equalities between constants.
2069 if (isa<Constant>(LHS) && isa<Constant>(RHS)) continue;
2071 // Prefer a constant on the right-hand side, or an Argument if no constants.
2072 if (isa<Constant>(LHS) || (isa<Argument>(LHS) && !isa<Constant>(RHS)))
2073 std::swap(LHS, RHS);
2074 assert((isa<Argument>(LHS) || isa<Instruction>(LHS)) && "Unexpected value!");
2076 // If there is no obvious reason to prefer the left-hand side over the right-
2077 // hand side, ensure the longest lived term is on the right-hand side, so the
2078 // shortest lived term will be replaced by the longest lived. This tends to
2079 // expose more simplifications.
2080 uint32_t LVN = VN.lookup_or_add(LHS);
2081 if ((isa<Argument>(LHS) && isa<Argument>(RHS)) ||
2082 (isa<Instruction>(LHS) && isa<Instruction>(RHS))) {
2083 // Move the 'oldest' value to the right-hand side, using the value number as
2085 uint32_t RVN = VN.lookup_or_add(RHS);
2087 std::swap(LHS, RHS);
2092 // If value numbering later sees that an instruction in the scope is equal
2093 // to 'LHS' then ensure it will be turned into 'RHS'. In order to preserve
2094 // the invariant that instructions only occur in the leader table for their
2095 // own value number (this is used by removeFromLeaderTable), do not do this
2096 // if RHS is an instruction (if an instruction in the scope is morphed into
2097 // LHS then it will be turned into RHS by the next GVN iteration anyway, so
2098 // using the leader table is about compiling faster, not optimizing better).
2099 // The leader table only tracks basic blocks, not edges. Only add to if we
2100 // have the simple case where the edge dominates the end.
2101 if (RootDominatesEnd && !isa<Instruction>(RHS))
2102 addToLeaderTable(LVN, RHS, Root.getEnd());
2104 // Replace all occurrences of 'LHS' with 'RHS' everywhere in the scope. As
2105 // LHS always has at least one use that is not dominated by Root, this will
2106 // never do anything if LHS has only one use.
2107 if (!LHS->hasOneUse()) {
2108 unsigned NumReplacements = replaceAllDominatedUsesWith(LHS, RHS, Root);
2109 Changed |= NumReplacements > 0;
2110 NumGVNEqProp += NumReplacements;
2113 // Now try to deduce additional equalities from this one. For example, if the
2114 // known equality was "(A != B)" == "false" then it follows that A and B are
2115 // equal in the scope. Only boolean equalities with an explicit true or false
2116 // RHS are currently supported.
2117 if (!RHS->getType()->isIntegerTy(1))
2118 // Not a boolean equality - bail out.
2120 ConstantInt *CI = dyn_cast<ConstantInt>(RHS);
2122 // RHS neither 'true' nor 'false' - bail out.
2124 // Whether RHS equals 'true'. Otherwise it equals 'false'.
2125 bool isKnownTrue = CI->isAllOnesValue();
2126 bool isKnownFalse = !isKnownTrue;
2128 // If "A && B" is known true then both A and B are known true. If "A || B"
2129 // is known false then both A and B are known false.
2131 if ((isKnownTrue && match(LHS, m_And(m_Value(A), m_Value(B)))) ||
2132 (isKnownFalse && match(LHS, m_Or(m_Value(A), m_Value(B))))) {
2133 Worklist.push_back(std::make_pair(A, RHS));
2134 Worklist.push_back(std::make_pair(B, RHS));
2138 // If we are propagating an equality like "(A == B)" == "true" then also
2139 // propagate the equality A == B. When propagating a comparison such as
2140 // "(A >= B)" == "true", replace all instances of "A < B" with "false".
2141 if (ICmpInst *Cmp = dyn_cast<ICmpInst>(LHS)) {
2142 Value *Op0 = Cmp->getOperand(0), *Op1 = Cmp->getOperand(1);
2144 // If "A == B" is known true, or "A != B" is known false, then replace
2145 // A with B everywhere in the scope.
2146 if ((isKnownTrue && Cmp->getPredicate() == CmpInst::ICMP_EQ) ||
2147 (isKnownFalse && Cmp->getPredicate() == CmpInst::ICMP_NE))
2148 Worklist.push_back(std::make_pair(Op0, Op1));
2150 // If "A >= B" is known true, replace "A < B" with false everywhere.
2151 CmpInst::Predicate NotPred = Cmp->getInversePredicate();
2152 Constant *NotVal = ConstantInt::get(Cmp->getType(), isKnownFalse);
2153 // Since we don't have the instruction "A < B" immediately to hand, work out
2154 // the value number that it would have and use that to find an appropriate
2155 // instruction (if any).
2156 uint32_t NextNum = VN.getNextUnusedValueNumber();
2157 uint32_t Num = VN.lookup_or_add_cmp(Cmp->getOpcode(), NotPred, Op0, Op1);
2158 // If the number we were assigned was brand new then there is no point in
2159 // looking for an instruction realizing it: there cannot be one!
2160 if (Num < NextNum) {
2161 Value *NotCmp = findLeader(Root.getEnd(), Num);
2162 if (NotCmp && isa<Instruction>(NotCmp)) {
2163 unsigned NumReplacements =
2164 replaceAllDominatedUsesWith(NotCmp, NotVal, Root);
2165 Changed |= NumReplacements > 0;
2166 NumGVNEqProp += NumReplacements;
2169 // Ensure that any instruction in scope that gets the "A < B" value number
2170 // is replaced with false.
2171 // The leader table only tracks basic blocks, not edges. Only add to if we
2172 // have the simple case where the edge dominates the end.
2173 if (RootDominatesEnd)
2174 addToLeaderTable(Num, NotVal, Root.getEnd());
2183 static bool normalOpAfterIntrinsic(Instruction *I, Value *Repl)
2185 switch (I->getOpcode()) {
2186 case Instruction::Add:
2187 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Repl))
2188 return II->getIntrinsicID() == Intrinsic::sadd_with_overflow
2189 || II->getIntrinsicID() == Intrinsic::uadd_with_overflow;
2191 case Instruction::Sub:
2192 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Repl))
2193 return II->getIntrinsicID() == Intrinsic::ssub_with_overflow
2194 || II->getIntrinsicID() == Intrinsic::usub_with_overflow;
2196 case Instruction::Mul:
2197 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Repl))
2198 return II->getIntrinsicID() == Intrinsic::smul_with_overflow
2199 || II->getIntrinsicID() == Intrinsic::umul_with_overflow;
2206 static bool intrinsicAterNormalOp(Instruction *I, Value *Repl)
2208 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
2212 Instruction *RI = dyn_cast<Instruction>(Repl);
2216 switch (RI->getOpcode()) {
2217 case Instruction::Add:
2218 return II->getIntrinsicID() == Intrinsic::sadd_with_overflow
2219 || II->getIntrinsicID() == Intrinsic::uadd_with_overflow;
2220 case Instruction::Sub:
2221 return II->getIntrinsicID() == Intrinsic::ssub_with_overflow
2222 || II->getIntrinsicID() == Intrinsic::usub_with_overflow;
2223 case Instruction::Mul:
2224 return II->getIntrinsicID() == Intrinsic::smul_with_overflow
2225 || II->getIntrinsicID() == Intrinsic::umul_with_overflow;
2231 /// processInstruction - When calculating availability, handle an instruction
2232 /// by inserting it into the appropriate sets
2233 bool GVN::processInstruction(Instruction *I) {
2234 // Ignore dbg info intrinsics.
2235 if (isa<DbgInfoIntrinsic>(I))
2238 // If the instruction can be easily simplified then do so now in preference
2239 // to value numbering it. Value numbering often exposes redundancies, for
2240 // example if it determines that %y is equal to %x then the instruction
2241 // "%z = and i32 %x, %y" becomes "%z = and i32 %x, %x" which we now simplify.
2242 if (Value *V = SimplifyInstruction(I, DL, TLI, DT)) {
2243 I->replaceAllUsesWith(V);
2244 if (MD && V->getType()->getScalarType()->isPointerTy())
2245 MD->invalidateCachedPointerInfo(V);
2246 markInstructionForDeletion(I);
2251 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
2252 if (processLoad(LI))
2255 unsigned Num = VN.lookup_or_add(LI);
2256 addToLeaderTable(Num, LI, LI->getParent());
2260 // For conditional branches, we can perform simple conditional propagation on
2261 // the condition value itself.
2262 if (BranchInst *BI = dyn_cast<BranchInst>(I)) {
2263 if (!BI->isConditional())
2266 if (isa<Constant>(BI->getCondition()))
2267 return processFoldableCondBr(BI);
2269 Value *BranchCond = BI->getCondition();
2270 BasicBlock *TrueSucc = BI->getSuccessor(0);
2271 BasicBlock *FalseSucc = BI->getSuccessor(1);
2272 // Avoid multiple edges early.
2273 if (TrueSucc == FalseSucc)
2276 BasicBlock *Parent = BI->getParent();
2277 bool Changed = false;
2279 Value *TrueVal = ConstantInt::getTrue(TrueSucc->getContext());
2280 BasicBlockEdge TrueE(Parent, TrueSucc);
2281 Changed |= propagateEquality(BranchCond, TrueVal, TrueE);
2283 Value *FalseVal = ConstantInt::getFalse(FalseSucc->getContext());
2284 BasicBlockEdge FalseE(Parent, FalseSucc);
2285 Changed |= propagateEquality(BranchCond, FalseVal, FalseE);
2290 // For switches, propagate the case values into the case destinations.
2291 if (SwitchInst *SI = dyn_cast<SwitchInst>(I)) {
2292 Value *SwitchCond = SI->getCondition();
2293 BasicBlock *Parent = SI->getParent();
2294 bool Changed = false;
2296 // Remember how many outgoing edges there are to every successor.
2297 SmallDenseMap<BasicBlock *, unsigned, 16> SwitchEdges;
2298 for (unsigned i = 0, n = SI->getNumSuccessors(); i != n; ++i)
2299 ++SwitchEdges[SI->getSuccessor(i)];
2301 for (SwitchInst::CaseIt i = SI->case_begin(), e = SI->case_end();
2303 BasicBlock *Dst = i.getCaseSuccessor();
2304 // If there is only a single edge, propagate the case value into it.
2305 if (SwitchEdges.lookup(Dst) == 1) {
2306 BasicBlockEdge E(Parent, Dst);
2307 Changed |= propagateEquality(SwitchCond, i.getCaseValue(), E);
2313 // Instructions with void type don't return a value, so there's
2314 // no point in trying to find redundancies in them.
2315 if (I->getType()->isVoidTy()) return false;
2317 uint32_t NextNum = VN.getNextUnusedValueNumber();
2318 unsigned Num = VN.lookup_or_add(I);
2320 // Allocations are always uniquely numbered, so we can save time and memory
2321 // by fast failing them.
2322 if (isa<AllocaInst>(I) || isa<TerminatorInst>(I) || isa<PHINode>(I)) {
2323 addToLeaderTable(Num, I, I->getParent());
2327 // If the number we were assigned was a brand new VN, then we don't
2328 // need to do a lookup to see if the number already exists
2329 // somewhere in the domtree: it can't!
2330 if (Num >= NextNum) {
2331 addToLeaderTable(Num, I, I->getParent());
2335 // Perform fast-path value-number based elimination of values inherited from
2337 Value *repl = findLeader(I->getParent(), Num);
2339 // Failure, just remember this instance for future use.
2340 addToLeaderTable(Num, I, I->getParent());
2344 if (normalOpAfterIntrinsic(I, repl)) {
2345 // An intrinsic followed by a normal operation (e.g. sadd_with_overflow
2346 // followed by a sadd): replace the second instruction with an extract.
2347 IntrinsicInst *II = cast<IntrinsicInst>(repl);
2349 repl = ExtractValueInst::Create(II, 0, I->getName() + ".repl", I);
2350 } else if (intrinsicAterNormalOp(I, repl)) {
2351 // A normal operation followed by an intrinsic (e.g. sadd followed by a
2352 // sadd_with_overflow).
2353 // Clone the intrinsic, and insert it before the replacing instruction. Then
2354 // replace the (current) instruction with the cloned one. In a subsequent
2355 // run, the original replacement (the non-intrinsic) will be be replaced by
2356 // the new intrinsic.
2357 Instruction *RI = dyn_cast<Instruction>(repl);
2359 Instruction *newIntrinsic = I->clone();
2360 newIntrinsic->setName(I->getName() + ".repl");
2361 newIntrinsic->insertBefore(RI);
2362 repl = newIntrinsic;
2366 patchAndReplaceAllUsesWith(I, repl);
2367 if (MD && repl->getType()->getScalarType()->isPointerTy())
2368 MD->invalidateCachedPointerInfo(repl);
2369 markInstructionForDeletion(I);
2373 /// runOnFunction - This is the main transformation entry point for a function.
2374 bool GVN::runOnFunction(Function& F) {
2375 if (skipOptnoneFunction(F))
2379 MD = &getAnalysis<MemoryDependenceAnalysis>();
2380 DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
2381 DataLayoutPass *DLP = getAnalysisIfAvailable<DataLayoutPass>();
2382 DL = DLP ? &DLP->getDataLayout() : 0;
2383 TLI = &getAnalysis<TargetLibraryInfo>();
2384 VN.setAliasAnalysis(&getAnalysis<AliasAnalysis>());
2388 bool Changed = false;
2389 bool ShouldContinue = true;
2391 // Merge unconditional branches, allowing PRE to catch more
2392 // optimization opportunities.
2393 for (Function::iterator FI = F.begin(), FE = F.end(); FI != FE; ) {
2394 BasicBlock *BB = FI++;
2396 bool removedBlock = MergeBlockIntoPredecessor(BB, this);
2397 if (removedBlock) ++NumGVNBlocks;
2399 Changed |= removedBlock;
2402 unsigned Iteration = 0;
2403 while (ShouldContinue) {
2404 DEBUG(dbgs() << "GVN iteration: " << Iteration << "\n");
2405 ShouldContinue = iterateOnFunction(F);
2406 Changed |= ShouldContinue;
2411 // Fabricate val-num for dead-code in order to suppress assertion in
2413 assignValNumForDeadCode();
2414 bool PREChanged = true;
2415 while (PREChanged) {
2416 PREChanged = performPRE(F);
2417 Changed |= PREChanged;
2421 // FIXME: Should perform GVN again after PRE does something. PRE can move
2422 // computations into blocks where they become fully redundant. Note that
2423 // we can't do this until PRE's critical edge splitting updates memdep.
2424 // Actually, when this happens, we should just fully integrate PRE into GVN.
2426 cleanupGlobalSets();
2427 // Do not cleanup DeadBlocks in cleanupGlobalSets() as it's called for each
2435 bool GVN::processBlock(BasicBlock *BB) {
2436 // FIXME: Kill off InstrsToErase by doing erasing eagerly in a helper function
2437 // (and incrementing BI before processing an instruction).
2438 assert(InstrsToErase.empty() &&
2439 "We expect InstrsToErase to be empty across iterations");
2440 if (DeadBlocks.count(BB))
2443 bool ChangedFunction = false;
2445 for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
2447 ChangedFunction |= processInstruction(BI);
2448 if (InstrsToErase.empty()) {
2453 // If we need some instructions deleted, do it now.
2454 NumGVNInstr += InstrsToErase.size();
2456 // Avoid iterator invalidation.
2457 bool AtStart = BI == BB->begin();
2461 for (SmallVectorImpl<Instruction *>::iterator I = InstrsToErase.begin(),
2462 E = InstrsToErase.end(); I != E; ++I) {
2463 DEBUG(dbgs() << "GVN removed: " << **I << '\n');
2464 if (MD) MD->removeInstruction(*I);
2465 DEBUG(verifyRemoved(*I));
2466 (*I)->eraseFromParent();
2468 InstrsToErase.clear();
2476 return ChangedFunction;
2479 /// performPRE - Perform a purely local form of PRE that looks for diamond
2480 /// control flow patterns and attempts to perform simple PRE at the join point.
2481 bool GVN::performPRE(Function &F) {
2482 bool Changed = false;
2483 SmallVector<std::pair<Value*, BasicBlock*>, 8> predMap;
2484 for (df_iterator<BasicBlock*> DI = df_begin(&F.getEntryBlock()),
2485 DE = df_end(&F.getEntryBlock()); DI != DE; ++DI) {
2486 BasicBlock *CurrentBlock = *DI;
2488 // Nothing to PRE in the entry block.
2489 if (CurrentBlock == &F.getEntryBlock()) continue;
2491 // Don't perform PRE on a landing pad.
2492 if (CurrentBlock->isLandingPad()) continue;
2494 for (BasicBlock::iterator BI = CurrentBlock->begin(),
2495 BE = CurrentBlock->end(); BI != BE; ) {
2496 Instruction *CurInst = BI++;
2498 if (isa<AllocaInst>(CurInst) ||
2499 isa<TerminatorInst>(CurInst) || isa<PHINode>(CurInst) ||
2500 CurInst->getType()->isVoidTy() ||
2501 CurInst->mayReadFromMemory() || CurInst->mayHaveSideEffects() ||
2502 isa<DbgInfoIntrinsic>(CurInst))
2505 // Don't do PRE on compares. The PHI would prevent CodeGenPrepare from
2506 // sinking the compare again, and it would force the code generator to
2507 // move the i1 from processor flags or predicate registers into a general
2508 // purpose register.
2509 if (isa<CmpInst>(CurInst))
2512 // We don't currently value number ANY inline asm calls.
2513 if (CallInst *CallI = dyn_cast<CallInst>(CurInst))
2514 if (CallI->isInlineAsm())
2517 uint32_t ValNo = VN.lookup(CurInst);
2519 // Look for the predecessors for PRE opportunities. We're
2520 // only trying to solve the basic diamond case, where
2521 // a value is computed in the successor and one predecessor,
2522 // but not the other. We also explicitly disallow cases
2523 // where the successor is its own predecessor, because they're
2524 // more complicated to get right.
2525 unsigned NumWith = 0;
2526 unsigned NumWithout = 0;
2527 BasicBlock *PREPred = 0;
2530 for (pred_iterator PI = pred_begin(CurrentBlock),
2531 PE = pred_end(CurrentBlock); PI != PE; ++PI) {
2532 BasicBlock *P = *PI;
2533 // We're not interested in PRE where the block is its
2534 // own predecessor, or in blocks with predecessors
2535 // that are not reachable.
2536 if (P == CurrentBlock) {
2539 } else if (!DT->isReachableFromEntry(P)) {
2544 Value* predV = findLeader(P, ValNo);
2546 predMap.push_back(std::make_pair(static_cast<Value *>(0), P));
2549 } else if (predV == CurInst) {
2550 /* CurInst dominates this predecessor. */
2554 predMap.push_back(std::make_pair(predV, P));
2559 // Don't do PRE when it might increase code size, i.e. when
2560 // we would need to insert instructions in more than one pred.
2561 if (NumWithout != 1 || NumWith == 0)
2564 // Don't do PRE across indirect branch.
2565 if (isa<IndirectBrInst>(PREPred->getTerminator()))
2568 // We can't do PRE safely on a critical edge, so instead we schedule
2569 // the edge to be split and perform the PRE the next time we iterate
2571 unsigned SuccNum = GetSuccessorNumber(PREPred, CurrentBlock);
2572 if (isCriticalEdge(PREPred->getTerminator(), SuccNum)) {
2573 toSplit.push_back(std::make_pair(PREPred->getTerminator(), SuccNum));
2577 // Instantiate the expression in the predecessor that lacked it.
2578 // Because we are going top-down through the block, all value numbers
2579 // will be available in the predecessor by the time we need them. Any
2580 // that weren't originally present will have been instantiated earlier
2582 Instruction *PREInstr = CurInst->clone();
2583 bool success = true;
2584 for (unsigned i = 0, e = CurInst->getNumOperands(); i != e; ++i) {
2585 Value *Op = PREInstr->getOperand(i);
2586 if (isa<Argument>(Op) || isa<Constant>(Op) || isa<GlobalValue>(Op))
2589 if (Value *V = findLeader(PREPred, VN.lookup(Op))) {
2590 PREInstr->setOperand(i, V);
2597 // Fail out if we encounter an operand that is not available in
2598 // the PRE predecessor. This is typically because of loads which
2599 // are not value numbered precisely.
2601 DEBUG(verifyRemoved(PREInstr));
2606 PREInstr->insertBefore(PREPred->getTerminator());
2607 PREInstr->setName(CurInst->getName() + ".pre");
2608 PREInstr->setDebugLoc(CurInst->getDebugLoc());
2609 VN.add(PREInstr, ValNo);
2612 // Update the availability map to include the new instruction.
2613 addToLeaderTable(ValNo, PREInstr, PREPred);
2615 // Create a PHI to make the value available in this block.
2616 PHINode* Phi = PHINode::Create(CurInst->getType(), predMap.size(),
2617 CurInst->getName() + ".pre-phi",
2618 CurrentBlock->begin());
2619 for (unsigned i = 0, e = predMap.size(); i != e; ++i) {
2620 if (Value *V = predMap[i].first)
2621 Phi->addIncoming(V, predMap[i].second);
2623 Phi->addIncoming(PREInstr, PREPred);
2627 addToLeaderTable(ValNo, Phi, CurrentBlock);
2628 Phi->setDebugLoc(CurInst->getDebugLoc());
2629 CurInst->replaceAllUsesWith(Phi);
2630 if (Phi->getType()->getScalarType()->isPointerTy()) {
2631 // Because we have added a PHI-use of the pointer value, it has now
2632 // "escaped" from alias analysis' perspective. We need to inform
2634 for (unsigned ii = 0, ee = Phi->getNumIncomingValues(); ii != ee;
2636 unsigned jj = PHINode::getOperandNumForIncomingValue(ii);
2637 VN.getAliasAnalysis()->addEscapingUse(Phi->getOperandUse(jj));
2641 MD->invalidateCachedPointerInfo(Phi);
2644 removeFromLeaderTable(ValNo, CurInst, CurrentBlock);
2646 DEBUG(dbgs() << "GVN PRE removed: " << *CurInst << '\n');
2647 if (MD) MD->removeInstruction(CurInst);
2648 DEBUG(verifyRemoved(CurInst));
2649 CurInst->eraseFromParent();
2654 if (splitCriticalEdges())
2660 /// Split the critical edge connecting the given two blocks, and return
2661 /// the block inserted to the critical edge.
2662 BasicBlock *GVN::splitCriticalEdges(BasicBlock *Pred, BasicBlock *Succ) {
2663 BasicBlock *BB = SplitCriticalEdge(Pred, Succ, this);
2665 MD->invalidateCachedPredecessors();
2669 /// splitCriticalEdges - Split critical edges found during the previous
2670 /// iteration that may enable further optimization.
2671 bool GVN::splitCriticalEdges() {
2672 if (toSplit.empty())
2675 std::pair<TerminatorInst*, unsigned> Edge = toSplit.pop_back_val();
2676 SplitCriticalEdge(Edge.first, Edge.second, this);
2677 } while (!toSplit.empty());
2678 if (MD) MD->invalidateCachedPredecessors();
2682 /// iterateOnFunction - Executes one iteration of GVN
2683 bool GVN::iterateOnFunction(Function &F) {
2684 cleanupGlobalSets();
2686 // Top-down walk of the dominator tree
2687 bool Changed = false;
2689 // Needed for value numbering with phi construction to work.
2690 ReversePostOrderTraversal<Function*> RPOT(&F);
2691 for (ReversePostOrderTraversal<Function*>::rpo_iterator RI = RPOT.begin(),
2692 RE = RPOT.end(); RI != RE; ++RI)
2693 Changed |= processBlock(*RI);
2695 // Save the blocks this function have before transformation begins. GVN may
2696 // split critical edge, and hence may invalidate the RPO/DT iterator.
2698 std::vector<BasicBlock *> BBVect;
2699 BBVect.reserve(256);
2700 for (df_iterator<DomTreeNode*> DI = df_begin(DT->getRootNode()),
2701 DE = df_end(DT->getRootNode()); DI != DE; ++DI)
2702 BBVect.push_back(DI->getBlock());
2704 for (std::vector<BasicBlock *>::iterator I = BBVect.begin(), E = BBVect.end();
2706 Changed |= processBlock(*I);
2712 void GVN::cleanupGlobalSets() {
2714 LeaderTable.clear();
2715 TableAllocator.Reset();
2718 /// verifyRemoved - Verify that the specified instruction does not occur in our
2719 /// internal data structures.
2720 void GVN::verifyRemoved(const Instruction *Inst) const {
2721 VN.verifyRemoved(Inst);
2723 // Walk through the value number scope to make sure the instruction isn't
2724 // ferreted away in it.
2725 for (DenseMap<uint32_t, LeaderTableEntry>::const_iterator
2726 I = LeaderTable.begin(), E = LeaderTable.end(); I != E; ++I) {
2727 const LeaderTableEntry *Node = &I->second;
2728 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2730 while (Node->Next) {
2732 assert(Node->Val != Inst && "Inst still in value numbering scope!");
2737 // BB is declared dead, which implied other blocks become dead as well. This
2738 // function is to add all these blocks to "DeadBlocks". For the dead blocks'
2739 // live successors, update their phi nodes by replacing the operands
2740 // corresponding to dead blocks with UndefVal.
2742 void GVN::addDeadBlock(BasicBlock *BB) {
2743 SmallVector<BasicBlock *, 4> NewDead;
2744 SmallSetVector<BasicBlock *, 4> DF;
2746 NewDead.push_back(BB);
2747 while (!NewDead.empty()) {
2748 BasicBlock *D = NewDead.pop_back_val();
2749 if (DeadBlocks.count(D))
2752 // All blocks dominated by D are dead.
2753 SmallVector<BasicBlock *, 8> Dom;
2754 DT->getDescendants(D, Dom);
2755 DeadBlocks.insert(Dom.begin(), Dom.end());
2757 // Figure out the dominance-frontier(D).
2758 for (SmallVectorImpl<BasicBlock *>::iterator I = Dom.begin(),
2759 E = Dom.end(); I != E; I++) {
2761 for (succ_iterator SI = succ_begin(B), SE = succ_end(B); SI != SE; SI++) {
2762 BasicBlock *S = *SI;
2763 if (DeadBlocks.count(S))
2766 bool AllPredDead = true;
2767 for (pred_iterator PI = pred_begin(S), PE = pred_end(S); PI != PE; PI++)
2768 if (!DeadBlocks.count(*PI)) {
2769 AllPredDead = false;
2774 // S could be proved dead later on. That is why we don't update phi
2775 // operands at this moment.
2778 // While S is not dominated by D, it is dead by now. This could take
2779 // place if S already have a dead predecessor before D is declared
2781 NewDead.push_back(S);
2787 // For the dead blocks' live successors, update their phi nodes by replacing
2788 // the operands corresponding to dead blocks with UndefVal.
2789 for(SmallSetVector<BasicBlock *, 4>::iterator I = DF.begin(), E = DF.end();
2792 if (DeadBlocks.count(B))
2795 SmallVector<BasicBlock *, 4> Preds(pred_begin(B), pred_end(B));
2796 for (SmallVectorImpl<BasicBlock *>::iterator PI = Preds.begin(),
2797 PE = Preds.end(); PI != PE; PI++) {
2798 BasicBlock *P = *PI;
2800 if (!DeadBlocks.count(P))
2803 if (isCriticalEdge(P->getTerminator(), GetSuccessorNumber(P, B))) {
2804 if (BasicBlock *S = splitCriticalEdges(P, B))
2805 DeadBlocks.insert(P = S);
2808 for (BasicBlock::iterator II = B->begin(); isa<PHINode>(II); ++II) {
2809 PHINode &Phi = cast<PHINode>(*II);
2810 Phi.setIncomingValue(Phi.getBasicBlockIndex(P),
2811 UndefValue::get(Phi.getType()));
2817 // If the given branch is recognized as a foldable branch (i.e. conditional
2818 // branch with constant condition), it will perform following analyses and
2820 // 1) If the dead out-coming edge is a critical-edge, split it. Let
2821 // R be the target of the dead out-coming edge.
2822 // 1) Identify the set of dead blocks implied by the branch's dead outcoming
2823 // edge. The result of this step will be {X| X is dominated by R}
2824 // 2) Identify those blocks which haves at least one dead prodecessor. The
2825 // result of this step will be dominance-frontier(R).
2826 // 3) Update the PHIs in DF(R) by replacing the operands corresponding to
2827 // dead blocks with "UndefVal" in an hope these PHIs will optimized away.
2829 // Return true iff *NEW* dead code are found.
2830 bool GVN::processFoldableCondBr(BranchInst *BI) {
2831 if (!BI || BI->isUnconditional())
2834 ConstantInt *Cond = dyn_cast<ConstantInt>(BI->getCondition());
2838 BasicBlock *DeadRoot = Cond->getZExtValue() ?
2839 BI->getSuccessor(1) : BI->getSuccessor(0);
2840 if (DeadBlocks.count(DeadRoot))
2843 if (!DeadRoot->getSinglePredecessor())
2844 DeadRoot = splitCriticalEdges(BI->getParent(), DeadRoot);
2846 addDeadBlock(DeadRoot);
2850 // performPRE() will trigger assert if it come across an instruciton without
2851 // associated val-num. As it normally has far more live instructions than dead
2852 // instructions, it makes more sense just to "fabricate" a val-number for the
2853 // dead code than checking if instruction involved is dead or not.
2854 void GVN::assignValNumForDeadCode() {
2855 for (SetVector<BasicBlock *>::iterator I = DeadBlocks.begin(),
2856 E = DeadBlocks.end(); I != E; I++) {
2857 BasicBlock *BB = *I;
2858 for (BasicBlock::iterator II = BB->begin(), EE = BB->end();
2860 Instruction *Inst = &*II;
2861 unsigned ValNum = VN.lookup_or_add(Inst);
2862 addToLeaderTable(ValNum, Inst, BB);